xref: /titanic_41/usr/src/uts/sfmmu/vm/hat_sfmmu.c (revision c39526b769298791ff5b0b6c5e761f49aabaeb4e)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * VM - Hardware Address Translation management for Spitfire MMU.
28  *
29  * This file implements the machine specific hardware translation
30  * needed by the VM system.  The machine independent interface is
31  * described in <vm/hat.h> while the machine dependent interface
32  * and data structures are described in <vm/hat_sfmmu.h>.
33  *
34  * The hat layer manages the address translation hardware as a cache
35  * driven by calls from the higher levels in the VM system.
36  */
37 
38 #include <sys/types.h>
39 #include <sys/kstat.h>
40 #include <vm/hat.h>
41 #include <vm/hat_sfmmu.h>
42 #include <vm/page.h>
43 #include <sys/pte.h>
44 #include <sys/systm.h>
45 #include <sys/mman.h>
46 #include <sys/sysmacros.h>
47 #include <sys/machparam.h>
48 #include <sys/vtrace.h>
49 #include <sys/kmem.h>
50 #include <sys/mmu.h>
51 #include <sys/cmn_err.h>
52 #include <sys/cpu.h>
53 #include <sys/cpuvar.h>
54 #include <sys/debug.h>
55 #include <sys/lgrp.h>
56 #include <sys/archsystm.h>
57 #include <sys/machsystm.h>
58 #include <sys/vmsystm.h>
59 #include <vm/as.h>
60 #include <vm/seg.h>
61 #include <vm/seg_kp.h>
62 #include <vm/seg_kmem.h>
63 #include <vm/seg_kpm.h>
64 #include <vm/rm.h>
65 #include <sys/t_lock.h>
66 #include <sys/obpdefs.h>
67 #include <sys/vm_machparam.h>
68 #include <sys/var.h>
69 #include <sys/trap.h>
70 #include <sys/machtrap.h>
71 #include <sys/scb.h>
72 #include <sys/bitmap.h>
73 #include <sys/machlock.h>
74 #include <sys/membar.h>
75 #include <sys/atomic.h>
76 #include <sys/cpu_module.h>
77 #include <sys/prom_debug.h>
78 #include <sys/ksynch.h>
79 #include <sys/mem_config.h>
80 #include <sys/mem_cage.h>
81 #include <vm/vm_dep.h>
82 #include <vm/xhat_sfmmu.h>
83 #include <sys/fpu/fpusystm.h>
84 #include <vm/mach_kpm.h>
85 #include <sys/callb.h>
86 
87 #ifdef	DEBUG
88 #define	SFMMU_VALIDATE_HMERID(hat, rid, saddr, len)			\
89 	if (SFMMU_IS_SHMERID_VALID(rid)) {				\
90 		caddr_t _eaddr = (saddr) + (len);			\
91 		sf_srd_t *_srdp;					\
92 		sf_region_t *_rgnp;					\
93 		ASSERT((rid) < SFMMU_MAX_HME_REGIONS);			\
94 		ASSERT(SF_RGNMAP_TEST(hat->sfmmu_hmeregion_map, rid));	\
95 		ASSERT((hat) != ksfmmup);				\
96 		_srdp = (hat)->sfmmu_srdp;				\
97 		ASSERT(_srdp != NULL);					\
98 		ASSERT(_srdp->srd_refcnt != 0);				\
99 		_rgnp = _srdp->srd_hmergnp[(rid)];			\
100 		ASSERT(_rgnp != NULL && _rgnp->rgn_id == rid);		\
101 		ASSERT(_rgnp->rgn_refcnt != 0);				\
102 		ASSERT(!(_rgnp->rgn_flags & SFMMU_REGION_FREE));	\
103 		ASSERT((_rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) ==	\
104 		    SFMMU_REGION_HME);					\
105 		ASSERT((saddr) >= _rgnp->rgn_saddr);			\
106 		ASSERT((saddr) < _rgnp->rgn_saddr + _rgnp->rgn_size);	\
107 		ASSERT(_eaddr > _rgnp->rgn_saddr);			\
108 		ASSERT(_eaddr <= _rgnp->rgn_saddr + _rgnp->rgn_size);	\
109 	}
110 
111 #define	SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid) 	 	 \
112 {						 			 \
113 		caddr_t _hsva;						 \
114 		caddr_t _heva;						 \
115 		caddr_t _rsva;					 	 \
116 		caddr_t _reva;					 	 \
117 		int	_ttesz = get_hblk_ttesz(hmeblkp);		 \
118 		int	_flagtte;					 \
119 		ASSERT((srdp)->srd_refcnt != 0);			 \
120 		ASSERT((rid) < SFMMU_MAX_HME_REGIONS);			 \
121 		ASSERT((rgnp)->rgn_id == rid);				 \
122 		ASSERT(!((rgnp)->rgn_flags & SFMMU_REGION_FREE));	 \
123 		ASSERT(((rgnp)->rgn_flags & SFMMU_REGION_TYPE_MASK) ==	 \
124 		    SFMMU_REGION_HME);					 \
125 		ASSERT(_ttesz <= (rgnp)->rgn_pgszc);			 \
126 		_hsva = (caddr_t)get_hblk_base(hmeblkp);		 \
127 		_heva = get_hblk_endaddr(hmeblkp);			 \
128 		_rsva = (caddr_t)P2ALIGN(				 \
129 		    (uintptr_t)(rgnp)->rgn_saddr, HBLK_MIN_BYTES);	 \
130 		_reva = (caddr_t)P2ROUNDUP(				 \
131 		    (uintptr_t)((rgnp)->rgn_saddr + (rgnp)->rgn_size),	 \
132 		    HBLK_MIN_BYTES);					 \
133 		ASSERT(_hsva >= _rsva);				 	 \
134 		ASSERT(_hsva < _reva);				 	 \
135 		ASSERT(_heva > _rsva);				 	 \
136 		ASSERT(_heva <= _reva);				 	 \
137 		_flagtte = (_ttesz < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ :  \
138 			_ttesz;						 \
139 		ASSERT(rgnp->rgn_hmeflags & (0x1 << _flagtte));		 \
140 }
141 
142 #else /* DEBUG */
143 #define	SFMMU_VALIDATE_HMERID(hat, rid, addr, len)
144 #define	SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid)
145 #endif /* DEBUG */
146 
147 #if defined(SF_ERRATA_57)
148 extern caddr_t errata57_limit;
149 #endif
150 
151 #define	HME8BLK_SZ_RND		((roundup(HME8BLK_SZ, sizeof (int64_t))) /  \
152 				(sizeof (int64_t)))
153 #define	HBLK_RESERVE		((struct hme_blk *)hblk_reserve)
154 
155 #define	HBLK_RESERVE_CNT	128
156 #define	HBLK_RESERVE_MIN	20
157 
158 static struct hme_blk		*freehblkp;
159 static kmutex_t			freehblkp_lock;
160 static int			freehblkcnt;
161 
162 static int64_t			hblk_reserve[HME8BLK_SZ_RND];
163 static kmutex_t			hblk_reserve_lock;
164 static kthread_t		*hblk_reserve_thread;
165 
166 static nucleus_hblk8_info_t	nucleus_hblk8;
167 static nucleus_hblk1_info_t	nucleus_hblk1;
168 
169 /*
170  * Data to manage per-cpu hmeblk pending queues, hmeblks are queued here
171  * after the initial phase of removing an hmeblk from the hash chain, see
172  * the detailed comment in sfmmu_hblk_hash_rm() for further details.
173  */
174 static cpu_hme_pend_t		*cpu_hme_pend;
175 static uint_t			cpu_hme_pend_thresh;
176 /*
177  * SFMMU specific hat functions
178  */
179 void	hat_pagecachectl(struct page *, int);
180 
181 /* flags for hat_pagecachectl */
182 #define	HAT_CACHE	0x1
183 #define	HAT_UNCACHE	0x2
184 #define	HAT_TMPNC	0x4
185 
186 /*
187  * Flag to allow the creation of non-cacheable translations
188  * to system memory. It is off by default. At the moment this
189  * flag is used by the ecache error injector. The error injector
190  * will turn it on when creating such a translation then shut it
191  * off when it's finished.
192  */
193 
194 int	sfmmu_allow_nc_trans = 0;
195 
196 /*
197  * Flag to disable large page support.
198  * 	value of 1 => disable all large pages.
199  *	bits 1, 2, and 3 are to disable 64K, 512K and 4M pages respectively.
200  *
201  * For example, use the value 0x4 to disable 512K pages.
202  *
203  */
204 #define	LARGE_PAGES_OFF		0x1
205 
206 /*
207  * The disable_large_pages and disable_ism_large_pages variables control
208  * hat_memload_array and the page sizes to be used by ISM and the kernel.
209  *
210  * The disable_auto_data_large_pages and disable_auto_text_large_pages variables
211  * are only used to control which OOB pages to use at upper VM segment creation
212  * time, and are set in hat_init_pagesizes and used in the map_pgsz* routines.
213  * Their values may come from platform or CPU specific code to disable page
214  * sizes that should not be used.
215  *
216  * WARNING: 512K pages are currently not supported for ISM/DISM.
217  */
218 uint_t	disable_large_pages = 0;
219 uint_t	disable_ism_large_pages = (1 << TTE512K);
220 uint_t	disable_auto_data_large_pages = 0;
221 uint_t	disable_auto_text_large_pages = 0;
222 
223 /*
224  * Private sfmmu data structures for hat management
225  */
226 static struct kmem_cache *sfmmuid_cache;
227 static struct kmem_cache *mmuctxdom_cache;
228 
229 /*
230  * Private sfmmu data structures for tsb management
231  */
232 static struct kmem_cache *sfmmu_tsbinfo_cache;
233 static struct kmem_cache *sfmmu_tsb8k_cache;
234 static struct kmem_cache *sfmmu_tsb_cache[NLGRPS_MAX];
235 static vmem_t *kmem_bigtsb_arena;
236 static vmem_t *kmem_tsb_arena;
237 
238 /*
239  * sfmmu static variables for hmeblk resource management.
240  */
241 static vmem_t *hat_memload1_arena; /* HAT translation arena for sfmmu1_cache */
242 static struct kmem_cache *sfmmu8_cache;
243 static struct kmem_cache *sfmmu1_cache;
244 static struct kmem_cache *pa_hment_cache;
245 
246 static kmutex_t 	ism_mlist_lock;	/* mutex for ism mapping list */
247 /*
248  * private data for ism
249  */
250 static struct kmem_cache *ism_blk_cache;
251 static struct kmem_cache *ism_ment_cache;
252 #define	ISMID_STARTADDR	NULL
253 
254 /*
255  * Region management data structures and function declarations.
256  */
257 
258 static void	sfmmu_leave_srd(sfmmu_t *);
259 static int	sfmmu_srdcache_constructor(void *, void *, int);
260 static void	sfmmu_srdcache_destructor(void *, void *);
261 static int	sfmmu_rgncache_constructor(void *, void *, int);
262 static void	sfmmu_rgncache_destructor(void *, void *);
263 static int	sfrgnmap_isnull(sf_region_map_t *);
264 static int	sfhmergnmap_isnull(sf_hmeregion_map_t *);
265 static int	sfmmu_scdcache_constructor(void *, void *, int);
266 static void	sfmmu_scdcache_destructor(void *, void *);
267 static void	sfmmu_rgn_cb_noop(caddr_t, caddr_t, caddr_t,
268     size_t, void *, u_offset_t);
269 
270 static uint_t srd_hashmask = SFMMU_MAX_SRD_BUCKETS - 1;
271 static sf_srd_bucket_t *srd_buckets;
272 static struct kmem_cache *srd_cache;
273 static uint_t srd_rgn_hashmask = SFMMU_MAX_REGION_BUCKETS - 1;
274 static struct kmem_cache *region_cache;
275 static struct kmem_cache *scd_cache;
276 
277 #ifdef sun4v
278 int use_bigtsb_arena = 1;
279 #else
280 int use_bigtsb_arena = 0;
281 #endif
282 
283 /* External /etc/system tunable, for turning on&off the shctx support */
284 int disable_shctx = 0;
285 /* Internal variable, set by MD if the HW supports shctx feature */
286 int shctx_on = 0;
287 
288 #ifdef DEBUG
289 static void check_scd_sfmmu_list(sfmmu_t **, sfmmu_t *, int);
290 #endif
291 static void sfmmu_to_scd_list(sfmmu_t **, sfmmu_t *);
292 static void sfmmu_from_scd_list(sfmmu_t **, sfmmu_t *);
293 
294 static sf_scd_t *sfmmu_alloc_scd(sf_srd_t *, sf_region_map_t *);
295 static void sfmmu_find_scd(sfmmu_t *);
296 static void sfmmu_join_scd(sf_scd_t *, sfmmu_t *);
297 static void sfmmu_finish_join_scd(sfmmu_t *);
298 static void sfmmu_leave_scd(sfmmu_t *, uchar_t);
299 static void sfmmu_destroy_scd(sf_srd_t *, sf_scd_t *, sf_region_map_t *);
300 static int sfmmu_alloc_scd_tsbs(sf_srd_t *, sf_scd_t *);
301 static void sfmmu_free_scd_tsbs(sfmmu_t *);
302 static void sfmmu_tsb_inv_ctx(sfmmu_t *);
303 static int find_ism_rid(sfmmu_t *, sfmmu_t *, caddr_t, uint_t *);
304 static void sfmmu_ism_hatflags(sfmmu_t *, int);
305 static int sfmmu_srd_lock_held(sf_srd_t *);
306 static void sfmmu_remove_scd(sf_scd_t **, sf_scd_t *);
307 static void sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *);
308 static void sfmmu_link_scd_to_regions(sf_srd_t *, sf_scd_t *);
309 static void sfmmu_unlink_scd_from_regions(sf_srd_t *, sf_scd_t *);
310 static void sfmmu_link_to_hmeregion(sfmmu_t *, sf_region_t *);
311 static void sfmmu_unlink_from_hmeregion(sfmmu_t *, sf_region_t *);
312 
313 /*
314  * ``hat_lock'' is a hashed mutex lock for protecting sfmmu TSB lists,
315  * HAT flags, synchronizing TLB/TSB coherency, and context management.
316  * The lock is hashed on the sfmmup since the case where we need to lock
317  * all processes is rare but does occur (e.g. we need to unload a shared
318  * mapping from all processes using the mapping).  We have a lot of buckets,
319  * and each slab of sfmmu_t's can use about a quarter of them, giving us
320  * a fairly good distribution without wasting too much space and overhead
321  * when we have to grab them all.
322  */
323 #define	SFMMU_NUM_LOCK	128		/* must be power of two */
324 hatlock_t	hat_lock[SFMMU_NUM_LOCK];
325 
326 /*
327  * Hash algorithm optimized for a small number of slabs.
328  *  7 is (highbit((sizeof sfmmu_t)) - 1)
329  * This hash algorithm is based upon the knowledge that sfmmu_t's come from a
330  * kmem_cache, and thus they will be sequential within that cache.  In
331  * addition, each new slab will have a different "color" up to cache_maxcolor
332  * which will skew the hashing for each successive slab which is allocated.
333  * If the size of sfmmu_t changed to a larger size, this algorithm may need
334  * to be revisited.
335  */
336 #define	TSB_HASH_SHIFT_BITS (7)
337 #define	PTR_HASH(x) ((uintptr_t)x >> TSB_HASH_SHIFT_BITS)
338 
339 #ifdef DEBUG
340 int tsb_hash_debug = 0;
341 #define	TSB_HASH(sfmmup)	\
342 	(tsb_hash_debug ? &hat_lock[0] : \
343 	&hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)])
344 #else	/* DEBUG */
345 #define	TSB_HASH(sfmmup)	&hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)]
346 #endif	/* DEBUG */
347 
348 
349 /* sfmmu_replace_tsb() return codes. */
350 typedef enum tsb_replace_rc {
351 	TSB_SUCCESS,
352 	TSB_ALLOCFAIL,
353 	TSB_LOSTRACE,
354 	TSB_ALREADY_SWAPPED,
355 	TSB_CANTGROW
356 } tsb_replace_rc_t;
357 
358 /*
359  * Flags for TSB allocation routines.
360  */
361 #define	TSB_ALLOC	0x01
362 #define	TSB_FORCEALLOC	0x02
363 #define	TSB_GROW	0x04
364 #define	TSB_SHRINK	0x08
365 #define	TSB_SWAPIN	0x10
366 
367 /*
368  * Support for HAT callbacks.
369  */
370 #define	SFMMU_MAX_RELOC_CALLBACKS	10
371 int sfmmu_max_cb_id = SFMMU_MAX_RELOC_CALLBACKS;
372 static id_t sfmmu_cb_nextid = 0;
373 static id_t sfmmu_tsb_cb_id;
374 struct sfmmu_callback *sfmmu_cb_table;
375 
376 /*
377  * Kernel page relocation is enabled by default for non-caged
378  * kernel pages.  This has little effect unless segkmem_reloc is
379  * set, since by default kernel memory comes from inside the
380  * kernel cage.
381  */
382 int hat_kpr_enabled = 1;
383 
384 kmutex_t	kpr_mutex;
385 kmutex_t	kpr_suspendlock;
386 kthread_t	*kreloc_thread;
387 
388 /*
389  * Enable VA->PA translation sanity checking on DEBUG kernels.
390  * Disabled by default.  This is incompatible with some
391  * drivers (error injector, RSM) so if it breaks you get
392  * to keep both pieces.
393  */
394 int hat_check_vtop = 0;
395 
396 /*
397  * Private sfmmu routines (prototypes)
398  */
399 static struct hme_blk *sfmmu_shadow_hcreate(sfmmu_t *, caddr_t, int, uint_t);
400 static struct 	hme_blk *sfmmu_hblk_alloc(sfmmu_t *, caddr_t,
401 			struct hmehash_bucket *, uint_t, hmeblk_tag, uint_t,
402 			uint_t);
403 static caddr_t	sfmmu_hblk_unload(struct hat *, struct hme_blk *, caddr_t,
404 			caddr_t, demap_range_t *, uint_t);
405 static caddr_t	sfmmu_hblk_sync(struct hat *, struct hme_blk *, caddr_t,
406 			caddr_t, int);
407 static void	sfmmu_hblk_free(struct hme_blk **);
408 static void	sfmmu_hblks_list_purge(struct hme_blk **, int);
409 static uint_t	sfmmu_get_free_hblk(struct hme_blk **, uint_t);
410 static uint_t	sfmmu_put_free_hblk(struct hme_blk *, uint_t);
411 static struct hme_blk *sfmmu_hblk_steal(int);
412 static int	sfmmu_steal_this_hblk(struct hmehash_bucket *,
413 			struct hme_blk *, uint64_t, struct hme_blk *);
414 static caddr_t	sfmmu_hblk_unlock(struct hme_blk *, caddr_t, caddr_t);
415 
416 static void	hat_do_memload_array(struct hat *, caddr_t, size_t,
417 		    struct page **, uint_t, uint_t, uint_t);
418 static void	hat_do_memload(struct hat *, caddr_t, struct page *,
419 		    uint_t, uint_t, uint_t);
420 static void	sfmmu_memload_batchsmall(struct hat *, caddr_t, page_t **,
421 		    uint_t, uint_t, pgcnt_t, uint_t);
422 void		sfmmu_tteload(struct hat *, tte_t *, caddr_t, page_t *,
423 			uint_t);
424 static int	sfmmu_tteload_array(sfmmu_t *, tte_t *, caddr_t, page_t **,
425 			uint_t, uint_t);
426 static struct hmehash_bucket *sfmmu_tteload_acquire_hashbucket(sfmmu_t *,
427 					caddr_t, int, uint_t);
428 static struct hme_blk *sfmmu_tteload_find_hmeblk(sfmmu_t *,
429 			struct hmehash_bucket *, caddr_t, uint_t, uint_t,
430 			uint_t);
431 static int	sfmmu_tteload_addentry(sfmmu_t *, struct hme_blk *, tte_t *,
432 			caddr_t, page_t **, uint_t, uint_t);
433 static void	sfmmu_tteload_release_hashbucket(struct hmehash_bucket *);
434 
435 static int	sfmmu_pagearray_setup(caddr_t, page_t **, tte_t *, int);
436 static pfn_t	sfmmu_uvatopfn(caddr_t, sfmmu_t *, tte_t *);
437 void		sfmmu_memtte(tte_t *, pfn_t, uint_t, int);
438 #ifdef VAC
439 static void	sfmmu_vac_conflict(struct hat *, caddr_t, page_t *);
440 static int	sfmmu_vacconflict_array(caddr_t, page_t *, int *);
441 int	tst_tnc(page_t *pp, pgcnt_t);
442 void	conv_tnc(page_t *pp, int);
443 #endif
444 
445 static void	sfmmu_get_ctx(sfmmu_t *);
446 static void	sfmmu_free_sfmmu(sfmmu_t *);
447 
448 static void	sfmmu_ttesync(struct hat *, caddr_t, tte_t *, page_t *);
449 static void	sfmmu_chgattr(struct hat *, caddr_t, size_t, uint_t, int);
450 
451 cpuset_t	sfmmu_pageunload(page_t *, struct sf_hment *, int);
452 static void	hat_pagereload(struct page *, struct page *);
453 static cpuset_t	sfmmu_pagesync(page_t *, struct sf_hment *, uint_t);
454 #ifdef VAC
455 void	sfmmu_page_cache_array(page_t *, int, int, pgcnt_t);
456 static void	sfmmu_page_cache(page_t *, int, int, int);
457 #endif
458 
459 cpuset_t	sfmmu_rgntlb_demap(caddr_t, sf_region_t *,
460     struct hme_blk *, int);
461 static void	sfmmu_tlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *,
462 			pfn_t, int, int, int, int);
463 static void	sfmmu_ismtlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *,
464 			pfn_t, int);
465 static void	sfmmu_tlb_demap(caddr_t, sfmmu_t *, struct hme_blk *, int, int);
466 static void	sfmmu_tlb_range_demap(demap_range_t *);
467 static void	sfmmu_invalidate_ctx(sfmmu_t *);
468 static void	sfmmu_sync_mmustate(sfmmu_t *);
469 
470 static void 	sfmmu_tsbinfo_setup_phys(struct tsb_info *, pfn_t);
471 static int	sfmmu_tsbinfo_alloc(struct tsb_info **, int, int, uint_t,
472 			sfmmu_t *);
473 static void	sfmmu_tsb_free(struct tsb_info *);
474 static void	sfmmu_tsbinfo_free(struct tsb_info *);
475 static int	sfmmu_init_tsbinfo(struct tsb_info *, int, int, uint_t,
476 			sfmmu_t *);
477 static void	sfmmu_tsb_chk_reloc(sfmmu_t *, hatlock_t *);
478 static void	sfmmu_tsb_swapin(sfmmu_t *, hatlock_t *);
479 static int	sfmmu_select_tsb_szc(pgcnt_t);
480 static void	sfmmu_mod_tsb(sfmmu_t *, caddr_t, tte_t *, int);
481 #define		sfmmu_load_tsb(sfmmup, vaddr, tte, szc) \
482 	sfmmu_mod_tsb(sfmmup, vaddr, tte, szc)
483 #define		sfmmu_unload_tsb(sfmmup, vaddr, szc)    \
484 	sfmmu_mod_tsb(sfmmup, vaddr, NULL, szc)
485 static void	sfmmu_copy_tsb(struct tsb_info *, struct tsb_info *);
486 static tsb_replace_rc_t sfmmu_replace_tsb(sfmmu_t *, struct tsb_info *, uint_t,
487     hatlock_t *, uint_t);
488 static void	sfmmu_size_tsb(sfmmu_t *, int, uint64_t, uint64_t, int);
489 
490 #ifdef VAC
491 void	sfmmu_cache_flush(pfn_t, int);
492 void	sfmmu_cache_flushcolor(int, pfn_t);
493 #endif
494 static caddr_t	sfmmu_hblk_chgattr(sfmmu_t *, struct hme_blk *, caddr_t,
495 			caddr_t, demap_range_t *, uint_t, int);
496 
497 static uint64_t	sfmmu_vtop_attr(uint_t, int mode, tte_t *);
498 static uint_t	sfmmu_ptov_attr(tte_t *);
499 static caddr_t	sfmmu_hblk_chgprot(sfmmu_t *, struct hme_blk *, caddr_t,
500 			caddr_t, demap_range_t *, uint_t);
501 static uint_t	sfmmu_vtop_prot(uint_t, uint_t *);
502 static int	sfmmu_idcache_constructor(void *, void *, int);
503 static void	sfmmu_idcache_destructor(void *, void *);
504 static int	sfmmu_hblkcache_constructor(void *, void *, int);
505 static void	sfmmu_hblkcache_destructor(void *, void *);
506 static void	sfmmu_hblkcache_reclaim(void *);
507 static void	sfmmu_shadow_hcleanup(sfmmu_t *, struct hme_blk *,
508 			struct hmehash_bucket *);
509 static void	sfmmu_hblk_hash_rm(struct hmehash_bucket *, struct hme_blk *,
510 			struct hme_blk *, struct hme_blk **, int);
511 static void	sfmmu_hblk_hash_add(struct hmehash_bucket *, struct hme_blk *,
512 			uint64_t);
513 static struct hme_blk *sfmmu_check_pending_hblks(int);
514 static void	sfmmu_free_hblks(sfmmu_t *, caddr_t, caddr_t, int);
515 static void	sfmmu_cleanup_rhblk(sf_srd_t *, caddr_t, uint_t, int);
516 static void	sfmmu_unload_hmeregion_va(sf_srd_t *, uint_t, caddr_t, caddr_t,
517 			int, caddr_t *);
518 static void	sfmmu_unload_hmeregion(sf_srd_t *, sf_region_t *);
519 
520 static void	sfmmu_rm_large_mappings(page_t *, int);
521 
522 static void	hat_lock_init(void);
523 static void	hat_kstat_init(void);
524 static int	sfmmu_kstat_percpu_update(kstat_t *ksp, int rw);
525 static void	sfmmu_set_scd_rttecnt(sf_srd_t *, sf_scd_t *);
526 static	int	sfmmu_is_rgnva(sf_srd_t *, caddr_t, ulong_t, ulong_t);
527 static void	sfmmu_check_page_sizes(sfmmu_t *, int);
528 int	fnd_mapping_sz(page_t *);
529 static void	iment_add(struct ism_ment *,  struct hat *);
530 static void	iment_sub(struct ism_ment *, struct hat *);
531 static pgcnt_t	ism_tsb_entries(sfmmu_t *, int szc);
532 extern void	sfmmu_setup_tsbinfo(sfmmu_t *);
533 extern void	sfmmu_clear_utsbinfo(void);
534 
535 static void		sfmmu_ctx_wrap_around(mmu_ctx_t *, boolean_t);
536 
537 extern int vpm_enable;
538 
539 /* kpm globals */
540 #ifdef	DEBUG
541 /*
542  * Enable trap level tsbmiss handling
543  */
544 int	kpm_tsbmtl = 1;
545 
546 /*
547  * Flush the TLB on kpm mapout. Note: Xcalls are used (again) for the
548  * required TLB shootdowns in this case, so handle w/ care. Off by default.
549  */
550 int	kpm_tlb_flush;
551 #endif	/* DEBUG */
552 
553 static void	*sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *, size_t, int);
554 
555 #ifdef DEBUG
556 static void	sfmmu_check_hblk_flist();
557 #endif
558 
559 /*
560  * Semi-private sfmmu data structures.  Some of them are initialize in
561  * startup or in hat_init. Some of them are private but accessed by
562  * assembly code or mach_sfmmu.c
563  */
564 struct hmehash_bucket *uhme_hash;	/* user hmeblk hash table */
565 struct hmehash_bucket *khme_hash;	/* kernel hmeblk hash table */
566 uint64_t	uhme_hash_pa;		/* PA of uhme_hash */
567 uint64_t	khme_hash_pa;		/* PA of khme_hash */
568 int 		uhmehash_num;		/* # of buckets in user hash table */
569 int 		khmehash_num;		/* # of buckets in kernel hash table */
570 
571 uint_t		max_mmu_ctxdoms = 0;	/* max context domains in the system */
572 mmu_ctx_t	**mmu_ctxs_tbl;		/* global array of context domains */
573 uint64_t	mmu_saved_gnum = 0;	/* to init incoming MMUs' gnums */
574 
575 #define	DEFAULT_NUM_CTXS_PER_MMU 8192
576 static uint_t	nctxs = DEFAULT_NUM_CTXS_PER_MMU;
577 
578 int		cache;			/* describes system cache */
579 
580 caddr_t		ktsb_base;		/* kernel 8k-indexed tsb base address */
581 uint64_t	ktsb_pbase;		/* kernel 8k-indexed tsb phys address */
582 int		ktsb_szcode;		/* kernel 8k-indexed tsb size code */
583 int		ktsb_sz;		/* kernel 8k-indexed tsb size */
584 
585 caddr_t		ktsb4m_base;		/* kernel 4m-indexed tsb base address */
586 uint64_t	ktsb4m_pbase;		/* kernel 4m-indexed tsb phys address */
587 int		ktsb4m_szcode;		/* kernel 4m-indexed tsb size code */
588 int		ktsb4m_sz;		/* kernel 4m-indexed tsb size */
589 
590 uint64_t	kpm_tsbbase;		/* kernel seg_kpm 4M TSB base address */
591 int		kpm_tsbsz;		/* kernel seg_kpm 4M TSB size code */
592 uint64_t	kpmsm_tsbbase;		/* kernel seg_kpm 8K TSB base address */
593 int		kpmsm_tsbsz;		/* kernel seg_kpm 8K TSB size code */
594 
595 #ifndef sun4v
596 int		utsb_dtlb_ttenum = -1;	/* index in TLB for utsb locked TTE */
597 int		utsb4m_dtlb_ttenum = -1; /* index in TLB for 4M TSB TTE */
598 int		dtlb_resv_ttenum;	/* index in TLB of first reserved TTE */
599 caddr_t		utsb_vabase;		/* reserved kernel virtual memory */
600 caddr_t		utsb4m_vabase;		/* for trap handler TSB accesses */
601 #endif /* sun4v */
602 uint64_t	tsb_alloc_bytes = 0;	/* bytes allocated to TSBs */
603 vmem_t		*kmem_tsb_default_arena[NLGRPS_MAX];	/* For dynamic TSBs */
604 vmem_t		*kmem_bigtsb_default_arena[NLGRPS_MAX]; /* dynamic 256M TSBs */
605 
606 /*
607  * Size to use for TSB slabs.  Future platforms that support page sizes
608  * larger than 4M may wish to change these values, and provide their own
609  * assembly macros for building and decoding the TSB base register contents.
610  * Note disable_large_pages will override the value set here.
611  */
612 static	uint_t tsb_slab_ttesz = TTE4M;
613 size_t	tsb_slab_size = MMU_PAGESIZE4M;
614 uint_t	tsb_slab_shift = MMU_PAGESHIFT4M;
615 /* PFN mask for TTE */
616 size_t	tsb_slab_mask = MMU_PAGEOFFSET4M >> MMU_PAGESHIFT;
617 
618 /*
619  * Size to use for TSB slabs.  These are used only when 256M tsb arenas
620  * exist.
621  */
622 static uint_t	bigtsb_slab_ttesz = TTE256M;
623 static size_t	bigtsb_slab_size = MMU_PAGESIZE256M;
624 static uint_t	bigtsb_slab_shift = MMU_PAGESHIFT256M;
625 /* 256M page alignment for 8K pfn */
626 static size_t	bigtsb_slab_mask = MMU_PAGEOFFSET256M >> MMU_PAGESHIFT;
627 
628 /* largest TSB size to grow to, will be smaller on smaller memory systems */
629 static int	tsb_max_growsize = 0;
630 
631 /*
632  * Tunable parameters dealing with TSB policies.
633  */
634 
635 /*
636  * This undocumented tunable forces all 8K TSBs to be allocated from
637  * the kernel heap rather than from the kmem_tsb_default_arena arenas.
638  */
639 #ifdef	DEBUG
640 int	tsb_forceheap = 0;
641 #endif	/* DEBUG */
642 
643 /*
644  * Decide whether to use per-lgroup arenas, or one global set of
645  * TSB arenas.  The default is not to break up per-lgroup, since
646  * most platforms don't recognize any tangible benefit from it.
647  */
648 int	tsb_lgrp_affinity = 0;
649 
650 /*
651  * Used for growing the TSB based on the process RSS.
652  * tsb_rss_factor is based on the smallest TSB, and is
653  * shifted by the TSB size to determine if we need to grow.
654  * The default will grow the TSB if the number of TTEs for
655  * this page size exceeds 75% of the number of TSB entries,
656  * which should _almost_ eliminate all conflict misses
657  * (at the expense of using up lots and lots of memory).
658  */
659 #define	TSB_RSS_FACTOR		(TSB_ENTRIES(TSB_MIN_SZCODE) * 0.75)
660 #define	SFMMU_RSS_TSBSIZE(tsbszc)	(tsb_rss_factor << tsbszc)
661 #define	SELECT_TSB_SIZECODE(pgcnt) ( \
662 	(enable_tsb_rss_sizing)? sfmmu_select_tsb_szc(pgcnt) : \
663 	default_tsb_size)
664 #define	TSB_OK_SHRINK()	\
665 	(tsb_alloc_bytes > tsb_alloc_hiwater || freemem < desfree)
666 #define	TSB_OK_GROW()	\
667 	(tsb_alloc_bytes < tsb_alloc_hiwater && freemem > desfree)
668 
669 int	enable_tsb_rss_sizing = 1;
670 int	tsb_rss_factor	= (int)TSB_RSS_FACTOR;
671 
672 /* which TSB size code to use for new address spaces or if rss sizing off */
673 int default_tsb_size = TSB_8K_SZCODE;
674 
675 static uint64_t tsb_alloc_hiwater; /* limit TSB reserved memory */
676 uint64_t tsb_alloc_hiwater_factor; /* tsb_alloc_hiwater = physmem / this */
677 #define	TSB_ALLOC_HIWATER_FACTOR_DEFAULT	32
678 
679 #ifdef DEBUG
680 static int tsb_random_size = 0;	/* set to 1 to test random tsb sizes on alloc */
681 static int tsb_grow_stress = 0;	/* if set to 1, keep replacing TSB w/ random */
682 static int tsb_alloc_mtbf = 0;	/* fail allocation every n attempts */
683 static int tsb_alloc_fail_mtbf = 0;
684 static int tsb_alloc_count = 0;
685 #endif /* DEBUG */
686 
687 /* if set to 1, will remap valid TTEs when growing TSB. */
688 int tsb_remap_ttes = 1;
689 
690 /*
691  * If we have more than this many mappings, allocate a second TSB.
692  * This default is chosen because the I/D fully associative TLBs are
693  * assumed to have at least 8 available entries. Platforms with a
694  * larger fully-associative TLB could probably override the default.
695  */
696 
697 #ifdef sun4v
698 int tsb_sectsb_threshold = 0;
699 #else
700 int tsb_sectsb_threshold = 8;
701 #endif
702 
703 /*
704  * kstat data
705  */
706 struct sfmmu_global_stat sfmmu_global_stat;
707 struct sfmmu_tsbsize_stat sfmmu_tsbsize_stat;
708 
709 /*
710  * Global data
711  */
712 sfmmu_t 	*ksfmmup;		/* kernel's hat id */
713 
714 #ifdef DEBUG
715 static void	chk_tte(tte_t *, tte_t *, tte_t *, struct hme_blk *);
716 #endif
717 
718 /* sfmmu locking operations */
719 static kmutex_t *sfmmu_mlspl_enter(struct page *, int);
720 static int	sfmmu_mlspl_held(struct page *, int);
721 
722 kmutex_t *sfmmu_page_enter(page_t *);
723 void	sfmmu_page_exit(kmutex_t *);
724 int	sfmmu_page_spl_held(struct page *);
725 
726 /* sfmmu internal locking operations - accessed directly */
727 static void	sfmmu_mlist_reloc_enter(page_t *, page_t *,
728 				kmutex_t **, kmutex_t **);
729 static void	sfmmu_mlist_reloc_exit(kmutex_t *, kmutex_t *);
730 static hatlock_t *
731 		sfmmu_hat_enter(sfmmu_t *);
732 static hatlock_t *
733 		sfmmu_hat_tryenter(sfmmu_t *);
734 static void	sfmmu_hat_exit(hatlock_t *);
735 static void	sfmmu_hat_lock_all(void);
736 static void	sfmmu_hat_unlock_all(void);
737 static void	sfmmu_ismhat_enter(sfmmu_t *, int);
738 static void	sfmmu_ismhat_exit(sfmmu_t *, int);
739 
740 /*
741  * Array of mutexes protecting a page's mapping list and p_nrm field.
742  *
743  * The hash function looks complicated, but is made up so that:
744  *
745  * "pp" not shifted, so adjacent pp values will hash to different cache lines
746  *  (8 byte alignment * 8 bytes/mutes == 64 byte coherency subblock)
747  *
748  * "pp" >> mml_shift, incorporates more source bits into the hash result
749  *
750  *  "& (mml_table_size - 1), should be faster than using remainder "%"
751  *
752  * Hopefully, mml_table, mml_table_size and mml_shift are all in the same
753  * cacheline, since they get declared next to each other below. We'll trust
754  * ld not to do something random.
755  */
756 #ifdef	DEBUG
757 int mlist_hash_debug = 0;
758 #define	MLIST_HASH(pp)	(mlist_hash_debug ? &mml_table[0] : \
759 	&mml_table[((uintptr_t)(pp) + \
760 	((uintptr_t)(pp) >> mml_shift)) & (mml_table_sz - 1)])
761 #else	/* !DEBUG */
762 #define	MLIST_HASH(pp)   &mml_table[ \
763 	((uintptr_t)(pp) + ((uintptr_t)(pp) >> mml_shift)) & (mml_table_sz - 1)]
764 #endif	/* !DEBUG */
765 
766 kmutex_t		*mml_table;
767 uint_t			mml_table_sz;	/* must be a power of 2 */
768 uint_t			mml_shift;	/* log2(mml_table_sz) + 3 for align */
769 
770 kpm_hlk_t	*kpmp_table;
771 uint_t		kpmp_table_sz;	/* must be a power of 2 */
772 uchar_t		kpmp_shift;
773 
774 kpm_shlk_t	*kpmp_stable;
775 uint_t		kpmp_stable_sz;	/* must be a power of 2 */
776 
777 /*
778  * SPL_HASH was improved to avoid false cache line sharing
779  */
780 #define	SPL_TABLE_SIZE	128
781 #define	SPL_MASK	(SPL_TABLE_SIZE - 1)
782 #define	SPL_SHIFT	7		/* log2(SPL_TABLE_SIZE) */
783 
784 #define	SPL_INDEX(pp) \
785 	((((uintptr_t)(pp) >> SPL_SHIFT) ^ \
786 	((uintptr_t)(pp) >> (SPL_SHIFT << 1))) & \
787 	(SPL_TABLE_SIZE - 1))
788 
789 #define	SPL_HASH(pp)    \
790 	(&sfmmu_page_lock[SPL_INDEX(pp) & SPL_MASK].pad_mutex)
791 
792 static	pad_mutex_t	sfmmu_page_lock[SPL_TABLE_SIZE];
793 
794 
795 /*
796  * hat_unload_callback() will group together callbacks in order
797  * to avoid xt_sync() calls.  This is the maximum size of the group.
798  */
799 #define	MAX_CB_ADDR	32
800 
801 tte_t	hw_tte;
802 static ulong_t sfmmu_dmr_maxbit = DMR_MAXBIT;
803 
804 static char	*mmu_ctx_kstat_names[] = {
805 	"mmu_ctx_tsb_exceptions",
806 	"mmu_ctx_tsb_raise_exception",
807 	"mmu_ctx_wrap_around",
808 };
809 
810 /*
811  * Wrapper for vmem_xalloc since vmem_create only allows limited
812  * parameters for vm_source_alloc functions.  This function allows us
813  * to specify alignment consistent with the size of the object being
814  * allocated.
815  */
816 static void *
817 sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag)
818 {
819 	return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag));
820 }
821 
822 /* Common code for setting tsb_alloc_hiwater. */
823 #define	SFMMU_SET_TSB_ALLOC_HIWATER(pages)	tsb_alloc_hiwater = \
824 		ptob(pages) / tsb_alloc_hiwater_factor
825 
826 /*
827  * Set tsb_max_growsize to allow at most all of physical memory to be mapped by
828  * a single TSB.  physmem is the number of physical pages so we need physmem 8K
829  * TTEs to represent all those physical pages.  We round this up by using
830  * 1<<highbit().  To figure out which size code to use, remember that the size
831  * code is just an amount to shift the smallest TSB size to get the size of
832  * this TSB.  So we subtract that size, TSB_START_SIZE, from highbit() (or
833  * highbit() - 1) to get the size code for the smallest TSB that can represent
834  * all of physical memory, while erring on the side of too much.
835  *
836  * Restrict tsb_max_growsize to make sure that:
837  *	1) TSBs can't grow larger than the TSB slab size
838  *	2) TSBs can't grow larger than UTSB_MAX_SZCODE.
839  */
840 #define	SFMMU_SET_TSB_MAX_GROWSIZE(pages) {				\
841 	int	_i, _szc, _slabszc, _tsbszc;				\
842 									\
843 	_i = highbit(pages);						\
844 	if ((1 << (_i - 1)) == (pages))					\
845 		_i--;		/* 2^n case, round down */              \
846 	_szc = _i - TSB_START_SIZE;					\
847 	_slabszc = bigtsb_slab_shift - (TSB_START_SIZE + TSB_ENTRY_SHIFT); \
848 	_tsbszc = MIN(_szc, _slabszc);                                  \
849 	tsb_max_growsize = MIN(_tsbszc, UTSB_MAX_SZCODE);               \
850 }
851 
852 /*
853  * Given a pointer to an sfmmu and a TTE size code, return a pointer to the
854  * tsb_info which handles that TTE size.
855  */
856 #define	SFMMU_GET_TSBINFO(tsbinfop, sfmmup, tte_szc) {			\
857 	(tsbinfop) = (sfmmup)->sfmmu_tsb;				\
858 	ASSERT(((tsbinfop)->tsb_flags & TSB_SHAREDCTX) ||		\
859 	    sfmmu_hat_lock_held(sfmmup));				\
860 	if ((tte_szc) >= TTE4M)	{					\
861 		ASSERT((tsbinfop) != NULL);				\
862 		(tsbinfop) = (tsbinfop)->tsb_next;			\
863 	}								\
864 }
865 
866 /*
867  * Macro to use to unload entries from the TSB.
868  * It has knowledge of which page sizes get replicated in the TSB
869  * and will call the appropriate unload routine for the appropriate size.
870  */
871 #define	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, ismhat)		\
872 {									\
873 	int ttesz = get_hblk_ttesz(hmeblkp);				\
874 	if (ttesz == TTE8K || ttesz == TTE4M) {				\
875 		sfmmu_unload_tsb(sfmmup, addr, ttesz);			\
876 	} else {							\
877 		caddr_t sva = ismhat ? addr : 				\
878 		    (caddr_t)get_hblk_base(hmeblkp);			\
879 		caddr_t eva = sva + get_hblk_span(hmeblkp);		\
880 		ASSERT(addr >= sva && addr < eva);			\
881 		sfmmu_unload_tsb_range(sfmmup, sva, eva, ttesz);	\
882 	}								\
883 }
884 
885 
886 /* Update tsb_alloc_hiwater after memory is configured. */
887 /*ARGSUSED*/
888 static void
889 sfmmu_update_post_add(void *arg, pgcnt_t delta_pages)
890 {
891 	/* Assumes physmem has already been updated. */
892 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
893 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
894 }
895 
896 /*
897  * Update tsb_alloc_hiwater before memory is deleted.  We'll do nothing here
898  * and update tsb_alloc_hiwater and tsb_max_growsize after the memory is
899  * deleted.
900  */
901 /*ARGSUSED*/
902 static int
903 sfmmu_update_pre_del(void *arg, pgcnt_t delta_pages)
904 {
905 	return (0);
906 }
907 
908 /* Update tsb_alloc_hiwater after memory fails to be unconfigured. */
909 /*ARGSUSED*/
910 static void
911 sfmmu_update_post_del(void *arg, pgcnt_t delta_pages, int cancelled)
912 {
913 	/*
914 	 * Whether the delete was cancelled or not, just go ahead and update
915 	 * tsb_alloc_hiwater and tsb_max_growsize.
916 	 */
917 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
918 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
919 }
920 
921 static kphysm_setup_vector_t sfmmu_update_vec = {
922 	KPHYSM_SETUP_VECTOR_VERSION,	/* version */
923 	sfmmu_update_post_add,		/* post_add */
924 	sfmmu_update_pre_del,		/* pre_del */
925 	sfmmu_update_post_del		/* post_del */
926 };
927 
928 
929 /*
930  * HME_BLK HASH PRIMITIVES
931  */
932 
933 /*
934  * Enter a hme on the mapping list for page pp.
935  * When large pages are more prevalent in the system we might want to
936  * keep the mapping list in ascending order by the hment size. For now,
937  * small pages are more frequent, so don't slow it down.
938  */
939 #define	HME_ADD(hme, pp)					\
940 {								\
941 	ASSERT(sfmmu_mlist_held(pp));				\
942 								\
943 	hme->hme_prev = NULL;					\
944 	hme->hme_next = pp->p_mapping;				\
945 	hme->hme_page = pp;					\
946 	if (pp->p_mapping) {					\
947 		((struct sf_hment *)(pp->p_mapping))->hme_prev = hme;\
948 		ASSERT(pp->p_share > 0);			\
949 	} else  {						\
950 		/* EMPTY */					\
951 		ASSERT(pp->p_share == 0);			\
952 	}							\
953 	pp->p_mapping = hme;					\
954 	pp->p_share++;						\
955 }
956 
957 /*
958  * Enter a hme on the mapping list for page pp.
959  * If we are unmapping a large translation, we need to make sure that the
960  * change is reflect in the corresponding bit of the p_index field.
961  */
962 #define	HME_SUB(hme, pp)					\
963 {								\
964 	ASSERT(sfmmu_mlist_held(pp));				\
965 	ASSERT(hme->hme_page == pp || IS_PAHME(hme));		\
966 								\
967 	if (pp->p_mapping == NULL) {				\
968 		panic("hme_remove - no mappings");		\
969 	}							\
970 								\
971 	membar_stst();	/* ensure previous stores finish */	\
972 								\
973 	ASSERT(pp->p_share > 0);				\
974 	pp->p_share--;						\
975 								\
976 	if (hme->hme_prev) {					\
977 		ASSERT(pp->p_mapping != hme);			\
978 		ASSERT(hme->hme_prev->hme_page == pp ||		\
979 			IS_PAHME(hme->hme_prev));		\
980 		hme->hme_prev->hme_next = hme->hme_next;	\
981 	} else {						\
982 		ASSERT(pp->p_mapping == hme);			\
983 		pp->p_mapping = hme->hme_next;			\
984 		ASSERT((pp->p_mapping == NULL) ?		\
985 			(pp->p_share == 0) : 1);		\
986 	}							\
987 								\
988 	if (hme->hme_next) {					\
989 		ASSERT(hme->hme_next->hme_page == pp ||		\
990 			IS_PAHME(hme->hme_next));		\
991 		hme->hme_next->hme_prev = hme->hme_prev;	\
992 	}							\
993 								\
994 	/* zero out the entry */				\
995 	hme->hme_next = NULL;					\
996 	hme->hme_prev = NULL;					\
997 	hme->hme_page = NULL;					\
998 								\
999 	if (hme_size(hme) > TTE8K) {				\
1000 		/* remove mappings for remainder of large pg */	\
1001 		sfmmu_rm_large_mappings(pp, hme_size(hme));	\
1002 	}							\
1003 }
1004 
1005 /*
1006  * This function returns the hment given the hme_blk and a vaddr.
1007  * It assumes addr has already been checked to belong to hme_blk's
1008  * range.
1009  */
1010 #define	HBLKTOHME(hment, hmeblkp, addr)					\
1011 {									\
1012 	int index;							\
1013 	HBLKTOHME_IDX(hment, hmeblkp, addr, index)			\
1014 }
1015 
1016 /*
1017  * Version of HBLKTOHME that also returns the index in hmeblkp
1018  * of the hment.
1019  */
1020 #define	HBLKTOHME_IDX(hment, hmeblkp, addr, idx)			\
1021 {									\
1022 	ASSERT(in_hblk_range((hmeblkp), (addr)));			\
1023 									\
1024 	if (get_hblk_ttesz(hmeblkp) == TTE8K) {				\
1025 		idx = (((uintptr_t)(addr) >> MMU_PAGESHIFT) & (NHMENTS-1)); \
1026 	} else								\
1027 		idx = 0;						\
1028 									\
1029 	(hment) = &(hmeblkp)->hblk_hme[idx];				\
1030 }
1031 
1032 /*
1033  * Disable any page sizes not supported by the CPU
1034  */
1035 void
1036 hat_init_pagesizes()
1037 {
1038 	int 		i;
1039 
1040 	mmu_exported_page_sizes = 0;
1041 	for (i = TTE8K; i < max_mmu_page_sizes; i++) {
1042 
1043 		szc_2_userszc[i] = (uint_t)-1;
1044 		userszc_2_szc[i] = (uint_t)-1;
1045 
1046 		if ((mmu_exported_pagesize_mask & (1 << i)) == 0) {
1047 			disable_large_pages |= (1 << i);
1048 		} else {
1049 			szc_2_userszc[i] = mmu_exported_page_sizes;
1050 			userszc_2_szc[mmu_exported_page_sizes] = i;
1051 			mmu_exported_page_sizes++;
1052 		}
1053 	}
1054 
1055 	disable_ism_large_pages |= disable_large_pages;
1056 	disable_auto_data_large_pages = disable_large_pages;
1057 	disable_auto_text_large_pages = disable_large_pages;
1058 
1059 	/*
1060 	 * Initialize mmu-specific large page sizes.
1061 	 */
1062 	if (&mmu_large_pages_disabled) {
1063 		disable_large_pages |= mmu_large_pages_disabled(HAT_LOAD);
1064 		disable_ism_large_pages |=
1065 		    mmu_large_pages_disabled(HAT_LOAD_SHARE);
1066 		disable_auto_data_large_pages |=
1067 		    mmu_large_pages_disabled(HAT_AUTO_DATA);
1068 		disable_auto_text_large_pages |=
1069 		    mmu_large_pages_disabled(HAT_AUTO_TEXT);
1070 	}
1071 }
1072 
1073 /*
1074  * Initialize the hardware address translation structures.
1075  */
1076 void
1077 hat_init(void)
1078 {
1079 	int 		i;
1080 	uint_t		sz;
1081 	size_t		size;
1082 
1083 	hat_lock_init();
1084 	hat_kstat_init();
1085 
1086 	/*
1087 	 * Hardware-only bits in a TTE
1088 	 */
1089 	MAKE_TTE_MASK(&hw_tte);
1090 
1091 	hat_init_pagesizes();
1092 
1093 	/* Initialize the hash locks */
1094 	for (i = 0; i < khmehash_num; i++) {
1095 		mutex_init(&khme_hash[i].hmehash_mutex, NULL,
1096 		    MUTEX_DEFAULT, NULL);
1097 		khme_hash[i].hmeh_nextpa = HMEBLK_ENDPA;
1098 	}
1099 	for (i = 0; i < uhmehash_num; i++) {
1100 		mutex_init(&uhme_hash[i].hmehash_mutex, NULL,
1101 		    MUTEX_DEFAULT, NULL);
1102 		uhme_hash[i].hmeh_nextpa = HMEBLK_ENDPA;
1103 	}
1104 	khmehash_num--;		/* make sure counter starts from 0 */
1105 	uhmehash_num--;		/* make sure counter starts from 0 */
1106 
1107 	/*
1108 	 * Allocate context domain structures.
1109 	 *
1110 	 * A platform may choose to modify max_mmu_ctxdoms in
1111 	 * set_platform_defaults(). If a platform does not define
1112 	 * a set_platform_defaults() or does not choose to modify
1113 	 * max_mmu_ctxdoms, it gets one MMU context domain for every CPU.
1114 	 *
1115 	 * For all platforms that have CPUs sharing MMUs, this
1116 	 * value must be defined.
1117 	 */
1118 	if (max_mmu_ctxdoms == 0)
1119 		max_mmu_ctxdoms = max_ncpus;
1120 
1121 	size = max_mmu_ctxdoms * sizeof (mmu_ctx_t *);
1122 	mmu_ctxs_tbl = kmem_zalloc(size, KM_SLEEP);
1123 
1124 	/* mmu_ctx_t is 64 bytes aligned */
1125 	mmuctxdom_cache = kmem_cache_create("mmuctxdom_cache",
1126 	    sizeof (mmu_ctx_t), 64, NULL, NULL, NULL, NULL, NULL, 0);
1127 	/*
1128 	 * MMU context domain initialization for the Boot CPU.
1129 	 * This needs the context domains array allocated above.
1130 	 */
1131 	mutex_enter(&cpu_lock);
1132 	sfmmu_cpu_init(CPU);
1133 	mutex_exit(&cpu_lock);
1134 
1135 	/*
1136 	 * Intialize ism mapping list lock.
1137 	 */
1138 
1139 	mutex_init(&ism_mlist_lock, NULL, MUTEX_DEFAULT, NULL);
1140 
1141 	/*
1142 	 * Each sfmmu structure carries an array of MMU context info
1143 	 * structures, one per context domain. The size of this array depends
1144 	 * on the maximum number of context domains. So, the size of the
1145 	 * sfmmu structure varies per platform.
1146 	 *
1147 	 * sfmmu is allocated from static arena, because trap
1148 	 * handler at TL > 0 is not allowed to touch kernel relocatable
1149 	 * memory. sfmmu's alignment is changed to 64 bytes from
1150 	 * default 8 bytes, as the lower 6 bits will be used to pass
1151 	 * pgcnt to vtag_flush_pgcnt_tl1.
1152 	 */
1153 	size = sizeof (sfmmu_t) + sizeof (sfmmu_ctx_t) * (max_mmu_ctxdoms - 1);
1154 
1155 	sfmmuid_cache = kmem_cache_create("sfmmuid_cache", size,
1156 	    64, sfmmu_idcache_constructor, sfmmu_idcache_destructor,
1157 	    NULL, NULL, static_arena, 0);
1158 
1159 	sfmmu_tsbinfo_cache = kmem_cache_create("sfmmu_tsbinfo_cache",
1160 	    sizeof (struct tsb_info), 0, NULL, NULL, NULL, NULL, NULL, 0);
1161 
1162 	/*
1163 	 * Since we only use the tsb8k cache to "borrow" pages for TSBs
1164 	 * from the heap when low on memory or when TSB_FORCEALLOC is
1165 	 * specified, don't use magazines to cache them--we want to return
1166 	 * them to the system as quickly as possible.
1167 	 */
1168 	sfmmu_tsb8k_cache = kmem_cache_create("sfmmu_tsb8k_cache",
1169 	    MMU_PAGESIZE, MMU_PAGESIZE, NULL, NULL, NULL, NULL,
1170 	    static_arena, KMC_NOMAGAZINE);
1171 
1172 	/*
1173 	 * Set tsb_alloc_hiwater to 1/tsb_alloc_hiwater_factor of physical
1174 	 * memory, which corresponds to the old static reserve for TSBs.
1175 	 * tsb_alloc_hiwater_factor defaults to 32.  This caps the amount of
1176 	 * memory we'll allocate for TSB slabs; beyond this point TSB
1177 	 * allocations will be taken from the kernel heap (via
1178 	 * sfmmu_tsb8k_cache) and will be throttled as would any other kmem
1179 	 * consumer.
1180 	 */
1181 	if (tsb_alloc_hiwater_factor == 0) {
1182 		tsb_alloc_hiwater_factor = TSB_ALLOC_HIWATER_FACTOR_DEFAULT;
1183 	}
1184 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
1185 
1186 	for (sz = tsb_slab_ttesz; sz > 0; sz--) {
1187 		if (!(disable_large_pages & (1 << sz)))
1188 			break;
1189 	}
1190 
1191 	if (sz < tsb_slab_ttesz) {
1192 		tsb_slab_ttesz = sz;
1193 		tsb_slab_shift = MMU_PAGESHIFT + (sz << 1) + sz;
1194 		tsb_slab_size = 1 << tsb_slab_shift;
1195 		tsb_slab_mask = (1 << (tsb_slab_shift - MMU_PAGESHIFT)) - 1;
1196 		use_bigtsb_arena = 0;
1197 	} else if (use_bigtsb_arena &&
1198 	    (disable_large_pages & (1 << bigtsb_slab_ttesz))) {
1199 		use_bigtsb_arena = 0;
1200 	}
1201 
1202 	if (!use_bigtsb_arena) {
1203 		bigtsb_slab_shift = tsb_slab_shift;
1204 	}
1205 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
1206 
1207 	/*
1208 	 * On smaller memory systems, allocate TSB memory in smaller chunks
1209 	 * than the default 4M slab size. We also honor disable_large_pages
1210 	 * here.
1211 	 *
1212 	 * The trap handlers need to be patched with the final slab shift,
1213 	 * since they need to be able to construct the TSB pointer at runtime.
1214 	 */
1215 	if ((tsb_max_growsize <= TSB_512K_SZCODE) &&
1216 	    !(disable_large_pages & (1 << TTE512K))) {
1217 		tsb_slab_ttesz = TTE512K;
1218 		tsb_slab_shift = MMU_PAGESHIFT512K;
1219 		tsb_slab_size = MMU_PAGESIZE512K;
1220 		tsb_slab_mask = MMU_PAGEOFFSET512K >> MMU_PAGESHIFT;
1221 		use_bigtsb_arena = 0;
1222 	}
1223 
1224 	if (!use_bigtsb_arena) {
1225 		bigtsb_slab_ttesz = tsb_slab_ttesz;
1226 		bigtsb_slab_shift = tsb_slab_shift;
1227 		bigtsb_slab_size = tsb_slab_size;
1228 		bigtsb_slab_mask = tsb_slab_mask;
1229 	}
1230 
1231 
1232 	/*
1233 	 * Set up memory callback to update tsb_alloc_hiwater and
1234 	 * tsb_max_growsize.
1235 	 */
1236 	i = kphysm_setup_func_register(&sfmmu_update_vec, (void *) 0);
1237 	ASSERT(i == 0);
1238 
1239 	/*
1240 	 * kmem_tsb_arena is the source from which large TSB slabs are
1241 	 * drawn.  The quantum of this arena corresponds to the largest
1242 	 * TSB size we can dynamically allocate for user processes.
1243 	 * Currently it must also be a supported page size since we
1244 	 * use exactly one translation entry to map each slab page.
1245 	 *
1246 	 * The per-lgroup kmem_tsb_default_arena arenas are the arenas from
1247 	 * which most TSBs are allocated.  Since most TSB allocations are
1248 	 * typically 8K we have a kmem cache we stack on top of each
1249 	 * kmem_tsb_default_arena to speed up those allocations.
1250 	 *
1251 	 * Note the two-level scheme of arenas is required only
1252 	 * because vmem_create doesn't allow us to specify alignment
1253 	 * requirements.  If this ever changes the code could be
1254 	 * simplified to use only one level of arenas.
1255 	 *
1256 	 * If 256M page support exists on sun4v, 256MB kmem_bigtsb_arena
1257 	 * will be provided in addition to the 4M kmem_tsb_arena.
1258 	 */
1259 	if (use_bigtsb_arena) {
1260 		kmem_bigtsb_arena = vmem_create("kmem_bigtsb", NULL, 0,
1261 		    bigtsb_slab_size, sfmmu_vmem_xalloc_aligned_wrapper,
1262 		    vmem_xfree, heap_arena, 0, VM_SLEEP);
1263 	}
1264 
1265 	kmem_tsb_arena = vmem_create("kmem_tsb", NULL, 0, tsb_slab_size,
1266 	    sfmmu_vmem_xalloc_aligned_wrapper,
1267 	    vmem_xfree, heap_arena, 0, VM_SLEEP);
1268 
1269 	if (tsb_lgrp_affinity) {
1270 		char s[50];
1271 		for (i = 0; i < NLGRPS_MAX; i++) {
1272 			if (use_bigtsb_arena) {
1273 				(void) sprintf(s, "kmem_bigtsb_lgrp%d", i);
1274 				kmem_bigtsb_default_arena[i] = vmem_create(s,
1275 				    NULL, 0, 2 * tsb_slab_size,
1276 				    sfmmu_tsb_segkmem_alloc,
1277 				    sfmmu_tsb_segkmem_free, kmem_bigtsb_arena,
1278 				    0, VM_SLEEP | VM_BESTFIT);
1279 			}
1280 
1281 			(void) sprintf(s, "kmem_tsb_lgrp%d", i);
1282 			kmem_tsb_default_arena[i] = vmem_create(s,
1283 			    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1284 			    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1285 			    VM_SLEEP | VM_BESTFIT);
1286 
1287 			(void) sprintf(s, "sfmmu_tsb_lgrp%d_cache", i);
1288 			sfmmu_tsb_cache[i] = kmem_cache_create(s,
1289 			    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1290 			    kmem_tsb_default_arena[i], 0);
1291 		}
1292 	} else {
1293 		if (use_bigtsb_arena) {
1294 			kmem_bigtsb_default_arena[0] =
1295 			    vmem_create("kmem_bigtsb_default", NULL, 0,
1296 			    2 * tsb_slab_size, sfmmu_tsb_segkmem_alloc,
1297 			    sfmmu_tsb_segkmem_free, kmem_bigtsb_arena, 0,
1298 			    VM_SLEEP | VM_BESTFIT);
1299 		}
1300 
1301 		kmem_tsb_default_arena[0] = vmem_create("kmem_tsb_default",
1302 		    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1303 		    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1304 		    VM_SLEEP | VM_BESTFIT);
1305 		sfmmu_tsb_cache[0] = kmem_cache_create("sfmmu_tsb_cache",
1306 		    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1307 		    kmem_tsb_default_arena[0], 0);
1308 	}
1309 
1310 	sfmmu8_cache = kmem_cache_create("sfmmu8_cache", HME8BLK_SZ,
1311 	    HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1312 	    sfmmu_hblkcache_destructor,
1313 	    sfmmu_hblkcache_reclaim, (void *)HME8BLK_SZ,
1314 	    hat_memload_arena, KMC_NOHASH);
1315 
1316 	hat_memload1_arena = vmem_create("hat_memload1", NULL, 0, PAGESIZE,
1317 	    segkmem_alloc_permanent, segkmem_free, heap_arena, 0,
1318 	    VMC_DUMPSAFE | VM_SLEEP);
1319 
1320 	sfmmu1_cache = kmem_cache_create("sfmmu1_cache", HME1BLK_SZ,
1321 	    HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1322 	    sfmmu_hblkcache_destructor,
1323 	    NULL, (void *)HME1BLK_SZ,
1324 	    hat_memload1_arena, KMC_NOHASH);
1325 
1326 	pa_hment_cache = kmem_cache_create("pa_hment_cache", PAHME_SZ,
1327 	    0, NULL, NULL, NULL, NULL, static_arena, KMC_NOHASH);
1328 
1329 	ism_blk_cache = kmem_cache_create("ism_blk_cache",
1330 	    sizeof (ism_blk_t), ecache_alignsize, NULL, NULL,
1331 	    NULL, NULL, static_arena, KMC_NOHASH);
1332 
1333 	ism_ment_cache = kmem_cache_create("ism_ment_cache",
1334 	    sizeof (ism_ment_t), 0, NULL, NULL,
1335 	    NULL, NULL, NULL, 0);
1336 
1337 	/*
1338 	 * We grab the first hat for the kernel,
1339 	 */
1340 	AS_LOCK_ENTER(&kas, &kas.a_lock, RW_WRITER);
1341 	kas.a_hat = hat_alloc(&kas);
1342 	AS_LOCK_EXIT(&kas, &kas.a_lock);
1343 
1344 	/*
1345 	 * Initialize hblk_reserve.
1346 	 */
1347 	((struct hme_blk *)hblk_reserve)->hblk_nextpa =
1348 	    va_to_pa((caddr_t)hblk_reserve);
1349 
1350 #ifndef UTSB_PHYS
1351 	/*
1352 	 * Reserve some kernel virtual address space for the locked TTEs
1353 	 * that allow us to probe the TSB from TL>0.
1354 	 */
1355 	utsb_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1356 	    0, 0, NULL, NULL, VM_SLEEP);
1357 	utsb4m_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1358 	    0, 0, NULL, NULL, VM_SLEEP);
1359 #endif
1360 
1361 #ifdef VAC
1362 	/*
1363 	 * The big page VAC handling code assumes VAC
1364 	 * will not be bigger than the smallest big
1365 	 * page- which is 64K.
1366 	 */
1367 	if (TTEPAGES(TTE64K) < CACHE_NUM_COLOR) {
1368 		cmn_err(CE_PANIC, "VAC too big!");
1369 	}
1370 #endif
1371 
1372 	(void) xhat_init();
1373 
1374 	uhme_hash_pa = va_to_pa(uhme_hash);
1375 	khme_hash_pa = va_to_pa(khme_hash);
1376 
1377 	/*
1378 	 * Initialize relocation locks. kpr_suspendlock is held
1379 	 * at PIL_MAX to prevent interrupts from pinning the holder
1380 	 * of a suspended TTE which may access it leading to a
1381 	 * deadlock condition.
1382 	 */
1383 	mutex_init(&kpr_mutex, NULL, MUTEX_DEFAULT, NULL);
1384 	mutex_init(&kpr_suspendlock, NULL, MUTEX_SPIN, (void *)PIL_MAX);
1385 
1386 	/*
1387 	 * If Shared context support is disabled via /etc/system
1388 	 * set shctx_on to 0 here if it was set to 1 earlier in boot
1389 	 * sequence by cpu module initialization code.
1390 	 */
1391 	if (shctx_on && disable_shctx) {
1392 		shctx_on = 0;
1393 	}
1394 
1395 	if (shctx_on) {
1396 		srd_buckets = kmem_zalloc(SFMMU_MAX_SRD_BUCKETS *
1397 		    sizeof (srd_buckets[0]), KM_SLEEP);
1398 		for (i = 0; i < SFMMU_MAX_SRD_BUCKETS; i++) {
1399 			mutex_init(&srd_buckets[i].srdb_lock, NULL,
1400 			    MUTEX_DEFAULT, NULL);
1401 		}
1402 
1403 		srd_cache = kmem_cache_create("srd_cache", sizeof (sf_srd_t),
1404 		    0, sfmmu_srdcache_constructor, sfmmu_srdcache_destructor,
1405 		    NULL, NULL, NULL, 0);
1406 		region_cache = kmem_cache_create("region_cache",
1407 		    sizeof (sf_region_t), 0, sfmmu_rgncache_constructor,
1408 		    sfmmu_rgncache_destructor, NULL, NULL, NULL, 0);
1409 		scd_cache = kmem_cache_create("scd_cache", sizeof (sf_scd_t),
1410 		    0, sfmmu_scdcache_constructor,  sfmmu_scdcache_destructor,
1411 		    NULL, NULL, NULL, 0);
1412 	}
1413 
1414 	/*
1415 	 * Pre-allocate hrm_hashtab before enabling the collection of
1416 	 * refmod statistics.  Allocating on the fly would mean us
1417 	 * running the risk of suffering recursive mutex enters or
1418 	 * deadlocks.
1419 	 */
1420 	hrm_hashtab = kmem_zalloc(HRM_HASHSIZE * sizeof (struct hrmstat *),
1421 	    KM_SLEEP);
1422 
1423 	/* Allocate per-cpu pending freelist of hmeblks */
1424 	cpu_hme_pend = kmem_zalloc((NCPU * sizeof (cpu_hme_pend_t)) + 64,
1425 	    KM_SLEEP);
1426 	cpu_hme_pend = (cpu_hme_pend_t *)P2ROUNDUP(
1427 	    (uintptr_t)cpu_hme_pend, 64);
1428 
1429 	for (i = 0; i < NCPU; i++) {
1430 		mutex_init(&cpu_hme_pend[i].chp_mutex, NULL, MUTEX_DEFAULT,
1431 		    NULL);
1432 	}
1433 
1434 	if (cpu_hme_pend_thresh == 0) {
1435 		cpu_hme_pend_thresh = CPU_HME_PEND_THRESH;
1436 	}
1437 }
1438 
1439 /*
1440  * Initialize locking for the hat layer, called early during boot.
1441  */
1442 static void
1443 hat_lock_init()
1444 {
1445 	int i;
1446 
1447 	/*
1448 	 * initialize the array of mutexes protecting a page's mapping
1449 	 * list and p_nrm field.
1450 	 */
1451 	for (i = 0; i < mml_table_sz; i++)
1452 		mutex_init(&mml_table[i], NULL, MUTEX_DEFAULT, NULL);
1453 
1454 	if (kpm_enable) {
1455 		for (i = 0; i < kpmp_table_sz; i++) {
1456 			mutex_init(&kpmp_table[i].khl_mutex, NULL,
1457 			    MUTEX_DEFAULT, NULL);
1458 		}
1459 	}
1460 
1461 	/*
1462 	 * Initialize array of mutex locks that protects sfmmu fields and
1463 	 * TSB lists.
1464 	 */
1465 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
1466 		mutex_init(HATLOCK_MUTEXP(&hat_lock[i]), NULL, MUTEX_DEFAULT,
1467 		    NULL);
1468 }
1469 
1470 #define	SFMMU_KERNEL_MAXVA \
1471 	(kmem64_base ? (uintptr_t)kmem64_end : (SYSLIMIT))
1472 
1473 /*
1474  * Allocate a hat structure.
1475  * Called when an address space first uses a hat.
1476  */
1477 struct hat *
1478 hat_alloc(struct as *as)
1479 {
1480 	sfmmu_t *sfmmup;
1481 	int i;
1482 	uint64_t cnum;
1483 	extern uint_t get_color_start(struct as *);
1484 
1485 	ASSERT(AS_WRITE_HELD(as, &as->a_lock));
1486 	sfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
1487 	sfmmup->sfmmu_as = as;
1488 	sfmmup->sfmmu_flags = 0;
1489 	sfmmup->sfmmu_tteflags = 0;
1490 	sfmmup->sfmmu_rtteflags = 0;
1491 	LOCK_INIT_CLEAR(&sfmmup->sfmmu_ctx_lock);
1492 
1493 	if (as == &kas) {
1494 		ksfmmup = sfmmup;
1495 		sfmmup->sfmmu_cext = 0;
1496 		cnum = KCONTEXT;
1497 
1498 		sfmmup->sfmmu_clrstart = 0;
1499 		sfmmup->sfmmu_tsb = NULL;
1500 		/*
1501 		 * hat_kern_setup() will call sfmmu_init_ktsbinfo()
1502 		 * to setup tsb_info for ksfmmup.
1503 		 */
1504 	} else {
1505 
1506 		/*
1507 		 * Just set to invalid ctx. When it faults, it will
1508 		 * get a valid ctx. This would avoid the situation
1509 		 * where we get a ctx, but it gets stolen and then
1510 		 * we fault when we try to run and so have to get
1511 		 * another ctx.
1512 		 */
1513 		sfmmup->sfmmu_cext = 0;
1514 		cnum = INVALID_CONTEXT;
1515 
1516 		/* initialize original physical page coloring bin */
1517 		sfmmup->sfmmu_clrstart = get_color_start(as);
1518 #ifdef DEBUG
1519 		if (tsb_random_size) {
1520 			uint32_t randval = (uint32_t)gettick() >> 4;
1521 			int size = randval % (tsb_max_growsize + 1);
1522 
1523 			/* chose a random tsb size for stress testing */
1524 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb, size,
1525 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1526 		} else
1527 #endif /* DEBUG */
1528 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb,
1529 			    default_tsb_size,
1530 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1531 		sfmmup->sfmmu_flags = HAT_SWAPPED | HAT_ALLCTX_INVALID;
1532 		ASSERT(sfmmup->sfmmu_tsb != NULL);
1533 	}
1534 
1535 	ASSERT(max_mmu_ctxdoms > 0);
1536 	for (i = 0; i < max_mmu_ctxdoms; i++) {
1537 		sfmmup->sfmmu_ctxs[i].cnum = cnum;
1538 		sfmmup->sfmmu_ctxs[i].gnum = 0;
1539 	}
1540 
1541 	for (i = 0; i < max_mmu_page_sizes; i++) {
1542 		sfmmup->sfmmu_ttecnt[i] = 0;
1543 		sfmmup->sfmmu_scdrttecnt[i] = 0;
1544 		sfmmup->sfmmu_ismttecnt[i] = 0;
1545 		sfmmup->sfmmu_scdismttecnt[i] = 0;
1546 		sfmmup->sfmmu_pgsz[i] = TTE8K;
1547 	}
1548 	sfmmup->sfmmu_tsb0_4minflcnt = 0;
1549 	sfmmup->sfmmu_iblk = NULL;
1550 	sfmmup->sfmmu_ismhat = 0;
1551 	sfmmup->sfmmu_scdhat = 0;
1552 	sfmmup->sfmmu_ismblkpa = (uint64_t)-1;
1553 	if (sfmmup == ksfmmup) {
1554 		CPUSET_ALL(sfmmup->sfmmu_cpusran);
1555 	} else {
1556 		CPUSET_ZERO(sfmmup->sfmmu_cpusran);
1557 	}
1558 	sfmmup->sfmmu_free = 0;
1559 	sfmmup->sfmmu_rmstat = 0;
1560 	sfmmup->sfmmu_clrbin = sfmmup->sfmmu_clrstart;
1561 	sfmmup->sfmmu_xhat_provider = NULL;
1562 	cv_init(&sfmmup->sfmmu_tsb_cv, NULL, CV_DEFAULT, NULL);
1563 	sfmmup->sfmmu_srdp = NULL;
1564 	SF_RGNMAP_ZERO(sfmmup->sfmmu_region_map);
1565 	bzero(sfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
1566 	sfmmup->sfmmu_scdp = NULL;
1567 	sfmmup->sfmmu_scd_link.next = NULL;
1568 	sfmmup->sfmmu_scd_link.prev = NULL;
1569 	return (sfmmup);
1570 }
1571 
1572 /*
1573  * Create per-MMU context domain kstats for a given MMU ctx.
1574  */
1575 static void
1576 sfmmu_mmu_kstat_create(mmu_ctx_t *mmu_ctxp)
1577 {
1578 	mmu_ctx_stat_t	stat;
1579 	kstat_t		*mmu_kstat;
1580 
1581 	ASSERT(MUTEX_HELD(&cpu_lock));
1582 	ASSERT(mmu_ctxp->mmu_kstat == NULL);
1583 
1584 	mmu_kstat = kstat_create("unix", mmu_ctxp->mmu_idx, "mmu_ctx",
1585 	    "hat", KSTAT_TYPE_NAMED, MMU_CTX_NUM_STATS, KSTAT_FLAG_VIRTUAL);
1586 
1587 	if (mmu_kstat == NULL) {
1588 		cmn_err(CE_WARN, "kstat_create for MMU %d failed",
1589 		    mmu_ctxp->mmu_idx);
1590 	} else {
1591 		mmu_kstat->ks_data = mmu_ctxp->mmu_kstat_data;
1592 		for (stat = 0; stat < MMU_CTX_NUM_STATS; stat++)
1593 			kstat_named_init(&mmu_ctxp->mmu_kstat_data[stat],
1594 			    mmu_ctx_kstat_names[stat], KSTAT_DATA_INT64);
1595 		mmu_ctxp->mmu_kstat = mmu_kstat;
1596 		kstat_install(mmu_kstat);
1597 	}
1598 }
1599 
1600 /*
1601  * plat_cpuid_to_mmu_ctx_info() is a platform interface that returns MMU
1602  * context domain information for a given CPU. If a platform does not
1603  * specify that interface, then the function below is used instead to return
1604  * default information. The defaults are as follows:
1605  *
1606  *	- The number of MMU context IDs supported on any CPU in the
1607  *	  system is 8K.
1608  *	- There is one MMU context domain per CPU.
1609  */
1610 /*ARGSUSED*/
1611 static void
1612 sfmmu_cpuid_to_mmu_ctx_info(processorid_t cpuid, mmu_ctx_info_t *infop)
1613 {
1614 	infop->mmu_nctxs = nctxs;
1615 	infop->mmu_idx = cpu[cpuid]->cpu_seqid;
1616 }
1617 
1618 /*
1619  * Called during CPU initialization to set the MMU context-related information
1620  * for a CPU.
1621  *
1622  * cpu_lock serializes accesses to mmu_ctxs and mmu_saved_gnum.
1623  */
1624 void
1625 sfmmu_cpu_init(cpu_t *cp)
1626 {
1627 	mmu_ctx_info_t	info;
1628 	mmu_ctx_t	*mmu_ctxp;
1629 
1630 	ASSERT(MUTEX_HELD(&cpu_lock));
1631 
1632 	if (&plat_cpuid_to_mmu_ctx_info == NULL)
1633 		sfmmu_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1634 	else
1635 		plat_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1636 
1637 	ASSERT(info.mmu_idx < max_mmu_ctxdoms);
1638 
1639 	if ((mmu_ctxp = mmu_ctxs_tbl[info.mmu_idx]) == NULL) {
1640 		/* Each mmu_ctx is cacheline aligned. */
1641 		mmu_ctxp = kmem_cache_alloc(mmuctxdom_cache, KM_SLEEP);
1642 		bzero(mmu_ctxp, sizeof (mmu_ctx_t));
1643 
1644 		mutex_init(&mmu_ctxp->mmu_lock, NULL, MUTEX_SPIN,
1645 		    (void *)ipltospl(DISP_LEVEL));
1646 		mmu_ctxp->mmu_idx = info.mmu_idx;
1647 		mmu_ctxp->mmu_nctxs = info.mmu_nctxs;
1648 		/*
1649 		 * Globally for lifetime of a system,
1650 		 * gnum must always increase.
1651 		 * mmu_saved_gnum is protected by the cpu_lock.
1652 		 */
1653 		mmu_ctxp->mmu_gnum = mmu_saved_gnum + 1;
1654 		mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
1655 
1656 		sfmmu_mmu_kstat_create(mmu_ctxp);
1657 
1658 		mmu_ctxs_tbl[info.mmu_idx] = mmu_ctxp;
1659 	} else {
1660 		ASSERT(mmu_ctxp->mmu_idx == info.mmu_idx);
1661 		ASSERT(mmu_ctxp->mmu_nctxs <= info.mmu_nctxs);
1662 	}
1663 
1664 	/*
1665 	 * The mmu_lock is acquired here to prevent races with
1666 	 * the wrap-around code.
1667 	 */
1668 	mutex_enter(&mmu_ctxp->mmu_lock);
1669 
1670 
1671 	mmu_ctxp->mmu_ncpus++;
1672 	CPUSET_ADD(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1673 	CPU_MMU_IDX(cp) = info.mmu_idx;
1674 	CPU_MMU_CTXP(cp) = mmu_ctxp;
1675 
1676 	mutex_exit(&mmu_ctxp->mmu_lock);
1677 }
1678 
1679 static void
1680 sfmmu_ctxdom_free(mmu_ctx_t *mmu_ctxp)
1681 {
1682 	ASSERT(MUTEX_HELD(&cpu_lock));
1683 	ASSERT(!MUTEX_HELD(&mmu_ctxp->mmu_lock));
1684 
1685 	mutex_destroy(&mmu_ctxp->mmu_lock);
1686 
1687 	if (mmu_ctxp->mmu_kstat)
1688 		kstat_delete(mmu_ctxp->mmu_kstat);
1689 
1690 	/* mmu_saved_gnum is protected by the cpu_lock. */
1691 	if (mmu_saved_gnum < mmu_ctxp->mmu_gnum)
1692 		mmu_saved_gnum = mmu_ctxp->mmu_gnum;
1693 
1694 	kmem_cache_free(mmuctxdom_cache, mmu_ctxp);
1695 }
1696 
1697 /*
1698  * Called to perform MMU context-related cleanup for a CPU.
1699  */
1700 void
1701 sfmmu_cpu_cleanup(cpu_t *cp)
1702 {
1703 	mmu_ctx_t	*mmu_ctxp;
1704 
1705 	ASSERT(MUTEX_HELD(&cpu_lock));
1706 
1707 	mmu_ctxp = CPU_MMU_CTXP(cp);
1708 	ASSERT(mmu_ctxp != NULL);
1709 
1710 	/*
1711 	 * The mmu_lock is acquired here to prevent races with
1712 	 * the wrap-around code.
1713 	 */
1714 	mutex_enter(&mmu_ctxp->mmu_lock);
1715 
1716 	CPU_MMU_CTXP(cp) = NULL;
1717 
1718 	CPUSET_DEL(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1719 	if (--mmu_ctxp->mmu_ncpus == 0) {
1720 		mmu_ctxs_tbl[mmu_ctxp->mmu_idx] = NULL;
1721 		mutex_exit(&mmu_ctxp->mmu_lock);
1722 		sfmmu_ctxdom_free(mmu_ctxp);
1723 		return;
1724 	}
1725 
1726 	mutex_exit(&mmu_ctxp->mmu_lock);
1727 }
1728 
1729 uint_t
1730 sfmmu_ctxdom_nctxs(int idx)
1731 {
1732 	return (mmu_ctxs_tbl[idx]->mmu_nctxs);
1733 }
1734 
1735 #ifdef sun4v
1736 /*
1737  * sfmmu_ctxdoms_* is an interface provided to help keep context domains
1738  * consistant after suspend/resume on system that can resume on a different
1739  * hardware than it was suspended.
1740  *
1741  * sfmmu_ctxdom_lock(void) locks all context domains and prevents new contexts
1742  * from being allocated.  It acquires all hat_locks, which blocks most access to
1743  * context data, except for a few cases that are handled separately or are
1744  * harmless.  It wraps each domain to increment gnum and invalidate on-CPU
1745  * contexts, and forces cnum to its max.  As a result of this call all user
1746  * threads that are running on CPUs trap and try to perform wrap around but
1747  * can't because hat_locks are taken.  Threads that were not on CPUs but started
1748  * by scheduler go to sfmmu_alloc_ctx() to aquire context without checking
1749  * hat_lock, but fail, because cnum == nctxs, and therefore also trap and block
1750  * on hat_lock trying to wrap.  sfmmu_ctxdom_lock() must be called before CPUs
1751  * are paused, else it could deadlock acquiring locks held by paused CPUs.
1752  *
1753  * sfmmu_ctxdoms_remove() removes context domains from every CPUs and records
1754  * the CPUs that had them.  It must be called after CPUs have been paused. This
1755  * ensures that no threads are in sfmmu_alloc_ctx() accessing domain data,
1756  * because pause_cpus sends a mondo interrupt to every CPU, and sfmmu_alloc_ctx
1757  * runs with interrupts disabled.  When CPUs are later resumed, they may enter
1758  * sfmmu_alloc_ctx, but it will check for CPU_MMU_CTXP = NULL and immediately
1759  * return failure.  Or, they will be blocked trying to acquire hat_lock. Thus
1760  * after sfmmu_ctxdoms_remove returns, we are guaranteed that no one is
1761  * accessing the old context domains.
1762  *
1763  * sfmmu_ctxdoms_update(void) frees space used by old context domains and
1764  * allocates new context domains based on hardware layout.  It initializes
1765  * every CPU that had context domain before migration to have one again.
1766  * sfmmu_ctxdoms_update must be called after CPUs are resumed, else it
1767  * could deadlock acquiring locks held by paused CPUs.
1768  *
1769  * sfmmu_ctxdoms_unlock(void) releases all hat_locks after which user threads
1770  * acquire new context ids and continue execution.
1771  *
1772  * Therefore functions should be called in the following order:
1773  *       suspend_routine()
1774  *		sfmmu_ctxdom_lock()
1775  *		pause_cpus()
1776  *		suspend()
1777  *			if (suspend failed)
1778  *				sfmmu_ctxdom_unlock()
1779  *		...
1780  *		sfmmu_ctxdom_remove()
1781  *		resume_cpus()
1782  *		sfmmu_ctxdom_update()
1783  *		sfmmu_ctxdom_unlock()
1784  */
1785 static cpuset_t sfmmu_ctxdoms_pset;
1786 
1787 void
1788 sfmmu_ctxdoms_remove()
1789 {
1790 	processorid_t	id;
1791 	cpu_t		*cp;
1792 
1793 	/*
1794 	 * Record the CPUs that have domains in sfmmu_ctxdoms_pset, so they can
1795 	 * be restored post-migration. A CPU may be powered off and not have a
1796 	 * domain, for example.
1797 	 */
1798 	CPUSET_ZERO(sfmmu_ctxdoms_pset);
1799 
1800 	for (id = 0; id < NCPU; id++) {
1801 		if ((cp = cpu[id]) != NULL && CPU_MMU_CTXP(cp) != NULL) {
1802 			CPUSET_ADD(sfmmu_ctxdoms_pset, id);
1803 			CPU_MMU_CTXP(cp) = NULL;
1804 		}
1805 	}
1806 }
1807 
1808 void
1809 sfmmu_ctxdoms_lock(void)
1810 {
1811 	int		idx;
1812 	mmu_ctx_t	*mmu_ctxp;
1813 
1814 	sfmmu_hat_lock_all();
1815 
1816 	/*
1817 	 * At this point, no thread can be in sfmmu_ctx_wrap_around, because
1818 	 * hat_lock is always taken before calling it.
1819 	 *
1820 	 * For each domain, set mmu_cnum to max so no more contexts can be
1821 	 * allocated, and wrap to flush on-CPU contexts and force threads to
1822 	 * acquire a new context when we later drop hat_lock after migration.
1823 	 * Setting mmu_cnum may race with sfmmu_alloc_ctx which also sets cnum,
1824 	 * but the latter uses CAS and will miscompare and not overwrite it.
1825 	 */
1826 	kpreempt_disable(); /* required by sfmmu_ctx_wrap_around */
1827 	for (idx = 0; idx < max_mmu_ctxdoms; idx++) {
1828 		if ((mmu_ctxp = mmu_ctxs_tbl[idx]) != NULL) {
1829 			mutex_enter(&mmu_ctxp->mmu_lock);
1830 			mmu_ctxp->mmu_cnum = mmu_ctxp->mmu_nctxs;
1831 			/* make sure updated cnum visible */
1832 			membar_enter();
1833 			mutex_exit(&mmu_ctxp->mmu_lock);
1834 			sfmmu_ctx_wrap_around(mmu_ctxp, B_FALSE);
1835 		}
1836 	}
1837 	kpreempt_enable();
1838 }
1839 
1840 void
1841 sfmmu_ctxdoms_unlock(void)
1842 {
1843 	sfmmu_hat_unlock_all();
1844 }
1845 
1846 void
1847 sfmmu_ctxdoms_update(void)
1848 {
1849 	processorid_t	id;
1850 	cpu_t		*cp;
1851 	uint_t		idx;
1852 	mmu_ctx_t	*mmu_ctxp;
1853 
1854 	/*
1855 	 * Free all context domains.  As side effect, this increases
1856 	 * mmu_saved_gnum to the maximum gnum over all domains, which is used to
1857 	 * init gnum in the new domains, which therefore will be larger than the
1858 	 * sfmmu gnum for any process, guaranteeing that every process will see
1859 	 * a new generation and allocate a new context regardless of what new
1860 	 * domain it runs in.
1861 	 */
1862 	mutex_enter(&cpu_lock);
1863 
1864 	for (idx = 0; idx < max_mmu_ctxdoms; idx++) {
1865 		if (mmu_ctxs_tbl[idx] != NULL) {
1866 			mmu_ctxp = mmu_ctxs_tbl[idx];
1867 			mmu_ctxs_tbl[idx] = NULL;
1868 			sfmmu_ctxdom_free(mmu_ctxp);
1869 		}
1870 	}
1871 
1872 	for (id = 0; id < NCPU; id++) {
1873 		if (CPU_IN_SET(sfmmu_ctxdoms_pset, id) &&
1874 		    (cp = cpu[id]) != NULL)
1875 			sfmmu_cpu_init(cp);
1876 	}
1877 	mutex_exit(&cpu_lock);
1878 }
1879 #endif
1880 
1881 /*
1882  * Hat_setup, makes an address space context the current active one.
1883  * In sfmmu this translates to setting the secondary context with the
1884  * corresponding context.
1885  */
1886 void
1887 hat_setup(struct hat *sfmmup, int allocflag)
1888 {
1889 	hatlock_t *hatlockp;
1890 
1891 	/* Init needs some special treatment. */
1892 	if (allocflag == HAT_INIT) {
1893 		/*
1894 		 * Make sure that we have
1895 		 * 1. a TSB
1896 		 * 2. a valid ctx that doesn't get stolen after this point.
1897 		 */
1898 		hatlockp = sfmmu_hat_enter(sfmmup);
1899 
1900 		/*
1901 		 * Swap in the TSB.  hat_init() allocates tsbinfos without
1902 		 * TSBs, but we need one for init, since the kernel does some
1903 		 * special things to set up its stack and needs the TSB to
1904 		 * resolve page faults.
1905 		 */
1906 		sfmmu_tsb_swapin(sfmmup, hatlockp);
1907 
1908 		sfmmu_get_ctx(sfmmup);
1909 
1910 		sfmmu_hat_exit(hatlockp);
1911 	} else {
1912 		ASSERT(allocflag == HAT_ALLOC);
1913 
1914 		hatlockp = sfmmu_hat_enter(sfmmup);
1915 		kpreempt_disable();
1916 
1917 		CPUSET_ADD(sfmmup->sfmmu_cpusran, CPU->cpu_id);
1918 		/*
1919 		 * sfmmu_setctx_sec takes <pgsz|cnum> as a parameter,
1920 		 * pagesize bits don't matter in this case since we are passing
1921 		 * INVALID_CONTEXT to it.
1922 		 * Compatibility Note: hw takes care of MMU_SCONTEXT1
1923 		 */
1924 		sfmmu_setctx_sec(INVALID_CONTEXT);
1925 		sfmmu_clear_utsbinfo();
1926 
1927 		kpreempt_enable();
1928 		sfmmu_hat_exit(hatlockp);
1929 	}
1930 }
1931 
1932 /*
1933  * Free all the translation resources for the specified address space.
1934  * Called from as_free when an address space is being destroyed.
1935  */
1936 void
1937 hat_free_start(struct hat *sfmmup)
1938 {
1939 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
1940 	ASSERT(sfmmup != ksfmmup);
1941 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1942 
1943 	sfmmup->sfmmu_free = 1;
1944 	if (sfmmup->sfmmu_scdp != NULL) {
1945 		sfmmu_leave_scd(sfmmup, 0);
1946 	}
1947 
1948 	ASSERT(sfmmup->sfmmu_scdp == NULL);
1949 }
1950 
1951 void
1952 hat_free_end(struct hat *sfmmup)
1953 {
1954 	int i;
1955 
1956 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1957 	ASSERT(sfmmup->sfmmu_free == 1);
1958 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
1959 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
1960 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
1961 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
1962 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
1963 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
1964 
1965 	if (sfmmup->sfmmu_rmstat) {
1966 		hat_freestat(sfmmup->sfmmu_as, NULL);
1967 	}
1968 
1969 	while (sfmmup->sfmmu_tsb != NULL) {
1970 		struct tsb_info *next = sfmmup->sfmmu_tsb->tsb_next;
1971 		sfmmu_tsbinfo_free(sfmmup->sfmmu_tsb);
1972 		sfmmup->sfmmu_tsb = next;
1973 	}
1974 
1975 	if (sfmmup->sfmmu_srdp != NULL) {
1976 		sfmmu_leave_srd(sfmmup);
1977 		ASSERT(sfmmup->sfmmu_srdp == NULL);
1978 		for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1979 			if (sfmmup->sfmmu_hmeregion_links[i] != NULL) {
1980 				kmem_free(sfmmup->sfmmu_hmeregion_links[i],
1981 				    SFMMU_L2_HMERLINKS_SIZE);
1982 				sfmmup->sfmmu_hmeregion_links[i] = NULL;
1983 			}
1984 		}
1985 	}
1986 	sfmmu_free_sfmmu(sfmmup);
1987 
1988 #ifdef DEBUG
1989 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1990 		ASSERT(sfmmup->sfmmu_hmeregion_links[i] == NULL);
1991 	}
1992 #endif
1993 
1994 	kmem_cache_free(sfmmuid_cache, sfmmup);
1995 }
1996 
1997 /*
1998  * Set up any translation structures, for the specified address space,
1999  * that are needed or preferred when the process is being swapped in.
2000  */
2001 /* ARGSUSED */
2002 void
2003 hat_swapin(struct hat *hat)
2004 {
2005 	ASSERT(hat->sfmmu_xhat_provider == NULL);
2006 }
2007 
2008 /*
2009  * Free all of the translation resources, for the specified address space,
2010  * that can be freed while the process is swapped out. Called from as_swapout.
2011  * Also, free up the ctx that this process was using.
2012  */
2013 void
2014 hat_swapout(struct hat *sfmmup)
2015 {
2016 	struct hmehash_bucket *hmebp;
2017 	struct hme_blk *hmeblkp;
2018 	struct hme_blk *pr_hblk = NULL;
2019 	struct hme_blk *nx_hblk;
2020 	int i;
2021 	struct hme_blk *list = NULL;
2022 	hatlock_t *hatlockp;
2023 	struct tsb_info *tsbinfop;
2024 	struct free_tsb {
2025 		struct free_tsb *next;
2026 		struct tsb_info *tsbinfop;
2027 	};			/* free list of TSBs */
2028 	struct free_tsb *freelist, *last, *next;
2029 
2030 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
2031 	SFMMU_STAT(sf_swapout);
2032 
2033 	/*
2034 	 * There is no way to go from an as to all its translations in sfmmu.
2035 	 * Here is one of the times when we take the big hit and traverse
2036 	 * the hash looking for hme_blks to free up.  Not only do we free up
2037 	 * this as hme_blks but all those that are free.  We are obviously
2038 	 * swapping because we need memory so let's free up as much
2039 	 * as we can.
2040 	 *
2041 	 * Note that we don't flush TLB/TSB here -- it's not necessary
2042 	 * because:
2043 	 *  1) we free the ctx we're using and throw away the TSB(s);
2044 	 *  2) processes aren't runnable while being swapped out.
2045 	 */
2046 	ASSERT(sfmmup != KHATID);
2047 	for (i = 0; i <= UHMEHASH_SZ; i++) {
2048 		hmebp = &uhme_hash[i];
2049 		SFMMU_HASH_LOCK(hmebp);
2050 		hmeblkp = hmebp->hmeblkp;
2051 		pr_hblk = NULL;
2052 		while (hmeblkp) {
2053 
2054 			ASSERT(!hmeblkp->hblk_xhat_bit);
2055 
2056 			if ((hmeblkp->hblk_tag.htag_id == sfmmup) &&
2057 			    !hmeblkp->hblk_shw_bit && !hmeblkp->hblk_lckcnt) {
2058 				ASSERT(!hmeblkp->hblk_shared);
2059 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
2060 				    (caddr_t)get_hblk_base(hmeblkp),
2061 				    get_hblk_endaddr(hmeblkp),
2062 				    NULL, HAT_UNLOAD);
2063 			}
2064 			nx_hblk = hmeblkp->hblk_next;
2065 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
2066 				ASSERT(!hmeblkp->hblk_lckcnt);
2067 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
2068 				    &list, 0);
2069 			} else {
2070 				pr_hblk = hmeblkp;
2071 			}
2072 			hmeblkp = nx_hblk;
2073 		}
2074 		SFMMU_HASH_UNLOCK(hmebp);
2075 	}
2076 
2077 	sfmmu_hblks_list_purge(&list, 0);
2078 
2079 	/*
2080 	 * Now free up the ctx so that others can reuse it.
2081 	 */
2082 	hatlockp = sfmmu_hat_enter(sfmmup);
2083 
2084 	sfmmu_invalidate_ctx(sfmmup);
2085 
2086 	/*
2087 	 * Free TSBs, but not tsbinfos, and set SWAPPED flag.
2088 	 * If TSBs were never swapped in, just return.
2089 	 * This implies that we don't support partial swapping
2090 	 * of TSBs -- either all are swapped out, or none are.
2091 	 *
2092 	 * We must hold the HAT lock here to prevent racing with another
2093 	 * thread trying to unmap TTEs from the TSB or running the post-
2094 	 * relocator after relocating the TSB's memory.  Unfortunately, we
2095 	 * can't free memory while holding the HAT lock or we could
2096 	 * deadlock, so we build a list of TSBs to be freed after marking
2097 	 * the tsbinfos as swapped out and free them after dropping the
2098 	 * lock.
2099 	 */
2100 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
2101 		sfmmu_hat_exit(hatlockp);
2102 		return;
2103 	}
2104 
2105 	SFMMU_FLAGS_SET(sfmmup, HAT_SWAPPED);
2106 	last = freelist = NULL;
2107 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
2108 	    tsbinfop = tsbinfop->tsb_next) {
2109 		ASSERT((tsbinfop->tsb_flags & TSB_SWAPPED) == 0);
2110 
2111 		/*
2112 		 * Cast the TSB into a struct free_tsb and put it on the free
2113 		 * list.
2114 		 */
2115 		if (freelist == NULL) {
2116 			last = freelist = (struct free_tsb *)tsbinfop->tsb_va;
2117 		} else {
2118 			last->next = (struct free_tsb *)tsbinfop->tsb_va;
2119 			last = last->next;
2120 		}
2121 		last->next = NULL;
2122 		last->tsbinfop = tsbinfop;
2123 		tsbinfop->tsb_flags |= TSB_SWAPPED;
2124 		/*
2125 		 * Zero out the TTE to clear the valid bit.
2126 		 * Note we can't use a value like 0xbad because we want to
2127 		 * ensure diagnostic bits are NEVER set on TTEs that might
2128 		 * be loaded.  The intent is to catch any invalid access
2129 		 * to the swapped TSB, such as a thread running with a valid
2130 		 * context without first calling sfmmu_tsb_swapin() to
2131 		 * allocate TSB memory.
2132 		 */
2133 		tsbinfop->tsb_tte.ll = 0;
2134 	}
2135 
2136 	/* Now we can drop the lock and free the TSB memory. */
2137 	sfmmu_hat_exit(hatlockp);
2138 	for (; freelist != NULL; freelist = next) {
2139 		next = freelist->next;
2140 		sfmmu_tsb_free(freelist->tsbinfop);
2141 	}
2142 }
2143 
2144 /*
2145  * Duplicate the translations of an as into another newas
2146  */
2147 /* ARGSUSED */
2148 int
2149 hat_dup(struct hat *hat, struct hat *newhat, caddr_t addr, size_t len,
2150 	uint_t flag)
2151 {
2152 	sf_srd_t *srdp;
2153 	sf_scd_t *scdp;
2154 	int i;
2155 	extern uint_t get_color_start(struct as *);
2156 
2157 	ASSERT(hat->sfmmu_xhat_provider == NULL);
2158 	ASSERT((flag == 0) || (flag == HAT_DUP_ALL) || (flag == HAT_DUP_COW) ||
2159 	    (flag == HAT_DUP_SRD));
2160 	ASSERT(hat != ksfmmup);
2161 	ASSERT(newhat != ksfmmup);
2162 	ASSERT(flag != HAT_DUP_ALL || hat->sfmmu_srdp == newhat->sfmmu_srdp);
2163 
2164 	if (flag == HAT_DUP_COW) {
2165 		panic("hat_dup: HAT_DUP_COW not supported");
2166 	}
2167 
2168 	if (flag == HAT_DUP_SRD && ((srdp = hat->sfmmu_srdp) != NULL)) {
2169 		ASSERT(srdp->srd_evp != NULL);
2170 		VN_HOLD(srdp->srd_evp);
2171 		ASSERT(srdp->srd_refcnt > 0);
2172 		newhat->sfmmu_srdp = srdp;
2173 		atomic_add_32((volatile uint_t *)&srdp->srd_refcnt, 1);
2174 	}
2175 
2176 	/*
2177 	 * HAT_DUP_ALL flag is used after as duplication is done.
2178 	 */
2179 	if (flag == HAT_DUP_ALL && ((srdp = newhat->sfmmu_srdp) != NULL)) {
2180 		ASSERT(newhat->sfmmu_srdp->srd_refcnt >= 2);
2181 		newhat->sfmmu_rtteflags = hat->sfmmu_rtteflags;
2182 		if (hat->sfmmu_flags & HAT_4MTEXT_FLAG) {
2183 			newhat->sfmmu_flags |= HAT_4MTEXT_FLAG;
2184 		}
2185 
2186 		/* check if need to join scd */
2187 		if ((scdp = hat->sfmmu_scdp) != NULL &&
2188 		    newhat->sfmmu_scdp != scdp) {
2189 			int ret;
2190 			SF_RGNMAP_IS_SUBSET(&newhat->sfmmu_region_map,
2191 			    &scdp->scd_region_map, ret);
2192 			ASSERT(ret);
2193 			sfmmu_join_scd(scdp, newhat);
2194 			ASSERT(newhat->sfmmu_scdp == scdp &&
2195 			    scdp->scd_refcnt >= 2);
2196 			for (i = 0; i < max_mmu_page_sizes; i++) {
2197 				newhat->sfmmu_ismttecnt[i] =
2198 				    hat->sfmmu_ismttecnt[i];
2199 				newhat->sfmmu_scdismttecnt[i] =
2200 				    hat->sfmmu_scdismttecnt[i];
2201 			}
2202 		}
2203 
2204 		sfmmu_check_page_sizes(newhat, 1);
2205 	}
2206 
2207 	if (flag == HAT_DUP_ALL && consistent_coloring == 0 &&
2208 	    update_proc_pgcolorbase_after_fork != 0) {
2209 		hat->sfmmu_clrbin = get_color_start(hat->sfmmu_as);
2210 	}
2211 	return (0);
2212 }
2213 
2214 void
2215 hat_memload(struct hat *hat, caddr_t addr, struct page *pp,
2216 	uint_t attr, uint_t flags)
2217 {
2218 	hat_do_memload(hat, addr, pp, attr, flags,
2219 	    SFMMU_INVALID_SHMERID);
2220 }
2221 
2222 void
2223 hat_memload_region(struct hat *hat, caddr_t addr, struct page *pp,
2224 	uint_t attr, uint_t flags, hat_region_cookie_t rcookie)
2225 {
2226 	uint_t rid;
2227 	if (rcookie == HAT_INVALID_REGION_COOKIE ||
2228 	    hat->sfmmu_xhat_provider != NULL) {
2229 		hat_do_memload(hat, addr, pp, attr, flags,
2230 		    SFMMU_INVALID_SHMERID);
2231 		return;
2232 	}
2233 	rid = (uint_t)((uint64_t)rcookie);
2234 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2235 	hat_do_memload(hat, addr, pp, attr, flags, rid);
2236 }
2237 
2238 /*
2239  * Set up addr to map to page pp with protection prot.
2240  * As an optimization we also load the TSB with the
2241  * corresponding tte but it is no big deal if  the tte gets kicked out.
2242  */
2243 static void
2244 hat_do_memload(struct hat *hat, caddr_t addr, struct page *pp,
2245 	uint_t attr, uint_t flags, uint_t rid)
2246 {
2247 	tte_t tte;
2248 
2249 
2250 	ASSERT(hat != NULL);
2251 	ASSERT(PAGE_LOCKED(pp));
2252 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2253 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2254 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2255 	SFMMU_VALIDATE_HMERID(hat, rid, addr, MMU_PAGESIZE);
2256 
2257 	if (PP_ISFREE(pp)) {
2258 		panic("hat_memload: loading a mapping to free page %p",
2259 		    (void *)pp);
2260 	}
2261 
2262 	if (hat->sfmmu_xhat_provider) {
2263 		/* no regions for xhats */
2264 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
2265 		XHAT_MEMLOAD(hat, addr, pp, attr, flags);
2266 		return;
2267 	}
2268 
2269 	ASSERT((hat == ksfmmup) ||
2270 	    AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock));
2271 
2272 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2273 		cmn_err(CE_NOTE, "hat_memload: unsupported flags %d",
2274 		    flags & ~SFMMU_LOAD_ALLFLAG);
2275 
2276 	if (hat->sfmmu_rmstat)
2277 		hat_resvstat(MMU_PAGESIZE, hat->sfmmu_as, addr);
2278 
2279 #if defined(SF_ERRATA_57)
2280 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2281 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2282 	    !(flags & HAT_LOAD_SHARE)) {
2283 		cmn_err(CE_WARN, "hat_memload: illegal attempt to make user "
2284 		    " page executable");
2285 		attr &= ~PROT_EXEC;
2286 	}
2287 #endif
2288 
2289 	sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2290 	(void) sfmmu_tteload_array(hat, &tte, addr, &pp, flags, rid);
2291 
2292 	/*
2293 	 * Check TSB and TLB page sizes.
2294 	 */
2295 	if ((flags & HAT_LOAD_SHARE) == 0) {
2296 		sfmmu_check_page_sizes(hat, 1);
2297 	}
2298 }
2299 
2300 /*
2301  * hat_devload can be called to map real memory (e.g.
2302  * /dev/kmem) and even though hat_devload will determine pf is
2303  * for memory, it will be unable to get a shared lock on the
2304  * page (because someone else has it exclusively) and will
2305  * pass dp = NULL.  If tteload doesn't get a non-NULL
2306  * page pointer it can't cache memory.
2307  */
2308 void
2309 hat_devload(struct hat *hat, caddr_t addr, size_t len, pfn_t pfn,
2310 	uint_t attr, int flags)
2311 {
2312 	tte_t tte;
2313 	struct page *pp = NULL;
2314 	int use_lgpg = 0;
2315 
2316 	ASSERT(hat != NULL);
2317 
2318 	if (hat->sfmmu_xhat_provider) {
2319 		XHAT_DEVLOAD(hat, addr, len, pfn, attr, flags);
2320 		return;
2321 	}
2322 
2323 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2324 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2325 	ASSERT((hat == ksfmmup) ||
2326 	    AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock));
2327 	if (len == 0)
2328 		panic("hat_devload: zero len");
2329 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2330 		cmn_err(CE_NOTE, "hat_devload: unsupported flags %d",
2331 		    flags & ~SFMMU_LOAD_ALLFLAG);
2332 
2333 #if defined(SF_ERRATA_57)
2334 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2335 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2336 	    !(flags & HAT_LOAD_SHARE)) {
2337 		cmn_err(CE_WARN, "hat_devload: illegal attempt to make user "
2338 		    " page executable");
2339 		attr &= ~PROT_EXEC;
2340 	}
2341 #endif
2342 
2343 	/*
2344 	 * If it's a memory page find its pp
2345 	 */
2346 	if (!(flags & HAT_LOAD_NOCONSIST) && pf_is_memory(pfn)) {
2347 		pp = page_numtopp_nolock(pfn);
2348 		if (pp == NULL) {
2349 			flags |= HAT_LOAD_NOCONSIST;
2350 		} else {
2351 			if (PP_ISFREE(pp)) {
2352 				panic("hat_memload: loading "
2353 				    "a mapping to free page %p",
2354 				    (void *)pp);
2355 			}
2356 			if (!PAGE_LOCKED(pp) && !PP_ISNORELOC(pp)) {
2357 				panic("hat_memload: loading a mapping "
2358 				    "to unlocked relocatable page %p",
2359 				    (void *)pp);
2360 			}
2361 			ASSERT(len == MMU_PAGESIZE);
2362 		}
2363 	}
2364 
2365 	if (hat->sfmmu_rmstat)
2366 		hat_resvstat(len, hat->sfmmu_as, addr);
2367 
2368 	if (flags & HAT_LOAD_NOCONSIST) {
2369 		attr |= SFMMU_UNCACHEVTTE;
2370 		use_lgpg = 1;
2371 	}
2372 	if (!pf_is_memory(pfn)) {
2373 		attr |= SFMMU_UNCACHEPTTE | HAT_NOSYNC;
2374 		use_lgpg = 1;
2375 		switch (attr & HAT_ORDER_MASK) {
2376 			case HAT_STRICTORDER:
2377 			case HAT_UNORDERED_OK:
2378 				/*
2379 				 * we set the side effect bit for all non
2380 				 * memory mappings unless merging is ok
2381 				 */
2382 				attr |= SFMMU_SIDEFFECT;
2383 				break;
2384 			case HAT_MERGING_OK:
2385 			case HAT_LOADCACHING_OK:
2386 			case HAT_STORECACHING_OK:
2387 				break;
2388 			default:
2389 				panic("hat_devload: bad attr");
2390 				break;
2391 		}
2392 	}
2393 	while (len) {
2394 		if (!use_lgpg) {
2395 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2396 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2397 			    flags, SFMMU_INVALID_SHMERID);
2398 			len -= MMU_PAGESIZE;
2399 			addr += MMU_PAGESIZE;
2400 			pfn++;
2401 			continue;
2402 		}
2403 		/*
2404 		 *  try to use large pages, check va/pa alignments
2405 		 *  Note that 32M/256M page sizes are not (yet) supported.
2406 		 */
2407 		if ((len >= MMU_PAGESIZE4M) &&
2408 		    !((uintptr_t)addr & MMU_PAGEOFFSET4M) &&
2409 		    !(disable_large_pages & (1 << TTE4M)) &&
2410 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET4M)) {
2411 			sfmmu_memtte(&tte, pfn, attr, TTE4M);
2412 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2413 			    flags, SFMMU_INVALID_SHMERID);
2414 			len -= MMU_PAGESIZE4M;
2415 			addr += MMU_PAGESIZE4M;
2416 			pfn += MMU_PAGESIZE4M / MMU_PAGESIZE;
2417 		} else if ((len >= MMU_PAGESIZE512K) &&
2418 		    !((uintptr_t)addr & MMU_PAGEOFFSET512K) &&
2419 		    !(disable_large_pages & (1 << TTE512K)) &&
2420 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET512K)) {
2421 			sfmmu_memtte(&tte, pfn, attr, TTE512K);
2422 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2423 			    flags, SFMMU_INVALID_SHMERID);
2424 			len -= MMU_PAGESIZE512K;
2425 			addr += MMU_PAGESIZE512K;
2426 			pfn += MMU_PAGESIZE512K / MMU_PAGESIZE;
2427 		} else if ((len >= MMU_PAGESIZE64K) &&
2428 		    !((uintptr_t)addr & MMU_PAGEOFFSET64K) &&
2429 		    !(disable_large_pages & (1 << TTE64K)) &&
2430 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET64K)) {
2431 			sfmmu_memtte(&tte, pfn, attr, TTE64K);
2432 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2433 			    flags, SFMMU_INVALID_SHMERID);
2434 			len -= MMU_PAGESIZE64K;
2435 			addr += MMU_PAGESIZE64K;
2436 			pfn += MMU_PAGESIZE64K / MMU_PAGESIZE;
2437 		} else {
2438 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2439 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2440 			    flags, SFMMU_INVALID_SHMERID);
2441 			len -= MMU_PAGESIZE;
2442 			addr += MMU_PAGESIZE;
2443 			pfn++;
2444 		}
2445 	}
2446 
2447 	/*
2448 	 * Check TSB and TLB page sizes.
2449 	 */
2450 	if ((flags & HAT_LOAD_SHARE) == 0) {
2451 		sfmmu_check_page_sizes(hat, 1);
2452 	}
2453 }
2454 
2455 void
2456 hat_memload_array(struct hat *hat, caddr_t addr, size_t len,
2457 	struct page **pps, uint_t attr, uint_t flags)
2458 {
2459 	hat_do_memload_array(hat, addr, len, pps, attr, flags,
2460 	    SFMMU_INVALID_SHMERID);
2461 }
2462 
2463 void
2464 hat_memload_array_region(struct hat *hat, caddr_t addr, size_t len,
2465 	struct page **pps, uint_t attr, uint_t flags,
2466 	hat_region_cookie_t rcookie)
2467 {
2468 	uint_t rid;
2469 	if (rcookie == HAT_INVALID_REGION_COOKIE ||
2470 	    hat->sfmmu_xhat_provider != NULL) {
2471 		hat_do_memload_array(hat, addr, len, pps, attr, flags,
2472 		    SFMMU_INVALID_SHMERID);
2473 		return;
2474 	}
2475 	rid = (uint_t)((uint64_t)rcookie);
2476 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2477 	hat_do_memload_array(hat, addr, len, pps, attr, flags, rid);
2478 }
2479 
2480 /*
2481  * Map the largest extend possible out of the page array. The array may NOT
2482  * be in order.  The largest possible mapping a page can have
2483  * is specified in the p_szc field.  The p_szc field
2484  * cannot change as long as there any mappings (large or small)
2485  * to any of the pages that make up the large page. (ie. any
2486  * promotion/demotion of page size is not up to the hat but up to
2487  * the page free list manager).  The array
2488  * should consist of properly aligned contigous pages that are
2489  * part of a big page for a large mapping to be created.
2490  */
2491 static void
2492 hat_do_memload_array(struct hat *hat, caddr_t addr, size_t len,
2493 	struct page **pps, uint_t attr, uint_t flags, uint_t rid)
2494 {
2495 	int  ttesz;
2496 	size_t mapsz;
2497 	pgcnt_t	numpg, npgs;
2498 	tte_t tte;
2499 	page_t *pp;
2500 	uint_t large_pages_disable;
2501 
2502 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2503 	SFMMU_VALIDATE_HMERID(hat, rid, addr, len);
2504 
2505 	if (hat->sfmmu_xhat_provider) {
2506 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
2507 		XHAT_MEMLOAD_ARRAY(hat, addr, len, pps, attr, flags);
2508 		return;
2509 	}
2510 
2511 	if (hat->sfmmu_rmstat)
2512 		hat_resvstat(len, hat->sfmmu_as, addr);
2513 
2514 #if defined(SF_ERRATA_57)
2515 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2516 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2517 	    !(flags & HAT_LOAD_SHARE)) {
2518 		cmn_err(CE_WARN, "hat_memload_array: illegal attempt to make "
2519 		    "user page executable");
2520 		attr &= ~PROT_EXEC;
2521 	}
2522 #endif
2523 
2524 	/* Get number of pages */
2525 	npgs = len >> MMU_PAGESHIFT;
2526 
2527 	if (flags & HAT_LOAD_SHARE) {
2528 		large_pages_disable = disable_ism_large_pages;
2529 	} else {
2530 		large_pages_disable = disable_large_pages;
2531 	}
2532 
2533 	if (npgs < NHMENTS || large_pages_disable == LARGE_PAGES_OFF) {
2534 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2535 		    rid);
2536 		return;
2537 	}
2538 
2539 	while (npgs >= NHMENTS) {
2540 		pp = *pps;
2541 		for (ttesz = pp->p_szc; ttesz != TTE8K; ttesz--) {
2542 			/*
2543 			 * Check if this page size is disabled.
2544 			 */
2545 			if (large_pages_disable & (1 << ttesz))
2546 				continue;
2547 
2548 			numpg = TTEPAGES(ttesz);
2549 			mapsz = numpg << MMU_PAGESHIFT;
2550 			if ((npgs >= numpg) &&
2551 			    IS_P2ALIGNED(addr, mapsz) &&
2552 			    IS_P2ALIGNED(pp->p_pagenum, numpg)) {
2553 				/*
2554 				 * At this point we have enough pages and
2555 				 * we know the virtual address and the pfn
2556 				 * are properly aligned.  We still need
2557 				 * to check for physical contiguity but since
2558 				 * it is very likely that this is the case
2559 				 * we will assume they are so and undo
2560 				 * the request if necessary.  It would
2561 				 * be great if we could get a hint flag
2562 				 * like HAT_CONTIG which would tell us
2563 				 * the pages are contigous for sure.
2564 				 */
2565 				sfmmu_memtte(&tte, (*pps)->p_pagenum,
2566 				    attr, ttesz);
2567 				if (!sfmmu_tteload_array(hat, &tte, addr,
2568 				    pps, flags, rid)) {
2569 					break;
2570 				}
2571 			}
2572 		}
2573 		if (ttesz == TTE8K) {
2574 			/*
2575 			 * We were not able to map array using a large page
2576 			 * batch a hmeblk or fraction at a time.
2577 			 */
2578 			numpg = ((uintptr_t)addr >> MMU_PAGESHIFT)
2579 			    & (NHMENTS-1);
2580 			numpg = NHMENTS - numpg;
2581 			ASSERT(numpg <= npgs);
2582 			mapsz = numpg * MMU_PAGESIZE;
2583 			sfmmu_memload_batchsmall(hat, addr, pps, attr, flags,
2584 			    numpg, rid);
2585 		}
2586 		addr += mapsz;
2587 		npgs -= numpg;
2588 		pps += numpg;
2589 	}
2590 
2591 	if (npgs) {
2592 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2593 		    rid);
2594 	}
2595 
2596 	/*
2597 	 * Check TSB and TLB page sizes.
2598 	 */
2599 	if ((flags & HAT_LOAD_SHARE) == 0) {
2600 		sfmmu_check_page_sizes(hat, 1);
2601 	}
2602 }
2603 
2604 /*
2605  * Function tries to batch 8K pages into the same hme blk.
2606  */
2607 static void
2608 sfmmu_memload_batchsmall(struct hat *hat, caddr_t vaddr, page_t **pps,
2609 		    uint_t attr, uint_t flags, pgcnt_t npgs, uint_t rid)
2610 {
2611 	tte_t	tte;
2612 	page_t *pp;
2613 	struct hmehash_bucket *hmebp;
2614 	struct hme_blk *hmeblkp;
2615 	int	index;
2616 
2617 	while (npgs) {
2618 		/*
2619 		 * Acquire the hash bucket.
2620 		 */
2621 		hmebp = sfmmu_tteload_acquire_hashbucket(hat, vaddr, TTE8K,
2622 		    rid);
2623 		ASSERT(hmebp);
2624 
2625 		/*
2626 		 * Find the hment block.
2627 		 */
2628 		hmeblkp = sfmmu_tteload_find_hmeblk(hat, hmebp, vaddr,
2629 		    TTE8K, flags, rid);
2630 		ASSERT(hmeblkp);
2631 
2632 		do {
2633 			/*
2634 			 * Make the tte.
2635 			 */
2636 			pp = *pps;
2637 			sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2638 
2639 			/*
2640 			 * Add the translation.
2641 			 */
2642 			(void) sfmmu_tteload_addentry(hat, hmeblkp, &tte,
2643 			    vaddr, pps, flags, rid);
2644 
2645 			/*
2646 			 * Goto next page.
2647 			 */
2648 			pps++;
2649 			npgs--;
2650 
2651 			/*
2652 			 * Goto next address.
2653 			 */
2654 			vaddr += MMU_PAGESIZE;
2655 
2656 			/*
2657 			 * Don't crossover into a different hmentblk.
2658 			 */
2659 			index = (int)(((uintptr_t)vaddr >> MMU_PAGESHIFT) &
2660 			    (NHMENTS-1));
2661 
2662 		} while (index != 0 && npgs != 0);
2663 
2664 		/*
2665 		 * Release the hash bucket.
2666 		 */
2667 
2668 		sfmmu_tteload_release_hashbucket(hmebp);
2669 	}
2670 }
2671 
2672 /*
2673  * Construct a tte for a page:
2674  *
2675  * tte_valid = 1
2676  * tte_size2 = size & TTE_SZ2_BITS (Panther and Olympus-C only)
2677  * tte_size = size
2678  * tte_nfo = attr & HAT_NOFAULT
2679  * tte_ie = attr & HAT_STRUCTURE_LE
2680  * tte_hmenum = hmenum
2681  * tte_pahi = pp->p_pagenum >> TTE_PASHIFT;
2682  * tte_palo = pp->p_pagenum & TTE_PALOMASK;
2683  * tte_ref = 1 (optimization)
2684  * tte_wr_perm = attr & PROT_WRITE;
2685  * tte_no_sync = attr & HAT_NOSYNC
2686  * tte_lock = attr & SFMMU_LOCKTTE
2687  * tte_cp = !(attr & SFMMU_UNCACHEPTTE)
2688  * tte_cv = !(attr & SFMMU_UNCACHEVTTE)
2689  * tte_e = attr & SFMMU_SIDEFFECT
2690  * tte_priv = !(attr & PROT_USER)
2691  * tte_hwwr = if nosync is set and it is writable we set the mod bit (opt)
2692  * tte_glb = 0
2693  */
2694 void
2695 sfmmu_memtte(tte_t *ttep, pfn_t pfn, uint_t attr, int tte_sz)
2696 {
2697 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2698 
2699 	ttep->tte_inthi = MAKE_TTE_INTHI(pfn, attr, tte_sz, 0 /* hmenum */);
2700 	ttep->tte_intlo = MAKE_TTE_INTLO(pfn, attr, tte_sz, 0 /* hmenum */);
2701 
2702 	if (TTE_IS_NOSYNC(ttep)) {
2703 		TTE_SET_REF(ttep);
2704 		if (TTE_IS_WRITABLE(ttep)) {
2705 			TTE_SET_MOD(ttep);
2706 		}
2707 	}
2708 	if (TTE_IS_NFO(ttep) && TTE_IS_EXECUTABLE(ttep)) {
2709 		panic("sfmmu_memtte: can't set both NFO and EXEC bits");
2710 	}
2711 }
2712 
2713 /*
2714  * This function will add a translation to the hme_blk and allocate the
2715  * hme_blk if one does not exist.
2716  * If a page structure is specified then it will add the
2717  * corresponding hment to the mapping list.
2718  * It will also update the hmenum field for the tte.
2719  *
2720  * Currently this function is only used for kernel mappings.
2721  * So pass invalid region to sfmmu_tteload_array().
2722  */
2723 void
2724 sfmmu_tteload(struct hat *sfmmup, tte_t *ttep, caddr_t vaddr, page_t *pp,
2725 	uint_t flags)
2726 {
2727 	ASSERT(sfmmup == ksfmmup);
2728 	(void) sfmmu_tteload_array(sfmmup, ttep, vaddr, &pp, flags,
2729 	    SFMMU_INVALID_SHMERID);
2730 }
2731 
2732 /*
2733  * Load (ttep != NULL) or unload (ttep == NULL) one entry in the TSB.
2734  * Assumes that a particular page size may only be resident in one TSB.
2735  */
2736 static void
2737 sfmmu_mod_tsb(sfmmu_t *sfmmup, caddr_t vaddr, tte_t *ttep, int ttesz)
2738 {
2739 	struct tsb_info *tsbinfop = NULL;
2740 	uint64_t tag;
2741 	struct tsbe *tsbe_addr;
2742 	uint64_t tsb_base;
2743 	uint_t tsb_size;
2744 	int vpshift = MMU_PAGESHIFT;
2745 	int phys = 0;
2746 
2747 	if (sfmmup == ksfmmup) { /* No support for 32/256M ksfmmu pages */
2748 		phys = ktsb_phys;
2749 		if (ttesz >= TTE4M) {
2750 #ifndef sun4v
2751 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2752 #endif
2753 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2754 			tsb_size = ktsb4m_szcode;
2755 		} else {
2756 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2757 			tsb_size = ktsb_szcode;
2758 		}
2759 	} else {
2760 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2761 
2762 		/*
2763 		 * If there isn't a TSB for this page size, or the TSB is
2764 		 * swapped out, there is nothing to do.  Note that the latter
2765 		 * case seems impossible but can occur if hat_pageunload()
2766 		 * is called on an ISM mapping while the process is swapped
2767 		 * out.
2768 		 */
2769 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2770 			return;
2771 
2772 		/*
2773 		 * If another thread is in the middle of relocating a TSB
2774 		 * we can't unload the entry so set a flag so that the
2775 		 * TSB will be flushed before it can be accessed by the
2776 		 * process.
2777 		 */
2778 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2779 			if (ttep == NULL)
2780 				tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2781 			return;
2782 		}
2783 #if defined(UTSB_PHYS)
2784 		phys = 1;
2785 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2786 #else
2787 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2788 #endif
2789 		tsb_size = tsbinfop->tsb_szc;
2790 	}
2791 	if (ttesz >= TTE4M)
2792 		vpshift = MMU_PAGESHIFT4M;
2793 
2794 	tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2795 	tag = sfmmu_make_tsbtag(vaddr);
2796 
2797 	if (ttep == NULL) {
2798 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2799 	} else {
2800 		if (ttesz >= TTE4M) {
2801 			SFMMU_STAT(sf_tsb_load4m);
2802 		} else {
2803 			SFMMU_STAT(sf_tsb_load8k);
2804 		}
2805 
2806 		sfmmu_load_tsbe(tsbe_addr, tag, ttep, phys);
2807 	}
2808 }
2809 
2810 /*
2811  * Unmap all entries from [start, end) matching the given page size.
2812  *
2813  * This function is used primarily to unmap replicated 64K or 512K entries
2814  * from the TSB that are inserted using the base page size TSB pointer, but
2815  * it may also be called to unmap a range of addresses from the TSB.
2816  */
2817 void
2818 sfmmu_unload_tsb_range(sfmmu_t *sfmmup, caddr_t start, caddr_t end, int ttesz)
2819 {
2820 	struct tsb_info *tsbinfop;
2821 	uint64_t tag;
2822 	struct tsbe *tsbe_addr;
2823 	caddr_t vaddr;
2824 	uint64_t tsb_base;
2825 	int vpshift, vpgsz;
2826 	uint_t tsb_size;
2827 	int phys = 0;
2828 
2829 	/*
2830 	 * Assumptions:
2831 	 *  If ttesz == 8K, 64K or 512K, we walk through the range 8K
2832 	 *  at a time shooting down any valid entries we encounter.
2833 	 *
2834 	 *  If ttesz >= 4M we walk the range 4M at a time shooting
2835 	 *  down any valid mappings we find.
2836 	 */
2837 	if (sfmmup == ksfmmup) {
2838 		phys = ktsb_phys;
2839 		if (ttesz >= TTE4M) {
2840 #ifndef sun4v
2841 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2842 #endif
2843 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2844 			tsb_size = ktsb4m_szcode;
2845 		} else {
2846 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2847 			tsb_size = ktsb_szcode;
2848 		}
2849 	} else {
2850 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2851 
2852 		/*
2853 		 * If there isn't a TSB for this page size, or the TSB is
2854 		 * swapped out, there is nothing to do.  Note that the latter
2855 		 * case seems impossible but can occur if hat_pageunload()
2856 		 * is called on an ISM mapping while the process is swapped
2857 		 * out.
2858 		 */
2859 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2860 			return;
2861 
2862 		/*
2863 		 * If another thread is in the middle of relocating a TSB
2864 		 * we can't unload the entry so set a flag so that the
2865 		 * TSB will be flushed before it can be accessed by the
2866 		 * process.
2867 		 */
2868 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2869 			tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2870 			return;
2871 		}
2872 #if defined(UTSB_PHYS)
2873 		phys = 1;
2874 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2875 #else
2876 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2877 #endif
2878 		tsb_size = tsbinfop->tsb_szc;
2879 	}
2880 	if (ttesz >= TTE4M) {
2881 		vpshift = MMU_PAGESHIFT4M;
2882 		vpgsz = MMU_PAGESIZE4M;
2883 	} else {
2884 		vpshift = MMU_PAGESHIFT;
2885 		vpgsz = MMU_PAGESIZE;
2886 	}
2887 
2888 	for (vaddr = start; vaddr < end; vaddr += vpgsz) {
2889 		tag = sfmmu_make_tsbtag(vaddr);
2890 		tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2891 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2892 	}
2893 }
2894 
2895 /*
2896  * Select the optimum TSB size given the number of mappings
2897  * that need to be cached.
2898  */
2899 static int
2900 sfmmu_select_tsb_szc(pgcnt_t pgcnt)
2901 {
2902 	int szc = 0;
2903 
2904 #ifdef DEBUG
2905 	if (tsb_grow_stress) {
2906 		uint32_t randval = (uint32_t)gettick() >> 4;
2907 		return (randval % (tsb_max_growsize + 1));
2908 	}
2909 #endif	/* DEBUG */
2910 
2911 	while ((szc < tsb_max_growsize) && (pgcnt > SFMMU_RSS_TSBSIZE(szc)))
2912 		szc++;
2913 	return (szc);
2914 }
2915 
2916 /*
2917  * This function will add a translation to the hme_blk and allocate the
2918  * hme_blk if one does not exist.
2919  * If a page structure is specified then it will add the
2920  * corresponding hment to the mapping list.
2921  * It will also update the hmenum field for the tte.
2922  * Furthermore, it attempts to create a large page translation
2923  * for <addr,hat> at page array pps.  It assumes addr and first
2924  * pp is correctly aligned.  It returns 0 if successful and 1 otherwise.
2925  */
2926 static int
2927 sfmmu_tteload_array(sfmmu_t *sfmmup, tte_t *ttep, caddr_t vaddr,
2928 	page_t **pps, uint_t flags, uint_t rid)
2929 {
2930 	struct hmehash_bucket *hmebp;
2931 	struct hme_blk *hmeblkp;
2932 	int 	ret;
2933 	uint_t	size;
2934 
2935 	/*
2936 	 * Get mapping size.
2937 	 */
2938 	size = TTE_CSZ(ttep);
2939 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
2940 
2941 	/*
2942 	 * Acquire the hash bucket.
2943 	 */
2944 	hmebp = sfmmu_tteload_acquire_hashbucket(sfmmup, vaddr, size, rid);
2945 	ASSERT(hmebp);
2946 
2947 	/*
2948 	 * Find the hment block.
2949 	 */
2950 	hmeblkp = sfmmu_tteload_find_hmeblk(sfmmup, hmebp, vaddr, size, flags,
2951 	    rid);
2952 	ASSERT(hmeblkp);
2953 
2954 	/*
2955 	 * Add the translation.
2956 	 */
2957 	ret = sfmmu_tteload_addentry(sfmmup, hmeblkp, ttep, vaddr, pps, flags,
2958 	    rid);
2959 
2960 	/*
2961 	 * Release the hash bucket.
2962 	 */
2963 	sfmmu_tteload_release_hashbucket(hmebp);
2964 
2965 	return (ret);
2966 }
2967 
2968 /*
2969  * Function locks and returns a pointer to the hash bucket for vaddr and size.
2970  */
2971 static struct hmehash_bucket *
2972 sfmmu_tteload_acquire_hashbucket(sfmmu_t *sfmmup, caddr_t vaddr, int size,
2973     uint_t rid)
2974 {
2975 	struct hmehash_bucket *hmebp;
2976 	int hmeshift;
2977 	void *htagid = sfmmutohtagid(sfmmup, rid);
2978 
2979 	ASSERT(htagid != NULL);
2980 
2981 	hmeshift = HME_HASH_SHIFT(size);
2982 
2983 	hmebp = HME_HASH_FUNCTION(htagid, vaddr, hmeshift);
2984 
2985 	SFMMU_HASH_LOCK(hmebp);
2986 
2987 	return (hmebp);
2988 }
2989 
2990 /*
2991  * Function returns a pointer to an hmeblk in the hash bucket, hmebp. If the
2992  * hmeblk doesn't exists for the [sfmmup, vaddr & size] signature, a hmeblk is
2993  * allocated.
2994  */
2995 static struct hme_blk *
2996 sfmmu_tteload_find_hmeblk(sfmmu_t *sfmmup, struct hmehash_bucket *hmebp,
2997 	caddr_t vaddr, uint_t size, uint_t flags, uint_t rid)
2998 {
2999 	hmeblk_tag hblktag;
3000 	int hmeshift;
3001 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
3002 
3003 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
3004 
3005 	hblktag.htag_id = sfmmutohtagid(sfmmup, rid);
3006 	ASSERT(hblktag.htag_id != NULL);
3007 	hmeshift = HME_HASH_SHIFT(size);
3008 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
3009 	hblktag.htag_rehash = HME_HASH_REHASH(size);
3010 	hblktag.htag_rid = rid;
3011 
3012 ttearray_realloc:
3013 
3014 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3015 
3016 	/*
3017 	 * We block until hblk_reserve_lock is released; it's held by
3018 	 * the thread, temporarily using hblk_reserve, until hblk_reserve is
3019 	 * replaced by a hblk from sfmmu8_cache.
3020 	 */
3021 	if (hmeblkp == (struct hme_blk *)hblk_reserve &&
3022 	    hblk_reserve_thread != curthread) {
3023 		SFMMU_HASH_UNLOCK(hmebp);
3024 		mutex_enter(&hblk_reserve_lock);
3025 		mutex_exit(&hblk_reserve_lock);
3026 		SFMMU_STAT(sf_hblk_reserve_hit);
3027 		SFMMU_HASH_LOCK(hmebp);
3028 		goto ttearray_realloc;
3029 	}
3030 
3031 	if (hmeblkp == NULL) {
3032 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
3033 		    hblktag, flags, rid);
3034 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3035 		ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3036 	} else {
3037 		/*
3038 		 * It is possible for 8k and 64k hblks to collide since they
3039 		 * have the same rehash value. This is because we
3040 		 * lazily free hblks and 8K/64K blks could be lingering.
3041 		 * If we find size mismatch we free the block and & try again.
3042 		 */
3043 		if (get_hblk_ttesz(hmeblkp) != size) {
3044 			ASSERT(!hmeblkp->hblk_vcnt);
3045 			ASSERT(!hmeblkp->hblk_hmecnt);
3046 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3047 			    &list, 0);
3048 			goto ttearray_realloc;
3049 		}
3050 		if (hmeblkp->hblk_shw_bit) {
3051 			/*
3052 			 * if the hblk was previously used as a shadow hblk then
3053 			 * we will change it to a normal hblk
3054 			 */
3055 			ASSERT(!hmeblkp->hblk_shared);
3056 			if (hmeblkp->hblk_shw_mask) {
3057 				sfmmu_shadow_hcleanup(sfmmup, hmeblkp, hmebp);
3058 				ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3059 				goto ttearray_realloc;
3060 			} else {
3061 				hmeblkp->hblk_shw_bit = 0;
3062 			}
3063 		}
3064 		SFMMU_STAT(sf_hblk_hit);
3065 	}
3066 
3067 	/*
3068 	 * hat_memload() should never call kmem_cache_free() for kernel hmeblks;
3069 	 * see block comment showing the stacktrace in sfmmu_hblk_alloc();
3070 	 * set the flag parameter to 1 so that sfmmu_hblks_list_purge() will
3071 	 * just add these hmeblks to the per-cpu pending queue.
3072 	 */
3073 	sfmmu_hblks_list_purge(&list, 1);
3074 
3075 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
3076 	ASSERT(!hmeblkp->hblk_shw_bit);
3077 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3078 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3079 	ASSERT(hmeblkp->hblk_tag.htag_rid == rid);
3080 
3081 	return (hmeblkp);
3082 }
3083 
3084 /*
3085  * Function adds a tte entry into the hmeblk. It returns 0 if successful and 1
3086  * otherwise.
3087  */
3088 static int
3089 sfmmu_tteload_addentry(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, tte_t *ttep,
3090 	caddr_t vaddr, page_t **pps, uint_t flags, uint_t rid)
3091 {
3092 	page_t *pp = *pps;
3093 	int hmenum, size, remap;
3094 	tte_t tteold, flush_tte;
3095 #ifdef DEBUG
3096 	tte_t orig_old;
3097 #endif /* DEBUG */
3098 	struct sf_hment *sfhme;
3099 	kmutex_t *pml, *pmtx;
3100 	hatlock_t *hatlockp;
3101 	int myflt;
3102 
3103 	/*
3104 	 * remove this panic when we decide to let user virtual address
3105 	 * space be >= USERLIMIT.
3106 	 */
3107 	if (!TTE_IS_PRIVILEGED(ttep) && vaddr >= (caddr_t)USERLIMIT)
3108 		panic("user addr %p in kernel space", (void *)vaddr);
3109 #if defined(TTE_IS_GLOBAL)
3110 	if (TTE_IS_GLOBAL(ttep))
3111 		panic("sfmmu_tteload: creating global tte");
3112 #endif
3113 
3114 #ifdef DEBUG
3115 	if (pf_is_memory(sfmmu_ttetopfn(ttep, vaddr)) &&
3116 	    !TTE_IS_PCACHEABLE(ttep) && !sfmmu_allow_nc_trans)
3117 		panic("sfmmu_tteload: non cacheable memory tte");
3118 #endif /* DEBUG */
3119 
3120 	/* don't simulate dirty bit for writeable ISM/DISM mappings */
3121 	if ((flags & HAT_LOAD_SHARE) && TTE_IS_WRITABLE(ttep)) {
3122 		TTE_SET_REF(ttep);
3123 		TTE_SET_MOD(ttep);
3124 	}
3125 
3126 	if ((flags & HAT_LOAD_SHARE) || !TTE_IS_REF(ttep) ||
3127 	    !TTE_IS_MOD(ttep)) {
3128 		/*
3129 		 * Don't load TSB for dummy as in ISM.  Also don't preload
3130 		 * the TSB if the TTE isn't writable since we're likely to
3131 		 * fault on it again -- preloading can be fairly expensive.
3132 		 */
3133 		flags |= SFMMU_NO_TSBLOAD;
3134 	}
3135 
3136 	size = TTE_CSZ(ttep);
3137 	switch (size) {
3138 	case TTE8K:
3139 		SFMMU_STAT(sf_tteload8k);
3140 		break;
3141 	case TTE64K:
3142 		SFMMU_STAT(sf_tteload64k);
3143 		break;
3144 	case TTE512K:
3145 		SFMMU_STAT(sf_tteload512k);
3146 		break;
3147 	case TTE4M:
3148 		SFMMU_STAT(sf_tteload4m);
3149 		break;
3150 	case (TTE32M):
3151 		SFMMU_STAT(sf_tteload32m);
3152 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3153 		break;
3154 	case (TTE256M):
3155 		SFMMU_STAT(sf_tteload256m);
3156 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3157 		break;
3158 	}
3159 
3160 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
3161 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
3162 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3163 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3164 
3165 	HBLKTOHME_IDX(sfhme, hmeblkp, vaddr, hmenum);
3166 
3167 	/*
3168 	 * Need to grab mlist lock here so that pageunload
3169 	 * will not change tte behind us.
3170 	 */
3171 	if (pp) {
3172 		pml = sfmmu_mlist_enter(pp);
3173 	}
3174 
3175 	sfmmu_copytte(&sfhme->hme_tte, &tteold);
3176 	/*
3177 	 * Look for corresponding hment and if valid verify
3178 	 * pfns are equal.
3179 	 */
3180 	remap = TTE_IS_VALID(&tteold);
3181 	if (remap) {
3182 		pfn_t	new_pfn, old_pfn;
3183 
3184 		old_pfn = TTE_TO_PFN(vaddr, &tteold);
3185 		new_pfn = TTE_TO_PFN(vaddr, ttep);
3186 
3187 		if (flags & HAT_LOAD_REMAP) {
3188 			/* make sure we are remapping same type of pages */
3189 			if (pf_is_memory(old_pfn) != pf_is_memory(new_pfn)) {
3190 				panic("sfmmu_tteload - tte remap io<->memory");
3191 			}
3192 			if (old_pfn != new_pfn &&
3193 			    (pp != NULL || sfhme->hme_page != NULL)) {
3194 				panic("sfmmu_tteload - tte remap pp != NULL");
3195 			}
3196 		} else if (old_pfn != new_pfn) {
3197 			panic("sfmmu_tteload - tte remap, hmeblkp 0x%p",
3198 			    (void *)hmeblkp);
3199 		}
3200 		ASSERT(TTE_CSZ(&tteold) == TTE_CSZ(ttep));
3201 	}
3202 
3203 	if (pp) {
3204 		if (size == TTE8K) {
3205 #ifdef VAC
3206 			/*
3207 			 * Handle VAC consistency
3208 			 */
3209 			if (!remap && (cache & CACHE_VAC) && !PP_ISNC(pp)) {
3210 				sfmmu_vac_conflict(sfmmup, vaddr, pp);
3211 			}
3212 #endif
3213 
3214 			if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3215 				pmtx = sfmmu_page_enter(pp);
3216 				PP_CLRRO(pp);
3217 				sfmmu_page_exit(pmtx);
3218 			} else if (!PP_ISMAPPED(pp) &&
3219 			    (!TTE_IS_WRITABLE(ttep)) && !(PP_ISMOD(pp))) {
3220 				pmtx = sfmmu_page_enter(pp);
3221 				if (!(PP_ISMOD(pp))) {
3222 					PP_SETRO(pp);
3223 				}
3224 				sfmmu_page_exit(pmtx);
3225 			}
3226 
3227 		} else if (sfmmu_pagearray_setup(vaddr, pps, ttep, remap)) {
3228 			/*
3229 			 * sfmmu_pagearray_setup failed so return
3230 			 */
3231 			sfmmu_mlist_exit(pml);
3232 			return (1);
3233 		}
3234 	}
3235 
3236 	/*
3237 	 * Make sure hment is not on a mapping list.
3238 	 */
3239 	ASSERT(remap || (sfhme->hme_page == NULL));
3240 
3241 	/* if it is not a remap then hme->next better be NULL */
3242 	ASSERT((!remap) ? sfhme->hme_next == NULL : 1);
3243 
3244 	if (flags & HAT_LOAD_LOCK) {
3245 		if ((hmeblkp->hblk_lckcnt + 1) >= MAX_HBLK_LCKCNT) {
3246 			panic("too high lckcnt-hmeblk %p",
3247 			    (void *)hmeblkp);
3248 		}
3249 		atomic_add_32(&hmeblkp->hblk_lckcnt, 1);
3250 
3251 		HBLK_STACK_TRACE(hmeblkp, HBLK_LOCK);
3252 	}
3253 
3254 #ifdef VAC
3255 	if (pp && PP_ISNC(pp)) {
3256 		/*
3257 		 * If the physical page is marked to be uncacheable, like
3258 		 * by a vac conflict, make sure the new mapping is also
3259 		 * uncacheable.
3260 		 */
3261 		TTE_CLR_VCACHEABLE(ttep);
3262 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
3263 	}
3264 #endif
3265 	ttep->tte_hmenum = hmenum;
3266 
3267 #ifdef DEBUG
3268 	orig_old = tteold;
3269 #endif /* DEBUG */
3270 
3271 	while (sfmmu_modifytte_try(&tteold, ttep, &sfhme->hme_tte) < 0) {
3272 		if ((sfmmup == KHATID) &&
3273 		    (flags & (HAT_LOAD_LOCK | HAT_LOAD_REMAP))) {
3274 			sfmmu_copytte(&sfhme->hme_tte, &tteold);
3275 		}
3276 #ifdef DEBUG
3277 		chk_tte(&orig_old, &tteold, ttep, hmeblkp);
3278 #endif /* DEBUG */
3279 	}
3280 	ASSERT(TTE_IS_VALID(&sfhme->hme_tte));
3281 
3282 	if (!TTE_IS_VALID(&tteold)) {
3283 
3284 		atomic_add_16(&hmeblkp->hblk_vcnt, 1);
3285 		if (rid == SFMMU_INVALID_SHMERID) {
3286 			atomic_add_long(&sfmmup->sfmmu_ttecnt[size], 1);
3287 		} else {
3288 			sf_srd_t *srdp = sfmmup->sfmmu_srdp;
3289 			sf_region_t *rgnp = srdp->srd_hmergnp[rid];
3290 			/*
3291 			 * We already accounted for region ttecnt's in sfmmu
3292 			 * during hat_join_region() processing. Here we
3293 			 * only update ttecnt's in region struture.
3294 			 */
3295 			atomic_add_long(&rgnp->rgn_ttecnt[size], 1);
3296 		}
3297 	}
3298 
3299 	myflt = (astosfmmu(curthread->t_procp->p_as) == sfmmup);
3300 	if (size > TTE8K && (flags & HAT_LOAD_SHARE) == 0 &&
3301 	    sfmmup != ksfmmup) {
3302 		uchar_t tteflag = 1 << size;
3303 		if (rid == SFMMU_INVALID_SHMERID) {
3304 			if (!(sfmmup->sfmmu_tteflags & tteflag)) {
3305 				hatlockp = sfmmu_hat_enter(sfmmup);
3306 				sfmmup->sfmmu_tteflags |= tteflag;
3307 				sfmmu_hat_exit(hatlockp);
3308 			}
3309 		} else if (!(sfmmup->sfmmu_rtteflags & tteflag)) {
3310 			hatlockp = sfmmu_hat_enter(sfmmup);
3311 			sfmmup->sfmmu_rtteflags |= tteflag;
3312 			sfmmu_hat_exit(hatlockp);
3313 		}
3314 		/*
3315 		 * Update the current CPU tsbmiss area, so the current thread
3316 		 * won't need to take the tsbmiss for the new pagesize.
3317 		 * The other threads in the process will update their tsb
3318 		 * miss area lazily in sfmmu_tsbmiss_exception() when they
3319 		 * fail to find the translation for a newly added pagesize.
3320 		 */
3321 		if (size > TTE64K && myflt) {
3322 			struct tsbmiss *tsbmp;
3323 			kpreempt_disable();
3324 			tsbmp = &tsbmiss_area[CPU->cpu_id];
3325 			if (rid == SFMMU_INVALID_SHMERID) {
3326 				if (!(tsbmp->uhat_tteflags & tteflag)) {
3327 					tsbmp->uhat_tteflags |= tteflag;
3328 				}
3329 			} else {
3330 				if (!(tsbmp->uhat_rtteflags & tteflag)) {
3331 					tsbmp->uhat_rtteflags |= tteflag;
3332 				}
3333 			}
3334 			kpreempt_enable();
3335 		}
3336 	}
3337 
3338 	if (size >= TTE4M && (flags & HAT_LOAD_TEXT) &&
3339 	    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
3340 		hatlockp = sfmmu_hat_enter(sfmmup);
3341 		SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
3342 		sfmmu_hat_exit(hatlockp);
3343 	}
3344 
3345 	flush_tte.tte_intlo = (tteold.tte_intlo ^ ttep->tte_intlo) &
3346 	    hw_tte.tte_intlo;
3347 	flush_tte.tte_inthi = (tteold.tte_inthi ^ ttep->tte_inthi) &
3348 	    hw_tte.tte_inthi;
3349 
3350 	if (remap && (flush_tte.tte_inthi || flush_tte.tte_intlo)) {
3351 		/*
3352 		 * If remap and new tte differs from old tte we need
3353 		 * to sync the mod bit and flush TLB/TSB.  We don't
3354 		 * need to sync ref bit because we currently always set
3355 		 * ref bit in tteload.
3356 		 */
3357 		ASSERT(TTE_IS_REF(ttep));
3358 		if (TTE_IS_MOD(&tteold)) {
3359 			sfmmu_ttesync(sfmmup, vaddr, &tteold, pp);
3360 		}
3361 		/*
3362 		 * hwtte bits shouldn't change for SRD hmeblks as long as SRD
3363 		 * hmes are only used for read only text. Adding this code for
3364 		 * completeness and future use of shared hmeblks with writable
3365 		 * mappings of VMODSORT vnodes.
3366 		 */
3367 		if (hmeblkp->hblk_shared) {
3368 			cpuset_t cpuset = sfmmu_rgntlb_demap(vaddr,
3369 			    sfmmup->sfmmu_srdp->srd_hmergnp[rid], hmeblkp, 1);
3370 			xt_sync(cpuset);
3371 			SFMMU_STAT_ADD(sf_region_remap_demap, 1);
3372 		} else {
3373 			sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 0);
3374 			xt_sync(sfmmup->sfmmu_cpusran);
3375 		}
3376 	}
3377 
3378 	if ((flags & SFMMU_NO_TSBLOAD) == 0) {
3379 		/*
3380 		 * We only preload 8K and 4M mappings into the TSB, since
3381 		 * 64K and 512K mappings are replicated and hence don't
3382 		 * have a single, unique TSB entry. Ditto for 32M/256M.
3383 		 */
3384 		if (size == TTE8K || size == TTE4M) {
3385 			sf_scd_t *scdp;
3386 			hatlockp = sfmmu_hat_enter(sfmmup);
3387 			/*
3388 			 * Don't preload private TSB if the mapping is used
3389 			 * by the shctx in the SCD.
3390 			 */
3391 			scdp = sfmmup->sfmmu_scdp;
3392 			if (rid == SFMMU_INVALID_SHMERID || scdp == NULL ||
3393 			    !SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
3394 				sfmmu_load_tsb(sfmmup, vaddr, &sfhme->hme_tte,
3395 				    size);
3396 			}
3397 			sfmmu_hat_exit(hatlockp);
3398 		}
3399 	}
3400 	if (pp) {
3401 		if (!remap) {
3402 			HME_ADD(sfhme, pp);
3403 			atomic_add_16(&hmeblkp->hblk_hmecnt, 1);
3404 			ASSERT(hmeblkp->hblk_hmecnt > 0);
3405 
3406 			/*
3407 			 * Cannot ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
3408 			 * see pageunload() for comment.
3409 			 */
3410 		}
3411 		sfmmu_mlist_exit(pml);
3412 	}
3413 
3414 	return (0);
3415 }
3416 /*
3417  * Function unlocks hash bucket.
3418  */
3419 static void
3420 sfmmu_tteload_release_hashbucket(struct hmehash_bucket *hmebp)
3421 {
3422 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3423 	SFMMU_HASH_UNLOCK(hmebp);
3424 }
3425 
3426 /*
3427  * function which checks and sets up page array for a large
3428  * translation.  Will set p_vcolor, p_index, p_ro fields.
3429  * Assumes addr and pfnum of first page are properly aligned.
3430  * Will check for physical contiguity. If check fails it return
3431  * non null.
3432  */
3433 static int
3434 sfmmu_pagearray_setup(caddr_t addr, page_t **pps, tte_t *ttep, int remap)
3435 {
3436 	int 	i, index, ttesz;
3437 	pfn_t	pfnum;
3438 	pgcnt_t	npgs;
3439 	page_t *pp, *pp1;
3440 	kmutex_t *pmtx;
3441 #ifdef VAC
3442 	int osz;
3443 	int cflags = 0;
3444 	int vac_err = 0;
3445 #endif
3446 	int newidx = 0;
3447 
3448 	ttesz = TTE_CSZ(ttep);
3449 
3450 	ASSERT(ttesz > TTE8K);
3451 
3452 	npgs = TTEPAGES(ttesz);
3453 	index = PAGESZ_TO_INDEX(ttesz);
3454 
3455 	pfnum = (*pps)->p_pagenum;
3456 	ASSERT(IS_P2ALIGNED(pfnum, npgs));
3457 
3458 	/*
3459 	 * Save the first pp so we can do HAT_TMPNC at the end.
3460 	 */
3461 	pp1 = *pps;
3462 #ifdef VAC
3463 	osz = fnd_mapping_sz(pp1);
3464 #endif
3465 
3466 	for (i = 0; i < npgs; i++, pps++) {
3467 		pp = *pps;
3468 		ASSERT(PAGE_LOCKED(pp));
3469 		ASSERT(pp->p_szc >= ttesz);
3470 		ASSERT(pp->p_szc == pp1->p_szc);
3471 		ASSERT(sfmmu_mlist_held(pp));
3472 
3473 		/*
3474 		 * XXX is it possible to maintain P_RO on the root only?
3475 		 */
3476 		if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3477 			pmtx = sfmmu_page_enter(pp);
3478 			PP_CLRRO(pp);
3479 			sfmmu_page_exit(pmtx);
3480 		} else if (!PP_ISMAPPED(pp) && !TTE_IS_WRITABLE(ttep) &&
3481 		    !PP_ISMOD(pp)) {
3482 			pmtx = sfmmu_page_enter(pp);
3483 			if (!(PP_ISMOD(pp))) {
3484 				PP_SETRO(pp);
3485 			}
3486 			sfmmu_page_exit(pmtx);
3487 		}
3488 
3489 		/*
3490 		 * If this is a remap we skip vac & contiguity checks.
3491 		 */
3492 		if (remap)
3493 			continue;
3494 
3495 		/*
3496 		 * set p_vcolor and detect any vac conflicts.
3497 		 */
3498 #ifdef VAC
3499 		if (vac_err == 0) {
3500 			vac_err = sfmmu_vacconflict_array(addr, pp, &cflags);
3501 
3502 		}
3503 #endif
3504 
3505 		/*
3506 		 * Save current index in case we need to undo it.
3507 		 * Note: "PAGESZ_TO_INDEX(sz)	(1 << (sz))"
3508 		 *	"SFMMU_INDEX_SHIFT	6"
3509 		 *	 "SFMMU_INDEX_MASK	((1 << SFMMU_INDEX_SHIFT) - 1)"
3510 		 *	 "PP_MAPINDEX(p_index)	(p_index & SFMMU_INDEX_MASK)"
3511 		 *
3512 		 * So:	index = PAGESZ_TO_INDEX(ttesz);
3513 		 *	if ttesz == 1 then index = 0x2
3514 		 *		    2 then index = 0x4
3515 		 *		    3 then index = 0x8
3516 		 *		    4 then index = 0x10
3517 		 *		    5 then index = 0x20
3518 		 * The code below checks if it's a new pagesize (ie, newidx)
3519 		 * in case we need to take it back out of p_index,
3520 		 * and then or's the new index into the existing index.
3521 		 */
3522 		if ((PP_MAPINDEX(pp) & index) == 0)
3523 			newidx = 1;
3524 		pp->p_index = (PP_MAPINDEX(pp) | index);
3525 
3526 		/*
3527 		 * contiguity check
3528 		 */
3529 		if (pp->p_pagenum != pfnum) {
3530 			/*
3531 			 * If we fail the contiguity test then
3532 			 * the only thing we need to fix is the p_index field.
3533 			 * We might get a few extra flushes but since this
3534 			 * path is rare that is ok.  The p_ro field will
3535 			 * get automatically fixed on the next tteload to
3536 			 * the page.  NO TNC bit is set yet.
3537 			 */
3538 			while (i >= 0) {
3539 				pp = *pps;
3540 				if (newidx)
3541 					pp->p_index = (PP_MAPINDEX(pp) &
3542 					    ~index);
3543 				pps--;
3544 				i--;
3545 			}
3546 			return (1);
3547 		}
3548 		pfnum++;
3549 		addr += MMU_PAGESIZE;
3550 	}
3551 
3552 #ifdef VAC
3553 	if (vac_err) {
3554 		if (ttesz > osz) {
3555 			/*
3556 			 * There are some smaller mappings that causes vac
3557 			 * conflicts. Convert all existing small mappings to
3558 			 * TNC.
3559 			 */
3560 			SFMMU_STAT_ADD(sf_uncache_conflict, npgs);
3561 			sfmmu_page_cache_array(pp1, HAT_TMPNC, CACHE_FLUSH,
3562 			    npgs);
3563 		} else {
3564 			/* EMPTY */
3565 			/*
3566 			 * If there exists an big page mapping,
3567 			 * that means the whole existing big page
3568 			 * has TNC setting already. No need to covert to
3569 			 * TNC again.
3570 			 */
3571 			ASSERT(PP_ISTNC(pp1));
3572 		}
3573 	}
3574 #endif	/* VAC */
3575 
3576 	return (0);
3577 }
3578 
3579 #ifdef VAC
3580 /*
3581  * Routine that detects vac consistency for a large page. It also
3582  * sets virtual color for all pp's for this big mapping.
3583  */
3584 static int
3585 sfmmu_vacconflict_array(caddr_t addr, page_t *pp, int *cflags)
3586 {
3587 	int vcolor, ocolor;
3588 
3589 	ASSERT(sfmmu_mlist_held(pp));
3590 
3591 	if (PP_ISNC(pp)) {
3592 		return (HAT_TMPNC);
3593 	}
3594 
3595 	vcolor = addr_to_vcolor(addr);
3596 	if (PP_NEWPAGE(pp)) {
3597 		PP_SET_VCOLOR(pp, vcolor);
3598 		return (0);
3599 	}
3600 
3601 	ocolor = PP_GET_VCOLOR(pp);
3602 	if (ocolor == vcolor) {
3603 		return (0);
3604 	}
3605 
3606 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
3607 		/*
3608 		 * Previous user of page had a differnet color
3609 		 * but since there are no current users
3610 		 * we just flush the cache and change the color.
3611 		 * As an optimization for large pages we flush the
3612 		 * entire cache of that color and set a flag.
3613 		 */
3614 		SFMMU_STAT(sf_pgcolor_conflict);
3615 		if (!CacheColor_IsFlushed(*cflags, ocolor)) {
3616 			CacheColor_SetFlushed(*cflags, ocolor);
3617 			sfmmu_cache_flushcolor(ocolor, pp->p_pagenum);
3618 		}
3619 		PP_SET_VCOLOR(pp, vcolor);
3620 		return (0);
3621 	}
3622 
3623 	/*
3624 	 * We got a real conflict with a current mapping.
3625 	 * set flags to start unencaching all mappings
3626 	 * and return failure so we restart looping
3627 	 * the pp array from the beginning.
3628 	 */
3629 	return (HAT_TMPNC);
3630 }
3631 #endif	/* VAC */
3632 
3633 /*
3634  * creates a large page shadow hmeblk for a tte.
3635  * The purpose of this routine is to allow us to do quick unloads because
3636  * the vm layer can easily pass a very large but sparsely populated range.
3637  */
3638 static struct hme_blk *
3639 sfmmu_shadow_hcreate(sfmmu_t *sfmmup, caddr_t vaddr, int ttesz, uint_t flags)
3640 {
3641 	struct hmehash_bucket *hmebp;
3642 	hmeblk_tag hblktag;
3643 	int hmeshift, size, vshift;
3644 	uint_t shw_mask, newshw_mask;
3645 	struct hme_blk *hmeblkp;
3646 
3647 	ASSERT(sfmmup != KHATID);
3648 	if (mmu_page_sizes == max_mmu_page_sizes) {
3649 		ASSERT(ttesz < TTE256M);
3650 	} else {
3651 		ASSERT(ttesz < TTE4M);
3652 		ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
3653 		ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
3654 	}
3655 
3656 	if (ttesz == TTE8K) {
3657 		size = TTE512K;
3658 	} else {
3659 		size = ++ttesz;
3660 	}
3661 
3662 	hblktag.htag_id = sfmmup;
3663 	hmeshift = HME_HASH_SHIFT(size);
3664 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
3665 	hblktag.htag_rehash = HME_HASH_REHASH(size);
3666 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3667 	hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
3668 
3669 	SFMMU_HASH_LOCK(hmebp);
3670 
3671 	HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
3672 	ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
3673 	if (hmeblkp == NULL) {
3674 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
3675 		    hblktag, flags, SFMMU_INVALID_SHMERID);
3676 	}
3677 	ASSERT(hmeblkp);
3678 	if (!hmeblkp->hblk_shw_mask) {
3679 		/*
3680 		 * if this is a unused hblk it was just allocated or could
3681 		 * potentially be a previous large page hblk so we need to
3682 		 * set the shadow bit.
3683 		 */
3684 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3685 		hmeblkp->hblk_shw_bit = 1;
3686 	} else if (hmeblkp->hblk_shw_bit == 0) {
3687 		panic("sfmmu_shadow_hcreate: shw bit not set in hmeblkp 0x%p",
3688 		    (void *)hmeblkp);
3689 	}
3690 	ASSERT(hmeblkp->hblk_shw_bit == 1);
3691 	ASSERT(!hmeblkp->hblk_shared);
3692 	vshift = vaddr_to_vshift(hblktag, vaddr, size);
3693 	ASSERT(vshift < 8);
3694 	/*
3695 	 * Atomically set shw mask bit
3696 	 */
3697 	do {
3698 		shw_mask = hmeblkp->hblk_shw_mask;
3699 		newshw_mask = shw_mask | (1 << vshift);
3700 		newshw_mask = cas32(&hmeblkp->hblk_shw_mask, shw_mask,
3701 		    newshw_mask);
3702 	} while (newshw_mask != shw_mask);
3703 
3704 	SFMMU_HASH_UNLOCK(hmebp);
3705 
3706 	return (hmeblkp);
3707 }
3708 
3709 /*
3710  * This routine cleanup a previous shadow hmeblk and changes it to
3711  * a regular hblk.  This happens rarely but it is possible
3712  * when a process wants to use large pages and there are hblks still
3713  * lying around from the previous as that used these hmeblks.
3714  * The alternative was to cleanup the shadow hblks at unload time
3715  * but since so few user processes actually use large pages, it is
3716  * better to be lazy and cleanup at this time.
3717  */
3718 static void
3719 sfmmu_shadow_hcleanup(sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
3720 	struct hmehash_bucket *hmebp)
3721 {
3722 	caddr_t addr, endaddr;
3723 	int hashno, size;
3724 
3725 	ASSERT(hmeblkp->hblk_shw_bit);
3726 	ASSERT(!hmeblkp->hblk_shared);
3727 
3728 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3729 
3730 	if (!hmeblkp->hblk_shw_mask) {
3731 		hmeblkp->hblk_shw_bit = 0;
3732 		return;
3733 	}
3734 	addr = (caddr_t)get_hblk_base(hmeblkp);
3735 	endaddr = get_hblk_endaddr(hmeblkp);
3736 	size = get_hblk_ttesz(hmeblkp);
3737 	hashno = size - 1;
3738 	ASSERT(hashno > 0);
3739 	SFMMU_HASH_UNLOCK(hmebp);
3740 
3741 	sfmmu_free_hblks(sfmmup, addr, endaddr, hashno);
3742 
3743 	SFMMU_HASH_LOCK(hmebp);
3744 }
3745 
3746 static void
3747 sfmmu_free_hblks(sfmmu_t *sfmmup, caddr_t addr, caddr_t endaddr,
3748 	int hashno)
3749 {
3750 	int hmeshift, shadow = 0;
3751 	hmeblk_tag hblktag;
3752 	struct hmehash_bucket *hmebp;
3753 	struct hme_blk *hmeblkp;
3754 	struct hme_blk *nx_hblk, *pr_hblk, *list = NULL;
3755 
3756 	ASSERT(hashno > 0);
3757 	hblktag.htag_id = sfmmup;
3758 	hblktag.htag_rehash = hashno;
3759 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3760 
3761 	hmeshift = HME_HASH_SHIFT(hashno);
3762 
3763 	while (addr < endaddr) {
3764 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3765 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3766 		SFMMU_HASH_LOCK(hmebp);
3767 		/* inline HME_HASH_SEARCH */
3768 		hmeblkp = hmebp->hmeblkp;
3769 		pr_hblk = NULL;
3770 		while (hmeblkp) {
3771 			if (HTAGS_EQ(hmeblkp->hblk_tag, hblktag)) {
3772 				/* found hme_blk */
3773 				ASSERT(!hmeblkp->hblk_shared);
3774 				if (hmeblkp->hblk_shw_bit) {
3775 					if (hmeblkp->hblk_shw_mask) {
3776 						shadow = 1;
3777 						sfmmu_shadow_hcleanup(sfmmup,
3778 						    hmeblkp, hmebp);
3779 						break;
3780 					} else {
3781 						hmeblkp->hblk_shw_bit = 0;
3782 					}
3783 				}
3784 
3785 				/*
3786 				 * Hblk_hmecnt and hblk_vcnt could be non zero
3787 				 * since hblk_unload() does not gurantee that.
3788 				 *
3789 				 * XXX - this could cause tteload() to spin
3790 				 * where sfmmu_shadow_hcleanup() is called.
3791 				 */
3792 			}
3793 
3794 			nx_hblk = hmeblkp->hblk_next;
3795 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
3796 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3797 				    &list, 0);
3798 			} else {
3799 				pr_hblk = hmeblkp;
3800 			}
3801 			hmeblkp = nx_hblk;
3802 		}
3803 
3804 		SFMMU_HASH_UNLOCK(hmebp);
3805 
3806 		if (shadow) {
3807 			/*
3808 			 * We found another shadow hblk so cleaned its
3809 			 * children.  We need to go back and cleanup
3810 			 * the original hblk so we don't change the
3811 			 * addr.
3812 			 */
3813 			shadow = 0;
3814 		} else {
3815 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
3816 			    (1 << hmeshift));
3817 		}
3818 	}
3819 	sfmmu_hblks_list_purge(&list, 0);
3820 }
3821 
3822 /*
3823  * This routine's job is to delete stale invalid shared hmeregions hmeblks that
3824  * may still linger on after pageunload.
3825  */
3826 static void
3827 sfmmu_cleanup_rhblk(sf_srd_t *srdp, caddr_t addr, uint_t rid, int ttesz)
3828 {
3829 	int hmeshift;
3830 	hmeblk_tag hblktag;
3831 	struct hmehash_bucket *hmebp;
3832 	struct hme_blk *hmeblkp;
3833 	struct hme_blk *pr_hblk;
3834 	struct hme_blk *list = NULL;
3835 
3836 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3837 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3838 
3839 	hmeshift = HME_HASH_SHIFT(ttesz);
3840 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3841 	hblktag.htag_rehash = ttesz;
3842 	hblktag.htag_rid = rid;
3843 	hblktag.htag_id = srdp;
3844 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3845 
3846 	SFMMU_HASH_LOCK(hmebp);
3847 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3848 	if (hmeblkp != NULL) {
3849 		ASSERT(hmeblkp->hblk_shared);
3850 		ASSERT(!hmeblkp->hblk_shw_bit);
3851 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3852 			panic("sfmmu_cleanup_rhblk: valid hmeblk");
3853 		}
3854 		ASSERT(!hmeblkp->hblk_lckcnt);
3855 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3856 		    &list, 0);
3857 	}
3858 	SFMMU_HASH_UNLOCK(hmebp);
3859 	sfmmu_hblks_list_purge(&list, 0);
3860 }
3861 
3862 /* ARGSUSED */
3863 static void
3864 sfmmu_rgn_cb_noop(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
3865     size_t r_size, void *r_obj, u_offset_t r_objoff)
3866 {
3867 }
3868 
3869 /*
3870  * Searches for an hmeblk which maps addr, then unloads this mapping
3871  * and updates *eaddrp, if the hmeblk is found.
3872  */
3873 static void
3874 sfmmu_unload_hmeregion_va(sf_srd_t *srdp, uint_t rid, caddr_t addr,
3875     caddr_t eaddr, int ttesz, caddr_t *eaddrp)
3876 {
3877 	int hmeshift;
3878 	hmeblk_tag hblktag;
3879 	struct hmehash_bucket *hmebp;
3880 	struct hme_blk *hmeblkp;
3881 	struct hme_blk *pr_hblk;
3882 	struct hme_blk *list = NULL;
3883 
3884 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3885 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3886 	ASSERT(ttesz >= HBLK_MIN_TTESZ);
3887 
3888 	hmeshift = HME_HASH_SHIFT(ttesz);
3889 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3890 	hblktag.htag_rehash = ttesz;
3891 	hblktag.htag_rid = rid;
3892 	hblktag.htag_id = srdp;
3893 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3894 
3895 	SFMMU_HASH_LOCK(hmebp);
3896 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3897 	if (hmeblkp != NULL) {
3898 		ASSERT(hmeblkp->hblk_shared);
3899 		ASSERT(!hmeblkp->hblk_lckcnt);
3900 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3901 			*eaddrp = sfmmu_hblk_unload(NULL, hmeblkp, addr,
3902 			    eaddr, NULL, HAT_UNLOAD);
3903 			ASSERT(*eaddrp > addr);
3904 		}
3905 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3906 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3907 		    &list, 0);
3908 	}
3909 	SFMMU_HASH_UNLOCK(hmebp);
3910 	sfmmu_hblks_list_purge(&list, 0);
3911 }
3912 
3913 static void
3914 sfmmu_unload_hmeregion(sf_srd_t *srdp, sf_region_t *rgnp)
3915 {
3916 	int ttesz = rgnp->rgn_pgszc;
3917 	size_t rsz = rgnp->rgn_size;
3918 	caddr_t rsaddr = rgnp->rgn_saddr;
3919 	caddr_t readdr = rsaddr + rsz;
3920 	caddr_t rhsaddr;
3921 	caddr_t va;
3922 	uint_t rid = rgnp->rgn_id;
3923 	caddr_t cbsaddr;
3924 	caddr_t cbeaddr;
3925 	hat_rgn_cb_func_t rcbfunc;
3926 	ulong_t cnt;
3927 
3928 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3929 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3930 
3931 	ASSERT(IS_P2ALIGNED(rsaddr, TTEBYTES(ttesz)));
3932 	ASSERT(IS_P2ALIGNED(rsz, TTEBYTES(ttesz)));
3933 	if (ttesz < HBLK_MIN_TTESZ) {
3934 		ttesz = HBLK_MIN_TTESZ;
3935 		rhsaddr = (caddr_t)P2ALIGN((uintptr_t)rsaddr, HBLK_MIN_BYTES);
3936 	} else {
3937 		rhsaddr = rsaddr;
3938 	}
3939 
3940 	if ((rcbfunc = rgnp->rgn_cb_function) == NULL) {
3941 		rcbfunc = sfmmu_rgn_cb_noop;
3942 	}
3943 
3944 	while (ttesz >= HBLK_MIN_TTESZ) {
3945 		cbsaddr = rsaddr;
3946 		cbeaddr = rsaddr;
3947 		if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
3948 			ttesz--;
3949 			continue;
3950 		}
3951 		cnt = 0;
3952 		va = rsaddr;
3953 		while (va < readdr) {
3954 			ASSERT(va >= rhsaddr);
3955 			if (va != cbeaddr) {
3956 				if (cbeaddr != cbsaddr) {
3957 					ASSERT(cbeaddr > cbsaddr);
3958 					(*rcbfunc)(cbsaddr, cbeaddr,
3959 					    rsaddr, rsz, rgnp->rgn_obj,
3960 					    rgnp->rgn_objoff);
3961 				}
3962 				cbsaddr = va;
3963 				cbeaddr = va;
3964 			}
3965 			sfmmu_unload_hmeregion_va(srdp, rid, va, readdr,
3966 			    ttesz, &cbeaddr);
3967 			cnt++;
3968 			va = rhsaddr + (cnt << TTE_PAGE_SHIFT(ttesz));
3969 		}
3970 		if (cbeaddr != cbsaddr) {
3971 			ASSERT(cbeaddr > cbsaddr);
3972 			(*rcbfunc)(cbsaddr, cbeaddr, rsaddr,
3973 			    rsz, rgnp->rgn_obj,
3974 			    rgnp->rgn_objoff);
3975 		}
3976 		ttesz--;
3977 	}
3978 }
3979 
3980 /*
3981  * Release one hardware address translation lock on the given address range.
3982  */
3983 void
3984 hat_unlock(struct hat *sfmmup, caddr_t addr, size_t len)
3985 {
3986 	struct hmehash_bucket *hmebp;
3987 	hmeblk_tag hblktag;
3988 	int hmeshift, hashno = 1;
3989 	struct hme_blk *hmeblkp, *list = NULL;
3990 	caddr_t endaddr;
3991 
3992 	ASSERT(sfmmup != NULL);
3993 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3994 
3995 	ASSERT((sfmmup == ksfmmup) ||
3996 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
3997 	ASSERT((len & MMU_PAGEOFFSET) == 0);
3998 	endaddr = addr + len;
3999 	hblktag.htag_id = sfmmup;
4000 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4001 
4002 	/*
4003 	 * Spitfire supports 4 page sizes.
4004 	 * Most pages are expected to be of the smallest page size (8K) and
4005 	 * these will not need to be rehashed. 64K pages also don't need to be
4006 	 * rehashed because an hmeblk spans 64K of address space. 512K pages
4007 	 * might need 1 rehash and and 4M pages might need 2 rehashes.
4008 	 */
4009 	while (addr < endaddr) {
4010 		hmeshift = HME_HASH_SHIFT(hashno);
4011 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
4012 		hblktag.htag_rehash = hashno;
4013 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
4014 
4015 		SFMMU_HASH_LOCK(hmebp);
4016 
4017 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
4018 		if (hmeblkp != NULL) {
4019 			ASSERT(!hmeblkp->hblk_shared);
4020 			/*
4021 			 * If we encounter a shadow hmeblk then
4022 			 * we know there are no valid hmeblks mapping
4023 			 * this address at this size or larger.
4024 			 * Just increment address by the smallest
4025 			 * page size.
4026 			 */
4027 			if (hmeblkp->hblk_shw_bit) {
4028 				addr += MMU_PAGESIZE;
4029 			} else {
4030 				addr = sfmmu_hblk_unlock(hmeblkp, addr,
4031 				    endaddr);
4032 			}
4033 			SFMMU_HASH_UNLOCK(hmebp);
4034 			hashno = 1;
4035 			continue;
4036 		}
4037 		SFMMU_HASH_UNLOCK(hmebp);
4038 
4039 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
4040 			/*
4041 			 * We have traversed the whole list and rehashed
4042 			 * if necessary without finding the address to unlock
4043 			 * which should never happen.
4044 			 */
4045 			panic("sfmmu_unlock: addr not found. "
4046 			    "addr %p hat %p", (void *)addr, (void *)sfmmup);
4047 		} else {
4048 			hashno++;
4049 		}
4050 	}
4051 
4052 	sfmmu_hblks_list_purge(&list, 0);
4053 }
4054 
4055 void
4056 hat_unlock_region(struct hat *sfmmup, caddr_t addr, size_t len,
4057     hat_region_cookie_t rcookie)
4058 {
4059 	sf_srd_t *srdp;
4060 	sf_region_t *rgnp;
4061 	int ttesz;
4062 	uint_t rid;
4063 	caddr_t eaddr;
4064 	caddr_t va;
4065 	int hmeshift;
4066 	hmeblk_tag hblktag;
4067 	struct hmehash_bucket *hmebp;
4068 	struct hme_blk *hmeblkp;
4069 	struct hme_blk *pr_hblk;
4070 	struct hme_blk *list;
4071 
4072 	if (rcookie == HAT_INVALID_REGION_COOKIE) {
4073 		hat_unlock(sfmmup, addr, len);
4074 		return;
4075 	}
4076 
4077 	ASSERT(sfmmup != NULL);
4078 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4079 	ASSERT(sfmmup != ksfmmup);
4080 
4081 	srdp = sfmmup->sfmmu_srdp;
4082 	rid = (uint_t)((uint64_t)rcookie);
4083 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
4084 	eaddr = addr + len;
4085 	va = addr;
4086 	list = NULL;
4087 	rgnp = srdp->srd_hmergnp[rid];
4088 	SFMMU_VALIDATE_HMERID(sfmmup, rid, addr, len);
4089 
4090 	ASSERT(IS_P2ALIGNED(addr, TTEBYTES(rgnp->rgn_pgszc)));
4091 	ASSERT(IS_P2ALIGNED(len, TTEBYTES(rgnp->rgn_pgszc)));
4092 	if (rgnp->rgn_pgszc < HBLK_MIN_TTESZ) {
4093 		ttesz = HBLK_MIN_TTESZ;
4094 	} else {
4095 		ttesz = rgnp->rgn_pgszc;
4096 	}
4097 	while (va < eaddr) {
4098 		while (ttesz < rgnp->rgn_pgszc &&
4099 		    IS_P2ALIGNED(va, TTEBYTES(ttesz + 1))) {
4100 			ttesz++;
4101 		}
4102 		while (ttesz >= HBLK_MIN_TTESZ) {
4103 			if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
4104 				ttesz--;
4105 				continue;
4106 			}
4107 			hmeshift = HME_HASH_SHIFT(ttesz);
4108 			hblktag.htag_bspage = HME_HASH_BSPAGE(va, hmeshift);
4109 			hblktag.htag_rehash = ttesz;
4110 			hblktag.htag_rid = rid;
4111 			hblktag.htag_id = srdp;
4112 			hmebp = HME_HASH_FUNCTION(srdp, va, hmeshift);
4113 			SFMMU_HASH_LOCK(hmebp);
4114 			HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk,
4115 			    &list);
4116 			if (hmeblkp == NULL) {
4117 				SFMMU_HASH_UNLOCK(hmebp);
4118 				ttesz--;
4119 				continue;
4120 			}
4121 			ASSERT(hmeblkp->hblk_shared);
4122 			va = sfmmu_hblk_unlock(hmeblkp, va, eaddr);
4123 			ASSERT(va >= eaddr ||
4124 			    IS_P2ALIGNED((uintptr_t)va, TTEBYTES(ttesz)));
4125 			SFMMU_HASH_UNLOCK(hmebp);
4126 			break;
4127 		}
4128 		if (ttesz < HBLK_MIN_TTESZ) {
4129 			panic("hat_unlock_region: addr not found "
4130 			    "addr %p hat %p", (void *)va, (void *)sfmmup);
4131 		}
4132 	}
4133 	sfmmu_hblks_list_purge(&list, 0);
4134 }
4135 
4136 /*
4137  * Function to unlock a range of addresses in an hmeblk.  It returns the
4138  * next address that needs to be unlocked.
4139  * Should be called with the hash lock held.
4140  */
4141 static caddr_t
4142 sfmmu_hblk_unlock(struct hme_blk *hmeblkp, caddr_t addr, caddr_t endaddr)
4143 {
4144 	struct sf_hment *sfhme;
4145 	tte_t tteold, ttemod;
4146 	int ttesz, ret;
4147 
4148 	ASSERT(in_hblk_range(hmeblkp, addr));
4149 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4150 
4151 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4152 	ttesz = get_hblk_ttesz(hmeblkp);
4153 
4154 	HBLKTOHME(sfhme, hmeblkp, addr);
4155 	while (addr < endaddr) {
4156 readtte:
4157 		sfmmu_copytte(&sfhme->hme_tte, &tteold);
4158 		if (TTE_IS_VALID(&tteold)) {
4159 
4160 			ttemod = tteold;
4161 
4162 			ret = sfmmu_modifytte_try(&tteold, &ttemod,
4163 			    &sfhme->hme_tte);
4164 
4165 			if (ret < 0)
4166 				goto readtte;
4167 
4168 			if (hmeblkp->hblk_lckcnt == 0)
4169 				panic("zero hblk lckcnt");
4170 
4171 			if (((uintptr_t)addr + TTEBYTES(ttesz)) >
4172 			    (uintptr_t)endaddr)
4173 				panic("can't unlock large tte");
4174 
4175 			ASSERT(hmeblkp->hblk_lckcnt > 0);
4176 			atomic_add_32(&hmeblkp->hblk_lckcnt, -1);
4177 			HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
4178 		} else {
4179 			panic("sfmmu_hblk_unlock: invalid tte");
4180 		}
4181 		addr += TTEBYTES(ttesz);
4182 		sfhme++;
4183 	}
4184 	return (addr);
4185 }
4186 
4187 /*
4188  * Physical Address Mapping Framework
4189  *
4190  * General rules:
4191  *
4192  * (1) Applies only to seg_kmem memory pages. To make things easier,
4193  *     seg_kpm addresses are also accepted by the routines, but nothing
4194  *     is done with them since by definition their PA mappings are static.
4195  * (2) hat_add_callback() may only be called while holding the page lock
4196  *     SE_SHARED or SE_EXCL of the underlying page (e.g., as_pagelock()),
4197  *     or passing HAC_PAGELOCK flag.
4198  * (3) prehandler() and posthandler() may not call hat_add_callback() or
4199  *     hat_delete_callback(), nor should they allocate memory. Post quiesce
4200  *     callbacks may not sleep or acquire adaptive mutex locks.
4201  * (4) Either prehandler() or posthandler() (but not both) may be specified
4202  *     as being NULL.  Specifying an errhandler() is optional.
4203  *
4204  * Details of using the framework:
4205  *
4206  * registering a callback (hat_register_callback())
4207  *
4208  *	Pass prehandler, posthandler, errhandler addresses
4209  *	as described below. If capture_cpus argument is nonzero,
4210  *	suspend callback to the prehandler will occur with CPUs
4211  *	captured and executing xc_loop() and CPUs will remain
4212  *	captured until after the posthandler suspend callback
4213  *	occurs.
4214  *
4215  * adding a callback (hat_add_callback())
4216  *
4217  *      as_pagelock();
4218  *	hat_add_callback();
4219  *      save returned pfn in private data structures or program registers;
4220  *      as_pageunlock();
4221  *
4222  * prehandler()
4223  *
4224  *	Stop all accesses by physical address to this memory page.
4225  *	Called twice: the first, PRESUSPEND, is a context safe to acquire
4226  *	adaptive locks. The second, SUSPEND, is called at high PIL with
4227  *	CPUs captured so adaptive locks may NOT be acquired (and all spin
4228  *	locks must be XCALL_PIL or higher locks).
4229  *
4230  *	May return the following errors:
4231  *		EIO:	A fatal error has occurred. This will result in panic.
4232  *		EAGAIN:	The page cannot be suspended. This will fail the
4233  *			relocation.
4234  *		0:	Success.
4235  *
4236  * posthandler()
4237  *
4238  *      Save new pfn in private data structures or program registers;
4239  *	not allowed to fail (non-zero return values will result in panic).
4240  *
4241  * errhandler()
4242  *
4243  *	called when an error occurs related to the callback.  Currently
4244  *	the only such error is HAT_CB_ERR_LEAKED which indicates that
4245  *	a page is being freed, but there are still outstanding callback(s)
4246  *	registered on the page.
4247  *
4248  * removing a callback (hat_delete_callback(); e.g., prior to freeing memory)
4249  *
4250  *	stop using physical address
4251  *	hat_delete_callback();
4252  *
4253  */
4254 
4255 /*
4256  * Register a callback class.  Each subsystem should do this once and
4257  * cache the id_t returned for use in setting up and tearing down callbacks.
4258  *
4259  * There is no facility for removing callback IDs once they are created;
4260  * the "key" should be unique for each module, so in case a module is unloaded
4261  * and subsequently re-loaded, we can recycle the module's previous entry.
4262  */
4263 id_t
4264 hat_register_callback(int key,
4265 	int (*prehandler)(caddr_t, uint_t, uint_t, void *),
4266 	int (*posthandler)(caddr_t, uint_t, uint_t, void *, pfn_t),
4267 	int (*errhandler)(caddr_t, uint_t, uint_t, void *),
4268 	int capture_cpus)
4269 {
4270 	id_t id;
4271 
4272 	/*
4273 	 * Search the table for a pre-existing callback associated with
4274 	 * the identifier "key".  If one exists, we re-use that entry in
4275 	 * the table for this instance, otherwise we assign the next
4276 	 * available table slot.
4277 	 */
4278 	for (id = 0; id < sfmmu_max_cb_id; id++) {
4279 		if (sfmmu_cb_table[id].key == key)
4280 			break;
4281 	}
4282 
4283 	if (id == sfmmu_max_cb_id) {
4284 		id = sfmmu_cb_nextid++;
4285 		if (id >= sfmmu_max_cb_id)
4286 			panic("hat_register_callback: out of callback IDs");
4287 	}
4288 
4289 	ASSERT(prehandler != NULL || posthandler != NULL);
4290 
4291 	sfmmu_cb_table[id].key = key;
4292 	sfmmu_cb_table[id].prehandler = prehandler;
4293 	sfmmu_cb_table[id].posthandler = posthandler;
4294 	sfmmu_cb_table[id].errhandler = errhandler;
4295 	sfmmu_cb_table[id].capture_cpus = capture_cpus;
4296 
4297 	return (id);
4298 }
4299 
4300 #define	HAC_COOKIE_NONE	(void *)-1
4301 
4302 /*
4303  * Add relocation callbacks to the specified addr/len which will be called
4304  * when relocating the associated page. See the description of pre and
4305  * posthandler above for more details.
4306  *
4307  * If HAC_PAGELOCK is included in flags, the underlying memory page is
4308  * locked internally so the caller must be able to deal with the callback
4309  * running even before this function has returned.  If HAC_PAGELOCK is not
4310  * set, it is assumed that the underlying memory pages are locked.
4311  *
4312  * Since the caller must track the individual page boundaries anyway,
4313  * we only allow a callback to be added to a single page (large
4314  * or small).  Thus [addr, addr + len) MUST be contained within a single
4315  * page.
4316  *
4317  * Registering multiple callbacks on the same [addr, addr+len) is supported,
4318  * _provided_that_ a unique parameter is specified for each callback.
4319  * If multiple callbacks are registered on the same range the callback will
4320  * be invoked with each unique parameter. Registering the same callback with
4321  * the same argument more than once will result in corrupted kernel state.
4322  *
4323  * Returns the pfn of the underlying kernel page in *rpfn
4324  * on success, or PFN_INVALID on failure.
4325  *
4326  * cookiep (if passed) provides storage space for an opaque cookie
4327  * to return later to hat_delete_callback(). This cookie makes the callback
4328  * deletion significantly quicker by avoiding a potentially lengthy hash
4329  * search.
4330  *
4331  * Returns values:
4332  *    0:      success
4333  *    ENOMEM: memory allocation failure (e.g. flags was passed as HAC_NOSLEEP)
4334  *    EINVAL: callback ID is not valid
4335  *    ENXIO:  ["vaddr", "vaddr" + len) is not mapped in the kernel's address
4336  *            space
4337  *    ERANGE: ["vaddr", "vaddr" + len) crosses a page boundary
4338  */
4339 int
4340 hat_add_callback(id_t callback_id, caddr_t vaddr, uint_t len, uint_t flags,
4341 	void *pvt, pfn_t *rpfn, void **cookiep)
4342 {
4343 	struct 		hmehash_bucket *hmebp;
4344 	hmeblk_tag 	hblktag;
4345 	struct hme_blk	*hmeblkp;
4346 	int 		hmeshift, hashno;
4347 	caddr_t 	saddr, eaddr, baseaddr;
4348 	struct pa_hment *pahmep;
4349 	struct sf_hment *sfhmep, *osfhmep;
4350 	kmutex_t	*pml;
4351 	tte_t   	tte;
4352 	page_t		*pp;
4353 	vnode_t		*vp;
4354 	u_offset_t	off;
4355 	pfn_t		pfn;
4356 	int		kmflags = (flags & HAC_SLEEP)? KM_SLEEP : KM_NOSLEEP;
4357 	int		locked = 0;
4358 
4359 	/*
4360 	 * For KPM mappings, just return the physical address since we
4361 	 * don't need to register any callbacks.
4362 	 */
4363 	if (IS_KPM_ADDR(vaddr)) {
4364 		uint64_t paddr;
4365 		SFMMU_KPM_VTOP(vaddr, paddr);
4366 		*rpfn = btop(paddr);
4367 		if (cookiep != NULL)
4368 			*cookiep = HAC_COOKIE_NONE;
4369 		return (0);
4370 	}
4371 
4372 	if (callback_id < (id_t)0 || callback_id >= sfmmu_cb_nextid) {
4373 		*rpfn = PFN_INVALID;
4374 		return (EINVAL);
4375 	}
4376 
4377 	if ((pahmep = kmem_cache_alloc(pa_hment_cache, kmflags)) == NULL) {
4378 		*rpfn = PFN_INVALID;
4379 		return (ENOMEM);
4380 	}
4381 
4382 	sfhmep = &pahmep->sfment;
4383 
4384 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4385 	eaddr = saddr + len;
4386 
4387 rehash:
4388 	/* Find the mapping(s) for this page */
4389 	for (hashno = TTE64K, hmeblkp = NULL;
4390 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4391 	    hashno++) {
4392 		hmeshift = HME_HASH_SHIFT(hashno);
4393 		hblktag.htag_id = ksfmmup;
4394 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4395 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4396 		hblktag.htag_rehash = hashno;
4397 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4398 
4399 		SFMMU_HASH_LOCK(hmebp);
4400 
4401 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4402 
4403 		if (hmeblkp == NULL)
4404 			SFMMU_HASH_UNLOCK(hmebp);
4405 	}
4406 
4407 	if (hmeblkp == NULL) {
4408 		kmem_cache_free(pa_hment_cache, pahmep);
4409 		*rpfn = PFN_INVALID;
4410 		return (ENXIO);
4411 	}
4412 
4413 	ASSERT(!hmeblkp->hblk_shared);
4414 
4415 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4416 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4417 
4418 	if (!TTE_IS_VALID(&tte)) {
4419 		SFMMU_HASH_UNLOCK(hmebp);
4420 		kmem_cache_free(pa_hment_cache, pahmep);
4421 		*rpfn = PFN_INVALID;
4422 		return (ENXIO);
4423 	}
4424 
4425 	/*
4426 	 * Make sure the boundaries for the callback fall within this
4427 	 * single mapping.
4428 	 */
4429 	baseaddr = (caddr_t)get_hblk_base(hmeblkp);
4430 	ASSERT(saddr >= baseaddr);
4431 	if (eaddr > saddr + TTEBYTES(TTE_CSZ(&tte))) {
4432 		SFMMU_HASH_UNLOCK(hmebp);
4433 		kmem_cache_free(pa_hment_cache, pahmep);
4434 		*rpfn = PFN_INVALID;
4435 		return (ERANGE);
4436 	}
4437 
4438 	pfn = sfmmu_ttetopfn(&tte, vaddr);
4439 
4440 	/*
4441 	 * The pfn may not have a page_t underneath in which case we
4442 	 * just return it. This can happen if we are doing I/O to a
4443 	 * static portion of the kernel's address space, for instance.
4444 	 */
4445 	pp = osfhmep->hme_page;
4446 	if (pp == NULL) {
4447 		SFMMU_HASH_UNLOCK(hmebp);
4448 		kmem_cache_free(pa_hment_cache, pahmep);
4449 		*rpfn = pfn;
4450 		if (cookiep)
4451 			*cookiep = HAC_COOKIE_NONE;
4452 		return (0);
4453 	}
4454 	ASSERT(pp == PP_PAGEROOT(pp));
4455 
4456 	vp = pp->p_vnode;
4457 	off = pp->p_offset;
4458 
4459 	pml = sfmmu_mlist_enter(pp);
4460 
4461 	if (flags & HAC_PAGELOCK) {
4462 		if (!page_trylock(pp, SE_SHARED)) {
4463 			/*
4464 			 * Somebody is holding SE_EXCL lock. Might
4465 			 * even be hat_page_relocate(). Drop all
4466 			 * our locks, lookup the page in &kvp, and
4467 			 * retry. If it doesn't exist in &kvp and &zvp,
4468 			 * then we must be dealing with a kernel mapped
4469 			 * page which doesn't actually belong to
4470 			 * segkmem so we punt.
4471 			 */
4472 			sfmmu_mlist_exit(pml);
4473 			SFMMU_HASH_UNLOCK(hmebp);
4474 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4475 
4476 			/* check zvp before giving up */
4477 			if (pp == NULL)
4478 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4479 				    SE_SHARED);
4480 
4481 			/* Okay, we didn't find it, give up */
4482 			if (pp == NULL) {
4483 				kmem_cache_free(pa_hment_cache, pahmep);
4484 				*rpfn = pfn;
4485 				if (cookiep)
4486 					*cookiep = HAC_COOKIE_NONE;
4487 				return (0);
4488 			}
4489 			page_unlock(pp);
4490 			goto rehash;
4491 		}
4492 		locked = 1;
4493 	}
4494 
4495 	if (!PAGE_LOCKED(pp) && !panicstr)
4496 		panic("hat_add_callback: page 0x%p not locked", (void *)pp);
4497 
4498 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4499 	    pp->p_offset != off) {
4500 		/*
4501 		 * The page moved before we got our hands on it.  Drop
4502 		 * all the locks and try again.
4503 		 */
4504 		ASSERT((flags & HAC_PAGELOCK) != 0);
4505 		sfmmu_mlist_exit(pml);
4506 		SFMMU_HASH_UNLOCK(hmebp);
4507 		page_unlock(pp);
4508 		locked = 0;
4509 		goto rehash;
4510 	}
4511 
4512 	if (!VN_ISKAS(vp)) {
4513 		/*
4514 		 * This is not a segkmem page but another page which
4515 		 * has been kernel mapped. It had better have at least
4516 		 * a share lock on it. Return the pfn.
4517 		 */
4518 		sfmmu_mlist_exit(pml);
4519 		SFMMU_HASH_UNLOCK(hmebp);
4520 		if (locked)
4521 			page_unlock(pp);
4522 		kmem_cache_free(pa_hment_cache, pahmep);
4523 		ASSERT(PAGE_LOCKED(pp));
4524 		*rpfn = pfn;
4525 		if (cookiep)
4526 			*cookiep = HAC_COOKIE_NONE;
4527 		return (0);
4528 	}
4529 
4530 	/*
4531 	 * Setup this pa_hment and link its embedded dummy sf_hment into
4532 	 * the mapping list.
4533 	 */
4534 	pp->p_share++;
4535 	pahmep->cb_id = callback_id;
4536 	pahmep->addr = vaddr;
4537 	pahmep->len = len;
4538 	pahmep->refcnt = 1;
4539 	pahmep->flags = 0;
4540 	pahmep->pvt = pvt;
4541 
4542 	sfhmep->hme_tte.ll = 0;
4543 	sfhmep->hme_data = pahmep;
4544 	sfhmep->hme_prev = osfhmep;
4545 	sfhmep->hme_next = osfhmep->hme_next;
4546 
4547 	if (osfhmep->hme_next)
4548 		osfhmep->hme_next->hme_prev = sfhmep;
4549 
4550 	osfhmep->hme_next = sfhmep;
4551 
4552 	sfmmu_mlist_exit(pml);
4553 	SFMMU_HASH_UNLOCK(hmebp);
4554 
4555 	if (locked)
4556 		page_unlock(pp);
4557 
4558 	*rpfn = pfn;
4559 	if (cookiep)
4560 		*cookiep = (void *)pahmep;
4561 
4562 	return (0);
4563 }
4564 
4565 /*
4566  * Remove the relocation callbacks from the specified addr/len.
4567  */
4568 void
4569 hat_delete_callback(caddr_t vaddr, uint_t len, void *pvt, uint_t flags,
4570 	void *cookie)
4571 {
4572 	struct		hmehash_bucket *hmebp;
4573 	hmeblk_tag	hblktag;
4574 	struct hme_blk	*hmeblkp;
4575 	int		hmeshift, hashno;
4576 	caddr_t		saddr;
4577 	struct pa_hment	*pahmep;
4578 	struct sf_hment	*sfhmep, *osfhmep;
4579 	kmutex_t	*pml;
4580 	tte_t		tte;
4581 	page_t		*pp;
4582 	vnode_t		*vp;
4583 	u_offset_t	off;
4584 	int		locked = 0;
4585 
4586 	/*
4587 	 * If the cookie is HAC_COOKIE_NONE then there is no pa_hment to
4588 	 * remove so just return.
4589 	 */
4590 	if (cookie == HAC_COOKIE_NONE || IS_KPM_ADDR(vaddr))
4591 		return;
4592 
4593 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4594 
4595 rehash:
4596 	/* Find the mapping(s) for this page */
4597 	for (hashno = TTE64K, hmeblkp = NULL;
4598 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4599 	    hashno++) {
4600 		hmeshift = HME_HASH_SHIFT(hashno);
4601 		hblktag.htag_id = ksfmmup;
4602 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4603 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4604 		hblktag.htag_rehash = hashno;
4605 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4606 
4607 		SFMMU_HASH_LOCK(hmebp);
4608 
4609 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4610 
4611 		if (hmeblkp == NULL)
4612 			SFMMU_HASH_UNLOCK(hmebp);
4613 	}
4614 
4615 	if (hmeblkp == NULL)
4616 		return;
4617 
4618 	ASSERT(!hmeblkp->hblk_shared);
4619 
4620 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4621 
4622 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4623 	if (!TTE_IS_VALID(&tte)) {
4624 		SFMMU_HASH_UNLOCK(hmebp);
4625 		return;
4626 	}
4627 
4628 	pp = osfhmep->hme_page;
4629 	if (pp == NULL) {
4630 		SFMMU_HASH_UNLOCK(hmebp);
4631 		ASSERT(cookie == NULL);
4632 		return;
4633 	}
4634 
4635 	vp = pp->p_vnode;
4636 	off = pp->p_offset;
4637 
4638 	pml = sfmmu_mlist_enter(pp);
4639 
4640 	if (flags & HAC_PAGELOCK) {
4641 		if (!page_trylock(pp, SE_SHARED)) {
4642 			/*
4643 			 * Somebody is holding SE_EXCL lock. Might
4644 			 * even be hat_page_relocate(). Drop all
4645 			 * our locks, lookup the page in &kvp, and
4646 			 * retry. If it doesn't exist in &kvp and &zvp,
4647 			 * then we must be dealing with a kernel mapped
4648 			 * page which doesn't actually belong to
4649 			 * segkmem so we punt.
4650 			 */
4651 			sfmmu_mlist_exit(pml);
4652 			SFMMU_HASH_UNLOCK(hmebp);
4653 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4654 			/* check zvp before giving up */
4655 			if (pp == NULL)
4656 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4657 				    SE_SHARED);
4658 
4659 			if (pp == NULL) {
4660 				ASSERT(cookie == NULL);
4661 				return;
4662 			}
4663 			page_unlock(pp);
4664 			goto rehash;
4665 		}
4666 		locked = 1;
4667 	}
4668 
4669 	ASSERT(PAGE_LOCKED(pp));
4670 
4671 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4672 	    pp->p_offset != off) {
4673 		/*
4674 		 * The page moved before we got our hands on it.  Drop
4675 		 * all the locks and try again.
4676 		 */
4677 		ASSERT((flags & HAC_PAGELOCK) != 0);
4678 		sfmmu_mlist_exit(pml);
4679 		SFMMU_HASH_UNLOCK(hmebp);
4680 		page_unlock(pp);
4681 		locked = 0;
4682 		goto rehash;
4683 	}
4684 
4685 	if (!VN_ISKAS(vp)) {
4686 		/*
4687 		 * This is not a segkmem page but another page which
4688 		 * has been kernel mapped.
4689 		 */
4690 		sfmmu_mlist_exit(pml);
4691 		SFMMU_HASH_UNLOCK(hmebp);
4692 		if (locked)
4693 			page_unlock(pp);
4694 		ASSERT(cookie == NULL);
4695 		return;
4696 	}
4697 
4698 	if (cookie != NULL) {
4699 		pahmep = (struct pa_hment *)cookie;
4700 		sfhmep = &pahmep->sfment;
4701 	} else {
4702 		for (sfhmep = pp->p_mapping; sfhmep != NULL;
4703 		    sfhmep = sfhmep->hme_next) {
4704 
4705 			/*
4706 			 * skip va<->pa mappings
4707 			 */
4708 			if (!IS_PAHME(sfhmep))
4709 				continue;
4710 
4711 			pahmep = sfhmep->hme_data;
4712 			ASSERT(pahmep != NULL);
4713 
4714 			/*
4715 			 * if pa_hment matches, remove it
4716 			 */
4717 			if ((pahmep->pvt == pvt) &&
4718 			    (pahmep->addr == vaddr) &&
4719 			    (pahmep->len == len)) {
4720 				break;
4721 			}
4722 		}
4723 	}
4724 
4725 	if (sfhmep == NULL) {
4726 		if (!panicstr) {
4727 			panic("hat_delete_callback: pa_hment not found, pp %p",
4728 			    (void *)pp);
4729 		}
4730 		return;
4731 	}
4732 
4733 	/*
4734 	 * Note: at this point a valid kernel mapping must still be
4735 	 * present on this page.
4736 	 */
4737 	pp->p_share--;
4738 	if (pp->p_share <= 0)
4739 		panic("hat_delete_callback: zero p_share");
4740 
4741 	if (--pahmep->refcnt == 0) {
4742 		if (pahmep->flags != 0)
4743 			panic("hat_delete_callback: pa_hment is busy");
4744 
4745 		/*
4746 		 * Remove sfhmep from the mapping list for the page.
4747 		 */
4748 		if (sfhmep->hme_prev) {
4749 			sfhmep->hme_prev->hme_next = sfhmep->hme_next;
4750 		} else {
4751 			pp->p_mapping = sfhmep->hme_next;
4752 		}
4753 
4754 		if (sfhmep->hme_next)
4755 			sfhmep->hme_next->hme_prev = sfhmep->hme_prev;
4756 
4757 		sfmmu_mlist_exit(pml);
4758 		SFMMU_HASH_UNLOCK(hmebp);
4759 
4760 		if (locked)
4761 			page_unlock(pp);
4762 
4763 		kmem_cache_free(pa_hment_cache, pahmep);
4764 		return;
4765 	}
4766 
4767 	sfmmu_mlist_exit(pml);
4768 	SFMMU_HASH_UNLOCK(hmebp);
4769 	if (locked)
4770 		page_unlock(pp);
4771 }
4772 
4773 /*
4774  * hat_probe returns 1 if the translation for the address 'addr' is
4775  * loaded, zero otherwise.
4776  *
4777  * hat_probe should be used only for advisorary purposes because it may
4778  * occasionally return the wrong value. The implementation must guarantee that
4779  * returning the wrong value is a very rare event. hat_probe is used
4780  * to implement optimizations in the segment drivers.
4781  *
4782  */
4783 int
4784 hat_probe(struct hat *sfmmup, caddr_t addr)
4785 {
4786 	pfn_t pfn;
4787 	tte_t tte;
4788 
4789 	ASSERT(sfmmup != NULL);
4790 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4791 
4792 	ASSERT((sfmmup == ksfmmup) ||
4793 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
4794 
4795 	if (sfmmup == ksfmmup) {
4796 		while ((pfn = sfmmu_vatopfn(addr, sfmmup, &tte))
4797 		    == PFN_SUSPENDED) {
4798 			sfmmu_vatopfn_suspended(addr, sfmmup, &tte);
4799 		}
4800 	} else {
4801 		pfn = sfmmu_uvatopfn(addr, sfmmup, NULL);
4802 	}
4803 
4804 	if (pfn != PFN_INVALID)
4805 		return (1);
4806 	else
4807 		return (0);
4808 }
4809 
4810 ssize_t
4811 hat_getpagesize(struct hat *sfmmup, caddr_t addr)
4812 {
4813 	tte_t tte;
4814 
4815 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4816 
4817 	if (sfmmup == ksfmmup) {
4818 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4819 			return (-1);
4820 		}
4821 	} else {
4822 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4823 			return (-1);
4824 		}
4825 	}
4826 
4827 	ASSERT(TTE_IS_VALID(&tte));
4828 	return (TTEBYTES(TTE_CSZ(&tte)));
4829 }
4830 
4831 uint_t
4832 hat_getattr(struct hat *sfmmup, caddr_t addr, uint_t *attr)
4833 {
4834 	tte_t tte;
4835 
4836 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4837 
4838 	if (sfmmup == ksfmmup) {
4839 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4840 			tte.ll = 0;
4841 		}
4842 	} else {
4843 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4844 			tte.ll = 0;
4845 		}
4846 	}
4847 	if (TTE_IS_VALID(&tte)) {
4848 		*attr = sfmmu_ptov_attr(&tte);
4849 		return (0);
4850 	}
4851 	*attr = 0;
4852 	return ((uint_t)0xffffffff);
4853 }
4854 
4855 /*
4856  * Enables more attributes on specified address range (ie. logical OR)
4857  */
4858 void
4859 hat_setattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4860 {
4861 	if (hat->sfmmu_xhat_provider) {
4862 		XHAT_SETATTR(hat, addr, len, attr);
4863 		return;
4864 	} else {
4865 		/*
4866 		 * This must be a CPU HAT. If the address space has
4867 		 * XHATs attached, change attributes for all of them,
4868 		 * just in case
4869 		 */
4870 		ASSERT(hat->sfmmu_as != NULL);
4871 		if (hat->sfmmu_as->a_xhat != NULL)
4872 			xhat_setattr_all(hat->sfmmu_as, addr, len, attr);
4873 	}
4874 
4875 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_SETATTR);
4876 }
4877 
4878 /*
4879  * Assigns attributes to the specified address range.  All the attributes
4880  * are specified.
4881  */
4882 void
4883 hat_chgattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4884 {
4885 	if (hat->sfmmu_xhat_provider) {
4886 		XHAT_CHGATTR(hat, addr, len, attr);
4887 		return;
4888 	} else {
4889 		/*
4890 		 * This must be a CPU HAT. If the address space has
4891 		 * XHATs attached, change attributes for all of them,
4892 		 * just in case
4893 		 */
4894 		ASSERT(hat->sfmmu_as != NULL);
4895 		if (hat->sfmmu_as->a_xhat != NULL)
4896 			xhat_chgattr_all(hat->sfmmu_as, addr, len, attr);
4897 	}
4898 
4899 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CHGATTR);
4900 }
4901 
4902 /*
4903  * Remove attributes on the specified address range (ie. loginal NAND)
4904  */
4905 void
4906 hat_clrattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4907 {
4908 	if (hat->sfmmu_xhat_provider) {
4909 		XHAT_CLRATTR(hat, addr, len, attr);
4910 		return;
4911 	} else {
4912 		/*
4913 		 * This must be a CPU HAT. If the address space has
4914 		 * XHATs attached, change attributes for all of them,
4915 		 * just in case
4916 		 */
4917 		ASSERT(hat->sfmmu_as != NULL);
4918 		if (hat->sfmmu_as->a_xhat != NULL)
4919 			xhat_clrattr_all(hat->sfmmu_as, addr, len, attr);
4920 	}
4921 
4922 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CLRATTR);
4923 }
4924 
4925 /*
4926  * Change attributes on an address range to that specified by attr and mode.
4927  */
4928 static void
4929 sfmmu_chgattr(struct hat *sfmmup, caddr_t addr, size_t len, uint_t attr,
4930 	int mode)
4931 {
4932 	struct hmehash_bucket *hmebp;
4933 	hmeblk_tag hblktag;
4934 	int hmeshift, hashno = 1;
4935 	struct hme_blk *hmeblkp, *list = NULL;
4936 	caddr_t endaddr;
4937 	cpuset_t cpuset;
4938 	demap_range_t dmr;
4939 
4940 	CPUSET_ZERO(cpuset);
4941 
4942 	ASSERT((sfmmup == ksfmmup) ||
4943 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
4944 	ASSERT((len & MMU_PAGEOFFSET) == 0);
4945 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
4946 
4947 	if ((attr & PROT_USER) && (mode != SFMMU_CLRATTR) &&
4948 	    ((addr + len) > (caddr_t)USERLIMIT)) {
4949 		panic("user addr %p in kernel space",
4950 		    (void *)addr);
4951 	}
4952 
4953 	endaddr = addr + len;
4954 	hblktag.htag_id = sfmmup;
4955 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4956 	DEMAP_RANGE_INIT(sfmmup, &dmr);
4957 
4958 	while (addr < endaddr) {
4959 		hmeshift = HME_HASH_SHIFT(hashno);
4960 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
4961 		hblktag.htag_rehash = hashno;
4962 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
4963 
4964 		SFMMU_HASH_LOCK(hmebp);
4965 
4966 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
4967 		if (hmeblkp != NULL) {
4968 			ASSERT(!hmeblkp->hblk_shared);
4969 			/*
4970 			 * We've encountered a shadow hmeblk so skip the range
4971 			 * of the next smaller mapping size.
4972 			 */
4973 			if (hmeblkp->hblk_shw_bit) {
4974 				ASSERT(sfmmup != ksfmmup);
4975 				ASSERT(hashno > 1);
4976 				addr = (caddr_t)P2END((uintptr_t)addr,
4977 				    TTEBYTES(hashno - 1));
4978 			} else {
4979 				addr = sfmmu_hblk_chgattr(sfmmup,
4980 				    hmeblkp, addr, endaddr, &dmr, attr, mode);
4981 			}
4982 			SFMMU_HASH_UNLOCK(hmebp);
4983 			hashno = 1;
4984 			continue;
4985 		}
4986 		SFMMU_HASH_UNLOCK(hmebp);
4987 
4988 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
4989 			/*
4990 			 * We have traversed the whole list and rehashed
4991 			 * if necessary without finding the address to chgattr.
4992 			 * This is ok, so we increment the address by the
4993 			 * smallest hmeblk range for kernel mappings or for
4994 			 * user mappings with no large pages, and the largest
4995 			 * hmeblk range, to account for shadow hmeblks, for
4996 			 * user mappings with large pages and continue.
4997 			 */
4998 			if (sfmmup == ksfmmup)
4999 				addr = (caddr_t)P2END((uintptr_t)addr,
5000 				    TTEBYTES(1));
5001 			else
5002 				addr = (caddr_t)P2END((uintptr_t)addr,
5003 				    TTEBYTES(hashno));
5004 			hashno = 1;
5005 		} else {
5006 			hashno++;
5007 		}
5008 	}
5009 
5010 	sfmmu_hblks_list_purge(&list, 0);
5011 	DEMAP_RANGE_FLUSH(&dmr);
5012 	cpuset = sfmmup->sfmmu_cpusran;
5013 	xt_sync(cpuset);
5014 }
5015 
5016 /*
5017  * This function chgattr on a range of addresses in an hmeblk.  It returns the
5018  * next addres that needs to be chgattr.
5019  * It should be called with the hash lock held.
5020  * XXX It should be possible to optimize chgattr by not flushing every time but
5021  * on the other hand:
5022  * 1. do one flush crosscall.
5023  * 2. only flush if we are increasing permissions (make sure this will work)
5024  */
5025 static caddr_t
5026 sfmmu_hblk_chgattr(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5027 	caddr_t endaddr, demap_range_t *dmrp, uint_t attr, int mode)
5028 {
5029 	tte_t tte, tteattr, tteflags, ttemod;
5030 	struct sf_hment *sfhmep;
5031 	int ttesz;
5032 	struct page *pp = NULL;
5033 	kmutex_t *pml, *pmtx;
5034 	int ret;
5035 	int use_demap_range;
5036 #if defined(SF_ERRATA_57)
5037 	int check_exec;
5038 #endif
5039 
5040 	ASSERT(in_hblk_range(hmeblkp, addr));
5041 	ASSERT(hmeblkp->hblk_shw_bit == 0);
5042 	ASSERT(!hmeblkp->hblk_shared);
5043 
5044 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5045 	ttesz = get_hblk_ttesz(hmeblkp);
5046 
5047 	/*
5048 	 * Flush the current demap region if addresses have been
5049 	 * skipped or the page size doesn't match.
5050 	 */
5051 	use_demap_range = (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp));
5052 	if (use_demap_range) {
5053 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5054 	} else {
5055 		DEMAP_RANGE_FLUSH(dmrp);
5056 	}
5057 
5058 	tteattr.ll = sfmmu_vtop_attr(attr, mode, &tteflags);
5059 #if defined(SF_ERRATA_57)
5060 	check_exec = (sfmmup != ksfmmup) &&
5061 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
5062 	    TTE_IS_EXECUTABLE(&tteattr);
5063 #endif
5064 	HBLKTOHME(sfhmep, hmeblkp, addr);
5065 	while (addr < endaddr) {
5066 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5067 		if (TTE_IS_VALID(&tte)) {
5068 			if ((tte.ll & tteflags.ll) == tteattr.ll) {
5069 				/*
5070 				 * if the new attr is the same as old
5071 				 * continue
5072 				 */
5073 				goto next_addr;
5074 			}
5075 			if (!TTE_IS_WRITABLE(&tteattr)) {
5076 				/*
5077 				 * make sure we clear hw modify bit if we
5078 				 * removing write protections
5079 				 */
5080 				tteflags.tte_intlo |= TTE_HWWR_INT;
5081 			}
5082 
5083 			pml = NULL;
5084 			pp = sfhmep->hme_page;
5085 			if (pp) {
5086 				pml = sfmmu_mlist_enter(pp);
5087 			}
5088 
5089 			if (pp != sfhmep->hme_page) {
5090 				/*
5091 				 * tte must have been unloaded.
5092 				 */
5093 				ASSERT(pml);
5094 				sfmmu_mlist_exit(pml);
5095 				continue;
5096 			}
5097 
5098 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5099 
5100 			ttemod = tte;
5101 			ttemod.ll = (ttemod.ll & ~tteflags.ll) | tteattr.ll;
5102 			ASSERT(TTE_TO_TTEPFN(&ttemod) == TTE_TO_TTEPFN(&tte));
5103 
5104 #if defined(SF_ERRATA_57)
5105 			if (check_exec && addr < errata57_limit)
5106 				ttemod.tte_exec_perm = 0;
5107 #endif
5108 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5109 			    &sfhmep->hme_tte);
5110 
5111 			if (ret < 0) {
5112 				/* tte changed underneath us */
5113 				if (pml) {
5114 					sfmmu_mlist_exit(pml);
5115 				}
5116 				continue;
5117 			}
5118 
5119 			if (tteflags.tte_intlo & TTE_HWWR_INT) {
5120 				/*
5121 				 * need to sync if we are clearing modify bit.
5122 				 */
5123 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5124 			}
5125 
5126 			if (pp && PP_ISRO(pp)) {
5127 				if (tteattr.tte_intlo & TTE_WRPRM_INT) {
5128 					pmtx = sfmmu_page_enter(pp);
5129 					PP_CLRRO(pp);
5130 					sfmmu_page_exit(pmtx);
5131 				}
5132 			}
5133 
5134 			if (ret > 0 && use_demap_range) {
5135 				DEMAP_RANGE_MARKPG(dmrp, addr);
5136 			} else if (ret > 0) {
5137 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5138 			}
5139 
5140 			if (pml) {
5141 				sfmmu_mlist_exit(pml);
5142 			}
5143 		}
5144 next_addr:
5145 		addr += TTEBYTES(ttesz);
5146 		sfhmep++;
5147 		DEMAP_RANGE_NEXTPG(dmrp);
5148 	}
5149 	return (addr);
5150 }
5151 
5152 /*
5153  * This routine converts virtual attributes to physical ones.  It will
5154  * update the tteflags field with the tte mask corresponding to the attributes
5155  * affected and it returns the new attributes.  It will also clear the modify
5156  * bit if we are taking away write permission.  This is necessary since the
5157  * modify bit is the hardware permission bit and we need to clear it in order
5158  * to detect write faults.
5159  */
5160 static uint64_t
5161 sfmmu_vtop_attr(uint_t attr, int mode, tte_t *ttemaskp)
5162 {
5163 	tte_t ttevalue;
5164 
5165 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
5166 
5167 	switch (mode) {
5168 	case SFMMU_CHGATTR:
5169 		/* all attributes specified */
5170 		ttevalue.tte_inthi = MAKE_TTEATTR_INTHI(attr);
5171 		ttevalue.tte_intlo = MAKE_TTEATTR_INTLO(attr);
5172 		ttemaskp->tte_inthi = TTEINTHI_ATTR;
5173 		ttemaskp->tte_intlo = TTEINTLO_ATTR;
5174 		break;
5175 	case SFMMU_SETATTR:
5176 		ASSERT(!(attr & ~HAT_PROT_MASK));
5177 		ttemaskp->ll = 0;
5178 		ttevalue.ll = 0;
5179 		/*
5180 		 * a valid tte implies exec and read for sfmmu
5181 		 * so no need to do anything about them.
5182 		 * since priviledged access implies user access
5183 		 * PROT_USER doesn't make sense either.
5184 		 */
5185 		if (attr & PROT_WRITE) {
5186 			ttemaskp->tte_intlo |= TTE_WRPRM_INT;
5187 			ttevalue.tte_intlo |= TTE_WRPRM_INT;
5188 		}
5189 		break;
5190 	case SFMMU_CLRATTR:
5191 		/* attributes will be nand with current ones */
5192 		if (attr & ~(PROT_WRITE | PROT_USER)) {
5193 			panic("sfmmu: attr %x not supported", attr);
5194 		}
5195 		ttemaskp->ll = 0;
5196 		ttevalue.ll = 0;
5197 		if (attr & PROT_WRITE) {
5198 			/* clear both writable and modify bit */
5199 			ttemaskp->tte_intlo |= TTE_WRPRM_INT | TTE_HWWR_INT;
5200 		}
5201 		if (attr & PROT_USER) {
5202 			ttemaskp->tte_intlo |= TTE_PRIV_INT;
5203 			ttevalue.tte_intlo |= TTE_PRIV_INT;
5204 		}
5205 		break;
5206 	default:
5207 		panic("sfmmu_vtop_attr: bad mode %x", mode);
5208 	}
5209 	ASSERT(TTE_TO_TTEPFN(&ttevalue) == 0);
5210 	return (ttevalue.ll);
5211 }
5212 
5213 static uint_t
5214 sfmmu_ptov_attr(tte_t *ttep)
5215 {
5216 	uint_t attr;
5217 
5218 	ASSERT(TTE_IS_VALID(ttep));
5219 
5220 	attr = PROT_READ;
5221 
5222 	if (TTE_IS_WRITABLE(ttep)) {
5223 		attr |= PROT_WRITE;
5224 	}
5225 	if (TTE_IS_EXECUTABLE(ttep)) {
5226 		attr |= PROT_EXEC;
5227 	}
5228 	if (!TTE_IS_PRIVILEGED(ttep)) {
5229 		attr |= PROT_USER;
5230 	}
5231 	if (TTE_IS_NFO(ttep)) {
5232 		attr |= HAT_NOFAULT;
5233 	}
5234 	if (TTE_IS_NOSYNC(ttep)) {
5235 		attr |= HAT_NOSYNC;
5236 	}
5237 	if (TTE_IS_SIDEFFECT(ttep)) {
5238 		attr |= SFMMU_SIDEFFECT;
5239 	}
5240 	if (!TTE_IS_VCACHEABLE(ttep)) {
5241 		attr |= SFMMU_UNCACHEVTTE;
5242 	}
5243 	if (!TTE_IS_PCACHEABLE(ttep)) {
5244 		attr |= SFMMU_UNCACHEPTTE;
5245 	}
5246 	return (attr);
5247 }
5248 
5249 /*
5250  * hat_chgprot is a deprecated hat call.  New segment drivers
5251  * should store all attributes and use hat_*attr calls.
5252  *
5253  * Change the protections in the virtual address range
5254  * given to the specified virtual protection.  If vprot is ~PROT_WRITE,
5255  * then remove write permission, leaving the other
5256  * permissions unchanged.  If vprot is ~PROT_USER, remove user permissions.
5257  *
5258  */
5259 void
5260 hat_chgprot(struct hat *sfmmup, caddr_t addr, size_t len, uint_t vprot)
5261 {
5262 	struct hmehash_bucket *hmebp;
5263 	hmeblk_tag hblktag;
5264 	int hmeshift, hashno = 1;
5265 	struct hme_blk *hmeblkp, *list = NULL;
5266 	caddr_t endaddr;
5267 	cpuset_t cpuset;
5268 	demap_range_t dmr;
5269 
5270 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5271 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
5272 
5273 	if (sfmmup->sfmmu_xhat_provider) {
5274 		XHAT_CHGPROT(sfmmup, addr, len, vprot);
5275 		return;
5276 	} else {
5277 		/*
5278 		 * This must be a CPU HAT. If the address space has
5279 		 * XHATs attached, change attributes for all of them,
5280 		 * just in case
5281 		 */
5282 		ASSERT(sfmmup->sfmmu_as != NULL);
5283 		if (sfmmup->sfmmu_as->a_xhat != NULL)
5284 			xhat_chgprot_all(sfmmup->sfmmu_as, addr, len, vprot);
5285 	}
5286 
5287 	CPUSET_ZERO(cpuset);
5288 
5289 	if ((vprot != (uint_t)~PROT_WRITE) && (vprot & PROT_USER) &&
5290 	    ((addr + len) > (caddr_t)USERLIMIT)) {
5291 		panic("user addr %p vprot %x in kernel space",
5292 		    (void *)addr, vprot);
5293 	}
5294 	endaddr = addr + len;
5295 	hblktag.htag_id = sfmmup;
5296 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5297 	DEMAP_RANGE_INIT(sfmmup, &dmr);
5298 
5299 	while (addr < endaddr) {
5300 		hmeshift = HME_HASH_SHIFT(hashno);
5301 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5302 		hblktag.htag_rehash = hashno;
5303 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5304 
5305 		SFMMU_HASH_LOCK(hmebp);
5306 
5307 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
5308 		if (hmeblkp != NULL) {
5309 			ASSERT(!hmeblkp->hblk_shared);
5310 			/*
5311 			 * We've encountered a shadow hmeblk so skip the range
5312 			 * of the next smaller mapping size.
5313 			 */
5314 			if (hmeblkp->hblk_shw_bit) {
5315 				ASSERT(sfmmup != ksfmmup);
5316 				ASSERT(hashno > 1);
5317 				addr = (caddr_t)P2END((uintptr_t)addr,
5318 				    TTEBYTES(hashno - 1));
5319 			} else {
5320 				addr = sfmmu_hblk_chgprot(sfmmup, hmeblkp,
5321 				    addr, endaddr, &dmr, vprot);
5322 			}
5323 			SFMMU_HASH_UNLOCK(hmebp);
5324 			hashno = 1;
5325 			continue;
5326 		}
5327 		SFMMU_HASH_UNLOCK(hmebp);
5328 
5329 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
5330 			/*
5331 			 * We have traversed the whole list and rehashed
5332 			 * if necessary without finding the address to chgprot.
5333 			 * This is ok so we increment the address by the
5334 			 * smallest hmeblk range for kernel mappings and the
5335 			 * largest hmeblk range, to account for shadow hmeblks,
5336 			 * for user mappings and continue.
5337 			 */
5338 			if (sfmmup == ksfmmup)
5339 				addr = (caddr_t)P2END((uintptr_t)addr,
5340 				    TTEBYTES(1));
5341 			else
5342 				addr = (caddr_t)P2END((uintptr_t)addr,
5343 				    TTEBYTES(hashno));
5344 			hashno = 1;
5345 		} else {
5346 			hashno++;
5347 		}
5348 	}
5349 
5350 	sfmmu_hblks_list_purge(&list, 0);
5351 	DEMAP_RANGE_FLUSH(&dmr);
5352 	cpuset = sfmmup->sfmmu_cpusran;
5353 	xt_sync(cpuset);
5354 }
5355 
5356 /*
5357  * This function chgprots a range of addresses in an hmeblk.  It returns the
5358  * next addres that needs to be chgprot.
5359  * It should be called with the hash lock held.
5360  * XXX It shold be possible to optimize chgprot by not flushing every time but
5361  * on the other hand:
5362  * 1. do one flush crosscall.
5363  * 2. only flush if we are increasing permissions (make sure this will work)
5364  */
5365 static caddr_t
5366 sfmmu_hblk_chgprot(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5367 	caddr_t endaddr, demap_range_t *dmrp, uint_t vprot)
5368 {
5369 	uint_t pprot;
5370 	tte_t tte, ttemod;
5371 	struct sf_hment *sfhmep;
5372 	uint_t tteflags;
5373 	int ttesz;
5374 	struct page *pp = NULL;
5375 	kmutex_t *pml, *pmtx;
5376 	int ret;
5377 	int use_demap_range;
5378 #if defined(SF_ERRATA_57)
5379 	int check_exec;
5380 #endif
5381 
5382 	ASSERT(in_hblk_range(hmeblkp, addr));
5383 	ASSERT(hmeblkp->hblk_shw_bit == 0);
5384 	ASSERT(!hmeblkp->hblk_shared);
5385 
5386 #ifdef DEBUG
5387 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5388 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5389 		panic("sfmmu_hblk_chgprot: partial chgprot of large page");
5390 	}
5391 #endif /* DEBUG */
5392 
5393 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5394 	ttesz = get_hblk_ttesz(hmeblkp);
5395 
5396 	pprot = sfmmu_vtop_prot(vprot, &tteflags);
5397 #if defined(SF_ERRATA_57)
5398 	check_exec = (sfmmup != ksfmmup) &&
5399 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
5400 	    ((vprot & PROT_EXEC) == PROT_EXEC);
5401 #endif
5402 	HBLKTOHME(sfhmep, hmeblkp, addr);
5403 
5404 	/*
5405 	 * Flush the current demap region if addresses have been
5406 	 * skipped or the page size doesn't match.
5407 	 */
5408 	use_demap_range = (TTEBYTES(ttesz) == MMU_PAGESIZE);
5409 	if (use_demap_range) {
5410 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5411 	} else {
5412 		DEMAP_RANGE_FLUSH(dmrp);
5413 	}
5414 
5415 	while (addr < endaddr) {
5416 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5417 		if (TTE_IS_VALID(&tte)) {
5418 			if (TTE_GET_LOFLAGS(&tte, tteflags) == pprot) {
5419 				/*
5420 				 * if the new protection is the same as old
5421 				 * continue
5422 				 */
5423 				goto next_addr;
5424 			}
5425 			pml = NULL;
5426 			pp = sfhmep->hme_page;
5427 			if (pp) {
5428 				pml = sfmmu_mlist_enter(pp);
5429 			}
5430 			if (pp != sfhmep->hme_page) {
5431 				/*
5432 				 * tte most have been unloaded
5433 				 * underneath us.  Recheck
5434 				 */
5435 				ASSERT(pml);
5436 				sfmmu_mlist_exit(pml);
5437 				continue;
5438 			}
5439 
5440 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5441 
5442 			ttemod = tte;
5443 			TTE_SET_LOFLAGS(&ttemod, tteflags, pprot);
5444 #if defined(SF_ERRATA_57)
5445 			if (check_exec && addr < errata57_limit)
5446 				ttemod.tte_exec_perm = 0;
5447 #endif
5448 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5449 			    &sfhmep->hme_tte);
5450 
5451 			if (ret < 0) {
5452 				/* tte changed underneath us */
5453 				if (pml) {
5454 					sfmmu_mlist_exit(pml);
5455 				}
5456 				continue;
5457 			}
5458 
5459 			if (tteflags & TTE_HWWR_INT) {
5460 				/*
5461 				 * need to sync if we are clearing modify bit.
5462 				 */
5463 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5464 			}
5465 
5466 			if (pp && PP_ISRO(pp)) {
5467 				if (pprot & TTE_WRPRM_INT) {
5468 					pmtx = sfmmu_page_enter(pp);
5469 					PP_CLRRO(pp);
5470 					sfmmu_page_exit(pmtx);
5471 				}
5472 			}
5473 
5474 			if (ret > 0 && use_demap_range) {
5475 				DEMAP_RANGE_MARKPG(dmrp, addr);
5476 			} else if (ret > 0) {
5477 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5478 			}
5479 
5480 			if (pml) {
5481 				sfmmu_mlist_exit(pml);
5482 			}
5483 		}
5484 next_addr:
5485 		addr += TTEBYTES(ttesz);
5486 		sfhmep++;
5487 		DEMAP_RANGE_NEXTPG(dmrp);
5488 	}
5489 	return (addr);
5490 }
5491 
5492 /*
5493  * This routine is deprecated and should only be used by hat_chgprot.
5494  * The correct routine is sfmmu_vtop_attr.
5495  * This routine converts virtual page protections to physical ones.  It will
5496  * update the tteflags field with the tte mask corresponding to the protections
5497  * affected and it returns the new protections.  It will also clear the modify
5498  * bit if we are taking away write permission.  This is necessary since the
5499  * modify bit is the hardware permission bit and we need to clear it in order
5500  * to detect write faults.
5501  * It accepts the following special protections:
5502  * ~PROT_WRITE = remove write permissions.
5503  * ~PROT_USER = remove user permissions.
5504  */
5505 static uint_t
5506 sfmmu_vtop_prot(uint_t vprot, uint_t *tteflagsp)
5507 {
5508 	if (vprot == (uint_t)~PROT_WRITE) {
5509 		*tteflagsp = TTE_WRPRM_INT | TTE_HWWR_INT;
5510 		return (0);		/* will cause wrprm to be cleared */
5511 	}
5512 	if (vprot == (uint_t)~PROT_USER) {
5513 		*tteflagsp = TTE_PRIV_INT;
5514 		return (0);		/* will cause privprm to be cleared */
5515 	}
5516 	if ((vprot == 0) || (vprot == PROT_USER) ||
5517 	    ((vprot & PROT_ALL) != vprot)) {
5518 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5519 	}
5520 
5521 	switch (vprot) {
5522 	case (PROT_READ):
5523 	case (PROT_EXEC):
5524 	case (PROT_EXEC | PROT_READ):
5525 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5526 		return (TTE_PRIV_INT); 		/* set prv and clr wrt */
5527 	case (PROT_WRITE):
5528 	case (PROT_WRITE | PROT_READ):
5529 	case (PROT_EXEC | PROT_WRITE):
5530 	case (PROT_EXEC | PROT_WRITE | PROT_READ):
5531 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5532 		return (TTE_PRIV_INT | TTE_WRPRM_INT); 	/* set prv and wrt */
5533 	case (PROT_USER | PROT_READ):
5534 	case (PROT_USER | PROT_EXEC):
5535 	case (PROT_USER | PROT_EXEC | PROT_READ):
5536 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5537 		return (0); 			/* clr prv and wrt */
5538 	case (PROT_USER | PROT_WRITE):
5539 	case (PROT_USER | PROT_WRITE | PROT_READ):
5540 	case (PROT_USER | PROT_EXEC | PROT_WRITE):
5541 	case (PROT_USER | PROT_EXEC | PROT_WRITE | PROT_READ):
5542 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5543 		return (TTE_WRPRM_INT); 	/* clr prv and set wrt */
5544 	default:
5545 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5546 	}
5547 	return (0);
5548 }
5549 
5550 /*
5551  * Alternate unload for very large virtual ranges. With a true 64 bit VA,
5552  * the normal algorithm would take too long for a very large VA range with
5553  * few real mappings. This routine just walks thru all HMEs in the global
5554  * hash table to find and remove mappings.
5555  */
5556 static void
5557 hat_unload_large_virtual(
5558 	struct hat		*sfmmup,
5559 	caddr_t			startaddr,
5560 	size_t			len,
5561 	uint_t			flags,
5562 	hat_callback_t		*callback)
5563 {
5564 	struct hmehash_bucket *hmebp;
5565 	struct hme_blk *hmeblkp;
5566 	struct hme_blk *pr_hblk = NULL;
5567 	struct hme_blk *nx_hblk;
5568 	struct hme_blk *list = NULL;
5569 	int i;
5570 	demap_range_t dmr, *dmrp;
5571 	cpuset_t cpuset;
5572 	caddr_t	endaddr = startaddr + len;
5573 	caddr_t	sa;
5574 	caddr_t	ea;
5575 	caddr_t	cb_sa[MAX_CB_ADDR];
5576 	caddr_t	cb_ea[MAX_CB_ADDR];
5577 	int	addr_cnt = 0;
5578 	int	a = 0;
5579 
5580 	if (sfmmup->sfmmu_free) {
5581 		dmrp = NULL;
5582 	} else {
5583 		dmrp = &dmr;
5584 		DEMAP_RANGE_INIT(sfmmup, dmrp);
5585 	}
5586 
5587 	/*
5588 	 * Loop through all the hash buckets of HME blocks looking for matches.
5589 	 */
5590 	for (i = 0; i <= UHMEHASH_SZ; i++) {
5591 		hmebp = &uhme_hash[i];
5592 		SFMMU_HASH_LOCK(hmebp);
5593 		hmeblkp = hmebp->hmeblkp;
5594 		pr_hblk = NULL;
5595 		while (hmeblkp) {
5596 			nx_hblk = hmeblkp->hblk_next;
5597 
5598 			/*
5599 			 * skip if not this context, if a shadow block or
5600 			 * if the mapping is not in the requested range
5601 			 */
5602 			if (hmeblkp->hblk_tag.htag_id != sfmmup ||
5603 			    hmeblkp->hblk_shw_bit ||
5604 			    (sa = (caddr_t)get_hblk_base(hmeblkp)) >= endaddr ||
5605 			    (ea = get_hblk_endaddr(hmeblkp)) <= startaddr) {
5606 				pr_hblk = hmeblkp;
5607 				goto next_block;
5608 			}
5609 
5610 			ASSERT(!hmeblkp->hblk_shared);
5611 			/*
5612 			 * unload if there are any current valid mappings
5613 			 */
5614 			if (hmeblkp->hblk_vcnt != 0 ||
5615 			    hmeblkp->hblk_hmecnt != 0)
5616 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
5617 				    sa, ea, dmrp, flags);
5618 
5619 			/*
5620 			 * on unmap we also release the HME block itself, once
5621 			 * all mappings are gone.
5622 			 */
5623 			if ((flags & HAT_UNLOAD_UNMAP) != 0 &&
5624 			    !hmeblkp->hblk_vcnt &&
5625 			    !hmeblkp->hblk_hmecnt) {
5626 				ASSERT(!hmeblkp->hblk_lckcnt);
5627 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
5628 				    &list, 0);
5629 			} else {
5630 				pr_hblk = hmeblkp;
5631 			}
5632 
5633 			if (callback == NULL)
5634 				goto next_block;
5635 
5636 			/*
5637 			 * HME blocks may span more than one page, but we may be
5638 			 * unmapping only one page, so check for a smaller range
5639 			 * for the callback
5640 			 */
5641 			if (sa < startaddr)
5642 				sa = startaddr;
5643 			if (--ea > endaddr)
5644 				ea = endaddr - 1;
5645 
5646 			cb_sa[addr_cnt] = sa;
5647 			cb_ea[addr_cnt] = ea;
5648 			if (++addr_cnt == MAX_CB_ADDR) {
5649 				if (dmrp != NULL) {
5650 					DEMAP_RANGE_FLUSH(dmrp);
5651 					cpuset = sfmmup->sfmmu_cpusran;
5652 					xt_sync(cpuset);
5653 				}
5654 
5655 				for (a = 0; a < MAX_CB_ADDR; ++a) {
5656 					callback->hcb_start_addr = cb_sa[a];
5657 					callback->hcb_end_addr = cb_ea[a];
5658 					callback->hcb_function(callback);
5659 				}
5660 				addr_cnt = 0;
5661 			}
5662 
5663 next_block:
5664 			hmeblkp = nx_hblk;
5665 		}
5666 		SFMMU_HASH_UNLOCK(hmebp);
5667 	}
5668 
5669 	sfmmu_hblks_list_purge(&list, 0);
5670 	if (dmrp != NULL) {
5671 		DEMAP_RANGE_FLUSH(dmrp);
5672 		cpuset = sfmmup->sfmmu_cpusran;
5673 		xt_sync(cpuset);
5674 	}
5675 
5676 	for (a = 0; a < addr_cnt; ++a) {
5677 		callback->hcb_start_addr = cb_sa[a];
5678 		callback->hcb_end_addr = cb_ea[a];
5679 		callback->hcb_function(callback);
5680 	}
5681 
5682 	/*
5683 	 * Check TSB and TLB page sizes if the process isn't exiting.
5684 	 */
5685 	if (!sfmmup->sfmmu_free)
5686 		sfmmu_check_page_sizes(sfmmup, 0);
5687 }
5688 
5689 /*
5690  * Unload all the mappings in the range [addr..addr+len). addr and len must
5691  * be MMU_PAGESIZE aligned.
5692  */
5693 
5694 extern struct seg *segkmap;
5695 #define	ISSEGKMAP(sfmmup, addr) (sfmmup == ksfmmup && \
5696 segkmap->s_base <= (addr) && (addr) < (segkmap->s_base + segkmap->s_size))
5697 
5698 
5699 void
5700 hat_unload_callback(
5701 	struct hat *sfmmup,
5702 	caddr_t addr,
5703 	size_t len,
5704 	uint_t flags,
5705 	hat_callback_t *callback)
5706 {
5707 	struct hmehash_bucket *hmebp;
5708 	hmeblk_tag hblktag;
5709 	int hmeshift, hashno, iskernel;
5710 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
5711 	caddr_t endaddr;
5712 	cpuset_t cpuset;
5713 	int addr_count = 0;
5714 	int a;
5715 	caddr_t cb_start_addr[MAX_CB_ADDR];
5716 	caddr_t cb_end_addr[MAX_CB_ADDR];
5717 	int issegkmap = ISSEGKMAP(sfmmup, addr);
5718 	demap_range_t dmr, *dmrp;
5719 
5720 	if (sfmmup->sfmmu_xhat_provider) {
5721 		XHAT_UNLOAD_CALLBACK(sfmmup, addr, len, flags, callback);
5722 		return;
5723 	} else {
5724 		/*
5725 		 * This must be a CPU HAT. If the address space has
5726 		 * XHATs attached, unload the mappings for all of them,
5727 		 * just in case
5728 		 */
5729 		ASSERT(sfmmup->sfmmu_as != NULL);
5730 		if (sfmmup->sfmmu_as->a_xhat != NULL)
5731 			xhat_unload_callback_all(sfmmup->sfmmu_as, addr,
5732 			    len, flags, callback);
5733 	}
5734 
5735 	ASSERT((sfmmup == ksfmmup) || (flags & HAT_UNLOAD_OTHER) || \
5736 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
5737 
5738 	ASSERT(sfmmup != NULL);
5739 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5740 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
5741 
5742 	/*
5743 	 * Probing through a large VA range (say 63 bits) will be slow, even
5744 	 * at 4 Meg steps between the probes. So, when the virtual address range
5745 	 * is very large, search the HME entries for what to unload.
5746 	 *
5747 	 *	len >> TTE_PAGE_SHIFT(TTE4M) is the # of 4Meg probes we'd need
5748 	 *
5749 	 *	UHMEHASH_SZ is number of hash buckets to examine
5750 	 *
5751 	 */
5752 	if (sfmmup != KHATID && (len >> TTE_PAGE_SHIFT(TTE4M)) > UHMEHASH_SZ) {
5753 		hat_unload_large_virtual(sfmmup, addr, len, flags, callback);
5754 		return;
5755 	}
5756 
5757 	CPUSET_ZERO(cpuset);
5758 
5759 	/*
5760 	 * If the process is exiting, we can save a lot of fuss since
5761 	 * we'll flush the TLB when we free the ctx anyway.
5762 	 */
5763 	if (sfmmup->sfmmu_free)
5764 		dmrp = NULL;
5765 	else
5766 		dmrp = &dmr;
5767 
5768 	DEMAP_RANGE_INIT(sfmmup, dmrp);
5769 	endaddr = addr + len;
5770 	hblktag.htag_id = sfmmup;
5771 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5772 
5773 	/*
5774 	 * It is likely for the vm to call unload over a wide range of
5775 	 * addresses that are actually very sparsely populated by
5776 	 * translations.  In order to speed this up the sfmmu hat supports
5777 	 * the concept of shadow hmeblks. Dummy large page hmeblks that
5778 	 * correspond to actual small translations are allocated at tteload
5779 	 * time and are referred to as shadow hmeblks.  Now, during unload
5780 	 * time, we first check if we have a shadow hmeblk for that
5781 	 * translation.  The absence of one means the corresponding address
5782 	 * range is empty and can be skipped.
5783 	 *
5784 	 * The kernel is an exception to above statement and that is why
5785 	 * we don't use shadow hmeblks and hash starting from the smallest
5786 	 * page size.
5787 	 */
5788 	if (sfmmup == KHATID) {
5789 		iskernel = 1;
5790 		hashno = TTE64K;
5791 	} else {
5792 		iskernel = 0;
5793 		if (mmu_page_sizes == max_mmu_page_sizes) {
5794 			hashno = TTE256M;
5795 		} else {
5796 			hashno = TTE4M;
5797 		}
5798 	}
5799 	while (addr < endaddr) {
5800 		hmeshift = HME_HASH_SHIFT(hashno);
5801 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5802 		hblktag.htag_rehash = hashno;
5803 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5804 
5805 		SFMMU_HASH_LOCK(hmebp);
5806 
5807 		HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
5808 		if (hmeblkp == NULL) {
5809 			/*
5810 			 * didn't find an hmeblk. skip the appropiate
5811 			 * address range.
5812 			 */
5813 			SFMMU_HASH_UNLOCK(hmebp);
5814 			if (iskernel) {
5815 				if (hashno < mmu_hashcnt) {
5816 					hashno++;
5817 					continue;
5818 				} else {
5819 					hashno = TTE64K;
5820 					addr = (caddr_t)roundup((uintptr_t)addr
5821 					    + 1, MMU_PAGESIZE64K);
5822 					continue;
5823 				}
5824 			}
5825 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5826 			    (1 << hmeshift));
5827 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5828 				ASSERT(hashno == TTE64K);
5829 				continue;
5830 			}
5831 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5832 				hashno = TTE512K;
5833 				continue;
5834 			}
5835 			if (mmu_page_sizes == max_mmu_page_sizes) {
5836 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5837 					hashno = TTE4M;
5838 					continue;
5839 				}
5840 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5841 					hashno = TTE32M;
5842 					continue;
5843 				}
5844 				hashno = TTE256M;
5845 				continue;
5846 			} else {
5847 				hashno = TTE4M;
5848 				continue;
5849 			}
5850 		}
5851 		ASSERT(hmeblkp);
5852 		ASSERT(!hmeblkp->hblk_shared);
5853 		if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5854 			/*
5855 			 * If the valid count is zero we can skip the range
5856 			 * mapped by this hmeblk.
5857 			 * We free hblks in the case of HAT_UNMAP.  HAT_UNMAP
5858 			 * is used by segment drivers as a hint
5859 			 * that the mapping resource won't be used any longer.
5860 			 * The best example of this is during exit().
5861 			 */
5862 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5863 			    get_hblk_span(hmeblkp));
5864 			if ((flags & HAT_UNLOAD_UNMAP) ||
5865 			    (iskernel && !issegkmap)) {
5866 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
5867 				    &list, 0);
5868 			}
5869 			SFMMU_HASH_UNLOCK(hmebp);
5870 
5871 			if (iskernel) {
5872 				hashno = TTE64K;
5873 				continue;
5874 			}
5875 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5876 				ASSERT(hashno == TTE64K);
5877 				continue;
5878 			}
5879 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5880 				hashno = TTE512K;
5881 				continue;
5882 			}
5883 			if (mmu_page_sizes == max_mmu_page_sizes) {
5884 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5885 					hashno = TTE4M;
5886 					continue;
5887 				}
5888 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5889 					hashno = TTE32M;
5890 					continue;
5891 				}
5892 				hashno = TTE256M;
5893 				continue;
5894 			} else {
5895 				hashno = TTE4M;
5896 				continue;
5897 			}
5898 		}
5899 		if (hmeblkp->hblk_shw_bit) {
5900 			/*
5901 			 * If we encounter a shadow hmeblk we know there is
5902 			 * smaller sized hmeblks mapping the same address space.
5903 			 * Decrement the hash size and rehash.
5904 			 */
5905 			ASSERT(sfmmup != KHATID);
5906 			hashno--;
5907 			SFMMU_HASH_UNLOCK(hmebp);
5908 			continue;
5909 		}
5910 
5911 		/*
5912 		 * track callback address ranges.
5913 		 * only start a new range when it's not contiguous
5914 		 */
5915 		if (callback != NULL) {
5916 			if (addr_count > 0 &&
5917 			    addr == cb_end_addr[addr_count - 1])
5918 				--addr_count;
5919 			else
5920 				cb_start_addr[addr_count] = addr;
5921 		}
5922 
5923 		addr = sfmmu_hblk_unload(sfmmup, hmeblkp, addr, endaddr,
5924 		    dmrp, flags);
5925 
5926 		if (callback != NULL)
5927 			cb_end_addr[addr_count++] = addr;
5928 
5929 		if (((flags & HAT_UNLOAD_UNMAP) || (iskernel && !issegkmap)) &&
5930 		    !hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5931 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 0);
5932 		}
5933 		SFMMU_HASH_UNLOCK(hmebp);
5934 
5935 		/*
5936 		 * Notify our caller as to exactly which pages
5937 		 * have been unloaded. We do these in clumps,
5938 		 * to minimize the number of xt_sync()s that need to occur.
5939 		 */
5940 		if (callback != NULL && addr_count == MAX_CB_ADDR) {
5941 			DEMAP_RANGE_FLUSH(dmrp);
5942 			if (dmrp != NULL) {
5943 				cpuset = sfmmup->sfmmu_cpusran;
5944 				xt_sync(cpuset);
5945 			}
5946 
5947 			for (a = 0; a < MAX_CB_ADDR; ++a) {
5948 				callback->hcb_start_addr = cb_start_addr[a];
5949 				callback->hcb_end_addr = cb_end_addr[a];
5950 				callback->hcb_function(callback);
5951 			}
5952 			addr_count = 0;
5953 		}
5954 		if (iskernel) {
5955 			hashno = TTE64K;
5956 			continue;
5957 		}
5958 		if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5959 			ASSERT(hashno == TTE64K);
5960 			continue;
5961 		}
5962 		if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5963 			hashno = TTE512K;
5964 			continue;
5965 		}
5966 		if (mmu_page_sizes == max_mmu_page_sizes) {
5967 			if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5968 				hashno = TTE4M;
5969 				continue;
5970 			}
5971 			if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5972 				hashno = TTE32M;
5973 				continue;
5974 			}
5975 			hashno = TTE256M;
5976 		} else {
5977 			hashno = TTE4M;
5978 		}
5979 	}
5980 
5981 	sfmmu_hblks_list_purge(&list, 0);
5982 	DEMAP_RANGE_FLUSH(dmrp);
5983 	if (dmrp != NULL) {
5984 		cpuset = sfmmup->sfmmu_cpusran;
5985 		xt_sync(cpuset);
5986 	}
5987 	if (callback && addr_count != 0) {
5988 		for (a = 0; a < addr_count; ++a) {
5989 			callback->hcb_start_addr = cb_start_addr[a];
5990 			callback->hcb_end_addr = cb_end_addr[a];
5991 			callback->hcb_function(callback);
5992 		}
5993 	}
5994 
5995 	/*
5996 	 * Check TSB and TLB page sizes if the process isn't exiting.
5997 	 */
5998 	if (!sfmmup->sfmmu_free)
5999 		sfmmu_check_page_sizes(sfmmup, 0);
6000 }
6001 
6002 /*
6003  * Unload all the mappings in the range [addr..addr+len). addr and len must
6004  * be MMU_PAGESIZE aligned.
6005  */
6006 void
6007 hat_unload(struct hat *sfmmup, caddr_t addr, size_t len, uint_t flags)
6008 {
6009 	if (sfmmup->sfmmu_xhat_provider) {
6010 		XHAT_UNLOAD(sfmmup, addr, len, flags);
6011 		return;
6012 	}
6013 	hat_unload_callback(sfmmup, addr, len, flags, NULL);
6014 }
6015 
6016 
6017 /*
6018  * Find the largest mapping size for this page.
6019  */
6020 int
6021 fnd_mapping_sz(page_t *pp)
6022 {
6023 	int sz;
6024 	int p_index;
6025 
6026 	p_index = PP_MAPINDEX(pp);
6027 
6028 	sz = 0;
6029 	p_index >>= 1;	/* don't care about 8K bit */
6030 	for (; p_index; p_index >>= 1) {
6031 		sz++;
6032 	}
6033 
6034 	return (sz);
6035 }
6036 
6037 /*
6038  * This function unloads a range of addresses for an hmeblk.
6039  * It returns the next address to be unloaded.
6040  * It should be called with the hash lock held.
6041  */
6042 static caddr_t
6043 sfmmu_hblk_unload(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
6044 	caddr_t endaddr, demap_range_t *dmrp, uint_t flags)
6045 {
6046 	tte_t	tte, ttemod;
6047 	struct	sf_hment *sfhmep;
6048 	int	ttesz;
6049 	long	ttecnt;
6050 	page_t *pp;
6051 	kmutex_t *pml;
6052 	int ret;
6053 	int use_demap_range;
6054 
6055 	ASSERT(in_hblk_range(hmeblkp, addr));
6056 	ASSERT(!hmeblkp->hblk_shw_bit);
6057 	ASSERT(sfmmup != NULL || hmeblkp->hblk_shared);
6058 	ASSERT(sfmmup == NULL || !hmeblkp->hblk_shared);
6059 	ASSERT(dmrp == NULL || !hmeblkp->hblk_shared);
6060 
6061 #ifdef DEBUG
6062 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
6063 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
6064 		panic("sfmmu_hblk_unload: partial unload of large page");
6065 	}
6066 #endif /* DEBUG */
6067 
6068 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
6069 	ttesz = get_hblk_ttesz(hmeblkp);
6070 
6071 	use_demap_range = ((dmrp == NULL) ||
6072 	    (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp)));
6073 
6074 	if (use_demap_range) {
6075 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
6076 	} else {
6077 		DEMAP_RANGE_FLUSH(dmrp);
6078 	}
6079 	ttecnt = 0;
6080 	HBLKTOHME(sfhmep, hmeblkp, addr);
6081 
6082 	while (addr < endaddr) {
6083 		pml = NULL;
6084 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6085 		if (TTE_IS_VALID(&tte)) {
6086 			pp = sfhmep->hme_page;
6087 			if (pp != NULL) {
6088 				pml = sfmmu_mlist_enter(pp);
6089 			}
6090 
6091 			/*
6092 			 * Verify if hme still points to 'pp' now that
6093 			 * we have p_mapping lock.
6094 			 */
6095 			if (sfhmep->hme_page != pp) {
6096 				if (pp != NULL && sfhmep->hme_page != NULL) {
6097 					ASSERT(pml != NULL);
6098 					sfmmu_mlist_exit(pml);
6099 					/* Re-start this iteration. */
6100 					continue;
6101 				}
6102 				ASSERT((pp != NULL) &&
6103 				    (sfhmep->hme_page == NULL));
6104 				goto tte_unloaded;
6105 			}
6106 
6107 			/*
6108 			 * This point on we have both HASH and p_mapping
6109 			 * lock.
6110 			 */
6111 			ASSERT(pp == sfhmep->hme_page);
6112 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
6113 
6114 			/*
6115 			 * We need to loop on modify tte because it is
6116 			 * possible for pagesync to come along and
6117 			 * change the software bits beneath us.
6118 			 *
6119 			 * Page_unload can also invalidate the tte after
6120 			 * we read tte outside of p_mapping lock.
6121 			 */
6122 again:
6123 			ttemod = tte;
6124 
6125 			TTE_SET_INVALID(&ttemod);
6126 			ret = sfmmu_modifytte_try(&tte, &ttemod,
6127 			    &sfhmep->hme_tte);
6128 
6129 			if (ret <= 0) {
6130 				if (TTE_IS_VALID(&tte)) {
6131 					ASSERT(ret < 0);
6132 					goto again;
6133 				}
6134 				if (pp != NULL) {
6135 					panic("sfmmu_hblk_unload: pp = 0x%p "
6136 					    "tte became invalid under mlist"
6137 					    " lock = 0x%p", (void *)pp,
6138 					    (void *)pml);
6139 				}
6140 				continue;
6141 			}
6142 
6143 			if (!(flags & HAT_UNLOAD_NOSYNC)) {
6144 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
6145 			}
6146 
6147 			/*
6148 			 * Ok- we invalidated the tte. Do the rest of the job.
6149 			 */
6150 			ttecnt++;
6151 
6152 			if (flags & HAT_UNLOAD_UNLOCK) {
6153 				ASSERT(hmeblkp->hblk_lckcnt > 0);
6154 				atomic_add_32(&hmeblkp->hblk_lckcnt, -1);
6155 				HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
6156 			}
6157 
6158 			/*
6159 			 * Normally we would need to flush the page
6160 			 * from the virtual cache at this point in
6161 			 * order to prevent a potential cache alias
6162 			 * inconsistency.
6163 			 * The particular scenario we need to worry
6164 			 * about is:
6165 			 * Given:  va1 and va2 are two virtual address
6166 			 * that alias and map the same physical
6167 			 * address.
6168 			 * 1.   mapping exists from va1 to pa and data
6169 			 * has been read into the cache.
6170 			 * 2.   unload va1.
6171 			 * 3.   load va2 and modify data using va2.
6172 			 * 4    unload va2.
6173 			 * 5.   load va1 and reference data.  Unless we
6174 			 * flush the data cache when we unload we will
6175 			 * get stale data.
6176 			 * Fortunately, page coloring eliminates the
6177 			 * above scenario by remembering the color a
6178 			 * physical page was last or is currently
6179 			 * mapped to.  Now, we delay the flush until
6180 			 * the loading of translations.  Only when the
6181 			 * new translation is of a different color
6182 			 * are we forced to flush.
6183 			 */
6184 			if (use_demap_range) {
6185 				/*
6186 				 * Mark this page as needing a demap.
6187 				 */
6188 				DEMAP_RANGE_MARKPG(dmrp, addr);
6189 			} else {
6190 				ASSERT(sfmmup != NULL);
6191 				ASSERT(!hmeblkp->hblk_shared);
6192 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
6193 				    sfmmup->sfmmu_free, 0);
6194 			}
6195 
6196 			if (pp) {
6197 				/*
6198 				 * Remove the hment from the mapping list
6199 				 */
6200 				ASSERT(hmeblkp->hblk_hmecnt > 0);
6201 
6202 				/*
6203 				 * Again, we cannot
6204 				 * ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS);
6205 				 */
6206 				HME_SUB(sfhmep, pp);
6207 				membar_stst();
6208 				atomic_add_16(&hmeblkp->hblk_hmecnt, -1);
6209 			}
6210 
6211 			ASSERT(hmeblkp->hblk_vcnt > 0);
6212 			atomic_add_16(&hmeblkp->hblk_vcnt, -1);
6213 
6214 			ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
6215 			    !hmeblkp->hblk_lckcnt);
6216 
6217 #ifdef VAC
6218 			if (pp && (pp->p_nrm & (P_KPMC | P_KPMS | P_TNC))) {
6219 				if (PP_ISTNC(pp)) {
6220 					/*
6221 					 * If page was temporary
6222 					 * uncached, try to recache
6223 					 * it. Note that HME_SUB() was
6224 					 * called above so p_index and
6225 					 * mlist had been updated.
6226 					 */
6227 					conv_tnc(pp, ttesz);
6228 				} else if (pp->p_mapping == NULL) {
6229 					ASSERT(kpm_enable);
6230 					/*
6231 					 * Page is marked to be in VAC conflict
6232 					 * to an existing kpm mapping and/or is
6233 					 * kpm mapped using only the regular
6234 					 * pagesize.
6235 					 */
6236 					sfmmu_kpm_hme_unload(pp);
6237 				}
6238 			}
6239 #endif	/* VAC */
6240 		} else if ((pp = sfhmep->hme_page) != NULL) {
6241 				/*
6242 				 * TTE is invalid but the hme
6243 				 * still exists. let pageunload
6244 				 * complete its job.
6245 				 */
6246 				ASSERT(pml == NULL);
6247 				pml = sfmmu_mlist_enter(pp);
6248 				if (sfhmep->hme_page != NULL) {
6249 					sfmmu_mlist_exit(pml);
6250 					continue;
6251 				}
6252 				ASSERT(sfhmep->hme_page == NULL);
6253 		} else if (hmeblkp->hblk_hmecnt != 0) {
6254 			/*
6255 			 * pageunload may have not finished decrementing
6256 			 * hblk_vcnt and hblk_hmecnt. Find page_t if any and
6257 			 * wait for pageunload to finish. Rely on pageunload
6258 			 * to decrement hblk_hmecnt after hblk_vcnt.
6259 			 */
6260 			pfn_t pfn = TTE_TO_TTEPFN(&tte);
6261 			ASSERT(pml == NULL);
6262 			if (pf_is_memory(pfn)) {
6263 				pp = page_numtopp_nolock(pfn);
6264 				if (pp != NULL) {
6265 					pml = sfmmu_mlist_enter(pp);
6266 					sfmmu_mlist_exit(pml);
6267 					pml = NULL;
6268 				}
6269 			}
6270 		}
6271 
6272 tte_unloaded:
6273 		/*
6274 		 * At this point, the tte we are looking at
6275 		 * should be unloaded, and hme has been unlinked
6276 		 * from page too. This is important because in
6277 		 * pageunload, it does ttesync() then HME_SUB.
6278 		 * We need to make sure HME_SUB has been completed
6279 		 * so we know ttesync() has been completed. Otherwise,
6280 		 * at exit time, after return from hat layer, VM will
6281 		 * release as structure which hat_setstat() (called
6282 		 * by ttesync()) needs.
6283 		 */
6284 #ifdef DEBUG
6285 		{
6286 			tte_t	dtte;
6287 
6288 			ASSERT(sfhmep->hme_page == NULL);
6289 
6290 			sfmmu_copytte(&sfhmep->hme_tte, &dtte);
6291 			ASSERT(!TTE_IS_VALID(&dtte));
6292 		}
6293 #endif
6294 
6295 		if (pml) {
6296 			sfmmu_mlist_exit(pml);
6297 		}
6298 
6299 		addr += TTEBYTES(ttesz);
6300 		sfhmep++;
6301 		DEMAP_RANGE_NEXTPG(dmrp);
6302 	}
6303 	/*
6304 	 * For shared hmeblks this routine is only called when region is freed
6305 	 * and no longer referenced.  So no need to decrement ttecnt
6306 	 * in the region structure here.
6307 	 */
6308 	if (ttecnt > 0 && sfmmup != NULL) {
6309 		atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -ttecnt);
6310 	}
6311 	return (addr);
6312 }
6313 
6314 /*
6315  * Invalidate a virtual address range for the local CPU.
6316  * For best performance ensure that the va range is completely
6317  * mapped, otherwise the entire TLB will be flushed.
6318  */
6319 void
6320 hat_flush_range(struct hat *sfmmup, caddr_t va, size_t size)
6321 {
6322 	ssize_t sz;
6323 	caddr_t endva = va + size;
6324 
6325 	while (va < endva) {
6326 		sz = hat_getpagesize(sfmmup, va);
6327 		if (sz < 0) {
6328 			vtag_flushall();
6329 			break;
6330 		}
6331 		vtag_flushpage(va, (uint64_t)sfmmup);
6332 		va += sz;
6333 	}
6334 }
6335 
6336 /*
6337  * Synchronize all the mappings in the range [addr..addr+len).
6338  * Can be called with clearflag having two states:
6339  * HAT_SYNC_DONTZERO means just return the rm stats
6340  * HAT_SYNC_ZERORM means zero rm bits in the tte and return the stats
6341  */
6342 void
6343 hat_sync(struct hat *sfmmup, caddr_t addr, size_t len, uint_t clearflag)
6344 {
6345 	struct hmehash_bucket *hmebp;
6346 	hmeblk_tag hblktag;
6347 	int hmeshift, hashno = 1;
6348 	struct hme_blk *hmeblkp, *list = NULL;
6349 	caddr_t endaddr;
6350 	cpuset_t cpuset;
6351 
6352 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
6353 	ASSERT((sfmmup == ksfmmup) ||
6354 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
6355 	ASSERT((len & MMU_PAGEOFFSET) == 0);
6356 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
6357 	    (clearflag == HAT_SYNC_ZERORM));
6358 
6359 	CPUSET_ZERO(cpuset);
6360 
6361 	endaddr = addr + len;
6362 	hblktag.htag_id = sfmmup;
6363 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
6364 
6365 	/*
6366 	 * Spitfire supports 4 page sizes.
6367 	 * Most pages are expected to be of the smallest page
6368 	 * size (8K) and these will not need to be rehashed. 64K
6369 	 * pages also don't need to be rehashed because the an hmeblk
6370 	 * spans 64K of address space. 512K pages might need 1 rehash and
6371 	 * and 4M pages 2 rehashes.
6372 	 */
6373 	while (addr < endaddr) {
6374 		hmeshift = HME_HASH_SHIFT(hashno);
6375 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
6376 		hblktag.htag_rehash = hashno;
6377 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
6378 
6379 		SFMMU_HASH_LOCK(hmebp);
6380 
6381 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
6382 		if (hmeblkp != NULL) {
6383 			ASSERT(!hmeblkp->hblk_shared);
6384 			/*
6385 			 * We've encountered a shadow hmeblk so skip the range
6386 			 * of the next smaller mapping size.
6387 			 */
6388 			if (hmeblkp->hblk_shw_bit) {
6389 				ASSERT(sfmmup != ksfmmup);
6390 				ASSERT(hashno > 1);
6391 				addr = (caddr_t)P2END((uintptr_t)addr,
6392 				    TTEBYTES(hashno - 1));
6393 			} else {
6394 				addr = sfmmu_hblk_sync(sfmmup, hmeblkp,
6395 				    addr, endaddr, clearflag);
6396 			}
6397 			SFMMU_HASH_UNLOCK(hmebp);
6398 			hashno = 1;
6399 			continue;
6400 		}
6401 		SFMMU_HASH_UNLOCK(hmebp);
6402 
6403 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
6404 			/*
6405 			 * We have traversed the whole list and rehashed
6406 			 * if necessary without finding the address to sync.
6407 			 * This is ok so we increment the address by the
6408 			 * smallest hmeblk range for kernel mappings and the
6409 			 * largest hmeblk range, to account for shadow hmeblks,
6410 			 * for user mappings and continue.
6411 			 */
6412 			if (sfmmup == ksfmmup)
6413 				addr = (caddr_t)P2END((uintptr_t)addr,
6414 				    TTEBYTES(1));
6415 			else
6416 				addr = (caddr_t)P2END((uintptr_t)addr,
6417 				    TTEBYTES(hashno));
6418 			hashno = 1;
6419 		} else {
6420 			hashno++;
6421 		}
6422 	}
6423 	sfmmu_hblks_list_purge(&list, 0);
6424 	cpuset = sfmmup->sfmmu_cpusran;
6425 	xt_sync(cpuset);
6426 }
6427 
6428 static caddr_t
6429 sfmmu_hblk_sync(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
6430 	caddr_t endaddr, int clearflag)
6431 {
6432 	tte_t	tte, ttemod;
6433 	struct sf_hment *sfhmep;
6434 	int ttesz;
6435 	struct page *pp;
6436 	kmutex_t *pml;
6437 	int ret;
6438 
6439 	ASSERT(hmeblkp->hblk_shw_bit == 0);
6440 	ASSERT(!hmeblkp->hblk_shared);
6441 
6442 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
6443 
6444 	ttesz = get_hblk_ttesz(hmeblkp);
6445 	HBLKTOHME(sfhmep, hmeblkp, addr);
6446 
6447 	while (addr < endaddr) {
6448 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6449 		if (TTE_IS_VALID(&tte)) {
6450 			pml = NULL;
6451 			pp = sfhmep->hme_page;
6452 			if (pp) {
6453 				pml = sfmmu_mlist_enter(pp);
6454 			}
6455 			if (pp != sfhmep->hme_page) {
6456 				/*
6457 				 * tte most have been unloaded
6458 				 * underneath us.  Recheck
6459 				 */
6460 				ASSERT(pml);
6461 				sfmmu_mlist_exit(pml);
6462 				continue;
6463 			}
6464 
6465 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
6466 
6467 			if (clearflag == HAT_SYNC_ZERORM) {
6468 				ttemod = tte;
6469 				TTE_CLR_RM(&ttemod);
6470 				ret = sfmmu_modifytte_try(&tte, &ttemod,
6471 				    &sfhmep->hme_tte);
6472 				if (ret < 0) {
6473 					if (pml) {
6474 						sfmmu_mlist_exit(pml);
6475 					}
6476 					continue;
6477 				}
6478 
6479 				if (ret > 0) {
6480 					sfmmu_tlb_demap(addr, sfmmup,
6481 					    hmeblkp, 0, 0);
6482 				}
6483 			}
6484 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
6485 			if (pml) {
6486 				sfmmu_mlist_exit(pml);
6487 			}
6488 		}
6489 		addr += TTEBYTES(ttesz);
6490 		sfhmep++;
6491 	}
6492 	return (addr);
6493 }
6494 
6495 /*
6496  * This function will sync a tte to the page struct and it will
6497  * update the hat stats. Currently it allows us to pass a NULL pp
6498  * and we will simply update the stats.  We may want to change this
6499  * so we only keep stats for pages backed by pp's.
6500  */
6501 static void
6502 sfmmu_ttesync(struct hat *sfmmup, caddr_t addr, tte_t *ttep, page_t *pp)
6503 {
6504 	uint_t rm = 0;
6505 	int   	sz;
6506 	pgcnt_t	npgs;
6507 
6508 	ASSERT(TTE_IS_VALID(ttep));
6509 
6510 	if (TTE_IS_NOSYNC(ttep)) {
6511 		return;
6512 	}
6513 
6514 	if (TTE_IS_REF(ttep))  {
6515 		rm = P_REF;
6516 	}
6517 	if (TTE_IS_MOD(ttep))  {
6518 		rm |= P_MOD;
6519 	}
6520 
6521 	if (rm == 0) {
6522 		return;
6523 	}
6524 
6525 	sz = TTE_CSZ(ttep);
6526 	if (sfmmup != NULL && sfmmup->sfmmu_rmstat) {
6527 		int i;
6528 		caddr_t	vaddr = addr;
6529 
6530 		for (i = 0; i < TTEPAGES(sz); i++, vaddr += MMU_PAGESIZE) {
6531 			hat_setstat(sfmmup->sfmmu_as, vaddr, MMU_PAGESIZE, rm);
6532 		}
6533 
6534 	}
6535 
6536 	/*
6537 	 * XXX I want to use cas to update nrm bits but they
6538 	 * currently belong in common/vm and not in hat where
6539 	 * they should be.
6540 	 * The nrm bits are protected by the same mutex as
6541 	 * the one that protects the page's mapping list.
6542 	 */
6543 	if (!pp)
6544 		return;
6545 	ASSERT(sfmmu_mlist_held(pp));
6546 	/*
6547 	 * If the tte is for a large page, we need to sync all the
6548 	 * pages covered by the tte.
6549 	 */
6550 	if (sz != TTE8K) {
6551 		ASSERT(pp->p_szc != 0);
6552 		pp = PP_GROUPLEADER(pp, sz);
6553 		ASSERT(sfmmu_mlist_held(pp));
6554 	}
6555 
6556 	/* Get number of pages from tte size. */
6557 	npgs = TTEPAGES(sz);
6558 
6559 	do {
6560 		ASSERT(pp);
6561 		ASSERT(sfmmu_mlist_held(pp));
6562 		if (((rm & P_REF) != 0 && !PP_ISREF(pp)) ||
6563 		    ((rm & P_MOD) != 0 && !PP_ISMOD(pp)))
6564 			hat_page_setattr(pp, rm);
6565 
6566 		/*
6567 		 * Are we done? If not, we must have a large mapping.
6568 		 * For large mappings we need to sync the rest of the pages
6569 		 * covered by this tte; goto the next page.
6570 		 */
6571 	} while (--npgs > 0 && (pp = PP_PAGENEXT(pp)));
6572 }
6573 
6574 /*
6575  * Execute pre-callback handler of each pa_hment linked to pp
6576  *
6577  * Inputs:
6578  *   flag: either HAT_PRESUSPEND or HAT_SUSPEND.
6579  *   capture_cpus: pointer to return value (below)
6580  *
6581  * Returns:
6582  *   Propagates the subsystem callback return values back to the caller;
6583  *   returns 0 on success.  If capture_cpus is non-NULL, the value returned
6584  *   is zero if all of the pa_hments are of a type that do not require
6585  *   capturing CPUs prior to suspending the mapping, else it is 1.
6586  */
6587 static int
6588 hat_pageprocess_precallbacks(struct page *pp, uint_t flag, int *capture_cpus)
6589 {
6590 	struct sf_hment	*sfhmep;
6591 	struct pa_hment *pahmep;
6592 	int (*f)(caddr_t, uint_t, uint_t, void *);
6593 	int		ret;
6594 	id_t		id;
6595 	int		locked = 0;
6596 	kmutex_t	*pml;
6597 
6598 	ASSERT(PAGE_EXCL(pp));
6599 	if (!sfmmu_mlist_held(pp)) {
6600 		pml = sfmmu_mlist_enter(pp);
6601 		locked = 1;
6602 	}
6603 
6604 	if (capture_cpus)
6605 		*capture_cpus = 0;
6606 
6607 top:
6608 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6609 		/*
6610 		 * skip sf_hments corresponding to VA<->PA mappings;
6611 		 * for pa_hment's, hme_tte.ll is zero
6612 		 */
6613 		if (!IS_PAHME(sfhmep))
6614 			continue;
6615 
6616 		pahmep = sfhmep->hme_data;
6617 		ASSERT(pahmep != NULL);
6618 
6619 		/*
6620 		 * skip if pre-handler has been called earlier in this loop
6621 		 */
6622 		if (pahmep->flags & flag)
6623 			continue;
6624 
6625 		id = pahmep->cb_id;
6626 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6627 		if (capture_cpus && sfmmu_cb_table[id].capture_cpus != 0)
6628 			*capture_cpus = 1;
6629 		if ((f = sfmmu_cb_table[id].prehandler) == NULL) {
6630 			pahmep->flags |= flag;
6631 			continue;
6632 		}
6633 
6634 		/*
6635 		 * Drop the mapping list lock to avoid locking order issues.
6636 		 */
6637 		if (locked)
6638 			sfmmu_mlist_exit(pml);
6639 
6640 		ret = f(pahmep->addr, pahmep->len, flag, pahmep->pvt);
6641 		if (ret != 0)
6642 			return (ret);	/* caller must do the cleanup */
6643 
6644 		if (locked) {
6645 			pml = sfmmu_mlist_enter(pp);
6646 			pahmep->flags |= flag;
6647 			goto top;
6648 		}
6649 
6650 		pahmep->flags |= flag;
6651 	}
6652 
6653 	if (locked)
6654 		sfmmu_mlist_exit(pml);
6655 
6656 	return (0);
6657 }
6658 
6659 /*
6660  * Execute post-callback handler of each pa_hment linked to pp
6661  *
6662  * Same overall assumptions and restrictions apply as for
6663  * hat_pageprocess_precallbacks().
6664  */
6665 static void
6666 hat_pageprocess_postcallbacks(struct page *pp, uint_t flag)
6667 {
6668 	pfn_t pgpfn = pp->p_pagenum;
6669 	pfn_t pgmask = btop(page_get_pagesize(pp->p_szc)) - 1;
6670 	pfn_t newpfn;
6671 	struct sf_hment *sfhmep;
6672 	struct pa_hment *pahmep;
6673 	int (*f)(caddr_t, uint_t, uint_t, void *, pfn_t);
6674 	id_t	id;
6675 	int	locked = 0;
6676 	kmutex_t *pml;
6677 
6678 	ASSERT(PAGE_EXCL(pp));
6679 	if (!sfmmu_mlist_held(pp)) {
6680 		pml = sfmmu_mlist_enter(pp);
6681 		locked = 1;
6682 	}
6683 
6684 top:
6685 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6686 		/*
6687 		 * skip sf_hments corresponding to VA<->PA mappings;
6688 		 * for pa_hment's, hme_tte.ll is zero
6689 		 */
6690 		if (!IS_PAHME(sfhmep))
6691 			continue;
6692 
6693 		pahmep = sfhmep->hme_data;
6694 		ASSERT(pahmep != NULL);
6695 
6696 		if ((pahmep->flags & flag) == 0)
6697 			continue;
6698 
6699 		pahmep->flags &= ~flag;
6700 
6701 		id = pahmep->cb_id;
6702 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6703 		if ((f = sfmmu_cb_table[id].posthandler) == NULL)
6704 			continue;
6705 
6706 		/*
6707 		 * Convert the base page PFN into the constituent PFN
6708 		 * which is needed by the callback handler.
6709 		 */
6710 		newpfn = pgpfn | (btop((uintptr_t)pahmep->addr) & pgmask);
6711 
6712 		/*
6713 		 * Drop the mapping list lock to avoid locking order issues.
6714 		 */
6715 		if (locked)
6716 			sfmmu_mlist_exit(pml);
6717 
6718 		if (f(pahmep->addr, pahmep->len, flag, pahmep->pvt, newpfn)
6719 		    != 0)
6720 			panic("sfmmu: posthandler failed");
6721 
6722 		if (locked) {
6723 			pml = sfmmu_mlist_enter(pp);
6724 			goto top;
6725 		}
6726 	}
6727 
6728 	if (locked)
6729 		sfmmu_mlist_exit(pml);
6730 }
6731 
6732 /*
6733  * Suspend locked kernel mapping
6734  */
6735 void
6736 hat_pagesuspend(struct page *pp)
6737 {
6738 	struct sf_hment *sfhmep;
6739 	sfmmu_t *sfmmup;
6740 	tte_t tte, ttemod;
6741 	struct hme_blk *hmeblkp;
6742 	caddr_t addr;
6743 	int index, cons;
6744 	cpuset_t cpuset;
6745 
6746 	ASSERT(PAGE_EXCL(pp));
6747 	ASSERT(sfmmu_mlist_held(pp));
6748 
6749 	mutex_enter(&kpr_suspendlock);
6750 
6751 	/*
6752 	 * We're about to suspend a kernel mapping so mark this thread as
6753 	 * non-traceable by DTrace. This prevents us from running into issues
6754 	 * with probe context trying to touch a suspended page
6755 	 * in the relocation codepath itself.
6756 	 */
6757 	curthread->t_flag |= T_DONTDTRACE;
6758 
6759 	index = PP_MAPINDEX(pp);
6760 	cons = TTE8K;
6761 
6762 retry:
6763 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6764 
6765 		if (IS_PAHME(sfhmep))
6766 			continue;
6767 
6768 		if (get_hblk_ttesz(sfmmu_hmetohblk(sfhmep)) != cons)
6769 			continue;
6770 
6771 		/*
6772 		 * Loop until we successfully set the suspend bit in
6773 		 * the TTE.
6774 		 */
6775 again:
6776 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6777 		ASSERT(TTE_IS_VALID(&tte));
6778 
6779 		ttemod = tte;
6780 		TTE_SET_SUSPEND(&ttemod);
6781 		if (sfmmu_modifytte_try(&tte, &ttemod,
6782 		    &sfhmep->hme_tte) < 0)
6783 			goto again;
6784 
6785 		/*
6786 		 * Invalidate TSB entry
6787 		 */
6788 		hmeblkp = sfmmu_hmetohblk(sfhmep);
6789 
6790 		sfmmup = hblktosfmmu(hmeblkp);
6791 		ASSERT(sfmmup == ksfmmup);
6792 		ASSERT(!hmeblkp->hblk_shared);
6793 
6794 		addr = tte_to_vaddr(hmeblkp, tte);
6795 
6796 		/*
6797 		 * No need to make sure that the TSB for this sfmmu is
6798 		 * not being relocated since it is ksfmmup and thus it
6799 		 * will never be relocated.
6800 		 */
6801 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
6802 
6803 		/*
6804 		 * Update xcall stats
6805 		 */
6806 		cpuset = cpu_ready_set;
6807 		CPUSET_DEL(cpuset, CPU->cpu_id);
6808 
6809 		/* LINTED: constant in conditional context */
6810 		SFMMU_XCALL_STATS(ksfmmup);
6811 
6812 		/*
6813 		 * Flush TLB entry on remote CPU's
6814 		 */
6815 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
6816 		    (uint64_t)ksfmmup);
6817 		xt_sync(cpuset);
6818 
6819 		/*
6820 		 * Flush TLB entry on local CPU
6821 		 */
6822 		vtag_flushpage(addr, (uint64_t)ksfmmup);
6823 	}
6824 
6825 	while (index != 0) {
6826 		index = index >> 1;
6827 		if (index != 0)
6828 			cons++;
6829 		if (index & 0x1) {
6830 			pp = PP_GROUPLEADER(pp, cons);
6831 			goto retry;
6832 		}
6833 	}
6834 }
6835 
6836 #ifdef	DEBUG
6837 
6838 #define	N_PRLE	1024
6839 struct prle {
6840 	page_t *targ;
6841 	page_t *repl;
6842 	int status;
6843 	int pausecpus;
6844 	hrtime_t whence;
6845 };
6846 
6847 static struct prle page_relocate_log[N_PRLE];
6848 static int prl_entry;
6849 static kmutex_t prl_mutex;
6850 
6851 #define	PAGE_RELOCATE_LOG(t, r, s, p)					\
6852 	mutex_enter(&prl_mutex);					\
6853 	page_relocate_log[prl_entry].targ = *(t);			\
6854 	page_relocate_log[prl_entry].repl = *(r);			\
6855 	page_relocate_log[prl_entry].status = (s);			\
6856 	page_relocate_log[prl_entry].pausecpus = (p);			\
6857 	page_relocate_log[prl_entry].whence = gethrtime();		\
6858 	prl_entry = (prl_entry == (N_PRLE - 1))? 0 : prl_entry + 1;	\
6859 	mutex_exit(&prl_mutex);
6860 
6861 #else	/* !DEBUG */
6862 #define	PAGE_RELOCATE_LOG(t, r, s, p)
6863 #endif
6864 
6865 /*
6866  * Core Kernel Page Relocation Algorithm
6867  *
6868  * Input:
6869  *
6870  * target : 	constituent pages are SE_EXCL locked.
6871  * replacement:	constituent pages are SE_EXCL locked.
6872  *
6873  * Output:
6874  *
6875  * nrelocp:	number of pages relocated
6876  */
6877 int
6878 hat_page_relocate(page_t **target, page_t **replacement, spgcnt_t *nrelocp)
6879 {
6880 	page_t		*targ, *repl;
6881 	page_t		*tpp, *rpp;
6882 	kmutex_t	*low, *high;
6883 	spgcnt_t	npages, i;
6884 	page_t		*pl = NULL;
6885 	int		old_pil;
6886 	cpuset_t	cpuset;
6887 	int		cap_cpus;
6888 	int		ret;
6889 #ifdef VAC
6890 	int		cflags = 0;
6891 #endif
6892 
6893 	if (hat_kpr_enabled == 0 || !kcage_on || PP_ISNORELOC(*target)) {
6894 		PAGE_RELOCATE_LOG(target, replacement, EAGAIN, -1);
6895 		return (EAGAIN);
6896 	}
6897 
6898 	mutex_enter(&kpr_mutex);
6899 	kreloc_thread = curthread;
6900 
6901 	targ = *target;
6902 	repl = *replacement;
6903 	ASSERT(repl != NULL);
6904 	ASSERT(targ->p_szc == repl->p_szc);
6905 
6906 	npages = page_get_pagecnt(targ->p_szc);
6907 
6908 	/*
6909 	 * unload VA<->PA mappings that are not locked
6910 	 */
6911 	tpp = targ;
6912 	for (i = 0; i < npages; i++) {
6913 		(void) hat_pageunload(tpp, SFMMU_KERNEL_RELOC);
6914 		tpp++;
6915 	}
6916 
6917 	/*
6918 	 * Do "presuspend" callbacks, in a context from which we can still
6919 	 * block as needed. Note that we don't hold the mapping list lock
6920 	 * of "targ" at this point due to potential locking order issues;
6921 	 * we assume that between the hat_pageunload() above and holding
6922 	 * the SE_EXCL lock that the mapping list *cannot* change at this
6923 	 * point.
6924 	 */
6925 	ret = hat_pageprocess_precallbacks(targ, HAT_PRESUSPEND, &cap_cpus);
6926 	if (ret != 0) {
6927 		/*
6928 		 * EIO translates to fatal error, for all others cleanup
6929 		 * and return EAGAIN.
6930 		 */
6931 		ASSERT(ret != EIO);
6932 		hat_pageprocess_postcallbacks(targ, HAT_POSTUNSUSPEND);
6933 		PAGE_RELOCATE_LOG(target, replacement, ret, -1);
6934 		kreloc_thread = NULL;
6935 		mutex_exit(&kpr_mutex);
6936 		return (EAGAIN);
6937 	}
6938 
6939 	/*
6940 	 * acquire p_mapping list lock for both the target and replacement
6941 	 * root pages.
6942 	 *
6943 	 * low and high refer to the need to grab the mlist locks in a
6944 	 * specific order in order to prevent race conditions.  Thus the
6945 	 * lower lock must be grabbed before the higher lock.
6946 	 *
6947 	 * This will block hat_unload's accessing p_mapping list.  Since
6948 	 * we have SE_EXCL lock, hat_memload and hat_pageunload will be
6949 	 * blocked.  Thus, no one else will be accessing the p_mapping list
6950 	 * while we suspend and reload the locked mapping below.
6951 	 */
6952 	tpp = targ;
6953 	rpp = repl;
6954 	sfmmu_mlist_reloc_enter(tpp, rpp, &low, &high);
6955 
6956 	kpreempt_disable();
6957 
6958 	/*
6959 	 * We raise our PIL to 13 so that we don't get captured by
6960 	 * another CPU or pinned by an interrupt thread.  We can't go to
6961 	 * PIL 14 since the nexus driver(s) may need to interrupt at
6962 	 * that level in the case of IOMMU pseudo mappings.
6963 	 */
6964 	cpuset = cpu_ready_set;
6965 	CPUSET_DEL(cpuset, CPU->cpu_id);
6966 	if (!cap_cpus || CPUSET_ISNULL(cpuset)) {
6967 		old_pil = splr(XCALL_PIL);
6968 	} else {
6969 		old_pil = -1;
6970 		xc_attention(cpuset);
6971 	}
6972 	ASSERT(getpil() == XCALL_PIL);
6973 
6974 	/*
6975 	 * Now do suspend callbacks. In the case of an IOMMU mapping
6976 	 * this will suspend all DMA activity to the page while it is
6977 	 * being relocated. Since we are well above LOCK_LEVEL and CPUs
6978 	 * may be captured at this point we should have acquired any needed
6979 	 * locks in the presuspend callback.
6980 	 */
6981 	ret = hat_pageprocess_precallbacks(targ, HAT_SUSPEND, NULL);
6982 	if (ret != 0) {
6983 		repl = targ;
6984 		goto suspend_fail;
6985 	}
6986 
6987 	/*
6988 	 * Raise the PIL yet again, this time to block all high-level
6989 	 * interrupts on this CPU. This is necessary to prevent an
6990 	 * interrupt routine from pinning the thread which holds the
6991 	 * mapping suspended and then touching the suspended page.
6992 	 *
6993 	 * Once the page is suspended we also need to be careful to
6994 	 * avoid calling any functions which touch any seg_kmem memory
6995 	 * since that memory may be backed by the very page we are
6996 	 * relocating in here!
6997 	 */
6998 	hat_pagesuspend(targ);
6999 
7000 	/*
7001 	 * Now that we are confident everybody has stopped using this page,
7002 	 * copy the page contents.  Note we use a physical copy to prevent
7003 	 * locking issues and to avoid fpRAS because we can't handle it in
7004 	 * this context.
7005 	 */
7006 	for (i = 0; i < npages; i++, tpp++, rpp++) {
7007 #ifdef VAC
7008 		/*
7009 		 * If the replacement has a different vcolor than
7010 		 * the one being replacd, we need to handle VAC
7011 		 * consistency for it just as we were setting up
7012 		 * a new mapping to it.
7013 		 */
7014 		if ((PP_GET_VCOLOR(rpp) != NO_VCOLOR) &&
7015 		    (tpp->p_vcolor != rpp->p_vcolor) &&
7016 		    !CacheColor_IsFlushed(cflags, PP_GET_VCOLOR(rpp))) {
7017 			CacheColor_SetFlushed(cflags, PP_GET_VCOLOR(rpp));
7018 			sfmmu_cache_flushcolor(PP_GET_VCOLOR(rpp),
7019 			    rpp->p_pagenum);
7020 		}
7021 #endif
7022 		/*
7023 		 * Copy the contents of the page.
7024 		 */
7025 		ppcopy_kernel(tpp, rpp);
7026 	}
7027 
7028 	tpp = targ;
7029 	rpp = repl;
7030 	for (i = 0; i < npages; i++, tpp++, rpp++) {
7031 		/*
7032 		 * Copy attributes.  VAC consistency was handled above,
7033 		 * if required.
7034 		 */
7035 		rpp->p_nrm = tpp->p_nrm;
7036 		tpp->p_nrm = 0;
7037 		rpp->p_index = tpp->p_index;
7038 		tpp->p_index = 0;
7039 #ifdef VAC
7040 		rpp->p_vcolor = tpp->p_vcolor;
7041 #endif
7042 	}
7043 
7044 	/*
7045 	 * First, unsuspend the page, if we set the suspend bit, and transfer
7046 	 * the mapping list from the target page to the replacement page.
7047 	 * Next process postcallbacks; since pa_hment's are linked only to the
7048 	 * p_mapping list of root page, we don't iterate over the constituent
7049 	 * pages.
7050 	 */
7051 	hat_pagereload(targ, repl);
7052 
7053 suspend_fail:
7054 	hat_pageprocess_postcallbacks(repl, HAT_UNSUSPEND);
7055 
7056 	/*
7057 	 * Now lower our PIL and release any captured CPUs since we
7058 	 * are out of the "danger zone".  After this it will again be
7059 	 * safe to acquire adaptive mutex locks, or to drop them...
7060 	 */
7061 	if (old_pil != -1) {
7062 		splx(old_pil);
7063 	} else {
7064 		xc_dismissed(cpuset);
7065 	}
7066 
7067 	kpreempt_enable();
7068 
7069 	sfmmu_mlist_reloc_exit(low, high);
7070 
7071 	/*
7072 	 * Postsuspend callbacks should drop any locks held across
7073 	 * the suspend callbacks.  As before, we don't hold the mapping
7074 	 * list lock at this point.. our assumption is that the mapping
7075 	 * list still can't change due to our holding SE_EXCL lock and
7076 	 * there being no unlocked mappings left. Hence the restriction
7077 	 * on calling context to hat_delete_callback()
7078 	 */
7079 	hat_pageprocess_postcallbacks(repl, HAT_POSTUNSUSPEND);
7080 	if (ret != 0) {
7081 		/*
7082 		 * The second presuspend call failed: we got here through
7083 		 * the suspend_fail label above.
7084 		 */
7085 		ASSERT(ret != EIO);
7086 		PAGE_RELOCATE_LOG(target, replacement, ret, cap_cpus);
7087 		kreloc_thread = NULL;
7088 		mutex_exit(&kpr_mutex);
7089 		return (EAGAIN);
7090 	}
7091 
7092 	/*
7093 	 * Now that we're out of the performance critical section we can
7094 	 * take care of updating the hash table, since we still
7095 	 * hold all the pages locked SE_EXCL at this point we
7096 	 * needn't worry about things changing out from under us.
7097 	 */
7098 	tpp = targ;
7099 	rpp = repl;
7100 	for (i = 0; i < npages; i++, tpp++, rpp++) {
7101 
7102 		/*
7103 		 * replace targ with replacement in page_hash table
7104 		 */
7105 		targ = tpp;
7106 		page_relocate_hash(rpp, targ);
7107 
7108 		/*
7109 		 * concatenate target; caller of platform_page_relocate()
7110 		 * expects target to be concatenated after returning.
7111 		 */
7112 		ASSERT(targ->p_next == targ);
7113 		ASSERT(targ->p_prev == targ);
7114 		page_list_concat(&pl, &targ);
7115 	}
7116 
7117 	ASSERT(*target == pl);
7118 	*nrelocp = npages;
7119 	PAGE_RELOCATE_LOG(target, replacement, 0, cap_cpus);
7120 	kreloc_thread = NULL;
7121 	mutex_exit(&kpr_mutex);
7122 	return (0);
7123 }
7124 
7125 /*
7126  * Called when stray pa_hments are found attached to a page which is
7127  * being freed.  Notify the subsystem which attached the pa_hment of
7128  * the error if it registered a suitable handler, else panic.
7129  */
7130 static void
7131 sfmmu_pahment_leaked(struct pa_hment *pahmep)
7132 {
7133 	id_t cb_id = pahmep->cb_id;
7134 
7135 	ASSERT(cb_id >= (id_t)0 && cb_id < sfmmu_cb_nextid);
7136 	if (sfmmu_cb_table[cb_id].errhandler != NULL) {
7137 		if (sfmmu_cb_table[cb_id].errhandler(pahmep->addr, pahmep->len,
7138 		    HAT_CB_ERR_LEAKED, pahmep->pvt) == 0)
7139 			return;		/* non-fatal */
7140 	}
7141 	panic("pa_hment leaked: 0x%p", (void *)pahmep);
7142 }
7143 
7144 /*
7145  * Remove all mappings to page 'pp'.
7146  */
7147 int
7148 hat_pageunload(struct page *pp, uint_t forceflag)
7149 {
7150 	struct page *origpp = pp;
7151 	struct sf_hment *sfhme, *tmphme;
7152 	struct hme_blk *hmeblkp;
7153 	kmutex_t *pml;
7154 #ifdef VAC
7155 	kmutex_t *pmtx;
7156 #endif
7157 	cpuset_t cpuset, tset;
7158 	int index, cons;
7159 	int xhme_blks;
7160 	int pa_hments;
7161 
7162 	ASSERT(PAGE_EXCL(pp));
7163 
7164 retry_xhat:
7165 	tmphme = NULL;
7166 	xhme_blks = 0;
7167 	pa_hments = 0;
7168 	CPUSET_ZERO(cpuset);
7169 
7170 	pml = sfmmu_mlist_enter(pp);
7171 
7172 #ifdef VAC
7173 	if (pp->p_kpmref)
7174 		sfmmu_kpm_pageunload(pp);
7175 	ASSERT(!PP_ISMAPPED_KPM(pp));
7176 #endif
7177 	/*
7178 	 * Clear vpm reference. Since the page is exclusively locked
7179 	 * vpm cannot be referencing it.
7180 	 */
7181 	if (vpm_enable) {
7182 		pp->p_vpmref = 0;
7183 	}
7184 
7185 	index = PP_MAPINDEX(pp);
7186 	cons = TTE8K;
7187 retry:
7188 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7189 		tmphme = sfhme->hme_next;
7190 
7191 		if (IS_PAHME(sfhme)) {
7192 			ASSERT(sfhme->hme_data != NULL);
7193 			pa_hments++;
7194 			continue;
7195 		}
7196 
7197 		hmeblkp = sfmmu_hmetohblk(sfhme);
7198 		if (hmeblkp->hblk_xhat_bit) {
7199 			struct xhat_hme_blk *xblk =
7200 			    (struct xhat_hme_blk *)hmeblkp;
7201 
7202 			(void) XHAT_PAGEUNLOAD(xblk->xhat_hme_blk_hat,
7203 			    pp, forceflag, XBLK2PROVBLK(xblk));
7204 
7205 			xhme_blks = 1;
7206 			continue;
7207 		}
7208 
7209 		/*
7210 		 * If there are kernel mappings don't unload them, they will
7211 		 * be suspended.
7212 		 */
7213 		if (forceflag == SFMMU_KERNEL_RELOC && hmeblkp->hblk_lckcnt &&
7214 		    hmeblkp->hblk_tag.htag_id == ksfmmup)
7215 			continue;
7216 
7217 		tset = sfmmu_pageunload(pp, sfhme, cons);
7218 		CPUSET_OR(cpuset, tset);
7219 	}
7220 
7221 	while (index != 0) {
7222 		index = index >> 1;
7223 		if (index != 0)
7224 			cons++;
7225 		if (index & 0x1) {
7226 			/* Go to leading page */
7227 			pp = PP_GROUPLEADER(pp, cons);
7228 			ASSERT(sfmmu_mlist_held(pp));
7229 			goto retry;
7230 		}
7231 	}
7232 
7233 	/*
7234 	 * cpuset may be empty if the page was only mapped by segkpm,
7235 	 * in which case we won't actually cross-trap.
7236 	 */
7237 	xt_sync(cpuset);
7238 
7239 	/*
7240 	 * The page should have no mappings at this point, unless
7241 	 * we were called from hat_page_relocate() in which case we
7242 	 * leave the locked mappings which will be suspended later.
7243 	 */
7244 	ASSERT(!PP_ISMAPPED(origpp) || xhme_blks || pa_hments ||
7245 	    (forceflag == SFMMU_KERNEL_RELOC));
7246 
7247 #ifdef VAC
7248 	if (PP_ISTNC(pp)) {
7249 		if (cons == TTE8K) {
7250 			pmtx = sfmmu_page_enter(pp);
7251 			PP_CLRTNC(pp);
7252 			sfmmu_page_exit(pmtx);
7253 		} else {
7254 			conv_tnc(pp, cons);
7255 		}
7256 	}
7257 #endif	/* VAC */
7258 
7259 	if (pa_hments && forceflag != SFMMU_KERNEL_RELOC) {
7260 		/*
7261 		 * Unlink any pa_hments and free them, calling back
7262 		 * the responsible subsystem to notify it of the error.
7263 		 * This can occur in situations such as drivers leaking
7264 		 * DMA handles: naughty, but common enough that we'd like
7265 		 * to keep the system running rather than bringing it
7266 		 * down with an obscure error like "pa_hment leaked"
7267 		 * which doesn't aid the user in debugging their driver.
7268 		 */
7269 		for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7270 			tmphme = sfhme->hme_next;
7271 			if (IS_PAHME(sfhme)) {
7272 				struct pa_hment *pahmep = sfhme->hme_data;
7273 				sfmmu_pahment_leaked(pahmep);
7274 				HME_SUB(sfhme, pp);
7275 				kmem_cache_free(pa_hment_cache, pahmep);
7276 			}
7277 		}
7278 
7279 		ASSERT(!PP_ISMAPPED(origpp) || xhme_blks);
7280 	}
7281 
7282 	sfmmu_mlist_exit(pml);
7283 
7284 	/*
7285 	 * XHAT may not have finished unloading pages
7286 	 * because some other thread was waiting for
7287 	 * mlist lock and XHAT_PAGEUNLOAD let it do
7288 	 * the job.
7289 	 */
7290 	if (xhme_blks) {
7291 		pp = origpp;
7292 		goto retry_xhat;
7293 	}
7294 
7295 	return (0);
7296 }
7297 
7298 cpuset_t
7299 sfmmu_pageunload(page_t *pp, struct sf_hment *sfhme, int cons)
7300 {
7301 	struct hme_blk *hmeblkp;
7302 	sfmmu_t *sfmmup;
7303 	tte_t tte, ttemod;
7304 #ifdef DEBUG
7305 	tte_t orig_old;
7306 #endif /* DEBUG */
7307 	caddr_t addr;
7308 	int ttesz;
7309 	int ret;
7310 	cpuset_t cpuset;
7311 
7312 	ASSERT(pp != NULL);
7313 	ASSERT(sfmmu_mlist_held(pp));
7314 	ASSERT(!PP_ISKAS(pp));
7315 
7316 	CPUSET_ZERO(cpuset);
7317 
7318 	hmeblkp = sfmmu_hmetohblk(sfhme);
7319 
7320 readtte:
7321 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7322 	if (TTE_IS_VALID(&tte)) {
7323 		sfmmup = hblktosfmmu(hmeblkp);
7324 		ttesz = get_hblk_ttesz(hmeblkp);
7325 		/*
7326 		 * Only unload mappings of 'cons' size.
7327 		 */
7328 		if (ttesz != cons)
7329 			return (cpuset);
7330 
7331 		/*
7332 		 * Note that we have p_mapping lock, but no hash lock here.
7333 		 * hblk_unload() has to have both hash lock AND p_mapping
7334 		 * lock before it tries to modify tte. So, the tte could
7335 		 * not become invalid in the sfmmu_modifytte_try() below.
7336 		 */
7337 		ttemod = tte;
7338 #ifdef DEBUG
7339 		orig_old = tte;
7340 #endif /* DEBUG */
7341 
7342 		TTE_SET_INVALID(&ttemod);
7343 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7344 		if (ret < 0) {
7345 #ifdef DEBUG
7346 			/* only R/M bits can change. */
7347 			chk_tte(&orig_old, &tte, &ttemod, hmeblkp);
7348 #endif /* DEBUG */
7349 			goto readtte;
7350 		}
7351 
7352 		if (ret == 0) {
7353 			panic("pageunload: cas failed?");
7354 		}
7355 
7356 		addr = tte_to_vaddr(hmeblkp, tte);
7357 
7358 		if (hmeblkp->hblk_shared) {
7359 			sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7360 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
7361 			sf_region_t *rgnp;
7362 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7363 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7364 			ASSERT(srdp != NULL);
7365 			rgnp = srdp->srd_hmergnp[rid];
7366 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
7367 			cpuset = sfmmu_rgntlb_demap(addr, rgnp, hmeblkp, 1);
7368 			sfmmu_ttesync(NULL, addr, &tte, pp);
7369 			ASSERT(rgnp->rgn_ttecnt[ttesz] > 0);
7370 			atomic_add_long(&rgnp->rgn_ttecnt[ttesz], -1);
7371 		} else {
7372 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
7373 			atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -1);
7374 
7375 			/*
7376 			 * We need to flush the page from the virtual cache
7377 			 * in order to prevent a virtual cache alias
7378 			 * inconsistency. The particular scenario we need
7379 			 * to worry about is:
7380 			 * Given:  va1 and va2 are two virtual address that
7381 			 * alias and will map the same physical address.
7382 			 * 1.   mapping exists from va1 to pa and data has
7383 			 *	been read into the cache.
7384 			 * 2.   unload va1.
7385 			 * 3.   load va2 and modify data using va2.
7386 			 * 4    unload va2.
7387 			 * 5.   load va1 and reference data.  Unless we flush
7388 			 *	the data cache when we unload we will get
7389 			 *	stale data.
7390 			 * This scenario is taken care of by using virtual
7391 			 * page coloring.
7392 			 */
7393 			if (sfmmup->sfmmu_ismhat) {
7394 				/*
7395 				 * Flush TSBs, TLBs and caches
7396 				 * of every process
7397 				 * sharing this ism segment.
7398 				 */
7399 				sfmmu_hat_lock_all();
7400 				mutex_enter(&ism_mlist_lock);
7401 				kpreempt_disable();
7402 				sfmmu_ismtlbcache_demap(addr, sfmmup, hmeblkp,
7403 				    pp->p_pagenum, CACHE_NO_FLUSH);
7404 				kpreempt_enable();
7405 				mutex_exit(&ism_mlist_lock);
7406 				sfmmu_hat_unlock_all();
7407 				cpuset = cpu_ready_set;
7408 			} else {
7409 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7410 				cpuset = sfmmup->sfmmu_cpusran;
7411 			}
7412 		}
7413 
7414 		/*
7415 		 * Hme_sub has to run after ttesync() and a_rss update.
7416 		 * See hblk_unload().
7417 		 */
7418 		HME_SUB(sfhme, pp);
7419 		membar_stst();
7420 
7421 		/*
7422 		 * We can not make ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
7423 		 * since pteload may have done a HME_ADD() right after
7424 		 * we did the HME_SUB() above. Hmecnt is now maintained
7425 		 * by cas only. no lock guranteed its value. The only
7426 		 * gurantee we have is the hmecnt should not be less than
7427 		 * what it should be so the hblk will not be taken away.
7428 		 * It's also important that we decremented the hmecnt after
7429 		 * we are done with hmeblkp so that this hmeblk won't be
7430 		 * stolen.
7431 		 */
7432 		ASSERT(hmeblkp->hblk_hmecnt > 0);
7433 		ASSERT(hmeblkp->hblk_vcnt > 0);
7434 		atomic_add_16(&hmeblkp->hblk_vcnt, -1);
7435 		atomic_add_16(&hmeblkp->hblk_hmecnt, -1);
7436 		/*
7437 		 * This is bug 4063182.
7438 		 * XXX: fixme
7439 		 * ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
7440 		 *	!hmeblkp->hblk_lckcnt);
7441 		 */
7442 	} else {
7443 		panic("invalid tte? pp %p &tte %p",
7444 		    (void *)pp, (void *)&tte);
7445 	}
7446 
7447 	return (cpuset);
7448 }
7449 
7450 /*
7451  * While relocating a kernel page, this function will move the mappings
7452  * from tpp to dpp and modify any associated data with these mappings.
7453  * It also unsuspends the suspended kernel mapping.
7454  */
7455 static void
7456 hat_pagereload(struct page *tpp, struct page *dpp)
7457 {
7458 	struct sf_hment *sfhme;
7459 	tte_t tte, ttemod;
7460 	int index, cons;
7461 
7462 	ASSERT(getpil() == PIL_MAX);
7463 	ASSERT(sfmmu_mlist_held(tpp));
7464 	ASSERT(sfmmu_mlist_held(dpp));
7465 
7466 	index = PP_MAPINDEX(tpp);
7467 	cons = TTE8K;
7468 
7469 	/* Update real mappings to the page */
7470 retry:
7471 	for (sfhme = tpp->p_mapping; sfhme != NULL; sfhme = sfhme->hme_next) {
7472 		if (IS_PAHME(sfhme))
7473 			continue;
7474 		sfmmu_copytte(&sfhme->hme_tte, &tte);
7475 		ttemod = tte;
7476 
7477 		/*
7478 		 * replace old pfn with new pfn in TTE
7479 		 */
7480 		PFN_TO_TTE(ttemod, dpp->p_pagenum);
7481 
7482 		/*
7483 		 * clear suspend bit
7484 		 */
7485 		ASSERT(TTE_IS_SUSPEND(&ttemod));
7486 		TTE_CLR_SUSPEND(&ttemod);
7487 
7488 		if (sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte) < 0)
7489 			panic("hat_pagereload(): sfmmu_modifytte_try() failed");
7490 
7491 		/*
7492 		 * set hme_page point to new page
7493 		 */
7494 		sfhme->hme_page = dpp;
7495 	}
7496 
7497 	/*
7498 	 * move p_mapping list from old page to new page
7499 	 */
7500 	dpp->p_mapping = tpp->p_mapping;
7501 	tpp->p_mapping = NULL;
7502 	dpp->p_share = tpp->p_share;
7503 	tpp->p_share = 0;
7504 
7505 	while (index != 0) {
7506 		index = index >> 1;
7507 		if (index != 0)
7508 			cons++;
7509 		if (index & 0x1) {
7510 			tpp = PP_GROUPLEADER(tpp, cons);
7511 			dpp = PP_GROUPLEADER(dpp, cons);
7512 			goto retry;
7513 		}
7514 	}
7515 
7516 	curthread->t_flag &= ~T_DONTDTRACE;
7517 	mutex_exit(&kpr_suspendlock);
7518 }
7519 
7520 uint_t
7521 hat_pagesync(struct page *pp, uint_t clearflag)
7522 {
7523 	struct sf_hment *sfhme, *tmphme = NULL;
7524 	struct hme_blk *hmeblkp;
7525 	kmutex_t *pml;
7526 	cpuset_t cpuset, tset;
7527 	int	index, cons;
7528 	extern	ulong_t po_share;
7529 	page_t	*save_pp = pp;
7530 	int	stop_on_sh = 0;
7531 	uint_t	shcnt;
7532 
7533 	CPUSET_ZERO(cpuset);
7534 
7535 	if (PP_ISRO(pp) && (clearflag & HAT_SYNC_STOPON_MOD)) {
7536 		return (PP_GENERIC_ATTR(pp));
7537 	}
7538 
7539 	if ((clearflag & HAT_SYNC_ZERORM) == 0) {
7540 		if ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(pp)) {
7541 			return (PP_GENERIC_ATTR(pp));
7542 		}
7543 		if ((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(pp)) {
7544 			return (PP_GENERIC_ATTR(pp));
7545 		}
7546 		if (clearflag & HAT_SYNC_STOPON_SHARED) {
7547 			if (pp->p_share > po_share) {
7548 				hat_page_setattr(pp, P_REF);
7549 				return (PP_GENERIC_ATTR(pp));
7550 			}
7551 			stop_on_sh = 1;
7552 			shcnt = 0;
7553 		}
7554 	}
7555 
7556 	clearflag &= ~HAT_SYNC_STOPON_SHARED;
7557 	pml = sfmmu_mlist_enter(pp);
7558 	index = PP_MAPINDEX(pp);
7559 	cons = TTE8K;
7560 retry:
7561 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7562 		/*
7563 		 * We need to save the next hment on the list since
7564 		 * it is possible for pagesync to remove an invalid hment
7565 		 * from the list.
7566 		 */
7567 		tmphme = sfhme->hme_next;
7568 		if (IS_PAHME(sfhme))
7569 			continue;
7570 		/*
7571 		 * If we are looking for large mappings and this hme doesn't
7572 		 * reach the range we are seeking, just ignore it.
7573 		 */
7574 		hmeblkp = sfmmu_hmetohblk(sfhme);
7575 		if (hmeblkp->hblk_xhat_bit)
7576 			continue;
7577 
7578 		if (hme_size(sfhme) < cons)
7579 			continue;
7580 
7581 		if (stop_on_sh) {
7582 			if (hmeblkp->hblk_shared) {
7583 				sf_srd_t *srdp = hblktosrd(hmeblkp);
7584 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7585 				sf_region_t *rgnp;
7586 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7587 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7588 				ASSERT(srdp != NULL);
7589 				rgnp = srdp->srd_hmergnp[rid];
7590 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
7591 				    rgnp, rid);
7592 				shcnt += rgnp->rgn_refcnt;
7593 			} else {
7594 				shcnt++;
7595 			}
7596 			if (shcnt > po_share) {
7597 				/*
7598 				 * tell the pager to spare the page this time
7599 				 * around.
7600 				 */
7601 				hat_page_setattr(save_pp, P_REF);
7602 				index = 0;
7603 				break;
7604 			}
7605 		}
7606 		tset = sfmmu_pagesync(pp, sfhme,
7607 		    clearflag & ~HAT_SYNC_STOPON_RM);
7608 		CPUSET_OR(cpuset, tset);
7609 
7610 		/*
7611 		 * If clearflag is HAT_SYNC_DONTZERO, break out as soon
7612 		 * as the "ref" or "mod" is set or share cnt exceeds po_share.
7613 		 */
7614 		if ((clearflag & ~HAT_SYNC_STOPON_RM) == HAT_SYNC_DONTZERO &&
7615 		    (((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(save_pp)) ||
7616 		    ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(save_pp)))) {
7617 			index = 0;
7618 			break;
7619 		}
7620 	}
7621 
7622 	while (index) {
7623 		index = index >> 1;
7624 		cons++;
7625 		if (index & 0x1) {
7626 			/* Go to leading page */
7627 			pp = PP_GROUPLEADER(pp, cons);
7628 			goto retry;
7629 		}
7630 	}
7631 
7632 	xt_sync(cpuset);
7633 	sfmmu_mlist_exit(pml);
7634 	return (PP_GENERIC_ATTR(save_pp));
7635 }
7636 
7637 /*
7638  * Get all the hardware dependent attributes for a page struct
7639  */
7640 static cpuset_t
7641 sfmmu_pagesync(struct page *pp, struct sf_hment *sfhme,
7642 	uint_t clearflag)
7643 {
7644 	caddr_t addr;
7645 	tte_t tte, ttemod;
7646 	struct hme_blk *hmeblkp;
7647 	int ret;
7648 	sfmmu_t *sfmmup;
7649 	cpuset_t cpuset;
7650 
7651 	ASSERT(pp != NULL);
7652 	ASSERT(sfmmu_mlist_held(pp));
7653 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
7654 	    (clearflag == HAT_SYNC_ZERORM));
7655 
7656 	SFMMU_STAT(sf_pagesync);
7657 
7658 	CPUSET_ZERO(cpuset);
7659 
7660 sfmmu_pagesync_retry:
7661 
7662 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7663 	if (TTE_IS_VALID(&tte)) {
7664 		hmeblkp = sfmmu_hmetohblk(sfhme);
7665 		sfmmup = hblktosfmmu(hmeblkp);
7666 		addr = tte_to_vaddr(hmeblkp, tte);
7667 		if (clearflag == HAT_SYNC_ZERORM) {
7668 			ttemod = tte;
7669 			TTE_CLR_RM(&ttemod);
7670 			ret = sfmmu_modifytte_try(&tte, &ttemod,
7671 			    &sfhme->hme_tte);
7672 			if (ret < 0) {
7673 				/*
7674 				 * cas failed and the new value is not what
7675 				 * we want.
7676 				 */
7677 				goto sfmmu_pagesync_retry;
7678 			}
7679 
7680 			if (ret > 0) {
7681 				/* we win the cas */
7682 				if (hmeblkp->hblk_shared) {
7683 					sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7684 					uint_t rid =
7685 					    hmeblkp->hblk_tag.htag_rid;
7686 					sf_region_t *rgnp;
7687 					ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7688 					ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7689 					ASSERT(srdp != NULL);
7690 					rgnp = srdp->srd_hmergnp[rid];
7691 					SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7692 					    srdp, rgnp, rid);
7693 					cpuset = sfmmu_rgntlb_demap(addr,
7694 					    rgnp, hmeblkp, 1);
7695 				} else {
7696 					sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
7697 					    0, 0);
7698 					cpuset = sfmmup->sfmmu_cpusran;
7699 				}
7700 			}
7701 		}
7702 		sfmmu_ttesync(hmeblkp->hblk_shared ? NULL : sfmmup, addr,
7703 		    &tte, pp);
7704 	}
7705 	return (cpuset);
7706 }
7707 
7708 /*
7709  * Remove write permission from a mappings to a page, so that
7710  * we can detect the next modification of it. This requires modifying
7711  * the TTE then invalidating (demap) any TLB entry using that TTE.
7712  * This code is similar to sfmmu_pagesync().
7713  */
7714 static cpuset_t
7715 sfmmu_pageclrwrt(struct page *pp, struct sf_hment *sfhme)
7716 {
7717 	caddr_t addr;
7718 	tte_t tte;
7719 	tte_t ttemod;
7720 	struct hme_blk *hmeblkp;
7721 	int ret;
7722 	sfmmu_t *sfmmup;
7723 	cpuset_t cpuset;
7724 
7725 	ASSERT(pp != NULL);
7726 	ASSERT(sfmmu_mlist_held(pp));
7727 
7728 	CPUSET_ZERO(cpuset);
7729 	SFMMU_STAT(sf_clrwrt);
7730 
7731 retry:
7732 
7733 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7734 	if (TTE_IS_VALID(&tte) && TTE_IS_WRITABLE(&tte)) {
7735 		hmeblkp = sfmmu_hmetohblk(sfhme);
7736 
7737 		/*
7738 		 * xhat mappings should never be to a VMODSORT page.
7739 		 */
7740 		ASSERT(hmeblkp->hblk_xhat_bit == 0);
7741 
7742 		sfmmup = hblktosfmmu(hmeblkp);
7743 		addr = tte_to_vaddr(hmeblkp, tte);
7744 
7745 		ttemod = tte;
7746 		TTE_CLR_WRT(&ttemod);
7747 		TTE_CLR_MOD(&ttemod);
7748 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7749 
7750 		/*
7751 		 * if cas failed and the new value is not what
7752 		 * we want retry
7753 		 */
7754 		if (ret < 0)
7755 			goto retry;
7756 
7757 		/* we win the cas */
7758 		if (ret > 0) {
7759 			if (hmeblkp->hblk_shared) {
7760 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7761 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7762 				sf_region_t *rgnp;
7763 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7764 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7765 				ASSERT(srdp != NULL);
7766 				rgnp = srdp->srd_hmergnp[rid];
7767 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7768 				    srdp, rgnp, rid);
7769 				cpuset = sfmmu_rgntlb_demap(addr,
7770 				    rgnp, hmeblkp, 1);
7771 			} else {
7772 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7773 				cpuset = sfmmup->sfmmu_cpusran;
7774 			}
7775 		}
7776 	}
7777 
7778 	return (cpuset);
7779 }
7780 
7781 /*
7782  * Walk all mappings of a page, removing write permission and clearing the
7783  * ref/mod bits. This code is similar to hat_pagesync()
7784  */
7785 static void
7786 hat_page_clrwrt(page_t *pp)
7787 {
7788 	struct sf_hment *sfhme;
7789 	struct sf_hment *tmphme = NULL;
7790 	kmutex_t *pml;
7791 	cpuset_t cpuset;
7792 	cpuset_t tset;
7793 	int	index;
7794 	int	 cons;
7795 
7796 	CPUSET_ZERO(cpuset);
7797 
7798 	pml = sfmmu_mlist_enter(pp);
7799 	index = PP_MAPINDEX(pp);
7800 	cons = TTE8K;
7801 retry:
7802 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7803 		tmphme = sfhme->hme_next;
7804 
7805 		/*
7806 		 * If we are looking for large mappings and this hme doesn't
7807 		 * reach the range we are seeking, just ignore its.
7808 		 */
7809 
7810 		if (hme_size(sfhme) < cons)
7811 			continue;
7812 
7813 		tset = sfmmu_pageclrwrt(pp, sfhme);
7814 		CPUSET_OR(cpuset, tset);
7815 	}
7816 
7817 	while (index) {
7818 		index = index >> 1;
7819 		cons++;
7820 		if (index & 0x1) {
7821 			/* Go to leading page */
7822 			pp = PP_GROUPLEADER(pp, cons);
7823 			goto retry;
7824 		}
7825 	}
7826 
7827 	xt_sync(cpuset);
7828 	sfmmu_mlist_exit(pml);
7829 }
7830 
7831 /*
7832  * Set the given REF/MOD/RO bits for the given page.
7833  * For a vnode with a sorted v_pages list, we need to change
7834  * the attributes and the v_pages list together under page_vnode_mutex.
7835  */
7836 void
7837 hat_page_setattr(page_t *pp, uint_t flag)
7838 {
7839 	vnode_t		*vp = pp->p_vnode;
7840 	page_t		**listp;
7841 	kmutex_t	*pmtx;
7842 	kmutex_t	*vphm = NULL;
7843 	int		noshuffle;
7844 
7845 	noshuffle = flag & P_NSH;
7846 	flag &= ~P_NSH;
7847 
7848 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7849 
7850 	/*
7851 	 * nothing to do if attribute already set
7852 	 */
7853 	if ((pp->p_nrm & flag) == flag)
7854 		return;
7855 
7856 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp) &&
7857 	    !noshuffle) {
7858 		vphm = page_vnode_mutex(vp);
7859 		mutex_enter(vphm);
7860 	}
7861 
7862 	pmtx = sfmmu_page_enter(pp);
7863 	pp->p_nrm |= flag;
7864 	sfmmu_page_exit(pmtx);
7865 
7866 	if (vphm != NULL) {
7867 		/*
7868 		 * Some File Systems examine v_pages for NULL w/o
7869 		 * grabbing the vphm mutex. Must not let it become NULL when
7870 		 * pp is the only page on the list.
7871 		 */
7872 		if (pp->p_vpnext != pp) {
7873 			page_vpsub(&vp->v_pages, pp);
7874 			if (vp->v_pages != NULL)
7875 				listp = &vp->v_pages->p_vpprev->p_vpnext;
7876 			else
7877 				listp = &vp->v_pages;
7878 			page_vpadd(listp, pp);
7879 		}
7880 		mutex_exit(vphm);
7881 	}
7882 }
7883 
7884 void
7885 hat_page_clrattr(page_t *pp, uint_t flag)
7886 {
7887 	vnode_t		*vp = pp->p_vnode;
7888 	kmutex_t	*pmtx;
7889 
7890 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7891 
7892 	pmtx = sfmmu_page_enter(pp);
7893 
7894 	/*
7895 	 * Caller is expected to hold page's io lock for VMODSORT to work
7896 	 * correctly with pvn_vplist_dirty() and pvn_getdirty() when mod
7897 	 * bit is cleared.
7898 	 * We don't have assert to avoid tripping some existing third party
7899 	 * code. The dirty page is moved back to top of the v_page list
7900 	 * after IO is done in pvn_write_done().
7901 	 */
7902 	pp->p_nrm &= ~flag;
7903 	sfmmu_page_exit(pmtx);
7904 
7905 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) {
7906 
7907 		/*
7908 		 * VMODSORT works by removing write permissions and getting
7909 		 * a fault when a page is made dirty. At this point
7910 		 * we need to remove write permission from all mappings
7911 		 * to this page.
7912 		 */
7913 		hat_page_clrwrt(pp);
7914 	}
7915 }
7916 
7917 uint_t
7918 hat_page_getattr(page_t *pp, uint_t flag)
7919 {
7920 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7921 	return ((uint_t)(pp->p_nrm & flag));
7922 }
7923 
7924 /*
7925  * DEBUG kernels: verify that a kernel va<->pa translation
7926  * is safe by checking the underlying page_t is in a page
7927  * relocation-safe state.
7928  */
7929 #ifdef	DEBUG
7930 void
7931 sfmmu_check_kpfn(pfn_t pfn)
7932 {
7933 	page_t *pp;
7934 	int index, cons;
7935 
7936 	if (hat_check_vtop == 0)
7937 		return;
7938 
7939 	if (hat_kpr_enabled == 0 || kvseg.s_base == NULL || panicstr)
7940 		return;
7941 
7942 	pp = page_numtopp_nolock(pfn);
7943 	if (!pp)
7944 		return;
7945 
7946 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7947 		return;
7948 
7949 	/*
7950 	 * Handed a large kernel page, we dig up the root page since we
7951 	 * know the root page might have the lock also.
7952 	 */
7953 	if (pp->p_szc != 0) {
7954 		index = PP_MAPINDEX(pp);
7955 		cons = TTE8K;
7956 again:
7957 		while (index != 0) {
7958 			index >>= 1;
7959 			if (index != 0)
7960 				cons++;
7961 			if (index & 0x1) {
7962 				pp = PP_GROUPLEADER(pp, cons);
7963 				goto again;
7964 			}
7965 		}
7966 	}
7967 
7968 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7969 		return;
7970 
7971 	/*
7972 	 * Pages need to be locked or allocated "permanent" (either from
7973 	 * static_arena arena or explicitly setting PG_NORELOC when calling
7974 	 * page_create_va()) for VA->PA translations to be valid.
7975 	 */
7976 	if (!PP_ISNORELOC(pp))
7977 		panic("Illegal VA->PA translation, pp 0x%p not permanent",
7978 		    (void *)pp);
7979 	else
7980 		panic("Illegal VA->PA translation, pp 0x%p not locked",
7981 		    (void *)pp);
7982 }
7983 #endif	/* DEBUG */
7984 
7985 /*
7986  * Returns a page frame number for a given virtual address.
7987  * Returns PFN_INVALID to indicate an invalid mapping
7988  */
7989 pfn_t
7990 hat_getpfnum(struct hat *hat, caddr_t addr)
7991 {
7992 	pfn_t pfn;
7993 	tte_t tte;
7994 
7995 	/*
7996 	 * We would like to
7997 	 * ASSERT(AS_LOCK_HELD(as, &as->a_lock));
7998 	 * but we can't because the iommu driver will call this
7999 	 * routine at interrupt time and it can't grab the as lock
8000 	 * or it will deadlock: A thread could have the as lock
8001 	 * and be waiting for io.  The io can't complete
8002 	 * because the interrupt thread is blocked trying to grab
8003 	 * the as lock.
8004 	 */
8005 
8006 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8007 
8008 	if (hat == ksfmmup) {
8009 		if (IS_KMEM_VA_LARGEPAGE(addr)) {
8010 			ASSERT(segkmem_lpszc > 0);
8011 			pfn = sfmmu_kvaszc2pfn(addr, segkmem_lpszc);
8012 			if (pfn != PFN_INVALID) {
8013 				sfmmu_check_kpfn(pfn);
8014 				return (pfn);
8015 			}
8016 		} else if (segkpm && IS_KPM_ADDR(addr)) {
8017 			return (sfmmu_kpm_vatopfn(addr));
8018 		}
8019 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
8020 		    == PFN_SUSPENDED) {
8021 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
8022 		}
8023 		sfmmu_check_kpfn(pfn);
8024 		return (pfn);
8025 	} else {
8026 		return (sfmmu_uvatopfn(addr, hat, NULL));
8027 	}
8028 }
8029 
8030 /*
8031  * hat_getkpfnum() is an obsolete DDI routine, and its use is discouraged.
8032  * Use hat_getpfnum(kas.a_hat, ...) instead.
8033  *
8034  * We'd like to return PFN_INVALID if the mappings have underlying page_t's
8035  * but can't right now due to the fact that some software has grown to use
8036  * this interface incorrectly. So for now when the interface is misused,
8037  * return a warning to the user that in the future it won't work in the
8038  * way they're abusing it, and carry on (after disabling page relocation).
8039  */
8040 pfn_t
8041 hat_getkpfnum(caddr_t addr)
8042 {
8043 	pfn_t pfn;
8044 	tte_t tte;
8045 	int badcaller = 0;
8046 	extern int segkmem_reloc;
8047 
8048 	if (segkpm && IS_KPM_ADDR(addr)) {
8049 		badcaller = 1;
8050 		pfn = sfmmu_kpm_vatopfn(addr);
8051 	} else {
8052 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
8053 		    == PFN_SUSPENDED) {
8054 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
8055 		}
8056 		badcaller = pf_is_memory(pfn);
8057 	}
8058 
8059 	if (badcaller) {
8060 		/*
8061 		 * We can't return PFN_INVALID or the caller may panic
8062 		 * or corrupt the system.  The only alternative is to
8063 		 * disable page relocation at this point for all kernel
8064 		 * memory.  This will impact any callers of page_relocate()
8065 		 * such as FMA or DR.
8066 		 *
8067 		 * RFE: Add junk here to spit out an ereport so the sysadmin
8068 		 * can be advised that he should upgrade his device driver
8069 		 * so that this doesn't happen.
8070 		 */
8071 		hat_getkpfnum_badcall(caller());
8072 		if (hat_kpr_enabled && segkmem_reloc) {
8073 			hat_kpr_enabled = 0;
8074 			segkmem_reloc = 0;
8075 			cmn_err(CE_WARN, "Kernel Page Relocation is DISABLED");
8076 		}
8077 	}
8078 	return (pfn);
8079 }
8080 
8081 /*
8082  * This routine will return both pfn and tte for the vaddr.
8083  */
8084 static pfn_t
8085 sfmmu_uvatopfn(caddr_t vaddr, struct hat *sfmmup, tte_t *ttep)
8086 {
8087 	struct hmehash_bucket *hmebp;
8088 	hmeblk_tag hblktag;
8089 	int hmeshift, hashno = 1;
8090 	struct hme_blk *hmeblkp = NULL;
8091 	tte_t tte;
8092 
8093 	struct sf_hment *sfhmep;
8094 	pfn_t pfn;
8095 
8096 	/* support for ISM */
8097 	ism_map_t	*ism_map;
8098 	ism_blk_t	*ism_blkp;
8099 	int		i;
8100 	sfmmu_t *ism_hatid = NULL;
8101 	sfmmu_t *locked_hatid = NULL;
8102 	sfmmu_t	*sv_sfmmup = sfmmup;
8103 	caddr_t	sv_vaddr = vaddr;
8104 	sf_srd_t *srdp;
8105 
8106 	if (ttep == NULL) {
8107 		ttep = &tte;
8108 	} else {
8109 		ttep->ll = 0;
8110 	}
8111 
8112 	ASSERT(sfmmup != ksfmmup);
8113 	SFMMU_STAT(sf_user_vtop);
8114 	/*
8115 	 * Set ism_hatid if vaddr falls in a ISM segment.
8116 	 */
8117 	ism_blkp = sfmmup->sfmmu_iblk;
8118 	if (ism_blkp != NULL) {
8119 		sfmmu_ismhat_enter(sfmmup, 0);
8120 		locked_hatid = sfmmup;
8121 	}
8122 	while (ism_blkp != NULL && ism_hatid == NULL) {
8123 		ism_map = ism_blkp->iblk_maps;
8124 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
8125 			if (vaddr >= ism_start(ism_map[i]) &&
8126 			    vaddr < ism_end(ism_map[i])) {
8127 				sfmmup = ism_hatid = ism_map[i].imap_ismhat;
8128 				vaddr = (caddr_t)(vaddr -
8129 				    ism_start(ism_map[i]));
8130 				break;
8131 			}
8132 		}
8133 		ism_blkp = ism_blkp->iblk_next;
8134 	}
8135 	if (locked_hatid) {
8136 		sfmmu_ismhat_exit(locked_hatid, 0);
8137 	}
8138 
8139 	hblktag.htag_id = sfmmup;
8140 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
8141 	do {
8142 		hmeshift = HME_HASH_SHIFT(hashno);
8143 		hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
8144 		hblktag.htag_rehash = hashno;
8145 		hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
8146 
8147 		SFMMU_HASH_LOCK(hmebp);
8148 
8149 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
8150 		if (hmeblkp != NULL) {
8151 			ASSERT(!hmeblkp->hblk_shared);
8152 			HBLKTOHME(sfhmep, hmeblkp, vaddr);
8153 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
8154 			SFMMU_HASH_UNLOCK(hmebp);
8155 			if (TTE_IS_VALID(ttep)) {
8156 				pfn = TTE_TO_PFN(vaddr, ttep);
8157 				return (pfn);
8158 			}
8159 			break;
8160 		}
8161 		SFMMU_HASH_UNLOCK(hmebp);
8162 		hashno++;
8163 	} while (HME_REHASH(sfmmup) && (hashno <= mmu_hashcnt));
8164 
8165 	if (SF_HMERGNMAP_ISNULL(sv_sfmmup)) {
8166 		return (PFN_INVALID);
8167 	}
8168 	srdp = sv_sfmmup->sfmmu_srdp;
8169 	ASSERT(srdp != NULL);
8170 	ASSERT(srdp->srd_refcnt != 0);
8171 	hblktag.htag_id = srdp;
8172 	hashno = 1;
8173 	do {
8174 		hmeshift = HME_HASH_SHIFT(hashno);
8175 		hblktag.htag_bspage = HME_HASH_BSPAGE(sv_vaddr, hmeshift);
8176 		hblktag.htag_rehash = hashno;
8177 		hmebp = HME_HASH_FUNCTION(srdp, sv_vaddr, hmeshift);
8178 
8179 		SFMMU_HASH_LOCK(hmebp);
8180 		for (hmeblkp = hmebp->hmeblkp; hmeblkp != NULL;
8181 		    hmeblkp = hmeblkp->hblk_next) {
8182 			uint_t rid;
8183 			sf_region_t *rgnp;
8184 			caddr_t rsaddr;
8185 			caddr_t readdr;
8186 
8187 			if (!HTAGS_EQ_SHME(hmeblkp->hblk_tag, hblktag,
8188 			    sv_sfmmup->sfmmu_hmeregion_map)) {
8189 				continue;
8190 			}
8191 			ASSERT(hmeblkp->hblk_shared);
8192 			rid = hmeblkp->hblk_tag.htag_rid;
8193 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8194 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8195 			rgnp = srdp->srd_hmergnp[rid];
8196 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
8197 			HBLKTOHME(sfhmep, hmeblkp, sv_vaddr);
8198 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
8199 			rsaddr = rgnp->rgn_saddr;
8200 			readdr = rsaddr + rgnp->rgn_size;
8201 #ifdef DEBUG
8202 			if (TTE_IS_VALID(ttep) ||
8203 			    get_hblk_ttesz(hmeblkp) > TTE8K) {
8204 				caddr_t eva = tte_to_evaddr(hmeblkp, ttep);
8205 				ASSERT(eva > sv_vaddr);
8206 				ASSERT(sv_vaddr >= rsaddr);
8207 				ASSERT(sv_vaddr < readdr);
8208 				ASSERT(eva <= readdr);
8209 			}
8210 #endif /* DEBUG */
8211 			/*
8212 			 * Continue the search if we
8213 			 * found an invalid 8K tte outside of the area
8214 			 * covered by this hmeblk's region.
8215 			 */
8216 			if (TTE_IS_VALID(ttep)) {
8217 				SFMMU_HASH_UNLOCK(hmebp);
8218 				pfn = TTE_TO_PFN(sv_vaddr, ttep);
8219 				return (pfn);
8220 			} else if (get_hblk_ttesz(hmeblkp) > TTE8K ||
8221 			    (sv_vaddr >= rsaddr && sv_vaddr < readdr)) {
8222 				SFMMU_HASH_UNLOCK(hmebp);
8223 				pfn = PFN_INVALID;
8224 				return (pfn);
8225 			}
8226 		}
8227 		SFMMU_HASH_UNLOCK(hmebp);
8228 		hashno++;
8229 	} while (hashno <= mmu_hashcnt);
8230 	return (PFN_INVALID);
8231 }
8232 
8233 
8234 /*
8235  * For compatability with AT&T and later optimizations
8236  */
8237 /* ARGSUSED */
8238 void
8239 hat_map(struct hat *hat, caddr_t addr, size_t len, uint_t flags)
8240 {
8241 	ASSERT(hat != NULL);
8242 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8243 }
8244 
8245 /*
8246  * Return the number of mappings to a particular page.  This number is an
8247  * approximation of the number of people sharing the page.
8248  *
8249  * shared hmeblks or ism hmeblks are counted as 1 mapping here.
8250  * hat_page_checkshare() can be used to compare threshold to share
8251  * count that reflects the number of region sharers albeit at higher cost.
8252  */
8253 ulong_t
8254 hat_page_getshare(page_t *pp)
8255 {
8256 	page_t *spp = pp;	/* start page */
8257 	kmutex_t *pml;
8258 	ulong_t	cnt;
8259 	int index, sz = TTE64K;
8260 
8261 	/*
8262 	 * We need to grab the mlist lock to make sure any outstanding
8263 	 * load/unloads complete.  Otherwise we could return zero
8264 	 * even though the unload(s) hasn't finished yet.
8265 	 */
8266 	pml = sfmmu_mlist_enter(spp);
8267 	cnt = spp->p_share;
8268 
8269 #ifdef VAC
8270 	if (kpm_enable)
8271 		cnt += spp->p_kpmref;
8272 #endif
8273 	if (vpm_enable && pp->p_vpmref) {
8274 		cnt += 1;
8275 	}
8276 
8277 	/*
8278 	 * If we have any large mappings, we count the number of
8279 	 * mappings that this large page is part of.
8280 	 */
8281 	index = PP_MAPINDEX(spp);
8282 	index >>= 1;
8283 	while (index) {
8284 		pp = PP_GROUPLEADER(spp, sz);
8285 		if ((index & 0x1) && pp != spp) {
8286 			cnt += pp->p_share;
8287 			spp = pp;
8288 		}
8289 		index >>= 1;
8290 		sz++;
8291 	}
8292 	sfmmu_mlist_exit(pml);
8293 	return (cnt);
8294 }
8295 
8296 /*
8297  * Return 1 if the number of mappings exceeds sh_thresh. Return 0
8298  * otherwise. Count shared hmeblks by region's refcnt.
8299  */
8300 int
8301 hat_page_checkshare(page_t *pp, ulong_t sh_thresh)
8302 {
8303 	kmutex_t *pml;
8304 	ulong_t	cnt = 0;
8305 	int index, sz = TTE8K;
8306 	struct sf_hment *sfhme, *tmphme = NULL;
8307 	struct hme_blk *hmeblkp;
8308 
8309 	pml = sfmmu_mlist_enter(pp);
8310 
8311 #ifdef VAC
8312 	if (kpm_enable)
8313 		cnt = pp->p_kpmref;
8314 #endif
8315 
8316 	if (vpm_enable && pp->p_vpmref) {
8317 		cnt += 1;
8318 	}
8319 
8320 	if (pp->p_share + cnt > sh_thresh) {
8321 		sfmmu_mlist_exit(pml);
8322 		return (1);
8323 	}
8324 
8325 	index = PP_MAPINDEX(pp);
8326 
8327 again:
8328 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
8329 		tmphme = sfhme->hme_next;
8330 		if (IS_PAHME(sfhme)) {
8331 			continue;
8332 		}
8333 
8334 		hmeblkp = sfmmu_hmetohblk(sfhme);
8335 		if (hmeblkp->hblk_xhat_bit) {
8336 			cnt++;
8337 			if (cnt > sh_thresh) {
8338 				sfmmu_mlist_exit(pml);
8339 				return (1);
8340 			}
8341 			continue;
8342 		}
8343 		if (hme_size(sfhme) != sz) {
8344 			continue;
8345 		}
8346 
8347 		if (hmeblkp->hblk_shared) {
8348 			sf_srd_t *srdp = hblktosrd(hmeblkp);
8349 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
8350 			sf_region_t *rgnp;
8351 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8352 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8353 			ASSERT(srdp != NULL);
8354 			rgnp = srdp->srd_hmergnp[rid];
8355 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
8356 			    rgnp, rid);
8357 			cnt += rgnp->rgn_refcnt;
8358 		} else {
8359 			cnt++;
8360 		}
8361 		if (cnt > sh_thresh) {
8362 			sfmmu_mlist_exit(pml);
8363 			return (1);
8364 		}
8365 	}
8366 
8367 	index >>= 1;
8368 	sz++;
8369 	while (index) {
8370 		pp = PP_GROUPLEADER(pp, sz);
8371 		ASSERT(sfmmu_mlist_held(pp));
8372 		if (index & 0x1) {
8373 			goto again;
8374 		}
8375 		index >>= 1;
8376 		sz++;
8377 	}
8378 	sfmmu_mlist_exit(pml);
8379 	return (0);
8380 }
8381 
8382 /*
8383  * Unload all large mappings to the pp and reset the p_szc field of every
8384  * constituent page according to the remaining mappings.
8385  *
8386  * pp must be locked SE_EXCL. Even though no other constituent pages are
8387  * locked it's legal to unload the large mappings to the pp because all
8388  * constituent pages of large locked mappings have to be locked SE_SHARED.
8389  * This means if we have SE_EXCL lock on one of constituent pages none of the
8390  * large mappings to pp are locked.
8391  *
8392  * Decrease p_szc field starting from the last constituent page and ending
8393  * with the root page. This method is used because other threads rely on the
8394  * root's p_szc to find the lock to syncronize on. After a root page_t's p_szc
8395  * is demoted then other threads will succeed in sfmmu_mlspl_enter(). This
8396  * ensures that p_szc changes of the constituent pages appears atomic for all
8397  * threads that use sfmmu_mlspl_enter() to examine p_szc field.
8398  *
8399  * This mechanism is only used for file system pages where it's not always
8400  * possible to get SE_EXCL locks on all constituent pages to demote the size
8401  * code (as is done for anonymous or kernel large pages).
8402  *
8403  * See more comments in front of sfmmu_mlspl_enter().
8404  */
8405 void
8406 hat_page_demote(page_t *pp)
8407 {
8408 	int index;
8409 	int sz;
8410 	cpuset_t cpuset;
8411 	int sync = 0;
8412 	page_t *rootpp;
8413 	struct sf_hment *sfhme;
8414 	struct sf_hment *tmphme = NULL;
8415 	struct hme_blk *hmeblkp;
8416 	uint_t pszc;
8417 	page_t *lastpp;
8418 	cpuset_t tset;
8419 	pgcnt_t npgs;
8420 	kmutex_t *pml;
8421 	kmutex_t *pmtx = NULL;
8422 
8423 	ASSERT(PAGE_EXCL(pp));
8424 	ASSERT(!PP_ISFREE(pp));
8425 	ASSERT(!PP_ISKAS(pp));
8426 	ASSERT(page_szc_lock_assert(pp));
8427 	pml = sfmmu_mlist_enter(pp);
8428 
8429 	pszc = pp->p_szc;
8430 	if (pszc == 0) {
8431 		goto out;
8432 	}
8433 
8434 	index = PP_MAPINDEX(pp) >> 1;
8435 
8436 	if (index) {
8437 		CPUSET_ZERO(cpuset);
8438 		sz = TTE64K;
8439 		sync = 1;
8440 	}
8441 
8442 	while (index) {
8443 		if (!(index & 0x1)) {
8444 			index >>= 1;
8445 			sz++;
8446 			continue;
8447 		}
8448 		ASSERT(sz <= pszc);
8449 		rootpp = PP_GROUPLEADER(pp, sz);
8450 		for (sfhme = rootpp->p_mapping; sfhme; sfhme = tmphme) {
8451 			tmphme = sfhme->hme_next;
8452 			ASSERT(!IS_PAHME(sfhme));
8453 			hmeblkp = sfmmu_hmetohblk(sfhme);
8454 			if (hme_size(sfhme) != sz) {
8455 				continue;
8456 			}
8457 			if (hmeblkp->hblk_xhat_bit) {
8458 				cmn_err(CE_PANIC,
8459 				    "hat_page_demote: xhat hmeblk");
8460 			}
8461 			tset = sfmmu_pageunload(rootpp, sfhme, sz);
8462 			CPUSET_OR(cpuset, tset);
8463 		}
8464 		if (index >>= 1) {
8465 			sz++;
8466 		}
8467 	}
8468 
8469 	ASSERT(!PP_ISMAPPED_LARGE(pp));
8470 
8471 	if (sync) {
8472 		xt_sync(cpuset);
8473 #ifdef VAC
8474 		if (PP_ISTNC(pp)) {
8475 			conv_tnc(rootpp, sz);
8476 		}
8477 #endif	/* VAC */
8478 	}
8479 
8480 	pmtx = sfmmu_page_enter(pp);
8481 
8482 	ASSERT(pp->p_szc == pszc);
8483 	rootpp = PP_PAGEROOT(pp);
8484 	ASSERT(rootpp->p_szc == pszc);
8485 	lastpp = PP_PAGENEXT_N(rootpp, TTEPAGES(pszc) - 1);
8486 
8487 	while (lastpp != rootpp) {
8488 		sz = PP_MAPINDEX(lastpp) ? fnd_mapping_sz(lastpp) : 0;
8489 		ASSERT(sz < pszc);
8490 		npgs = (sz == 0) ? 1 : TTEPAGES(sz);
8491 		ASSERT(P2PHASE(lastpp->p_pagenum, npgs) == npgs - 1);
8492 		while (--npgs > 0) {
8493 			lastpp->p_szc = (uchar_t)sz;
8494 			lastpp = PP_PAGEPREV(lastpp);
8495 		}
8496 		if (sz) {
8497 			/*
8498 			 * make sure before current root's pszc
8499 			 * is updated all updates to constituent pages pszc
8500 			 * fields are globally visible.
8501 			 */
8502 			membar_producer();
8503 		}
8504 		lastpp->p_szc = sz;
8505 		ASSERT(IS_P2ALIGNED(lastpp->p_pagenum, TTEPAGES(sz)));
8506 		if (lastpp != rootpp) {
8507 			lastpp = PP_PAGEPREV(lastpp);
8508 		}
8509 	}
8510 	if (sz == 0) {
8511 		/* the loop above doesn't cover this case */
8512 		rootpp->p_szc = 0;
8513 	}
8514 out:
8515 	ASSERT(pp->p_szc == 0);
8516 	if (pmtx != NULL) {
8517 		sfmmu_page_exit(pmtx);
8518 	}
8519 	sfmmu_mlist_exit(pml);
8520 }
8521 
8522 /*
8523  * Refresh the HAT ismttecnt[] element for size szc.
8524  * Caller must have set ISM busy flag to prevent mapping
8525  * lists from changing while we're traversing them.
8526  */
8527 pgcnt_t
8528 ism_tsb_entries(sfmmu_t *sfmmup, int szc)
8529 {
8530 	ism_blk_t	*ism_blkp = sfmmup->sfmmu_iblk;
8531 	ism_map_t	*ism_map;
8532 	pgcnt_t		npgs = 0;
8533 	pgcnt_t		npgs_scd = 0;
8534 	int		j;
8535 	sf_scd_t	*scdp;
8536 	uchar_t		rid;
8537 
8538 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
8539 	scdp = sfmmup->sfmmu_scdp;
8540 
8541 	for (; ism_blkp != NULL; ism_blkp = ism_blkp->iblk_next) {
8542 		ism_map = ism_blkp->iblk_maps;
8543 		for (j = 0; ism_map[j].imap_ismhat && j < ISM_MAP_SLOTS; j++) {
8544 			rid = ism_map[j].imap_rid;
8545 			ASSERT(rid == SFMMU_INVALID_ISMRID ||
8546 			    rid < sfmmup->sfmmu_srdp->srd_next_ismrid);
8547 
8548 			if (scdp != NULL && rid != SFMMU_INVALID_ISMRID &&
8549 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
8550 				/* ISM is in sfmmup's SCD */
8551 				npgs_scd +=
8552 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8553 			} else {
8554 				/* ISMs is not in SCD */
8555 				npgs +=
8556 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8557 			}
8558 		}
8559 	}
8560 	sfmmup->sfmmu_ismttecnt[szc] = npgs;
8561 	sfmmup->sfmmu_scdismttecnt[szc] = npgs_scd;
8562 	return (npgs);
8563 }
8564 
8565 /*
8566  * Yield the memory claim requirement for an address space.
8567  *
8568  * This is currently implemented as the number of bytes that have active
8569  * hardware translations that have page structures.  Therefore, it can
8570  * underestimate the traditional resident set size, eg, if the
8571  * physical page is present and the hardware translation is missing;
8572  * and it can overestimate the rss, eg, if there are active
8573  * translations to a frame buffer with page structs.
8574  * Also, it does not take sharing into account.
8575  *
8576  * Note that we don't acquire locks here since this function is most often
8577  * called from the clock thread.
8578  */
8579 size_t
8580 hat_get_mapped_size(struct hat *hat)
8581 {
8582 	size_t		assize = 0;
8583 	int 		i;
8584 
8585 	if (hat == NULL)
8586 		return (0);
8587 
8588 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8589 
8590 	for (i = 0; i < mmu_page_sizes; i++)
8591 		assize += ((pgcnt_t)hat->sfmmu_ttecnt[i] +
8592 		    (pgcnt_t)hat->sfmmu_scdrttecnt[i]) * TTEBYTES(i);
8593 
8594 	if (hat->sfmmu_iblk == NULL)
8595 		return (assize);
8596 
8597 	for (i = 0; i < mmu_page_sizes; i++)
8598 		assize += ((pgcnt_t)hat->sfmmu_ismttecnt[i] +
8599 		    (pgcnt_t)hat->sfmmu_scdismttecnt[i]) * TTEBYTES(i);
8600 
8601 	return (assize);
8602 }
8603 
8604 int
8605 hat_stats_enable(struct hat *hat)
8606 {
8607 	hatlock_t	*hatlockp;
8608 
8609 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8610 
8611 	hatlockp = sfmmu_hat_enter(hat);
8612 	hat->sfmmu_rmstat++;
8613 	sfmmu_hat_exit(hatlockp);
8614 	return (1);
8615 }
8616 
8617 void
8618 hat_stats_disable(struct hat *hat)
8619 {
8620 	hatlock_t	*hatlockp;
8621 
8622 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8623 
8624 	hatlockp = sfmmu_hat_enter(hat);
8625 	hat->sfmmu_rmstat--;
8626 	sfmmu_hat_exit(hatlockp);
8627 }
8628 
8629 /*
8630  * Routines for entering or removing  ourselves from the
8631  * ism_hat's mapping list. This is used for both private and
8632  * SCD hats.
8633  */
8634 static void
8635 iment_add(struct ism_ment *iment,  struct hat *ism_hat)
8636 {
8637 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8638 
8639 	iment->iment_prev = NULL;
8640 	iment->iment_next = ism_hat->sfmmu_iment;
8641 	if (ism_hat->sfmmu_iment) {
8642 		ism_hat->sfmmu_iment->iment_prev = iment;
8643 	}
8644 	ism_hat->sfmmu_iment = iment;
8645 }
8646 
8647 static void
8648 iment_sub(struct ism_ment *iment, struct hat *ism_hat)
8649 {
8650 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8651 
8652 	if (ism_hat->sfmmu_iment == NULL) {
8653 		panic("ism map entry remove - no entries");
8654 	}
8655 
8656 	if (iment->iment_prev) {
8657 		ASSERT(ism_hat->sfmmu_iment != iment);
8658 		iment->iment_prev->iment_next = iment->iment_next;
8659 	} else {
8660 		ASSERT(ism_hat->sfmmu_iment == iment);
8661 		ism_hat->sfmmu_iment = iment->iment_next;
8662 	}
8663 
8664 	if (iment->iment_next) {
8665 		iment->iment_next->iment_prev = iment->iment_prev;
8666 	}
8667 
8668 	/*
8669 	 * zero out the entry
8670 	 */
8671 	iment->iment_next = NULL;
8672 	iment->iment_prev = NULL;
8673 	iment->iment_hat =  NULL;
8674 	iment->iment_base_va = 0;
8675 }
8676 
8677 /*
8678  * Hat_share()/unshare() return an (non-zero) error
8679  * when saddr and daddr are not properly aligned.
8680  *
8681  * The top level mapping element determines the alignment
8682  * requirement for saddr and daddr, depending on different
8683  * architectures.
8684  *
8685  * When hat_share()/unshare() are not supported,
8686  * HATOP_SHARE()/UNSHARE() return 0
8687  */
8688 int
8689 hat_share(struct hat *sfmmup, caddr_t addr,
8690 	struct hat *ism_hatid, caddr_t sptaddr, size_t len, uint_t ismszc)
8691 {
8692 	ism_blk_t	*ism_blkp;
8693 	ism_blk_t	*new_iblk;
8694 	ism_map_t 	*ism_map;
8695 	ism_ment_t	*ism_ment;
8696 	int		i, added;
8697 	hatlock_t	*hatlockp;
8698 	int		reload_mmu = 0;
8699 	uint_t		ismshift = page_get_shift(ismszc);
8700 	size_t		ismpgsz = page_get_pagesize(ismszc);
8701 	uint_t		ismmask = (uint_t)ismpgsz - 1;
8702 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8703 	ushort_t	ismhatflag;
8704 	hat_region_cookie_t rcookie;
8705 	sf_scd_t	*old_scdp;
8706 
8707 #ifdef DEBUG
8708 	caddr_t		eaddr = addr + len;
8709 #endif /* DEBUG */
8710 
8711 	ASSERT(ism_hatid != NULL && sfmmup != NULL);
8712 	ASSERT(sptaddr == ISMID_STARTADDR);
8713 	/*
8714 	 * Check the alignment.
8715 	 */
8716 	if (!ISM_ALIGNED(ismshift, addr) || !ISM_ALIGNED(ismshift, sptaddr))
8717 		return (EINVAL);
8718 
8719 	/*
8720 	 * Check size alignment.
8721 	 */
8722 	if (!ISM_ALIGNED(ismshift, len))
8723 		return (EINVAL);
8724 
8725 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
8726 
8727 	/*
8728 	 * Allocate ism_ment for the ism_hat's mapping list, and an
8729 	 * ism map blk in case we need one.  We must do our
8730 	 * allocations before acquiring locks to prevent a deadlock
8731 	 * in the kmem allocator on the mapping list lock.
8732 	 */
8733 	new_iblk = kmem_cache_alloc(ism_blk_cache, KM_SLEEP);
8734 	ism_ment = kmem_cache_alloc(ism_ment_cache, KM_SLEEP);
8735 
8736 	/*
8737 	 * Serialize ISM mappings with the ISM busy flag, and also the
8738 	 * trap handlers.
8739 	 */
8740 	sfmmu_ismhat_enter(sfmmup, 0);
8741 
8742 	/*
8743 	 * Allocate an ism map blk if necessary.
8744 	 */
8745 	if (sfmmup->sfmmu_iblk == NULL) {
8746 		sfmmup->sfmmu_iblk = new_iblk;
8747 		bzero(new_iblk, sizeof (*new_iblk));
8748 		new_iblk->iblk_nextpa = (uint64_t)-1;
8749 		membar_stst();	/* make sure next ptr visible to all CPUs */
8750 		sfmmup->sfmmu_ismblkpa = va_to_pa((caddr_t)new_iblk);
8751 		reload_mmu = 1;
8752 		new_iblk = NULL;
8753 	}
8754 
8755 #ifdef DEBUG
8756 	/*
8757 	 * Make sure mapping does not already exist.
8758 	 */
8759 	ism_blkp = sfmmup->sfmmu_iblk;
8760 	while (ism_blkp != NULL) {
8761 		ism_map = ism_blkp->iblk_maps;
8762 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
8763 			if ((addr >= ism_start(ism_map[i]) &&
8764 			    addr < ism_end(ism_map[i])) ||
8765 			    eaddr > ism_start(ism_map[i]) &&
8766 			    eaddr <= ism_end(ism_map[i])) {
8767 				panic("sfmmu_share: Already mapped!");
8768 			}
8769 		}
8770 		ism_blkp = ism_blkp->iblk_next;
8771 	}
8772 #endif /* DEBUG */
8773 
8774 	ASSERT(ismszc >= TTE4M);
8775 	if (ismszc == TTE4M) {
8776 		ismhatflag = HAT_4M_FLAG;
8777 	} else if (ismszc == TTE32M) {
8778 		ismhatflag = HAT_32M_FLAG;
8779 	} else if (ismszc == TTE256M) {
8780 		ismhatflag = HAT_256M_FLAG;
8781 	}
8782 	/*
8783 	 * Add mapping to first available mapping slot.
8784 	 */
8785 	ism_blkp = sfmmup->sfmmu_iblk;
8786 	added = 0;
8787 	while (!added) {
8788 		ism_map = ism_blkp->iblk_maps;
8789 		for (i = 0; i < ISM_MAP_SLOTS; i++)  {
8790 			if (ism_map[i].imap_ismhat == NULL) {
8791 
8792 				ism_map[i].imap_ismhat = ism_hatid;
8793 				ism_map[i].imap_vb_shift = (uchar_t)ismshift;
8794 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8795 				ism_map[i].imap_hatflags = ismhatflag;
8796 				ism_map[i].imap_sz_mask = ismmask;
8797 				/*
8798 				 * imap_seg is checked in ISM_CHECK to see if
8799 				 * non-NULL, then other info assumed valid.
8800 				 */
8801 				membar_stst();
8802 				ism_map[i].imap_seg = (uintptr_t)addr | sh_size;
8803 				ism_map[i].imap_ment = ism_ment;
8804 
8805 				/*
8806 				 * Now add ourselves to the ism_hat's
8807 				 * mapping list.
8808 				 */
8809 				ism_ment->iment_hat = sfmmup;
8810 				ism_ment->iment_base_va = addr;
8811 				ism_hatid->sfmmu_ismhat = 1;
8812 				mutex_enter(&ism_mlist_lock);
8813 				iment_add(ism_ment, ism_hatid);
8814 				mutex_exit(&ism_mlist_lock);
8815 				added = 1;
8816 				break;
8817 			}
8818 		}
8819 		if (!added && ism_blkp->iblk_next == NULL) {
8820 			ism_blkp->iblk_next = new_iblk;
8821 			new_iblk = NULL;
8822 			bzero(ism_blkp->iblk_next,
8823 			    sizeof (*ism_blkp->iblk_next));
8824 			ism_blkp->iblk_next->iblk_nextpa = (uint64_t)-1;
8825 			membar_stst();
8826 			ism_blkp->iblk_nextpa =
8827 			    va_to_pa((caddr_t)ism_blkp->iblk_next);
8828 		}
8829 		ism_blkp = ism_blkp->iblk_next;
8830 	}
8831 
8832 	/*
8833 	 * After calling hat_join_region, sfmmup may join a new SCD or
8834 	 * move from the old scd to a new scd, in which case, we want to
8835 	 * shrink the sfmmup's private tsb size, i.e., pass shrink to
8836 	 * sfmmu_check_page_sizes at the end of this routine.
8837 	 */
8838 	old_scdp = sfmmup->sfmmu_scdp;
8839 
8840 	rcookie = hat_join_region(sfmmup, addr, len, (void *)ism_hatid, 0,
8841 	    PROT_ALL, ismszc, NULL, HAT_REGION_ISM);
8842 	if (rcookie != HAT_INVALID_REGION_COOKIE) {
8843 		ism_map[i].imap_rid = (uchar_t)((uint64_t)rcookie);
8844 	}
8845 	/*
8846 	 * Update our counters for this sfmmup's ism mappings.
8847 	 */
8848 	for (i = 0; i <= ismszc; i++) {
8849 		if (!(disable_ism_large_pages & (1 << i)))
8850 			(void) ism_tsb_entries(sfmmup, i);
8851 	}
8852 
8853 	/*
8854 	 * For ISM and DISM we do not support 512K pages, so we only only
8855 	 * search the 4M and 8K/64K hashes for 4 pagesize cpus, and search the
8856 	 * 256M or 32M, and 4M and 8K/64K hashes for 6 pagesize cpus.
8857 	 *
8858 	 * Need to set 32M/256M ISM flags to make sure
8859 	 * sfmmu_check_page_sizes() enables them on Panther.
8860 	 */
8861 	ASSERT((disable_ism_large_pages & (1 << TTE512K)) != 0);
8862 
8863 	switch (ismszc) {
8864 	case TTE256M:
8865 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_256M_ISM)) {
8866 			hatlockp = sfmmu_hat_enter(sfmmup);
8867 			SFMMU_FLAGS_SET(sfmmup, HAT_256M_ISM);
8868 			sfmmu_hat_exit(hatlockp);
8869 		}
8870 		break;
8871 	case TTE32M:
8872 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_32M_ISM)) {
8873 			hatlockp = sfmmu_hat_enter(sfmmup);
8874 			SFMMU_FLAGS_SET(sfmmup, HAT_32M_ISM);
8875 			sfmmu_hat_exit(hatlockp);
8876 		}
8877 		break;
8878 	default:
8879 		break;
8880 	}
8881 
8882 	/*
8883 	 * If we updated the ismblkpa for this HAT we must make
8884 	 * sure all CPUs running this process reload their tsbmiss area.
8885 	 * Otherwise they will fail to load the mappings in the tsbmiss
8886 	 * handler and will loop calling pagefault().
8887 	 */
8888 	if (reload_mmu) {
8889 		hatlockp = sfmmu_hat_enter(sfmmup);
8890 		sfmmu_sync_mmustate(sfmmup);
8891 		sfmmu_hat_exit(hatlockp);
8892 	}
8893 
8894 	sfmmu_ismhat_exit(sfmmup, 0);
8895 
8896 	/*
8897 	 * Free up ismblk if we didn't use it.
8898 	 */
8899 	if (new_iblk != NULL)
8900 		kmem_cache_free(ism_blk_cache, new_iblk);
8901 
8902 	/*
8903 	 * Check TSB and TLB page sizes.
8904 	 */
8905 	if (sfmmup->sfmmu_scdp != NULL && old_scdp != sfmmup->sfmmu_scdp) {
8906 		sfmmu_check_page_sizes(sfmmup, 0);
8907 	} else {
8908 		sfmmu_check_page_sizes(sfmmup, 1);
8909 	}
8910 	return (0);
8911 }
8912 
8913 /*
8914  * hat_unshare removes exactly one ism_map from
8915  * this process's as.  It expects multiple calls
8916  * to hat_unshare for multiple shm segments.
8917  */
8918 void
8919 hat_unshare(struct hat *sfmmup, caddr_t addr, size_t len, uint_t ismszc)
8920 {
8921 	ism_map_t 	*ism_map;
8922 	ism_ment_t	*free_ment = NULL;
8923 	ism_blk_t	*ism_blkp;
8924 	struct hat	*ism_hatid;
8925 	int 		found, i;
8926 	hatlock_t	*hatlockp;
8927 	struct tsb_info	*tsbinfo;
8928 	uint_t		ismshift = page_get_shift(ismszc);
8929 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8930 	uchar_t		ism_rid;
8931 	sf_scd_t	*old_scdp;
8932 
8933 	ASSERT(ISM_ALIGNED(ismshift, addr));
8934 	ASSERT(ISM_ALIGNED(ismshift, len));
8935 	ASSERT(sfmmup != NULL);
8936 	ASSERT(sfmmup != ksfmmup);
8937 
8938 	if (sfmmup->sfmmu_xhat_provider) {
8939 		XHAT_UNSHARE(sfmmup, addr, len);
8940 		return;
8941 	} else {
8942 		/*
8943 		 * This must be a CPU HAT. If the address space has
8944 		 * XHATs attached, inform all XHATs that ISM segment
8945 		 * is going away
8946 		 */
8947 		ASSERT(sfmmup->sfmmu_as != NULL);
8948 		if (sfmmup->sfmmu_as->a_xhat != NULL)
8949 			xhat_unshare_all(sfmmup->sfmmu_as, addr, len);
8950 	}
8951 
8952 	/*
8953 	 * Make sure that during the entire time ISM mappings are removed,
8954 	 * the trap handlers serialize behind us, and that no one else
8955 	 * can be mucking with ISM mappings.  This also lets us get away
8956 	 * with not doing expensive cross calls to flush the TLB -- we
8957 	 * just discard the context, flush the entire TSB, and call it
8958 	 * a day.
8959 	 */
8960 	sfmmu_ismhat_enter(sfmmup, 0);
8961 
8962 	/*
8963 	 * Remove the mapping.
8964 	 *
8965 	 * We can't have any holes in the ism map.
8966 	 * The tsb miss code while searching the ism map will
8967 	 * stop on an empty map slot.  So we must move
8968 	 * everyone past the hole up 1 if any.
8969 	 *
8970 	 * Also empty ism map blks are not freed until the
8971 	 * process exits. This is to prevent a MT race condition
8972 	 * between sfmmu_unshare() and sfmmu_tsbmiss_exception().
8973 	 */
8974 	found = 0;
8975 	ism_blkp = sfmmup->sfmmu_iblk;
8976 	while (!found && ism_blkp != NULL) {
8977 		ism_map = ism_blkp->iblk_maps;
8978 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
8979 			if (addr == ism_start(ism_map[i]) &&
8980 			    sh_size == (size_t)(ism_size(ism_map[i]))) {
8981 				found = 1;
8982 				break;
8983 			}
8984 		}
8985 		if (!found)
8986 			ism_blkp = ism_blkp->iblk_next;
8987 	}
8988 
8989 	if (found) {
8990 		ism_hatid = ism_map[i].imap_ismhat;
8991 		ism_rid = ism_map[i].imap_rid;
8992 		ASSERT(ism_hatid != NULL);
8993 		ASSERT(ism_hatid->sfmmu_ismhat == 1);
8994 
8995 		/*
8996 		 * After hat_leave_region, the sfmmup may leave SCD,
8997 		 * in which case, we want to grow the private tsb size when
8998 		 * calling sfmmu_check_page_sizes at the end of the routine.
8999 		 */
9000 		old_scdp = sfmmup->sfmmu_scdp;
9001 		/*
9002 		 * Then remove ourselves from the region.
9003 		 */
9004 		if (ism_rid != SFMMU_INVALID_ISMRID) {
9005 			hat_leave_region(sfmmup, (void *)((uint64_t)ism_rid),
9006 			    HAT_REGION_ISM);
9007 		}
9008 
9009 		/*
9010 		 * And now guarantee that any other cpu
9011 		 * that tries to process an ISM miss
9012 		 * will go to tl=0.
9013 		 */
9014 		hatlockp = sfmmu_hat_enter(sfmmup);
9015 		sfmmu_invalidate_ctx(sfmmup);
9016 		sfmmu_hat_exit(hatlockp);
9017 
9018 		/*
9019 		 * Remove ourselves from the ism mapping list.
9020 		 */
9021 		mutex_enter(&ism_mlist_lock);
9022 		iment_sub(ism_map[i].imap_ment, ism_hatid);
9023 		mutex_exit(&ism_mlist_lock);
9024 		free_ment = ism_map[i].imap_ment;
9025 
9026 		/*
9027 		 * We delete the ism map by copying
9028 		 * the next map over the current one.
9029 		 * We will take the next one in the maps
9030 		 * array or from the next ism_blk.
9031 		 */
9032 		while (ism_blkp != NULL) {
9033 			ism_map = ism_blkp->iblk_maps;
9034 			while (i < (ISM_MAP_SLOTS - 1)) {
9035 				ism_map[i] = ism_map[i + 1];
9036 				i++;
9037 			}
9038 			/* i == (ISM_MAP_SLOTS - 1) */
9039 			ism_blkp = ism_blkp->iblk_next;
9040 			if (ism_blkp != NULL) {
9041 				ism_map[i] = ism_blkp->iblk_maps[0];
9042 				i = 0;
9043 			} else {
9044 				ism_map[i].imap_seg = 0;
9045 				ism_map[i].imap_vb_shift = 0;
9046 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
9047 				ism_map[i].imap_hatflags = 0;
9048 				ism_map[i].imap_sz_mask = 0;
9049 				ism_map[i].imap_ismhat = NULL;
9050 				ism_map[i].imap_ment = NULL;
9051 			}
9052 		}
9053 
9054 		/*
9055 		 * Now flush entire TSB for the process, since
9056 		 * demapping page by page can be too expensive.
9057 		 * We don't have to flush the TLB here anymore
9058 		 * since we switch to a new TLB ctx instead.
9059 		 * Also, there is no need to flush if the process
9060 		 * is exiting since the TSB will be freed later.
9061 		 */
9062 		if (!sfmmup->sfmmu_free) {
9063 			hatlockp = sfmmu_hat_enter(sfmmup);
9064 			for (tsbinfo = sfmmup->sfmmu_tsb; tsbinfo != NULL;
9065 			    tsbinfo = tsbinfo->tsb_next) {
9066 				if (tsbinfo->tsb_flags & TSB_SWAPPED)
9067 					continue;
9068 				if (tsbinfo->tsb_flags & TSB_RELOC_FLAG) {
9069 					tsbinfo->tsb_flags |=
9070 					    TSB_FLUSH_NEEDED;
9071 					continue;
9072 				}
9073 
9074 				sfmmu_inv_tsb(tsbinfo->tsb_va,
9075 				    TSB_BYTES(tsbinfo->tsb_szc));
9076 			}
9077 			sfmmu_hat_exit(hatlockp);
9078 		}
9079 	}
9080 
9081 	/*
9082 	 * Update our counters for this sfmmup's ism mappings.
9083 	 */
9084 	for (i = 0; i <= ismszc; i++) {
9085 		if (!(disable_ism_large_pages & (1 << i)))
9086 			(void) ism_tsb_entries(sfmmup, i);
9087 	}
9088 
9089 	sfmmu_ismhat_exit(sfmmup, 0);
9090 
9091 	/*
9092 	 * We must do our freeing here after dropping locks
9093 	 * to prevent a deadlock in the kmem allocator on the
9094 	 * mapping list lock.
9095 	 */
9096 	if (free_ment != NULL)
9097 		kmem_cache_free(ism_ment_cache, free_ment);
9098 
9099 	/*
9100 	 * Check TSB and TLB page sizes if the process isn't exiting.
9101 	 */
9102 	if (!sfmmup->sfmmu_free) {
9103 		if (found && old_scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
9104 			sfmmu_check_page_sizes(sfmmup, 1);
9105 		} else {
9106 			sfmmu_check_page_sizes(sfmmup, 0);
9107 		}
9108 	}
9109 }
9110 
9111 /* ARGSUSED */
9112 static int
9113 sfmmu_idcache_constructor(void *buf, void *cdrarg, int kmflags)
9114 {
9115 	/* void *buf is sfmmu_t pointer */
9116 	bzero(buf, sizeof (sfmmu_t));
9117 
9118 	return (0);
9119 }
9120 
9121 /* ARGSUSED */
9122 static void
9123 sfmmu_idcache_destructor(void *buf, void *cdrarg)
9124 {
9125 	/* void *buf is sfmmu_t pointer */
9126 }
9127 
9128 /*
9129  * setup kmem hmeblks by bzeroing all members and initializing the nextpa
9130  * field to be the pa of this hmeblk
9131  */
9132 /* ARGSUSED */
9133 static int
9134 sfmmu_hblkcache_constructor(void *buf, void *cdrarg, int kmflags)
9135 {
9136 	struct hme_blk *hmeblkp;
9137 
9138 	bzero(buf, (size_t)cdrarg);
9139 	hmeblkp = (struct hme_blk *)buf;
9140 	hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
9141 
9142 #ifdef	HBLK_TRACE
9143 	mutex_init(&hmeblkp->hblk_audit_lock, NULL, MUTEX_DEFAULT, NULL);
9144 #endif	/* HBLK_TRACE */
9145 
9146 	return (0);
9147 }
9148 
9149 /* ARGSUSED */
9150 static void
9151 sfmmu_hblkcache_destructor(void *buf, void *cdrarg)
9152 {
9153 
9154 #ifdef	HBLK_TRACE
9155 
9156 	struct hme_blk *hmeblkp;
9157 
9158 	hmeblkp = (struct hme_blk *)buf;
9159 	mutex_destroy(&hmeblkp->hblk_audit_lock);
9160 
9161 #endif	/* HBLK_TRACE */
9162 }
9163 
9164 #define	SFMMU_CACHE_RECLAIM_SCAN_RATIO 8
9165 static int sfmmu_cache_reclaim_scan_ratio = SFMMU_CACHE_RECLAIM_SCAN_RATIO;
9166 /*
9167  * The kmem allocator will callback into our reclaim routine when the system
9168  * is running low in memory.  We traverse the hash and free up all unused but
9169  * still cached hme_blks.  We also traverse the free list and free them up
9170  * as well.
9171  */
9172 /*ARGSUSED*/
9173 static void
9174 sfmmu_hblkcache_reclaim(void *cdrarg)
9175 {
9176 	int i;
9177 	struct hmehash_bucket *hmebp;
9178 	struct hme_blk *hmeblkp, *nx_hblk, *pr_hblk = NULL;
9179 	static struct hmehash_bucket *uhmehash_reclaim_hand;
9180 	static struct hmehash_bucket *khmehash_reclaim_hand;
9181 	struct hme_blk *list = NULL, *last_hmeblkp;
9182 	cpuset_t cpuset = cpu_ready_set;
9183 	cpu_hme_pend_t *cpuhp;
9184 
9185 	/* Free up hmeblks on the cpu pending lists */
9186 	for (i = 0; i < NCPU; i++) {
9187 		cpuhp = &cpu_hme_pend[i];
9188 		if (cpuhp->chp_listp != NULL)  {
9189 			mutex_enter(&cpuhp->chp_mutex);
9190 			if (cpuhp->chp_listp == NULL) {
9191 				mutex_exit(&cpuhp->chp_mutex);
9192 				continue;
9193 			}
9194 			for (last_hmeblkp = cpuhp->chp_listp;
9195 			    last_hmeblkp->hblk_next != NULL;
9196 			    last_hmeblkp = last_hmeblkp->hblk_next)
9197 				;
9198 			last_hmeblkp->hblk_next = list;
9199 			list = cpuhp->chp_listp;
9200 			cpuhp->chp_listp = NULL;
9201 			cpuhp->chp_count = 0;
9202 			mutex_exit(&cpuhp->chp_mutex);
9203 		}
9204 
9205 	}
9206 
9207 	if (list != NULL) {
9208 		kpreempt_disable();
9209 		CPUSET_DEL(cpuset, CPU->cpu_id);
9210 		xt_sync(cpuset);
9211 		xt_sync(cpuset);
9212 		kpreempt_enable();
9213 		sfmmu_hblk_free(&list);
9214 		list = NULL;
9215 	}
9216 
9217 	hmebp = uhmehash_reclaim_hand;
9218 	if (hmebp == NULL || hmebp > &uhme_hash[UHMEHASH_SZ])
9219 		uhmehash_reclaim_hand = hmebp = uhme_hash;
9220 	uhmehash_reclaim_hand += UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
9221 
9222 	for (i = UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
9223 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
9224 			hmeblkp = hmebp->hmeblkp;
9225 			pr_hblk = NULL;
9226 			while (hmeblkp) {
9227 				nx_hblk = hmeblkp->hblk_next;
9228 				if (!hmeblkp->hblk_vcnt &&
9229 				    !hmeblkp->hblk_hmecnt) {
9230 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9231 					    pr_hblk, &list, 0);
9232 				} else {
9233 					pr_hblk = hmeblkp;
9234 				}
9235 				hmeblkp = nx_hblk;
9236 			}
9237 			SFMMU_HASH_UNLOCK(hmebp);
9238 		}
9239 		if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
9240 			hmebp = uhme_hash;
9241 	}
9242 
9243 	hmebp = khmehash_reclaim_hand;
9244 	if (hmebp == NULL || hmebp > &khme_hash[KHMEHASH_SZ])
9245 		khmehash_reclaim_hand = hmebp = khme_hash;
9246 	khmehash_reclaim_hand += KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
9247 
9248 	for (i = KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
9249 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
9250 			hmeblkp = hmebp->hmeblkp;
9251 			pr_hblk = NULL;
9252 			while (hmeblkp) {
9253 				nx_hblk = hmeblkp->hblk_next;
9254 				if (!hmeblkp->hblk_vcnt &&
9255 				    !hmeblkp->hblk_hmecnt) {
9256 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9257 					    pr_hblk, &list, 0);
9258 				} else {
9259 					pr_hblk = hmeblkp;
9260 				}
9261 				hmeblkp = nx_hblk;
9262 			}
9263 			SFMMU_HASH_UNLOCK(hmebp);
9264 		}
9265 		if (hmebp++ == &khme_hash[KHMEHASH_SZ])
9266 			hmebp = khme_hash;
9267 	}
9268 	sfmmu_hblks_list_purge(&list, 0);
9269 }
9270 
9271 /*
9272  * sfmmu_get_ppvcolor should become a vm_machdep or hatop interface.
9273  * same goes for sfmmu_get_addrvcolor().
9274  *
9275  * This function will return the virtual color for the specified page. The
9276  * virtual color corresponds to this page current mapping or its last mapping.
9277  * It is used by memory allocators to choose addresses with the correct
9278  * alignment so vac consistency is automatically maintained.  If the page
9279  * has no color it returns -1.
9280  */
9281 /*ARGSUSED*/
9282 int
9283 sfmmu_get_ppvcolor(struct page *pp)
9284 {
9285 #ifdef VAC
9286 	int color;
9287 
9288 	if (!(cache & CACHE_VAC) || PP_NEWPAGE(pp)) {
9289 		return (-1);
9290 	}
9291 	color = PP_GET_VCOLOR(pp);
9292 	ASSERT(color < mmu_btop(shm_alignment));
9293 	return (color);
9294 #else
9295 	return (-1);
9296 #endif	/* VAC */
9297 }
9298 
9299 /*
9300  * This function will return the desired alignment for vac consistency
9301  * (vac color) given a virtual address.  If no vac is present it returns -1.
9302  */
9303 /*ARGSUSED*/
9304 int
9305 sfmmu_get_addrvcolor(caddr_t vaddr)
9306 {
9307 #ifdef VAC
9308 	if (cache & CACHE_VAC) {
9309 		return (addr_to_vcolor(vaddr));
9310 	} else {
9311 		return (-1);
9312 	}
9313 #else
9314 	return (-1);
9315 #endif	/* VAC */
9316 }
9317 
9318 #ifdef VAC
9319 /*
9320  * Check for conflicts.
9321  * A conflict exists if the new and existent mappings do not match in
9322  * their "shm_alignment fields. If conflicts exist, the existant mappings
9323  * are flushed unless one of them is locked. If one of them is locked, then
9324  * the mappings are flushed and converted to non-cacheable mappings.
9325  */
9326 static void
9327 sfmmu_vac_conflict(struct hat *hat, caddr_t addr, page_t *pp)
9328 {
9329 	struct hat *tmphat;
9330 	struct sf_hment *sfhmep, *tmphme = NULL;
9331 	struct hme_blk *hmeblkp;
9332 	int vcolor;
9333 	tte_t tte;
9334 
9335 	ASSERT(sfmmu_mlist_held(pp));
9336 	ASSERT(!PP_ISNC(pp));		/* page better be cacheable */
9337 
9338 	vcolor = addr_to_vcolor(addr);
9339 	if (PP_NEWPAGE(pp)) {
9340 		PP_SET_VCOLOR(pp, vcolor);
9341 		return;
9342 	}
9343 
9344 	if (PP_GET_VCOLOR(pp) == vcolor) {
9345 		return;
9346 	}
9347 
9348 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
9349 		/*
9350 		 * Previous user of page had a different color
9351 		 * but since there are no current users
9352 		 * we just flush the cache and change the color.
9353 		 */
9354 		SFMMU_STAT(sf_pgcolor_conflict);
9355 		sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9356 		PP_SET_VCOLOR(pp, vcolor);
9357 		return;
9358 	}
9359 
9360 	/*
9361 	 * If we get here we have a vac conflict with a current
9362 	 * mapping.  VAC conflict policy is as follows.
9363 	 * - The default is to unload the other mappings unless:
9364 	 * - If we have a large mapping we uncache the page.
9365 	 * We need to uncache the rest of the large page too.
9366 	 * - If any of the mappings are locked we uncache the page.
9367 	 * - If the requested mapping is inconsistent
9368 	 * with another mapping and that mapping
9369 	 * is in the same address space we have to
9370 	 * make it non-cached.  The default thing
9371 	 * to do is unload the inconsistent mapping
9372 	 * but if they are in the same address space
9373 	 * we run the risk of unmapping the pc or the
9374 	 * stack which we will use as we return to the user,
9375 	 * in which case we can then fault on the thing
9376 	 * we just unloaded and get into an infinite loop.
9377 	 */
9378 	if (PP_ISMAPPED_LARGE(pp)) {
9379 		int sz;
9380 
9381 		/*
9382 		 * Existing mapping is for big pages. We don't unload
9383 		 * existing big mappings to satisfy new mappings.
9384 		 * Always convert all mappings to TNC.
9385 		 */
9386 		sz = fnd_mapping_sz(pp);
9387 		pp = PP_GROUPLEADER(pp, sz);
9388 		SFMMU_STAT_ADD(sf_uncache_conflict, TTEPAGES(sz));
9389 		sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH,
9390 		    TTEPAGES(sz));
9391 
9392 		return;
9393 	}
9394 
9395 	/*
9396 	 * check if any mapping is in same as or if it is locked
9397 	 * since in that case we need to uncache.
9398 	 */
9399 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9400 		tmphme = sfhmep->hme_next;
9401 		if (IS_PAHME(sfhmep))
9402 			continue;
9403 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9404 		if (hmeblkp->hblk_xhat_bit)
9405 			continue;
9406 		tmphat = hblktosfmmu(hmeblkp);
9407 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
9408 		ASSERT(TTE_IS_VALID(&tte));
9409 		if (hmeblkp->hblk_shared || tmphat == hat ||
9410 		    hmeblkp->hblk_lckcnt) {
9411 			/*
9412 			 * We have an uncache conflict
9413 			 */
9414 			SFMMU_STAT(sf_uncache_conflict);
9415 			sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH, 1);
9416 			return;
9417 		}
9418 	}
9419 
9420 	/*
9421 	 * We have an unload conflict
9422 	 * We have already checked for LARGE mappings, therefore
9423 	 * the remaining mapping(s) must be TTE8K.
9424 	 */
9425 	SFMMU_STAT(sf_unload_conflict);
9426 
9427 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9428 		tmphme = sfhmep->hme_next;
9429 		if (IS_PAHME(sfhmep))
9430 			continue;
9431 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9432 		if (hmeblkp->hblk_xhat_bit)
9433 			continue;
9434 		ASSERT(!hmeblkp->hblk_shared);
9435 		(void) sfmmu_pageunload(pp, sfhmep, TTE8K);
9436 	}
9437 
9438 	if (PP_ISMAPPED_KPM(pp))
9439 		sfmmu_kpm_vac_unload(pp, addr);
9440 
9441 	/*
9442 	 * Unloads only do TLB flushes so we need to flush the
9443 	 * cache here.
9444 	 */
9445 	sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9446 	PP_SET_VCOLOR(pp, vcolor);
9447 }
9448 
9449 /*
9450  * Whenever a mapping is unloaded and the page is in TNC state,
9451  * we see if the page can be made cacheable again. 'pp' is
9452  * the page that we just unloaded a mapping from, the size
9453  * of mapping that was unloaded is 'ottesz'.
9454  * Remark:
9455  * The recache policy for mpss pages can leave a performance problem
9456  * under the following circumstances:
9457  * . A large page in uncached mode has just been unmapped.
9458  * . All constituent pages are TNC due to a conflicting small mapping.
9459  * . There are many other, non conflicting, small mappings around for
9460  *   a lot of the constituent pages.
9461  * . We're called w/ the "old" groupleader page and the old ottesz,
9462  *   but this is irrelevant, since we're no more "PP_ISMAPPED_LARGE", so
9463  *   we end up w/ TTE8K or npages == 1.
9464  * . We call tst_tnc w/ the old groupleader only, and if there is no
9465  *   conflict, we re-cache only this page.
9466  * . All other small mappings are not checked and will be left in TNC mode.
9467  * The problem is not very serious because:
9468  * . mpss is actually only defined for heap and stack, so the probability
9469  *   is not very high that a large page mapping exists in parallel to a small
9470  *   one (this is possible, but seems to be bad programming style in the
9471  *   appl).
9472  * . The problem gets a little bit more serious, when those TNC pages
9473  *   have to be mapped into kernel space, e.g. for networking.
9474  * . When VAC alias conflicts occur in applications, this is regarded
9475  *   as an application bug. So if kstat's show them, the appl should
9476  *   be changed anyway.
9477  */
9478 void
9479 conv_tnc(page_t *pp, int ottesz)
9480 {
9481 	int cursz, dosz;
9482 	pgcnt_t curnpgs, dopgs;
9483 	pgcnt_t pg64k;
9484 	page_t *pp2;
9485 
9486 	/*
9487 	 * Determine how big a range we check for TNC and find
9488 	 * leader page. cursz is the size of the biggest
9489 	 * mapping that still exist on 'pp'.
9490 	 */
9491 	if (PP_ISMAPPED_LARGE(pp)) {
9492 		cursz = fnd_mapping_sz(pp);
9493 	} else {
9494 		cursz = TTE8K;
9495 	}
9496 
9497 	if (ottesz >= cursz) {
9498 		dosz = ottesz;
9499 		pp2 = pp;
9500 	} else {
9501 		dosz = cursz;
9502 		pp2 = PP_GROUPLEADER(pp, dosz);
9503 	}
9504 
9505 	pg64k = TTEPAGES(TTE64K);
9506 	dopgs = TTEPAGES(dosz);
9507 
9508 	ASSERT(dopgs == 1 || ((dopgs & (pg64k - 1)) == 0));
9509 
9510 	while (dopgs != 0) {
9511 		curnpgs = TTEPAGES(cursz);
9512 		if (tst_tnc(pp2, curnpgs)) {
9513 			SFMMU_STAT_ADD(sf_recache, curnpgs);
9514 			sfmmu_page_cache_array(pp2, HAT_CACHE, CACHE_NO_FLUSH,
9515 			    curnpgs);
9516 		}
9517 
9518 		ASSERT(dopgs >= curnpgs);
9519 		dopgs -= curnpgs;
9520 
9521 		if (dopgs == 0) {
9522 			break;
9523 		}
9524 
9525 		pp2 = PP_PAGENEXT_N(pp2, curnpgs);
9526 		if (((dopgs & (pg64k - 1)) == 0) && PP_ISMAPPED_LARGE(pp2)) {
9527 			cursz = fnd_mapping_sz(pp2);
9528 		} else {
9529 			cursz = TTE8K;
9530 		}
9531 	}
9532 }
9533 
9534 /*
9535  * Returns 1 if page(s) can be converted from TNC to cacheable setting,
9536  * returns 0 otherwise. Note that oaddr argument is valid for only
9537  * 8k pages.
9538  */
9539 int
9540 tst_tnc(page_t *pp, pgcnt_t npages)
9541 {
9542 	struct	sf_hment *sfhme;
9543 	struct	hme_blk *hmeblkp;
9544 	tte_t	tte;
9545 	caddr_t	vaddr;
9546 	int	clr_valid = 0;
9547 	int 	color, color1, bcolor;
9548 	int	i, ncolors;
9549 
9550 	ASSERT(pp != NULL);
9551 	ASSERT(!(cache & CACHE_WRITEBACK));
9552 
9553 	if (npages > 1) {
9554 		ncolors = CACHE_NUM_COLOR;
9555 	}
9556 
9557 	for (i = 0; i < npages; i++) {
9558 		ASSERT(sfmmu_mlist_held(pp));
9559 		ASSERT(PP_ISTNC(pp));
9560 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
9561 
9562 		if (PP_ISPNC(pp)) {
9563 			return (0);
9564 		}
9565 
9566 		clr_valid = 0;
9567 		if (PP_ISMAPPED_KPM(pp)) {
9568 			caddr_t kpmvaddr;
9569 
9570 			ASSERT(kpm_enable);
9571 			kpmvaddr = hat_kpm_page2va(pp, 1);
9572 			ASSERT(!(npages > 1 && IS_KPM_ALIAS_RANGE(kpmvaddr)));
9573 			color1 = addr_to_vcolor(kpmvaddr);
9574 			clr_valid = 1;
9575 		}
9576 
9577 		for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9578 			if (IS_PAHME(sfhme))
9579 				continue;
9580 			hmeblkp = sfmmu_hmetohblk(sfhme);
9581 			if (hmeblkp->hblk_xhat_bit)
9582 				continue;
9583 
9584 			sfmmu_copytte(&sfhme->hme_tte, &tte);
9585 			ASSERT(TTE_IS_VALID(&tte));
9586 
9587 			vaddr = tte_to_vaddr(hmeblkp, tte);
9588 			color = addr_to_vcolor(vaddr);
9589 
9590 			if (npages > 1) {
9591 				/*
9592 				 * If there is a big mapping, make sure
9593 				 * 8K mapping is consistent with the big
9594 				 * mapping.
9595 				 */
9596 				bcolor = i % ncolors;
9597 				if (color != bcolor) {
9598 					return (0);
9599 				}
9600 			}
9601 			if (!clr_valid) {
9602 				clr_valid = 1;
9603 				color1 = color;
9604 			}
9605 
9606 			if (color1 != color) {
9607 				return (0);
9608 			}
9609 		}
9610 
9611 		pp = PP_PAGENEXT(pp);
9612 	}
9613 
9614 	return (1);
9615 }
9616 
9617 void
9618 sfmmu_page_cache_array(page_t *pp, int flags, int cache_flush_flag,
9619 	pgcnt_t npages)
9620 {
9621 	kmutex_t *pmtx;
9622 	int i, ncolors, bcolor;
9623 	kpm_hlk_t *kpmp;
9624 	cpuset_t cpuset;
9625 
9626 	ASSERT(pp != NULL);
9627 	ASSERT(!(cache & CACHE_WRITEBACK));
9628 
9629 	kpmp = sfmmu_kpm_kpmp_enter(pp, npages);
9630 	pmtx = sfmmu_page_enter(pp);
9631 
9632 	/*
9633 	 * Fast path caching single unmapped page
9634 	 */
9635 	if (npages == 1 && !PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp) &&
9636 	    flags == HAT_CACHE) {
9637 		PP_CLRTNC(pp);
9638 		PP_CLRPNC(pp);
9639 		sfmmu_page_exit(pmtx);
9640 		sfmmu_kpm_kpmp_exit(kpmp);
9641 		return;
9642 	}
9643 
9644 	/*
9645 	 * We need to capture all cpus in order to change cacheability
9646 	 * because we can't allow one cpu to access the same physical
9647 	 * page using a cacheable and a non-cachebale mapping at the same
9648 	 * time. Since we may end up walking the ism mapping list
9649 	 * have to grab it's lock now since we can't after all the
9650 	 * cpus have been captured.
9651 	 */
9652 	sfmmu_hat_lock_all();
9653 	mutex_enter(&ism_mlist_lock);
9654 	kpreempt_disable();
9655 	cpuset = cpu_ready_set;
9656 	xc_attention(cpuset);
9657 
9658 	if (npages > 1) {
9659 		/*
9660 		 * Make sure all colors are flushed since the
9661 		 * sfmmu_page_cache() only flushes one color-
9662 		 * it does not know big pages.
9663 		 */
9664 		ncolors = CACHE_NUM_COLOR;
9665 		if (flags & HAT_TMPNC) {
9666 			for (i = 0; i < ncolors; i++) {
9667 				sfmmu_cache_flushcolor(i, pp->p_pagenum);
9668 			}
9669 			cache_flush_flag = CACHE_NO_FLUSH;
9670 		}
9671 	}
9672 
9673 	for (i = 0; i < npages; i++) {
9674 
9675 		ASSERT(sfmmu_mlist_held(pp));
9676 
9677 		if (!(flags == HAT_TMPNC && PP_ISTNC(pp))) {
9678 
9679 			if (npages > 1) {
9680 				bcolor = i % ncolors;
9681 			} else {
9682 				bcolor = NO_VCOLOR;
9683 			}
9684 
9685 			sfmmu_page_cache(pp, flags, cache_flush_flag,
9686 			    bcolor);
9687 		}
9688 
9689 		pp = PP_PAGENEXT(pp);
9690 	}
9691 
9692 	xt_sync(cpuset);
9693 	xc_dismissed(cpuset);
9694 	mutex_exit(&ism_mlist_lock);
9695 	sfmmu_hat_unlock_all();
9696 	sfmmu_page_exit(pmtx);
9697 	sfmmu_kpm_kpmp_exit(kpmp);
9698 	kpreempt_enable();
9699 }
9700 
9701 /*
9702  * This function changes the virtual cacheability of all mappings to a
9703  * particular page.  When changing from uncache to cacheable the mappings will
9704  * only be changed if all of them have the same virtual color.
9705  * We need to flush the cache in all cpus.  It is possible that
9706  * a process referenced a page as cacheable but has sinced exited
9707  * and cleared the mapping list.  We still to flush it but have no
9708  * state so all cpus is the only alternative.
9709  */
9710 static void
9711 sfmmu_page_cache(page_t *pp, int flags, int cache_flush_flag, int bcolor)
9712 {
9713 	struct	sf_hment *sfhme;
9714 	struct	hme_blk *hmeblkp;
9715 	sfmmu_t *sfmmup;
9716 	tte_t	tte, ttemod;
9717 	caddr_t	vaddr;
9718 	int	ret, color;
9719 	pfn_t	pfn;
9720 
9721 	color = bcolor;
9722 	pfn = pp->p_pagenum;
9723 
9724 	for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9725 
9726 		if (IS_PAHME(sfhme))
9727 			continue;
9728 		hmeblkp = sfmmu_hmetohblk(sfhme);
9729 
9730 		if (hmeblkp->hblk_xhat_bit)
9731 			continue;
9732 
9733 		sfmmu_copytte(&sfhme->hme_tte, &tte);
9734 		ASSERT(TTE_IS_VALID(&tte));
9735 		vaddr = tte_to_vaddr(hmeblkp, tte);
9736 		color = addr_to_vcolor(vaddr);
9737 
9738 #ifdef DEBUG
9739 		if ((flags & HAT_CACHE) && bcolor != NO_VCOLOR) {
9740 			ASSERT(color == bcolor);
9741 		}
9742 #endif
9743 
9744 		ASSERT(flags != HAT_TMPNC || color == PP_GET_VCOLOR(pp));
9745 
9746 		ttemod = tte;
9747 		if (flags & (HAT_UNCACHE | HAT_TMPNC)) {
9748 			TTE_CLR_VCACHEABLE(&ttemod);
9749 		} else {	/* flags & HAT_CACHE */
9750 			TTE_SET_VCACHEABLE(&ttemod);
9751 		}
9752 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
9753 		if (ret < 0) {
9754 			/*
9755 			 * Since all cpus are captured modifytte should not
9756 			 * fail.
9757 			 */
9758 			panic("sfmmu_page_cache: write to tte failed");
9759 		}
9760 
9761 		sfmmup = hblktosfmmu(hmeblkp);
9762 		if (cache_flush_flag == CACHE_FLUSH) {
9763 			/*
9764 			 * Flush TSBs, TLBs and caches
9765 			 */
9766 			if (hmeblkp->hblk_shared) {
9767 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9768 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9769 				sf_region_t *rgnp;
9770 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9771 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9772 				ASSERT(srdp != NULL);
9773 				rgnp = srdp->srd_hmergnp[rid];
9774 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9775 				    srdp, rgnp, rid);
9776 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9777 				    hmeblkp, 0);
9778 				sfmmu_cache_flush(pfn, addr_to_vcolor(vaddr));
9779 			} else if (sfmmup->sfmmu_ismhat) {
9780 				if (flags & HAT_CACHE) {
9781 					SFMMU_STAT(sf_ism_recache);
9782 				} else {
9783 					SFMMU_STAT(sf_ism_uncache);
9784 				}
9785 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9786 				    pfn, CACHE_FLUSH);
9787 			} else {
9788 				sfmmu_tlbcache_demap(vaddr, sfmmup, hmeblkp,
9789 				    pfn, 0, FLUSH_ALL_CPUS, CACHE_FLUSH, 1);
9790 			}
9791 
9792 			/*
9793 			 * all cache entries belonging to this pfn are
9794 			 * now flushed.
9795 			 */
9796 			cache_flush_flag = CACHE_NO_FLUSH;
9797 		} else {
9798 			/*
9799 			 * Flush only TSBs and TLBs.
9800 			 */
9801 			if (hmeblkp->hblk_shared) {
9802 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9803 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9804 				sf_region_t *rgnp;
9805 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9806 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9807 				ASSERT(srdp != NULL);
9808 				rgnp = srdp->srd_hmergnp[rid];
9809 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9810 				    srdp, rgnp, rid);
9811 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9812 				    hmeblkp, 0);
9813 			} else if (sfmmup->sfmmu_ismhat) {
9814 				if (flags & HAT_CACHE) {
9815 					SFMMU_STAT(sf_ism_recache);
9816 				} else {
9817 					SFMMU_STAT(sf_ism_uncache);
9818 				}
9819 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9820 				    pfn, CACHE_NO_FLUSH);
9821 			} else {
9822 				sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 1);
9823 			}
9824 		}
9825 	}
9826 
9827 	if (PP_ISMAPPED_KPM(pp))
9828 		sfmmu_kpm_page_cache(pp, flags, cache_flush_flag);
9829 
9830 	switch (flags) {
9831 
9832 		default:
9833 			panic("sfmmu_pagecache: unknown flags");
9834 			break;
9835 
9836 		case HAT_CACHE:
9837 			PP_CLRTNC(pp);
9838 			PP_CLRPNC(pp);
9839 			PP_SET_VCOLOR(pp, color);
9840 			break;
9841 
9842 		case HAT_TMPNC:
9843 			PP_SETTNC(pp);
9844 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9845 			break;
9846 
9847 		case HAT_UNCACHE:
9848 			PP_SETPNC(pp);
9849 			PP_CLRTNC(pp);
9850 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9851 			break;
9852 	}
9853 }
9854 #endif	/* VAC */
9855 
9856 
9857 /*
9858  * Wrapper routine used to return a context.
9859  *
9860  * It's the responsibility of the caller to guarantee that the
9861  * process serializes on calls here by taking the HAT lock for
9862  * the hat.
9863  *
9864  */
9865 static void
9866 sfmmu_get_ctx(sfmmu_t *sfmmup)
9867 {
9868 	mmu_ctx_t *mmu_ctxp;
9869 	uint_t pstate_save;
9870 	int ret;
9871 
9872 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9873 	ASSERT(sfmmup != ksfmmup);
9874 
9875 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID)) {
9876 		sfmmu_setup_tsbinfo(sfmmup);
9877 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ALLCTX_INVALID);
9878 	}
9879 
9880 	kpreempt_disable();
9881 
9882 	mmu_ctxp = CPU_MMU_CTXP(CPU);
9883 	ASSERT(mmu_ctxp);
9884 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
9885 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
9886 
9887 	/*
9888 	 * Do a wrap-around if cnum reaches the max # cnum supported by a MMU.
9889 	 */
9890 	if (mmu_ctxp->mmu_cnum == mmu_ctxp->mmu_nctxs)
9891 		sfmmu_ctx_wrap_around(mmu_ctxp, B_TRUE);
9892 
9893 	/*
9894 	 * Let the MMU set up the page sizes to use for
9895 	 * this context in the TLB. Don't program 2nd dtlb for ism hat.
9896 	 */
9897 	if ((&mmu_set_ctx_page_sizes) && (sfmmup->sfmmu_ismhat == 0)) {
9898 		mmu_set_ctx_page_sizes(sfmmup);
9899 	}
9900 
9901 	/*
9902 	 * sfmmu_alloc_ctx and sfmmu_load_mmustate will be performed with
9903 	 * interrupts disabled to prevent race condition with wrap-around
9904 	 * ctx invalidatation. In sun4v, ctx invalidation also involves
9905 	 * a HV call to set the number of TSBs to 0. If interrupts are not
9906 	 * disabled until after sfmmu_load_mmustate is complete TSBs may
9907 	 * become assigned to INVALID_CONTEXT. This is not allowed.
9908 	 */
9909 	pstate_save = sfmmu_disable_intrs();
9910 
9911 	if (sfmmu_alloc_ctx(sfmmup, 1, CPU, SFMMU_PRIVATE) &&
9912 	    sfmmup->sfmmu_scdp != NULL) {
9913 		sf_scd_t *scdp = sfmmup->sfmmu_scdp;
9914 		sfmmu_t *scsfmmup = scdp->scd_sfmmup;
9915 		ret = sfmmu_alloc_ctx(scsfmmup, 1, CPU, SFMMU_SHARED);
9916 		/* debug purpose only */
9917 		ASSERT(!ret || scsfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
9918 		    != INVALID_CONTEXT);
9919 	}
9920 	sfmmu_load_mmustate(sfmmup);
9921 
9922 	sfmmu_enable_intrs(pstate_save);
9923 
9924 	kpreempt_enable();
9925 }
9926 
9927 /*
9928  * When all cnums are used up in a MMU, cnum will wrap around to the
9929  * next generation and start from 2.
9930  */
9931 static void
9932 sfmmu_ctx_wrap_around(mmu_ctx_t *mmu_ctxp, boolean_t reset_cnum)
9933 {
9934 
9935 	/* caller must have disabled the preemption */
9936 	ASSERT(curthread->t_preempt >= 1);
9937 	ASSERT(mmu_ctxp != NULL);
9938 
9939 	/* acquire Per-MMU (PM) spin lock */
9940 	mutex_enter(&mmu_ctxp->mmu_lock);
9941 
9942 	/* re-check to see if wrap-around is needed */
9943 	if (mmu_ctxp->mmu_cnum < mmu_ctxp->mmu_nctxs)
9944 		goto done;
9945 
9946 	SFMMU_MMU_STAT(mmu_wrap_around);
9947 
9948 	/* update gnum */
9949 	ASSERT(mmu_ctxp->mmu_gnum != 0);
9950 	mmu_ctxp->mmu_gnum++;
9951 	if (mmu_ctxp->mmu_gnum == 0 ||
9952 	    mmu_ctxp->mmu_gnum > MAX_SFMMU_GNUM_VAL) {
9953 		cmn_err(CE_PANIC, "mmu_gnum of mmu_ctx 0x%p is out of bound.",
9954 		    (void *)mmu_ctxp);
9955 	}
9956 
9957 	if (mmu_ctxp->mmu_ncpus > 1) {
9958 		cpuset_t cpuset;
9959 
9960 		membar_enter(); /* make sure updated gnum visible */
9961 
9962 		SFMMU_XCALL_STATS(NULL);
9963 
9964 		/* xcall to others on the same MMU to invalidate ctx */
9965 		cpuset = mmu_ctxp->mmu_cpuset;
9966 		ASSERT(CPU_IN_SET(cpuset, CPU->cpu_id) || !reset_cnum);
9967 		CPUSET_DEL(cpuset, CPU->cpu_id);
9968 		CPUSET_AND(cpuset, cpu_ready_set);
9969 
9970 		/*
9971 		 * Pass in INVALID_CONTEXT as the first parameter to
9972 		 * sfmmu_raise_tsb_exception, which invalidates the context
9973 		 * of any process running on the CPUs in the MMU.
9974 		 */
9975 		xt_some(cpuset, sfmmu_raise_tsb_exception,
9976 		    INVALID_CONTEXT, INVALID_CONTEXT);
9977 		xt_sync(cpuset);
9978 
9979 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
9980 	}
9981 
9982 	if (sfmmu_getctx_sec() != INVALID_CONTEXT) {
9983 		sfmmu_setctx_sec(INVALID_CONTEXT);
9984 		sfmmu_clear_utsbinfo();
9985 	}
9986 
9987 	/*
9988 	 * No xcall is needed here. For sun4u systems all CPUs in context
9989 	 * domain share a single physical MMU therefore it's enough to flush
9990 	 * TLB on local CPU. On sun4v systems we use 1 global context
9991 	 * domain and flush all remote TLBs in sfmmu_raise_tsb_exception
9992 	 * handler. Note that vtag_flushall_uctxs() is called
9993 	 * for Ultra II machine, where the equivalent flushall functionality
9994 	 * is implemented in SW, and only user ctx TLB entries are flushed.
9995 	 */
9996 	if (&vtag_flushall_uctxs != NULL) {
9997 		vtag_flushall_uctxs();
9998 	} else {
9999 		vtag_flushall();
10000 	}
10001 
10002 	/* reset mmu cnum, skips cnum 0 and 1 */
10003 	if (reset_cnum == B_TRUE)
10004 		mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
10005 
10006 done:
10007 	mutex_exit(&mmu_ctxp->mmu_lock);
10008 }
10009 
10010 
10011 /*
10012  * For multi-threaded process, set the process context to INVALID_CONTEXT
10013  * so that it faults and reloads the MMU state from TL=0. For single-threaded
10014  * process, we can just load the MMU state directly without having to
10015  * set context invalid. Caller must hold the hat lock since we don't
10016  * acquire it here.
10017  */
10018 static void
10019 sfmmu_sync_mmustate(sfmmu_t *sfmmup)
10020 {
10021 	uint_t cnum;
10022 	uint_t pstate_save;
10023 
10024 	ASSERT(sfmmup != ksfmmup);
10025 	ASSERT(sfmmu_hat_lock_held(sfmmup));
10026 
10027 	kpreempt_disable();
10028 
10029 	/*
10030 	 * We check whether the pass'ed-in sfmmup is the same as the
10031 	 * current running proc. This is to makes sure the current proc
10032 	 * stays single-threaded if it already is.
10033 	 */
10034 	if ((sfmmup == curthread->t_procp->p_as->a_hat) &&
10035 	    (curthread->t_procp->p_lwpcnt == 1)) {
10036 		/* single-thread */
10037 		cnum = sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum;
10038 		if (cnum != INVALID_CONTEXT) {
10039 			uint_t curcnum;
10040 			/*
10041 			 * Disable interrupts to prevent race condition
10042 			 * with sfmmu_ctx_wrap_around ctx invalidation.
10043 			 * In sun4v, ctx invalidation involves setting
10044 			 * TSB to NULL, hence, interrupts should be disabled
10045 			 * untill after sfmmu_load_mmustate is completed.
10046 			 */
10047 			pstate_save = sfmmu_disable_intrs();
10048 			curcnum = sfmmu_getctx_sec();
10049 			if (curcnum == cnum)
10050 				sfmmu_load_mmustate(sfmmup);
10051 			sfmmu_enable_intrs(pstate_save);
10052 			ASSERT(curcnum == cnum || curcnum == INVALID_CONTEXT);
10053 		}
10054 	} else {
10055 		/*
10056 		 * multi-thread
10057 		 * or when sfmmup is not the same as the curproc.
10058 		 */
10059 		sfmmu_invalidate_ctx(sfmmup);
10060 	}
10061 
10062 	kpreempt_enable();
10063 }
10064 
10065 
10066 /*
10067  * Replace the specified TSB with a new TSB.  This function gets called when
10068  * we grow, shrink or swapin a TSB.  When swapping in a TSB (TSB_SWAPIN), the
10069  * TSB_FORCEALLOC flag may be used to force allocation of a minimum-sized TSB
10070  * (8K).
10071  *
10072  * Caller must hold the HAT lock, but should assume any tsb_info
10073  * pointers it has are no longer valid after calling this function.
10074  *
10075  * Return values:
10076  *	TSB_ALLOCFAIL	Failed to allocate a TSB, due to memory constraints
10077  *	TSB_LOSTRACE	HAT is busy, i.e. another thread is already doing
10078  *			something to this tsbinfo/TSB
10079  *	TSB_SUCCESS	Operation succeeded
10080  */
10081 static tsb_replace_rc_t
10082 sfmmu_replace_tsb(sfmmu_t *sfmmup, struct tsb_info *old_tsbinfo, uint_t szc,
10083     hatlock_t *hatlockp, uint_t flags)
10084 {
10085 	struct tsb_info *new_tsbinfo = NULL;
10086 	struct tsb_info *curtsb, *prevtsb;
10087 	uint_t tte_sz_mask;
10088 	int i;
10089 
10090 	ASSERT(sfmmup != ksfmmup);
10091 	ASSERT(sfmmup->sfmmu_ismhat == 0);
10092 	ASSERT(sfmmu_hat_lock_held(sfmmup));
10093 	ASSERT(szc <= tsb_max_growsize);
10094 
10095 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_BUSY))
10096 		return (TSB_LOSTRACE);
10097 
10098 	/*
10099 	 * Find the tsb_info ahead of this one in the list, and
10100 	 * also make sure that the tsb_info passed in really
10101 	 * exists!
10102 	 */
10103 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
10104 	    curtsb != old_tsbinfo && curtsb != NULL;
10105 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
10106 		;
10107 	ASSERT(curtsb != NULL);
10108 
10109 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
10110 		/*
10111 		 * The process is swapped out, so just set the new size
10112 		 * code.  When it swaps back in, we'll allocate a new one
10113 		 * of the new chosen size.
10114 		 */
10115 		curtsb->tsb_szc = szc;
10116 		return (TSB_SUCCESS);
10117 	}
10118 	SFMMU_FLAGS_SET(sfmmup, HAT_BUSY);
10119 
10120 	tte_sz_mask = old_tsbinfo->tsb_ttesz_mask;
10121 
10122 	/*
10123 	 * All initialization is done inside of sfmmu_tsbinfo_alloc().
10124 	 * If we fail to allocate a TSB, exit.
10125 	 *
10126 	 * If tsb grows with new tsb size > 4M and old tsb size < 4M,
10127 	 * then try 4M slab after the initial alloc fails.
10128 	 *
10129 	 * If tsb swapin with tsb size > 4M, then try 4M after the
10130 	 * initial alloc fails.
10131 	 */
10132 	sfmmu_hat_exit(hatlockp);
10133 	if (sfmmu_tsbinfo_alloc(&new_tsbinfo, szc,
10134 	    tte_sz_mask, flags, sfmmup) &&
10135 	    (!(flags & (TSB_GROW | TSB_SWAPIN)) || (szc <= TSB_4M_SZCODE) ||
10136 	    (!(flags & TSB_SWAPIN) &&
10137 	    (old_tsbinfo->tsb_szc >= TSB_4M_SZCODE)) ||
10138 	    sfmmu_tsbinfo_alloc(&new_tsbinfo, TSB_4M_SZCODE,
10139 	    tte_sz_mask, flags, sfmmup))) {
10140 		(void) sfmmu_hat_enter(sfmmup);
10141 		if (!(flags & TSB_SWAPIN))
10142 			SFMMU_STAT(sf_tsb_resize_failures);
10143 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10144 		return (TSB_ALLOCFAIL);
10145 	}
10146 	(void) sfmmu_hat_enter(sfmmup);
10147 
10148 	/*
10149 	 * Re-check to make sure somebody else didn't muck with us while we
10150 	 * didn't hold the HAT lock.  If the process swapped out, fine, just
10151 	 * exit; this can happen if we try to shrink the TSB from the context
10152 	 * of another process (such as on an ISM unmap), though it is rare.
10153 	 */
10154 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
10155 		SFMMU_STAT(sf_tsb_resize_failures);
10156 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10157 		sfmmu_hat_exit(hatlockp);
10158 		sfmmu_tsbinfo_free(new_tsbinfo);
10159 		(void) sfmmu_hat_enter(sfmmup);
10160 		return (TSB_LOSTRACE);
10161 	}
10162 
10163 #ifdef	DEBUG
10164 	/* Reverify that the tsb_info still exists.. for debugging only */
10165 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
10166 	    curtsb != old_tsbinfo && curtsb != NULL;
10167 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
10168 		;
10169 	ASSERT(curtsb != NULL);
10170 #endif	/* DEBUG */
10171 
10172 	/*
10173 	 * Quiesce any CPUs running this process on their next TLB miss
10174 	 * so they atomically see the new tsb_info.  We temporarily set the
10175 	 * context to invalid context so new threads that come on processor
10176 	 * after we do the xcall to cpusran will also serialize behind the
10177 	 * HAT lock on TLB miss and will see the new TSB.  Since this short
10178 	 * race with a new thread coming on processor is relatively rare,
10179 	 * this synchronization mechanism should be cheaper than always
10180 	 * pausing all CPUs for the duration of the setup, which is what
10181 	 * the old implementation did.  This is particuarly true if we are
10182 	 * copying a huge chunk of memory around during that window.
10183 	 *
10184 	 * The memory barriers are to make sure things stay consistent
10185 	 * with resume() since it does not hold the HAT lock while
10186 	 * walking the list of tsb_info structures.
10187 	 */
10188 	if ((flags & TSB_SWAPIN) != TSB_SWAPIN) {
10189 		/* The TSB is either growing or shrinking. */
10190 		sfmmu_invalidate_ctx(sfmmup);
10191 	} else {
10192 		/*
10193 		 * It is illegal to swap in TSBs from a process other
10194 		 * than a process being swapped in.  This in turn
10195 		 * implies we do not have a valid MMU context here
10196 		 * since a process needs one to resolve translation
10197 		 * misses.
10198 		 */
10199 		ASSERT(curthread->t_procp->p_as->a_hat == sfmmup);
10200 	}
10201 
10202 #ifdef DEBUG
10203 	ASSERT(max_mmu_ctxdoms > 0);
10204 
10205 	/*
10206 	 * Process should have INVALID_CONTEXT on all MMUs
10207 	 */
10208 	for (i = 0; i < max_mmu_ctxdoms; i++) {
10209 
10210 		ASSERT(sfmmup->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
10211 	}
10212 #endif
10213 
10214 	new_tsbinfo->tsb_next = old_tsbinfo->tsb_next;
10215 	membar_stst();	/* strict ordering required */
10216 	if (prevtsb)
10217 		prevtsb->tsb_next = new_tsbinfo;
10218 	else
10219 		sfmmup->sfmmu_tsb = new_tsbinfo;
10220 	membar_enter();	/* make sure new TSB globally visible */
10221 
10222 	/*
10223 	 * We need to migrate TSB entries from the old TSB to the new TSB
10224 	 * if tsb_remap_ttes is set and the TSB is growing.
10225 	 */
10226 	if (tsb_remap_ttes && ((flags & TSB_GROW) == TSB_GROW))
10227 		sfmmu_copy_tsb(old_tsbinfo, new_tsbinfo);
10228 
10229 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10230 
10231 	/*
10232 	 * Drop the HAT lock to free our old tsb_info.
10233 	 */
10234 	sfmmu_hat_exit(hatlockp);
10235 
10236 	if ((flags & TSB_GROW) == TSB_GROW) {
10237 		SFMMU_STAT(sf_tsb_grow);
10238 	} else if ((flags & TSB_SHRINK) == TSB_SHRINK) {
10239 		SFMMU_STAT(sf_tsb_shrink);
10240 	}
10241 
10242 	sfmmu_tsbinfo_free(old_tsbinfo);
10243 
10244 	(void) sfmmu_hat_enter(sfmmup);
10245 	return (TSB_SUCCESS);
10246 }
10247 
10248 /*
10249  * This function will re-program hat pgsz array, and invalidate the
10250  * process' context, forcing the process to switch to another
10251  * context on the next TLB miss, and therefore start using the
10252  * TLB that is reprogrammed for the new page sizes.
10253  */
10254 void
10255 sfmmu_reprog_pgsz_arr(sfmmu_t *sfmmup, uint8_t *tmp_pgsz)
10256 {
10257 	int i;
10258 	hatlock_t *hatlockp = NULL;
10259 
10260 	hatlockp = sfmmu_hat_enter(sfmmup);
10261 	/* USIII+-IV+ optimization, requires hat lock */
10262 	if (tmp_pgsz) {
10263 		for (i = 0; i < mmu_page_sizes; i++)
10264 			sfmmup->sfmmu_pgsz[i] = tmp_pgsz[i];
10265 	}
10266 	SFMMU_STAT(sf_tlb_reprog_pgsz);
10267 
10268 	sfmmu_invalidate_ctx(sfmmup);
10269 
10270 	sfmmu_hat_exit(hatlockp);
10271 }
10272 
10273 /*
10274  * The scd_rttecnt field in the SCD must be updated to take account of the
10275  * regions which it contains.
10276  */
10277 static void
10278 sfmmu_set_scd_rttecnt(sf_srd_t *srdp, sf_scd_t *scdp)
10279 {
10280 	uint_t rid;
10281 	uint_t i, j;
10282 	ulong_t w;
10283 	sf_region_t *rgnp;
10284 
10285 	ASSERT(srdp != NULL);
10286 
10287 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
10288 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
10289 			continue;
10290 		}
10291 
10292 		j = 0;
10293 		while (w) {
10294 			if (!(w & 0x1)) {
10295 				j++;
10296 				w >>= 1;
10297 				continue;
10298 			}
10299 			rid = (i << BT_ULSHIFT) | j;
10300 			j++;
10301 			w >>= 1;
10302 
10303 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
10304 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
10305 			rgnp = srdp->srd_hmergnp[rid];
10306 			ASSERT(rgnp->rgn_refcnt > 0);
10307 			ASSERT(rgnp->rgn_id == rid);
10308 
10309 			scdp->scd_rttecnt[rgnp->rgn_pgszc] +=
10310 			    rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
10311 
10312 			/*
10313 			 * Maintain the tsb0 inflation cnt for the regions
10314 			 * in the SCD.
10315 			 */
10316 			if (rgnp->rgn_pgszc >= TTE4M) {
10317 				scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt +=
10318 				    rgnp->rgn_size >>
10319 				    (TTE_PAGE_SHIFT(TTE8K) + 2);
10320 			}
10321 		}
10322 	}
10323 }
10324 
10325 /*
10326  * This function assumes that there are either four or six supported page
10327  * sizes and at most two programmable TLBs, so we need to decide which
10328  * page sizes are most important and then tell the MMU layer so it
10329  * can adjust the TLB page sizes accordingly (if supported).
10330  *
10331  * If these assumptions change, this function will need to be
10332  * updated to support whatever the new limits are.
10333  *
10334  * The growing flag is nonzero if we are growing the address space,
10335  * and zero if it is shrinking.  This allows us to decide whether
10336  * to grow or shrink our TSB, depending upon available memory
10337  * conditions.
10338  */
10339 static void
10340 sfmmu_check_page_sizes(sfmmu_t *sfmmup, int growing)
10341 {
10342 	uint64_t ttecnt[MMU_PAGE_SIZES];
10343 	uint64_t tte8k_cnt, tte4m_cnt;
10344 	uint8_t i;
10345 	int sectsb_thresh;
10346 
10347 	/*
10348 	 * Kernel threads, processes with small address spaces not using
10349 	 * large pages, and dummy ISM HATs need not apply.
10350 	 */
10351 	if (sfmmup == ksfmmup || sfmmup->sfmmu_ismhat != NULL)
10352 		return;
10353 
10354 	if (!SFMMU_LGPGS_INUSE(sfmmup) &&
10355 	    sfmmup->sfmmu_ttecnt[TTE8K] <= tsb_rss_factor)
10356 		return;
10357 
10358 	for (i = 0; i < mmu_page_sizes; i++) {
10359 		ttecnt[i] = sfmmup->sfmmu_ttecnt[i] +
10360 		    sfmmup->sfmmu_ismttecnt[i];
10361 	}
10362 
10363 	/* Check pagesizes in use, and possibly reprogram DTLB. */
10364 	if (&mmu_check_page_sizes)
10365 		mmu_check_page_sizes(sfmmup, ttecnt);
10366 
10367 	/*
10368 	 * Calculate the number of 8k ttes to represent the span of these
10369 	 * pages.
10370 	 */
10371 	tte8k_cnt = ttecnt[TTE8K] +
10372 	    (ttecnt[TTE64K] << (MMU_PAGESHIFT64K - MMU_PAGESHIFT)) +
10373 	    (ttecnt[TTE512K] << (MMU_PAGESHIFT512K - MMU_PAGESHIFT));
10374 	if (mmu_page_sizes == max_mmu_page_sizes) {
10375 		tte4m_cnt = ttecnt[TTE4M] +
10376 		    (ttecnt[TTE32M] << (MMU_PAGESHIFT32M - MMU_PAGESHIFT4M)) +
10377 		    (ttecnt[TTE256M] << (MMU_PAGESHIFT256M - MMU_PAGESHIFT4M));
10378 	} else {
10379 		tte4m_cnt = ttecnt[TTE4M];
10380 	}
10381 
10382 	/*
10383 	 * Inflate tte8k_cnt to allow for region large page allocation failure.
10384 	 */
10385 	tte8k_cnt += sfmmup->sfmmu_tsb0_4minflcnt;
10386 
10387 	/*
10388 	 * Inflate TSB sizes by a factor of 2 if this process
10389 	 * uses 4M text pages to minimize extra conflict misses
10390 	 * in the first TSB since without counting text pages
10391 	 * 8K TSB may become too small.
10392 	 *
10393 	 * Also double the size of the second TSB to minimize
10394 	 * extra conflict misses due to competition between 4M text pages
10395 	 * and data pages.
10396 	 *
10397 	 * We need to adjust the second TSB allocation threshold by the
10398 	 * inflation factor, since there is no point in creating a second
10399 	 * TSB when we know all the mappings can fit in the I/D TLBs.
10400 	 */
10401 	sectsb_thresh = tsb_sectsb_threshold;
10402 	if (sfmmup->sfmmu_flags & HAT_4MTEXT_FLAG) {
10403 		tte8k_cnt <<= 1;
10404 		tte4m_cnt <<= 1;
10405 		sectsb_thresh <<= 1;
10406 	}
10407 
10408 	/*
10409 	 * Check to see if our TSB is the right size; we may need to
10410 	 * grow or shrink it.  If the process is small, our work is
10411 	 * finished at this point.
10412 	 */
10413 	if (tte8k_cnt <= tsb_rss_factor && tte4m_cnt <= sectsb_thresh) {
10414 		return;
10415 	}
10416 	sfmmu_size_tsb(sfmmup, growing, tte8k_cnt, tte4m_cnt, sectsb_thresh);
10417 }
10418 
10419 static void
10420 sfmmu_size_tsb(sfmmu_t *sfmmup, int growing, uint64_t tte8k_cnt,
10421 	uint64_t tte4m_cnt, int sectsb_thresh)
10422 {
10423 	int tsb_bits;
10424 	uint_t tsb_szc;
10425 	struct tsb_info *tsbinfop;
10426 	hatlock_t *hatlockp = NULL;
10427 
10428 	hatlockp = sfmmu_hat_enter(sfmmup);
10429 	ASSERT(hatlockp != NULL);
10430 	tsbinfop = sfmmup->sfmmu_tsb;
10431 	ASSERT(tsbinfop != NULL);
10432 
10433 	/*
10434 	 * If we're growing, select the size based on RSS.  If we're
10435 	 * shrinking, leave some room so we don't have to turn around and
10436 	 * grow again immediately.
10437 	 */
10438 	if (growing)
10439 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
10440 	else
10441 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt << 1);
10442 
10443 	if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10444 	    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10445 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10446 		    hatlockp, TSB_SHRINK);
10447 	} else if (growing && tsb_szc > tsbinfop->tsb_szc && TSB_OK_GROW()) {
10448 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10449 		    hatlockp, TSB_GROW);
10450 	}
10451 	tsbinfop = sfmmup->sfmmu_tsb;
10452 
10453 	/*
10454 	 * With the TLB and first TSB out of the way, we need to see if
10455 	 * we need a second TSB for 4M pages.  If we managed to reprogram
10456 	 * the TLB page sizes above, the process will start using this new
10457 	 * TSB right away; otherwise, it will start using it on the next
10458 	 * context switch.  Either way, it's no big deal so there's no
10459 	 * synchronization with the trap handlers here unless we grow the
10460 	 * TSB (in which case it's required to prevent using the old one
10461 	 * after it's freed). Note: second tsb is required for 32M/256M
10462 	 * page sizes.
10463 	 */
10464 	if (tte4m_cnt > sectsb_thresh) {
10465 		/*
10466 		 * If we're growing, select the size based on RSS.  If we're
10467 		 * shrinking, leave some room so we don't have to turn
10468 		 * around and grow again immediately.
10469 		 */
10470 		if (growing)
10471 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
10472 		else
10473 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt << 1);
10474 		if (tsbinfop->tsb_next == NULL) {
10475 			struct tsb_info *newtsb;
10476 			int allocflags = SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)?
10477 			    0 : TSB_ALLOC;
10478 
10479 			sfmmu_hat_exit(hatlockp);
10480 
10481 			/*
10482 			 * Try to allocate a TSB for 4[32|256]M pages.  If we
10483 			 * can't get the size we want, retry w/a minimum sized
10484 			 * TSB.  If that still didn't work, give up; we can
10485 			 * still run without one.
10486 			 */
10487 			tsb_bits = (mmu_page_sizes == max_mmu_page_sizes)?
10488 			    TSB4M|TSB32M|TSB256M:TSB4M;
10489 			if ((sfmmu_tsbinfo_alloc(&newtsb, tsb_szc, tsb_bits,
10490 			    allocflags, sfmmup)) &&
10491 			    (tsb_szc <= TSB_4M_SZCODE ||
10492 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
10493 			    tsb_bits, allocflags, sfmmup)) &&
10494 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_MIN_SZCODE,
10495 			    tsb_bits, allocflags, sfmmup)) {
10496 				return;
10497 			}
10498 
10499 			hatlockp = sfmmu_hat_enter(sfmmup);
10500 
10501 			sfmmu_invalidate_ctx(sfmmup);
10502 
10503 			if (sfmmup->sfmmu_tsb->tsb_next == NULL) {
10504 				sfmmup->sfmmu_tsb->tsb_next = newtsb;
10505 				SFMMU_STAT(sf_tsb_sectsb_create);
10506 				sfmmu_hat_exit(hatlockp);
10507 				return;
10508 			} else {
10509 				/*
10510 				 * It's annoying, but possible for us
10511 				 * to get here.. we dropped the HAT lock
10512 				 * because of locking order in the kmem
10513 				 * allocator, and while we were off getting
10514 				 * our memory, some other thread decided to
10515 				 * do us a favor and won the race to get a
10516 				 * second TSB for this process.  Sigh.
10517 				 */
10518 				sfmmu_hat_exit(hatlockp);
10519 				sfmmu_tsbinfo_free(newtsb);
10520 				return;
10521 			}
10522 		}
10523 
10524 		/*
10525 		 * We have a second TSB, see if it's big enough.
10526 		 */
10527 		tsbinfop = tsbinfop->tsb_next;
10528 
10529 		/*
10530 		 * Check to see if our second TSB is the right size;
10531 		 * we may need to grow or shrink it.
10532 		 * To prevent thrashing (e.g. growing the TSB on a
10533 		 * subsequent map operation), only try to shrink if
10534 		 * the TSB reach exceeds twice the virtual address
10535 		 * space size.
10536 		 */
10537 		if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10538 		    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10539 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10540 			    tsb_szc, hatlockp, TSB_SHRINK);
10541 		} else if (growing && tsb_szc > tsbinfop->tsb_szc &&
10542 		    TSB_OK_GROW()) {
10543 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10544 			    tsb_szc, hatlockp, TSB_GROW);
10545 		}
10546 	}
10547 
10548 	sfmmu_hat_exit(hatlockp);
10549 }
10550 
10551 /*
10552  * Free up a sfmmu
10553  * Since the sfmmu is currently embedded in the hat struct we simply zero
10554  * out our fields and free up the ism map blk list if any.
10555  */
10556 static void
10557 sfmmu_free_sfmmu(sfmmu_t *sfmmup)
10558 {
10559 	ism_blk_t	*blkp, *nx_blkp;
10560 #ifdef	DEBUG
10561 	ism_map_t	*map;
10562 	int 		i;
10563 #endif
10564 
10565 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
10566 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
10567 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
10568 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
10569 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
10570 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
10571 	ASSERT(SF_RGNMAP_ISNULL(sfmmup));
10572 
10573 	sfmmup->sfmmu_free = 0;
10574 	sfmmup->sfmmu_ismhat = 0;
10575 
10576 	blkp = sfmmup->sfmmu_iblk;
10577 	sfmmup->sfmmu_iblk = NULL;
10578 
10579 	while (blkp) {
10580 #ifdef	DEBUG
10581 		map = blkp->iblk_maps;
10582 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
10583 			ASSERT(map[i].imap_seg == 0);
10584 			ASSERT(map[i].imap_ismhat == NULL);
10585 			ASSERT(map[i].imap_ment == NULL);
10586 		}
10587 #endif
10588 		nx_blkp = blkp->iblk_next;
10589 		blkp->iblk_next = NULL;
10590 		blkp->iblk_nextpa = (uint64_t)-1;
10591 		kmem_cache_free(ism_blk_cache, blkp);
10592 		blkp = nx_blkp;
10593 	}
10594 }
10595 
10596 /*
10597  * Locking primitves accessed by HATLOCK macros
10598  */
10599 
10600 #define	SFMMU_SPL_MTX	(0x0)
10601 #define	SFMMU_ML_MTX	(0x1)
10602 
10603 #define	SFMMU_MLSPL_MTX(type, pg)	(((type) == SFMMU_SPL_MTX) ? \
10604 					    SPL_HASH(pg) : MLIST_HASH(pg))
10605 
10606 kmutex_t *
10607 sfmmu_page_enter(struct page *pp)
10608 {
10609 	return (sfmmu_mlspl_enter(pp, SFMMU_SPL_MTX));
10610 }
10611 
10612 void
10613 sfmmu_page_exit(kmutex_t *spl)
10614 {
10615 	mutex_exit(spl);
10616 }
10617 
10618 int
10619 sfmmu_page_spl_held(struct page *pp)
10620 {
10621 	return (sfmmu_mlspl_held(pp, SFMMU_SPL_MTX));
10622 }
10623 
10624 kmutex_t *
10625 sfmmu_mlist_enter(struct page *pp)
10626 {
10627 	return (sfmmu_mlspl_enter(pp, SFMMU_ML_MTX));
10628 }
10629 
10630 void
10631 sfmmu_mlist_exit(kmutex_t *mml)
10632 {
10633 	mutex_exit(mml);
10634 }
10635 
10636 int
10637 sfmmu_mlist_held(struct page *pp)
10638 {
10639 
10640 	return (sfmmu_mlspl_held(pp, SFMMU_ML_MTX));
10641 }
10642 
10643 /*
10644  * Common code for sfmmu_mlist_enter() and sfmmu_page_enter().  For
10645  * sfmmu_mlist_enter() case mml_table lock array is used and for
10646  * sfmmu_page_enter() sfmmu_page_lock lock array is used.
10647  *
10648  * The lock is taken on a root page so that it protects an operation on all
10649  * constituent pages of a large page pp belongs to.
10650  *
10651  * The routine takes a lock from the appropriate array. The lock is determined
10652  * by hashing the root page. After taking the lock this routine checks if the
10653  * root page has the same size code that was used to determine the root (i.e
10654  * that root hasn't changed).  If root page has the expected p_szc field we
10655  * have the right lock and it's returned to the caller. If root's p_szc
10656  * decreased we release the lock and retry from the beginning.  This case can
10657  * happen due to hat_page_demote() decreasing p_szc between our load of p_szc
10658  * value and taking the lock. The number of retries due to p_szc decrease is
10659  * limited by the maximum p_szc value. If p_szc is 0 we return the lock
10660  * determined by hashing pp itself.
10661  *
10662  * If our caller doesn't hold a SE_SHARED or SE_EXCL lock on pp it's also
10663  * possible that p_szc can increase. To increase p_szc a thread has to lock
10664  * all constituent pages EXCL and do hat_pageunload() on all of them. All the
10665  * callers that don't hold a page locked recheck if hmeblk through which pp
10666  * was found still maps this pp.  If it doesn't map it anymore returned lock
10667  * is immediately dropped. Therefore if sfmmu_mlspl_enter() hits the case of
10668  * p_szc increase after taking the lock it returns this lock without further
10669  * retries because in this case the caller doesn't care about which lock was
10670  * taken. The caller will drop it right away.
10671  *
10672  * After the routine returns it's guaranteed that hat_page_demote() can't
10673  * change p_szc field of any of constituent pages of a large page pp belongs
10674  * to as long as pp was either locked at least SHARED prior to this call or
10675  * the caller finds that hment that pointed to this pp still references this
10676  * pp (this also assumes that the caller holds hme hash bucket lock so that
10677  * the same pp can't be remapped into the same hmeblk after it was unmapped by
10678  * hat_pageunload()).
10679  */
10680 static kmutex_t *
10681 sfmmu_mlspl_enter(struct page *pp, int type)
10682 {
10683 	kmutex_t	*mtx;
10684 	uint_t		prev_rszc = UINT_MAX;
10685 	page_t		*rootpp;
10686 	uint_t		szc;
10687 	uint_t		rszc;
10688 	uint_t		pszc = pp->p_szc;
10689 
10690 	ASSERT(pp != NULL);
10691 
10692 again:
10693 	if (pszc == 0) {
10694 		mtx = SFMMU_MLSPL_MTX(type, pp);
10695 		mutex_enter(mtx);
10696 		return (mtx);
10697 	}
10698 
10699 	/* The lock lives in the root page */
10700 	rootpp = PP_GROUPLEADER(pp, pszc);
10701 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10702 	mutex_enter(mtx);
10703 
10704 	/*
10705 	 * Return mml in the following 3 cases:
10706 	 *
10707 	 * 1) If pp itself is root since if its p_szc decreased before we took
10708 	 * the lock pp is still the root of smaller szc page. And if its p_szc
10709 	 * increased it doesn't matter what lock we return (see comment in
10710 	 * front of this routine).
10711 	 *
10712 	 * 2) If pp's not root but rootpp is the root of a rootpp->p_szc size
10713 	 * large page we have the right lock since any previous potential
10714 	 * hat_page_demote() is done demoting from greater than current root's
10715 	 * p_szc because hat_page_demote() changes root's p_szc last. No
10716 	 * further hat_page_demote() can start or be in progress since it
10717 	 * would need the same lock we currently hold.
10718 	 *
10719 	 * 3) If rootpp's p_szc increased since previous iteration it doesn't
10720 	 * matter what lock we return (see comment in front of this routine).
10721 	 */
10722 	if (pp == rootpp || (rszc = rootpp->p_szc) == pszc ||
10723 	    rszc >= prev_rszc) {
10724 		return (mtx);
10725 	}
10726 
10727 	/*
10728 	 * hat_page_demote() could have decreased root's p_szc.
10729 	 * In this case pp's p_szc must also be smaller than pszc.
10730 	 * Retry.
10731 	 */
10732 	if (rszc < pszc) {
10733 		szc = pp->p_szc;
10734 		if (szc < pszc) {
10735 			mutex_exit(mtx);
10736 			pszc = szc;
10737 			goto again;
10738 		}
10739 		/*
10740 		 * pp's p_szc increased after it was decreased.
10741 		 * page cannot be mapped. Return current lock. The caller
10742 		 * will drop it right away.
10743 		 */
10744 		return (mtx);
10745 	}
10746 
10747 	/*
10748 	 * root's p_szc is greater than pp's p_szc.
10749 	 * hat_page_demote() is not done with all pages
10750 	 * yet. Wait for it to complete.
10751 	 */
10752 	mutex_exit(mtx);
10753 	rootpp = PP_GROUPLEADER(rootpp, rszc);
10754 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10755 	mutex_enter(mtx);
10756 	mutex_exit(mtx);
10757 	prev_rszc = rszc;
10758 	goto again;
10759 }
10760 
10761 static int
10762 sfmmu_mlspl_held(struct page *pp, int type)
10763 {
10764 	kmutex_t	*mtx;
10765 
10766 	ASSERT(pp != NULL);
10767 	/* The lock lives in the root page */
10768 	pp = PP_PAGEROOT(pp);
10769 	ASSERT(pp != NULL);
10770 
10771 	mtx = SFMMU_MLSPL_MTX(type, pp);
10772 	return (MUTEX_HELD(mtx));
10773 }
10774 
10775 static uint_t
10776 sfmmu_get_free_hblk(struct hme_blk **hmeblkpp, uint_t critical)
10777 {
10778 	struct  hme_blk *hblkp;
10779 
10780 
10781 	if (freehblkp != NULL) {
10782 		mutex_enter(&freehblkp_lock);
10783 		if (freehblkp != NULL) {
10784 			/*
10785 			 * If the current thread is owning hblk_reserve OR
10786 			 * critical request from sfmmu_hblk_steal()
10787 			 * let it succeed even if freehblkcnt is really low.
10788 			 */
10789 			if (freehblkcnt <= HBLK_RESERVE_MIN && !critical) {
10790 				SFMMU_STAT(sf_get_free_throttle);
10791 				mutex_exit(&freehblkp_lock);
10792 				return (0);
10793 			}
10794 			freehblkcnt--;
10795 			*hmeblkpp = freehblkp;
10796 			hblkp = *hmeblkpp;
10797 			freehblkp = hblkp->hblk_next;
10798 			mutex_exit(&freehblkp_lock);
10799 			hblkp->hblk_next = NULL;
10800 			SFMMU_STAT(sf_get_free_success);
10801 
10802 			ASSERT(hblkp->hblk_hmecnt == 0);
10803 			ASSERT(hblkp->hblk_vcnt == 0);
10804 			ASSERT(hblkp->hblk_nextpa == va_to_pa((caddr_t)hblkp));
10805 
10806 			return (1);
10807 		}
10808 		mutex_exit(&freehblkp_lock);
10809 	}
10810 
10811 	/* Check cpu hblk pending queues */
10812 	if ((*hmeblkpp = sfmmu_check_pending_hblks(TTE8K)) != NULL) {
10813 		hblkp = *hmeblkpp;
10814 		hblkp->hblk_next = NULL;
10815 		hblkp->hblk_nextpa = va_to_pa((caddr_t)hblkp);
10816 
10817 		ASSERT(hblkp->hblk_hmecnt == 0);
10818 		ASSERT(hblkp->hblk_vcnt == 0);
10819 
10820 		return (1);
10821 	}
10822 
10823 	SFMMU_STAT(sf_get_free_fail);
10824 	return (0);
10825 }
10826 
10827 static uint_t
10828 sfmmu_put_free_hblk(struct hme_blk *hmeblkp, uint_t critical)
10829 {
10830 	struct  hme_blk *hblkp;
10831 
10832 	ASSERT(hmeblkp->hblk_hmecnt == 0);
10833 	ASSERT(hmeblkp->hblk_vcnt == 0);
10834 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
10835 
10836 	/*
10837 	 * If the current thread is mapping into kernel space,
10838 	 * let it succede even if freehblkcnt is max
10839 	 * so that it will avoid freeing it to kmem.
10840 	 * This will prevent stack overflow due to
10841 	 * possible recursion since kmem_cache_free()
10842 	 * might require creation of a slab which
10843 	 * in turn needs an hmeblk to map that slab;
10844 	 * let's break this vicious chain at the first
10845 	 * opportunity.
10846 	 */
10847 	if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10848 		mutex_enter(&freehblkp_lock);
10849 		if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10850 			SFMMU_STAT(sf_put_free_success);
10851 			freehblkcnt++;
10852 			hmeblkp->hblk_next = freehblkp;
10853 			freehblkp = hmeblkp;
10854 			mutex_exit(&freehblkp_lock);
10855 			return (1);
10856 		}
10857 		mutex_exit(&freehblkp_lock);
10858 	}
10859 
10860 	/*
10861 	 * Bring down freehblkcnt to HBLK_RESERVE_CNT. We are here
10862 	 * only if freehblkcnt is at least HBLK_RESERVE_CNT *and*
10863 	 * we are not in the process of mapping into kernel space.
10864 	 */
10865 	ASSERT(!critical);
10866 	while (freehblkcnt > HBLK_RESERVE_CNT) {
10867 		mutex_enter(&freehblkp_lock);
10868 		if (freehblkcnt > HBLK_RESERVE_CNT) {
10869 			freehblkcnt--;
10870 			hblkp = freehblkp;
10871 			freehblkp = hblkp->hblk_next;
10872 			mutex_exit(&freehblkp_lock);
10873 			ASSERT(get_hblk_cache(hblkp) == sfmmu8_cache);
10874 			kmem_cache_free(sfmmu8_cache, hblkp);
10875 			continue;
10876 		}
10877 		mutex_exit(&freehblkp_lock);
10878 	}
10879 	SFMMU_STAT(sf_put_free_fail);
10880 	return (0);
10881 }
10882 
10883 static void
10884 sfmmu_hblk_swap(struct hme_blk *new)
10885 {
10886 	struct hme_blk *old, *hblkp, *prev;
10887 	uint64_t newpa;
10888 	caddr_t	base, vaddr, endaddr;
10889 	struct hmehash_bucket *hmebp;
10890 	struct sf_hment *osfhme, *nsfhme;
10891 	page_t *pp;
10892 	kmutex_t *pml;
10893 	tte_t tte;
10894 	struct hme_blk *list = NULL;
10895 
10896 #ifdef	DEBUG
10897 	hmeblk_tag		hblktag;
10898 	struct hme_blk		*found;
10899 #endif
10900 	old = HBLK_RESERVE;
10901 	ASSERT(!old->hblk_shared);
10902 
10903 	/*
10904 	 * save pa before bcopy clobbers it
10905 	 */
10906 	newpa = new->hblk_nextpa;
10907 
10908 	base = (caddr_t)get_hblk_base(old);
10909 	endaddr = base + get_hblk_span(old);
10910 
10911 	/*
10912 	 * acquire hash bucket lock.
10913 	 */
10914 	hmebp = sfmmu_tteload_acquire_hashbucket(ksfmmup, base, TTE8K,
10915 	    SFMMU_INVALID_SHMERID);
10916 
10917 	/*
10918 	 * copy contents from old to new
10919 	 */
10920 	bcopy((void *)old, (void *)new, HME8BLK_SZ);
10921 
10922 	/*
10923 	 * add new to hash chain
10924 	 */
10925 	sfmmu_hblk_hash_add(hmebp, new, newpa);
10926 
10927 	/*
10928 	 * search hash chain for hblk_reserve; this needs to be performed
10929 	 * after adding new, otherwise prev won't correspond to the hblk which
10930 	 * is prior to old in hash chain when we call sfmmu_hblk_hash_rm to
10931 	 * remove old later.
10932 	 */
10933 	for (prev = NULL,
10934 	    hblkp = hmebp->hmeblkp; hblkp != NULL && hblkp != old;
10935 	    prev = hblkp, hblkp = hblkp->hblk_next)
10936 		;
10937 
10938 	if (hblkp != old)
10939 		panic("sfmmu_hblk_swap: hblk_reserve not found");
10940 
10941 	/*
10942 	 * p_mapping list is still pointing to hments in hblk_reserve;
10943 	 * fix up p_mapping list so that they point to hments in new.
10944 	 *
10945 	 * Since all these mappings are created by hblk_reserve_thread
10946 	 * on the way and it's using at least one of the buffers from each of
10947 	 * the newly minted slabs, there is no danger of any of these
10948 	 * mappings getting unloaded by another thread.
10949 	 *
10950 	 * tsbmiss could only modify ref/mod bits of hments in old/new.
10951 	 * Since all of these hments hold mappings established by segkmem
10952 	 * and mappings in segkmem are setup with HAT_NOSYNC, ref/mod bits
10953 	 * have no meaning for the mappings in hblk_reserve.  hments in
10954 	 * old and new are identical except for ref/mod bits.
10955 	 */
10956 	for (vaddr = base; vaddr < endaddr; vaddr += TTEBYTES(TTE8K)) {
10957 
10958 		HBLKTOHME(osfhme, old, vaddr);
10959 		sfmmu_copytte(&osfhme->hme_tte, &tte);
10960 
10961 		if (TTE_IS_VALID(&tte)) {
10962 			if ((pp = osfhme->hme_page) == NULL)
10963 				panic("sfmmu_hblk_swap: page not mapped");
10964 
10965 			pml = sfmmu_mlist_enter(pp);
10966 
10967 			if (pp != osfhme->hme_page)
10968 				panic("sfmmu_hblk_swap: mapping changed");
10969 
10970 			HBLKTOHME(nsfhme, new, vaddr);
10971 
10972 			HME_ADD(nsfhme, pp);
10973 			HME_SUB(osfhme, pp);
10974 
10975 			sfmmu_mlist_exit(pml);
10976 		}
10977 	}
10978 
10979 	/*
10980 	 * remove old from hash chain
10981 	 */
10982 	sfmmu_hblk_hash_rm(hmebp, old, prev, &list, 1);
10983 
10984 #ifdef	DEBUG
10985 
10986 	hblktag.htag_id = ksfmmup;
10987 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
10988 	hblktag.htag_bspage = HME_HASH_BSPAGE(base, HME_HASH_SHIFT(TTE8K));
10989 	hblktag.htag_rehash = HME_HASH_REHASH(TTE8K);
10990 	HME_HASH_FAST_SEARCH(hmebp, hblktag, found);
10991 
10992 	if (found != new)
10993 		panic("sfmmu_hblk_swap: new hblk not found");
10994 #endif
10995 
10996 	SFMMU_HASH_UNLOCK(hmebp);
10997 
10998 	/*
10999 	 * Reset hblk_reserve
11000 	 */
11001 	bzero((void *)old, HME8BLK_SZ);
11002 	old->hblk_nextpa = va_to_pa((caddr_t)old);
11003 }
11004 
11005 /*
11006  * Grab the mlist mutex for both pages passed in.
11007  *
11008  * low and high will be returned as pointers to the mutexes for these pages.
11009  * low refers to the mutex residing in the lower bin of the mlist hash, while
11010  * high refers to the mutex residing in the higher bin of the mlist hash.  This
11011  * is due to the locking order restrictions on the same thread grabbing
11012  * multiple mlist mutexes.  The low lock must be acquired before the high lock.
11013  *
11014  * If both pages hash to the same mutex, only grab that single mutex, and
11015  * high will be returned as NULL
11016  * If the pages hash to different bins in the hash, grab the lower addressed
11017  * lock first and then the higher addressed lock in order to follow the locking
11018  * rules involved with the same thread grabbing multiple mlist mutexes.
11019  * low and high will both have non-NULL values.
11020  */
11021 static void
11022 sfmmu_mlist_reloc_enter(struct page *targ, struct page *repl,
11023     kmutex_t **low, kmutex_t **high)
11024 {
11025 	kmutex_t	*mml_targ, *mml_repl;
11026 
11027 	/*
11028 	 * no need to do the dance around szc as in sfmmu_mlist_enter()
11029 	 * because this routine is only called by hat_page_relocate() and all
11030 	 * targ and repl pages are already locked EXCL so szc can't change.
11031 	 */
11032 
11033 	mml_targ = MLIST_HASH(PP_PAGEROOT(targ));
11034 	mml_repl = MLIST_HASH(PP_PAGEROOT(repl));
11035 
11036 	if (mml_targ == mml_repl) {
11037 		*low = mml_targ;
11038 		*high = NULL;
11039 	} else {
11040 		if (mml_targ < mml_repl) {
11041 			*low = mml_targ;
11042 			*high = mml_repl;
11043 		} else {
11044 			*low = mml_repl;
11045 			*high = mml_targ;
11046 		}
11047 	}
11048 
11049 	mutex_enter(*low);
11050 	if (*high)
11051 		mutex_enter(*high);
11052 }
11053 
11054 static void
11055 sfmmu_mlist_reloc_exit(kmutex_t *low, kmutex_t *high)
11056 {
11057 	if (high)
11058 		mutex_exit(high);
11059 	mutex_exit(low);
11060 }
11061 
11062 static hatlock_t *
11063 sfmmu_hat_enter(sfmmu_t *sfmmup)
11064 {
11065 	hatlock_t	*hatlockp;
11066 
11067 	if (sfmmup != ksfmmup) {
11068 		hatlockp = TSB_HASH(sfmmup);
11069 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
11070 		return (hatlockp);
11071 	}
11072 	return (NULL);
11073 }
11074 
11075 static hatlock_t *
11076 sfmmu_hat_tryenter(sfmmu_t *sfmmup)
11077 {
11078 	hatlock_t	*hatlockp;
11079 
11080 	if (sfmmup != ksfmmup) {
11081 		hatlockp = TSB_HASH(sfmmup);
11082 		if (mutex_tryenter(HATLOCK_MUTEXP(hatlockp)) == 0)
11083 			return (NULL);
11084 		return (hatlockp);
11085 	}
11086 	return (NULL);
11087 }
11088 
11089 static void
11090 sfmmu_hat_exit(hatlock_t *hatlockp)
11091 {
11092 	if (hatlockp != NULL)
11093 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
11094 }
11095 
11096 static void
11097 sfmmu_hat_lock_all(void)
11098 {
11099 	int i;
11100 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
11101 		mutex_enter(HATLOCK_MUTEXP(&hat_lock[i]));
11102 }
11103 
11104 static void
11105 sfmmu_hat_unlock_all(void)
11106 {
11107 	int i;
11108 	for (i = SFMMU_NUM_LOCK - 1; i >= 0; i--)
11109 		mutex_exit(HATLOCK_MUTEXP(&hat_lock[i]));
11110 }
11111 
11112 int
11113 sfmmu_hat_lock_held(sfmmu_t *sfmmup)
11114 {
11115 	ASSERT(sfmmup != ksfmmup);
11116 	return (MUTEX_HELD(HATLOCK_MUTEXP(TSB_HASH(sfmmup))));
11117 }
11118 
11119 /*
11120  * Locking primitives to provide consistency between ISM unmap
11121  * and other operations.  Since ISM unmap can take a long time, we
11122  * use HAT_ISMBUSY flag (protected by the hatlock) to avoid creating
11123  * contention on the hatlock buckets while ISM segments are being
11124  * unmapped.  The tradeoff is that the flags don't prevent priority
11125  * inversion from occurring, so we must request kernel priority in
11126  * case we have to sleep to keep from getting buried while holding
11127  * the HAT_ISMBUSY flag set, which in turn could block other kernel
11128  * threads from running (for example, in sfmmu_uvatopfn()).
11129  */
11130 static void
11131 sfmmu_ismhat_enter(sfmmu_t *sfmmup, int hatlock_held)
11132 {
11133 	hatlock_t *hatlockp;
11134 
11135 	THREAD_KPRI_REQUEST();
11136 	if (!hatlock_held)
11137 		hatlockp = sfmmu_hat_enter(sfmmup);
11138 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY))
11139 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
11140 	SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
11141 	if (!hatlock_held)
11142 		sfmmu_hat_exit(hatlockp);
11143 }
11144 
11145 static void
11146 sfmmu_ismhat_exit(sfmmu_t *sfmmup, int hatlock_held)
11147 {
11148 	hatlock_t *hatlockp;
11149 
11150 	if (!hatlock_held)
11151 		hatlockp = sfmmu_hat_enter(sfmmup);
11152 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
11153 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
11154 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11155 	if (!hatlock_held)
11156 		sfmmu_hat_exit(hatlockp);
11157 	THREAD_KPRI_RELEASE();
11158 }
11159 
11160 /*
11161  *
11162  * Algorithm:
11163  *
11164  * (1) if segkmem is not ready, allocate hblk from an array of pre-alloc'ed
11165  *	hblks.
11166  *
11167  * (2) if we are allocating an hblk for mapping a slab in sfmmu_cache,
11168  *
11169  * 		(a) try to return an hblk from reserve pool of free hblks;
11170  *		(b) if the reserve pool is empty, acquire hblk_reserve_lock
11171  *		    and return hblk_reserve.
11172  *
11173  * (3) call kmem_cache_alloc() to allocate hblk;
11174  *
11175  *		(a) if hblk_reserve_lock is held by the current thread,
11176  *		    atomically replace hblk_reserve by the hblk that is
11177  *		    returned by kmem_cache_alloc; release hblk_reserve_lock
11178  *		    and call kmem_cache_alloc() again.
11179  *		(b) if reserve pool is not full, add the hblk that is
11180  *		    returned by kmem_cache_alloc to reserve pool and
11181  *		    call kmem_cache_alloc again.
11182  *
11183  */
11184 static struct hme_blk *
11185 sfmmu_hblk_alloc(sfmmu_t *sfmmup, caddr_t vaddr,
11186 	struct hmehash_bucket *hmebp, uint_t size, hmeblk_tag hblktag,
11187 	uint_t flags, uint_t rid)
11188 {
11189 	struct hme_blk *hmeblkp = NULL;
11190 	struct hme_blk *newhblkp;
11191 	struct hme_blk *shw_hblkp = NULL;
11192 	struct kmem_cache *sfmmu_cache = NULL;
11193 	uint64_t hblkpa;
11194 	ulong_t index;
11195 	uint_t owner;		/* set to 1 if using hblk_reserve */
11196 	uint_t forcefree;
11197 	int sleep;
11198 	sf_srd_t *srdp;
11199 	sf_region_t *rgnp;
11200 
11201 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11202 	ASSERT(hblktag.htag_rid == rid);
11203 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
11204 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
11205 	    IS_P2ALIGNED(vaddr, TTEBYTES(size)));
11206 
11207 	/*
11208 	 * If segkmem is not created yet, allocate from static hmeblks
11209 	 * created at the end of startup_modules().  See the block comment
11210 	 * in startup_modules() describing how we estimate the number of
11211 	 * static hmeblks that will be needed during re-map.
11212 	 */
11213 	if (!hblk_alloc_dynamic) {
11214 
11215 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11216 
11217 		if (size == TTE8K) {
11218 			index = nucleus_hblk8.index;
11219 			if (index >= nucleus_hblk8.len) {
11220 				/*
11221 				 * If we panic here, see startup_modules() to
11222 				 * make sure that we are calculating the
11223 				 * number of hblk8's that we need correctly.
11224 				 */
11225 				prom_panic("no nucleus hblk8 to allocate");
11226 			}
11227 			hmeblkp =
11228 			    (struct hme_blk *)&nucleus_hblk8.list[index];
11229 			nucleus_hblk8.index++;
11230 			SFMMU_STAT(sf_hblk8_nalloc);
11231 		} else {
11232 			index = nucleus_hblk1.index;
11233 			if (nucleus_hblk1.index >= nucleus_hblk1.len) {
11234 				/*
11235 				 * If we panic here, see startup_modules().
11236 				 * Most likely you need to update the
11237 				 * calculation of the number of hblk1 elements
11238 				 * that the kernel needs to boot.
11239 				 */
11240 				prom_panic("no nucleus hblk1 to allocate");
11241 			}
11242 			hmeblkp =
11243 			    (struct hme_blk *)&nucleus_hblk1.list[index];
11244 			nucleus_hblk1.index++;
11245 			SFMMU_STAT(sf_hblk1_nalloc);
11246 		}
11247 
11248 		goto hblk_init;
11249 	}
11250 
11251 	SFMMU_HASH_UNLOCK(hmebp);
11252 
11253 	if (sfmmup != KHATID && !SFMMU_IS_SHMERID_VALID(rid)) {
11254 		if (mmu_page_sizes == max_mmu_page_sizes) {
11255 			if (size < TTE256M)
11256 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
11257 				    size, flags);
11258 		} else {
11259 			if (size < TTE4M)
11260 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
11261 				    size, flags);
11262 		}
11263 	} else if (SFMMU_IS_SHMERID_VALID(rid)) {
11264 		/*
11265 		 * Shared hmes use per region bitmaps in rgn_hmeflag
11266 		 * rather than shadow hmeblks to keep track of the
11267 		 * mapping sizes which have been allocated for the region.
11268 		 * Here we cleanup old invalid hmeblks with this rid,
11269 		 * which may be left around by pageunload().
11270 		 */
11271 		int ttesz;
11272 		caddr_t va;
11273 		caddr_t	eva = vaddr + TTEBYTES(size);
11274 
11275 		ASSERT(sfmmup != KHATID);
11276 
11277 		srdp = sfmmup->sfmmu_srdp;
11278 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11279 		rgnp = srdp->srd_hmergnp[rid];
11280 		ASSERT(rgnp != NULL && rgnp->rgn_id == rid);
11281 		ASSERT(rgnp->rgn_refcnt != 0);
11282 		ASSERT(size <= rgnp->rgn_pgszc);
11283 
11284 		ttesz = HBLK_MIN_TTESZ;
11285 		do {
11286 			if (!(rgnp->rgn_hmeflags & (0x1 << ttesz))) {
11287 				continue;
11288 			}
11289 
11290 			if (ttesz > size && ttesz != HBLK_MIN_TTESZ) {
11291 				sfmmu_cleanup_rhblk(srdp, vaddr, rid, ttesz);
11292 			} else if (ttesz < size) {
11293 				for (va = vaddr; va < eva;
11294 				    va += TTEBYTES(ttesz)) {
11295 					sfmmu_cleanup_rhblk(srdp, va, rid,
11296 					    ttesz);
11297 				}
11298 			}
11299 		} while (++ttesz <= rgnp->rgn_pgszc);
11300 	}
11301 
11302 fill_hblk:
11303 	owner = (hblk_reserve_thread == curthread) ? 1 : 0;
11304 
11305 	if (owner && size == TTE8K) {
11306 
11307 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11308 		/*
11309 		 * We are really in a tight spot. We already own
11310 		 * hblk_reserve and we need another hblk.  In anticipation
11311 		 * of this kind of scenario, we specifically set aside
11312 		 * HBLK_RESERVE_MIN number of hblks to be used exclusively
11313 		 * by owner of hblk_reserve.
11314 		 */
11315 		SFMMU_STAT(sf_hblk_recurse_cnt);
11316 
11317 		if (!sfmmu_get_free_hblk(&hmeblkp, 1))
11318 			panic("sfmmu_hblk_alloc: reserve list is empty");
11319 
11320 		goto hblk_verify;
11321 	}
11322 
11323 	ASSERT(!owner);
11324 
11325 	if ((flags & HAT_NO_KALLOC) == 0) {
11326 
11327 		sfmmu_cache = ((size == TTE8K) ? sfmmu8_cache : sfmmu1_cache);
11328 		sleep = ((sfmmup == KHATID) ? KM_NOSLEEP : KM_SLEEP);
11329 
11330 		if ((hmeblkp = kmem_cache_alloc(sfmmu_cache, sleep)) == NULL) {
11331 			hmeblkp = sfmmu_hblk_steal(size);
11332 		} else {
11333 			/*
11334 			 * if we are the owner of hblk_reserve,
11335 			 * swap hblk_reserve with hmeblkp and
11336 			 * start a fresh life.  Hope things go
11337 			 * better this time.
11338 			 */
11339 			if (hblk_reserve_thread == curthread) {
11340 				ASSERT(sfmmu_cache == sfmmu8_cache);
11341 				sfmmu_hblk_swap(hmeblkp);
11342 				hblk_reserve_thread = NULL;
11343 				mutex_exit(&hblk_reserve_lock);
11344 				goto fill_hblk;
11345 			}
11346 			/*
11347 			 * let's donate this hblk to our reserve list if
11348 			 * we are not mapping kernel range
11349 			 */
11350 			if (size == TTE8K && sfmmup != KHATID) {
11351 				if (sfmmu_put_free_hblk(hmeblkp, 0))
11352 					goto fill_hblk;
11353 			}
11354 		}
11355 	} else {
11356 		/*
11357 		 * We are here to map the slab in sfmmu8_cache; let's
11358 		 * check if we could tap our reserve list; if successful,
11359 		 * this will avoid the pain of going thru sfmmu_hblk_swap
11360 		 */
11361 		SFMMU_STAT(sf_hblk_slab_cnt);
11362 		if (!sfmmu_get_free_hblk(&hmeblkp, 0)) {
11363 			/*
11364 			 * let's start hblk_reserve dance
11365 			 */
11366 			SFMMU_STAT(sf_hblk_reserve_cnt);
11367 			owner = 1;
11368 			mutex_enter(&hblk_reserve_lock);
11369 			hmeblkp = HBLK_RESERVE;
11370 			hblk_reserve_thread = curthread;
11371 		}
11372 	}
11373 
11374 hblk_verify:
11375 	ASSERT(hmeblkp != NULL);
11376 	set_hblk_sz(hmeblkp, size);
11377 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
11378 	SFMMU_HASH_LOCK(hmebp);
11379 	HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11380 	if (newhblkp != NULL) {
11381 		SFMMU_HASH_UNLOCK(hmebp);
11382 		if (hmeblkp != HBLK_RESERVE) {
11383 			/*
11384 			 * This is really tricky!
11385 			 *
11386 			 * vmem_alloc(vmem_seg_arena)
11387 			 *  vmem_alloc(vmem_internal_arena)
11388 			 *   segkmem_alloc(heap_arena)
11389 			 *    vmem_alloc(heap_arena)
11390 			 *    page_create()
11391 			 *    hat_memload()
11392 			 *	kmem_cache_free()
11393 			 *	 kmem_cache_alloc()
11394 			 *	  kmem_slab_create()
11395 			 *	   vmem_alloc(kmem_internal_arena)
11396 			 *	    segkmem_alloc(heap_arena)
11397 			 *		vmem_alloc(heap_arena)
11398 			 *		page_create()
11399 			 *		hat_memload()
11400 			 *		  kmem_cache_free()
11401 			 *		...
11402 			 *
11403 			 * Thus, hat_memload() could call kmem_cache_free
11404 			 * for enough number of times that we could easily
11405 			 * hit the bottom of the stack or run out of reserve
11406 			 * list of vmem_seg structs.  So, we must donate
11407 			 * this hblk to reserve list if it's allocated
11408 			 * from sfmmu8_cache *and* mapping kernel range.
11409 			 * We don't need to worry about freeing hmeblk1's
11410 			 * to kmem since they don't map any kmem slabs.
11411 			 *
11412 			 * Note: When segkmem supports largepages, we must
11413 			 * free hmeblk1's to reserve list as well.
11414 			 */
11415 			forcefree = (sfmmup == KHATID) ? 1 : 0;
11416 			if (size == TTE8K &&
11417 			    sfmmu_put_free_hblk(hmeblkp, forcefree)) {
11418 				goto re_verify;
11419 			}
11420 			ASSERT(sfmmup != KHATID);
11421 			kmem_cache_free(get_hblk_cache(hmeblkp), hmeblkp);
11422 		} else {
11423 			/*
11424 			 * Hey! we don't need hblk_reserve any more.
11425 			 */
11426 			ASSERT(owner);
11427 			hblk_reserve_thread = NULL;
11428 			mutex_exit(&hblk_reserve_lock);
11429 			owner = 0;
11430 		}
11431 re_verify:
11432 		/*
11433 		 * let's check if the goodies are still present
11434 		 */
11435 		SFMMU_HASH_LOCK(hmebp);
11436 		HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11437 		if (newhblkp != NULL) {
11438 			/*
11439 			 * return newhblkp if it's not hblk_reserve;
11440 			 * if newhblkp is hblk_reserve, return it
11441 			 * _only if_ we are the owner of hblk_reserve.
11442 			 */
11443 			if (newhblkp != HBLK_RESERVE || owner) {
11444 				ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
11445 				    newhblkp->hblk_shared);
11446 				ASSERT(SFMMU_IS_SHMERID_VALID(rid) ||
11447 				    !newhblkp->hblk_shared);
11448 				return (newhblkp);
11449 			} else {
11450 				/*
11451 				 * we just hit hblk_reserve in the hash and
11452 				 * we are not the owner of that;
11453 				 *
11454 				 * block until hblk_reserve_thread completes
11455 				 * swapping hblk_reserve and try the dance
11456 				 * once again.
11457 				 */
11458 				SFMMU_HASH_UNLOCK(hmebp);
11459 				mutex_enter(&hblk_reserve_lock);
11460 				mutex_exit(&hblk_reserve_lock);
11461 				SFMMU_STAT(sf_hblk_reserve_hit);
11462 				goto fill_hblk;
11463 			}
11464 		} else {
11465 			/*
11466 			 * it's no more! try the dance once again.
11467 			 */
11468 			SFMMU_HASH_UNLOCK(hmebp);
11469 			goto fill_hblk;
11470 		}
11471 	}
11472 
11473 hblk_init:
11474 	if (SFMMU_IS_SHMERID_VALID(rid)) {
11475 		uint16_t tteflag = 0x1 <<
11476 		    ((size < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ : size);
11477 
11478 		if (!(rgnp->rgn_hmeflags & tteflag)) {
11479 			atomic_or_16(&rgnp->rgn_hmeflags, tteflag);
11480 		}
11481 		hmeblkp->hblk_shared = 1;
11482 	} else {
11483 		hmeblkp->hblk_shared = 0;
11484 	}
11485 	set_hblk_sz(hmeblkp, size);
11486 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11487 	hmeblkp->hblk_next = (struct hme_blk *)NULL;
11488 	hmeblkp->hblk_tag = hblktag;
11489 	hmeblkp->hblk_shadow = shw_hblkp;
11490 	hblkpa = hmeblkp->hblk_nextpa;
11491 	hmeblkp->hblk_nextpa = HMEBLK_ENDPA;
11492 
11493 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
11494 	ASSERT(get_hblk_span(hmeblkp) == HMEBLK_SPAN(size));
11495 	ASSERT(hmeblkp->hblk_hmecnt == 0);
11496 	ASSERT(hmeblkp->hblk_vcnt == 0);
11497 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11498 	ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
11499 	sfmmu_hblk_hash_add(hmebp, hmeblkp, hblkpa);
11500 	return (hmeblkp);
11501 }
11502 
11503 /*
11504  * This function cleans up the hme_blk and returns it to the free list.
11505  */
11506 /* ARGSUSED */
11507 static void
11508 sfmmu_hblk_free(struct hme_blk **listp)
11509 {
11510 	struct hme_blk *hmeblkp, *next_hmeblkp;
11511 	int		size;
11512 	uint_t		critical;
11513 	uint64_t	hblkpa;
11514 
11515 	ASSERT(*listp != NULL);
11516 
11517 	hmeblkp = *listp;
11518 	while (hmeblkp != NULL) {
11519 		next_hmeblkp = hmeblkp->hblk_next;
11520 		ASSERT(!hmeblkp->hblk_hmecnt);
11521 		ASSERT(!hmeblkp->hblk_vcnt);
11522 		ASSERT(!hmeblkp->hblk_lckcnt);
11523 		ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
11524 		ASSERT(hmeblkp->hblk_shared == 0);
11525 		ASSERT(hmeblkp->hblk_shw_bit == 0);
11526 		ASSERT(hmeblkp->hblk_shadow == NULL);
11527 
11528 		hblkpa = va_to_pa((caddr_t)hmeblkp);
11529 		ASSERT(hblkpa != (uint64_t)-1);
11530 		critical = (hblktosfmmu(hmeblkp) == KHATID) ? 1 : 0;
11531 
11532 		size = get_hblk_ttesz(hmeblkp);
11533 		hmeblkp->hblk_next = NULL;
11534 		hmeblkp->hblk_nextpa = hblkpa;
11535 
11536 		if (hmeblkp->hblk_nuc_bit == 0) {
11537 
11538 			if (size != TTE8K ||
11539 			    !sfmmu_put_free_hblk(hmeblkp, critical))
11540 				kmem_cache_free(get_hblk_cache(hmeblkp),
11541 				    hmeblkp);
11542 		}
11543 		hmeblkp = next_hmeblkp;
11544 	}
11545 }
11546 
11547 #define	BUCKETS_TO_SEARCH_BEFORE_UNLOAD	30
11548 #define	SFMMU_HBLK_STEAL_THRESHOLD 5
11549 
11550 static uint_t sfmmu_hblk_steal_twice;
11551 static uint_t sfmmu_hblk_steal_count, sfmmu_hblk_steal_unload_count;
11552 
11553 /*
11554  * Steal a hmeblk from user or kernel hme hash lists.
11555  * For 8K tte grab one from reserve pool (freehblkp) before proceeding to
11556  * steal and if we fail to steal after SFMMU_HBLK_STEAL_THRESHOLD attempts
11557  * tap into critical reserve of freehblkp.
11558  * Note: We remain looping in this routine until we find one.
11559  */
11560 static struct hme_blk *
11561 sfmmu_hblk_steal(int size)
11562 {
11563 	static struct hmehash_bucket *uhmehash_steal_hand = NULL;
11564 	struct hmehash_bucket *hmebp;
11565 	struct hme_blk *hmeblkp = NULL, *pr_hblk;
11566 	uint64_t hblkpa;
11567 	int i;
11568 	uint_t loop_cnt = 0, critical;
11569 
11570 	for (;;) {
11571 		/* Check cpu hblk pending queues */
11572 		if ((hmeblkp = sfmmu_check_pending_hblks(size)) != NULL) {
11573 			hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
11574 			ASSERT(hmeblkp->hblk_hmecnt == 0);
11575 			ASSERT(hmeblkp->hblk_vcnt == 0);
11576 			return (hmeblkp);
11577 		}
11578 
11579 		if (size == TTE8K) {
11580 			critical =
11581 			    (++loop_cnt > SFMMU_HBLK_STEAL_THRESHOLD) ? 1 : 0;
11582 			if (sfmmu_get_free_hblk(&hmeblkp, critical))
11583 				return (hmeblkp);
11584 		}
11585 
11586 		hmebp = (uhmehash_steal_hand == NULL) ? uhme_hash :
11587 		    uhmehash_steal_hand;
11588 		ASSERT(hmebp >= uhme_hash && hmebp <= &uhme_hash[UHMEHASH_SZ]);
11589 
11590 		for (i = 0; hmeblkp == NULL && i <= UHMEHASH_SZ +
11591 		    BUCKETS_TO_SEARCH_BEFORE_UNLOAD; i++) {
11592 			SFMMU_HASH_LOCK(hmebp);
11593 			hmeblkp = hmebp->hmeblkp;
11594 			hblkpa = hmebp->hmeh_nextpa;
11595 			pr_hblk = NULL;
11596 			while (hmeblkp) {
11597 				/*
11598 				 * check if it is a hmeblk that is not locked
11599 				 * and not shared. skip shadow hmeblks with
11600 				 * shadow_mask set i.e valid count non zero.
11601 				 */
11602 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11603 				    (hmeblkp->hblk_shw_bit == 0 ||
11604 				    hmeblkp->hblk_vcnt == 0) &&
11605 				    (hmeblkp->hblk_lckcnt == 0)) {
11606 					/*
11607 					 * there is a high probability that we
11608 					 * will find a free one. search some
11609 					 * buckets for a free hmeblk initially
11610 					 * before unloading a valid hmeblk.
11611 					 */
11612 					if ((hmeblkp->hblk_vcnt == 0 &&
11613 					    hmeblkp->hblk_hmecnt == 0) || (i >=
11614 					    BUCKETS_TO_SEARCH_BEFORE_UNLOAD)) {
11615 						if (sfmmu_steal_this_hblk(hmebp,
11616 						    hmeblkp, hblkpa, pr_hblk)) {
11617 							/*
11618 							 * Hblk is unloaded
11619 							 * successfully
11620 							 */
11621 							break;
11622 						}
11623 					}
11624 				}
11625 				pr_hblk = hmeblkp;
11626 				hblkpa = hmeblkp->hblk_nextpa;
11627 				hmeblkp = hmeblkp->hblk_next;
11628 			}
11629 
11630 			SFMMU_HASH_UNLOCK(hmebp);
11631 			if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
11632 				hmebp = uhme_hash;
11633 		}
11634 		uhmehash_steal_hand = hmebp;
11635 
11636 		if (hmeblkp != NULL)
11637 			break;
11638 
11639 		/*
11640 		 * in the worst case, look for a free one in the kernel
11641 		 * hash table.
11642 		 */
11643 		for (i = 0, hmebp = khme_hash; i <= KHMEHASH_SZ; i++) {
11644 			SFMMU_HASH_LOCK(hmebp);
11645 			hmeblkp = hmebp->hmeblkp;
11646 			hblkpa = hmebp->hmeh_nextpa;
11647 			pr_hblk = NULL;
11648 			while (hmeblkp) {
11649 				/*
11650 				 * check if it is free hmeblk
11651 				 */
11652 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11653 				    (hmeblkp->hblk_lckcnt == 0) &&
11654 				    (hmeblkp->hblk_vcnt == 0) &&
11655 				    (hmeblkp->hblk_hmecnt == 0)) {
11656 					if (sfmmu_steal_this_hblk(hmebp,
11657 					    hmeblkp, hblkpa, pr_hblk)) {
11658 						break;
11659 					} else {
11660 						/*
11661 						 * Cannot fail since we have
11662 						 * hash lock.
11663 						 */
11664 						panic("fail to steal?");
11665 					}
11666 				}
11667 
11668 				pr_hblk = hmeblkp;
11669 				hblkpa = hmeblkp->hblk_nextpa;
11670 				hmeblkp = hmeblkp->hblk_next;
11671 			}
11672 
11673 			SFMMU_HASH_UNLOCK(hmebp);
11674 			if (hmebp++ == &khme_hash[KHMEHASH_SZ])
11675 				hmebp = khme_hash;
11676 		}
11677 
11678 		if (hmeblkp != NULL)
11679 			break;
11680 		sfmmu_hblk_steal_twice++;
11681 	}
11682 	return (hmeblkp);
11683 }
11684 
11685 /*
11686  * This routine does real work to prepare a hblk to be "stolen" by
11687  * unloading the mappings, updating shadow counts ....
11688  * It returns 1 if the block is ready to be reused (stolen), or 0
11689  * means the block cannot be stolen yet- pageunload is still working
11690  * on this hblk.
11691  */
11692 static int
11693 sfmmu_steal_this_hblk(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
11694 	uint64_t hblkpa, struct hme_blk *pr_hblk)
11695 {
11696 	int shw_size, vshift;
11697 	struct hme_blk *shw_hblkp;
11698 	caddr_t vaddr;
11699 	uint_t shw_mask, newshw_mask;
11700 	struct hme_blk *list = NULL;
11701 
11702 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11703 
11704 	/*
11705 	 * check if the hmeblk is free, unload if necessary
11706 	 */
11707 	if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11708 		sfmmu_t *sfmmup;
11709 		demap_range_t dmr;
11710 
11711 		sfmmup = hblktosfmmu(hmeblkp);
11712 		if (hmeblkp->hblk_shared || sfmmup->sfmmu_ismhat) {
11713 			return (0);
11714 		}
11715 		DEMAP_RANGE_INIT(sfmmup, &dmr);
11716 		(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
11717 		    (caddr_t)get_hblk_base(hmeblkp),
11718 		    get_hblk_endaddr(hmeblkp), &dmr, HAT_UNLOAD);
11719 		DEMAP_RANGE_FLUSH(&dmr);
11720 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11721 			/*
11722 			 * Pageunload is working on the same hblk.
11723 			 */
11724 			return (0);
11725 		}
11726 
11727 		sfmmu_hblk_steal_unload_count++;
11728 	}
11729 
11730 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11731 	ASSERT(hmeblkp->hblk_vcnt == 0 && hmeblkp->hblk_hmecnt == 0);
11732 
11733 	sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 1);
11734 	hmeblkp->hblk_nextpa = hblkpa;
11735 
11736 	shw_hblkp = hmeblkp->hblk_shadow;
11737 	if (shw_hblkp) {
11738 		ASSERT(!hmeblkp->hblk_shared);
11739 		shw_size = get_hblk_ttesz(shw_hblkp);
11740 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
11741 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
11742 		ASSERT(vshift < 8);
11743 		/*
11744 		 * Atomically clear shadow mask bit
11745 		 */
11746 		do {
11747 			shw_mask = shw_hblkp->hblk_shw_mask;
11748 			ASSERT(shw_mask & (1 << vshift));
11749 			newshw_mask = shw_mask & ~(1 << vshift);
11750 			newshw_mask = cas32(&shw_hblkp->hblk_shw_mask,
11751 			    shw_mask, newshw_mask);
11752 		} while (newshw_mask != shw_mask);
11753 		hmeblkp->hblk_shadow = NULL;
11754 	}
11755 
11756 	/*
11757 	 * remove shadow bit if we are stealing an unused shadow hmeblk.
11758 	 * sfmmu_hblk_alloc needs it that way, will set shadow bit later if
11759 	 * we are indeed allocating a shadow hmeblk.
11760 	 */
11761 	hmeblkp->hblk_shw_bit = 0;
11762 
11763 	if (hmeblkp->hblk_shared) {
11764 		sf_srd_t	*srdp;
11765 		sf_region_t	*rgnp;
11766 		uint_t		rid;
11767 
11768 		srdp = hblktosrd(hmeblkp);
11769 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11770 		rid = hmeblkp->hblk_tag.htag_rid;
11771 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
11772 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
11773 		rgnp = srdp->srd_hmergnp[rid];
11774 		ASSERT(rgnp != NULL);
11775 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
11776 		hmeblkp->hblk_shared = 0;
11777 	}
11778 
11779 	sfmmu_hblk_steal_count++;
11780 	SFMMU_STAT(sf_steal_count);
11781 
11782 	return (1);
11783 }
11784 
11785 struct hme_blk *
11786 sfmmu_hmetohblk(struct sf_hment *sfhme)
11787 {
11788 	struct hme_blk *hmeblkp;
11789 	struct sf_hment *sfhme0;
11790 	struct hme_blk *hblk_dummy = 0;
11791 
11792 	/*
11793 	 * No dummy sf_hments, please.
11794 	 */
11795 	ASSERT(sfhme->hme_tte.ll != 0);
11796 
11797 	sfhme0 = sfhme - sfhme->hme_tte.tte_hmenum;
11798 	hmeblkp = (struct hme_blk *)((uintptr_t)sfhme0 -
11799 	    (uintptr_t)&hblk_dummy->hblk_hme[0]);
11800 
11801 	return (hmeblkp);
11802 }
11803 
11804 /*
11805  * On swapin, get appropriately sized TSB(s) and clear the HAT_SWAPPED flag.
11806  * If we can't get appropriately sized TSB(s), try for 8K TSB(s) using
11807  * KM_SLEEP allocation.
11808  *
11809  * Return 0 on success, -1 otherwise.
11810  */
11811 static void
11812 sfmmu_tsb_swapin(sfmmu_t *sfmmup, hatlock_t *hatlockp)
11813 {
11814 	struct tsb_info *tsbinfop, *next;
11815 	tsb_replace_rc_t rc;
11816 	boolean_t gotfirst = B_FALSE;
11817 
11818 	ASSERT(sfmmup != ksfmmup);
11819 	ASSERT(sfmmu_hat_lock_held(sfmmup));
11820 
11821 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPIN)) {
11822 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
11823 	}
11824 
11825 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
11826 		SFMMU_FLAGS_SET(sfmmup, HAT_SWAPIN);
11827 	} else {
11828 		return;
11829 	}
11830 
11831 	ASSERT(sfmmup->sfmmu_tsb != NULL);
11832 
11833 	/*
11834 	 * Loop over all tsbinfo's replacing them with ones that actually have
11835 	 * a TSB.  If any of the replacements ever fail, bail out of the loop.
11836 	 */
11837 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; tsbinfop = next) {
11838 		ASSERT(tsbinfop->tsb_flags & TSB_SWAPPED);
11839 		next = tsbinfop->tsb_next;
11840 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, tsbinfop->tsb_szc,
11841 		    hatlockp, TSB_SWAPIN);
11842 		if (rc != TSB_SUCCESS) {
11843 			break;
11844 		}
11845 		gotfirst = B_TRUE;
11846 	}
11847 
11848 	switch (rc) {
11849 	case TSB_SUCCESS:
11850 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11851 		cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11852 		return;
11853 	case TSB_LOSTRACE:
11854 		break;
11855 	case TSB_ALLOCFAIL:
11856 		break;
11857 	default:
11858 		panic("sfmmu_replace_tsb returned unrecognized failure code "
11859 		    "%d", rc);
11860 	}
11861 
11862 	/*
11863 	 * In this case, we failed to get one of our TSBs.  If we failed to
11864 	 * get the first TSB, get one of minimum size (8KB).  Walk the list
11865 	 * and throw away the tsbinfos, starting where the allocation failed;
11866 	 * we can get by with just one TSB as long as we don't leave the
11867 	 * SWAPPED tsbinfo structures lying around.
11868 	 */
11869 	tsbinfop = sfmmup->sfmmu_tsb;
11870 	next = tsbinfop->tsb_next;
11871 	tsbinfop->tsb_next = NULL;
11872 
11873 	sfmmu_hat_exit(hatlockp);
11874 	for (tsbinfop = next; tsbinfop != NULL; tsbinfop = next) {
11875 		next = tsbinfop->tsb_next;
11876 		sfmmu_tsbinfo_free(tsbinfop);
11877 	}
11878 	hatlockp = sfmmu_hat_enter(sfmmup);
11879 
11880 	/*
11881 	 * If we don't have any TSBs, get a single 8K TSB for 8K, 64K and 512K
11882 	 * pages.
11883 	 */
11884 	if (!gotfirst) {
11885 		tsbinfop = sfmmup->sfmmu_tsb;
11886 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, TSB_MIN_SZCODE,
11887 		    hatlockp, TSB_SWAPIN | TSB_FORCEALLOC);
11888 		ASSERT(rc == TSB_SUCCESS);
11889 	}
11890 
11891 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11892 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11893 }
11894 
11895 static int
11896 sfmmu_is_rgnva(sf_srd_t *srdp, caddr_t addr, ulong_t w, ulong_t bmw)
11897 {
11898 	ulong_t bix = 0;
11899 	uint_t rid;
11900 	sf_region_t *rgnp;
11901 
11902 	ASSERT(srdp != NULL);
11903 	ASSERT(srdp->srd_refcnt != 0);
11904 
11905 	w <<= BT_ULSHIFT;
11906 	while (bmw) {
11907 		if (!(bmw & 0x1)) {
11908 			bix++;
11909 			bmw >>= 1;
11910 			continue;
11911 		}
11912 		rid = w | bix;
11913 		rgnp = srdp->srd_hmergnp[rid];
11914 		ASSERT(rgnp->rgn_refcnt > 0);
11915 		ASSERT(rgnp->rgn_id == rid);
11916 		if (addr < rgnp->rgn_saddr ||
11917 		    addr >= (rgnp->rgn_saddr + rgnp->rgn_size)) {
11918 			bix++;
11919 			bmw >>= 1;
11920 		} else {
11921 			return (1);
11922 		}
11923 	}
11924 	return (0);
11925 }
11926 
11927 /*
11928  * Handle exceptions for low level tsb_handler.
11929  *
11930  * There are many scenarios that could land us here:
11931  *
11932  * If the context is invalid we land here. The context can be invalid
11933  * for 3 reasons: 1) we couldn't allocate a new context and now need to
11934  * perform a wrap around operation in order to allocate a new context.
11935  * 2) Context was invalidated to change pagesize programming 3) ISMs or
11936  * TSBs configuration is changeing for this process and we are forced into
11937  * here to do a syncronization operation. If the context is valid we can
11938  * be here from window trap hanlder. In this case just call trap to handle
11939  * the fault.
11940  *
11941  * Note that the process will run in INVALID_CONTEXT before
11942  * faulting into here and subsequently loading the MMU registers
11943  * (including the TSB base register) associated with this process.
11944  * For this reason, the trap handlers must all test for
11945  * INVALID_CONTEXT before attempting to access any registers other
11946  * than the context registers.
11947  */
11948 void
11949 sfmmu_tsbmiss_exception(struct regs *rp, uintptr_t tagaccess, uint_t traptype)
11950 {
11951 	sfmmu_t *sfmmup, *shsfmmup;
11952 	uint_t ctxtype;
11953 	klwp_id_t lwp;
11954 	char lwp_save_state;
11955 	hatlock_t *hatlockp, *shatlockp;
11956 	struct tsb_info *tsbinfop;
11957 	struct tsbmiss *tsbmp;
11958 	sf_scd_t *scdp;
11959 
11960 	SFMMU_STAT(sf_tsb_exceptions);
11961 	SFMMU_MMU_STAT(mmu_tsb_exceptions);
11962 	sfmmup = astosfmmu(curthread->t_procp->p_as);
11963 	/*
11964 	 * note that in sun4u, tagacces register contains ctxnum
11965 	 * while sun4v passes ctxtype in the tagaccess register.
11966 	 */
11967 	ctxtype = tagaccess & TAGACC_CTX_MASK;
11968 
11969 	ASSERT(sfmmup != ksfmmup && ctxtype != KCONTEXT);
11970 	ASSERT(sfmmup->sfmmu_ismhat == 0);
11971 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED) ||
11972 	    ctxtype == INVALID_CONTEXT);
11973 
11974 	if (ctxtype != INVALID_CONTEXT && traptype != T_DATA_PROT) {
11975 		/*
11976 		 * We may land here because shme bitmap and pagesize
11977 		 * flags are updated lazily in tsbmiss area on other cpus.
11978 		 * If we detect here that tsbmiss area is out of sync with
11979 		 * sfmmu update it and retry the trapped instruction.
11980 		 * Otherwise call trap().
11981 		 */
11982 		int ret = 0;
11983 		uchar_t tteflag_mask = (1 << TTE64K) | (1 << TTE8K);
11984 		caddr_t addr = (caddr_t)(tagaccess & TAGACC_VADDR_MASK);
11985 
11986 		/*
11987 		 * Must set lwp state to LWP_SYS before
11988 		 * trying to acquire any adaptive lock
11989 		 */
11990 		lwp = ttolwp(curthread);
11991 		ASSERT(lwp);
11992 		lwp_save_state = lwp->lwp_state;
11993 		lwp->lwp_state = LWP_SYS;
11994 
11995 		hatlockp = sfmmu_hat_enter(sfmmup);
11996 		kpreempt_disable();
11997 		tsbmp = &tsbmiss_area[CPU->cpu_id];
11998 		ASSERT(sfmmup == tsbmp->usfmmup);
11999 		if (((tsbmp->uhat_tteflags ^ sfmmup->sfmmu_tteflags) &
12000 		    ~tteflag_mask) ||
12001 		    ((tsbmp->uhat_rtteflags ^  sfmmup->sfmmu_rtteflags) &
12002 		    ~tteflag_mask)) {
12003 			tsbmp->uhat_tteflags = sfmmup->sfmmu_tteflags;
12004 			tsbmp->uhat_rtteflags = sfmmup->sfmmu_rtteflags;
12005 			ret = 1;
12006 		}
12007 		if (sfmmup->sfmmu_srdp != NULL) {
12008 			ulong_t *sm = sfmmup->sfmmu_hmeregion_map.bitmap;
12009 			ulong_t *tm = tsbmp->shmermap;
12010 			ulong_t i;
12011 			for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
12012 				ulong_t d = tm[i] ^ sm[i];
12013 				if (d) {
12014 					if (d & sm[i]) {
12015 						if (!ret && sfmmu_is_rgnva(
12016 						    sfmmup->sfmmu_srdp,
12017 						    addr, i, d & sm[i])) {
12018 							ret = 1;
12019 						}
12020 					}
12021 					tm[i] = sm[i];
12022 				}
12023 			}
12024 		}
12025 		kpreempt_enable();
12026 		sfmmu_hat_exit(hatlockp);
12027 		lwp->lwp_state = lwp_save_state;
12028 		if (ret) {
12029 			return;
12030 		}
12031 	} else if (ctxtype == INVALID_CONTEXT) {
12032 		/*
12033 		 * First, make sure we come out of here with a valid ctx,
12034 		 * since if we don't get one we'll simply loop on the
12035 		 * faulting instruction.
12036 		 *
12037 		 * If the ISM mappings are changing, the TSB is relocated,
12038 		 * the process is swapped, the process is joining SCD or
12039 		 * leaving SCD or shared regions we serialize behind the
12040 		 * controlling thread with hat lock, sfmmu_flags and
12041 		 * sfmmu_tsb_cv condition variable.
12042 		 */
12043 
12044 		/*
12045 		 * Must set lwp state to LWP_SYS before
12046 		 * trying to acquire any adaptive lock
12047 		 */
12048 		lwp = ttolwp(curthread);
12049 		ASSERT(lwp);
12050 		lwp_save_state = lwp->lwp_state;
12051 		lwp->lwp_state = LWP_SYS;
12052 
12053 		hatlockp = sfmmu_hat_enter(sfmmup);
12054 retry:
12055 		if ((scdp = sfmmup->sfmmu_scdp) != NULL) {
12056 			shsfmmup = scdp->scd_sfmmup;
12057 			ASSERT(shsfmmup != NULL);
12058 
12059 			for (tsbinfop = shsfmmup->sfmmu_tsb; tsbinfop != NULL;
12060 			    tsbinfop = tsbinfop->tsb_next) {
12061 				if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
12062 					/* drop the private hat lock */
12063 					sfmmu_hat_exit(hatlockp);
12064 					/* acquire the shared hat lock */
12065 					shatlockp = sfmmu_hat_enter(shsfmmup);
12066 					/*
12067 					 * recheck to see if anything changed
12068 					 * after we drop the private hat lock.
12069 					 */
12070 					if (sfmmup->sfmmu_scdp == scdp &&
12071 					    shsfmmup == scdp->scd_sfmmup) {
12072 						sfmmu_tsb_chk_reloc(shsfmmup,
12073 						    shatlockp);
12074 					}
12075 					sfmmu_hat_exit(shatlockp);
12076 					hatlockp = sfmmu_hat_enter(sfmmup);
12077 					goto retry;
12078 				}
12079 			}
12080 		}
12081 
12082 		for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
12083 		    tsbinfop = tsbinfop->tsb_next) {
12084 			if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
12085 				cv_wait(&sfmmup->sfmmu_tsb_cv,
12086 				    HATLOCK_MUTEXP(hatlockp));
12087 				goto retry;
12088 			}
12089 		}
12090 
12091 		/*
12092 		 * Wait for ISM maps to be updated.
12093 		 */
12094 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
12095 			cv_wait(&sfmmup->sfmmu_tsb_cv,
12096 			    HATLOCK_MUTEXP(hatlockp));
12097 			goto retry;
12098 		}
12099 
12100 		/* Is this process joining an SCD? */
12101 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
12102 			/*
12103 			 * Flush private TSB and setup shared TSB.
12104 			 * sfmmu_finish_join_scd() does not drop the
12105 			 * hat lock.
12106 			 */
12107 			sfmmu_finish_join_scd(sfmmup);
12108 			SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
12109 		}
12110 
12111 		/*
12112 		 * If we're swapping in, get TSB(s).  Note that we must do
12113 		 * this before we get a ctx or load the MMU state.  Once
12114 		 * we swap in we have to recheck to make sure the TSB(s) and
12115 		 * ISM mappings didn't change while we slept.
12116 		 */
12117 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
12118 			sfmmu_tsb_swapin(sfmmup, hatlockp);
12119 			goto retry;
12120 		}
12121 
12122 		sfmmu_get_ctx(sfmmup);
12123 
12124 		sfmmu_hat_exit(hatlockp);
12125 		/*
12126 		 * Must restore lwp_state if not calling
12127 		 * trap() for further processing. Restore
12128 		 * it anyway.
12129 		 */
12130 		lwp->lwp_state = lwp_save_state;
12131 		return;
12132 	}
12133 	trap(rp, (caddr_t)tagaccess, traptype, 0);
12134 }
12135 
12136 static void
12137 sfmmu_tsb_chk_reloc(sfmmu_t *sfmmup, hatlock_t *hatlockp)
12138 {
12139 	struct tsb_info *tp;
12140 
12141 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12142 
12143 	for (tp = sfmmup->sfmmu_tsb; tp != NULL; tp = tp->tsb_next) {
12144 		if (tp->tsb_flags & TSB_RELOC_FLAG) {
12145 			cv_wait(&sfmmup->sfmmu_tsb_cv,
12146 			    HATLOCK_MUTEXP(hatlockp));
12147 			break;
12148 		}
12149 	}
12150 }
12151 
12152 /*
12153  * sfmmu_vatopfn_suspended is called from GET_TTE when TL=0 and
12154  * TTE_SUSPENDED bit set in tte we block on aquiring a page lock
12155  * rather than spinning to avoid send mondo timeouts with
12156  * interrupts enabled. When the lock is acquired it is immediately
12157  * released and we return back to sfmmu_vatopfn just after
12158  * the GET_TTE call.
12159  */
12160 void
12161 sfmmu_vatopfn_suspended(caddr_t vaddr, sfmmu_t *sfmmu, tte_t *ttep)
12162 {
12163 	struct page	**pp;
12164 
12165 	(void) as_pagelock(sfmmu->sfmmu_as, &pp, vaddr, TTE_CSZ(ttep), S_WRITE);
12166 	as_pageunlock(sfmmu->sfmmu_as, pp, vaddr, TTE_CSZ(ttep), S_WRITE);
12167 }
12168 
12169 /*
12170  * sfmmu_tsbmiss_suspended is called from GET_TTE when TL>0 and
12171  * TTE_SUSPENDED bit set in tte. We do this so that we can handle
12172  * cross traps which cannot be handled while spinning in the
12173  * trap handlers. Simply enter and exit the kpr_suspendlock spin
12174  * mutex, which is held by the holder of the suspend bit, and then
12175  * retry the trapped instruction after unwinding.
12176  */
12177 /*ARGSUSED*/
12178 void
12179 sfmmu_tsbmiss_suspended(struct regs *rp, uintptr_t tagacc, uint_t traptype)
12180 {
12181 	ASSERT(curthread != kreloc_thread);
12182 	mutex_enter(&kpr_suspendlock);
12183 	mutex_exit(&kpr_suspendlock);
12184 }
12185 
12186 /*
12187  * This routine could be optimized to reduce the number of xcalls by flushing
12188  * the entire TLBs if region reference count is above some threshold but the
12189  * tradeoff will depend on the size of the TLB. So for now flush the specific
12190  * page a context at a time.
12191  *
12192  * If uselocks is 0 then it's called after all cpus were captured and all the
12193  * hat locks were taken. In this case don't take the region lock by relying on
12194  * the order of list region update operations in hat_join_region(),
12195  * hat_leave_region() and hat_dup_region(). The ordering in those routines
12196  * guarantees that list is always forward walkable and reaches active sfmmus
12197  * regardless of where xc_attention() captures a cpu.
12198  */
12199 cpuset_t
12200 sfmmu_rgntlb_demap(caddr_t addr, sf_region_t *rgnp,
12201     struct hme_blk *hmeblkp, int uselocks)
12202 {
12203 	sfmmu_t	*sfmmup;
12204 	cpuset_t cpuset;
12205 	cpuset_t rcpuset;
12206 	hatlock_t *hatlockp;
12207 	uint_t rid = rgnp->rgn_id;
12208 	sf_rgn_link_t *rlink;
12209 	sf_scd_t *scdp;
12210 
12211 	ASSERT(hmeblkp->hblk_shared);
12212 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
12213 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
12214 
12215 	CPUSET_ZERO(rcpuset);
12216 	if (uselocks) {
12217 		mutex_enter(&rgnp->rgn_mutex);
12218 	}
12219 	sfmmup = rgnp->rgn_sfmmu_head;
12220 	while (sfmmup != NULL) {
12221 		if (uselocks) {
12222 			hatlockp = sfmmu_hat_enter(sfmmup);
12223 		}
12224 
12225 		/*
12226 		 * When an SCD is created the SCD hat is linked on the sfmmu
12227 		 * region lists for each hme region which is part of the
12228 		 * SCD. If we find an SCD hat, when walking these lists,
12229 		 * then we flush the shared TSBs, if we find a private hat,
12230 		 * which is part of an SCD, but where the region
12231 		 * is not part of the SCD then we flush the private TSBs.
12232 		 */
12233 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12234 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
12235 			scdp = sfmmup->sfmmu_scdp;
12236 			if (SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
12237 				if (uselocks) {
12238 					sfmmu_hat_exit(hatlockp);
12239 				}
12240 				goto next;
12241 			}
12242 		}
12243 
12244 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12245 
12246 		kpreempt_disable();
12247 		cpuset = sfmmup->sfmmu_cpusran;
12248 		CPUSET_AND(cpuset, cpu_ready_set);
12249 		CPUSET_DEL(cpuset, CPU->cpu_id);
12250 		SFMMU_XCALL_STATS(sfmmup);
12251 		xt_some(cpuset, vtag_flushpage_tl1,
12252 		    (uint64_t)addr, (uint64_t)sfmmup);
12253 		vtag_flushpage(addr, (uint64_t)sfmmup);
12254 		if (uselocks) {
12255 			sfmmu_hat_exit(hatlockp);
12256 		}
12257 		kpreempt_enable();
12258 		CPUSET_OR(rcpuset, cpuset);
12259 
12260 next:
12261 		/* LINTED: constant in conditional context */
12262 		SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
12263 		ASSERT(rlink != NULL);
12264 		sfmmup = rlink->next;
12265 	}
12266 	if (uselocks) {
12267 		mutex_exit(&rgnp->rgn_mutex);
12268 	}
12269 	return (rcpuset);
12270 }
12271 
12272 /*
12273  * This routine takes an sfmmu pointer and the va for an adddress in an
12274  * ISM region as input and returns the corresponding region id in ism_rid.
12275  * The return value of 1 indicates that a region has been found and ism_rid
12276  * is valid, otherwise 0 is returned.
12277  */
12278 static int
12279 find_ism_rid(sfmmu_t *sfmmup, sfmmu_t *ism_sfmmup, caddr_t va, uint_t *ism_rid)
12280 {
12281 	ism_blk_t	*ism_blkp;
12282 	int		i;
12283 	ism_map_t	*ism_map;
12284 #ifdef DEBUG
12285 	struct hat	*ism_hatid;
12286 #endif
12287 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12288 
12289 	ism_blkp = sfmmup->sfmmu_iblk;
12290 	while (ism_blkp != NULL) {
12291 		ism_map = ism_blkp->iblk_maps;
12292 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
12293 			if ((va >= ism_start(ism_map[i])) &&
12294 			    (va < ism_end(ism_map[i]))) {
12295 
12296 				*ism_rid = ism_map[i].imap_rid;
12297 #ifdef DEBUG
12298 				ism_hatid = ism_map[i].imap_ismhat;
12299 				ASSERT(ism_hatid == ism_sfmmup);
12300 				ASSERT(ism_hatid->sfmmu_ismhat);
12301 #endif
12302 				return (1);
12303 			}
12304 		}
12305 		ism_blkp = ism_blkp->iblk_next;
12306 	}
12307 	return (0);
12308 }
12309 
12310 /*
12311  * Special routine to flush out ism mappings- TSBs, TLBs and D-caches.
12312  * This routine may be called with all cpu's captured. Therefore, the
12313  * caller is responsible for holding all locks and disabling kernel
12314  * preemption.
12315  */
12316 /* ARGSUSED */
12317 static void
12318 sfmmu_ismtlbcache_demap(caddr_t addr, sfmmu_t *ism_sfmmup,
12319 	struct hme_blk *hmeblkp, pfn_t pfnum, int cache_flush_flag)
12320 {
12321 	cpuset_t 	cpuset;
12322 	caddr_t 	va;
12323 	ism_ment_t	*ment;
12324 	sfmmu_t		*sfmmup;
12325 #ifdef VAC
12326 	int 		vcolor;
12327 #endif
12328 
12329 	sf_scd_t	*scdp;
12330 	uint_t		ism_rid;
12331 
12332 	ASSERT(!hmeblkp->hblk_shared);
12333 	/*
12334 	 * Walk the ism_hat's mapping list and flush the page
12335 	 * from every hat sharing this ism_hat. This routine
12336 	 * may be called while all cpu's have been captured.
12337 	 * Therefore we can't attempt to grab any locks. For now
12338 	 * this means we will protect the ism mapping list under
12339 	 * a single lock which will be grabbed by the caller.
12340 	 * If hat_share/unshare scalibility becomes a performance
12341 	 * problem then we may need to re-think ism mapping list locking.
12342 	 */
12343 	ASSERT(ism_sfmmup->sfmmu_ismhat);
12344 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
12345 	addr = addr - ISMID_STARTADDR;
12346 
12347 	for (ment = ism_sfmmup->sfmmu_iment; ment; ment = ment->iment_next) {
12348 
12349 		sfmmup = ment->iment_hat;
12350 
12351 		va = ment->iment_base_va;
12352 		va = (caddr_t)((uintptr_t)va  + (uintptr_t)addr);
12353 
12354 		/*
12355 		 * When an SCD is created the SCD hat is linked on the ism
12356 		 * mapping lists for each ISM segment which is part of the
12357 		 * SCD. If we find an SCD hat, when walking these lists,
12358 		 * then we flush the shared TSBs, if we find a private hat,
12359 		 * which is part of an SCD, but where the region
12360 		 * corresponding to this va is not part of the SCD then we
12361 		 * flush the private TSBs.
12362 		 */
12363 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12364 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD) &&
12365 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
12366 			if (!find_ism_rid(sfmmup, ism_sfmmup, va,
12367 			    &ism_rid)) {
12368 				cmn_err(CE_PANIC,
12369 				    "can't find matching ISM rid!");
12370 			}
12371 
12372 			scdp = sfmmup->sfmmu_scdp;
12373 			if (SFMMU_IS_ISMRID_VALID(ism_rid) &&
12374 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map,
12375 			    ism_rid)) {
12376 				continue;
12377 			}
12378 		}
12379 		SFMMU_UNLOAD_TSB(va, sfmmup, hmeblkp, 1);
12380 
12381 		cpuset = sfmmup->sfmmu_cpusran;
12382 		CPUSET_AND(cpuset, cpu_ready_set);
12383 		CPUSET_DEL(cpuset, CPU->cpu_id);
12384 		SFMMU_XCALL_STATS(sfmmup);
12385 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)va,
12386 		    (uint64_t)sfmmup);
12387 		vtag_flushpage(va, (uint64_t)sfmmup);
12388 
12389 #ifdef VAC
12390 		/*
12391 		 * Flush D$
12392 		 * When flushing D$ we must flush all
12393 		 * cpu's. See sfmmu_cache_flush().
12394 		 */
12395 		if (cache_flush_flag == CACHE_FLUSH) {
12396 			cpuset = cpu_ready_set;
12397 			CPUSET_DEL(cpuset, CPU->cpu_id);
12398 
12399 			SFMMU_XCALL_STATS(sfmmup);
12400 			vcolor = addr_to_vcolor(va);
12401 			xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12402 			vac_flushpage(pfnum, vcolor);
12403 		}
12404 #endif	/* VAC */
12405 	}
12406 }
12407 
12408 /*
12409  * Demaps the TSB, CPU caches, and flushes all TLBs on all CPUs of
12410  * a particular virtual address and ctx.  If noflush is set we do not
12411  * flush the TLB/TSB.  This function may or may not be called with the
12412  * HAT lock held.
12413  */
12414 static void
12415 sfmmu_tlbcache_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12416 	pfn_t pfnum, int tlb_noflush, int cpu_flag, int cache_flush_flag,
12417 	int hat_lock_held)
12418 {
12419 #ifdef VAC
12420 	int vcolor;
12421 #endif
12422 	cpuset_t cpuset;
12423 	hatlock_t *hatlockp;
12424 
12425 	ASSERT(!hmeblkp->hblk_shared);
12426 
12427 #if defined(lint) && !defined(VAC)
12428 	pfnum = pfnum;
12429 	cpu_flag = cpu_flag;
12430 	cache_flush_flag = cache_flush_flag;
12431 #endif
12432 
12433 	/*
12434 	 * There is no longer a need to protect against ctx being
12435 	 * stolen here since we don't store the ctx in the TSB anymore.
12436 	 */
12437 #ifdef VAC
12438 	vcolor = addr_to_vcolor(addr);
12439 #endif
12440 
12441 	/*
12442 	 * We must hold the hat lock during the flush of TLB,
12443 	 * to avoid a race with sfmmu_invalidate_ctx(), where
12444 	 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12445 	 * causing TLB demap routine to skip flush on that MMU.
12446 	 * If the context on a MMU has already been set to
12447 	 * INVALID_CONTEXT, we just get an extra flush on
12448 	 * that MMU.
12449 	 */
12450 	if (!hat_lock_held && !tlb_noflush)
12451 		hatlockp = sfmmu_hat_enter(sfmmup);
12452 
12453 	kpreempt_disable();
12454 	if (!tlb_noflush) {
12455 		/*
12456 		 * Flush the TSB and TLB.
12457 		 */
12458 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12459 
12460 		cpuset = sfmmup->sfmmu_cpusran;
12461 		CPUSET_AND(cpuset, cpu_ready_set);
12462 		CPUSET_DEL(cpuset, CPU->cpu_id);
12463 
12464 		SFMMU_XCALL_STATS(sfmmup);
12465 
12466 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
12467 		    (uint64_t)sfmmup);
12468 
12469 		vtag_flushpage(addr, (uint64_t)sfmmup);
12470 	}
12471 
12472 	if (!hat_lock_held && !tlb_noflush)
12473 		sfmmu_hat_exit(hatlockp);
12474 
12475 #ifdef VAC
12476 	/*
12477 	 * Flush the D$
12478 	 *
12479 	 * Even if the ctx is stolen, we need to flush the
12480 	 * cache. Our ctx stealer only flushes the TLBs.
12481 	 */
12482 	if (cache_flush_flag == CACHE_FLUSH) {
12483 		if (cpu_flag & FLUSH_ALL_CPUS) {
12484 			cpuset = cpu_ready_set;
12485 		} else {
12486 			cpuset = sfmmup->sfmmu_cpusran;
12487 			CPUSET_AND(cpuset, cpu_ready_set);
12488 		}
12489 		CPUSET_DEL(cpuset, CPU->cpu_id);
12490 		SFMMU_XCALL_STATS(sfmmup);
12491 		xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12492 		vac_flushpage(pfnum, vcolor);
12493 	}
12494 #endif	/* VAC */
12495 	kpreempt_enable();
12496 }
12497 
12498 /*
12499  * Demaps the TSB and flushes all TLBs on all cpus for a particular virtual
12500  * address and ctx.  If noflush is set we do not currently do anything.
12501  * This function may or may not be called with the HAT lock held.
12502  */
12503 static void
12504 sfmmu_tlb_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12505 	int tlb_noflush, int hat_lock_held)
12506 {
12507 	cpuset_t cpuset;
12508 	hatlock_t *hatlockp;
12509 
12510 	ASSERT(!hmeblkp->hblk_shared);
12511 
12512 	/*
12513 	 * If the process is exiting we have nothing to do.
12514 	 */
12515 	if (tlb_noflush)
12516 		return;
12517 
12518 	/*
12519 	 * Flush TSB.
12520 	 */
12521 	if (!hat_lock_held)
12522 		hatlockp = sfmmu_hat_enter(sfmmup);
12523 	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12524 
12525 	kpreempt_disable();
12526 
12527 	cpuset = sfmmup->sfmmu_cpusran;
12528 	CPUSET_AND(cpuset, cpu_ready_set);
12529 	CPUSET_DEL(cpuset, CPU->cpu_id);
12530 
12531 	SFMMU_XCALL_STATS(sfmmup);
12532 	xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, (uint64_t)sfmmup);
12533 
12534 	vtag_flushpage(addr, (uint64_t)sfmmup);
12535 
12536 	if (!hat_lock_held)
12537 		sfmmu_hat_exit(hatlockp);
12538 
12539 	kpreempt_enable();
12540 
12541 }
12542 
12543 /*
12544  * Special case of sfmmu_tlb_demap for MMU_PAGESIZE hblks. Use the xcall
12545  * call handler that can flush a range of pages to save on xcalls.
12546  */
12547 static int sfmmu_xcall_save;
12548 
12549 /*
12550  * this routine is never used for demaping addresses backed by SRD hmeblks.
12551  */
12552 static void
12553 sfmmu_tlb_range_demap(demap_range_t *dmrp)
12554 {
12555 	sfmmu_t *sfmmup = dmrp->dmr_sfmmup;
12556 	hatlock_t *hatlockp;
12557 	cpuset_t cpuset;
12558 	uint64_t sfmmu_pgcnt;
12559 	pgcnt_t pgcnt = 0;
12560 	int pgunload = 0;
12561 	int dirtypg = 0;
12562 	caddr_t addr = dmrp->dmr_addr;
12563 	caddr_t eaddr;
12564 	uint64_t bitvec = dmrp->dmr_bitvec;
12565 
12566 	ASSERT(bitvec & 1);
12567 
12568 	/*
12569 	 * Flush TSB and calculate number of pages to flush.
12570 	 */
12571 	while (bitvec != 0) {
12572 		dirtypg = 0;
12573 		/*
12574 		 * Find the first page to flush and then count how many
12575 		 * pages there are after it that also need to be flushed.
12576 		 * This way the number of TSB flushes is minimized.
12577 		 */
12578 		while ((bitvec & 1) == 0) {
12579 			pgcnt++;
12580 			addr += MMU_PAGESIZE;
12581 			bitvec >>= 1;
12582 		}
12583 		while (bitvec & 1) {
12584 			dirtypg++;
12585 			bitvec >>= 1;
12586 		}
12587 		eaddr = addr + ptob(dirtypg);
12588 		hatlockp = sfmmu_hat_enter(sfmmup);
12589 		sfmmu_unload_tsb_range(sfmmup, addr, eaddr, TTE8K);
12590 		sfmmu_hat_exit(hatlockp);
12591 		pgunload += dirtypg;
12592 		addr = eaddr;
12593 		pgcnt += dirtypg;
12594 	}
12595 
12596 	ASSERT((pgcnt<<MMU_PAGESHIFT) <= dmrp->dmr_endaddr - dmrp->dmr_addr);
12597 	if (sfmmup->sfmmu_free == 0) {
12598 		addr = dmrp->dmr_addr;
12599 		bitvec = dmrp->dmr_bitvec;
12600 
12601 		/*
12602 		 * make sure it has SFMMU_PGCNT_SHIFT bits only,
12603 		 * as it will be used to pack argument for xt_some
12604 		 */
12605 		ASSERT((pgcnt > 0) &&
12606 		    (pgcnt <= (1 << SFMMU_PGCNT_SHIFT)));
12607 
12608 		/*
12609 		 * Encode pgcnt as (pgcnt -1 ), and pass (pgcnt - 1) in
12610 		 * the low 6 bits of sfmmup. This is doable since pgcnt
12611 		 * always >= 1.
12612 		 */
12613 		ASSERT(!((uint64_t)sfmmup & SFMMU_PGCNT_MASK));
12614 		sfmmu_pgcnt = (uint64_t)sfmmup |
12615 		    ((pgcnt - 1) & SFMMU_PGCNT_MASK);
12616 
12617 		/*
12618 		 * We must hold the hat lock during the flush of TLB,
12619 		 * to avoid a race with sfmmu_invalidate_ctx(), where
12620 		 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12621 		 * causing TLB demap routine to skip flush on that MMU.
12622 		 * If the context on a MMU has already been set to
12623 		 * INVALID_CONTEXT, we just get an extra flush on
12624 		 * that MMU.
12625 		 */
12626 		hatlockp = sfmmu_hat_enter(sfmmup);
12627 		kpreempt_disable();
12628 
12629 		cpuset = sfmmup->sfmmu_cpusran;
12630 		CPUSET_AND(cpuset, cpu_ready_set);
12631 		CPUSET_DEL(cpuset, CPU->cpu_id);
12632 
12633 		SFMMU_XCALL_STATS(sfmmup);
12634 		xt_some(cpuset, vtag_flush_pgcnt_tl1, (uint64_t)addr,
12635 		    sfmmu_pgcnt);
12636 
12637 		for (; bitvec != 0; bitvec >>= 1) {
12638 			if (bitvec & 1)
12639 				vtag_flushpage(addr, (uint64_t)sfmmup);
12640 			addr += MMU_PAGESIZE;
12641 		}
12642 		kpreempt_enable();
12643 		sfmmu_hat_exit(hatlockp);
12644 
12645 		sfmmu_xcall_save += (pgunload-1);
12646 	}
12647 	dmrp->dmr_bitvec = 0;
12648 }
12649 
12650 /*
12651  * In cases where we need to synchronize with TLB/TSB miss trap
12652  * handlers, _and_ need to flush the TLB, it's a lot easier to
12653  * throw away the context from the process than to do a
12654  * special song and dance to keep things consistent for the
12655  * handlers.
12656  *
12657  * Since the process suddenly ends up without a context and our caller
12658  * holds the hat lock, threads that fault after this function is called
12659  * will pile up on the lock.  We can then do whatever we need to
12660  * atomically from the context of the caller.  The first blocked thread
12661  * to resume executing will get the process a new context, and the
12662  * process will resume executing.
12663  *
12664  * One added advantage of this approach is that on MMUs that
12665  * support a "flush all" operation, we will delay the flush until
12666  * cnum wrap-around, and then flush the TLB one time.  This
12667  * is rather rare, so it's a lot less expensive than making 8000
12668  * x-calls to flush the TLB 8000 times.
12669  *
12670  * A per-process (PP) lock is used to synchronize ctx allocations in
12671  * resume() and ctx invalidations here.
12672  */
12673 static void
12674 sfmmu_invalidate_ctx(sfmmu_t *sfmmup)
12675 {
12676 	cpuset_t cpuset;
12677 	int cnum, currcnum;
12678 	mmu_ctx_t *mmu_ctxp;
12679 	int i;
12680 	uint_t pstate_save;
12681 
12682 	SFMMU_STAT(sf_ctx_inv);
12683 
12684 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12685 	ASSERT(sfmmup != ksfmmup);
12686 
12687 	kpreempt_disable();
12688 
12689 	mmu_ctxp = CPU_MMU_CTXP(CPU);
12690 	ASSERT(mmu_ctxp);
12691 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
12692 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
12693 
12694 	currcnum = sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum;
12695 
12696 	pstate_save = sfmmu_disable_intrs();
12697 
12698 	lock_set(&sfmmup->sfmmu_ctx_lock);	/* acquire PP lock */
12699 	/* set HAT cnum invalid across all context domains. */
12700 	for (i = 0; i < max_mmu_ctxdoms; i++) {
12701 
12702 		cnum = 	sfmmup->sfmmu_ctxs[i].cnum;
12703 		if (cnum == INVALID_CONTEXT) {
12704 			continue;
12705 		}
12706 
12707 		sfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
12708 	}
12709 	membar_enter();	/* make sure globally visible to all CPUs */
12710 	lock_clear(&sfmmup->sfmmu_ctx_lock);	/* release PP lock */
12711 
12712 	sfmmu_enable_intrs(pstate_save);
12713 
12714 	cpuset = sfmmup->sfmmu_cpusran;
12715 	CPUSET_DEL(cpuset, CPU->cpu_id);
12716 	CPUSET_AND(cpuset, cpu_ready_set);
12717 	if (!CPUSET_ISNULL(cpuset)) {
12718 		SFMMU_XCALL_STATS(sfmmup);
12719 		xt_some(cpuset, sfmmu_raise_tsb_exception,
12720 		    (uint64_t)sfmmup, INVALID_CONTEXT);
12721 		xt_sync(cpuset);
12722 		SFMMU_STAT(sf_tsb_raise_exception);
12723 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
12724 	}
12725 
12726 	/*
12727 	 * If the hat to-be-invalidated is the same as the current
12728 	 * process on local CPU we need to invalidate
12729 	 * this CPU context as well.
12730 	 */
12731 	if ((sfmmu_getctx_sec() == currcnum) &&
12732 	    (currcnum != INVALID_CONTEXT)) {
12733 		/* sets shared context to INVALID too */
12734 		sfmmu_setctx_sec(INVALID_CONTEXT);
12735 		sfmmu_clear_utsbinfo();
12736 	}
12737 
12738 	SFMMU_FLAGS_SET(sfmmup, HAT_ALLCTX_INVALID);
12739 
12740 	kpreempt_enable();
12741 
12742 	/*
12743 	 * we hold the hat lock, so nobody should allocate a context
12744 	 * for us yet
12745 	 */
12746 	ASSERT(sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum == INVALID_CONTEXT);
12747 }
12748 
12749 #ifdef VAC
12750 /*
12751  * We need to flush the cache in all cpus.  It is possible that
12752  * a process referenced a page as cacheable but has sinced exited
12753  * and cleared the mapping list.  We still to flush it but have no
12754  * state so all cpus is the only alternative.
12755  */
12756 void
12757 sfmmu_cache_flush(pfn_t pfnum, int vcolor)
12758 {
12759 	cpuset_t cpuset;
12760 
12761 	kpreempt_disable();
12762 	cpuset = cpu_ready_set;
12763 	CPUSET_DEL(cpuset, CPU->cpu_id);
12764 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12765 	xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12766 	xt_sync(cpuset);
12767 	vac_flushpage(pfnum, vcolor);
12768 	kpreempt_enable();
12769 }
12770 
12771 void
12772 sfmmu_cache_flushcolor(int vcolor, pfn_t pfnum)
12773 {
12774 	cpuset_t cpuset;
12775 
12776 	ASSERT(vcolor >= 0);
12777 
12778 	kpreempt_disable();
12779 	cpuset = cpu_ready_set;
12780 	CPUSET_DEL(cpuset, CPU->cpu_id);
12781 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12782 	xt_some(cpuset, vac_flushcolor_tl1, vcolor, pfnum);
12783 	xt_sync(cpuset);
12784 	vac_flushcolor(vcolor, pfnum);
12785 	kpreempt_enable();
12786 }
12787 #endif	/* VAC */
12788 
12789 /*
12790  * We need to prevent processes from accessing the TSB using a cached physical
12791  * address.  It's alright if they try to access the TSB via virtual address
12792  * since they will just fault on that virtual address once the mapping has
12793  * been suspended.
12794  */
12795 #pragma weak sendmondo_in_recover
12796 
12797 /* ARGSUSED */
12798 static int
12799 sfmmu_tsb_pre_relocator(caddr_t va, uint_t tsbsz, uint_t flags, void *tsbinfo)
12800 {
12801 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12802 	sfmmu_t *sfmmup = tsbinfop->tsb_sfmmu;
12803 	hatlock_t *hatlockp;
12804 	sf_scd_t *scdp;
12805 
12806 	if (flags != HAT_PRESUSPEND)
12807 		return (0);
12808 
12809 	/*
12810 	 * If tsb is a shared TSB with TSB_SHAREDCTX set, sfmmup must
12811 	 * be a shared hat, then set SCD's tsbinfo's flag.
12812 	 * If tsb is not shared, sfmmup is a private hat, then set
12813 	 * its private tsbinfo's flag.
12814 	 */
12815 	hatlockp = sfmmu_hat_enter(sfmmup);
12816 	tsbinfop->tsb_flags |= TSB_RELOC_FLAG;
12817 
12818 	if (!(tsbinfop->tsb_flags & TSB_SHAREDCTX)) {
12819 		sfmmu_tsb_inv_ctx(sfmmup);
12820 		sfmmu_hat_exit(hatlockp);
12821 	} else {
12822 		/* release lock on the shared hat */
12823 		sfmmu_hat_exit(hatlockp);
12824 		/* sfmmup is a shared hat */
12825 		ASSERT(sfmmup->sfmmu_scdhat);
12826 		scdp = sfmmup->sfmmu_scdp;
12827 		ASSERT(scdp != NULL);
12828 		/* get private hat from the scd list */
12829 		mutex_enter(&scdp->scd_mutex);
12830 		sfmmup = scdp->scd_sf_list;
12831 		while (sfmmup != NULL) {
12832 			hatlockp = sfmmu_hat_enter(sfmmup);
12833 			/*
12834 			 * We do not call sfmmu_tsb_inv_ctx here because
12835 			 * sendmondo_in_recover check is only needed for
12836 			 * sun4u.
12837 			 */
12838 			sfmmu_invalidate_ctx(sfmmup);
12839 			sfmmu_hat_exit(hatlockp);
12840 			sfmmup = sfmmup->sfmmu_scd_link.next;
12841 
12842 		}
12843 		mutex_exit(&scdp->scd_mutex);
12844 	}
12845 	return (0);
12846 }
12847 
12848 static void
12849 sfmmu_tsb_inv_ctx(sfmmu_t *sfmmup)
12850 {
12851 	extern uint32_t sendmondo_in_recover;
12852 
12853 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12854 
12855 	/*
12856 	 * For Cheetah+ Erratum 25:
12857 	 * Wait for any active recovery to finish.  We can't risk
12858 	 * relocating the TSB of the thread running mondo_recover_proc()
12859 	 * since, if we did that, we would deadlock.  The scenario we are
12860 	 * trying to avoid is as follows:
12861 	 *
12862 	 * THIS CPU			RECOVER CPU
12863 	 * --------			-----------
12864 	 *				Begins recovery, walking through TSB
12865 	 * hat_pagesuspend() TSB TTE
12866 	 *				TLB miss on TSB TTE, spins at TL1
12867 	 * xt_sync()
12868 	 *	send_mondo_timeout()
12869 	 *	mondo_recover_proc()
12870 	 *	((deadlocked))
12871 	 *
12872 	 * The second half of the workaround is that mondo_recover_proc()
12873 	 * checks to see if the tsb_info has the RELOC flag set, and if it
12874 	 * does, it skips over that TSB without ever touching tsbinfop->tsb_va
12875 	 * and hence avoiding the TLB miss that could result in a deadlock.
12876 	 */
12877 	if (&sendmondo_in_recover) {
12878 		membar_enter();	/* make sure RELOC flag visible */
12879 		while (sendmondo_in_recover) {
12880 			drv_usecwait(1);
12881 			membar_consumer();
12882 		}
12883 	}
12884 
12885 	sfmmu_invalidate_ctx(sfmmup);
12886 }
12887 
12888 /* ARGSUSED */
12889 static int
12890 sfmmu_tsb_post_relocator(caddr_t va, uint_t tsbsz, uint_t flags,
12891 	void *tsbinfo, pfn_t newpfn)
12892 {
12893 	hatlock_t *hatlockp;
12894 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12895 	sfmmu_t	*sfmmup = tsbinfop->tsb_sfmmu;
12896 
12897 	if (flags != HAT_POSTUNSUSPEND)
12898 		return (0);
12899 
12900 	hatlockp = sfmmu_hat_enter(sfmmup);
12901 
12902 	SFMMU_STAT(sf_tsb_reloc);
12903 
12904 	/*
12905 	 * The process may have swapped out while we were relocating one
12906 	 * of its TSBs.  If so, don't bother doing the setup since the
12907 	 * process can't be using the memory anymore.
12908 	 */
12909 	if ((tsbinfop->tsb_flags & TSB_SWAPPED) == 0) {
12910 		ASSERT(va == tsbinfop->tsb_va);
12911 		sfmmu_tsbinfo_setup_phys(tsbinfop, newpfn);
12912 
12913 		if (tsbinfop->tsb_flags & TSB_FLUSH_NEEDED) {
12914 			sfmmu_inv_tsb(tsbinfop->tsb_va,
12915 			    TSB_BYTES(tsbinfop->tsb_szc));
12916 			tsbinfop->tsb_flags &= ~TSB_FLUSH_NEEDED;
12917 		}
12918 	}
12919 
12920 	membar_exit();
12921 	tsbinfop->tsb_flags &= ~TSB_RELOC_FLAG;
12922 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
12923 
12924 	sfmmu_hat_exit(hatlockp);
12925 
12926 	return (0);
12927 }
12928 
12929 /*
12930  * Allocate and initialize a tsb_info structure.  Note that we may or may not
12931  * allocate a TSB here, depending on the flags passed in.
12932  */
12933 static int
12934 sfmmu_tsbinfo_alloc(struct tsb_info **tsbinfopp, int tsb_szc, int tte_sz_mask,
12935 	uint_t flags, sfmmu_t *sfmmup)
12936 {
12937 	int err;
12938 
12939 	*tsbinfopp = (struct tsb_info *)kmem_cache_alloc(
12940 	    sfmmu_tsbinfo_cache, KM_SLEEP);
12941 
12942 	if ((err = sfmmu_init_tsbinfo(*tsbinfopp, tte_sz_mask,
12943 	    tsb_szc, flags, sfmmup)) != 0) {
12944 		kmem_cache_free(sfmmu_tsbinfo_cache, *tsbinfopp);
12945 		SFMMU_STAT(sf_tsb_allocfail);
12946 		*tsbinfopp = NULL;
12947 		return (err);
12948 	}
12949 	SFMMU_STAT(sf_tsb_alloc);
12950 
12951 	/*
12952 	 * Bump the TSB size counters for this TSB size.
12953 	 */
12954 	(*(((int *)&sfmmu_tsbsize_stat) + tsb_szc))++;
12955 	return (0);
12956 }
12957 
12958 static void
12959 sfmmu_tsb_free(struct tsb_info *tsbinfo)
12960 {
12961 	caddr_t tsbva = tsbinfo->tsb_va;
12962 	uint_t tsb_size = TSB_BYTES(tsbinfo->tsb_szc);
12963 	struct kmem_cache *kmem_cachep = tsbinfo->tsb_cache;
12964 	vmem_t	*vmp = tsbinfo->tsb_vmp;
12965 
12966 	/*
12967 	 * If we allocated this TSB from relocatable kernel memory, then we
12968 	 * need to uninstall the callback handler.
12969 	 */
12970 	if (tsbinfo->tsb_cache != sfmmu_tsb8k_cache) {
12971 		uintptr_t slab_mask;
12972 		caddr_t slab_vaddr;
12973 		page_t **ppl;
12974 		int ret;
12975 
12976 		ASSERT(tsb_size <= MMU_PAGESIZE4M || use_bigtsb_arena);
12977 		if (tsb_size > MMU_PAGESIZE4M)
12978 			slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
12979 		else
12980 			slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
12981 		slab_vaddr = (caddr_t)((uintptr_t)tsbva & slab_mask);
12982 
12983 		ret = as_pagelock(&kas, &ppl, slab_vaddr, PAGESIZE, S_WRITE);
12984 		ASSERT(ret == 0);
12985 		hat_delete_callback(tsbva, (uint_t)tsb_size, (void *)tsbinfo,
12986 		    0, NULL);
12987 		as_pageunlock(&kas, ppl, slab_vaddr, PAGESIZE, S_WRITE);
12988 	}
12989 
12990 	if (kmem_cachep != NULL) {
12991 		kmem_cache_free(kmem_cachep, tsbva);
12992 	} else {
12993 		vmem_xfree(vmp, (void *)tsbva, tsb_size);
12994 	}
12995 	tsbinfo->tsb_va = (caddr_t)0xbad00bad;
12996 	atomic_add_64(&tsb_alloc_bytes, -(int64_t)tsb_size);
12997 }
12998 
12999 static void
13000 sfmmu_tsbinfo_free(struct tsb_info *tsbinfo)
13001 {
13002 	if ((tsbinfo->tsb_flags & TSB_SWAPPED) == 0) {
13003 		sfmmu_tsb_free(tsbinfo);
13004 	}
13005 	kmem_cache_free(sfmmu_tsbinfo_cache, tsbinfo);
13006 
13007 }
13008 
13009 /*
13010  * Setup all the references to physical memory for this tsbinfo.
13011  * The underlying page(s) must be locked.
13012  */
13013 static void
13014 sfmmu_tsbinfo_setup_phys(struct tsb_info *tsbinfo, pfn_t pfn)
13015 {
13016 	ASSERT(pfn != PFN_INVALID);
13017 	ASSERT(pfn == va_to_pfn(tsbinfo->tsb_va));
13018 
13019 #ifndef sun4v
13020 	if (tsbinfo->tsb_szc == 0) {
13021 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn,
13022 		    PROT_WRITE|PROT_READ, TTE8K);
13023 	} else {
13024 		/*
13025 		 * Round down PA and use a large mapping; the handlers will
13026 		 * compute the TSB pointer at the correct offset into the
13027 		 * big virtual page.  NOTE: this assumes all TSBs larger
13028 		 * than 8K must come from physically contiguous slabs of
13029 		 * size tsb_slab_size.
13030 		 */
13031 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn & ~tsb_slab_mask,
13032 		    PROT_WRITE|PROT_READ, tsb_slab_ttesz);
13033 	}
13034 	tsbinfo->tsb_pa = ptob(pfn);
13035 
13036 	TTE_SET_LOCKED(&tsbinfo->tsb_tte); /* lock the tte into dtlb */
13037 	TTE_SET_MOD(&tsbinfo->tsb_tte);    /* enable writes */
13038 
13039 	ASSERT(TTE_IS_PRIVILEGED(&tsbinfo->tsb_tte));
13040 	ASSERT(TTE_IS_LOCKED(&tsbinfo->tsb_tte));
13041 #else /* sun4v */
13042 	tsbinfo->tsb_pa = ptob(pfn);
13043 #endif /* sun4v */
13044 }
13045 
13046 
13047 /*
13048  * Returns zero on success, ENOMEM if over the high water mark,
13049  * or EAGAIN if the caller needs to retry with a smaller TSB
13050  * size (or specify TSB_FORCEALLOC if the allocation can't fail).
13051  *
13052  * This call cannot fail to allocate a TSB if TSB_FORCEALLOC
13053  * is specified and the TSB requested is PAGESIZE, though it
13054  * may sleep waiting for memory if sufficient memory is not
13055  * available.
13056  */
13057 static int
13058 sfmmu_init_tsbinfo(struct tsb_info *tsbinfo, int tteszmask,
13059     int tsbcode, uint_t flags, sfmmu_t *sfmmup)
13060 {
13061 	caddr_t vaddr = NULL;
13062 	caddr_t slab_vaddr;
13063 	uintptr_t slab_mask;
13064 	int tsbbytes = TSB_BYTES(tsbcode);
13065 	int lowmem = 0;
13066 	struct kmem_cache *kmem_cachep = NULL;
13067 	vmem_t *vmp = NULL;
13068 	lgrp_id_t lgrpid = LGRP_NONE;
13069 	pfn_t pfn;
13070 	uint_t cbflags = HAC_SLEEP;
13071 	page_t **pplist;
13072 	int ret;
13073 
13074 	ASSERT(tsbbytes <= MMU_PAGESIZE4M || use_bigtsb_arena);
13075 	if (tsbbytes > MMU_PAGESIZE4M)
13076 		slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
13077 	else
13078 		slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
13079 
13080 	if (flags & (TSB_FORCEALLOC | TSB_SWAPIN | TSB_GROW | TSB_SHRINK))
13081 		flags |= TSB_ALLOC;
13082 
13083 	ASSERT((flags & TSB_FORCEALLOC) == 0 || tsbcode == TSB_MIN_SZCODE);
13084 
13085 	tsbinfo->tsb_sfmmu = sfmmup;
13086 
13087 	/*
13088 	 * If not allocating a TSB, set up the tsbinfo, set TSB_SWAPPED, and
13089 	 * return.
13090 	 */
13091 	if ((flags & TSB_ALLOC) == 0) {
13092 		tsbinfo->tsb_szc = tsbcode;
13093 		tsbinfo->tsb_ttesz_mask = tteszmask;
13094 		tsbinfo->tsb_va = (caddr_t)0xbadbadbeef;
13095 		tsbinfo->tsb_pa = -1;
13096 		tsbinfo->tsb_tte.ll = 0;
13097 		tsbinfo->tsb_next = NULL;
13098 		tsbinfo->tsb_flags = TSB_SWAPPED;
13099 		tsbinfo->tsb_cache = NULL;
13100 		tsbinfo->tsb_vmp = NULL;
13101 		return (0);
13102 	}
13103 
13104 #ifdef DEBUG
13105 	/*
13106 	 * For debugging:
13107 	 * Randomly force allocation failures every tsb_alloc_mtbf
13108 	 * tries if TSB_FORCEALLOC is not specified.  This will
13109 	 * return ENOMEM if tsb_alloc_mtbf is odd, or EAGAIN if
13110 	 * it is even, to allow testing of both failure paths...
13111 	 */
13112 	if (tsb_alloc_mtbf && ((flags & TSB_FORCEALLOC) == 0) &&
13113 	    (tsb_alloc_count++ == tsb_alloc_mtbf)) {
13114 		tsb_alloc_count = 0;
13115 		tsb_alloc_fail_mtbf++;
13116 		return ((tsb_alloc_mtbf & 1)? ENOMEM : EAGAIN);
13117 	}
13118 #endif	/* DEBUG */
13119 
13120 	/*
13121 	 * Enforce high water mark if we are not doing a forced allocation
13122 	 * and are not shrinking a process' TSB.
13123 	 */
13124 	if ((flags & TSB_SHRINK) == 0 &&
13125 	    (tsbbytes + tsb_alloc_bytes) > tsb_alloc_hiwater) {
13126 		if ((flags & TSB_FORCEALLOC) == 0)
13127 			return (ENOMEM);
13128 		lowmem = 1;
13129 	}
13130 
13131 	/*
13132 	 * Allocate from the correct location based upon the size of the TSB
13133 	 * compared to the base page size, and what memory conditions dictate.
13134 	 * Note we always do nonblocking allocations from the TSB arena since
13135 	 * we don't want memory fragmentation to cause processes to block
13136 	 * indefinitely waiting for memory; until the kernel algorithms that
13137 	 * coalesce large pages are improved this is our best option.
13138 	 *
13139 	 * Algorithm:
13140 	 *	If allocating a "large" TSB (>8K), allocate from the
13141 	 *		appropriate kmem_tsb_default_arena vmem arena
13142 	 *	else if low on memory or the TSB_FORCEALLOC flag is set or
13143 	 *	tsb_forceheap is set
13144 	 *		Allocate from kernel heap via sfmmu_tsb8k_cache with
13145 	 *		KM_SLEEP (never fails)
13146 	 *	else
13147 	 *		Allocate from appropriate sfmmu_tsb_cache with
13148 	 *		KM_NOSLEEP
13149 	 *	endif
13150 	 */
13151 	if (tsb_lgrp_affinity)
13152 		lgrpid = lgrp_home_id(curthread);
13153 	if (lgrpid == LGRP_NONE)
13154 		lgrpid = 0;	/* use lgrp of boot CPU */
13155 
13156 	if (tsbbytes > MMU_PAGESIZE) {
13157 		if (tsbbytes > MMU_PAGESIZE4M) {
13158 			vmp = kmem_bigtsb_default_arena[lgrpid];
13159 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
13160 			    0, 0, NULL, NULL, VM_NOSLEEP);
13161 		} else {
13162 			vmp = kmem_tsb_default_arena[lgrpid];
13163 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
13164 			    0, 0, NULL, NULL, VM_NOSLEEP);
13165 		}
13166 #ifdef	DEBUG
13167 	} else if (lowmem || (flags & TSB_FORCEALLOC) || tsb_forceheap) {
13168 #else	/* !DEBUG */
13169 	} else if (lowmem || (flags & TSB_FORCEALLOC)) {
13170 #endif	/* DEBUG */
13171 		kmem_cachep = sfmmu_tsb8k_cache;
13172 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_SLEEP);
13173 		ASSERT(vaddr != NULL);
13174 	} else {
13175 		kmem_cachep = sfmmu_tsb_cache[lgrpid];
13176 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_NOSLEEP);
13177 	}
13178 
13179 	tsbinfo->tsb_cache = kmem_cachep;
13180 	tsbinfo->tsb_vmp = vmp;
13181 
13182 	if (vaddr == NULL) {
13183 		return (EAGAIN);
13184 	}
13185 
13186 	atomic_add_64(&tsb_alloc_bytes, (int64_t)tsbbytes);
13187 	kmem_cachep = tsbinfo->tsb_cache;
13188 
13189 	/*
13190 	 * If we are allocating from outside the cage, then we need to
13191 	 * register a relocation callback handler.  Note that for now
13192 	 * since pseudo mappings always hang off of the slab's root page,
13193 	 * we need only lock the first 8K of the TSB slab.  This is a bit
13194 	 * hacky but it is good for performance.
13195 	 */
13196 	if (kmem_cachep != sfmmu_tsb8k_cache) {
13197 		slab_vaddr = (caddr_t)((uintptr_t)vaddr & slab_mask);
13198 		ret = as_pagelock(&kas, &pplist, slab_vaddr, PAGESIZE, S_WRITE);
13199 		ASSERT(ret == 0);
13200 		ret = hat_add_callback(sfmmu_tsb_cb_id, vaddr, (uint_t)tsbbytes,
13201 		    cbflags, (void *)tsbinfo, &pfn, NULL);
13202 
13203 		/*
13204 		 * Need to free up resources if we could not successfully
13205 		 * add the callback function and return an error condition.
13206 		 */
13207 		if (ret != 0) {
13208 			if (kmem_cachep) {
13209 				kmem_cache_free(kmem_cachep, vaddr);
13210 			} else {
13211 				vmem_xfree(vmp, (void *)vaddr, tsbbytes);
13212 			}
13213 			as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE,
13214 			    S_WRITE);
13215 			return (EAGAIN);
13216 		}
13217 	} else {
13218 		/*
13219 		 * Since allocation of 8K TSBs from heap is rare and occurs
13220 		 * during memory pressure we allocate them from permanent
13221 		 * memory rather than using callbacks to get the PFN.
13222 		 */
13223 		pfn = hat_getpfnum(kas.a_hat, vaddr);
13224 	}
13225 
13226 	tsbinfo->tsb_va = vaddr;
13227 	tsbinfo->tsb_szc = tsbcode;
13228 	tsbinfo->tsb_ttesz_mask = tteszmask;
13229 	tsbinfo->tsb_next = NULL;
13230 	tsbinfo->tsb_flags = 0;
13231 
13232 	sfmmu_tsbinfo_setup_phys(tsbinfo, pfn);
13233 
13234 	sfmmu_inv_tsb(vaddr, tsbbytes);
13235 
13236 	if (kmem_cachep != sfmmu_tsb8k_cache) {
13237 		as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE, S_WRITE);
13238 	}
13239 
13240 	return (0);
13241 }
13242 
13243 /*
13244  * Initialize per cpu tsb and per cpu tsbmiss_area
13245  */
13246 void
13247 sfmmu_init_tsbs(void)
13248 {
13249 	int i;
13250 	struct tsbmiss	*tsbmissp;
13251 	struct kpmtsbm	*kpmtsbmp;
13252 #ifndef sun4v
13253 	extern int	dcache_line_mask;
13254 #endif /* sun4v */
13255 	extern uint_t	vac_colors;
13256 
13257 	/*
13258 	 * Init. tsb miss area.
13259 	 */
13260 	tsbmissp = tsbmiss_area;
13261 
13262 	for (i = 0; i < NCPU; tsbmissp++, i++) {
13263 		/*
13264 		 * initialize the tsbmiss area.
13265 		 * Do this for all possible CPUs as some may be added
13266 		 * while the system is running. There is no cost to this.
13267 		 */
13268 		tsbmissp->ksfmmup = ksfmmup;
13269 #ifndef sun4v
13270 		tsbmissp->dcache_line_mask = (uint16_t)dcache_line_mask;
13271 #endif /* sun4v */
13272 		tsbmissp->khashstart =
13273 		    (struct hmehash_bucket *)va_to_pa((caddr_t)khme_hash);
13274 		tsbmissp->uhashstart =
13275 		    (struct hmehash_bucket *)va_to_pa((caddr_t)uhme_hash);
13276 		tsbmissp->khashsz = khmehash_num;
13277 		tsbmissp->uhashsz = uhmehash_num;
13278 	}
13279 
13280 	sfmmu_tsb_cb_id = hat_register_callback('T'<<16 | 'S' << 8 | 'B',
13281 	    sfmmu_tsb_pre_relocator, sfmmu_tsb_post_relocator, NULL, 0);
13282 
13283 	if (kpm_enable == 0)
13284 		return;
13285 
13286 	/* -- Begin KPM specific init -- */
13287 
13288 	if (kpm_smallpages) {
13289 		/*
13290 		 * If we're using base pagesize pages for seg_kpm
13291 		 * mappings, we use the kernel TSB since we can't afford
13292 		 * to allocate a second huge TSB for these mappings.
13293 		 */
13294 		kpm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13295 		kpm_tsbsz = ktsb_szcode;
13296 		kpmsm_tsbbase = kpm_tsbbase;
13297 		kpmsm_tsbsz = kpm_tsbsz;
13298 	} else {
13299 		/*
13300 		 * In VAC conflict case, just put the entries in the
13301 		 * kernel 8K indexed TSB for now so we can find them.
13302 		 * This could really be changed in the future if we feel
13303 		 * the need...
13304 		 */
13305 		kpmsm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13306 		kpmsm_tsbsz = ktsb_szcode;
13307 		kpm_tsbbase = ktsb_phys? ktsb4m_pbase : (uint64_t)ktsb4m_base;
13308 		kpm_tsbsz = ktsb4m_szcode;
13309 	}
13310 
13311 	kpmtsbmp = kpmtsbm_area;
13312 	for (i = 0; i < NCPU; kpmtsbmp++, i++) {
13313 		/*
13314 		 * Initialize the kpmtsbm area.
13315 		 * Do this for all possible CPUs as some may be added
13316 		 * while the system is running. There is no cost to this.
13317 		 */
13318 		kpmtsbmp->vbase = kpm_vbase;
13319 		kpmtsbmp->vend = kpm_vbase + kpm_size * vac_colors;
13320 		kpmtsbmp->sz_shift = kpm_size_shift;
13321 		kpmtsbmp->kpmp_shift = kpmp_shift;
13322 		kpmtsbmp->kpmp2pshft = (uchar_t)kpmp2pshft;
13323 		if (kpm_smallpages == 0) {
13324 			kpmtsbmp->kpmp_table_sz = kpmp_table_sz;
13325 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_table);
13326 		} else {
13327 			kpmtsbmp->kpmp_table_sz = kpmp_stable_sz;
13328 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_stable);
13329 		}
13330 		kpmtsbmp->msegphashpa = va_to_pa(memseg_phash);
13331 		kpmtsbmp->flags = KPMTSBM_ENABLE_FLAG;
13332 #ifdef	DEBUG
13333 		kpmtsbmp->flags |= (kpm_tsbmtl) ?  KPMTSBM_TLTSBM_FLAG : 0;
13334 #endif	/* DEBUG */
13335 		if (ktsb_phys)
13336 			kpmtsbmp->flags |= KPMTSBM_TSBPHYS_FLAG;
13337 	}
13338 
13339 	/* -- End KPM specific init -- */
13340 }
13341 
13342 /* Avoid using sfmmu_tsbinfo_alloc() to avoid kmem_alloc - no real reason */
13343 struct tsb_info ktsb_info[2];
13344 
13345 /*
13346  * Called from hat_kern_setup() to setup the tsb_info for ksfmmup.
13347  */
13348 void
13349 sfmmu_init_ktsbinfo()
13350 {
13351 	ASSERT(ksfmmup != NULL);
13352 	ASSERT(ksfmmup->sfmmu_tsb == NULL);
13353 	/*
13354 	 * Allocate tsbinfos for kernel and copy in data
13355 	 * to make debug easier and sun4v setup easier.
13356 	 */
13357 	ktsb_info[0].tsb_sfmmu = ksfmmup;
13358 	ktsb_info[0].tsb_szc = ktsb_szcode;
13359 	ktsb_info[0].tsb_ttesz_mask = TSB8K|TSB64K|TSB512K;
13360 	ktsb_info[0].tsb_va = ktsb_base;
13361 	ktsb_info[0].tsb_pa = ktsb_pbase;
13362 	ktsb_info[0].tsb_flags = 0;
13363 	ktsb_info[0].tsb_tte.ll = 0;
13364 	ktsb_info[0].tsb_cache = NULL;
13365 
13366 	ktsb_info[1].tsb_sfmmu = ksfmmup;
13367 	ktsb_info[1].tsb_szc = ktsb4m_szcode;
13368 	ktsb_info[1].tsb_ttesz_mask = TSB4M;
13369 	ktsb_info[1].tsb_va = ktsb4m_base;
13370 	ktsb_info[1].tsb_pa = ktsb4m_pbase;
13371 	ktsb_info[1].tsb_flags = 0;
13372 	ktsb_info[1].tsb_tte.ll = 0;
13373 	ktsb_info[1].tsb_cache = NULL;
13374 
13375 	/* Link them into ksfmmup. */
13376 	ktsb_info[0].tsb_next = &ktsb_info[1];
13377 	ktsb_info[1].tsb_next = NULL;
13378 	ksfmmup->sfmmu_tsb = &ktsb_info[0];
13379 
13380 	sfmmu_setup_tsbinfo(ksfmmup);
13381 }
13382 
13383 /*
13384  * Cache the last value returned from va_to_pa().  If the VA specified
13385  * in the current call to cached_va_to_pa() maps to the same Page (as the
13386  * previous call to cached_va_to_pa()), then compute the PA using
13387  * cached info, else call va_to_pa().
13388  *
13389  * Note: this function is neither MT-safe nor consistent in the presence
13390  * of multiple, interleaved threads.  This function was created to enable
13391  * an optimization used during boot (at a point when there's only one thread
13392  * executing on the "boot CPU", and before startup_vm() has been called).
13393  */
13394 static uint64_t
13395 cached_va_to_pa(void *vaddr)
13396 {
13397 	static uint64_t prev_vaddr_base = 0;
13398 	static uint64_t prev_pfn = 0;
13399 
13400 	if ((((uint64_t)vaddr) & MMU_PAGEMASK) == prev_vaddr_base) {
13401 		return (prev_pfn | ((uint64_t)vaddr & MMU_PAGEOFFSET));
13402 	} else {
13403 		uint64_t pa = va_to_pa(vaddr);
13404 
13405 		if (pa != ((uint64_t)-1)) {
13406 			/*
13407 			 * Computed physical address is valid.  Cache its
13408 			 * related info for the next cached_va_to_pa() call.
13409 			 */
13410 			prev_pfn = pa & MMU_PAGEMASK;
13411 			prev_vaddr_base = ((uint64_t)vaddr) & MMU_PAGEMASK;
13412 		}
13413 
13414 		return (pa);
13415 	}
13416 }
13417 
13418 /*
13419  * Carve up our nucleus hblk region.  We may allocate more hblks than
13420  * asked due to rounding errors but we are guaranteed to have at least
13421  * enough space to allocate the requested number of hblk8's and hblk1's.
13422  */
13423 void
13424 sfmmu_init_nucleus_hblks(caddr_t addr, size_t size, int nhblk8, int nhblk1)
13425 {
13426 	struct hme_blk *hmeblkp;
13427 	size_t hme8blk_sz, hme1blk_sz;
13428 	size_t i;
13429 	size_t hblk8_bound;
13430 	ulong_t j = 0, k = 0;
13431 
13432 	ASSERT(addr != NULL && size != 0);
13433 
13434 	/* Need to use proper structure alignment */
13435 	hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t));
13436 	hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t));
13437 
13438 	nucleus_hblk8.list = (void *)addr;
13439 	nucleus_hblk8.index = 0;
13440 
13441 	/*
13442 	 * Use as much memory as possible for hblk8's since we
13443 	 * expect all bop_alloc'ed memory to be allocated in 8k chunks.
13444 	 * We need to hold back enough space for the hblk1's which
13445 	 * we'll allocate next.
13446 	 */
13447 	hblk8_bound = size - (nhblk1 * hme1blk_sz) - hme8blk_sz;
13448 	for (i = 0; i <= hblk8_bound; i += hme8blk_sz, j++) {
13449 		hmeblkp = (struct hme_blk *)addr;
13450 		addr += hme8blk_sz;
13451 		hmeblkp->hblk_nuc_bit = 1;
13452 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13453 	}
13454 	nucleus_hblk8.len = j;
13455 	ASSERT(j >= nhblk8);
13456 	SFMMU_STAT_ADD(sf_hblk8_ncreate, j);
13457 
13458 	nucleus_hblk1.list = (void *)addr;
13459 	nucleus_hblk1.index = 0;
13460 	for (; i <= (size - hme1blk_sz); i += hme1blk_sz, k++) {
13461 		hmeblkp = (struct hme_blk *)addr;
13462 		addr += hme1blk_sz;
13463 		hmeblkp->hblk_nuc_bit = 1;
13464 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13465 	}
13466 	ASSERT(k >= nhblk1);
13467 	nucleus_hblk1.len = k;
13468 	SFMMU_STAT_ADD(sf_hblk1_ncreate, k);
13469 }
13470 
13471 /*
13472  * This function is currently not supported on this platform. For what
13473  * it's supposed to do, see hat.c and hat_srmmu.c
13474  */
13475 /* ARGSUSED */
13476 faultcode_t
13477 hat_softlock(struct hat *hat, caddr_t addr, size_t *lenp, page_t **ppp,
13478     uint_t flags)
13479 {
13480 	ASSERT(hat->sfmmu_xhat_provider == NULL);
13481 	return (FC_NOSUPPORT);
13482 }
13483 
13484 /*
13485  * Searchs the mapping list of the page for a mapping of the same size. If not
13486  * found the corresponding bit is cleared in the p_index field. When large
13487  * pages are more prevalent in the system, we can maintain the mapping list
13488  * in order and we don't have to traverse the list each time. Just check the
13489  * next and prev entries, and if both are of different size, we clear the bit.
13490  */
13491 static void
13492 sfmmu_rm_large_mappings(page_t *pp, int ttesz)
13493 {
13494 	struct sf_hment *sfhmep;
13495 	struct hme_blk *hmeblkp;
13496 	int	index;
13497 	pgcnt_t	npgs;
13498 
13499 	ASSERT(ttesz > TTE8K);
13500 
13501 	ASSERT(sfmmu_mlist_held(pp));
13502 
13503 	ASSERT(PP_ISMAPPED_LARGE(pp));
13504 
13505 	/*
13506 	 * Traverse mapping list looking for another mapping of same size.
13507 	 * since we only want to clear index field if all mappings of
13508 	 * that size are gone.
13509 	 */
13510 
13511 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
13512 		if (IS_PAHME(sfhmep))
13513 			continue;
13514 		hmeblkp = sfmmu_hmetohblk(sfhmep);
13515 		if (hmeblkp->hblk_xhat_bit)
13516 			continue;
13517 		if (hme_size(sfhmep) == ttesz) {
13518 			/*
13519 			 * another mapping of the same size. don't clear index.
13520 			 */
13521 			return;
13522 		}
13523 	}
13524 
13525 	/*
13526 	 * Clear the p_index bit for large page.
13527 	 */
13528 	index = PAGESZ_TO_INDEX(ttesz);
13529 	npgs = TTEPAGES(ttesz);
13530 	while (npgs-- > 0) {
13531 		ASSERT(pp->p_index & index);
13532 		pp->p_index &= ~index;
13533 		pp = PP_PAGENEXT(pp);
13534 	}
13535 }
13536 
13537 /*
13538  * return supported features
13539  */
13540 /* ARGSUSED */
13541 int
13542 hat_supported(enum hat_features feature, void *arg)
13543 {
13544 	switch (feature) {
13545 	case    HAT_SHARED_PT:
13546 	case	HAT_DYNAMIC_ISM_UNMAP:
13547 	case	HAT_VMODSORT:
13548 		return (1);
13549 	case	HAT_SHARED_REGIONS:
13550 		if (shctx_on)
13551 			return (1);
13552 		else
13553 			return (0);
13554 	default:
13555 		return (0);
13556 	}
13557 }
13558 
13559 void
13560 hat_enter(struct hat *hat)
13561 {
13562 	hatlock_t	*hatlockp;
13563 
13564 	if (hat != ksfmmup) {
13565 		hatlockp = TSB_HASH(hat);
13566 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
13567 	}
13568 }
13569 
13570 void
13571 hat_exit(struct hat *hat)
13572 {
13573 	hatlock_t	*hatlockp;
13574 
13575 	if (hat != ksfmmup) {
13576 		hatlockp = TSB_HASH(hat);
13577 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
13578 	}
13579 }
13580 
13581 /*ARGSUSED*/
13582 void
13583 hat_reserve(struct as *as, caddr_t addr, size_t len)
13584 {
13585 }
13586 
13587 static void
13588 hat_kstat_init(void)
13589 {
13590 	kstat_t *ksp;
13591 
13592 	ksp = kstat_create("unix", 0, "sfmmu_global_stat", "hat",
13593 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_global_stat),
13594 	    KSTAT_FLAG_VIRTUAL);
13595 	if (ksp) {
13596 		ksp->ks_data = (void *) &sfmmu_global_stat;
13597 		kstat_install(ksp);
13598 	}
13599 	ksp = kstat_create("unix", 0, "sfmmu_tsbsize_stat", "hat",
13600 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_tsbsize_stat),
13601 	    KSTAT_FLAG_VIRTUAL);
13602 	if (ksp) {
13603 		ksp->ks_data = (void *) &sfmmu_tsbsize_stat;
13604 		kstat_install(ksp);
13605 	}
13606 	ksp = kstat_create("unix", 0, "sfmmu_percpu_stat", "hat",
13607 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_percpu_stat) * NCPU,
13608 	    KSTAT_FLAG_WRITABLE);
13609 	if (ksp) {
13610 		ksp->ks_update = sfmmu_kstat_percpu_update;
13611 		kstat_install(ksp);
13612 	}
13613 }
13614 
13615 /* ARGSUSED */
13616 static int
13617 sfmmu_kstat_percpu_update(kstat_t *ksp, int rw)
13618 {
13619 	struct sfmmu_percpu_stat *cpu_kstat = ksp->ks_data;
13620 	struct tsbmiss *tsbm = tsbmiss_area;
13621 	struct kpmtsbm *kpmtsbm = kpmtsbm_area;
13622 	int i;
13623 
13624 	ASSERT(cpu_kstat);
13625 	if (rw == KSTAT_READ) {
13626 		for (i = 0; i < NCPU; cpu_kstat++, tsbm++, kpmtsbm++, i++) {
13627 			cpu_kstat->sf_itlb_misses = 0;
13628 			cpu_kstat->sf_dtlb_misses = 0;
13629 			cpu_kstat->sf_utsb_misses = tsbm->utsb_misses -
13630 			    tsbm->uprot_traps;
13631 			cpu_kstat->sf_ktsb_misses = tsbm->ktsb_misses +
13632 			    kpmtsbm->kpm_tsb_misses - tsbm->kprot_traps;
13633 			cpu_kstat->sf_tsb_hits = 0;
13634 			cpu_kstat->sf_umod_faults = tsbm->uprot_traps;
13635 			cpu_kstat->sf_kmod_faults = tsbm->kprot_traps;
13636 		}
13637 	} else {
13638 		/* KSTAT_WRITE is used to clear stats */
13639 		for (i = 0; i < NCPU; tsbm++, kpmtsbm++, i++) {
13640 			tsbm->utsb_misses = 0;
13641 			tsbm->ktsb_misses = 0;
13642 			tsbm->uprot_traps = 0;
13643 			tsbm->kprot_traps = 0;
13644 			kpmtsbm->kpm_dtlb_misses = 0;
13645 			kpmtsbm->kpm_tsb_misses = 0;
13646 		}
13647 	}
13648 	return (0);
13649 }
13650 
13651 #ifdef	DEBUG
13652 
13653 tte_t  *gorig[NCPU], *gcur[NCPU], *gnew[NCPU];
13654 
13655 /*
13656  * A tte checker. *orig_old is the value we read before cas.
13657  *	*cur is the value returned by cas.
13658  *	*new is the desired value when we do the cas.
13659  *
13660  *	*hmeblkp is currently unused.
13661  */
13662 
13663 /* ARGSUSED */
13664 void
13665 chk_tte(tte_t *orig_old, tte_t *cur, tte_t *new, struct hme_blk *hmeblkp)
13666 {
13667 	pfn_t i, j, k;
13668 	int cpuid = CPU->cpu_id;
13669 
13670 	gorig[cpuid] = orig_old;
13671 	gcur[cpuid] = cur;
13672 	gnew[cpuid] = new;
13673 
13674 #ifdef lint
13675 	hmeblkp = hmeblkp;
13676 #endif
13677 
13678 	if (TTE_IS_VALID(orig_old)) {
13679 		if (TTE_IS_VALID(cur)) {
13680 			i = TTE_TO_TTEPFN(orig_old);
13681 			j = TTE_TO_TTEPFN(cur);
13682 			k = TTE_TO_TTEPFN(new);
13683 			if (i != j) {
13684 				/* remap error? */
13685 				panic("chk_tte: bad pfn, 0x%lx, 0x%lx", i, j);
13686 			}
13687 
13688 			if (i != k) {
13689 				/* remap error? */
13690 				panic("chk_tte: bad pfn2, 0x%lx, 0x%lx", i, k);
13691 			}
13692 		} else {
13693 			if (TTE_IS_VALID(new)) {
13694 				panic("chk_tte: invalid cur? ");
13695 			}
13696 
13697 			i = TTE_TO_TTEPFN(orig_old);
13698 			k = TTE_TO_TTEPFN(new);
13699 			if (i != k) {
13700 				panic("chk_tte: bad pfn3, 0x%lx, 0x%lx", i, k);
13701 			}
13702 		}
13703 	} else {
13704 		if (TTE_IS_VALID(cur)) {
13705 			j = TTE_TO_TTEPFN(cur);
13706 			if (TTE_IS_VALID(new)) {
13707 				k = TTE_TO_TTEPFN(new);
13708 				if (j != k) {
13709 					panic("chk_tte: bad pfn4, 0x%lx, 0x%lx",
13710 					    j, k);
13711 				}
13712 			} else {
13713 				panic("chk_tte: why here?");
13714 			}
13715 		} else {
13716 			if (!TTE_IS_VALID(new)) {
13717 				panic("chk_tte: why here2 ?");
13718 			}
13719 		}
13720 	}
13721 }
13722 
13723 #endif /* DEBUG */
13724 
13725 extern void prefetch_tsbe_read(struct tsbe *);
13726 extern void prefetch_tsbe_write(struct tsbe *);
13727 
13728 
13729 /*
13730  * We want to prefetch 7 cache lines ahead for our read prefetch.  This gives
13731  * us optimal performance on Cheetah+.  You can only have 8 outstanding
13732  * prefetches at any one time, so we opted for 7 read prefetches and 1 write
13733  * prefetch to make the most utilization of the prefetch capability.
13734  */
13735 #define	TSBE_PREFETCH_STRIDE (7)
13736 
13737 void
13738 sfmmu_copy_tsb(struct tsb_info *old_tsbinfo, struct tsb_info *new_tsbinfo)
13739 {
13740 	int old_bytes = TSB_BYTES(old_tsbinfo->tsb_szc);
13741 	int new_bytes = TSB_BYTES(new_tsbinfo->tsb_szc);
13742 	int old_entries = TSB_ENTRIES(old_tsbinfo->tsb_szc);
13743 	int new_entries = TSB_ENTRIES(new_tsbinfo->tsb_szc);
13744 	struct tsbe *old;
13745 	struct tsbe *new;
13746 	struct tsbe *new_base = (struct tsbe *)new_tsbinfo->tsb_va;
13747 	uint64_t va;
13748 	int new_offset;
13749 	int i;
13750 	int vpshift;
13751 	int last_prefetch;
13752 
13753 	if (old_bytes == new_bytes) {
13754 		bcopy(old_tsbinfo->tsb_va, new_tsbinfo->tsb_va, new_bytes);
13755 	} else {
13756 
13757 		/*
13758 		 * A TSBE is 16 bytes which means there are four TSBE's per
13759 		 * P$ line (64 bytes), thus every 4 TSBE's we prefetch.
13760 		 */
13761 		old = (struct tsbe *)old_tsbinfo->tsb_va;
13762 		last_prefetch = old_entries - (4*(TSBE_PREFETCH_STRIDE+1));
13763 		for (i = 0; i < old_entries; i++, old++) {
13764 			if (((i & (4-1)) == 0) && (i < last_prefetch))
13765 				prefetch_tsbe_read(old);
13766 			if (!old->tte_tag.tag_invalid) {
13767 				/*
13768 				 * We have a valid TTE to remap.  Check the
13769 				 * size.  We won't remap 64K or 512K TTEs
13770 				 * because they span more than one TSB entry
13771 				 * and are indexed using an 8K virt. page.
13772 				 * Ditto for 32M and 256M TTEs.
13773 				 */
13774 				if (TTE_CSZ(&old->tte_data) == TTE64K ||
13775 				    TTE_CSZ(&old->tte_data) == TTE512K)
13776 					continue;
13777 				if (mmu_page_sizes == max_mmu_page_sizes) {
13778 					if (TTE_CSZ(&old->tte_data) == TTE32M ||
13779 					    TTE_CSZ(&old->tte_data) == TTE256M)
13780 						continue;
13781 				}
13782 
13783 				/* clear the lower 22 bits of the va */
13784 				va = *(uint64_t *)old << 22;
13785 				/* turn va into a virtual pfn */
13786 				va >>= 22 - TSB_START_SIZE;
13787 				/*
13788 				 * or in bits from the offset in the tsb
13789 				 * to get the real virtual pfn. These
13790 				 * correspond to bits [21:13] in the va
13791 				 */
13792 				vpshift =
13793 				    TTE_BSZS_SHIFT(TTE_CSZ(&old->tte_data)) &
13794 				    0x1ff;
13795 				va |= (i << vpshift);
13796 				va >>= vpshift;
13797 				new_offset = va & (new_entries - 1);
13798 				new = new_base + new_offset;
13799 				prefetch_tsbe_write(new);
13800 				*new = *old;
13801 			}
13802 		}
13803 	}
13804 }
13805 
13806 /*
13807  * unused in sfmmu
13808  */
13809 void
13810 hat_dump(void)
13811 {
13812 }
13813 
13814 /*
13815  * Called when a thread is exiting and we have switched to the kernel address
13816  * space.  Perform the same VM initialization resume() uses when switching
13817  * processes.
13818  *
13819  * Note that sfmmu_load_mmustate() is currently a no-op for kernel threads, but
13820  * we call it anyway in case the semantics change in the future.
13821  */
13822 /*ARGSUSED*/
13823 void
13824 hat_thread_exit(kthread_t *thd)
13825 {
13826 	uint_t pgsz_cnum;
13827 	uint_t pstate_save;
13828 
13829 	ASSERT(thd->t_procp->p_as == &kas);
13830 
13831 	pgsz_cnum = KCONTEXT;
13832 #ifdef sun4u
13833 	pgsz_cnum |= (ksfmmup->sfmmu_cext << CTXREG_EXT_SHIFT);
13834 #endif
13835 
13836 	/*
13837 	 * Note that sfmmu_load_mmustate() is currently a no-op for
13838 	 * kernel threads. We need to disable interrupts here,
13839 	 * simply because otherwise sfmmu_load_mmustate() would panic
13840 	 * if the caller does not disable interrupts.
13841 	 */
13842 	pstate_save = sfmmu_disable_intrs();
13843 
13844 	/* Compatibility Note: hw takes care of MMU_SCONTEXT1 */
13845 	sfmmu_setctx_sec(pgsz_cnum);
13846 	sfmmu_load_mmustate(ksfmmup);
13847 	sfmmu_enable_intrs(pstate_save);
13848 }
13849 
13850 
13851 /*
13852  * SRD support
13853  */
13854 #define	SRD_HASH_FUNCTION(vp)	(((((uintptr_t)(vp)) >> 4) ^ \
13855 				    (((uintptr_t)(vp)) >> 11)) & \
13856 				    srd_hashmask)
13857 
13858 /*
13859  * Attach the process to the srd struct associated with the exec vnode
13860  * from which the process is started.
13861  */
13862 void
13863 hat_join_srd(struct hat *sfmmup, vnode_t *evp)
13864 {
13865 	uint_t hash = SRD_HASH_FUNCTION(evp);
13866 	sf_srd_t *srdp;
13867 	sf_srd_t *newsrdp;
13868 
13869 	ASSERT(sfmmup != ksfmmup);
13870 	ASSERT(sfmmup->sfmmu_srdp == NULL);
13871 
13872 	if (!shctx_on) {
13873 		return;
13874 	}
13875 
13876 	VN_HOLD(evp);
13877 
13878 	if (srd_buckets[hash].srdb_srdp != NULL) {
13879 		mutex_enter(&srd_buckets[hash].srdb_lock);
13880 		for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13881 		    srdp = srdp->srd_hash) {
13882 			if (srdp->srd_evp == evp) {
13883 				ASSERT(srdp->srd_refcnt >= 0);
13884 				sfmmup->sfmmu_srdp = srdp;
13885 				atomic_add_32(
13886 				    (volatile uint_t *)&srdp->srd_refcnt, 1);
13887 				mutex_exit(&srd_buckets[hash].srdb_lock);
13888 				return;
13889 			}
13890 		}
13891 		mutex_exit(&srd_buckets[hash].srdb_lock);
13892 	}
13893 	newsrdp = kmem_cache_alloc(srd_cache, KM_SLEEP);
13894 	ASSERT(newsrdp->srd_next_ismrid == 0 && newsrdp->srd_next_hmerid == 0);
13895 
13896 	newsrdp->srd_evp = evp;
13897 	newsrdp->srd_refcnt = 1;
13898 	newsrdp->srd_hmergnfree = NULL;
13899 	newsrdp->srd_ismrgnfree = NULL;
13900 
13901 	mutex_enter(&srd_buckets[hash].srdb_lock);
13902 	for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13903 	    srdp = srdp->srd_hash) {
13904 		if (srdp->srd_evp == evp) {
13905 			ASSERT(srdp->srd_refcnt >= 0);
13906 			sfmmup->sfmmu_srdp = srdp;
13907 			atomic_add_32((volatile uint_t *)&srdp->srd_refcnt, 1);
13908 			mutex_exit(&srd_buckets[hash].srdb_lock);
13909 			kmem_cache_free(srd_cache, newsrdp);
13910 			return;
13911 		}
13912 	}
13913 	newsrdp->srd_hash = srd_buckets[hash].srdb_srdp;
13914 	srd_buckets[hash].srdb_srdp = newsrdp;
13915 	sfmmup->sfmmu_srdp = newsrdp;
13916 
13917 	mutex_exit(&srd_buckets[hash].srdb_lock);
13918 
13919 }
13920 
13921 static void
13922 sfmmu_leave_srd(sfmmu_t *sfmmup)
13923 {
13924 	vnode_t *evp;
13925 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
13926 	uint_t hash;
13927 	sf_srd_t **prev_srdpp;
13928 	sf_region_t *rgnp;
13929 	sf_region_t *nrgnp;
13930 #ifdef DEBUG
13931 	int rgns = 0;
13932 #endif
13933 	int i;
13934 
13935 	ASSERT(sfmmup != ksfmmup);
13936 	ASSERT(srdp != NULL);
13937 	ASSERT(srdp->srd_refcnt > 0);
13938 	ASSERT(sfmmup->sfmmu_scdp == NULL);
13939 	ASSERT(sfmmup->sfmmu_free == 1);
13940 
13941 	sfmmup->sfmmu_srdp = NULL;
13942 	evp = srdp->srd_evp;
13943 	ASSERT(evp != NULL);
13944 	if (atomic_add_32_nv(
13945 	    (volatile uint_t *)&srdp->srd_refcnt, -1)) {
13946 		VN_RELE(evp);
13947 		return;
13948 	}
13949 
13950 	hash = SRD_HASH_FUNCTION(evp);
13951 	mutex_enter(&srd_buckets[hash].srdb_lock);
13952 	for (prev_srdpp = &srd_buckets[hash].srdb_srdp;
13953 	    (srdp = *prev_srdpp) != NULL; prev_srdpp = &srdp->srd_hash) {
13954 		if (srdp->srd_evp == evp) {
13955 			break;
13956 		}
13957 	}
13958 	if (srdp == NULL || srdp->srd_refcnt) {
13959 		mutex_exit(&srd_buckets[hash].srdb_lock);
13960 		VN_RELE(evp);
13961 		return;
13962 	}
13963 	*prev_srdpp = srdp->srd_hash;
13964 	mutex_exit(&srd_buckets[hash].srdb_lock);
13965 
13966 	ASSERT(srdp->srd_refcnt == 0);
13967 	VN_RELE(evp);
13968 
13969 #ifdef DEBUG
13970 	for (i = 0; i < SFMMU_MAX_REGION_BUCKETS; i++) {
13971 		ASSERT(srdp->srd_rgnhash[i] == NULL);
13972 	}
13973 #endif /* DEBUG */
13974 
13975 	/* free each hme regions in the srd */
13976 	for (rgnp = srdp->srd_hmergnfree; rgnp != NULL; rgnp = nrgnp) {
13977 		nrgnp = rgnp->rgn_next;
13978 		ASSERT(rgnp->rgn_id < srdp->srd_next_hmerid);
13979 		ASSERT(rgnp->rgn_refcnt == 0);
13980 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13981 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13982 		ASSERT(rgnp->rgn_hmeflags == 0);
13983 		ASSERT(srdp->srd_hmergnp[rgnp->rgn_id] == rgnp);
13984 #ifdef DEBUG
13985 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13986 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13987 		}
13988 		rgns++;
13989 #endif /* DEBUG */
13990 		kmem_cache_free(region_cache, rgnp);
13991 	}
13992 	ASSERT(rgns == srdp->srd_next_hmerid);
13993 
13994 #ifdef DEBUG
13995 	rgns = 0;
13996 #endif
13997 	/* free each ism rgns in the srd */
13998 	for (rgnp = srdp->srd_ismrgnfree; rgnp != NULL; rgnp = nrgnp) {
13999 		nrgnp = rgnp->rgn_next;
14000 		ASSERT(rgnp->rgn_id < srdp->srd_next_ismrid);
14001 		ASSERT(rgnp->rgn_refcnt == 0);
14002 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
14003 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
14004 		ASSERT(srdp->srd_ismrgnp[rgnp->rgn_id] == rgnp);
14005 #ifdef DEBUG
14006 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
14007 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
14008 		}
14009 		rgns++;
14010 #endif /* DEBUG */
14011 		kmem_cache_free(region_cache, rgnp);
14012 	}
14013 	ASSERT(rgns == srdp->srd_next_ismrid);
14014 	ASSERT(srdp->srd_ismbusyrgns == 0);
14015 	ASSERT(srdp->srd_hmebusyrgns == 0);
14016 
14017 	srdp->srd_next_ismrid = 0;
14018 	srdp->srd_next_hmerid = 0;
14019 
14020 	bzero((void *)srdp->srd_ismrgnp,
14021 	    sizeof (sf_region_t *) * SFMMU_MAX_ISM_REGIONS);
14022 	bzero((void *)srdp->srd_hmergnp,
14023 	    sizeof (sf_region_t *) * SFMMU_MAX_HME_REGIONS);
14024 
14025 	ASSERT(srdp->srd_scdp == NULL);
14026 	kmem_cache_free(srd_cache, srdp);
14027 }
14028 
14029 /* ARGSUSED */
14030 static int
14031 sfmmu_srdcache_constructor(void *buf, void *cdrarg, int kmflags)
14032 {
14033 	sf_srd_t *srdp = (sf_srd_t *)buf;
14034 	bzero(buf, sizeof (*srdp));
14035 
14036 	mutex_init(&srdp->srd_mutex, NULL, MUTEX_DEFAULT, NULL);
14037 	mutex_init(&srdp->srd_scd_mutex, NULL, MUTEX_DEFAULT, NULL);
14038 	return (0);
14039 }
14040 
14041 /* ARGSUSED */
14042 static void
14043 sfmmu_srdcache_destructor(void *buf, void *cdrarg)
14044 {
14045 	sf_srd_t *srdp = (sf_srd_t *)buf;
14046 
14047 	mutex_destroy(&srdp->srd_mutex);
14048 	mutex_destroy(&srdp->srd_scd_mutex);
14049 }
14050 
14051 /*
14052  * The caller makes sure hat_join_region()/hat_leave_region() can't be called
14053  * at the same time for the same process and address range. This is ensured by
14054  * the fact that address space is locked as writer when a process joins the
14055  * regions. Therefore there's no need to hold an srd lock during the entire
14056  * execution of hat_join_region()/hat_leave_region().
14057  */
14058 
14059 #define	RGN_HASH_FUNCTION(obj)	(((((uintptr_t)(obj)) >> 4) ^ \
14060 				    (((uintptr_t)(obj)) >> 11)) & \
14061 					srd_rgn_hashmask)
14062 /*
14063  * This routine implements the shared context functionality required when
14064  * attaching a segment to an address space. It must be called from
14065  * hat_share() for D(ISM) segments and from segvn_create() for segments
14066  * with the MAP_PRIVATE and MAP_TEXT flags set. It returns a region_cookie
14067  * which is saved in the private segment data for hme segments and
14068  * the ism_map structure for ism segments.
14069  */
14070 hat_region_cookie_t
14071 hat_join_region(struct hat *sfmmup,
14072 	caddr_t r_saddr,
14073 	size_t r_size,
14074 	void *r_obj,
14075 	u_offset_t r_objoff,
14076 	uchar_t r_perm,
14077 	uchar_t r_pgszc,
14078 	hat_rgn_cb_func_t r_cb_function,
14079 	uint_t flags)
14080 {
14081 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14082 	uint_t rhash;
14083 	uint_t rid;
14084 	hatlock_t *hatlockp;
14085 	sf_region_t *rgnp;
14086 	sf_region_t *new_rgnp = NULL;
14087 	int i;
14088 	uint16_t *nextidp;
14089 	sf_region_t **freelistp;
14090 	int maxids;
14091 	sf_region_t **rarrp;
14092 	uint16_t *busyrgnsp;
14093 	ulong_t rttecnt;
14094 	uchar_t tteflag;
14095 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
14096 	int text = (r_type == HAT_REGION_TEXT);
14097 
14098 	if (srdp == NULL || r_size == 0) {
14099 		return (HAT_INVALID_REGION_COOKIE);
14100 	}
14101 
14102 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
14103 	ASSERT(sfmmup != ksfmmup);
14104 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
14105 	ASSERT(srdp->srd_refcnt > 0);
14106 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
14107 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
14108 	ASSERT(r_pgszc < mmu_page_sizes);
14109 	if (!IS_P2ALIGNED(r_saddr, TTEBYTES(r_pgszc)) ||
14110 	    !IS_P2ALIGNED(r_size, TTEBYTES(r_pgszc))) {
14111 		panic("hat_join_region: region addr or size is not aligned\n");
14112 	}
14113 
14114 
14115 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
14116 	    SFMMU_REGION_HME;
14117 	/*
14118 	 * Currently only support shared hmes for the read only main text
14119 	 * region.
14120 	 */
14121 	if (r_type == SFMMU_REGION_HME && ((r_obj != srdp->srd_evp) ||
14122 	    (r_perm & PROT_WRITE))) {
14123 		return (HAT_INVALID_REGION_COOKIE);
14124 	}
14125 
14126 	rhash = RGN_HASH_FUNCTION(r_obj);
14127 
14128 	if (r_type == SFMMU_REGION_ISM) {
14129 		nextidp = &srdp->srd_next_ismrid;
14130 		freelistp = &srdp->srd_ismrgnfree;
14131 		maxids = SFMMU_MAX_ISM_REGIONS;
14132 		rarrp = srdp->srd_ismrgnp;
14133 		busyrgnsp = &srdp->srd_ismbusyrgns;
14134 	} else {
14135 		nextidp = &srdp->srd_next_hmerid;
14136 		freelistp = &srdp->srd_hmergnfree;
14137 		maxids = SFMMU_MAX_HME_REGIONS;
14138 		rarrp = srdp->srd_hmergnp;
14139 		busyrgnsp = &srdp->srd_hmebusyrgns;
14140 	}
14141 
14142 	mutex_enter(&srdp->srd_mutex);
14143 
14144 	for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
14145 	    rgnp = rgnp->rgn_hash) {
14146 		if (rgnp->rgn_saddr == r_saddr && rgnp->rgn_size == r_size &&
14147 		    rgnp->rgn_obj == r_obj && rgnp->rgn_objoff == r_objoff &&
14148 		    rgnp->rgn_perm == r_perm && rgnp->rgn_pgszc == r_pgszc) {
14149 			break;
14150 		}
14151 	}
14152 
14153 rfound:
14154 	if (rgnp != NULL) {
14155 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14156 		ASSERT(rgnp->rgn_cb_function == r_cb_function);
14157 		ASSERT(rgnp->rgn_refcnt >= 0);
14158 		rid = rgnp->rgn_id;
14159 		ASSERT(rid < maxids);
14160 		ASSERT(rarrp[rid] == rgnp);
14161 		ASSERT(rid < *nextidp);
14162 		atomic_add_32((volatile uint_t *)&rgnp->rgn_refcnt, 1);
14163 		mutex_exit(&srdp->srd_mutex);
14164 		if (new_rgnp != NULL) {
14165 			kmem_cache_free(region_cache, new_rgnp);
14166 		}
14167 		if (r_type == SFMMU_REGION_HME) {
14168 			int myjoin =
14169 			    (sfmmup == astosfmmu(curthread->t_procp->p_as));
14170 
14171 			sfmmu_link_to_hmeregion(sfmmup, rgnp);
14172 			/*
14173 			 * bitmap should be updated after linking sfmmu on
14174 			 * region list so that pageunload() doesn't skip
14175 			 * TSB/TLB flush. As soon as bitmap is updated another
14176 			 * thread in this process can already start accessing
14177 			 * this region.
14178 			 */
14179 			/*
14180 			 * Normally ttecnt accounting is done as part of
14181 			 * pagefault handling. But a process may not take any
14182 			 * pagefaults on shared hmeblks created by some other
14183 			 * process. To compensate for this assume that the
14184 			 * entire region will end up faulted in using
14185 			 * the region's pagesize.
14186 			 *
14187 			 */
14188 			if (r_pgszc > TTE8K) {
14189 				tteflag = 1 << r_pgszc;
14190 				if (disable_large_pages & tteflag) {
14191 					tteflag = 0;
14192 				}
14193 			} else {
14194 				tteflag = 0;
14195 			}
14196 			if (tteflag && !(sfmmup->sfmmu_rtteflags & tteflag)) {
14197 				hatlockp = sfmmu_hat_enter(sfmmup);
14198 				sfmmup->sfmmu_rtteflags |= tteflag;
14199 				sfmmu_hat_exit(hatlockp);
14200 			}
14201 			hatlockp = sfmmu_hat_enter(sfmmup);
14202 
14203 			/*
14204 			 * Preallocate 1/4 of ttecnt's in 8K TSB for >= 4M
14205 			 * region to allow for large page allocation failure.
14206 			 */
14207 			if (r_pgszc >= TTE4M) {
14208 				sfmmup->sfmmu_tsb0_4minflcnt +=
14209 				    r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14210 			}
14211 
14212 			/* update sfmmu_ttecnt with the shme rgn ttecnt */
14213 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14214 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14215 			    rttecnt);
14216 
14217 			if (text && r_pgszc >= TTE4M &&
14218 			    (tteflag || ((disable_large_pages >> TTE4M) &
14219 			    ((1 << (r_pgszc - TTE4M + 1)) - 1))) &&
14220 			    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
14221 				SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
14222 			}
14223 
14224 			sfmmu_hat_exit(hatlockp);
14225 			/*
14226 			 * On Panther we need to make sure TLB is programmed
14227 			 * to accept 32M/256M pages.  Call
14228 			 * sfmmu_check_page_sizes() now to make sure TLB is
14229 			 * setup before making hmeregions visible to other
14230 			 * threads.
14231 			 */
14232 			sfmmu_check_page_sizes(sfmmup, 1);
14233 			hatlockp = sfmmu_hat_enter(sfmmup);
14234 			SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14235 
14236 			/*
14237 			 * if context is invalid tsb miss exception code will
14238 			 * call sfmmu_check_page_sizes() and update tsbmiss
14239 			 * area later.
14240 			 */
14241 			kpreempt_disable();
14242 			if (myjoin &&
14243 			    (sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
14244 			    != INVALID_CONTEXT)) {
14245 				struct tsbmiss *tsbmp;
14246 
14247 				tsbmp = &tsbmiss_area[CPU->cpu_id];
14248 				ASSERT(sfmmup == tsbmp->usfmmup);
14249 				BT_SET(tsbmp->shmermap, rid);
14250 				if (r_pgszc > TTE64K) {
14251 					tsbmp->uhat_rtteflags |= tteflag;
14252 				}
14253 
14254 			}
14255 			kpreempt_enable();
14256 
14257 			sfmmu_hat_exit(hatlockp);
14258 			ASSERT((hat_region_cookie_t)((uint64_t)rid) !=
14259 			    HAT_INVALID_REGION_COOKIE);
14260 		} else {
14261 			hatlockp = sfmmu_hat_enter(sfmmup);
14262 			SF_RGNMAP_ADD(sfmmup->sfmmu_ismregion_map, rid);
14263 			sfmmu_hat_exit(hatlockp);
14264 		}
14265 		ASSERT(rid < maxids);
14266 
14267 		if (r_type == SFMMU_REGION_ISM) {
14268 			sfmmu_find_scd(sfmmup);
14269 		}
14270 		return ((hat_region_cookie_t)((uint64_t)rid));
14271 	}
14272 
14273 	ASSERT(new_rgnp == NULL);
14274 
14275 	if (*busyrgnsp >= maxids) {
14276 		mutex_exit(&srdp->srd_mutex);
14277 		return (HAT_INVALID_REGION_COOKIE);
14278 	}
14279 
14280 	ASSERT(MUTEX_HELD(&srdp->srd_mutex));
14281 	if (*freelistp != NULL) {
14282 		rgnp = *freelistp;
14283 		*freelistp = rgnp->rgn_next;
14284 		ASSERT(rgnp->rgn_id < *nextidp);
14285 		ASSERT(rgnp->rgn_id < maxids);
14286 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
14287 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK)
14288 		    == r_type);
14289 		ASSERT(rarrp[rgnp->rgn_id] == rgnp);
14290 		ASSERT(rgnp->rgn_hmeflags == 0);
14291 	} else {
14292 		/*
14293 		 * release local locks before memory allocation.
14294 		 */
14295 		mutex_exit(&srdp->srd_mutex);
14296 
14297 		new_rgnp = kmem_cache_alloc(region_cache, KM_SLEEP);
14298 
14299 		mutex_enter(&srdp->srd_mutex);
14300 		for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
14301 		    rgnp = rgnp->rgn_hash) {
14302 			if (rgnp->rgn_saddr == r_saddr &&
14303 			    rgnp->rgn_size == r_size &&
14304 			    rgnp->rgn_obj == r_obj &&
14305 			    rgnp->rgn_objoff == r_objoff &&
14306 			    rgnp->rgn_perm == r_perm &&
14307 			    rgnp->rgn_pgszc == r_pgszc) {
14308 				break;
14309 			}
14310 		}
14311 		if (rgnp != NULL) {
14312 			goto rfound;
14313 		}
14314 
14315 		if (*nextidp >= maxids) {
14316 			mutex_exit(&srdp->srd_mutex);
14317 			goto fail;
14318 		}
14319 		rgnp = new_rgnp;
14320 		new_rgnp = NULL;
14321 		rgnp->rgn_id = (*nextidp)++;
14322 		ASSERT(rgnp->rgn_id < maxids);
14323 		ASSERT(rarrp[rgnp->rgn_id] == NULL);
14324 		rarrp[rgnp->rgn_id] = rgnp;
14325 	}
14326 
14327 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14328 	ASSERT(rgnp->rgn_hmeflags == 0);
14329 #ifdef DEBUG
14330 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14331 		ASSERT(rgnp->rgn_ttecnt[i] == 0);
14332 	}
14333 #endif
14334 	rgnp->rgn_saddr = r_saddr;
14335 	rgnp->rgn_size = r_size;
14336 	rgnp->rgn_obj = r_obj;
14337 	rgnp->rgn_objoff = r_objoff;
14338 	rgnp->rgn_perm = r_perm;
14339 	rgnp->rgn_pgszc = r_pgszc;
14340 	rgnp->rgn_flags = r_type;
14341 	rgnp->rgn_refcnt = 0;
14342 	rgnp->rgn_cb_function = r_cb_function;
14343 	rgnp->rgn_hash = srdp->srd_rgnhash[rhash];
14344 	srdp->srd_rgnhash[rhash] = rgnp;
14345 	(*busyrgnsp)++;
14346 	ASSERT(*busyrgnsp <= maxids);
14347 	goto rfound;
14348 
14349 fail:
14350 	ASSERT(new_rgnp != NULL);
14351 	kmem_cache_free(region_cache, new_rgnp);
14352 	return (HAT_INVALID_REGION_COOKIE);
14353 }
14354 
14355 /*
14356  * This function implements the shared context functionality required
14357  * when detaching a segment from an address space. It must be called
14358  * from hat_unshare() for all D(ISM) segments and from segvn_unmap(),
14359  * for segments with a valid region_cookie.
14360  * It will also be called from all seg_vn routines which change a
14361  * segment's attributes such as segvn_setprot(), segvn_setpagesize(),
14362  * segvn_clrszc() & segvn_advise(), as well as in the case of COW fault
14363  * from segvn_fault().
14364  */
14365 void
14366 hat_leave_region(struct hat *sfmmup, hat_region_cookie_t rcookie, uint_t flags)
14367 {
14368 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14369 	sf_scd_t *scdp;
14370 	uint_t rhash;
14371 	uint_t rid = (uint_t)((uint64_t)rcookie);
14372 	hatlock_t *hatlockp = NULL;
14373 	sf_region_t *rgnp;
14374 	sf_region_t **prev_rgnpp;
14375 	sf_region_t *cur_rgnp;
14376 	void *r_obj;
14377 	int i;
14378 	caddr_t	r_saddr;
14379 	caddr_t r_eaddr;
14380 	size_t	r_size;
14381 	uchar_t	r_pgszc;
14382 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
14383 
14384 	ASSERT(sfmmup != ksfmmup);
14385 	ASSERT(srdp != NULL);
14386 	ASSERT(srdp->srd_refcnt > 0);
14387 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
14388 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
14389 	ASSERT(!sfmmup->sfmmu_free || sfmmup->sfmmu_scdp == NULL);
14390 
14391 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
14392 	    SFMMU_REGION_HME;
14393 
14394 	if (r_type == SFMMU_REGION_ISM) {
14395 		ASSERT(SFMMU_IS_ISMRID_VALID(rid));
14396 		ASSERT(rid < SFMMU_MAX_ISM_REGIONS);
14397 		rgnp = srdp->srd_ismrgnp[rid];
14398 	} else {
14399 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14400 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14401 		rgnp = srdp->srd_hmergnp[rid];
14402 	}
14403 	ASSERT(rgnp != NULL);
14404 	ASSERT(rgnp->rgn_id == rid);
14405 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14406 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14407 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
14408 
14409 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
14410 	if (r_type == SFMMU_REGION_HME && sfmmup->sfmmu_as->a_xhat != NULL) {
14411 		xhat_unload_callback_all(sfmmup->sfmmu_as, rgnp->rgn_saddr,
14412 		    rgnp->rgn_size, 0, NULL);
14413 	}
14414 
14415 	if (sfmmup->sfmmu_free) {
14416 		ulong_t rttecnt;
14417 		r_pgszc = rgnp->rgn_pgszc;
14418 		r_size = rgnp->rgn_size;
14419 
14420 		ASSERT(sfmmup->sfmmu_scdp == NULL);
14421 		if (r_type == SFMMU_REGION_ISM) {
14422 			SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14423 		} else {
14424 			/* update shme rgns ttecnt in sfmmu_ttecnt */
14425 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14426 			ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14427 
14428 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14429 			    -rttecnt);
14430 
14431 			SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14432 		}
14433 	} else if (r_type == SFMMU_REGION_ISM) {
14434 		hatlockp = sfmmu_hat_enter(sfmmup);
14435 		ASSERT(rid < srdp->srd_next_ismrid);
14436 		SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14437 		scdp = sfmmup->sfmmu_scdp;
14438 		if (scdp != NULL &&
14439 		    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
14440 			sfmmu_leave_scd(sfmmup, r_type);
14441 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14442 		}
14443 		sfmmu_hat_exit(hatlockp);
14444 	} else {
14445 		ulong_t rttecnt;
14446 		r_pgszc = rgnp->rgn_pgszc;
14447 		r_saddr = rgnp->rgn_saddr;
14448 		r_size = rgnp->rgn_size;
14449 		r_eaddr = r_saddr + r_size;
14450 
14451 		ASSERT(r_type == SFMMU_REGION_HME);
14452 		hatlockp = sfmmu_hat_enter(sfmmup);
14453 		ASSERT(rid < srdp->srd_next_hmerid);
14454 		SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14455 
14456 		/*
14457 		 * If region is part of an SCD call sfmmu_leave_scd().
14458 		 * Otherwise if process is not exiting and has valid context
14459 		 * just drop the context on the floor to lose stale TLB
14460 		 * entries and force the update of tsb miss area to reflect
14461 		 * the new region map. After that clean our TSB entries.
14462 		 */
14463 		scdp = sfmmup->sfmmu_scdp;
14464 		if (scdp != NULL &&
14465 		    SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
14466 			sfmmu_leave_scd(sfmmup, r_type);
14467 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14468 		}
14469 		sfmmu_invalidate_ctx(sfmmup);
14470 
14471 		i = TTE8K;
14472 		while (i < mmu_page_sizes) {
14473 			if (rgnp->rgn_ttecnt[i] != 0) {
14474 				sfmmu_unload_tsb_range(sfmmup, r_saddr,
14475 				    r_eaddr, i);
14476 				if (i < TTE4M) {
14477 					i = TTE4M;
14478 					continue;
14479 				} else {
14480 					break;
14481 				}
14482 			}
14483 			i++;
14484 		}
14485 		/* Remove the preallocated 1/4 8k ttecnt for 4M regions. */
14486 		if (r_pgszc >= TTE4M) {
14487 			rttecnt = r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14488 			ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
14489 			    rttecnt);
14490 			sfmmup->sfmmu_tsb0_4minflcnt -= rttecnt;
14491 		}
14492 
14493 		/* update shme rgns ttecnt in sfmmu_ttecnt */
14494 		rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14495 		ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14496 		atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc], -rttecnt);
14497 
14498 		sfmmu_hat_exit(hatlockp);
14499 		if (scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
14500 			/* sfmmup left the scd, grow private tsb */
14501 			sfmmu_check_page_sizes(sfmmup, 1);
14502 		} else {
14503 			sfmmu_check_page_sizes(sfmmup, 0);
14504 		}
14505 	}
14506 
14507 	if (r_type == SFMMU_REGION_HME) {
14508 		sfmmu_unlink_from_hmeregion(sfmmup, rgnp);
14509 	}
14510 
14511 	r_obj = rgnp->rgn_obj;
14512 	if (atomic_add_32_nv((volatile uint_t *)&rgnp->rgn_refcnt, -1)) {
14513 		return;
14514 	}
14515 
14516 	/*
14517 	 * looks like nobody uses this region anymore. Free it.
14518 	 */
14519 	rhash = RGN_HASH_FUNCTION(r_obj);
14520 	mutex_enter(&srdp->srd_mutex);
14521 	for (prev_rgnpp = &srdp->srd_rgnhash[rhash];
14522 	    (cur_rgnp = *prev_rgnpp) != NULL;
14523 	    prev_rgnpp = &cur_rgnp->rgn_hash) {
14524 		if (cur_rgnp == rgnp && cur_rgnp->rgn_refcnt == 0) {
14525 			break;
14526 		}
14527 	}
14528 
14529 	if (cur_rgnp == NULL) {
14530 		mutex_exit(&srdp->srd_mutex);
14531 		return;
14532 	}
14533 
14534 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14535 	*prev_rgnpp = rgnp->rgn_hash;
14536 	if (r_type == SFMMU_REGION_ISM) {
14537 		rgnp->rgn_flags |= SFMMU_REGION_FREE;
14538 		ASSERT(rid < srdp->srd_next_ismrid);
14539 		rgnp->rgn_next = srdp->srd_ismrgnfree;
14540 		srdp->srd_ismrgnfree = rgnp;
14541 		ASSERT(srdp->srd_ismbusyrgns > 0);
14542 		srdp->srd_ismbusyrgns--;
14543 		mutex_exit(&srdp->srd_mutex);
14544 		return;
14545 	}
14546 	mutex_exit(&srdp->srd_mutex);
14547 
14548 	/*
14549 	 * Destroy region's hmeblks.
14550 	 */
14551 	sfmmu_unload_hmeregion(srdp, rgnp);
14552 
14553 	rgnp->rgn_hmeflags = 0;
14554 
14555 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14556 	ASSERT(rgnp->rgn_id == rid);
14557 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14558 		rgnp->rgn_ttecnt[i] = 0;
14559 	}
14560 	rgnp->rgn_flags |= SFMMU_REGION_FREE;
14561 	mutex_enter(&srdp->srd_mutex);
14562 	ASSERT(rid < srdp->srd_next_hmerid);
14563 	rgnp->rgn_next = srdp->srd_hmergnfree;
14564 	srdp->srd_hmergnfree = rgnp;
14565 	ASSERT(srdp->srd_hmebusyrgns > 0);
14566 	srdp->srd_hmebusyrgns--;
14567 	mutex_exit(&srdp->srd_mutex);
14568 }
14569 
14570 /*
14571  * For now only called for hmeblk regions and not for ISM regions.
14572  */
14573 void
14574 hat_dup_region(struct hat *sfmmup, hat_region_cookie_t rcookie)
14575 {
14576 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14577 	uint_t rid = (uint_t)((uint64_t)rcookie);
14578 	sf_region_t *rgnp;
14579 	sf_rgn_link_t *rlink;
14580 	sf_rgn_link_t *hrlink;
14581 	ulong_t	rttecnt;
14582 
14583 	ASSERT(sfmmup != ksfmmup);
14584 	ASSERT(srdp != NULL);
14585 	ASSERT(srdp->srd_refcnt > 0);
14586 
14587 	ASSERT(rid < srdp->srd_next_hmerid);
14588 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14589 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14590 
14591 	rgnp = srdp->srd_hmergnp[rid];
14592 	ASSERT(rgnp->rgn_refcnt > 0);
14593 	ASSERT(rgnp->rgn_id == rid);
14594 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == SFMMU_REGION_HME);
14595 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14596 
14597 	atomic_add_32((volatile uint_t *)&rgnp->rgn_refcnt, 1);
14598 
14599 	/* LINTED: constant in conditional context */
14600 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 0);
14601 	ASSERT(rlink != NULL);
14602 	mutex_enter(&rgnp->rgn_mutex);
14603 	ASSERT(rgnp->rgn_sfmmu_head != NULL);
14604 	/* LINTED: constant in conditional context */
14605 	SFMMU_HMERID2RLINKP(rgnp->rgn_sfmmu_head, rid, hrlink, 0, 0);
14606 	ASSERT(hrlink != NULL);
14607 	ASSERT(hrlink->prev == NULL);
14608 	rlink->next = rgnp->rgn_sfmmu_head;
14609 	rlink->prev = NULL;
14610 	hrlink->prev = sfmmup;
14611 	/*
14612 	 * make sure rlink's next field is correct
14613 	 * before making this link visible.
14614 	 */
14615 	membar_stst();
14616 	rgnp->rgn_sfmmu_head = sfmmup;
14617 	mutex_exit(&rgnp->rgn_mutex);
14618 
14619 	/* update sfmmu_ttecnt with the shme rgn ttecnt */
14620 	rttecnt = rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
14621 	atomic_add_long(&sfmmup->sfmmu_ttecnt[rgnp->rgn_pgszc], rttecnt);
14622 	/* update tsb0 inflation count */
14623 	if (rgnp->rgn_pgszc >= TTE4M) {
14624 		sfmmup->sfmmu_tsb0_4minflcnt +=
14625 		    rgnp->rgn_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14626 	}
14627 	/*
14628 	 * Update regionid bitmask without hat lock since no other thread
14629 	 * can update this region bitmask right now.
14630 	 */
14631 	SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14632 }
14633 
14634 /* ARGSUSED */
14635 static int
14636 sfmmu_rgncache_constructor(void *buf, void *cdrarg, int kmflags)
14637 {
14638 	sf_region_t *rgnp = (sf_region_t *)buf;
14639 	bzero(buf, sizeof (*rgnp));
14640 
14641 	mutex_init(&rgnp->rgn_mutex, NULL, MUTEX_DEFAULT, NULL);
14642 
14643 	return (0);
14644 }
14645 
14646 /* ARGSUSED */
14647 static void
14648 sfmmu_rgncache_destructor(void *buf, void *cdrarg)
14649 {
14650 	sf_region_t *rgnp = (sf_region_t *)buf;
14651 	mutex_destroy(&rgnp->rgn_mutex);
14652 }
14653 
14654 static int
14655 sfrgnmap_isnull(sf_region_map_t *map)
14656 {
14657 	int i;
14658 
14659 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14660 		if (map->bitmap[i] != 0) {
14661 			return (0);
14662 		}
14663 	}
14664 	return (1);
14665 }
14666 
14667 static int
14668 sfhmergnmap_isnull(sf_hmeregion_map_t *map)
14669 {
14670 	int i;
14671 
14672 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
14673 		if (map->bitmap[i] != 0) {
14674 			return (0);
14675 		}
14676 	}
14677 	return (1);
14678 }
14679 
14680 #ifdef DEBUG
14681 static void
14682 check_scd_sfmmu_list(sfmmu_t **headp, sfmmu_t *sfmmup, int onlist)
14683 {
14684 	sfmmu_t *sp;
14685 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14686 
14687 	for (sp = *headp; sp != NULL; sp = sp->sfmmu_scd_link.next) {
14688 		ASSERT(srdp == sp->sfmmu_srdp);
14689 		if (sp == sfmmup) {
14690 			if (onlist) {
14691 				return;
14692 			} else {
14693 				panic("shctx: sfmmu 0x%p found on scd"
14694 				    "list 0x%p", (void *)sfmmup,
14695 				    (void *)*headp);
14696 			}
14697 		}
14698 	}
14699 	if (onlist) {
14700 		panic("shctx: sfmmu 0x%p not found on scd list 0x%p",
14701 		    (void *)sfmmup, (void *)*headp);
14702 	} else {
14703 		return;
14704 	}
14705 }
14706 #else /* DEBUG */
14707 #define	check_scd_sfmmu_list(headp, sfmmup, onlist)
14708 #endif /* DEBUG */
14709 
14710 /*
14711  * Removes an sfmmu from the SCD sfmmu list.
14712  */
14713 static void
14714 sfmmu_from_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14715 {
14716 	ASSERT(sfmmup->sfmmu_srdp != NULL);
14717 	check_scd_sfmmu_list(headp, sfmmup, 1);
14718 	if (sfmmup->sfmmu_scd_link.prev != NULL) {
14719 		ASSERT(*headp != sfmmup);
14720 		sfmmup->sfmmu_scd_link.prev->sfmmu_scd_link.next =
14721 		    sfmmup->sfmmu_scd_link.next;
14722 	} else {
14723 		ASSERT(*headp == sfmmup);
14724 		*headp = sfmmup->sfmmu_scd_link.next;
14725 	}
14726 	if (sfmmup->sfmmu_scd_link.next != NULL) {
14727 		sfmmup->sfmmu_scd_link.next->sfmmu_scd_link.prev =
14728 		    sfmmup->sfmmu_scd_link.prev;
14729 	}
14730 }
14731 
14732 
14733 /*
14734  * Adds an sfmmu to the start of the queue.
14735  */
14736 static void
14737 sfmmu_to_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14738 {
14739 	check_scd_sfmmu_list(headp, sfmmup, 0);
14740 	sfmmup->sfmmu_scd_link.prev = NULL;
14741 	sfmmup->sfmmu_scd_link.next = *headp;
14742 	if (*headp != NULL)
14743 		(*headp)->sfmmu_scd_link.prev = sfmmup;
14744 	*headp = sfmmup;
14745 }
14746 
14747 /*
14748  * Remove an scd from the start of the queue.
14749  */
14750 static void
14751 sfmmu_remove_scd(sf_scd_t **headp, sf_scd_t *scdp)
14752 {
14753 	if (scdp->scd_prev != NULL) {
14754 		ASSERT(*headp != scdp);
14755 		scdp->scd_prev->scd_next = scdp->scd_next;
14756 	} else {
14757 		ASSERT(*headp == scdp);
14758 		*headp = scdp->scd_next;
14759 	}
14760 
14761 	if (scdp->scd_next != NULL) {
14762 		scdp->scd_next->scd_prev = scdp->scd_prev;
14763 	}
14764 }
14765 
14766 /*
14767  * Add an scd to the start of the queue.
14768  */
14769 static void
14770 sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *scdp)
14771 {
14772 	scdp->scd_prev = NULL;
14773 	scdp->scd_next = *headp;
14774 	if (*headp != NULL) {
14775 		(*headp)->scd_prev = scdp;
14776 	}
14777 	*headp = scdp;
14778 }
14779 
14780 static int
14781 sfmmu_alloc_scd_tsbs(sf_srd_t *srdp, sf_scd_t *scdp)
14782 {
14783 	uint_t rid;
14784 	uint_t i;
14785 	uint_t j;
14786 	ulong_t w;
14787 	sf_region_t *rgnp;
14788 	ulong_t tte8k_cnt = 0;
14789 	ulong_t tte4m_cnt = 0;
14790 	uint_t tsb_szc;
14791 	sfmmu_t *scsfmmup = scdp->scd_sfmmup;
14792 	sfmmu_t	*ism_hatid;
14793 	struct tsb_info *newtsb;
14794 	int szc;
14795 
14796 	ASSERT(srdp != NULL);
14797 
14798 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14799 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14800 			continue;
14801 		}
14802 		j = 0;
14803 		while (w) {
14804 			if (!(w & 0x1)) {
14805 				j++;
14806 				w >>= 1;
14807 				continue;
14808 			}
14809 			rid = (i << BT_ULSHIFT) | j;
14810 			j++;
14811 			w >>= 1;
14812 
14813 			if (rid < SFMMU_MAX_HME_REGIONS) {
14814 				rgnp = srdp->srd_hmergnp[rid];
14815 				ASSERT(rgnp->rgn_id == rid);
14816 				ASSERT(rgnp->rgn_refcnt > 0);
14817 
14818 				if (rgnp->rgn_pgszc < TTE4M) {
14819 					tte8k_cnt += rgnp->rgn_size >>
14820 					    TTE_PAGE_SHIFT(TTE8K);
14821 				} else {
14822 					ASSERT(rgnp->rgn_pgszc >= TTE4M);
14823 					tte4m_cnt += rgnp->rgn_size >>
14824 					    TTE_PAGE_SHIFT(TTE4M);
14825 					/*
14826 					 * Inflate SCD tsb0 by preallocating
14827 					 * 1/4 8k ttecnt for 4M regions to
14828 					 * allow for lgpg alloc failure.
14829 					 */
14830 					tte8k_cnt += rgnp->rgn_size >>
14831 					    (TTE_PAGE_SHIFT(TTE8K) + 2);
14832 				}
14833 			} else {
14834 				rid -= SFMMU_MAX_HME_REGIONS;
14835 				rgnp = srdp->srd_ismrgnp[rid];
14836 				ASSERT(rgnp->rgn_id == rid);
14837 				ASSERT(rgnp->rgn_refcnt > 0);
14838 
14839 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14840 				ASSERT(ism_hatid->sfmmu_ismhat);
14841 
14842 				for (szc = 0; szc < TTE4M; szc++) {
14843 					tte8k_cnt +=
14844 					    ism_hatid->sfmmu_ttecnt[szc] <<
14845 					    TTE_BSZS_SHIFT(szc);
14846 				}
14847 
14848 				ASSERT(rgnp->rgn_pgszc >= TTE4M);
14849 				if (rgnp->rgn_pgszc >= TTE4M) {
14850 					tte4m_cnt += rgnp->rgn_size >>
14851 					    TTE_PAGE_SHIFT(TTE4M);
14852 				}
14853 			}
14854 		}
14855 	}
14856 
14857 	tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
14858 
14859 	/* Allocate both the SCD TSBs here. */
14860 	if (sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14861 	    tsb_szc, TSB8K|TSB64K|TSB512K, TSB_ALLOC, scsfmmup) &&
14862 	    (tsb_szc <= TSB_4M_SZCODE ||
14863 	    sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14864 	    TSB_4M_SZCODE, TSB8K|TSB64K|TSB512K,
14865 	    TSB_ALLOC, scsfmmup))) {
14866 
14867 		SFMMU_STAT(sf_scd_1sttsb_allocfail);
14868 		return (TSB_ALLOCFAIL);
14869 	} else {
14870 		scsfmmup->sfmmu_tsb->tsb_flags |= TSB_SHAREDCTX;
14871 
14872 		if (tte4m_cnt) {
14873 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
14874 			if (sfmmu_tsbinfo_alloc(&newtsb, tsb_szc,
14875 			    TSB4M|TSB32M|TSB256M, TSB_ALLOC, scsfmmup) &&
14876 			    (tsb_szc <= TSB_4M_SZCODE ||
14877 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
14878 			    TSB4M|TSB32M|TSB256M,
14879 			    TSB_ALLOC, scsfmmup))) {
14880 				/*
14881 				 * If we fail to allocate the 2nd shared tsb,
14882 				 * just free the 1st tsb, return failure.
14883 				 */
14884 				sfmmu_tsbinfo_free(scsfmmup->sfmmu_tsb);
14885 				SFMMU_STAT(sf_scd_2ndtsb_allocfail);
14886 				return (TSB_ALLOCFAIL);
14887 			} else {
14888 				ASSERT(scsfmmup->sfmmu_tsb->tsb_next == NULL);
14889 				newtsb->tsb_flags |= TSB_SHAREDCTX;
14890 				scsfmmup->sfmmu_tsb->tsb_next = newtsb;
14891 				SFMMU_STAT(sf_scd_2ndtsb_alloc);
14892 			}
14893 		}
14894 		SFMMU_STAT(sf_scd_1sttsb_alloc);
14895 	}
14896 	return (TSB_SUCCESS);
14897 }
14898 
14899 static void
14900 sfmmu_free_scd_tsbs(sfmmu_t *scd_sfmmu)
14901 {
14902 	while (scd_sfmmu->sfmmu_tsb != NULL) {
14903 		struct tsb_info *next = scd_sfmmu->sfmmu_tsb->tsb_next;
14904 		sfmmu_tsbinfo_free(scd_sfmmu->sfmmu_tsb);
14905 		scd_sfmmu->sfmmu_tsb = next;
14906 	}
14907 }
14908 
14909 /*
14910  * Link the sfmmu onto the hme region list.
14911  */
14912 void
14913 sfmmu_link_to_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14914 {
14915 	uint_t rid;
14916 	sf_rgn_link_t *rlink;
14917 	sfmmu_t *head;
14918 	sf_rgn_link_t *hrlink;
14919 
14920 	rid = rgnp->rgn_id;
14921 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14922 
14923 	/* LINTED: constant in conditional context */
14924 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 1);
14925 	ASSERT(rlink != NULL);
14926 	mutex_enter(&rgnp->rgn_mutex);
14927 	if ((head = rgnp->rgn_sfmmu_head) == NULL) {
14928 		rlink->next = NULL;
14929 		rlink->prev = NULL;
14930 		/*
14931 		 * make sure rlink's next field is NULL
14932 		 * before making this link visible.
14933 		 */
14934 		membar_stst();
14935 		rgnp->rgn_sfmmu_head = sfmmup;
14936 	} else {
14937 		/* LINTED: constant in conditional context */
14938 		SFMMU_HMERID2RLINKP(head, rid, hrlink, 0, 0);
14939 		ASSERT(hrlink != NULL);
14940 		ASSERT(hrlink->prev == NULL);
14941 		rlink->next = head;
14942 		rlink->prev = NULL;
14943 		hrlink->prev = sfmmup;
14944 		/*
14945 		 * make sure rlink's next field is correct
14946 		 * before making this link visible.
14947 		 */
14948 		membar_stst();
14949 		rgnp->rgn_sfmmu_head = sfmmup;
14950 	}
14951 	mutex_exit(&rgnp->rgn_mutex);
14952 }
14953 
14954 /*
14955  * Unlink the sfmmu from the hme region list.
14956  */
14957 void
14958 sfmmu_unlink_from_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14959 {
14960 	uint_t rid;
14961 	sf_rgn_link_t *rlink;
14962 
14963 	rid = rgnp->rgn_id;
14964 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14965 
14966 	/* LINTED: constant in conditional context */
14967 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
14968 	ASSERT(rlink != NULL);
14969 	mutex_enter(&rgnp->rgn_mutex);
14970 	if (rgnp->rgn_sfmmu_head == sfmmup) {
14971 		sfmmu_t *next = rlink->next;
14972 		rgnp->rgn_sfmmu_head = next;
14973 		/*
14974 		 * if we are stopped by xc_attention() after this
14975 		 * point the forward link walking in
14976 		 * sfmmu_rgntlb_demap() will work correctly since the
14977 		 * head correctly points to the next element.
14978 		 */
14979 		membar_stst();
14980 		rlink->next = NULL;
14981 		ASSERT(rlink->prev == NULL);
14982 		if (next != NULL) {
14983 			sf_rgn_link_t *nrlink;
14984 			/* LINTED: constant in conditional context */
14985 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14986 			ASSERT(nrlink != NULL);
14987 			ASSERT(nrlink->prev == sfmmup);
14988 			nrlink->prev = NULL;
14989 		}
14990 	} else {
14991 		sfmmu_t *next = rlink->next;
14992 		sfmmu_t *prev = rlink->prev;
14993 		sf_rgn_link_t *prlink;
14994 
14995 		ASSERT(prev != NULL);
14996 		/* LINTED: constant in conditional context */
14997 		SFMMU_HMERID2RLINKP(prev, rid, prlink, 0, 0);
14998 		ASSERT(prlink != NULL);
14999 		ASSERT(prlink->next == sfmmup);
15000 		prlink->next = next;
15001 		/*
15002 		 * if we are stopped by xc_attention()
15003 		 * after this point the forward link walking
15004 		 * will work correctly since the prev element
15005 		 * correctly points to the next element.
15006 		 */
15007 		membar_stst();
15008 		rlink->next = NULL;
15009 		rlink->prev = NULL;
15010 		if (next != NULL) {
15011 			sf_rgn_link_t *nrlink;
15012 			/* LINTED: constant in conditional context */
15013 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
15014 			ASSERT(nrlink != NULL);
15015 			ASSERT(nrlink->prev == sfmmup);
15016 			nrlink->prev = prev;
15017 		}
15018 	}
15019 	mutex_exit(&rgnp->rgn_mutex);
15020 }
15021 
15022 /*
15023  * Link scd sfmmu onto ism or hme region list for each region in the
15024  * scd region map.
15025  */
15026 void
15027 sfmmu_link_scd_to_regions(sf_srd_t *srdp, sf_scd_t *scdp)
15028 {
15029 	uint_t rid;
15030 	uint_t i;
15031 	uint_t j;
15032 	ulong_t w;
15033 	sf_region_t *rgnp;
15034 	sfmmu_t *scsfmmup;
15035 
15036 	scsfmmup = scdp->scd_sfmmup;
15037 	ASSERT(scsfmmup->sfmmu_scdhat);
15038 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
15039 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
15040 			continue;
15041 		}
15042 		j = 0;
15043 		while (w) {
15044 			if (!(w & 0x1)) {
15045 				j++;
15046 				w >>= 1;
15047 				continue;
15048 			}
15049 			rid = (i << BT_ULSHIFT) | j;
15050 			j++;
15051 			w >>= 1;
15052 
15053 			if (rid < SFMMU_MAX_HME_REGIONS) {
15054 				rgnp = srdp->srd_hmergnp[rid];
15055 				ASSERT(rgnp->rgn_id == rid);
15056 				ASSERT(rgnp->rgn_refcnt > 0);
15057 				sfmmu_link_to_hmeregion(scsfmmup, rgnp);
15058 			} else {
15059 				sfmmu_t *ism_hatid = NULL;
15060 				ism_ment_t *ism_ment;
15061 				rid -= SFMMU_MAX_HME_REGIONS;
15062 				rgnp = srdp->srd_ismrgnp[rid];
15063 				ASSERT(rgnp->rgn_id == rid);
15064 				ASSERT(rgnp->rgn_refcnt > 0);
15065 
15066 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
15067 				ASSERT(ism_hatid->sfmmu_ismhat);
15068 				ism_ment = &scdp->scd_ism_links[rid];
15069 				ism_ment->iment_hat = scsfmmup;
15070 				ism_ment->iment_base_va = rgnp->rgn_saddr;
15071 				mutex_enter(&ism_mlist_lock);
15072 				iment_add(ism_ment, ism_hatid);
15073 				mutex_exit(&ism_mlist_lock);
15074 
15075 			}
15076 		}
15077 	}
15078 }
15079 /*
15080  * Unlink scd sfmmu from ism or hme region list for each region in the
15081  * scd region map.
15082  */
15083 void
15084 sfmmu_unlink_scd_from_regions(sf_srd_t *srdp, sf_scd_t *scdp)
15085 {
15086 	uint_t rid;
15087 	uint_t i;
15088 	uint_t j;
15089 	ulong_t w;
15090 	sf_region_t *rgnp;
15091 	sfmmu_t *scsfmmup;
15092 
15093 	scsfmmup = scdp->scd_sfmmup;
15094 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
15095 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
15096 			continue;
15097 		}
15098 		j = 0;
15099 		while (w) {
15100 			if (!(w & 0x1)) {
15101 				j++;
15102 				w >>= 1;
15103 				continue;
15104 			}
15105 			rid = (i << BT_ULSHIFT) | j;
15106 			j++;
15107 			w >>= 1;
15108 
15109 			if (rid < SFMMU_MAX_HME_REGIONS) {
15110 				rgnp = srdp->srd_hmergnp[rid];
15111 				ASSERT(rgnp->rgn_id == rid);
15112 				ASSERT(rgnp->rgn_refcnt > 0);
15113 				sfmmu_unlink_from_hmeregion(scsfmmup,
15114 				    rgnp);
15115 
15116 			} else {
15117 				sfmmu_t *ism_hatid = NULL;
15118 				ism_ment_t *ism_ment;
15119 				rid -= SFMMU_MAX_HME_REGIONS;
15120 				rgnp = srdp->srd_ismrgnp[rid];
15121 				ASSERT(rgnp->rgn_id == rid);
15122 				ASSERT(rgnp->rgn_refcnt > 0);
15123 
15124 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
15125 				ASSERT(ism_hatid->sfmmu_ismhat);
15126 				ism_ment = &scdp->scd_ism_links[rid];
15127 				ASSERT(ism_ment->iment_hat == scdp->scd_sfmmup);
15128 				ASSERT(ism_ment->iment_base_va ==
15129 				    rgnp->rgn_saddr);
15130 				mutex_enter(&ism_mlist_lock);
15131 				iment_sub(ism_ment, ism_hatid);
15132 				mutex_exit(&ism_mlist_lock);
15133 
15134 			}
15135 		}
15136 	}
15137 }
15138 /*
15139  * Allocates and initialises a new SCD structure, this is called with
15140  * the srd_scd_mutex held and returns with the reference count
15141  * initialised to 1.
15142  */
15143 static sf_scd_t *
15144 sfmmu_alloc_scd(sf_srd_t *srdp, sf_region_map_t *new_map)
15145 {
15146 	sf_scd_t *new_scdp;
15147 	sfmmu_t *scsfmmup;
15148 	int i;
15149 
15150 	ASSERT(MUTEX_HELD(&srdp->srd_scd_mutex));
15151 	new_scdp = kmem_cache_alloc(scd_cache, KM_SLEEP);
15152 
15153 	scsfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
15154 	new_scdp->scd_sfmmup = scsfmmup;
15155 	scsfmmup->sfmmu_srdp = srdp;
15156 	scsfmmup->sfmmu_scdp = new_scdp;
15157 	scsfmmup->sfmmu_tsb0_4minflcnt = 0;
15158 	scsfmmup->sfmmu_scdhat = 1;
15159 	CPUSET_ALL(scsfmmup->sfmmu_cpusran);
15160 	bzero(scsfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
15161 
15162 	ASSERT(max_mmu_ctxdoms > 0);
15163 	for (i = 0; i < max_mmu_ctxdoms; i++) {
15164 		scsfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
15165 		scsfmmup->sfmmu_ctxs[i].gnum = 0;
15166 	}
15167 
15168 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
15169 		new_scdp->scd_rttecnt[i] = 0;
15170 	}
15171 
15172 	new_scdp->scd_region_map = *new_map;
15173 	new_scdp->scd_refcnt = 1;
15174 	if (sfmmu_alloc_scd_tsbs(srdp, new_scdp) != TSB_SUCCESS) {
15175 		kmem_cache_free(scd_cache, new_scdp);
15176 		kmem_cache_free(sfmmuid_cache, scsfmmup);
15177 		return (NULL);
15178 	}
15179 	if (&mmu_init_scd) {
15180 		mmu_init_scd(new_scdp);
15181 	}
15182 	return (new_scdp);
15183 }
15184 
15185 /*
15186  * The first phase of a process joining an SCD. The hat structure is
15187  * linked to the SCD queue and then the HAT_JOIN_SCD sfmmu flag is set
15188  * and a cross-call with context invalidation is used to cause the
15189  * remaining work to be carried out in the sfmmu_tsbmiss_exception()
15190  * routine.
15191  */
15192 static void
15193 sfmmu_join_scd(sf_scd_t *scdp, sfmmu_t *sfmmup)
15194 {
15195 	hatlock_t *hatlockp;
15196 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15197 	int i;
15198 	sf_scd_t *old_scdp;
15199 
15200 	ASSERT(srdp != NULL);
15201 	ASSERT(scdp != NULL);
15202 	ASSERT(scdp->scd_refcnt > 0);
15203 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15204 
15205 	if ((old_scdp = sfmmup->sfmmu_scdp) != NULL) {
15206 		ASSERT(old_scdp != scdp);
15207 
15208 		mutex_enter(&old_scdp->scd_mutex);
15209 		sfmmu_from_scd_list(&old_scdp->scd_sf_list, sfmmup);
15210 		mutex_exit(&old_scdp->scd_mutex);
15211 		/*
15212 		 * sfmmup leaves the old scd. Update sfmmu_ttecnt to
15213 		 * include the shme rgn ttecnt for rgns that
15214 		 * were in the old SCD
15215 		 */
15216 		for (i = 0; i < mmu_page_sizes; i++) {
15217 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15218 			    old_scdp->scd_rttecnt[i]);
15219 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15220 			    sfmmup->sfmmu_scdrttecnt[i]);
15221 		}
15222 	}
15223 
15224 	/*
15225 	 * Move sfmmu to the scd lists.
15226 	 */
15227 	mutex_enter(&scdp->scd_mutex);
15228 	sfmmu_to_scd_list(&scdp->scd_sf_list, sfmmup);
15229 	mutex_exit(&scdp->scd_mutex);
15230 	SF_SCD_INCR_REF(scdp);
15231 
15232 	hatlockp = sfmmu_hat_enter(sfmmup);
15233 	/*
15234 	 * For a multi-thread process, we must stop
15235 	 * all the other threads before joining the scd.
15236 	 */
15237 
15238 	SFMMU_FLAGS_SET(sfmmup, HAT_JOIN_SCD);
15239 
15240 	sfmmu_invalidate_ctx(sfmmup);
15241 	sfmmup->sfmmu_scdp = scdp;
15242 
15243 	/*
15244 	 * Copy scd_rttecnt into sfmmup's sfmmu_scdrttecnt, and update
15245 	 * sfmmu_ttecnt to not include the rgn ttecnt just joined in SCD.
15246 	 */
15247 	for (i = 0; i < mmu_page_sizes; i++) {
15248 		sfmmup->sfmmu_scdrttecnt[i] = scdp->scd_rttecnt[i];
15249 		ASSERT(sfmmup->sfmmu_ttecnt[i] >= scdp->scd_rttecnt[i]);
15250 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15251 		    -sfmmup->sfmmu_scdrttecnt[i]);
15252 	}
15253 	/* update tsb0 inflation count */
15254 	if (old_scdp != NULL) {
15255 		sfmmup->sfmmu_tsb0_4minflcnt +=
15256 		    old_scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15257 	}
15258 	ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
15259 	    scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt);
15260 	sfmmup->sfmmu_tsb0_4minflcnt -= scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15261 
15262 	sfmmu_hat_exit(hatlockp);
15263 
15264 	if (old_scdp != NULL) {
15265 		SF_SCD_DECR_REF(srdp, old_scdp);
15266 	}
15267 
15268 }
15269 
15270 /*
15271  * This routine is called by a process to become part of an SCD. It is called
15272  * from sfmmu_tsbmiss_exception() once most of the initial work has been
15273  * done by sfmmu_join_scd(). This routine must not drop the hat lock.
15274  */
15275 static void
15276 sfmmu_finish_join_scd(sfmmu_t *sfmmup)
15277 {
15278 	struct tsb_info	*tsbinfop;
15279 
15280 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15281 	ASSERT(sfmmup->sfmmu_scdp != NULL);
15282 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD));
15283 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15284 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID));
15285 
15286 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
15287 	    tsbinfop = tsbinfop->tsb_next) {
15288 		if (tsbinfop->tsb_flags & TSB_SWAPPED) {
15289 			continue;
15290 		}
15291 		ASSERT(!(tsbinfop->tsb_flags & TSB_RELOC_FLAG));
15292 
15293 		sfmmu_inv_tsb(tsbinfop->tsb_va,
15294 		    TSB_BYTES(tsbinfop->tsb_szc));
15295 	}
15296 
15297 	/* Set HAT_CTX1_FLAG for all SCD ISMs */
15298 	sfmmu_ism_hatflags(sfmmup, 1);
15299 
15300 	SFMMU_STAT(sf_join_scd);
15301 }
15302 
15303 /*
15304  * This routine is called in order to check if there is an SCD which matches
15305  * the process's region map if not then a new SCD may be created.
15306  */
15307 static void
15308 sfmmu_find_scd(sfmmu_t *sfmmup)
15309 {
15310 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15311 	sf_scd_t *scdp, *new_scdp;
15312 	int ret;
15313 
15314 	ASSERT(srdp != NULL);
15315 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15316 
15317 	mutex_enter(&srdp->srd_scd_mutex);
15318 	for (scdp = srdp->srd_scdp; scdp != NULL;
15319 	    scdp = scdp->scd_next) {
15320 		SF_RGNMAP_EQUAL(&scdp->scd_region_map,
15321 		    &sfmmup->sfmmu_region_map, ret);
15322 		if (ret == 1) {
15323 			SF_SCD_INCR_REF(scdp);
15324 			mutex_exit(&srdp->srd_scd_mutex);
15325 			sfmmu_join_scd(scdp, sfmmup);
15326 			ASSERT(scdp->scd_refcnt >= 2);
15327 			atomic_add_32((volatile uint32_t *)
15328 			    &scdp->scd_refcnt, -1);
15329 			return;
15330 		} else {
15331 			/*
15332 			 * If the sfmmu region map is a subset of the scd
15333 			 * region map, then the assumption is that this process
15334 			 * will continue attaching to ISM segments until the
15335 			 * region maps are equal.
15336 			 */
15337 			SF_RGNMAP_IS_SUBSET(&scdp->scd_region_map,
15338 			    &sfmmup->sfmmu_region_map, ret);
15339 			if (ret == 1) {
15340 				mutex_exit(&srdp->srd_scd_mutex);
15341 				return;
15342 			}
15343 		}
15344 	}
15345 
15346 	ASSERT(scdp == NULL);
15347 	/*
15348 	 * No matching SCD has been found, create a new one.
15349 	 */
15350 	if ((new_scdp = sfmmu_alloc_scd(srdp, &sfmmup->sfmmu_region_map)) ==
15351 	    NULL) {
15352 		mutex_exit(&srdp->srd_scd_mutex);
15353 		return;
15354 	}
15355 
15356 	/*
15357 	 * sfmmu_alloc_scd() returns with a ref count of 1 on the scd.
15358 	 */
15359 
15360 	/* Set scd_rttecnt for shme rgns in SCD */
15361 	sfmmu_set_scd_rttecnt(srdp, new_scdp);
15362 
15363 	/*
15364 	 * Link scd onto srd_scdp list and scd sfmmu onto region/iment lists.
15365 	 */
15366 	sfmmu_link_scd_to_regions(srdp, new_scdp);
15367 	sfmmu_add_scd(&srdp->srd_scdp, new_scdp);
15368 	SFMMU_STAT_ADD(sf_create_scd, 1);
15369 
15370 	mutex_exit(&srdp->srd_scd_mutex);
15371 	sfmmu_join_scd(new_scdp, sfmmup);
15372 	ASSERT(new_scdp->scd_refcnt >= 2);
15373 	atomic_add_32((volatile uint32_t *)&new_scdp->scd_refcnt, -1);
15374 }
15375 
15376 /*
15377  * This routine is called by a process to remove itself from an SCD. It is
15378  * either called when the processes has detached from a segment or from
15379  * hat_free_start() as a result of calling exit.
15380  */
15381 static void
15382 sfmmu_leave_scd(sfmmu_t *sfmmup, uchar_t r_type)
15383 {
15384 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15385 	sf_srd_t *srdp =  sfmmup->sfmmu_srdp;
15386 	hatlock_t *hatlockp = TSB_HASH(sfmmup);
15387 	int i;
15388 
15389 	ASSERT(scdp != NULL);
15390 	ASSERT(srdp != NULL);
15391 
15392 	if (sfmmup->sfmmu_free) {
15393 		/*
15394 		 * If the process is part of an SCD the sfmmu is unlinked
15395 		 * from scd_sf_list.
15396 		 */
15397 		mutex_enter(&scdp->scd_mutex);
15398 		sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15399 		mutex_exit(&scdp->scd_mutex);
15400 		/*
15401 		 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15402 		 * are about to leave the SCD
15403 		 */
15404 		for (i = 0; i < mmu_page_sizes; i++) {
15405 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15406 			    scdp->scd_rttecnt[i]);
15407 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15408 			    sfmmup->sfmmu_scdrttecnt[i]);
15409 			sfmmup->sfmmu_scdrttecnt[i] = 0;
15410 		}
15411 		sfmmup->sfmmu_scdp = NULL;
15412 
15413 		SF_SCD_DECR_REF(srdp, scdp);
15414 		return;
15415 	}
15416 
15417 	ASSERT(r_type != SFMMU_REGION_ISM ||
15418 	    SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15419 	ASSERT(scdp->scd_refcnt);
15420 	ASSERT(!sfmmup->sfmmu_free);
15421 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15422 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15423 
15424 	/*
15425 	 * Wait for ISM maps to be updated.
15426 	 */
15427 	if (r_type != SFMMU_REGION_ISM) {
15428 		while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY) &&
15429 		    sfmmup->sfmmu_scdp != NULL) {
15430 			cv_wait(&sfmmup->sfmmu_tsb_cv,
15431 			    HATLOCK_MUTEXP(hatlockp));
15432 		}
15433 
15434 		if (sfmmup->sfmmu_scdp == NULL) {
15435 			sfmmu_hat_exit(hatlockp);
15436 			return;
15437 		}
15438 		SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
15439 	}
15440 
15441 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
15442 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
15443 		/*
15444 		 * Since HAT_JOIN_SCD was set our context
15445 		 * is still invalid.
15446 		 */
15447 	} else {
15448 		/*
15449 		 * For a multi-thread process, we must stop
15450 		 * all the other threads before leaving the scd.
15451 		 */
15452 
15453 		sfmmu_invalidate_ctx(sfmmup);
15454 	}
15455 
15456 	/* Clear all the rid's for ISM, delete flags, etc */
15457 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15458 	sfmmu_ism_hatflags(sfmmup, 0);
15459 
15460 	/*
15461 	 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15462 	 * are in SCD before this sfmmup leaves the SCD.
15463 	 */
15464 	for (i = 0; i < mmu_page_sizes; i++) {
15465 		ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15466 		    scdp->scd_rttecnt[i]);
15467 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15468 		    sfmmup->sfmmu_scdrttecnt[i]);
15469 		sfmmup->sfmmu_scdrttecnt[i] = 0;
15470 		/* update ismttecnt to include SCD ism before hat leaves SCD */
15471 		sfmmup->sfmmu_ismttecnt[i] += sfmmup->sfmmu_scdismttecnt[i];
15472 		sfmmup->sfmmu_scdismttecnt[i] = 0;
15473 	}
15474 	/* update tsb0 inflation count */
15475 	sfmmup->sfmmu_tsb0_4minflcnt += scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15476 
15477 	if (r_type != SFMMU_REGION_ISM) {
15478 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
15479 	}
15480 	sfmmup->sfmmu_scdp = NULL;
15481 
15482 	sfmmu_hat_exit(hatlockp);
15483 
15484 	/*
15485 	 * Unlink sfmmu from scd_sf_list this can be done without holding
15486 	 * the hat lock as we hold the sfmmu_as lock which prevents
15487 	 * hat_join_region from adding this thread to the scd again. Other
15488 	 * threads check if sfmmu_scdp is NULL under hat lock and if it's NULL
15489 	 * they won't get here, since sfmmu_leave_scd() clears sfmmu_scdp
15490 	 * while holding the hat lock.
15491 	 */
15492 	mutex_enter(&scdp->scd_mutex);
15493 	sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15494 	mutex_exit(&scdp->scd_mutex);
15495 	SFMMU_STAT(sf_leave_scd);
15496 
15497 	SF_SCD_DECR_REF(srdp, scdp);
15498 	hatlockp = sfmmu_hat_enter(sfmmup);
15499 
15500 }
15501 
15502 /*
15503  * Unlink and free up an SCD structure with a reference count of 0.
15504  */
15505 static void
15506 sfmmu_destroy_scd(sf_srd_t *srdp, sf_scd_t *scdp, sf_region_map_t *scd_rmap)
15507 {
15508 	sfmmu_t *scsfmmup;
15509 	sf_scd_t *sp;
15510 	hatlock_t *shatlockp;
15511 	int i, ret;
15512 
15513 	mutex_enter(&srdp->srd_scd_mutex);
15514 	for (sp = srdp->srd_scdp; sp != NULL; sp = sp->scd_next) {
15515 		if (sp == scdp)
15516 			break;
15517 	}
15518 	if (sp == NULL || sp->scd_refcnt) {
15519 		mutex_exit(&srdp->srd_scd_mutex);
15520 		return;
15521 	}
15522 
15523 	/*
15524 	 * It is possible that the scd has been freed and reallocated with a
15525 	 * different region map while we've been waiting for the srd_scd_mutex.
15526 	 */
15527 	SF_RGNMAP_EQUAL(scd_rmap, &sp->scd_region_map, ret);
15528 	if (ret != 1) {
15529 		mutex_exit(&srdp->srd_scd_mutex);
15530 		return;
15531 	}
15532 
15533 	ASSERT(scdp->scd_sf_list == NULL);
15534 	/*
15535 	 * Unlink scd from srd_scdp list.
15536 	 */
15537 	sfmmu_remove_scd(&srdp->srd_scdp, scdp);
15538 	mutex_exit(&srdp->srd_scd_mutex);
15539 
15540 	sfmmu_unlink_scd_from_regions(srdp, scdp);
15541 
15542 	/* Clear shared context tsb and release ctx */
15543 	scsfmmup = scdp->scd_sfmmup;
15544 
15545 	/*
15546 	 * create a barrier so that scd will not be destroyed
15547 	 * if other thread still holds the same shared hat lock.
15548 	 * E.g., sfmmu_tsbmiss_exception() needs to acquire the
15549 	 * shared hat lock before checking the shared tsb reloc flag.
15550 	 */
15551 	shatlockp = sfmmu_hat_enter(scsfmmup);
15552 	sfmmu_hat_exit(shatlockp);
15553 
15554 	sfmmu_free_scd_tsbs(scsfmmup);
15555 
15556 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
15557 		if (scsfmmup->sfmmu_hmeregion_links[i] != NULL) {
15558 			kmem_free(scsfmmup->sfmmu_hmeregion_links[i],
15559 			    SFMMU_L2_HMERLINKS_SIZE);
15560 			scsfmmup->sfmmu_hmeregion_links[i] = NULL;
15561 		}
15562 	}
15563 	kmem_cache_free(sfmmuid_cache, scsfmmup);
15564 	kmem_cache_free(scd_cache, scdp);
15565 	SFMMU_STAT(sf_destroy_scd);
15566 }
15567 
15568 /*
15569  * Modifies the HAT_CTX1_FLAG for each of the ISM segments which correspond to
15570  * bits which are set in the ism_region_map parameter. This flag indicates to
15571  * the tsbmiss handler that mapping for these segments should be loaded using
15572  * the shared context.
15573  */
15574 static void
15575 sfmmu_ism_hatflags(sfmmu_t *sfmmup, int addflag)
15576 {
15577 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15578 	ism_blk_t *ism_blkp;
15579 	ism_map_t *ism_map;
15580 	int i, rid;
15581 
15582 	ASSERT(sfmmup->sfmmu_iblk != NULL);
15583 	ASSERT(scdp != NULL);
15584 	/*
15585 	 * Note that the caller either set HAT_ISMBUSY flag or checked
15586 	 * under hat lock that HAT_ISMBUSY was not set by another thread.
15587 	 */
15588 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15589 
15590 	ism_blkp = sfmmup->sfmmu_iblk;
15591 	while (ism_blkp != NULL) {
15592 		ism_map = ism_blkp->iblk_maps;
15593 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
15594 			rid = ism_map[i].imap_rid;
15595 			if (rid == SFMMU_INVALID_ISMRID) {
15596 				continue;
15597 			}
15598 			ASSERT(rid >= 0 && rid < SFMMU_MAX_ISM_REGIONS);
15599 			if (SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid) &&
15600 			    addflag) {
15601 				ism_map[i].imap_hatflags |=
15602 				    HAT_CTX1_FLAG;
15603 			} else {
15604 				ism_map[i].imap_hatflags &=
15605 				    ~HAT_CTX1_FLAG;
15606 			}
15607 		}
15608 		ism_blkp = ism_blkp->iblk_next;
15609 	}
15610 }
15611 
15612 static int
15613 sfmmu_srd_lock_held(sf_srd_t *srdp)
15614 {
15615 	return (MUTEX_HELD(&srdp->srd_mutex));
15616 }
15617 
15618 /* ARGSUSED */
15619 static int
15620 sfmmu_scdcache_constructor(void *buf, void *cdrarg, int kmflags)
15621 {
15622 	sf_scd_t *scdp = (sf_scd_t *)buf;
15623 
15624 	bzero(buf, sizeof (sf_scd_t));
15625 	mutex_init(&scdp->scd_mutex, NULL, MUTEX_DEFAULT, NULL);
15626 	return (0);
15627 }
15628 
15629 /* ARGSUSED */
15630 static void
15631 sfmmu_scdcache_destructor(void *buf, void *cdrarg)
15632 {
15633 	sf_scd_t *scdp = (sf_scd_t *)buf;
15634 
15635 	mutex_destroy(&scdp->scd_mutex);
15636 }
15637 
15638 /*
15639  * The listp parameter is a pointer to a list of hmeblks which are partially
15640  * freed as result of calling sfmmu_hblk_hash_rm(), the last phase of the
15641  * freeing process is to cross-call all cpus to ensure that there are no
15642  * remaining cached references.
15643  *
15644  * If the local generation number is less than the global then we can free
15645  * hmeblks which are already on the pending queue as another cpu has completed
15646  * the cross-call.
15647  *
15648  * We cross-call to make sure that there are no threads on other cpus accessing
15649  * these hmblks and then complete the process of freeing them under the
15650  * following conditions:
15651  * 	The total number of pending hmeblks is greater than the threshold
15652  *	The reserve list has fewer than HBLK_RESERVE_CNT hmeblks
15653  *	It is at least 1 second since the last time we cross-called
15654  *
15655  * Otherwise, we add the hmeblks to the per-cpu pending queue.
15656  */
15657 static void
15658 sfmmu_hblks_list_purge(struct hme_blk **listp, int dontfree)
15659 {
15660 	struct hme_blk *hblkp, *pr_hblkp = NULL;
15661 	int		count = 0;
15662 	cpuset_t	cpuset = cpu_ready_set;
15663 	cpu_hme_pend_t	*cpuhp;
15664 	timestruc_t	now;
15665 	int		one_second_expired = 0;
15666 
15667 	gethrestime_lasttick(&now);
15668 
15669 	for (hblkp = *listp; hblkp != NULL; hblkp = hblkp->hblk_next) {
15670 		ASSERT(hblkp->hblk_shw_bit == 0);
15671 		ASSERT(hblkp->hblk_shared == 0);
15672 		count++;
15673 		pr_hblkp = hblkp;
15674 	}
15675 
15676 	cpuhp = &cpu_hme_pend[CPU->cpu_seqid];
15677 	mutex_enter(&cpuhp->chp_mutex);
15678 
15679 	if ((cpuhp->chp_count + count) == 0) {
15680 		mutex_exit(&cpuhp->chp_mutex);
15681 		return;
15682 	}
15683 
15684 	if ((now.tv_sec - cpuhp->chp_timestamp) > 1) {
15685 		one_second_expired  = 1;
15686 	}
15687 
15688 	if (!dontfree && (freehblkcnt < HBLK_RESERVE_CNT ||
15689 	    (cpuhp->chp_count + count) > cpu_hme_pend_thresh ||
15690 	    one_second_expired)) {
15691 		/* Append global list to local */
15692 		if (pr_hblkp == NULL) {
15693 			*listp = cpuhp->chp_listp;
15694 		} else {
15695 			pr_hblkp->hblk_next = cpuhp->chp_listp;
15696 		}
15697 		cpuhp->chp_listp = NULL;
15698 		cpuhp->chp_count = 0;
15699 		cpuhp->chp_timestamp = now.tv_sec;
15700 		mutex_exit(&cpuhp->chp_mutex);
15701 
15702 		kpreempt_disable();
15703 		CPUSET_DEL(cpuset, CPU->cpu_id);
15704 		xt_sync(cpuset);
15705 		xt_sync(cpuset);
15706 		kpreempt_enable();
15707 
15708 		/*
15709 		 * At this stage we know that no trap handlers on other
15710 		 * cpus can have references to hmeblks on the list.
15711 		 */
15712 		sfmmu_hblk_free(listp);
15713 	} else if (*listp != NULL) {
15714 		pr_hblkp->hblk_next = cpuhp->chp_listp;
15715 		cpuhp->chp_listp = *listp;
15716 		cpuhp->chp_count += count;
15717 		*listp = NULL;
15718 		mutex_exit(&cpuhp->chp_mutex);
15719 	} else {
15720 		mutex_exit(&cpuhp->chp_mutex);
15721 	}
15722 }
15723 
15724 /*
15725  * Add an hmeblk to the the hash list.
15726  */
15727 void
15728 sfmmu_hblk_hash_add(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
15729 	uint64_t hblkpa)
15730 {
15731 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
15732 #ifdef	DEBUG
15733 	if (hmebp->hmeblkp == NULL) {
15734 		ASSERT(hmebp->hmeh_nextpa == HMEBLK_ENDPA);
15735 	}
15736 #endif /* DEBUG */
15737 
15738 	hmeblkp->hblk_nextpa = hmebp->hmeh_nextpa;
15739 	/*
15740 	 * Since the TSB miss handler now does not lock the hash chain before
15741 	 * walking it, make sure that the hmeblks nextpa is globally visible
15742 	 * before we make the hmeblk globally visible by updating the chain root
15743 	 * pointer in the hash bucket.
15744 	 */
15745 	membar_producer();
15746 	hmebp->hmeh_nextpa = hblkpa;
15747 	hmeblkp->hblk_next = hmebp->hmeblkp;
15748 	hmebp->hmeblkp = hmeblkp;
15749 
15750 }
15751 
15752 /*
15753  * This function is the first part of a 2 part process to remove an hmeblk
15754  * from the hash chain. In this phase we unlink the hmeblk from the hash chain
15755  * but leave the next physical pointer unchanged. The hmeblk is then linked onto
15756  * a per-cpu pending list using the virtual address pointer.
15757  *
15758  * TSB miss trap handlers that start after this phase will no longer see
15759  * this hmeblk. TSB miss handlers that still cache this hmeblk in a register
15760  * can still use it for further chain traversal because we haven't yet modifed
15761  * the next physical pointer or freed it.
15762  *
15763  * In the second phase of hmeblk removal we'll issue a barrier xcall before
15764  * we reuse or free this hmeblk. This will make sure all lingering references to
15765  * the hmeblk after first phase disappear before we finally reclaim it.
15766  * This scheme eliminates the need for TSB miss handlers to lock hmeblk chains
15767  * during their traversal.
15768  *
15769  * The hmehash_mutex must be held when calling this function.
15770  *
15771  * Input:
15772  *	 hmebp - hme hash bucket pointer
15773  *	 hmeblkp - address of hmeblk to be removed
15774  *	 pr_hblk - virtual address of previous hmeblkp
15775  *	 listp - pointer to list of hmeblks linked by virtual address
15776  *	 free_now flag - indicates that a complete removal from the hash chains
15777  *			 is necessary.
15778  *
15779  * It is inefficient to use the free_now flag as a cross-call is required to
15780  * remove a single hmeblk from the hash chain but is necessary when hmeblks are
15781  * in short supply.
15782  */
15783 void
15784 sfmmu_hblk_hash_rm(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
15785     struct hme_blk *pr_hblk, struct hme_blk **listp,
15786     int free_now)
15787 {
15788 	int shw_size, vshift;
15789 	struct hme_blk *shw_hblkp;
15790 	uint_t		shw_mask, newshw_mask;
15791 	caddr_t		vaddr;
15792 	int		size;
15793 	cpuset_t cpuset = cpu_ready_set;
15794 
15795 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
15796 
15797 	if (hmebp->hmeblkp == hmeblkp) {
15798 		hmebp->hmeh_nextpa = hmeblkp->hblk_nextpa;
15799 		hmebp->hmeblkp = hmeblkp->hblk_next;
15800 	} else {
15801 		pr_hblk->hblk_nextpa = hmeblkp->hblk_nextpa;
15802 		pr_hblk->hblk_next = hmeblkp->hblk_next;
15803 	}
15804 
15805 	size = get_hblk_ttesz(hmeblkp);
15806 	shw_hblkp = hmeblkp->hblk_shadow;
15807 	if (shw_hblkp) {
15808 		ASSERT(hblktosfmmu(hmeblkp) != KHATID);
15809 		ASSERT(!hmeblkp->hblk_shared);
15810 #ifdef	DEBUG
15811 		if (mmu_page_sizes == max_mmu_page_sizes) {
15812 			ASSERT(size < TTE256M);
15813 		} else {
15814 			ASSERT(size < TTE4M);
15815 		}
15816 #endif /* DEBUG */
15817 
15818 		shw_size = get_hblk_ttesz(shw_hblkp);
15819 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
15820 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
15821 		ASSERT(vshift < 8);
15822 		/*
15823 		 * Atomically clear shadow mask bit
15824 		 */
15825 		do {
15826 			shw_mask = shw_hblkp->hblk_shw_mask;
15827 			ASSERT(shw_mask & (1 << vshift));
15828 			newshw_mask = shw_mask & ~(1 << vshift);
15829 			newshw_mask = cas32(&shw_hblkp->hblk_shw_mask,
15830 			    shw_mask, newshw_mask);
15831 		} while (newshw_mask != shw_mask);
15832 		hmeblkp->hblk_shadow = NULL;
15833 	}
15834 	hmeblkp->hblk_shw_bit = 0;
15835 
15836 	if (hmeblkp->hblk_shared) {
15837 #ifdef	DEBUG
15838 		sf_srd_t	*srdp;
15839 		sf_region_t	*rgnp;
15840 		uint_t		rid;
15841 
15842 		srdp = hblktosrd(hmeblkp);
15843 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
15844 		rid = hmeblkp->hblk_tag.htag_rid;
15845 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
15846 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
15847 		rgnp = srdp->srd_hmergnp[rid];
15848 		ASSERT(rgnp != NULL);
15849 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
15850 #endif /* DEBUG */
15851 		hmeblkp->hblk_shared = 0;
15852 	}
15853 	if (free_now) {
15854 		kpreempt_disable();
15855 		CPUSET_DEL(cpuset, CPU->cpu_id);
15856 		xt_sync(cpuset);
15857 		xt_sync(cpuset);
15858 		kpreempt_enable();
15859 
15860 		hmeblkp->hblk_nextpa = HMEBLK_ENDPA;
15861 		hmeblkp->hblk_next = NULL;
15862 	} else {
15863 		/* Append hmeblkp to listp for processing later. */
15864 		hmeblkp->hblk_next = *listp;
15865 		*listp = hmeblkp;
15866 	}
15867 }
15868 
15869 /*
15870  * This routine is called when memory is in short supply and returns a free
15871  * hmeblk of the requested size from the cpu pending lists.
15872  */
15873 static struct hme_blk *
15874 sfmmu_check_pending_hblks(int size)
15875 {
15876 	int i;
15877 	struct hme_blk *hmeblkp = NULL, *last_hmeblkp;
15878 	int found_hmeblk;
15879 	cpuset_t cpuset = cpu_ready_set;
15880 	cpu_hme_pend_t *cpuhp;
15881 
15882 	/* Flush cpu hblk pending queues */
15883 	for (i = 0; i < NCPU; i++) {
15884 		cpuhp = &cpu_hme_pend[i];
15885 		if (cpuhp->chp_listp != NULL)  {
15886 			mutex_enter(&cpuhp->chp_mutex);
15887 			if (cpuhp->chp_listp == NULL)  {
15888 				mutex_exit(&cpuhp->chp_mutex);
15889 				continue;
15890 			}
15891 			found_hmeblk = 0;
15892 			last_hmeblkp = NULL;
15893 			for (hmeblkp = cpuhp->chp_listp; hmeblkp != NULL;
15894 			    hmeblkp = hmeblkp->hblk_next) {
15895 				if (get_hblk_ttesz(hmeblkp) == size) {
15896 					if (last_hmeblkp == NULL) {
15897 						cpuhp->chp_listp =
15898 						    hmeblkp->hblk_next;
15899 					} else {
15900 						last_hmeblkp->hblk_next =
15901 						    hmeblkp->hblk_next;
15902 					}
15903 					ASSERT(cpuhp->chp_count > 0);
15904 					cpuhp->chp_count--;
15905 					found_hmeblk = 1;
15906 					break;
15907 				} else {
15908 					last_hmeblkp = hmeblkp;
15909 				}
15910 			}
15911 			mutex_exit(&cpuhp->chp_mutex);
15912 
15913 			if (found_hmeblk) {
15914 				kpreempt_disable();
15915 				CPUSET_DEL(cpuset, CPU->cpu_id);
15916 				xt_sync(cpuset);
15917 				xt_sync(cpuset);
15918 				kpreempt_enable();
15919 				return (hmeblkp);
15920 			}
15921 		}
15922 	}
15923 	return (NULL);
15924 }
15925