xref: /titanic_41/usr/src/uts/sfmmu/vm/hat_sfmmu.c (revision d24234c24aeaca4ca56ee3ac2794507968f274c4)
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 2009 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 *);
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 sun4v, there will be one global context domain, this is to
1116 	 * avoid the ldom cpu substitution problem.
1117 	 *
1118 	 * For all platforms that have CPUs sharing MMUs, this
1119 	 * value must be defined.
1120 	 */
1121 	if (max_mmu_ctxdoms == 0) {
1122 #ifndef sun4v
1123 		max_mmu_ctxdoms = max_ncpus;
1124 #else /* sun4v */
1125 		max_mmu_ctxdoms = 1;
1126 #endif /* sun4v */
1127 	}
1128 
1129 	size = max_mmu_ctxdoms * sizeof (mmu_ctx_t *);
1130 	mmu_ctxs_tbl = kmem_zalloc(size, KM_SLEEP);
1131 
1132 	/* mmu_ctx_t is 64 bytes aligned */
1133 	mmuctxdom_cache = kmem_cache_create("mmuctxdom_cache",
1134 	    sizeof (mmu_ctx_t), 64, NULL, NULL, NULL, NULL, NULL, 0);
1135 	/*
1136 	 * MMU context domain initialization for the Boot CPU.
1137 	 * This needs the context domains array allocated above.
1138 	 */
1139 	mutex_enter(&cpu_lock);
1140 	sfmmu_cpu_init(CPU);
1141 	mutex_exit(&cpu_lock);
1142 
1143 	/*
1144 	 * Intialize ism mapping list lock.
1145 	 */
1146 
1147 	mutex_init(&ism_mlist_lock, NULL, MUTEX_DEFAULT, NULL);
1148 
1149 	/*
1150 	 * Each sfmmu structure carries an array of MMU context info
1151 	 * structures, one per context domain. The size of this array depends
1152 	 * on the maximum number of context domains. So, the size of the
1153 	 * sfmmu structure varies per platform.
1154 	 *
1155 	 * sfmmu is allocated from static arena, because trap
1156 	 * handler at TL > 0 is not allowed to touch kernel relocatable
1157 	 * memory. sfmmu's alignment is changed to 64 bytes from
1158 	 * default 8 bytes, as the lower 6 bits will be used to pass
1159 	 * pgcnt to vtag_flush_pgcnt_tl1.
1160 	 */
1161 	size = sizeof (sfmmu_t) + sizeof (sfmmu_ctx_t) * (max_mmu_ctxdoms - 1);
1162 
1163 	sfmmuid_cache = kmem_cache_create("sfmmuid_cache", size,
1164 	    64, sfmmu_idcache_constructor, sfmmu_idcache_destructor,
1165 	    NULL, NULL, static_arena, 0);
1166 
1167 	sfmmu_tsbinfo_cache = kmem_cache_create("sfmmu_tsbinfo_cache",
1168 	    sizeof (struct tsb_info), 0, NULL, NULL, NULL, NULL, NULL, 0);
1169 
1170 	/*
1171 	 * Since we only use the tsb8k cache to "borrow" pages for TSBs
1172 	 * from the heap when low on memory or when TSB_FORCEALLOC is
1173 	 * specified, don't use magazines to cache them--we want to return
1174 	 * them to the system as quickly as possible.
1175 	 */
1176 	sfmmu_tsb8k_cache = kmem_cache_create("sfmmu_tsb8k_cache",
1177 	    MMU_PAGESIZE, MMU_PAGESIZE, NULL, NULL, NULL, NULL,
1178 	    static_arena, KMC_NOMAGAZINE);
1179 
1180 	/*
1181 	 * Set tsb_alloc_hiwater to 1/tsb_alloc_hiwater_factor of physical
1182 	 * memory, which corresponds to the old static reserve for TSBs.
1183 	 * tsb_alloc_hiwater_factor defaults to 32.  This caps the amount of
1184 	 * memory we'll allocate for TSB slabs; beyond this point TSB
1185 	 * allocations will be taken from the kernel heap (via
1186 	 * sfmmu_tsb8k_cache) and will be throttled as would any other kmem
1187 	 * consumer.
1188 	 */
1189 	if (tsb_alloc_hiwater_factor == 0) {
1190 		tsb_alloc_hiwater_factor = TSB_ALLOC_HIWATER_FACTOR_DEFAULT;
1191 	}
1192 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
1193 
1194 	for (sz = tsb_slab_ttesz; sz > 0; sz--) {
1195 		if (!(disable_large_pages & (1 << sz)))
1196 			break;
1197 	}
1198 
1199 	if (sz < tsb_slab_ttesz) {
1200 		tsb_slab_ttesz = sz;
1201 		tsb_slab_shift = MMU_PAGESHIFT + (sz << 1) + sz;
1202 		tsb_slab_size = 1 << tsb_slab_shift;
1203 		tsb_slab_mask = (1 << (tsb_slab_shift - MMU_PAGESHIFT)) - 1;
1204 		use_bigtsb_arena = 0;
1205 	} else if (use_bigtsb_arena &&
1206 	    (disable_large_pages & (1 << bigtsb_slab_ttesz))) {
1207 		use_bigtsb_arena = 0;
1208 	}
1209 
1210 	if (!use_bigtsb_arena) {
1211 		bigtsb_slab_shift = tsb_slab_shift;
1212 	}
1213 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
1214 
1215 	/*
1216 	 * On smaller memory systems, allocate TSB memory in smaller chunks
1217 	 * than the default 4M slab size. We also honor disable_large_pages
1218 	 * here.
1219 	 *
1220 	 * The trap handlers need to be patched with the final slab shift,
1221 	 * since they need to be able to construct the TSB pointer at runtime.
1222 	 */
1223 	if ((tsb_max_growsize <= TSB_512K_SZCODE) &&
1224 	    !(disable_large_pages & (1 << TTE512K))) {
1225 		tsb_slab_ttesz = TTE512K;
1226 		tsb_slab_shift = MMU_PAGESHIFT512K;
1227 		tsb_slab_size = MMU_PAGESIZE512K;
1228 		tsb_slab_mask = MMU_PAGEOFFSET512K >> MMU_PAGESHIFT;
1229 		use_bigtsb_arena = 0;
1230 	}
1231 
1232 	if (!use_bigtsb_arena) {
1233 		bigtsb_slab_ttesz = tsb_slab_ttesz;
1234 		bigtsb_slab_shift = tsb_slab_shift;
1235 		bigtsb_slab_size = tsb_slab_size;
1236 		bigtsb_slab_mask = tsb_slab_mask;
1237 	}
1238 
1239 
1240 	/*
1241 	 * Set up memory callback to update tsb_alloc_hiwater and
1242 	 * tsb_max_growsize.
1243 	 */
1244 	i = kphysm_setup_func_register(&sfmmu_update_vec, (void *) 0);
1245 	ASSERT(i == 0);
1246 
1247 	/*
1248 	 * kmem_tsb_arena is the source from which large TSB slabs are
1249 	 * drawn.  The quantum of this arena corresponds to the largest
1250 	 * TSB size we can dynamically allocate for user processes.
1251 	 * Currently it must also be a supported page size since we
1252 	 * use exactly one translation entry to map each slab page.
1253 	 *
1254 	 * The per-lgroup kmem_tsb_default_arena arenas are the arenas from
1255 	 * which most TSBs are allocated.  Since most TSB allocations are
1256 	 * typically 8K we have a kmem cache we stack on top of each
1257 	 * kmem_tsb_default_arena to speed up those allocations.
1258 	 *
1259 	 * Note the two-level scheme of arenas is required only
1260 	 * because vmem_create doesn't allow us to specify alignment
1261 	 * requirements.  If this ever changes the code could be
1262 	 * simplified to use only one level of arenas.
1263 	 *
1264 	 * If 256M page support exists on sun4v, 256MB kmem_bigtsb_arena
1265 	 * will be provided in addition to the 4M kmem_tsb_arena.
1266 	 */
1267 	if (use_bigtsb_arena) {
1268 		kmem_bigtsb_arena = vmem_create("kmem_bigtsb", NULL, 0,
1269 		    bigtsb_slab_size, sfmmu_vmem_xalloc_aligned_wrapper,
1270 		    vmem_xfree, heap_arena, 0, VM_SLEEP);
1271 	}
1272 
1273 	kmem_tsb_arena = vmem_create("kmem_tsb", NULL, 0, tsb_slab_size,
1274 	    sfmmu_vmem_xalloc_aligned_wrapper,
1275 	    vmem_xfree, heap_arena, 0, VM_SLEEP);
1276 
1277 	if (tsb_lgrp_affinity) {
1278 		char s[50];
1279 		for (i = 0; i < NLGRPS_MAX; i++) {
1280 			if (use_bigtsb_arena) {
1281 				(void) sprintf(s, "kmem_bigtsb_lgrp%d", i);
1282 				kmem_bigtsb_default_arena[i] = vmem_create(s,
1283 				    NULL, 0, 2 * tsb_slab_size,
1284 				    sfmmu_tsb_segkmem_alloc,
1285 				    sfmmu_tsb_segkmem_free, kmem_bigtsb_arena,
1286 				    0, VM_SLEEP | VM_BESTFIT);
1287 			}
1288 
1289 			(void) sprintf(s, "kmem_tsb_lgrp%d", i);
1290 			kmem_tsb_default_arena[i] = vmem_create(s,
1291 			    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1292 			    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1293 			    VM_SLEEP | VM_BESTFIT);
1294 
1295 			(void) sprintf(s, "sfmmu_tsb_lgrp%d_cache", i);
1296 			sfmmu_tsb_cache[i] = kmem_cache_create(s,
1297 			    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1298 			    kmem_tsb_default_arena[i], 0);
1299 		}
1300 	} else {
1301 		if (use_bigtsb_arena) {
1302 			kmem_bigtsb_default_arena[0] =
1303 			    vmem_create("kmem_bigtsb_default", NULL, 0,
1304 			    2 * tsb_slab_size, sfmmu_tsb_segkmem_alloc,
1305 			    sfmmu_tsb_segkmem_free, kmem_bigtsb_arena, 0,
1306 			    VM_SLEEP | VM_BESTFIT);
1307 		}
1308 
1309 		kmem_tsb_default_arena[0] = vmem_create("kmem_tsb_default",
1310 		    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1311 		    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1312 		    VM_SLEEP | VM_BESTFIT);
1313 		sfmmu_tsb_cache[0] = kmem_cache_create("sfmmu_tsb_cache",
1314 		    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1315 		    kmem_tsb_default_arena[0], 0);
1316 	}
1317 
1318 	sfmmu8_cache = kmem_cache_create("sfmmu8_cache", HME8BLK_SZ,
1319 	    HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1320 	    sfmmu_hblkcache_destructor,
1321 	    sfmmu_hblkcache_reclaim, (void *)HME8BLK_SZ,
1322 	    hat_memload_arena, KMC_NOHASH);
1323 
1324 	hat_memload1_arena = vmem_create("hat_memload1", NULL, 0, PAGESIZE,
1325 	    segkmem_alloc_permanent, segkmem_free, heap_arena, 0,
1326 	    VMC_DUMPSAFE | VM_SLEEP);
1327 
1328 	sfmmu1_cache = kmem_cache_create("sfmmu1_cache", HME1BLK_SZ,
1329 	    HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1330 	    sfmmu_hblkcache_destructor,
1331 	    NULL, (void *)HME1BLK_SZ,
1332 	    hat_memload1_arena, KMC_NOHASH);
1333 
1334 	pa_hment_cache = kmem_cache_create("pa_hment_cache", PAHME_SZ,
1335 	    0, NULL, NULL, NULL, NULL, static_arena, KMC_NOHASH);
1336 
1337 	ism_blk_cache = kmem_cache_create("ism_blk_cache",
1338 	    sizeof (ism_blk_t), ecache_alignsize, NULL, NULL,
1339 	    NULL, NULL, static_arena, KMC_NOHASH);
1340 
1341 	ism_ment_cache = kmem_cache_create("ism_ment_cache",
1342 	    sizeof (ism_ment_t), 0, NULL, NULL,
1343 	    NULL, NULL, NULL, 0);
1344 
1345 	/*
1346 	 * We grab the first hat for the kernel,
1347 	 */
1348 	AS_LOCK_ENTER(&kas, &kas.a_lock, RW_WRITER);
1349 	kas.a_hat = hat_alloc(&kas);
1350 	AS_LOCK_EXIT(&kas, &kas.a_lock);
1351 
1352 	/*
1353 	 * Initialize hblk_reserve.
1354 	 */
1355 	((struct hme_blk *)hblk_reserve)->hblk_nextpa =
1356 	    va_to_pa((caddr_t)hblk_reserve);
1357 
1358 #ifndef UTSB_PHYS
1359 	/*
1360 	 * Reserve some kernel virtual address space for the locked TTEs
1361 	 * that allow us to probe the TSB from TL>0.
1362 	 */
1363 	utsb_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1364 	    0, 0, NULL, NULL, VM_SLEEP);
1365 	utsb4m_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1366 	    0, 0, NULL, NULL, VM_SLEEP);
1367 #endif
1368 
1369 #ifdef VAC
1370 	/*
1371 	 * The big page VAC handling code assumes VAC
1372 	 * will not be bigger than the smallest big
1373 	 * page- which is 64K.
1374 	 */
1375 	if (TTEPAGES(TTE64K) < CACHE_NUM_COLOR) {
1376 		cmn_err(CE_PANIC, "VAC too big!");
1377 	}
1378 #endif
1379 
1380 	(void) xhat_init();
1381 
1382 	uhme_hash_pa = va_to_pa(uhme_hash);
1383 	khme_hash_pa = va_to_pa(khme_hash);
1384 
1385 	/*
1386 	 * Initialize relocation locks. kpr_suspendlock is held
1387 	 * at PIL_MAX to prevent interrupts from pinning the holder
1388 	 * of a suspended TTE which may access it leading to a
1389 	 * deadlock condition.
1390 	 */
1391 	mutex_init(&kpr_mutex, NULL, MUTEX_DEFAULT, NULL);
1392 	mutex_init(&kpr_suspendlock, NULL, MUTEX_SPIN, (void *)PIL_MAX);
1393 
1394 	/*
1395 	 * If Shared context support is disabled via /etc/system
1396 	 * set shctx_on to 0 here if it was set to 1 earlier in boot
1397 	 * sequence by cpu module initialization code.
1398 	 */
1399 	if (shctx_on && disable_shctx) {
1400 		shctx_on = 0;
1401 	}
1402 
1403 	if (shctx_on) {
1404 		srd_buckets = kmem_zalloc(SFMMU_MAX_SRD_BUCKETS *
1405 		    sizeof (srd_buckets[0]), KM_SLEEP);
1406 		for (i = 0; i < SFMMU_MAX_SRD_BUCKETS; i++) {
1407 			mutex_init(&srd_buckets[i].srdb_lock, NULL,
1408 			    MUTEX_DEFAULT, NULL);
1409 		}
1410 
1411 		srd_cache = kmem_cache_create("srd_cache", sizeof (sf_srd_t),
1412 		    0, sfmmu_srdcache_constructor, sfmmu_srdcache_destructor,
1413 		    NULL, NULL, NULL, 0);
1414 		region_cache = kmem_cache_create("region_cache",
1415 		    sizeof (sf_region_t), 0, sfmmu_rgncache_constructor,
1416 		    sfmmu_rgncache_destructor, NULL, NULL, NULL, 0);
1417 		scd_cache = kmem_cache_create("scd_cache", sizeof (sf_scd_t),
1418 		    0, sfmmu_scdcache_constructor,  sfmmu_scdcache_destructor,
1419 		    NULL, NULL, NULL, 0);
1420 	}
1421 
1422 	/*
1423 	 * Pre-allocate hrm_hashtab before enabling the collection of
1424 	 * refmod statistics.  Allocating on the fly would mean us
1425 	 * running the risk of suffering recursive mutex enters or
1426 	 * deadlocks.
1427 	 */
1428 	hrm_hashtab = kmem_zalloc(HRM_HASHSIZE * sizeof (struct hrmstat *),
1429 	    KM_SLEEP);
1430 
1431 	/* Allocate per-cpu pending freelist of hmeblks */
1432 	cpu_hme_pend = kmem_zalloc((NCPU * sizeof (cpu_hme_pend_t)) + 64,
1433 	    KM_SLEEP);
1434 	cpu_hme_pend = (cpu_hme_pend_t *)P2ROUNDUP(
1435 	    (uintptr_t)cpu_hme_pend, 64);
1436 
1437 	for (i = 0; i < NCPU; i++) {
1438 		mutex_init(&cpu_hme_pend[i].chp_mutex, NULL, MUTEX_DEFAULT,
1439 		    NULL);
1440 	}
1441 
1442 	if (cpu_hme_pend_thresh == 0) {
1443 		cpu_hme_pend_thresh = CPU_HME_PEND_THRESH;
1444 	}
1445 }
1446 
1447 /*
1448  * Initialize locking for the hat layer, called early during boot.
1449  */
1450 static void
1451 hat_lock_init()
1452 {
1453 	int i;
1454 
1455 	/*
1456 	 * initialize the array of mutexes protecting a page's mapping
1457 	 * list and p_nrm field.
1458 	 */
1459 	for (i = 0; i < mml_table_sz; i++)
1460 		mutex_init(&mml_table[i], NULL, MUTEX_DEFAULT, NULL);
1461 
1462 	if (kpm_enable) {
1463 		for (i = 0; i < kpmp_table_sz; i++) {
1464 			mutex_init(&kpmp_table[i].khl_mutex, NULL,
1465 			    MUTEX_DEFAULT, NULL);
1466 		}
1467 	}
1468 
1469 	/*
1470 	 * Initialize array of mutex locks that protects sfmmu fields and
1471 	 * TSB lists.
1472 	 */
1473 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
1474 		mutex_init(HATLOCK_MUTEXP(&hat_lock[i]), NULL, MUTEX_DEFAULT,
1475 		    NULL);
1476 }
1477 
1478 #define	SFMMU_KERNEL_MAXVA \
1479 	(kmem64_base ? (uintptr_t)kmem64_end : (SYSLIMIT))
1480 
1481 /*
1482  * Allocate a hat structure.
1483  * Called when an address space first uses a hat.
1484  */
1485 struct hat *
1486 hat_alloc(struct as *as)
1487 {
1488 	sfmmu_t *sfmmup;
1489 	int i;
1490 	uint64_t cnum;
1491 	extern uint_t get_color_start(struct as *);
1492 
1493 	ASSERT(AS_WRITE_HELD(as, &as->a_lock));
1494 	sfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
1495 	sfmmup->sfmmu_as = as;
1496 	sfmmup->sfmmu_flags = 0;
1497 	sfmmup->sfmmu_tteflags = 0;
1498 	sfmmup->sfmmu_rtteflags = 0;
1499 	LOCK_INIT_CLEAR(&sfmmup->sfmmu_ctx_lock);
1500 
1501 	if (as == &kas) {
1502 		ksfmmup = sfmmup;
1503 		sfmmup->sfmmu_cext = 0;
1504 		cnum = KCONTEXT;
1505 
1506 		sfmmup->sfmmu_clrstart = 0;
1507 		sfmmup->sfmmu_tsb = NULL;
1508 		/*
1509 		 * hat_kern_setup() will call sfmmu_init_ktsbinfo()
1510 		 * to setup tsb_info for ksfmmup.
1511 		 */
1512 	} else {
1513 
1514 		/*
1515 		 * Just set to invalid ctx. When it faults, it will
1516 		 * get a valid ctx. This would avoid the situation
1517 		 * where we get a ctx, but it gets stolen and then
1518 		 * we fault when we try to run and so have to get
1519 		 * another ctx.
1520 		 */
1521 		sfmmup->sfmmu_cext = 0;
1522 		cnum = INVALID_CONTEXT;
1523 
1524 		/* initialize original physical page coloring bin */
1525 		sfmmup->sfmmu_clrstart = get_color_start(as);
1526 #ifdef DEBUG
1527 		if (tsb_random_size) {
1528 			uint32_t randval = (uint32_t)gettick() >> 4;
1529 			int size = randval % (tsb_max_growsize + 1);
1530 
1531 			/* chose a random tsb size for stress testing */
1532 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb, size,
1533 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1534 		} else
1535 #endif /* DEBUG */
1536 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb,
1537 			    default_tsb_size,
1538 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1539 		sfmmup->sfmmu_flags = HAT_SWAPPED | HAT_ALLCTX_INVALID;
1540 		ASSERT(sfmmup->sfmmu_tsb != NULL);
1541 	}
1542 
1543 	ASSERT(max_mmu_ctxdoms > 0);
1544 	for (i = 0; i < max_mmu_ctxdoms; i++) {
1545 		sfmmup->sfmmu_ctxs[i].cnum = cnum;
1546 		sfmmup->sfmmu_ctxs[i].gnum = 0;
1547 	}
1548 
1549 	for (i = 0; i < max_mmu_page_sizes; i++) {
1550 		sfmmup->sfmmu_ttecnt[i] = 0;
1551 		sfmmup->sfmmu_scdrttecnt[i] = 0;
1552 		sfmmup->sfmmu_ismttecnt[i] = 0;
1553 		sfmmup->sfmmu_scdismttecnt[i] = 0;
1554 		sfmmup->sfmmu_pgsz[i] = TTE8K;
1555 	}
1556 	sfmmup->sfmmu_tsb0_4minflcnt = 0;
1557 	sfmmup->sfmmu_iblk = NULL;
1558 	sfmmup->sfmmu_ismhat = 0;
1559 	sfmmup->sfmmu_scdhat = 0;
1560 	sfmmup->sfmmu_ismblkpa = (uint64_t)-1;
1561 	if (sfmmup == ksfmmup) {
1562 		CPUSET_ALL(sfmmup->sfmmu_cpusran);
1563 	} else {
1564 		CPUSET_ZERO(sfmmup->sfmmu_cpusran);
1565 	}
1566 	sfmmup->sfmmu_free = 0;
1567 	sfmmup->sfmmu_rmstat = 0;
1568 	sfmmup->sfmmu_clrbin = sfmmup->sfmmu_clrstart;
1569 	sfmmup->sfmmu_xhat_provider = NULL;
1570 	cv_init(&sfmmup->sfmmu_tsb_cv, NULL, CV_DEFAULT, NULL);
1571 	sfmmup->sfmmu_srdp = NULL;
1572 	SF_RGNMAP_ZERO(sfmmup->sfmmu_region_map);
1573 	bzero(sfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
1574 	sfmmup->sfmmu_scdp = NULL;
1575 	sfmmup->sfmmu_scd_link.next = NULL;
1576 	sfmmup->sfmmu_scd_link.prev = NULL;
1577 	return (sfmmup);
1578 }
1579 
1580 /*
1581  * Create per-MMU context domain kstats for a given MMU ctx.
1582  */
1583 static void
1584 sfmmu_mmu_kstat_create(mmu_ctx_t *mmu_ctxp)
1585 {
1586 	mmu_ctx_stat_t	stat;
1587 	kstat_t		*mmu_kstat;
1588 
1589 	ASSERT(MUTEX_HELD(&cpu_lock));
1590 	ASSERT(mmu_ctxp->mmu_kstat == NULL);
1591 
1592 	mmu_kstat = kstat_create("unix", mmu_ctxp->mmu_idx, "mmu_ctx",
1593 	    "hat", KSTAT_TYPE_NAMED, MMU_CTX_NUM_STATS, KSTAT_FLAG_VIRTUAL);
1594 
1595 	if (mmu_kstat == NULL) {
1596 		cmn_err(CE_WARN, "kstat_create for MMU %d failed",
1597 		    mmu_ctxp->mmu_idx);
1598 	} else {
1599 		mmu_kstat->ks_data = mmu_ctxp->mmu_kstat_data;
1600 		for (stat = 0; stat < MMU_CTX_NUM_STATS; stat++)
1601 			kstat_named_init(&mmu_ctxp->mmu_kstat_data[stat],
1602 			    mmu_ctx_kstat_names[stat], KSTAT_DATA_INT64);
1603 		mmu_ctxp->mmu_kstat = mmu_kstat;
1604 		kstat_install(mmu_kstat);
1605 	}
1606 }
1607 
1608 /*
1609  * plat_cpuid_to_mmu_ctx_info() is a platform interface that returns MMU
1610  * context domain information for a given CPU. If a platform does not
1611  * specify that interface, then the function below is used instead to return
1612  * default information. The defaults are as follows:
1613  *
1614  *	- For sun4u systems there's one MMU context domain per CPU.
1615  *	  This default is used by all sun4u systems except OPL. OPL systems
1616  *	  provide platform specific interface to map CPU ids to MMU ids
1617  *	  because on OPL more than 1 CPU shares a single MMU.
1618  *        Note that on sun4v, there is one global context domain for
1619  *	  the entire system. This is to avoid running into potential problem
1620  *	  with ldom physical cpu substitution feature.
1621  *	- The number of MMU context IDs supported on any CPU in the
1622  *	  system is 8K.
1623  */
1624 /*ARGSUSED*/
1625 static void
1626 sfmmu_cpuid_to_mmu_ctx_info(processorid_t cpuid, mmu_ctx_info_t *infop)
1627 {
1628 	infop->mmu_nctxs = nctxs;
1629 #ifndef sun4v
1630 	infop->mmu_idx = cpu[cpuid]->cpu_seqid;
1631 #else /* sun4v */
1632 	infop->mmu_idx = 0;
1633 #endif /* sun4v */
1634 }
1635 
1636 /*
1637  * Called during CPU initialization to set the MMU context-related information
1638  * for a CPU.
1639  *
1640  * cpu_lock serializes accesses to mmu_ctxs and mmu_saved_gnum.
1641  */
1642 void
1643 sfmmu_cpu_init(cpu_t *cp)
1644 {
1645 	mmu_ctx_info_t	info;
1646 	mmu_ctx_t	*mmu_ctxp;
1647 
1648 	ASSERT(MUTEX_HELD(&cpu_lock));
1649 
1650 	if (&plat_cpuid_to_mmu_ctx_info == NULL)
1651 		sfmmu_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1652 	else
1653 		plat_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1654 
1655 	ASSERT(info.mmu_idx < max_mmu_ctxdoms);
1656 
1657 	if ((mmu_ctxp = mmu_ctxs_tbl[info.mmu_idx]) == NULL) {
1658 		/* Each mmu_ctx is cacheline aligned. */
1659 		mmu_ctxp = kmem_cache_alloc(mmuctxdom_cache, KM_SLEEP);
1660 		bzero(mmu_ctxp, sizeof (mmu_ctx_t));
1661 
1662 		mutex_init(&mmu_ctxp->mmu_lock, NULL, MUTEX_SPIN,
1663 		    (void *)ipltospl(DISP_LEVEL));
1664 		mmu_ctxp->mmu_idx = info.mmu_idx;
1665 		mmu_ctxp->mmu_nctxs = info.mmu_nctxs;
1666 		/*
1667 		 * Globally for lifetime of a system,
1668 		 * gnum must always increase.
1669 		 * mmu_saved_gnum is protected by the cpu_lock.
1670 		 */
1671 		mmu_ctxp->mmu_gnum = mmu_saved_gnum + 1;
1672 		mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
1673 
1674 		sfmmu_mmu_kstat_create(mmu_ctxp);
1675 
1676 		mmu_ctxs_tbl[info.mmu_idx] = mmu_ctxp;
1677 	} else {
1678 		ASSERT(mmu_ctxp->mmu_idx == info.mmu_idx);
1679 	}
1680 
1681 	/*
1682 	 * The mmu_lock is acquired here to prevent races with
1683 	 * the wrap-around code.
1684 	 */
1685 	mutex_enter(&mmu_ctxp->mmu_lock);
1686 
1687 
1688 	mmu_ctxp->mmu_ncpus++;
1689 	CPUSET_ADD(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1690 	CPU_MMU_IDX(cp) = info.mmu_idx;
1691 	CPU_MMU_CTXP(cp) = mmu_ctxp;
1692 
1693 	mutex_exit(&mmu_ctxp->mmu_lock);
1694 }
1695 
1696 /*
1697  * Called to perform MMU context-related cleanup for a CPU.
1698  */
1699 void
1700 sfmmu_cpu_cleanup(cpu_t *cp)
1701 {
1702 	mmu_ctx_t	*mmu_ctxp;
1703 
1704 	ASSERT(MUTEX_HELD(&cpu_lock));
1705 
1706 	mmu_ctxp = CPU_MMU_CTXP(cp);
1707 	ASSERT(mmu_ctxp != NULL);
1708 
1709 	/*
1710 	 * The mmu_lock is acquired here to prevent races with
1711 	 * the wrap-around code.
1712 	 */
1713 	mutex_enter(&mmu_ctxp->mmu_lock);
1714 
1715 	CPU_MMU_CTXP(cp) = NULL;
1716 
1717 	CPUSET_DEL(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1718 	if (--mmu_ctxp->mmu_ncpus == 0) {
1719 		mmu_ctxs_tbl[mmu_ctxp->mmu_idx] = NULL;
1720 		mutex_exit(&mmu_ctxp->mmu_lock);
1721 		mutex_destroy(&mmu_ctxp->mmu_lock);
1722 
1723 		if (mmu_ctxp->mmu_kstat)
1724 			kstat_delete(mmu_ctxp->mmu_kstat);
1725 
1726 		/* mmu_saved_gnum is protected by the cpu_lock. */
1727 		if (mmu_saved_gnum < mmu_ctxp->mmu_gnum)
1728 			mmu_saved_gnum = mmu_ctxp->mmu_gnum;
1729 
1730 		kmem_cache_free(mmuctxdom_cache, mmu_ctxp);
1731 
1732 		return;
1733 	}
1734 
1735 	mutex_exit(&mmu_ctxp->mmu_lock);
1736 }
1737 
1738 /*
1739  * Hat_setup, makes an address space context the current active one.
1740  * In sfmmu this translates to setting the secondary context with the
1741  * corresponding context.
1742  */
1743 void
1744 hat_setup(struct hat *sfmmup, int allocflag)
1745 {
1746 	hatlock_t *hatlockp;
1747 
1748 	/* Init needs some special treatment. */
1749 	if (allocflag == HAT_INIT) {
1750 		/*
1751 		 * Make sure that we have
1752 		 * 1. a TSB
1753 		 * 2. a valid ctx that doesn't get stolen after this point.
1754 		 */
1755 		hatlockp = sfmmu_hat_enter(sfmmup);
1756 
1757 		/*
1758 		 * Swap in the TSB.  hat_init() allocates tsbinfos without
1759 		 * TSBs, but we need one for init, since the kernel does some
1760 		 * special things to set up its stack and needs the TSB to
1761 		 * resolve page faults.
1762 		 */
1763 		sfmmu_tsb_swapin(sfmmup, hatlockp);
1764 
1765 		sfmmu_get_ctx(sfmmup);
1766 
1767 		sfmmu_hat_exit(hatlockp);
1768 	} else {
1769 		ASSERT(allocflag == HAT_ALLOC);
1770 
1771 		hatlockp = sfmmu_hat_enter(sfmmup);
1772 		kpreempt_disable();
1773 
1774 		CPUSET_ADD(sfmmup->sfmmu_cpusran, CPU->cpu_id);
1775 		/*
1776 		 * sfmmu_setctx_sec takes <pgsz|cnum> as a parameter,
1777 		 * pagesize bits don't matter in this case since we are passing
1778 		 * INVALID_CONTEXT to it.
1779 		 * Compatibility Note: hw takes care of MMU_SCONTEXT1
1780 		 */
1781 		sfmmu_setctx_sec(INVALID_CONTEXT);
1782 		sfmmu_clear_utsbinfo();
1783 
1784 		kpreempt_enable();
1785 		sfmmu_hat_exit(hatlockp);
1786 	}
1787 }
1788 
1789 /*
1790  * Free all the translation resources for the specified address space.
1791  * Called from as_free when an address space is being destroyed.
1792  */
1793 void
1794 hat_free_start(struct hat *sfmmup)
1795 {
1796 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
1797 	ASSERT(sfmmup != ksfmmup);
1798 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1799 
1800 	sfmmup->sfmmu_free = 1;
1801 	if (sfmmup->sfmmu_scdp != NULL) {
1802 		sfmmu_leave_scd(sfmmup, 0);
1803 	}
1804 
1805 	ASSERT(sfmmup->sfmmu_scdp == NULL);
1806 }
1807 
1808 void
1809 hat_free_end(struct hat *sfmmup)
1810 {
1811 	int i;
1812 
1813 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1814 	ASSERT(sfmmup->sfmmu_free == 1);
1815 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
1816 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
1817 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
1818 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
1819 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
1820 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
1821 
1822 	if (sfmmup->sfmmu_rmstat) {
1823 		hat_freestat(sfmmup->sfmmu_as, NULL);
1824 	}
1825 
1826 	while (sfmmup->sfmmu_tsb != NULL) {
1827 		struct tsb_info *next = sfmmup->sfmmu_tsb->tsb_next;
1828 		sfmmu_tsbinfo_free(sfmmup->sfmmu_tsb);
1829 		sfmmup->sfmmu_tsb = next;
1830 	}
1831 
1832 	if (sfmmup->sfmmu_srdp != NULL) {
1833 		sfmmu_leave_srd(sfmmup);
1834 		ASSERT(sfmmup->sfmmu_srdp == NULL);
1835 		for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1836 			if (sfmmup->sfmmu_hmeregion_links[i] != NULL) {
1837 				kmem_free(sfmmup->sfmmu_hmeregion_links[i],
1838 				    SFMMU_L2_HMERLINKS_SIZE);
1839 				sfmmup->sfmmu_hmeregion_links[i] = NULL;
1840 			}
1841 		}
1842 	}
1843 	sfmmu_free_sfmmu(sfmmup);
1844 
1845 #ifdef DEBUG
1846 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1847 		ASSERT(sfmmup->sfmmu_hmeregion_links[i] == NULL);
1848 	}
1849 #endif
1850 
1851 	kmem_cache_free(sfmmuid_cache, sfmmup);
1852 }
1853 
1854 /*
1855  * Set up any translation structures, for the specified address space,
1856  * that are needed or preferred when the process is being swapped in.
1857  */
1858 /* ARGSUSED */
1859 void
1860 hat_swapin(struct hat *hat)
1861 {
1862 	ASSERT(hat->sfmmu_xhat_provider == NULL);
1863 }
1864 
1865 /*
1866  * Free all of the translation resources, for the specified address space,
1867  * that can be freed while the process is swapped out. Called from as_swapout.
1868  * Also, free up the ctx that this process was using.
1869  */
1870 void
1871 hat_swapout(struct hat *sfmmup)
1872 {
1873 	struct hmehash_bucket *hmebp;
1874 	struct hme_blk *hmeblkp;
1875 	struct hme_blk *pr_hblk = NULL;
1876 	struct hme_blk *nx_hblk;
1877 	int i;
1878 	struct hme_blk *list = NULL;
1879 	hatlock_t *hatlockp;
1880 	struct tsb_info *tsbinfop;
1881 	struct free_tsb {
1882 		struct free_tsb *next;
1883 		struct tsb_info *tsbinfop;
1884 	};			/* free list of TSBs */
1885 	struct free_tsb *freelist, *last, *next;
1886 
1887 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1888 	SFMMU_STAT(sf_swapout);
1889 
1890 	/*
1891 	 * There is no way to go from an as to all its translations in sfmmu.
1892 	 * Here is one of the times when we take the big hit and traverse
1893 	 * the hash looking for hme_blks to free up.  Not only do we free up
1894 	 * this as hme_blks but all those that are free.  We are obviously
1895 	 * swapping because we need memory so let's free up as much
1896 	 * as we can.
1897 	 *
1898 	 * Note that we don't flush TLB/TSB here -- it's not necessary
1899 	 * because:
1900 	 *  1) we free the ctx we're using and throw away the TSB(s);
1901 	 *  2) processes aren't runnable while being swapped out.
1902 	 */
1903 	ASSERT(sfmmup != KHATID);
1904 	for (i = 0; i <= UHMEHASH_SZ; i++) {
1905 		hmebp = &uhme_hash[i];
1906 		SFMMU_HASH_LOCK(hmebp);
1907 		hmeblkp = hmebp->hmeblkp;
1908 		pr_hblk = NULL;
1909 		while (hmeblkp) {
1910 
1911 			ASSERT(!hmeblkp->hblk_xhat_bit);
1912 
1913 			if ((hmeblkp->hblk_tag.htag_id == sfmmup) &&
1914 			    !hmeblkp->hblk_shw_bit && !hmeblkp->hblk_lckcnt) {
1915 				ASSERT(!hmeblkp->hblk_shared);
1916 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
1917 				    (caddr_t)get_hblk_base(hmeblkp),
1918 				    get_hblk_endaddr(hmeblkp),
1919 				    NULL, HAT_UNLOAD);
1920 			}
1921 			nx_hblk = hmeblkp->hblk_next;
1922 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
1923 				ASSERT(!hmeblkp->hblk_lckcnt);
1924 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
1925 				    &list, 0);
1926 			} else {
1927 				pr_hblk = hmeblkp;
1928 			}
1929 			hmeblkp = nx_hblk;
1930 		}
1931 		SFMMU_HASH_UNLOCK(hmebp);
1932 	}
1933 
1934 	sfmmu_hblks_list_purge(&list, 0);
1935 
1936 	/*
1937 	 * Now free up the ctx so that others can reuse it.
1938 	 */
1939 	hatlockp = sfmmu_hat_enter(sfmmup);
1940 
1941 	sfmmu_invalidate_ctx(sfmmup);
1942 
1943 	/*
1944 	 * Free TSBs, but not tsbinfos, and set SWAPPED flag.
1945 	 * If TSBs were never swapped in, just return.
1946 	 * This implies that we don't support partial swapping
1947 	 * of TSBs -- either all are swapped out, or none are.
1948 	 *
1949 	 * We must hold the HAT lock here to prevent racing with another
1950 	 * thread trying to unmap TTEs from the TSB or running the post-
1951 	 * relocator after relocating the TSB's memory.  Unfortunately, we
1952 	 * can't free memory while holding the HAT lock or we could
1953 	 * deadlock, so we build a list of TSBs to be freed after marking
1954 	 * the tsbinfos as swapped out and free them after dropping the
1955 	 * lock.
1956 	 */
1957 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
1958 		sfmmu_hat_exit(hatlockp);
1959 		return;
1960 	}
1961 
1962 	SFMMU_FLAGS_SET(sfmmup, HAT_SWAPPED);
1963 	last = freelist = NULL;
1964 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
1965 	    tsbinfop = tsbinfop->tsb_next) {
1966 		ASSERT((tsbinfop->tsb_flags & TSB_SWAPPED) == 0);
1967 
1968 		/*
1969 		 * Cast the TSB into a struct free_tsb and put it on the free
1970 		 * list.
1971 		 */
1972 		if (freelist == NULL) {
1973 			last = freelist = (struct free_tsb *)tsbinfop->tsb_va;
1974 		} else {
1975 			last->next = (struct free_tsb *)tsbinfop->tsb_va;
1976 			last = last->next;
1977 		}
1978 		last->next = NULL;
1979 		last->tsbinfop = tsbinfop;
1980 		tsbinfop->tsb_flags |= TSB_SWAPPED;
1981 		/*
1982 		 * Zero out the TTE to clear the valid bit.
1983 		 * Note we can't use a value like 0xbad because we want to
1984 		 * ensure diagnostic bits are NEVER set on TTEs that might
1985 		 * be loaded.  The intent is to catch any invalid access
1986 		 * to the swapped TSB, such as a thread running with a valid
1987 		 * context without first calling sfmmu_tsb_swapin() to
1988 		 * allocate TSB memory.
1989 		 */
1990 		tsbinfop->tsb_tte.ll = 0;
1991 	}
1992 
1993 	/* Now we can drop the lock and free the TSB memory. */
1994 	sfmmu_hat_exit(hatlockp);
1995 	for (; freelist != NULL; freelist = next) {
1996 		next = freelist->next;
1997 		sfmmu_tsb_free(freelist->tsbinfop);
1998 	}
1999 }
2000 
2001 /*
2002  * Duplicate the translations of an as into another newas
2003  */
2004 /* ARGSUSED */
2005 int
2006 hat_dup(struct hat *hat, struct hat *newhat, caddr_t addr, size_t len,
2007 	uint_t flag)
2008 {
2009 	sf_srd_t *srdp;
2010 	sf_scd_t *scdp;
2011 	int i;
2012 	extern uint_t get_color_start(struct as *);
2013 
2014 	ASSERT(hat->sfmmu_xhat_provider == NULL);
2015 	ASSERT((flag == 0) || (flag == HAT_DUP_ALL) || (flag == HAT_DUP_COW) ||
2016 	    (flag == HAT_DUP_SRD));
2017 	ASSERT(hat != ksfmmup);
2018 	ASSERT(newhat != ksfmmup);
2019 	ASSERT(flag != HAT_DUP_ALL || hat->sfmmu_srdp == newhat->sfmmu_srdp);
2020 
2021 	if (flag == HAT_DUP_COW) {
2022 		panic("hat_dup: HAT_DUP_COW not supported");
2023 	}
2024 
2025 	if (flag == HAT_DUP_SRD && ((srdp = hat->sfmmu_srdp) != NULL)) {
2026 		ASSERT(srdp->srd_evp != NULL);
2027 		VN_HOLD(srdp->srd_evp);
2028 		ASSERT(srdp->srd_refcnt > 0);
2029 		newhat->sfmmu_srdp = srdp;
2030 		atomic_add_32((volatile uint_t *)&srdp->srd_refcnt, 1);
2031 	}
2032 
2033 	/*
2034 	 * HAT_DUP_ALL flag is used after as duplication is done.
2035 	 */
2036 	if (flag == HAT_DUP_ALL && ((srdp = newhat->sfmmu_srdp) != NULL)) {
2037 		ASSERT(newhat->sfmmu_srdp->srd_refcnt >= 2);
2038 		newhat->sfmmu_rtteflags = hat->sfmmu_rtteflags;
2039 		if (hat->sfmmu_flags & HAT_4MTEXT_FLAG) {
2040 			newhat->sfmmu_flags |= HAT_4MTEXT_FLAG;
2041 		}
2042 
2043 		/* check if need to join scd */
2044 		if ((scdp = hat->sfmmu_scdp) != NULL &&
2045 		    newhat->sfmmu_scdp != scdp) {
2046 			int ret;
2047 			SF_RGNMAP_IS_SUBSET(&newhat->sfmmu_region_map,
2048 			    &scdp->scd_region_map, ret);
2049 			ASSERT(ret);
2050 			sfmmu_join_scd(scdp, newhat);
2051 			ASSERT(newhat->sfmmu_scdp == scdp &&
2052 			    scdp->scd_refcnt >= 2);
2053 			for (i = 0; i < max_mmu_page_sizes; i++) {
2054 				newhat->sfmmu_ismttecnt[i] =
2055 				    hat->sfmmu_ismttecnt[i];
2056 				newhat->sfmmu_scdismttecnt[i] =
2057 				    hat->sfmmu_scdismttecnt[i];
2058 			}
2059 		}
2060 
2061 		sfmmu_check_page_sizes(newhat, 1);
2062 	}
2063 
2064 	if (flag == HAT_DUP_ALL && consistent_coloring == 0 &&
2065 	    update_proc_pgcolorbase_after_fork != 0) {
2066 		hat->sfmmu_clrbin = get_color_start(hat->sfmmu_as);
2067 	}
2068 	return (0);
2069 }
2070 
2071 void
2072 hat_memload(struct hat *hat, caddr_t addr, struct page *pp,
2073 	uint_t attr, uint_t flags)
2074 {
2075 	hat_do_memload(hat, addr, pp, attr, flags,
2076 	    SFMMU_INVALID_SHMERID);
2077 }
2078 
2079 void
2080 hat_memload_region(struct hat *hat, caddr_t addr, struct page *pp,
2081 	uint_t attr, uint_t flags, hat_region_cookie_t rcookie)
2082 {
2083 	uint_t rid;
2084 	if (rcookie == HAT_INVALID_REGION_COOKIE ||
2085 	    hat->sfmmu_xhat_provider != NULL) {
2086 		hat_do_memload(hat, addr, pp, attr, flags,
2087 		    SFMMU_INVALID_SHMERID);
2088 		return;
2089 	}
2090 	rid = (uint_t)((uint64_t)rcookie);
2091 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2092 	hat_do_memload(hat, addr, pp, attr, flags, rid);
2093 }
2094 
2095 /*
2096  * Set up addr to map to page pp with protection prot.
2097  * As an optimization we also load the TSB with the
2098  * corresponding tte but it is no big deal if  the tte gets kicked out.
2099  */
2100 static void
2101 hat_do_memload(struct hat *hat, caddr_t addr, struct page *pp,
2102 	uint_t attr, uint_t flags, uint_t rid)
2103 {
2104 	tte_t tte;
2105 
2106 
2107 	ASSERT(hat != NULL);
2108 	ASSERT(PAGE_LOCKED(pp));
2109 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2110 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2111 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2112 	SFMMU_VALIDATE_HMERID(hat, rid, addr, MMU_PAGESIZE);
2113 
2114 	if (PP_ISFREE(pp)) {
2115 		panic("hat_memload: loading a mapping to free page %p",
2116 		    (void *)pp);
2117 	}
2118 
2119 	if (hat->sfmmu_xhat_provider) {
2120 		/* no regions for xhats */
2121 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
2122 		XHAT_MEMLOAD(hat, addr, pp, attr, flags);
2123 		return;
2124 	}
2125 
2126 	ASSERT((hat == ksfmmup) ||
2127 	    AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock));
2128 
2129 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2130 		cmn_err(CE_NOTE, "hat_memload: unsupported flags %d",
2131 		    flags & ~SFMMU_LOAD_ALLFLAG);
2132 
2133 	if (hat->sfmmu_rmstat)
2134 		hat_resvstat(MMU_PAGESIZE, hat->sfmmu_as, addr);
2135 
2136 #if defined(SF_ERRATA_57)
2137 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2138 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2139 	    !(flags & HAT_LOAD_SHARE)) {
2140 		cmn_err(CE_WARN, "hat_memload: illegal attempt to make user "
2141 		    " page executable");
2142 		attr &= ~PROT_EXEC;
2143 	}
2144 #endif
2145 
2146 	sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2147 	(void) sfmmu_tteload_array(hat, &tte, addr, &pp, flags, rid);
2148 
2149 	/*
2150 	 * Check TSB and TLB page sizes.
2151 	 */
2152 	if ((flags & HAT_LOAD_SHARE) == 0) {
2153 		sfmmu_check_page_sizes(hat, 1);
2154 	}
2155 }
2156 
2157 /*
2158  * hat_devload can be called to map real memory (e.g.
2159  * /dev/kmem) and even though hat_devload will determine pf is
2160  * for memory, it will be unable to get a shared lock on the
2161  * page (because someone else has it exclusively) and will
2162  * pass dp = NULL.  If tteload doesn't get a non-NULL
2163  * page pointer it can't cache memory.
2164  */
2165 void
2166 hat_devload(struct hat *hat, caddr_t addr, size_t len, pfn_t pfn,
2167 	uint_t attr, int flags)
2168 {
2169 	tte_t tte;
2170 	struct page *pp = NULL;
2171 	int use_lgpg = 0;
2172 
2173 	ASSERT(hat != NULL);
2174 
2175 	if (hat->sfmmu_xhat_provider) {
2176 		XHAT_DEVLOAD(hat, addr, len, pfn, attr, flags);
2177 		return;
2178 	}
2179 
2180 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2181 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2182 	ASSERT((hat == ksfmmup) ||
2183 	    AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock));
2184 	if (len == 0)
2185 		panic("hat_devload: zero len");
2186 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2187 		cmn_err(CE_NOTE, "hat_devload: unsupported flags %d",
2188 		    flags & ~SFMMU_LOAD_ALLFLAG);
2189 
2190 #if defined(SF_ERRATA_57)
2191 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2192 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2193 	    !(flags & HAT_LOAD_SHARE)) {
2194 		cmn_err(CE_WARN, "hat_devload: illegal attempt to make user "
2195 		    " page executable");
2196 		attr &= ~PROT_EXEC;
2197 	}
2198 #endif
2199 
2200 	/*
2201 	 * If it's a memory page find its pp
2202 	 */
2203 	if (!(flags & HAT_LOAD_NOCONSIST) && pf_is_memory(pfn)) {
2204 		pp = page_numtopp_nolock(pfn);
2205 		if (pp == NULL) {
2206 			flags |= HAT_LOAD_NOCONSIST;
2207 		} else {
2208 			if (PP_ISFREE(pp)) {
2209 				panic("hat_memload: loading "
2210 				    "a mapping to free page %p",
2211 				    (void *)pp);
2212 			}
2213 			if (!PAGE_LOCKED(pp) && !PP_ISNORELOC(pp)) {
2214 				panic("hat_memload: loading a mapping "
2215 				    "to unlocked relocatable page %p",
2216 				    (void *)pp);
2217 			}
2218 			ASSERT(len == MMU_PAGESIZE);
2219 		}
2220 	}
2221 
2222 	if (hat->sfmmu_rmstat)
2223 		hat_resvstat(len, hat->sfmmu_as, addr);
2224 
2225 	if (flags & HAT_LOAD_NOCONSIST) {
2226 		attr |= SFMMU_UNCACHEVTTE;
2227 		use_lgpg = 1;
2228 	}
2229 	if (!pf_is_memory(pfn)) {
2230 		attr |= SFMMU_UNCACHEPTTE | HAT_NOSYNC;
2231 		use_lgpg = 1;
2232 		switch (attr & HAT_ORDER_MASK) {
2233 			case HAT_STRICTORDER:
2234 			case HAT_UNORDERED_OK:
2235 				/*
2236 				 * we set the side effect bit for all non
2237 				 * memory mappings unless merging is ok
2238 				 */
2239 				attr |= SFMMU_SIDEFFECT;
2240 				break;
2241 			case HAT_MERGING_OK:
2242 			case HAT_LOADCACHING_OK:
2243 			case HAT_STORECACHING_OK:
2244 				break;
2245 			default:
2246 				panic("hat_devload: bad attr");
2247 				break;
2248 		}
2249 	}
2250 	while (len) {
2251 		if (!use_lgpg) {
2252 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2253 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2254 			    flags, SFMMU_INVALID_SHMERID);
2255 			len -= MMU_PAGESIZE;
2256 			addr += MMU_PAGESIZE;
2257 			pfn++;
2258 			continue;
2259 		}
2260 		/*
2261 		 *  try to use large pages, check va/pa alignments
2262 		 *  Note that 32M/256M page sizes are not (yet) supported.
2263 		 */
2264 		if ((len >= MMU_PAGESIZE4M) &&
2265 		    !((uintptr_t)addr & MMU_PAGEOFFSET4M) &&
2266 		    !(disable_large_pages & (1 << TTE4M)) &&
2267 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET4M)) {
2268 			sfmmu_memtte(&tte, pfn, attr, TTE4M);
2269 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2270 			    flags, SFMMU_INVALID_SHMERID);
2271 			len -= MMU_PAGESIZE4M;
2272 			addr += MMU_PAGESIZE4M;
2273 			pfn += MMU_PAGESIZE4M / MMU_PAGESIZE;
2274 		} else if ((len >= MMU_PAGESIZE512K) &&
2275 		    !((uintptr_t)addr & MMU_PAGEOFFSET512K) &&
2276 		    !(disable_large_pages & (1 << TTE512K)) &&
2277 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET512K)) {
2278 			sfmmu_memtte(&tte, pfn, attr, TTE512K);
2279 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2280 			    flags, SFMMU_INVALID_SHMERID);
2281 			len -= MMU_PAGESIZE512K;
2282 			addr += MMU_PAGESIZE512K;
2283 			pfn += MMU_PAGESIZE512K / MMU_PAGESIZE;
2284 		} else if ((len >= MMU_PAGESIZE64K) &&
2285 		    !((uintptr_t)addr & MMU_PAGEOFFSET64K) &&
2286 		    !(disable_large_pages & (1 << TTE64K)) &&
2287 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET64K)) {
2288 			sfmmu_memtte(&tte, pfn, attr, TTE64K);
2289 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2290 			    flags, SFMMU_INVALID_SHMERID);
2291 			len -= MMU_PAGESIZE64K;
2292 			addr += MMU_PAGESIZE64K;
2293 			pfn += MMU_PAGESIZE64K / MMU_PAGESIZE;
2294 		} else {
2295 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2296 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2297 			    flags, SFMMU_INVALID_SHMERID);
2298 			len -= MMU_PAGESIZE;
2299 			addr += MMU_PAGESIZE;
2300 			pfn++;
2301 		}
2302 	}
2303 
2304 	/*
2305 	 * Check TSB and TLB page sizes.
2306 	 */
2307 	if ((flags & HAT_LOAD_SHARE) == 0) {
2308 		sfmmu_check_page_sizes(hat, 1);
2309 	}
2310 }
2311 
2312 void
2313 hat_memload_array(struct hat *hat, caddr_t addr, size_t len,
2314 	struct page **pps, uint_t attr, uint_t flags)
2315 {
2316 	hat_do_memload_array(hat, addr, len, pps, attr, flags,
2317 	    SFMMU_INVALID_SHMERID);
2318 }
2319 
2320 void
2321 hat_memload_array_region(struct hat *hat, caddr_t addr, size_t len,
2322 	struct page **pps, uint_t attr, uint_t flags,
2323 	hat_region_cookie_t rcookie)
2324 {
2325 	uint_t rid;
2326 	if (rcookie == HAT_INVALID_REGION_COOKIE ||
2327 	    hat->sfmmu_xhat_provider != NULL) {
2328 		hat_do_memload_array(hat, addr, len, pps, attr, flags,
2329 		    SFMMU_INVALID_SHMERID);
2330 		return;
2331 	}
2332 	rid = (uint_t)((uint64_t)rcookie);
2333 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2334 	hat_do_memload_array(hat, addr, len, pps, attr, flags, rid);
2335 }
2336 
2337 /*
2338  * Map the largest extend possible out of the page array. The array may NOT
2339  * be in order.  The largest possible mapping a page can have
2340  * is specified in the p_szc field.  The p_szc field
2341  * cannot change as long as there any mappings (large or small)
2342  * to any of the pages that make up the large page. (ie. any
2343  * promotion/demotion of page size is not up to the hat but up to
2344  * the page free list manager).  The array
2345  * should consist of properly aligned contigous pages that are
2346  * part of a big page for a large mapping to be created.
2347  */
2348 static void
2349 hat_do_memload_array(struct hat *hat, caddr_t addr, size_t len,
2350 	struct page **pps, uint_t attr, uint_t flags, uint_t rid)
2351 {
2352 	int  ttesz;
2353 	size_t mapsz;
2354 	pgcnt_t	numpg, npgs;
2355 	tte_t tte;
2356 	page_t *pp;
2357 	uint_t large_pages_disable;
2358 
2359 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2360 	SFMMU_VALIDATE_HMERID(hat, rid, addr, len);
2361 
2362 	if (hat->sfmmu_xhat_provider) {
2363 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
2364 		XHAT_MEMLOAD_ARRAY(hat, addr, len, pps, attr, flags);
2365 		return;
2366 	}
2367 
2368 	if (hat->sfmmu_rmstat)
2369 		hat_resvstat(len, hat->sfmmu_as, addr);
2370 
2371 #if defined(SF_ERRATA_57)
2372 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2373 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2374 	    !(flags & HAT_LOAD_SHARE)) {
2375 		cmn_err(CE_WARN, "hat_memload_array: illegal attempt to make "
2376 		    "user page executable");
2377 		attr &= ~PROT_EXEC;
2378 	}
2379 #endif
2380 
2381 	/* Get number of pages */
2382 	npgs = len >> MMU_PAGESHIFT;
2383 
2384 	if (flags & HAT_LOAD_SHARE) {
2385 		large_pages_disable = disable_ism_large_pages;
2386 	} else {
2387 		large_pages_disable = disable_large_pages;
2388 	}
2389 
2390 	if (npgs < NHMENTS || large_pages_disable == LARGE_PAGES_OFF) {
2391 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2392 		    rid);
2393 		return;
2394 	}
2395 
2396 	while (npgs >= NHMENTS) {
2397 		pp = *pps;
2398 		for (ttesz = pp->p_szc; ttesz != TTE8K; ttesz--) {
2399 			/*
2400 			 * Check if this page size is disabled.
2401 			 */
2402 			if (large_pages_disable & (1 << ttesz))
2403 				continue;
2404 
2405 			numpg = TTEPAGES(ttesz);
2406 			mapsz = numpg << MMU_PAGESHIFT;
2407 			if ((npgs >= numpg) &&
2408 			    IS_P2ALIGNED(addr, mapsz) &&
2409 			    IS_P2ALIGNED(pp->p_pagenum, numpg)) {
2410 				/*
2411 				 * At this point we have enough pages and
2412 				 * we know the virtual address and the pfn
2413 				 * are properly aligned.  We still need
2414 				 * to check for physical contiguity but since
2415 				 * it is very likely that this is the case
2416 				 * we will assume they are so and undo
2417 				 * the request if necessary.  It would
2418 				 * be great if we could get a hint flag
2419 				 * like HAT_CONTIG which would tell us
2420 				 * the pages are contigous for sure.
2421 				 */
2422 				sfmmu_memtte(&tte, (*pps)->p_pagenum,
2423 				    attr, ttesz);
2424 				if (!sfmmu_tteload_array(hat, &tte, addr,
2425 				    pps, flags, rid)) {
2426 					break;
2427 				}
2428 			}
2429 		}
2430 		if (ttesz == TTE8K) {
2431 			/*
2432 			 * We were not able to map array using a large page
2433 			 * batch a hmeblk or fraction at a time.
2434 			 */
2435 			numpg = ((uintptr_t)addr >> MMU_PAGESHIFT)
2436 			    & (NHMENTS-1);
2437 			numpg = NHMENTS - numpg;
2438 			ASSERT(numpg <= npgs);
2439 			mapsz = numpg * MMU_PAGESIZE;
2440 			sfmmu_memload_batchsmall(hat, addr, pps, attr, flags,
2441 			    numpg, rid);
2442 		}
2443 		addr += mapsz;
2444 		npgs -= numpg;
2445 		pps += numpg;
2446 	}
2447 
2448 	if (npgs) {
2449 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2450 		    rid);
2451 	}
2452 
2453 	/*
2454 	 * Check TSB and TLB page sizes.
2455 	 */
2456 	if ((flags & HAT_LOAD_SHARE) == 0) {
2457 		sfmmu_check_page_sizes(hat, 1);
2458 	}
2459 }
2460 
2461 /*
2462  * Function tries to batch 8K pages into the same hme blk.
2463  */
2464 static void
2465 sfmmu_memload_batchsmall(struct hat *hat, caddr_t vaddr, page_t **pps,
2466 		    uint_t attr, uint_t flags, pgcnt_t npgs, uint_t rid)
2467 {
2468 	tte_t	tte;
2469 	page_t *pp;
2470 	struct hmehash_bucket *hmebp;
2471 	struct hme_blk *hmeblkp;
2472 	int	index;
2473 
2474 	while (npgs) {
2475 		/*
2476 		 * Acquire the hash bucket.
2477 		 */
2478 		hmebp = sfmmu_tteload_acquire_hashbucket(hat, vaddr, TTE8K,
2479 		    rid);
2480 		ASSERT(hmebp);
2481 
2482 		/*
2483 		 * Find the hment block.
2484 		 */
2485 		hmeblkp = sfmmu_tteload_find_hmeblk(hat, hmebp, vaddr,
2486 		    TTE8K, flags, rid);
2487 		ASSERT(hmeblkp);
2488 
2489 		do {
2490 			/*
2491 			 * Make the tte.
2492 			 */
2493 			pp = *pps;
2494 			sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2495 
2496 			/*
2497 			 * Add the translation.
2498 			 */
2499 			(void) sfmmu_tteload_addentry(hat, hmeblkp, &tte,
2500 			    vaddr, pps, flags, rid);
2501 
2502 			/*
2503 			 * Goto next page.
2504 			 */
2505 			pps++;
2506 			npgs--;
2507 
2508 			/*
2509 			 * Goto next address.
2510 			 */
2511 			vaddr += MMU_PAGESIZE;
2512 
2513 			/*
2514 			 * Don't crossover into a different hmentblk.
2515 			 */
2516 			index = (int)(((uintptr_t)vaddr >> MMU_PAGESHIFT) &
2517 			    (NHMENTS-1));
2518 
2519 		} while (index != 0 && npgs != 0);
2520 
2521 		/*
2522 		 * Release the hash bucket.
2523 		 */
2524 
2525 		sfmmu_tteload_release_hashbucket(hmebp);
2526 	}
2527 }
2528 
2529 /*
2530  * Construct a tte for a page:
2531  *
2532  * tte_valid = 1
2533  * tte_size2 = size & TTE_SZ2_BITS (Panther and Olympus-C only)
2534  * tte_size = size
2535  * tte_nfo = attr & HAT_NOFAULT
2536  * tte_ie = attr & HAT_STRUCTURE_LE
2537  * tte_hmenum = hmenum
2538  * tte_pahi = pp->p_pagenum >> TTE_PASHIFT;
2539  * tte_palo = pp->p_pagenum & TTE_PALOMASK;
2540  * tte_ref = 1 (optimization)
2541  * tte_wr_perm = attr & PROT_WRITE;
2542  * tte_no_sync = attr & HAT_NOSYNC
2543  * tte_lock = attr & SFMMU_LOCKTTE
2544  * tte_cp = !(attr & SFMMU_UNCACHEPTTE)
2545  * tte_cv = !(attr & SFMMU_UNCACHEVTTE)
2546  * tte_e = attr & SFMMU_SIDEFFECT
2547  * tte_priv = !(attr & PROT_USER)
2548  * tte_hwwr = if nosync is set and it is writable we set the mod bit (opt)
2549  * tte_glb = 0
2550  */
2551 void
2552 sfmmu_memtte(tte_t *ttep, pfn_t pfn, uint_t attr, int tte_sz)
2553 {
2554 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2555 
2556 	ttep->tte_inthi = MAKE_TTE_INTHI(pfn, attr, tte_sz, 0 /* hmenum */);
2557 	ttep->tte_intlo = MAKE_TTE_INTLO(pfn, attr, tte_sz, 0 /* hmenum */);
2558 
2559 	if (TTE_IS_NOSYNC(ttep)) {
2560 		TTE_SET_REF(ttep);
2561 		if (TTE_IS_WRITABLE(ttep)) {
2562 			TTE_SET_MOD(ttep);
2563 		}
2564 	}
2565 	if (TTE_IS_NFO(ttep) && TTE_IS_EXECUTABLE(ttep)) {
2566 		panic("sfmmu_memtte: can't set both NFO and EXEC bits");
2567 	}
2568 }
2569 
2570 /*
2571  * This function will add a translation to the hme_blk and allocate the
2572  * hme_blk if one does not exist.
2573  * If a page structure is specified then it will add the
2574  * corresponding hment to the mapping list.
2575  * It will also update the hmenum field for the tte.
2576  *
2577  * Currently this function is only used for kernel mappings.
2578  * So pass invalid region to sfmmu_tteload_array().
2579  */
2580 void
2581 sfmmu_tteload(struct hat *sfmmup, tte_t *ttep, caddr_t vaddr, page_t *pp,
2582 	uint_t flags)
2583 {
2584 	ASSERT(sfmmup == ksfmmup);
2585 	(void) sfmmu_tteload_array(sfmmup, ttep, vaddr, &pp, flags,
2586 	    SFMMU_INVALID_SHMERID);
2587 }
2588 
2589 /*
2590  * Load (ttep != NULL) or unload (ttep == NULL) one entry in the TSB.
2591  * Assumes that a particular page size may only be resident in one TSB.
2592  */
2593 static void
2594 sfmmu_mod_tsb(sfmmu_t *sfmmup, caddr_t vaddr, tte_t *ttep, int ttesz)
2595 {
2596 	struct tsb_info *tsbinfop = NULL;
2597 	uint64_t tag;
2598 	struct tsbe *tsbe_addr;
2599 	uint64_t tsb_base;
2600 	uint_t tsb_size;
2601 	int vpshift = MMU_PAGESHIFT;
2602 	int phys = 0;
2603 
2604 	if (sfmmup == ksfmmup) { /* No support for 32/256M ksfmmu pages */
2605 		phys = ktsb_phys;
2606 		if (ttesz >= TTE4M) {
2607 #ifndef sun4v
2608 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2609 #endif
2610 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2611 			tsb_size = ktsb4m_szcode;
2612 		} else {
2613 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2614 			tsb_size = ktsb_szcode;
2615 		}
2616 	} else {
2617 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2618 
2619 		/*
2620 		 * If there isn't a TSB for this page size, or the TSB is
2621 		 * swapped out, there is nothing to do.  Note that the latter
2622 		 * case seems impossible but can occur if hat_pageunload()
2623 		 * is called on an ISM mapping while the process is swapped
2624 		 * out.
2625 		 */
2626 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2627 			return;
2628 
2629 		/*
2630 		 * If another thread is in the middle of relocating a TSB
2631 		 * we can't unload the entry so set a flag so that the
2632 		 * TSB will be flushed before it can be accessed by the
2633 		 * process.
2634 		 */
2635 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2636 			if (ttep == NULL)
2637 				tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2638 			return;
2639 		}
2640 #if defined(UTSB_PHYS)
2641 		phys = 1;
2642 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2643 #else
2644 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2645 #endif
2646 		tsb_size = tsbinfop->tsb_szc;
2647 	}
2648 	if (ttesz >= TTE4M)
2649 		vpshift = MMU_PAGESHIFT4M;
2650 
2651 	tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2652 	tag = sfmmu_make_tsbtag(vaddr);
2653 
2654 	if (ttep == NULL) {
2655 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2656 	} else {
2657 		if (ttesz >= TTE4M) {
2658 			SFMMU_STAT(sf_tsb_load4m);
2659 		} else {
2660 			SFMMU_STAT(sf_tsb_load8k);
2661 		}
2662 
2663 		sfmmu_load_tsbe(tsbe_addr, tag, ttep, phys);
2664 	}
2665 }
2666 
2667 /*
2668  * Unmap all entries from [start, end) matching the given page size.
2669  *
2670  * This function is used primarily to unmap replicated 64K or 512K entries
2671  * from the TSB that are inserted using the base page size TSB pointer, but
2672  * it may also be called to unmap a range of addresses from the TSB.
2673  */
2674 void
2675 sfmmu_unload_tsb_range(sfmmu_t *sfmmup, caddr_t start, caddr_t end, int ttesz)
2676 {
2677 	struct tsb_info *tsbinfop;
2678 	uint64_t tag;
2679 	struct tsbe *tsbe_addr;
2680 	caddr_t vaddr;
2681 	uint64_t tsb_base;
2682 	int vpshift, vpgsz;
2683 	uint_t tsb_size;
2684 	int phys = 0;
2685 
2686 	/*
2687 	 * Assumptions:
2688 	 *  If ttesz == 8K, 64K or 512K, we walk through the range 8K
2689 	 *  at a time shooting down any valid entries we encounter.
2690 	 *
2691 	 *  If ttesz >= 4M we walk the range 4M at a time shooting
2692 	 *  down any valid mappings we find.
2693 	 */
2694 	if (sfmmup == ksfmmup) {
2695 		phys = ktsb_phys;
2696 		if (ttesz >= TTE4M) {
2697 #ifndef sun4v
2698 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2699 #endif
2700 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2701 			tsb_size = ktsb4m_szcode;
2702 		} else {
2703 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2704 			tsb_size = ktsb_szcode;
2705 		}
2706 	} else {
2707 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2708 
2709 		/*
2710 		 * If there isn't a TSB for this page size, or the TSB is
2711 		 * swapped out, there is nothing to do.  Note that the latter
2712 		 * case seems impossible but can occur if hat_pageunload()
2713 		 * is called on an ISM mapping while the process is swapped
2714 		 * out.
2715 		 */
2716 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2717 			return;
2718 
2719 		/*
2720 		 * If another thread is in the middle of relocating a TSB
2721 		 * we can't unload the entry so set a flag so that the
2722 		 * TSB will be flushed before it can be accessed by the
2723 		 * process.
2724 		 */
2725 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2726 			tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2727 			return;
2728 		}
2729 #if defined(UTSB_PHYS)
2730 		phys = 1;
2731 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2732 #else
2733 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2734 #endif
2735 		tsb_size = tsbinfop->tsb_szc;
2736 	}
2737 	if (ttesz >= TTE4M) {
2738 		vpshift = MMU_PAGESHIFT4M;
2739 		vpgsz = MMU_PAGESIZE4M;
2740 	} else {
2741 		vpshift = MMU_PAGESHIFT;
2742 		vpgsz = MMU_PAGESIZE;
2743 	}
2744 
2745 	for (vaddr = start; vaddr < end; vaddr += vpgsz) {
2746 		tag = sfmmu_make_tsbtag(vaddr);
2747 		tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2748 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2749 	}
2750 }
2751 
2752 /*
2753  * Select the optimum TSB size given the number of mappings
2754  * that need to be cached.
2755  */
2756 static int
2757 sfmmu_select_tsb_szc(pgcnt_t pgcnt)
2758 {
2759 	int szc = 0;
2760 
2761 #ifdef DEBUG
2762 	if (tsb_grow_stress) {
2763 		uint32_t randval = (uint32_t)gettick() >> 4;
2764 		return (randval % (tsb_max_growsize + 1));
2765 	}
2766 #endif	/* DEBUG */
2767 
2768 	while ((szc < tsb_max_growsize) && (pgcnt > SFMMU_RSS_TSBSIZE(szc)))
2769 		szc++;
2770 	return (szc);
2771 }
2772 
2773 /*
2774  * This function will add a translation to the hme_blk and allocate the
2775  * hme_blk if one does not exist.
2776  * If a page structure is specified then it will add the
2777  * corresponding hment to the mapping list.
2778  * It will also update the hmenum field for the tte.
2779  * Furthermore, it attempts to create a large page translation
2780  * for <addr,hat> at page array pps.  It assumes addr and first
2781  * pp is correctly aligned.  It returns 0 if successful and 1 otherwise.
2782  */
2783 static int
2784 sfmmu_tteload_array(sfmmu_t *sfmmup, tte_t *ttep, caddr_t vaddr,
2785 	page_t **pps, uint_t flags, uint_t rid)
2786 {
2787 	struct hmehash_bucket *hmebp;
2788 	struct hme_blk *hmeblkp;
2789 	int 	ret;
2790 	uint_t	size;
2791 
2792 	/*
2793 	 * Get mapping size.
2794 	 */
2795 	size = TTE_CSZ(ttep);
2796 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
2797 
2798 	/*
2799 	 * Acquire the hash bucket.
2800 	 */
2801 	hmebp = sfmmu_tteload_acquire_hashbucket(sfmmup, vaddr, size, rid);
2802 	ASSERT(hmebp);
2803 
2804 	/*
2805 	 * Find the hment block.
2806 	 */
2807 	hmeblkp = sfmmu_tteload_find_hmeblk(sfmmup, hmebp, vaddr, size, flags,
2808 	    rid);
2809 	ASSERT(hmeblkp);
2810 
2811 	/*
2812 	 * Add the translation.
2813 	 */
2814 	ret = sfmmu_tteload_addentry(sfmmup, hmeblkp, ttep, vaddr, pps, flags,
2815 	    rid);
2816 
2817 	/*
2818 	 * Release the hash bucket.
2819 	 */
2820 	sfmmu_tteload_release_hashbucket(hmebp);
2821 
2822 	return (ret);
2823 }
2824 
2825 /*
2826  * Function locks and returns a pointer to the hash bucket for vaddr and size.
2827  */
2828 static struct hmehash_bucket *
2829 sfmmu_tteload_acquire_hashbucket(sfmmu_t *sfmmup, caddr_t vaddr, int size,
2830     uint_t rid)
2831 {
2832 	struct hmehash_bucket *hmebp;
2833 	int hmeshift;
2834 	void *htagid = sfmmutohtagid(sfmmup, rid);
2835 
2836 	ASSERT(htagid != NULL);
2837 
2838 	hmeshift = HME_HASH_SHIFT(size);
2839 
2840 	hmebp = HME_HASH_FUNCTION(htagid, vaddr, hmeshift);
2841 
2842 	SFMMU_HASH_LOCK(hmebp);
2843 
2844 	return (hmebp);
2845 }
2846 
2847 /*
2848  * Function returns a pointer to an hmeblk in the hash bucket, hmebp. If the
2849  * hmeblk doesn't exists for the [sfmmup, vaddr & size] signature, a hmeblk is
2850  * allocated.
2851  */
2852 static struct hme_blk *
2853 sfmmu_tteload_find_hmeblk(sfmmu_t *sfmmup, struct hmehash_bucket *hmebp,
2854 	caddr_t vaddr, uint_t size, uint_t flags, uint_t rid)
2855 {
2856 	hmeblk_tag hblktag;
2857 	int hmeshift;
2858 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
2859 
2860 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
2861 
2862 	hblktag.htag_id = sfmmutohtagid(sfmmup, rid);
2863 	ASSERT(hblktag.htag_id != NULL);
2864 	hmeshift = HME_HASH_SHIFT(size);
2865 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
2866 	hblktag.htag_rehash = HME_HASH_REHASH(size);
2867 	hblktag.htag_rid = rid;
2868 
2869 ttearray_realloc:
2870 
2871 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
2872 
2873 	/*
2874 	 * We block until hblk_reserve_lock is released; it's held by
2875 	 * the thread, temporarily using hblk_reserve, until hblk_reserve is
2876 	 * replaced by a hblk from sfmmu8_cache.
2877 	 */
2878 	if (hmeblkp == (struct hme_blk *)hblk_reserve &&
2879 	    hblk_reserve_thread != curthread) {
2880 		SFMMU_HASH_UNLOCK(hmebp);
2881 		mutex_enter(&hblk_reserve_lock);
2882 		mutex_exit(&hblk_reserve_lock);
2883 		SFMMU_STAT(sf_hblk_reserve_hit);
2884 		SFMMU_HASH_LOCK(hmebp);
2885 		goto ttearray_realloc;
2886 	}
2887 
2888 	if (hmeblkp == NULL) {
2889 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
2890 		    hblktag, flags, rid);
2891 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
2892 		ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
2893 	} else {
2894 		/*
2895 		 * It is possible for 8k and 64k hblks to collide since they
2896 		 * have the same rehash value. This is because we
2897 		 * lazily free hblks and 8K/64K blks could be lingering.
2898 		 * If we find size mismatch we free the block and & try again.
2899 		 */
2900 		if (get_hblk_ttesz(hmeblkp) != size) {
2901 			ASSERT(!hmeblkp->hblk_vcnt);
2902 			ASSERT(!hmeblkp->hblk_hmecnt);
2903 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
2904 			    &list, 0);
2905 			goto ttearray_realloc;
2906 		}
2907 		if (hmeblkp->hblk_shw_bit) {
2908 			/*
2909 			 * if the hblk was previously used as a shadow hblk then
2910 			 * we will change it to a normal hblk
2911 			 */
2912 			ASSERT(!hmeblkp->hblk_shared);
2913 			if (hmeblkp->hblk_shw_mask) {
2914 				sfmmu_shadow_hcleanup(sfmmup, hmeblkp, hmebp);
2915 				ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
2916 				goto ttearray_realloc;
2917 			} else {
2918 				hmeblkp->hblk_shw_bit = 0;
2919 			}
2920 		}
2921 		SFMMU_STAT(sf_hblk_hit);
2922 	}
2923 
2924 	/*
2925 	 * hat_memload() should never call kmem_cache_free() for kernel hmeblks;
2926 	 * see block comment showing the stacktrace in sfmmu_hblk_alloc();
2927 	 * set the flag parameter to 1 so that sfmmu_hblks_list_purge() will
2928 	 * just add these hmeblks to the per-cpu pending queue.
2929 	 */
2930 	sfmmu_hblks_list_purge(&list, 1);
2931 
2932 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
2933 	ASSERT(!hmeblkp->hblk_shw_bit);
2934 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
2935 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
2936 	ASSERT(hmeblkp->hblk_tag.htag_rid == rid);
2937 
2938 	return (hmeblkp);
2939 }
2940 
2941 /*
2942  * Function adds a tte entry into the hmeblk. It returns 0 if successful and 1
2943  * otherwise.
2944  */
2945 static int
2946 sfmmu_tteload_addentry(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, tte_t *ttep,
2947 	caddr_t vaddr, page_t **pps, uint_t flags, uint_t rid)
2948 {
2949 	page_t *pp = *pps;
2950 	int hmenum, size, remap;
2951 	tte_t tteold, flush_tte;
2952 #ifdef DEBUG
2953 	tte_t orig_old;
2954 #endif /* DEBUG */
2955 	struct sf_hment *sfhme;
2956 	kmutex_t *pml, *pmtx;
2957 	hatlock_t *hatlockp;
2958 	int myflt;
2959 
2960 	/*
2961 	 * remove this panic when we decide to let user virtual address
2962 	 * space be >= USERLIMIT.
2963 	 */
2964 	if (!TTE_IS_PRIVILEGED(ttep) && vaddr >= (caddr_t)USERLIMIT)
2965 		panic("user addr %p in kernel space", (void *)vaddr);
2966 #if defined(TTE_IS_GLOBAL)
2967 	if (TTE_IS_GLOBAL(ttep))
2968 		panic("sfmmu_tteload: creating global tte");
2969 #endif
2970 
2971 #ifdef DEBUG
2972 	if (pf_is_memory(sfmmu_ttetopfn(ttep, vaddr)) &&
2973 	    !TTE_IS_PCACHEABLE(ttep) && !sfmmu_allow_nc_trans)
2974 		panic("sfmmu_tteload: non cacheable memory tte");
2975 #endif /* DEBUG */
2976 
2977 	/* don't simulate dirty bit for writeable ISM/DISM mappings */
2978 	if ((flags & HAT_LOAD_SHARE) && TTE_IS_WRITABLE(ttep)) {
2979 		TTE_SET_REF(ttep);
2980 		TTE_SET_MOD(ttep);
2981 	}
2982 
2983 	if ((flags & HAT_LOAD_SHARE) || !TTE_IS_REF(ttep) ||
2984 	    !TTE_IS_MOD(ttep)) {
2985 		/*
2986 		 * Don't load TSB for dummy as in ISM.  Also don't preload
2987 		 * the TSB if the TTE isn't writable since we're likely to
2988 		 * fault on it again -- preloading can be fairly expensive.
2989 		 */
2990 		flags |= SFMMU_NO_TSBLOAD;
2991 	}
2992 
2993 	size = TTE_CSZ(ttep);
2994 	switch (size) {
2995 	case TTE8K:
2996 		SFMMU_STAT(sf_tteload8k);
2997 		break;
2998 	case TTE64K:
2999 		SFMMU_STAT(sf_tteload64k);
3000 		break;
3001 	case TTE512K:
3002 		SFMMU_STAT(sf_tteload512k);
3003 		break;
3004 	case TTE4M:
3005 		SFMMU_STAT(sf_tteload4m);
3006 		break;
3007 	case (TTE32M):
3008 		SFMMU_STAT(sf_tteload32m);
3009 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3010 		break;
3011 	case (TTE256M):
3012 		SFMMU_STAT(sf_tteload256m);
3013 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3014 		break;
3015 	}
3016 
3017 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
3018 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
3019 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3020 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3021 
3022 	HBLKTOHME_IDX(sfhme, hmeblkp, vaddr, hmenum);
3023 
3024 	/*
3025 	 * Need to grab mlist lock here so that pageunload
3026 	 * will not change tte behind us.
3027 	 */
3028 	if (pp) {
3029 		pml = sfmmu_mlist_enter(pp);
3030 	}
3031 
3032 	sfmmu_copytte(&sfhme->hme_tte, &tteold);
3033 	/*
3034 	 * Look for corresponding hment and if valid verify
3035 	 * pfns are equal.
3036 	 */
3037 	remap = TTE_IS_VALID(&tteold);
3038 	if (remap) {
3039 		pfn_t	new_pfn, old_pfn;
3040 
3041 		old_pfn = TTE_TO_PFN(vaddr, &tteold);
3042 		new_pfn = TTE_TO_PFN(vaddr, ttep);
3043 
3044 		if (flags & HAT_LOAD_REMAP) {
3045 			/* make sure we are remapping same type of pages */
3046 			if (pf_is_memory(old_pfn) != pf_is_memory(new_pfn)) {
3047 				panic("sfmmu_tteload - tte remap io<->memory");
3048 			}
3049 			if (old_pfn != new_pfn &&
3050 			    (pp != NULL || sfhme->hme_page != NULL)) {
3051 				panic("sfmmu_tteload - tte remap pp != NULL");
3052 			}
3053 		} else if (old_pfn != new_pfn) {
3054 			panic("sfmmu_tteload - tte remap, hmeblkp 0x%p",
3055 			    (void *)hmeblkp);
3056 		}
3057 		ASSERT(TTE_CSZ(&tteold) == TTE_CSZ(ttep));
3058 	}
3059 
3060 	if (pp) {
3061 		if (size == TTE8K) {
3062 #ifdef VAC
3063 			/*
3064 			 * Handle VAC consistency
3065 			 */
3066 			if (!remap && (cache & CACHE_VAC) && !PP_ISNC(pp)) {
3067 				sfmmu_vac_conflict(sfmmup, vaddr, pp);
3068 			}
3069 #endif
3070 
3071 			if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3072 				pmtx = sfmmu_page_enter(pp);
3073 				PP_CLRRO(pp);
3074 				sfmmu_page_exit(pmtx);
3075 			} else if (!PP_ISMAPPED(pp) &&
3076 			    (!TTE_IS_WRITABLE(ttep)) && !(PP_ISMOD(pp))) {
3077 				pmtx = sfmmu_page_enter(pp);
3078 				if (!(PP_ISMOD(pp))) {
3079 					PP_SETRO(pp);
3080 				}
3081 				sfmmu_page_exit(pmtx);
3082 			}
3083 
3084 		} else if (sfmmu_pagearray_setup(vaddr, pps, ttep, remap)) {
3085 			/*
3086 			 * sfmmu_pagearray_setup failed so return
3087 			 */
3088 			sfmmu_mlist_exit(pml);
3089 			return (1);
3090 		}
3091 	}
3092 
3093 	/*
3094 	 * Make sure hment is not on a mapping list.
3095 	 */
3096 	ASSERT(remap || (sfhme->hme_page == NULL));
3097 
3098 	/* if it is not a remap then hme->next better be NULL */
3099 	ASSERT((!remap) ? sfhme->hme_next == NULL : 1);
3100 
3101 	if (flags & HAT_LOAD_LOCK) {
3102 		if ((hmeblkp->hblk_lckcnt + 1) >= MAX_HBLK_LCKCNT) {
3103 			panic("too high lckcnt-hmeblk %p",
3104 			    (void *)hmeblkp);
3105 		}
3106 		atomic_add_32(&hmeblkp->hblk_lckcnt, 1);
3107 
3108 		HBLK_STACK_TRACE(hmeblkp, HBLK_LOCK);
3109 	}
3110 
3111 #ifdef VAC
3112 	if (pp && PP_ISNC(pp)) {
3113 		/*
3114 		 * If the physical page is marked to be uncacheable, like
3115 		 * by a vac conflict, make sure the new mapping is also
3116 		 * uncacheable.
3117 		 */
3118 		TTE_CLR_VCACHEABLE(ttep);
3119 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
3120 	}
3121 #endif
3122 	ttep->tte_hmenum = hmenum;
3123 
3124 #ifdef DEBUG
3125 	orig_old = tteold;
3126 #endif /* DEBUG */
3127 
3128 	while (sfmmu_modifytte_try(&tteold, ttep, &sfhme->hme_tte) < 0) {
3129 		if ((sfmmup == KHATID) &&
3130 		    (flags & (HAT_LOAD_LOCK | HAT_LOAD_REMAP))) {
3131 			sfmmu_copytte(&sfhme->hme_tte, &tteold);
3132 		}
3133 #ifdef DEBUG
3134 		chk_tte(&orig_old, &tteold, ttep, hmeblkp);
3135 #endif /* DEBUG */
3136 	}
3137 	ASSERT(TTE_IS_VALID(&sfhme->hme_tte));
3138 
3139 	if (!TTE_IS_VALID(&tteold)) {
3140 
3141 		atomic_add_16(&hmeblkp->hblk_vcnt, 1);
3142 		if (rid == SFMMU_INVALID_SHMERID) {
3143 			atomic_add_long(&sfmmup->sfmmu_ttecnt[size], 1);
3144 		} else {
3145 			sf_srd_t *srdp = sfmmup->sfmmu_srdp;
3146 			sf_region_t *rgnp = srdp->srd_hmergnp[rid];
3147 			/*
3148 			 * We already accounted for region ttecnt's in sfmmu
3149 			 * during hat_join_region() processing. Here we
3150 			 * only update ttecnt's in region struture.
3151 			 */
3152 			atomic_add_long(&rgnp->rgn_ttecnt[size], 1);
3153 		}
3154 	}
3155 
3156 	myflt = (astosfmmu(curthread->t_procp->p_as) == sfmmup);
3157 	if (size > TTE8K && (flags & HAT_LOAD_SHARE) == 0 &&
3158 	    sfmmup != ksfmmup) {
3159 		uchar_t tteflag = 1 << size;
3160 		if (rid == SFMMU_INVALID_SHMERID) {
3161 			if (!(sfmmup->sfmmu_tteflags & tteflag)) {
3162 				hatlockp = sfmmu_hat_enter(sfmmup);
3163 				sfmmup->sfmmu_tteflags |= tteflag;
3164 				sfmmu_hat_exit(hatlockp);
3165 			}
3166 		} else if (!(sfmmup->sfmmu_rtteflags & tteflag)) {
3167 			hatlockp = sfmmu_hat_enter(sfmmup);
3168 			sfmmup->sfmmu_rtteflags |= tteflag;
3169 			sfmmu_hat_exit(hatlockp);
3170 		}
3171 		/*
3172 		 * Update the current CPU tsbmiss area, so the current thread
3173 		 * won't need to take the tsbmiss for the new pagesize.
3174 		 * The other threads in the process will update their tsb
3175 		 * miss area lazily in sfmmu_tsbmiss_exception() when they
3176 		 * fail to find the translation for a newly added pagesize.
3177 		 */
3178 		if (size > TTE64K && myflt) {
3179 			struct tsbmiss *tsbmp;
3180 			kpreempt_disable();
3181 			tsbmp = &tsbmiss_area[CPU->cpu_id];
3182 			if (rid == SFMMU_INVALID_SHMERID) {
3183 				if (!(tsbmp->uhat_tteflags & tteflag)) {
3184 					tsbmp->uhat_tteflags |= tteflag;
3185 				}
3186 			} else {
3187 				if (!(tsbmp->uhat_rtteflags & tteflag)) {
3188 					tsbmp->uhat_rtteflags |= tteflag;
3189 				}
3190 			}
3191 			kpreempt_enable();
3192 		}
3193 	}
3194 
3195 	if (size >= TTE4M && (flags & HAT_LOAD_TEXT) &&
3196 	    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
3197 		hatlockp = sfmmu_hat_enter(sfmmup);
3198 		SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
3199 		sfmmu_hat_exit(hatlockp);
3200 	}
3201 
3202 	flush_tte.tte_intlo = (tteold.tte_intlo ^ ttep->tte_intlo) &
3203 	    hw_tte.tte_intlo;
3204 	flush_tte.tte_inthi = (tteold.tte_inthi ^ ttep->tte_inthi) &
3205 	    hw_tte.tte_inthi;
3206 
3207 	if (remap && (flush_tte.tte_inthi || flush_tte.tte_intlo)) {
3208 		/*
3209 		 * If remap and new tte differs from old tte we need
3210 		 * to sync the mod bit and flush TLB/TSB.  We don't
3211 		 * need to sync ref bit because we currently always set
3212 		 * ref bit in tteload.
3213 		 */
3214 		ASSERT(TTE_IS_REF(ttep));
3215 		if (TTE_IS_MOD(&tteold)) {
3216 			sfmmu_ttesync(sfmmup, vaddr, &tteold, pp);
3217 		}
3218 		/*
3219 		 * hwtte bits shouldn't change for SRD hmeblks as long as SRD
3220 		 * hmes are only used for read only text. Adding this code for
3221 		 * completeness and future use of shared hmeblks with writable
3222 		 * mappings of VMODSORT vnodes.
3223 		 */
3224 		if (hmeblkp->hblk_shared) {
3225 			cpuset_t cpuset = sfmmu_rgntlb_demap(vaddr,
3226 			    sfmmup->sfmmu_srdp->srd_hmergnp[rid], hmeblkp, 1);
3227 			xt_sync(cpuset);
3228 			SFMMU_STAT_ADD(sf_region_remap_demap, 1);
3229 		} else {
3230 			sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 0);
3231 			xt_sync(sfmmup->sfmmu_cpusran);
3232 		}
3233 	}
3234 
3235 	if ((flags & SFMMU_NO_TSBLOAD) == 0) {
3236 		/*
3237 		 * We only preload 8K and 4M mappings into the TSB, since
3238 		 * 64K and 512K mappings are replicated and hence don't
3239 		 * have a single, unique TSB entry. Ditto for 32M/256M.
3240 		 */
3241 		if (size == TTE8K || size == TTE4M) {
3242 			sf_scd_t *scdp;
3243 			hatlockp = sfmmu_hat_enter(sfmmup);
3244 			/*
3245 			 * Don't preload private TSB if the mapping is used
3246 			 * by the shctx in the SCD.
3247 			 */
3248 			scdp = sfmmup->sfmmu_scdp;
3249 			if (rid == SFMMU_INVALID_SHMERID || scdp == NULL ||
3250 			    !SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
3251 				sfmmu_load_tsb(sfmmup, vaddr, &sfhme->hme_tte,
3252 				    size);
3253 			}
3254 			sfmmu_hat_exit(hatlockp);
3255 		}
3256 	}
3257 	if (pp) {
3258 		if (!remap) {
3259 			HME_ADD(sfhme, pp);
3260 			atomic_add_16(&hmeblkp->hblk_hmecnt, 1);
3261 			ASSERT(hmeblkp->hblk_hmecnt > 0);
3262 
3263 			/*
3264 			 * Cannot ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
3265 			 * see pageunload() for comment.
3266 			 */
3267 		}
3268 		sfmmu_mlist_exit(pml);
3269 	}
3270 
3271 	return (0);
3272 }
3273 /*
3274  * Function unlocks hash bucket.
3275  */
3276 static void
3277 sfmmu_tteload_release_hashbucket(struct hmehash_bucket *hmebp)
3278 {
3279 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3280 	SFMMU_HASH_UNLOCK(hmebp);
3281 }
3282 
3283 /*
3284  * function which checks and sets up page array for a large
3285  * translation.  Will set p_vcolor, p_index, p_ro fields.
3286  * Assumes addr and pfnum of first page are properly aligned.
3287  * Will check for physical contiguity. If check fails it return
3288  * non null.
3289  */
3290 static int
3291 sfmmu_pagearray_setup(caddr_t addr, page_t **pps, tte_t *ttep, int remap)
3292 {
3293 	int 	i, index, ttesz;
3294 	pfn_t	pfnum;
3295 	pgcnt_t	npgs;
3296 	page_t *pp, *pp1;
3297 	kmutex_t *pmtx;
3298 #ifdef VAC
3299 	int osz;
3300 	int cflags = 0;
3301 	int vac_err = 0;
3302 #endif
3303 	int newidx = 0;
3304 
3305 	ttesz = TTE_CSZ(ttep);
3306 
3307 	ASSERT(ttesz > TTE8K);
3308 
3309 	npgs = TTEPAGES(ttesz);
3310 	index = PAGESZ_TO_INDEX(ttesz);
3311 
3312 	pfnum = (*pps)->p_pagenum;
3313 	ASSERT(IS_P2ALIGNED(pfnum, npgs));
3314 
3315 	/*
3316 	 * Save the first pp so we can do HAT_TMPNC at the end.
3317 	 */
3318 	pp1 = *pps;
3319 #ifdef VAC
3320 	osz = fnd_mapping_sz(pp1);
3321 #endif
3322 
3323 	for (i = 0; i < npgs; i++, pps++) {
3324 		pp = *pps;
3325 		ASSERT(PAGE_LOCKED(pp));
3326 		ASSERT(pp->p_szc >= ttesz);
3327 		ASSERT(pp->p_szc == pp1->p_szc);
3328 		ASSERT(sfmmu_mlist_held(pp));
3329 
3330 		/*
3331 		 * XXX is it possible to maintain P_RO on the root only?
3332 		 */
3333 		if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3334 			pmtx = sfmmu_page_enter(pp);
3335 			PP_CLRRO(pp);
3336 			sfmmu_page_exit(pmtx);
3337 		} else if (!PP_ISMAPPED(pp) && !TTE_IS_WRITABLE(ttep) &&
3338 		    !PP_ISMOD(pp)) {
3339 			pmtx = sfmmu_page_enter(pp);
3340 			if (!(PP_ISMOD(pp))) {
3341 				PP_SETRO(pp);
3342 			}
3343 			sfmmu_page_exit(pmtx);
3344 		}
3345 
3346 		/*
3347 		 * If this is a remap we skip vac & contiguity checks.
3348 		 */
3349 		if (remap)
3350 			continue;
3351 
3352 		/*
3353 		 * set p_vcolor and detect any vac conflicts.
3354 		 */
3355 #ifdef VAC
3356 		if (vac_err == 0) {
3357 			vac_err = sfmmu_vacconflict_array(addr, pp, &cflags);
3358 
3359 		}
3360 #endif
3361 
3362 		/*
3363 		 * Save current index in case we need to undo it.
3364 		 * Note: "PAGESZ_TO_INDEX(sz)	(1 << (sz))"
3365 		 *	"SFMMU_INDEX_SHIFT	6"
3366 		 *	 "SFMMU_INDEX_MASK	((1 << SFMMU_INDEX_SHIFT) - 1)"
3367 		 *	 "PP_MAPINDEX(p_index)	(p_index & SFMMU_INDEX_MASK)"
3368 		 *
3369 		 * So:	index = PAGESZ_TO_INDEX(ttesz);
3370 		 *	if ttesz == 1 then index = 0x2
3371 		 *		    2 then index = 0x4
3372 		 *		    3 then index = 0x8
3373 		 *		    4 then index = 0x10
3374 		 *		    5 then index = 0x20
3375 		 * The code below checks if it's a new pagesize (ie, newidx)
3376 		 * in case we need to take it back out of p_index,
3377 		 * and then or's the new index into the existing index.
3378 		 */
3379 		if ((PP_MAPINDEX(pp) & index) == 0)
3380 			newidx = 1;
3381 		pp->p_index = (PP_MAPINDEX(pp) | index);
3382 
3383 		/*
3384 		 * contiguity check
3385 		 */
3386 		if (pp->p_pagenum != pfnum) {
3387 			/*
3388 			 * If we fail the contiguity test then
3389 			 * the only thing we need to fix is the p_index field.
3390 			 * We might get a few extra flushes but since this
3391 			 * path is rare that is ok.  The p_ro field will
3392 			 * get automatically fixed on the next tteload to
3393 			 * the page.  NO TNC bit is set yet.
3394 			 */
3395 			while (i >= 0) {
3396 				pp = *pps;
3397 				if (newidx)
3398 					pp->p_index = (PP_MAPINDEX(pp) &
3399 					    ~index);
3400 				pps--;
3401 				i--;
3402 			}
3403 			return (1);
3404 		}
3405 		pfnum++;
3406 		addr += MMU_PAGESIZE;
3407 	}
3408 
3409 #ifdef VAC
3410 	if (vac_err) {
3411 		if (ttesz > osz) {
3412 			/*
3413 			 * There are some smaller mappings that causes vac
3414 			 * conflicts. Convert all existing small mappings to
3415 			 * TNC.
3416 			 */
3417 			SFMMU_STAT_ADD(sf_uncache_conflict, npgs);
3418 			sfmmu_page_cache_array(pp1, HAT_TMPNC, CACHE_FLUSH,
3419 			    npgs);
3420 		} else {
3421 			/* EMPTY */
3422 			/*
3423 			 * If there exists an big page mapping,
3424 			 * that means the whole existing big page
3425 			 * has TNC setting already. No need to covert to
3426 			 * TNC again.
3427 			 */
3428 			ASSERT(PP_ISTNC(pp1));
3429 		}
3430 	}
3431 #endif	/* VAC */
3432 
3433 	return (0);
3434 }
3435 
3436 #ifdef VAC
3437 /*
3438  * Routine that detects vac consistency for a large page. It also
3439  * sets virtual color for all pp's for this big mapping.
3440  */
3441 static int
3442 sfmmu_vacconflict_array(caddr_t addr, page_t *pp, int *cflags)
3443 {
3444 	int vcolor, ocolor;
3445 
3446 	ASSERT(sfmmu_mlist_held(pp));
3447 
3448 	if (PP_ISNC(pp)) {
3449 		return (HAT_TMPNC);
3450 	}
3451 
3452 	vcolor = addr_to_vcolor(addr);
3453 	if (PP_NEWPAGE(pp)) {
3454 		PP_SET_VCOLOR(pp, vcolor);
3455 		return (0);
3456 	}
3457 
3458 	ocolor = PP_GET_VCOLOR(pp);
3459 	if (ocolor == vcolor) {
3460 		return (0);
3461 	}
3462 
3463 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
3464 		/*
3465 		 * Previous user of page had a differnet color
3466 		 * but since there are no current users
3467 		 * we just flush the cache and change the color.
3468 		 * As an optimization for large pages we flush the
3469 		 * entire cache of that color and set a flag.
3470 		 */
3471 		SFMMU_STAT(sf_pgcolor_conflict);
3472 		if (!CacheColor_IsFlushed(*cflags, ocolor)) {
3473 			CacheColor_SetFlushed(*cflags, ocolor);
3474 			sfmmu_cache_flushcolor(ocolor, pp->p_pagenum);
3475 		}
3476 		PP_SET_VCOLOR(pp, vcolor);
3477 		return (0);
3478 	}
3479 
3480 	/*
3481 	 * We got a real conflict with a current mapping.
3482 	 * set flags to start unencaching all mappings
3483 	 * and return failure so we restart looping
3484 	 * the pp array from the beginning.
3485 	 */
3486 	return (HAT_TMPNC);
3487 }
3488 #endif	/* VAC */
3489 
3490 /*
3491  * creates a large page shadow hmeblk for a tte.
3492  * The purpose of this routine is to allow us to do quick unloads because
3493  * the vm layer can easily pass a very large but sparsely populated range.
3494  */
3495 static struct hme_blk *
3496 sfmmu_shadow_hcreate(sfmmu_t *sfmmup, caddr_t vaddr, int ttesz, uint_t flags)
3497 {
3498 	struct hmehash_bucket *hmebp;
3499 	hmeblk_tag hblktag;
3500 	int hmeshift, size, vshift;
3501 	uint_t shw_mask, newshw_mask;
3502 	struct hme_blk *hmeblkp;
3503 
3504 	ASSERT(sfmmup != KHATID);
3505 	if (mmu_page_sizes == max_mmu_page_sizes) {
3506 		ASSERT(ttesz < TTE256M);
3507 	} else {
3508 		ASSERT(ttesz < TTE4M);
3509 		ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
3510 		ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
3511 	}
3512 
3513 	if (ttesz == TTE8K) {
3514 		size = TTE512K;
3515 	} else {
3516 		size = ++ttesz;
3517 	}
3518 
3519 	hblktag.htag_id = sfmmup;
3520 	hmeshift = HME_HASH_SHIFT(size);
3521 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
3522 	hblktag.htag_rehash = HME_HASH_REHASH(size);
3523 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3524 	hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
3525 
3526 	SFMMU_HASH_LOCK(hmebp);
3527 
3528 	HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
3529 	ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
3530 	if (hmeblkp == NULL) {
3531 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
3532 		    hblktag, flags, SFMMU_INVALID_SHMERID);
3533 	}
3534 	ASSERT(hmeblkp);
3535 	if (!hmeblkp->hblk_shw_mask) {
3536 		/*
3537 		 * if this is a unused hblk it was just allocated or could
3538 		 * potentially be a previous large page hblk so we need to
3539 		 * set the shadow bit.
3540 		 */
3541 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3542 		hmeblkp->hblk_shw_bit = 1;
3543 	} else if (hmeblkp->hblk_shw_bit == 0) {
3544 		panic("sfmmu_shadow_hcreate: shw bit not set in hmeblkp 0x%p",
3545 		    (void *)hmeblkp);
3546 	}
3547 	ASSERT(hmeblkp->hblk_shw_bit == 1);
3548 	ASSERT(!hmeblkp->hblk_shared);
3549 	vshift = vaddr_to_vshift(hblktag, vaddr, size);
3550 	ASSERT(vshift < 8);
3551 	/*
3552 	 * Atomically set shw mask bit
3553 	 */
3554 	do {
3555 		shw_mask = hmeblkp->hblk_shw_mask;
3556 		newshw_mask = shw_mask | (1 << vshift);
3557 		newshw_mask = cas32(&hmeblkp->hblk_shw_mask, shw_mask,
3558 		    newshw_mask);
3559 	} while (newshw_mask != shw_mask);
3560 
3561 	SFMMU_HASH_UNLOCK(hmebp);
3562 
3563 	return (hmeblkp);
3564 }
3565 
3566 /*
3567  * This routine cleanup a previous shadow hmeblk and changes it to
3568  * a regular hblk.  This happens rarely but it is possible
3569  * when a process wants to use large pages and there are hblks still
3570  * lying around from the previous as that used these hmeblks.
3571  * The alternative was to cleanup the shadow hblks at unload time
3572  * but since so few user processes actually use large pages, it is
3573  * better to be lazy and cleanup at this time.
3574  */
3575 static void
3576 sfmmu_shadow_hcleanup(sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
3577 	struct hmehash_bucket *hmebp)
3578 {
3579 	caddr_t addr, endaddr;
3580 	int hashno, size;
3581 
3582 	ASSERT(hmeblkp->hblk_shw_bit);
3583 	ASSERT(!hmeblkp->hblk_shared);
3584 
3585 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3586 
3587 	if (!hmeblkp->hblk_shw_mask) {
3588 		hmeblkp->hblk_shw_bit = 0;
3589 		return;
3590 	}
3591 	addr = (caddr_t)get_hblk_base(hmeblkp);
3592 	endaddr = get_hblk_endaddr(hmeblkp);
3593 	size = get_hblk_ttesz(hmeblkp);
3594 	hashno = size - 1;
3595 	ASSERT(hashno > 0);
3596 	SFMMU_HASH_UNLOCK(hmebp);
3597 
3598 	sfmmu_free_hblks(sfmmup, addr, endaddr, hashno);
3599 
3600 	SFMMU_HASH_LOCK(hmebp);
3601 }
3602 
3603 static void
3604 sfmmu_free_hblks(sfmmu_t *sfmmup, caddr_t addr, caddr_t endaddr,
3605 	int hashno)
3606 {
3607 	int hmeshift, shadow = 0;
3608 	hmeblk_tag hblktag;
3609 	struct hmehash_bucket *hmebp;
3610 	struct hme_blk *hmeblkp;
3611 	struct hme_blk *nx_hblk, *pr_hblk, *list = NULL;
3612 
3613 	ASSERT(hashno > 0);
3614 	hblktag.htag_id = sfmmup;
3615 	hblktag.htag_rehash = hashno;
3616 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3617 
3618 	hmeshift = HME_HASH_SHIFT(hashno);
3619 
3620 	while (addr < endaddr) {
3621 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3622 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3623 		SFMMU_HASH_LOCK(hmebp);
3624 		/* inline HME_HASH_SEARCH */
3625 		hmeblkp = hmebp->hmeblkp;
3626 		pr_hblk = NULL;
3627 		while (hmeblkp) {
3628 			if (HTAGS_EQ(hmeblkp->hblk_tag, hblktag)) {
3629 				/* found hme_blk */
3630 				ASSERT(!hmeblkp->hblk_shared);
3631 				if (hmeblkp->hblk_shw_bit) {
3632 					if (hmeblkp->hblk_shw_mask) {
3633 						shadow = 1;
3634 						sfmmu_shadow_hcleanup(sfmmup,
3635 						    hmeblkp, hmebp);
3636 						break;
3637 					} else {
3638 						hmeblkp->hblk_shw_bit = 0;
3639 					}
3640 				}
3641 
3642 				/*
3643 				 * Hblk_hmecnt and hblk_vcnt could be non zero
3644 				 * since hblk_unload() does not gurantee that.
3645 				 *
3646 				 * XXX - this could cause tteload() to spin
3647 				 * where sfmmu_shadow_hcleanup() is called.
3648 				 */
3649 			}
3650 
3651 			nx_hblk = hmeblkp->hblk_next;
3652 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
3653 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3654 				    &list, 0);
3655 			} else {
3656 				pr_hblk = hmeblkp;
3657 			}
3658 			hmeblkp = nx_hblk;
3659 		}
3660 
3661 		SFMMU_HASH_UNLOCK(hmebp);
3662 
3663 		if (shadow) {
3664 			/*
3665 			 * We found another shadow hblk so cleaned its
3666 			 * children.  We need to go back and cleanup
3667 			 * the original hblk so we don't change the
3668 			 * addr.
3669 			 */
3670 			shadow = 0;
3671 		} else {
3672 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
3673 			    (1 << hmeshift));
3674 		}
3675 	}
3676 	sfmmu_hblks_list_purge(&list, 0);
3677 }
3678 
3679 /*
3680  * This routine's job is to delete stale invalid shared hmeregions hmeblks that
3681  * may still linger on after pageunload.
3682  */
3683 static void
3684 sfmmu_cleanup_rhblk(sf_srd_t *srdp, caddr_t addr, uint_t rid, int ttesz)
3685 {
3686 	int hmeshift;
3687 	hmeblk_tag hblktag;
3688 	struct hmehash_bucket *hmebp;
3689 	struct hme_blk *hmeblkp;
3690 	struct hme_blk *pr_hblk;
3691 	struct hme_blk *list = NULL;
3692 
3693 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3694 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3695 
3696 	hmeshift = HME_HASH_SHIFT(ttesz);
3697 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3698 	hblktag.htag_rehash = ttesz;
3699 	hblktag.htag_rid = rid;
3700 	hblktag.htag_id = srdp;
3701 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3702 
3703 	SFMMU_HASH_LOCK(hmebp);
3704 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3705 	if (hmeblkp != NULL) {
3706 		ASSERT(hmeblkp->hblk_shared);
3707 		ASSERT(!hmeblkp->hblk_shw_bit);
3708 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3709 			panic("sfmmu_cleanup_rhblk: valid hmeblk");
3710 		}
3711 		ASSERT(!hmeblkp->hblk_lckcnt);
3712 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3713 		    &list, 0);
3714 	}
3715 	SFMMU_HASH_UNLOCK(hmebp);
3716 	sfmmu_hblks_list_purge(&list, 0);
3717 }
3718 
3719 /* ARGSUSED */
3720 static void
3721 sfmmu_rgn_cb_noop(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
3722     size_t r_size, void *r_obj, u_offset_t r_objoff)
3723 {
3724 }
3725 
3726 /*
3727  * Searches for an hmeblk which maps addr, then unloads this mapping
3728  * and updates *eaddrp, if the hmeblk is found.
3729  */
3730 static void
3731 sfmmu_unload_hmeregion_va(sf_srd_t *srdp, uint_t rid, caddr_t addr,
3732     caddr_t eaddr, int ttesz, caddr_t *eaddrp)
3733 {
3734 	int hmeshift;
3735 	hmeblk_tag hblktag;
3736 	struct hmehash_bucket *hmebp;
3737 	struct hme_blk *hmeblkp;
3738 	struct hme_blk *pr_hblk;
3739 	struct hme_blk *list = NULL;
3740 
3741 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3742 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3743 	ASSERT(ttesz >= HBLK_MIN_TTESZ);
3744 
3745 	hmeshift = HME_HASH_SHIFT(ttesz);
3746 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3747 	hblktag.htag_rehash = ttesz;
3748 	hblktag.htag_rid = rid;
3749 	hblktag.htag_id = srdp;
3750 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3751 
3752 	SFMMU_HASH_LOCK(hmebp);
3753 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3754 	if (hmeblkp != NULL) {
3755 		ASSERT(hmeblkp->hblk_shared);
3756 		ASSERT(!hmeblkp->hblk_lckcnt);
3757 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3758 			*eaddrp = sfmmu_hblk_unload(NULL, hmeblkp, addr,
3759 			    eaddr, NULL, HAT_UNLOAD);
3760 			ASSERT(*eaddrp > addr);
3761 		}
3762 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3763 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3764 		    &list, 0);
3765 	}
3766 	SFMMU_HASH_UNLOCK(hmebp);
3767 	sfmmu_hblks_list_purge(&list, 0);
3768 }
3769 
3770 static void
3771 sfmmu_unload_hmeregion(sf_srd_t *srdp, sf_region_t *rgnp)
3772 {
3773 	int ttesz = rgnp->rgn_pgszc;
3774 	size_t rsz = rgnp->rgn_size;
3775 	caddr_t rsaddr = rgnp->rgn_saddr;
3776 	caddr_t readdr = rsaddr + rsz;
3777 	caddr_t rhsaddr;
3778 	caddr_t va;
3779 	uint_t rid = rgnp->rgn_id;
3780 	caddr_t cbsaddr;
3781 	caddr_t cbeaddr;
3782 	hat_rgn_cb_func_t rcbfunc;
3783 	ulong_t cnt;
3784 
3785 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3786 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3787 
3788 	ASSERT(IS_P2ALIGNED(rsaddr, TTEBYTES(ttesz)));
3789 	ASSERT(IS_P2ALIGNED(rsz, TTEBYTES(ttesz)));
3790 	if (ttesz < HBLK_MIN_TTESZ) {
3791 		ttesz = HBLK_MIN_TTESZ;
3792 		rhsaddr = (caddr_t)P2ALIGN((uintptr_t)rsaddr, HBLK_MIN_BYTES);
3793 	} else {
3794 		rhsaddr = rsaddr;
3795 	}
3796 
3797 	if ((rcbfunc = rgnp->rgn_cb_function) == NULL) {
3798 		rcbfunc = sfmmu_rgn_cb_noop;
3799 	}
3800 
3801 	while (ttesz >= HBLK_MIN_TTESZ) {
3802 		cbsaddr = rsaddr;
3803 		cbeaddr = rsaddr;
3804 		if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
3805 			ttesz--;
3806 			continue;
3807 		}
3808 		cnt = 0;
3809 		va = rsaddr;
3810 		while (va < readdr) {
3811 			ASSERT(va >= rhsaddr);
3812 			if (va != cbeaddr) {
3813 				if (cbeaddr != cbsaddr) {
3814 					ASSERT(cbeaddr > cbsaddr);
3815 					(*rcbfunc)(cbsaddr, cbeaddr,
3816 					    rsaddr, rsz, rgnp->rgn_obj,
3817 					    rgnp->rgn_objoff);
3818 				}
3819 				cbsaddr = va;
3820 				cbeaddr = va;
3821 			}
3822 			sfmmu_unload_hmeregion_va(srdp, rid, va, readdr,
3823 			    ttesz, &cbeaddr);
3824 			cnt++;
3825 			va = rhsaddr + (cnt << TTE_PAGE_SHIFT(ttesz));
3826 		}
3827 		if (cbeaddr != cbsaddr) {
3828 			ASSERT(cbeaddr > cbsaddr);
3829 			(*rcbfunc)(cbsaddr, cbeaddr, rsaddr,
3830 			    rsz, rgnp->rgn_obj,
3831 			    rgnp->rgn_objoff);
3832 		}
3833 		ttesz--;
3834 	}
3835 }
3836 
3837 /*
3838  * Release one hardware address translation lock on the given address range.
3839  */
3840 void
3841 hat_unlock(struct hat *sfmmup, caddr_t addr, size_t len)
3842 {
3843 	struct hmehash_bucket *hmebp;
3844 	hmeblk_tag hblktag;
3845 	int hmeshift, hashno = 1;
3846 	struct hme_blk *hmeblkp, *list = NULL;
3847 	caddr_t endaddr;
3848 
3849 	ASSERT(sfmmup != NULL);
3850 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3851 
3852 	ASSERT((sfmmup == ksfmmup) ||
3853 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
3854 	ASSERT((len & MMU_PAGEOFFSET) == 0);
3855 	endaddr = addr + len;
3856 	hblktag.htag_id = sfmmup;
3857 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3858 
3859 	/*
3860 	 * Spitfire supports 4 page sizes.
3861 	 * Most pages are expected to be of the smallest page size (8K) and
3862 	 * these will not need to be rehashed. 64K pages also don't need to be
3863 	 * rehashed because an hmeblk spans 64K of address space. 512K pages
3864 	 * might need 1 rehash and and 4M pages might need 2 rehashes.
3865 	 */
3866 	while (addr < endaddr) {
3867 		hmeshift = HME_HASH_SHIFT(hashno);
3868 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3869 		hblktag.htag_rehash = hashno;
3870 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3871 
3872 		SFMMU_HASH_LOCK(hmebp);
3873 
3874 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
3875 		if (hmeblkp != NULL) {
3876 			ASSERT(!hmeblkp->hblk_shared);
3877 			/*
3878 			 * If we encounter a shadow hmeblk then
3879 			 * we know there are no valid hmeblks mapping
3880 			 * this address at this size or larger.
3881 			 * Just increment address by the smallest
3882 			 * page size.
3883 			 */
3884 			if (hmeblkp->hblk_shw_bit) {
3885 				addr += MMU_PAGESIZE;
3886 			} else {
3887 				addr = sfmmu_hblk_unlock(hmeblkp, addr,
3888 				    endaddr);
3889 			}
3890 			SFMMU_HASH_UNLOCK(hmebp);
3891 			hashno = 1;
3892 			continue;
3893 		}
3894 		SFMMU_HASH_UNLOCK(hmebp);
3895 
3896 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
3897 			/*
3898 			 * We have traversed the whole list and rehashed
3899 			 * if necessary without finding the address to unlock
3900 			 * which should never happen.
3901 			 */
3902 			panic("sfmmu_unlock: addr not found. "
3903 			    "addr %p hat %p", (void *)addr, (void *)sfmmup);
3904 		} else {
3905 			hashno++;
3906 		}
3907 	}
3908 
3909 	sfmmu_hblks_list_purge(&list, 0);
3910 }
3911 
3912 void
3913 hat_unlock_region(struct hat *sfmmup, caddr_t addr, size_t len,
3914     hat_region_cookie_t rcookie)
3915 {
3916 	sf_srd_t *srdp;
3917 	sf_region_t *rgnp;
3918 	int ttesz;
3919 	uint_t rid;
3920 	caddr_t eaddr;
3921 	caddr_t va;
3922 	int hmeshift;
3923 	hmeblk_tag hblktag;
3924 	struct hmehash_bucket *hmebp;
3925 	struct hme_blk *hmeblkp;
3926 	struct hme_blk *pr_hblk;
3927 	struct hme_blk *list;
3928 
3929 	if (rcookie == HAT_INVALID_REGION_COOKIE) {
3930 		hat_unlock(sfmmup, addr, len);
3931 		return;
3932 	}
3933 
3934 	ASSERT(sfmmup != NULL);
3935 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3936 	ASSERT(sfmmup != ksfmmup);
3937 
3938 	srdp = sfmmup->sfmmu_srdp;
3939 	rid = (uint_t)((uint64_t)rcookie);
3940 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3941 	eaddr = addr + len;
3942 	va = addr;
3943 	list = NULL;
3944 	rgnp = srdp->srd_hmergnp[rid];
3945 	SFMMU_VALIDATE_HMERID(sfmmup, rid, addr, len);
3946 
3947 	ASSERT(IS_P2ALIGNED(addr, TTEBYTES(rgnp->rgn_pgszc)));
3948 	ASSERT(IS_P2ALIGNED(len, TTEBYTES(rgnp->rgn_pgszc)));
3949 	if (rgnp->rgn_pgszc < HBLK_MIN_TTESZ) {
3950 		ttesz = HBLK_MIN_TTESZ;
3951 	} else {
3952 		ttesz = rgnp->rgn_pgszc;
3953 	}
3954 	while (va < eaddr) {
3955 		while (ttesz < rgnp->rgn_pgszc &&
3956 		    IS_P2ALIGNED(va, TTEBYTES(ttesz + 1))) {
3957 			ttesz++;
3958 		}
3959 		while (ttesz >= HBLK_MIN_TTESZ) {
3960 			if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
3961 				ttesz--;
3962 				continue;
3963 			}
3964 			hmeshift = HME_HASH_SHIFT(ttesz);
3965 			hblktag.htag_bspage = HME_HASH_BSPAGE(va, hmeshift);
3966 			hblktag.htag_rehash = ttesz;
3967 			hblktag.htag_rid = rid;
3968 			hblktag.htag_id = srdp;
3969 			hmebp = HME_HASH_FUNCTION(srdp, va, hmeshift);
3970 			SFMMU_HASH_LOCK(hmebp);
3971 			HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk,
3972 			    &list);
3973 			if (hmeblkp == NULL) {
3974 				SFMMU_HASH_UNLOCK(hmebp);
3975 				ttesz--;
3976 				continue;
3977 			}
3978 			ASSERT(hmeblkp->hblk_shared);
3979 			va = sfmmu_hblk_unlock(hmeblkp, va, eaddr);
3980 			ASSERT(va >= eaddr ||
3981 			    IS_P2ALIGNED((uintptr_t)va, TTEBYTES(ttesz)));
3982 			SFMMU_HASH_UNLOCK(hmebp);
3983 			break;
3984 		}
3985 		if (ttesz < HBLK_MIN_TTESZ) {
3986 			panic("hat_unlock_region: addr not found "
3987 			    "addr %p hat %p", (void *)va, (void *)sfmmup);
3988 		}
3989 	}
3990 	sfmmu_hblks_list_purge(&list, 0);
3991 }
3992 
3993 /*
3994  * Function to unlock a range of addresses in an hmeblk.  It returns the
3995  * next address that needs to be unlocked.
3996  * Should be called with the hash lock held.
3997  */
3998 static caddr_t
3999 sfmmu_hblk_unlock(struct hme_blk *hmeblkp, caddr_t addr, caddr_t endaddr)
4000 {
4001 	struct sf_hment *sfhme;
4002 	tte_t tteold, ttemod;
4003 	int ttesz, ret;
4004 
4005 	ASSERT(in_hblk_range(hmeblkp, addr));
4006 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4007 
4008 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4009 	ttesz = get_hblk_ttesz(hmeblkp);
4010 
4011 	HBLKTOHME(sfhme, hmeblkp, addr);
4012 	while (addr < endaddr) {
4013 readtte:
4014 		sfmmu_copytte(&sfhme->hme_tte, &tteold);
4015 		if (TTE_IS_VALID(&tteold)) {
4016 
4017 			ttemod = tteold;
4018 
4019 			ret = sfmmu_modifytte_try(&tteold, &ttemod,
4020 			    &sfhme->hme_tte);
4021 
4022 			if (ret < 0)
4023 				goto readtte;
4024 
4025 			if (hmeblkp->hblk_lckcnt == 0)
4026 				panic("zero hblk lckcnt");
4027 
4028 			if (((uintptr_t)addr + TTEBYTES(ttesz)) >
4029 			    (uintptr_t)endaddr)
4030 				panic("can't unlock large tte");
4031 
4032 			ASSERT(hmeblkp->hblk_lckcnt > 0);
4033 			atomic_add_32(&hmeblkp->hblk_lckcnt, -1);
4034 			HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
4035 		} else {
4036 			panic("sfmmu_hblk_unlock: invalid tte");
4037 		}
4038 		addr += TTEBYTES(ttesz);
4039 		sfhme++;
4040 	}
4041 	return (addr);
4042 }
4043 
4044 /*
4045  * Physical Address Mapping Framework
4046  *
4047  * General rules:
4048  *
4049  * (1) Applies only to seg_kmem memory pages. To make things easier,
4050  *     seg_kpm addresses are also accepted by the routines, but nothing
4051  *     is done with them since by definition their PA mappings are static.
4052  * (2) hat_add_callback() may only be called while holding the page lock
4053  *     SE_SHARED or SE_EXCL of the underlying page (e.g., as_pagelock()),
4054  *     or passing HAC_PAGELOCK flag.
4055  * (3) prehandler() and posthandler() may not call hat_add_callback() or
4056  *     hat_delete_callback(), nor should they allocate memory. Post quiesce
4057  *     callbacks may not sleep or acquire adaptive mutex locks.
4058  * (4) Either prehandler() or posthandler() (but not both) may be specified
4059  *     as being NULL.  Specifying an errhandler() is optional.
4060  *
4061  * Details of using the framework:
4062  *
4063  * registering a callback (hat_register_callback())
4064  *
4065  *	Pass prehandler, posthandler, errhandler addresses
4066  *	as described below. If capture_cpus argument is nonzero,
4067  *	suspend callback to the prehandler will occur with CPUs
4068  *	captured and executing xc_loop() and CPUs will remain
4069  *	captured until after the posthandler suspend callback
4070  *	occurs.
4071  *
4072  * adding a callback (hat_add_callback())
4073  *
4074  *      as_pagelock();
4075  *	hat_add_callback();
4076  *      save returned pfn in private data structures or program registers;
4077  *      as_pageunlock();
4078  *
4079  * prehandler()
4080  *
4081  *	Stop all accesses by physical address to this memory page.
4082  *	Called twice: the first, PRESUSPEND, is a context safe to acquire
4083  *	adaptive locks. The second, SUSPEND, is called at high PIL with
4084  *	CPUs captured so adaptive locks may NOT be acquired (and all spin
4085  *	locks must be XCALL_PIL or higher locks).
4086  *
4087  *	May return the following errors:
4088  *		EIO:	A fatal error has occurred. This will result in panic.
4089  *		EAGAIN:	The page cannot be suspended. This will fail the
4090  *			relocation.
4091  *		0:	Success.
4092  *
4093  * posthandler()
4094  *
4095  *      Save new pfn in private data structures or program registers;
4096  *	not allowed to fail (non-zero return values will result in panic).
4097  *
4098  * errhandler()
4099  *
4100  *	called when an error occurs related to the callback.  Currently
4101  *	the only such error is HAT_CB_ERR_LEAKED which indicates that
4102  *	a page is being freed, but there are still outstanding callback(s)
4103  *	registered on the page.
4104  *
4105  * removing a callback (hat_delete_callback(); e.g., prior to freeing memory)
4106  *
4107  *	stop using physical address
4108  *	hat_delete_callback();
4109  *
4110  */
4111 
4112 /*
4113  * Register a callback class.  Each subsystem should do this once and
4114  * cache the id_t returned for use in setting up and tearing down callbacks.
4115  *
4116  * There is no facility for removing callback IDs once they are created;
4117  * the "key" should be unique for each module, so in case a module is unloaded
4118  * and subsequently re-loaded, we can recycle the module's previous entry.
4119  */
4120 id_t
4121 hat_register_callback(int key,
4122 	int (*prehandler)(caddr_t, uint_t, uint_t, void *),
4123 	int (*posthandler)(caddr_t, uint_t, uint_t, void *, pfn_t),
4124 	int (*errhandler)(caddr_t, uint_t, uint_t, void *),
4125 	int capture_cpus)
4126 {
4127 	id_t id;
4128 
4129 	/*
4130 	 * Search the table for a pre-existing callback associated with
4131 	 * the identifier "key".  If one exists, we re-use that entry in
4132 	 * the table for this instance, otherwise we assign the next
4133 	 * available table slot.
4134 	 */
4135 	for (id = 0; id < sfmmu_max_cb_id; id++) {
4136 		if (sfmmu_cb_table[id].key == key)
4137 			break;
4138 	}
4139 
4140 	if (id == sfmmu_max_cb_id) {
4141 		id = sfmmu_cb_nextid++;
4142 		if (id >= sfmmu_max_cb_id)
4143 			panic("hat_register_callback: out of callback IDs");
4144 	}
4145 
4146 	ASSERT(prehandler != NULL || posthandler != NULL);
4147 
4148 	sfmmu_cb_table[id].key = key;
4149 	sfmmu_cb_table[id].prehandler = prehandler;
4150 	sfmmu_cb_table[id].posthandler = posthandler;
4151 	sfmmu_cb_table[id].errhandler = errhandler;
4152 	sfmmu_cb_table[id].capture_cpus = capture_cpus;
4153 
4154 	return (id);
4155 }
4156 
4157 #define	HAC_COOKIE_NONE	(void *)-1
4158 
4159 /*
4160  * Add relocation callbacks to the specified addr/len which will be called
4161  * when relocating the associated page. See the description of pre and
4162  * posthandler above for more details.
4163  *
4164  * If HAC_PAGELOCK is included in flags, the underlying memory page is
4165  * locked internally so the caller must be able to deal with the callback
4166  * running even before this function has returned.  If HAC_PAGELOCK is not
4167  * set, it is assumed that the underlying memory pages are locked.
4168  *
4169  * Since the caller must track the individual page boundaries anyway,
4170  * we only allow a callback to be added to a single page (large
4171  * or small).  Thus [addr, addr + len) MUST be contained within a single
4172  * page.
4173  *
4174  * Registering multiple callbacks on the same [addr, addr+len) is supported,
4175  * _provided_that_ a unique parameter is specified for each callback.
4176  * If multiple callbacks are registered on the same range the callback will
4177  * be invoked with each unique parameter. Registering the same callback with
4178  * the same argument more than once will result in corrupted kernel state.
4179  *
4180  * Returns the pfn of the underlying kernel page in *rpfn
4181  * on success, or PFN_INVALID on failure.
4182  *
4183  * cookiep (if passed) provides storage space for an opaque cookie
4184  * to return later to hat_delete_callback(). This cookie makes the callback
4185  * deletion significantly quicker by avoiding a potentially lengthy hash
4186  * search.
4187  *
4188  * Returns values:
4189  *    0:      success
4190  *    ENOMEM: memory allocation failure (e.g. flags was passed as HAC_NOSLEEP)
4191  *    EINVAL: callback ID is not valid
4192  *    ENXIO:  ["vaddr", "vaddr" + len) is not mapped in the kernel's address
4193  *            space
4194  *    ERANGE: ["vaddr", "vaddr" + len) crosses a page boundary
4195  */
4196 int
4197 hat_add_callback(id_t callback_id, caddr_t vaddr, uint_t len, uint_t flags,
4198 	void *pvt, pfn_t *rpfn, void **cookiep)
4199 {
4200 	struct 		hmehash_bucket *hmebp;
4201 	hmeblk_tag 	hblktag;
4202 	struct hme_blk	*hmeblkp;
4203 	int 		hmeshift, hashno;
4204 	caddr_t 	saddr, eaddr, baseaddr;
4205 	struct pa_hment *pahmep;
4206 	struct sf_hment *sfhmep, *osfhmep;
4207 	kmutex_t	*pml;
4208 	tte_t   	tte;
4209 	page_t		*pp;
4210 	vnode_t		*vp;
4211 	u_offset_t	off;
4212 	pfn_t		pfn;
4213 	int		kmflags = (flags & HAC_SLEEP)? KM_SLEEP : KM_NOSLEEP;
4214 	int		locked = 0;
4215 
4216 	/*
4217 	 * For KPM mappings, just return the physical address since we
4218 	 * don't need to register any callbacks.
4219 	 */
4220 	if (IS_KPM_ADDR(vaddr)) {
4221 		uint64_t paddr;
4222 		SFMMU_KPM_VTOP(vaddr, paddr);
4223 		*rpfn = btop(paddr);
4224 		if (cookiep != NULL)
4225 			*cookiep = HAC_COOKIE_NONE;
4226 		return (0);
4227 	}
4228 
4229 	if (callback_id < (id_t)0 || callback_id >= sfmmu_cb_nextid) {
4230 		*rpfn = PFN_INVALID;
4231 		return (EINVAL);
4232 	}
4233 
4234 	if ((pahmep = kmem_cache_alloc(pa_hment_cache, kmflags)) == NULL) {
4235 		*rpfn = PFN_INVALID;
4236 		return (ENOMEM);
4237 	}
4238 
4239 	sfhmep = &pahmep->sfment;
4240 
4241 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4242 	eaddr = saddr + len;
4243 
4244 rehash:
4245 	/* Find the mapping(s) for this page */
4246 	for (hashno = TTE64K, hmeblkp = NULL;
4247 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4248 	    hashno++) {
4249 		hmeshift = HME_HASH_SHIFT(hashno);
4250 		hblktag.htag_id = ksfmmup;
4251 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4252 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4253 		hblktag.htag_rehash = hashno;
4254 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4255 
4256 		SFMMU_HASH_LOCK(hmebp);
4257 
4258 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4259 
4260 		if (hmeblkp == NULL)
4261 			SFMMU_HASH_UNLOCK(hmebp);
4262 	}
4263 
4264 	if (hmeblkp == NULL) {
4265 		kmem_cache_free(pa_hment_cache, pahmep);
4266 		*rpfn = PFN_INVALID;
4267 		return (ENXIO);
4268 	}
4269 
4270 	ASSERT(!hmeblkp->hblk_shared);
4271 
4272 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4273 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4274 
4275 	if (!TTE_IS_VALID(&tte)) {
4276 		SFMMU_HASH_UNLOCK(hmebp);
4277 		kmem_cache_free(pa_hment_cache, pahmep);
4278 		*rpfn = PFN_INVALID;
4279 		return (ENXIO);
4280 	}
4281 
4282 	/*
4283 	 * Make sure the boundaries for the callback fall within this
4284 	 * single mapping.
4285 	 */
4286 	baseaddr = (caddr_t)get_hblk_base(hmeblkp);
4287 	ASSERT(saddr >= baseaddr);
4288 	if (eaddr > saddr + TTEBYTES(TTE_CSZ(&tte))) {
4289 		SFMMU_HASH_UNLOCK(hmebp);
4290 		kmem_cache_free(pa_hment_cache, pahmep);
4291 		*rpfn = PFN_INVALID;
4292 		return (ERANGE);
4293 	}
4294 
4295 	pfn = sfmmu_ttetopfn(&tte, vaddr);
4296 
4297 	/*
4298 	 * The pfn may not have a page_t underneath in which case we
4299 	 * just return it. This can happen if we are doing I/O to a
4300 	 * static portion of the kernel's address space, for instance.
4301 	 */
4302 	pp = osfhmep->hme_page;
4303 	if (pp == NULL) {
4304 		SFMMU_HASH_UNLOCK(hmebp);
4305 		kmem_cache_free(pa_hment_cache, pahmep);
4306 		*rpfn = pfn;
4307 		if (cookiep)
4308 			*cookiep = HAC_COOKIE_NONE;
4309 		return (0);
4310 	}
4311 	ASSERT(pp == PP_PAGEROOT(pp));
4312 
4313 	vp = pp->p_vnode;
4314 	off = pp->p_offset;
4315 
4316 	pml = sfmmu_mlist_enter(pp);
4317 
4318 	if (flags & HAC_PAGELOCK) {
4319 		if (!page_trylock(pp, SE_SHARED)) {
4320 			/*
4321 			 * Somebody is holding SE_EXCL lock. Might
4322 			 * even be hat_page_relocate(). Drop all
4323 			 * our locks, lookup the page in &kvp, and
4324 			 * retry. If it doesn't exist in &kvp and &zvp,
4325 			 * then we must be dealing with a kernel mapped
4326 			 * page which doesn't actually belong to
4327 			 * segkmem so we punt.
4328 			 */
4329 			sfmmu_mlist_exit(pml);
4330 			SFMMU_HASH_UNLOCK(hmebp);
4331 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4332 
4333 			/* check zvp before giving up */
4334 			if (pp == NULL)
4335 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4336 				    SE_SHARED);
4337 
4338 			/* Okay, we didn't find it, give up */
4339 			if (pp == NULL) {
4340 				kmem_cache_free(pa_hment_cache, pahmep);
4341 				*rpfn = pfn;
4342 				if (cookiep)
4343 					*cookiep = HAC_COOKIE_NONE;
4344 				return (0);
4345 			}
4346 			page_unlock(pp);
4347 			goto rehash;
4348 		}
4349 		locked = 1;
4350 	}
4351 
4352 	if (!PAGE_LOCKED(pp) && !panicstr)
4353 		panic("hat_add_callback: page 0x%p not locked", (void *)pp);
4354 
4355 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4356 	    pp->p_offset != off) {
4357 		/*
4358 		 * The page moved before we got our hands on it.  Drop
4359 		 * all the locks and try again.
4360 		 */
4361 		ASSERT((flags & HAC_PAGELOCK) != 0);
4362 		sfmmu_mlist_exit(pml);
4363 		SFMMU_HASH_UNLOCK(hmebp);
4364 		page_unlock(pp);
4365 		locked = 0;
4366 		goto rehash;
4367 	}
4368 
4369 	if (!VN_ISKAS(vp)) {
4370 		/*
4371 		 * This is not a segkmem page but another page which
4372 		 * has been kernel mapped. It had better have at least
4373 		 * a share lock on it. Return the pfn.
4374 		 */
4375 		sfmmu_mlist_exit(pml);
4376 		SFMMU_HASH_UNLOCK(hmebp);
4377 		if (locked)
4378 			page_unlock(pp);
4379 		kmem_cache_free(pa_hment_cache, pahmep);
4380 		ASSERT(PAGE_LOCKED(pp));
4381 		*rpfn = pfn;
4382 		if (cookiep)
4383 			*cookiep = HAC_COOKIE_NONE;
4384 		return (0);
4385 	}
4386 
4387 	/*
4388 	 * Setup this pa_hment and link its embedded dummy sf_hment into
4389 	 * the mapping list.
4390 	 */
4391 	pp->p_share++;
4392 	pahmep->cb_id = callback_id;
4393 	pahmep->addr = vaddr;
4394 	pahmep->len = len;
4395 	pahmep->refcnt = 1;
4396 	pahmep->flags = 0;
4397 	pahmep->pvt = pvt;
4398 
4399 	sfhmep->hme_tte.ll = 0;
4400 	sfhmep->hme_data = pahmep;
4401 	sfhmep->hme_prev = osfhmep;
4402 	sfhmep->hme_next = osfhmep->hme_next;
4403 
4404 	if (osfhmep->hme_next)
4405 		osfhmep->hme_next->hme_prev = sfhmep;
4406 
4407 	osfhmep->hme_next = sfhmep;
4408 
4409 	sfmmu_mlist_exit(pml);
4410 	SFMMU_HASH_UNLOCK(hmebp);
4411 
4412 	if (locked)
4413 		page_unlock(pp);
4414 
4415 	*rpfn = pfn;
4416 	if (cookiep)
4417 		*cookiep = (void *)pahmep;
4418 
4419 	return (0);
4420 }
4421 
4422 /*
4423  * Remove the relocation callbacks from the specified addr/len.
4424  */
4425 void
4426 hat_delete_callback(caddr_t vaddr, uint_t len, void *pvt, uint_t flags,
4427 	void *cookie)
4428 {
4429 	struct		hmehash_bucket *hmebp;
4430 	hmeblk_tag	hblktag;
4431 	struct hme_blk	*hmeblkp;
4432 	int		hmeshift, hashno;
4433 	caddr_t		saddr;
4434 	struct pa_hment	*pahmep;
4435 	struct sf_hment	*sfhmep, *osfhmep;
4436 	kmutex_t	*pml;
4437 	tte_t		tte;
4438 	page_t		*pp;
4439 	vnode_t		*vp;
4440 	u_offset_t	off;
4441 	int		locked = 0;
4442 
4443 	/*
4444 	 * If the cookie is HAC_COOKIE_NONE then there is no pa_hment to
4445 	 * remove so just return.
4446 	 */
4447 	if (cookie == HAC_COOKIE_NONE || IS_KPM_ADDR(vaddr))
4448 		return;
4449 
4450 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4451 
4452 rehash:
4453 	/* Find the mapping(s) for this page */
4454 	for (hashno = TTE64K, hmeblkp = NULL;
4455 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4456 	    hashno++) {
4457 		hmeshift = HME_HASH_SHIFT(hashno);
4458 		hblktag.htag_id = ksfmmup;
4459 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4460 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4461 		hblktag.htag_rehash = hashno;
4462 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4463 
4464 		SFMMU_HASH_LOCK(hmebp);
4465 
4466 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4467 
4468 		if (hmeblkp == NULL)
4469 			SFMMU_HASH_UNLOCK(hmebp);
4470 	}
4471 
4472 	if (hmeblkp == NULL)
4473 		return;
4474 
4475 	ASSERT(!hmeblkp->hblk_shared);
4476 
4477 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4478 
4479 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4480 	if (!TTE_IS_VALID(&tte)) {
4481 		SFMMU_HASH_UNLOCK(hmebp);
4482 		return;
4483 	}
4484 
4485 	pp = osfhmep->hme_page;
4486 	if (pp == NULL) {
4487 		SFMMU_HASH_UNLOCK(hmebp);
4488 		ASSERT(cookie == NULL);
4489 		return;
4490 	}
4491 
4492 	vp = pp->p_vnode;
4493 	off = pp->p_offset;
4494 
4495 	pml = sfmmu_mlist_enter(pp);
4496 
4497 	if (flags & HAC_PAGELOCK) {
4498 		if (!page_trylock(pp, SE_SHARED)) {
4499 			/*
4500 			 * Somebody is holding SE_EXCL lock. Might
4501 			 * even be hat_page_relocate(). Drop all
4502 			 * our locks, lookup the page in &kvp, and
4503 			 * retry. If it doesn't exist in &kvp and &zvp,
4504 			 * then we must be dealing with a kernel mapped
4505 			 * page which doesn't actually belong to
4506 			 * segkmem so we punt.
4507 			 */
4508 			sfmmu_mlist_exit(pml);
4509 			SFMMU_HASH_UNLOCK(hmebp);
4510 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4511 			/* check zvp before giving up */
4512 			if (pp == NULL)
4513 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4514 				    SE_SHARED);
4515 
4516 			if (pp == NULL) {
4517 				ASSERT(cookie == NULL);
4518 				return;
4519 			}
4520 			page_unlock(pp);
4521 			goto rehash;
4522 		}
4523 		locked = 1;
4524 	}
4525 
4526 	ASSERT(PAGE_LOCKED(pp));
4527 
4528 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4529 	    pp->p_offset != off) {
4530 		/*
4531 		 * The page moved before we got our hands on it.  Drop
4532 		 * all the locks and try again.
4533 		 */
4534 		ASSERT((flags & HAC_PAGELOCK) != 0);
4535 		sfmmu_mlist_exit(pml);
4536 		SFMMU_HASH_UNLOCK(hmebp);
4537 		page_unlock(pp);
4538 		locked = 0;
4539 		goto rehash;
4540 	}
4541 
4542 	if (!VN_ISKAS(vp)) {
4543 		/*
4544 		 * This is not a segkmem page but another page which
4545 		 * has been kernel mapped.
4546 		 */
4547 		sfmmu_mlist_exit(pml);
4548 		SFMMU_HASH_UNLOCK(hmebp);
4549 		if (locked)
4550 			page_unlock(pp);
4551 		ASSERT(cookie == NULL);
4552 		return;
4553 	}
4554 
4555 	if (cookie != NULL) {
4556 		pahmep = (struct pa_hment *)cookie;
4557 		sfhmep = &pahmep->sfment;
4558 	} else {
4559 		for (sfhmep = pp->p_mapping; sfhmep != NULL;
4560 		    sfhmep = sfhmep->hme_next) {
4561 
4562 			/*
4563 			 * skip va<->pa mappings
4564 			 */
4565 			if (!IS_PAHME(sfhmep))
4566 				continue;
4567 
4568 			pahmep = sfhmep->hme_data;
4569 			ASSERT(pahmep != NULL);
4570 
4571 			/*
4572 			 * if pa_hment matches, remove it
4573 			 */
4574 			if ((pahmep->pvt == pvt) &&
4575 			    (pahmep->addr == vaddr) &&
4576 			    (pahmep->len == len)) {
4577 				break;
4578 			}
4579 		}
4580 	}
4581 
4582 	if (sfhmep == NULL) {
4583 		if (!panicstr) {
4584 			panic("hat_delete_callback: pa_hment not found, pp %p",
4585 			    (void *)pp);
4586 		}
4587 		return;
4588 	}
4589 
4590 	/*
4591 	 * Note: at this point a valid kernel mapping must still be
4592 	 * present on this page.
4593 	 */
4594 	pp->p_share--;
4595 	if (pp->p_share <= 0)
4596 		panic("hat_delete_callback: zero p_share");
4597 
4598 	if (--pahmep->refcnt == 0) {
4599 		if (pahmep->flags != 0)
4600 			panic("hat_delete_callback: pa_hment is busy");
4601 
4602 		/*
4603 		 * Remove sfhmep from the mapping list for the page.
4604 		 */
4605 		if (sfhmep->hme_prev) {
4606 			sfhmep->hme_prev->hme_next = sfhmep->hme_next;
4607 		} else {
4608 			pp->p_mapping = sfhmep->hme_next;
4609 		}
4610 
4611 		if (sfhmep->hme_next)
4612 			sfhmep->hme_next->hme_prev = sfhmep->hme_prev;
4613 
4614 		sfmmu_mlist_exit(pml);
4615 		SFMMU_HASH_UNLOCK(hmebp);
4616 
4617 		if (locked)
4618 			page_unlock(pp);
4619 
4620 		kmem_cache_free(pa_hment_cache, pahmep);
4621 		return;
4622 	}
4623 
4624 	sfmmu_mlist_exit(pml);
4625 	SFMMU_HASH_UNLOCK(hmebp);
4626 	if (locked)
4627 		page_unlock(pp);
4628 }
4629 
4630 /*
4631  * hat_probe returns 1 if the translation for the address 'addr' is
4632  * loaded, zero otherwise.
4633  *
4634  * hat_probe should be used only for advisorary purposes because it may
4635  * occasionally return the wrong value. The implementation must guarantee that
4636  * returning the wrong value is a very rare event. hat_probe is used
4637  * to implement optimizations in the segment drivers.
4638  *
4639  */
4640 int
4641 hat_probe(struct hat *sfmmup, caddr_t addr)
4642 {
4643 	pfn_t pfn;
4644 	tte_t tte;
4645 
4646 	ASSERT(sfmmup != NULL);
4647 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4648 
4649 	ASSERT((sfmmup == ksfmmup) ||
4650 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
4651 
4652 	if (sfmmup == ksfmmup) {
4653 		while ((pfn = sfmmu_vatopfn(addr, sfmmup, &tte))
4654 		    == PFN_SUSPENDED) {
4655 			sfmmu_vatopfn_suspended(addr, sfmmup, &tte);
4656 		}
4657 	} else {
4658 		pfn = sfmmu_uvatopfn(addr, sfmmup, NULL);
4659 	}
4660 
4661 	if (pfn != PFN_INVALID)
4662 		return (1);
4663 	else
4664 		return (0);
4665 }
4666 
4667 ssize_t
4668 hat_getpagesize(struct hat *sfmmup, caddr_t addr)
4669 {
4670 	tte_t tte;
4671 
4672 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4673 
4674 	if (sfmmup == ksfmmup) {
4675 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4676 			return (-1);
4677 		}
4678 	} else {
4679 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4680 			return (-1);
4681 		}
4682 	}
4683 
4684 	ASSERT(TTE_IS_VALID(&tte));
4685 	return (TTEBYTES(TTE_CSZ(&tte)));
4686 }
4687 
4688 uint_t
4689 hat_getattr(struct hat *sfmmup, caddr_t addr, uint_t *attr)
4690 {
4691 	tte_t tte;
4692 
4693 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4694 
4695 	if (sfmmup == ksfmmup) {
4696 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4697 			tte.ll = 0;
4698 		}
4699 	} else {
4700 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4701 			tte.ll = 0;
4702 		}
4703 	}
4704 	if (TTE_IS_VALID(&tte)) {
4705 		*attr = sfmmu_ptov_attr(&tte);
4706 		return (0);
4707 	}
4708 	*attr = 0;
4709 	return ((uint_t)0xffffffff);
4710 }
4711 
4712 /*
4713  * Enables more attributes on specified address range (ie. logical OR)
4714  */
4715 void
4716 hat_setattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4717 {
4718 	if (hat->sfmmu_xhat_provider) {
4719 		XHAT_SETATTR(hat, addr, len, attr);
4720 		return;
4721 	} else {
4722 		/*
4723 		 * This must be a CPU HAT. If the address space has
4724 		 * XHATs attached, change attributes for all of them,
4725 		 * just in case
4726 		 */
4727 		ASSERT(hat->sfmmu_as != NULL);
4728 		if (hat->sfmmu_as->a_xhat != NULL)
4729 			xhat_setattr_all(hat->sfmmu_as, addr, len, attr);
4730 	}
4731 
4732 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_SETATTR);
4733 }
4734 
4735 /*
4736  * Assigns attributes to the specified address range.  All the attributes
4737  * are specified.
4738  */
4739 void
4740 hat_chgattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4741 {
4742 	if (hat->sfmmu_xhat_provider) {
4743 		XHAT_CHGATTR(hat, addr, len, attr);
4744 		return;
4745 	} else {
4746 		/*
4747 		 * This must be a CPU HAT. If the address space has
4748 		 * XHATs attached, change attributes for all of them,
4749 		 * just in case
4750 		 */
4751 		ASSERT(hat->sfmmu_as != NULL);
4752 		if (hat->sfmmu_as->a_xhat != NULL)
4753 			xhat_chgattr_all(hat->sfmmu_as, addr, len, attr);
4754 	}
4755 
4756 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CHGATTR);
4757 }
4758 
4759 /*
4760  * Remove attributes on the specified address range (ie. loginal NAND)
4761  */
4762 void
4763 hat_clrattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4764 {
4765 	if (hat->sfmmu_xhat_provider) {
4766 		XHAT_CLRATTR(hat, addr, len, attr);
4767 		return;
4768 	} else {
4769 		/*
4770 		 * This must be a CPU HAT. If the address space has
4771 		 * XHATs attached, change attributes for all of them,
4772 		 * just in case
4773 		 */
4774 		ASSERT(hat->sfmmu_as != NULL);
4775 		if (hat->sfmmu_as->a_xhat != NULL)
4776 			xhat_clrattr_all(hat->sfmmu_as, addr, len, attr);
4777 	}
4778 
4779 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CLRATTR);
4780 }
4781 
4782 /*
4783  * Change attributes on an address range to that specified by attr and mode.
4784  */
4785 static void
4786 sfmmu_chgattr(struct hat *sfmmup, caddr_t addr, size_t len, uint_t attr,
4787 	int mode)
4788 {
4789 	struct hmehash_bucket *hmebp;
4790 	hmeblk_tag hblktag;
4791 	int hmeshift, hashno = 1;
4792 	struct hme_blk *hmeblkp, *list = NULL;
4793 	caddr_t endaddr;
4794 	cpuset_t cpuset;
4795 	demap_range_t dmr;
4796 
4797 	CPUSET_ZERO(cpuset);
4798 
4799 	ASSERT((sfmmup == ksfmmup) ||
4800 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
4801 	ASSERT((len & MMU_PAGEOFFSET) == 0);
4802 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
4803 
4804 	if ((attr & PROT_USER) && (mode != SFMMU_CLRATTR) &&
4805 	    ((addr + len) > (caddr_t)USERLIMIT)) {
4806 		panic("user addr %p in kernel space",
4807 		    (void *)addr);
4808 	}
4809 
4810 	endaddr = addr + len;
4811 	hblktag.htag_id = sfmmup;
4812 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4813 	DEMAP_RANGE_INIT(sfmmup, &dmr);
4814 
4815 	while (addr < endaddr) {
4816 		hmeshift = HME_HASH_SHIFT(hashno);
4817 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
4818 		hblktag.htag_rehash = hashno;
4819 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
4820 
4821 		SFMMU_HASH_LOCK(hmebp);
4822 
4823 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
4824 		if (hmeblkp != NULL) {
4825 			ASSERT(!hmeblkp->hblk_shared);
4826 			/*
4827 			 * We've encountered a shadow hmeblk so skip the range
4828 			 * of the next smaller mapping size.
4829 			 */
4830 			if (hmeblkp->hblk_shw_bit) {
4831 				ASSERT(sfmmup != ksfmmup);
4832 				ASSERT(hashno > 1);
4833 				addr = (caddr_t)P2END((uintptr_t)addr,
4834 				    TTEBYTES(hashno - 1));
4835 			} else {
4836 				addr = sfmmu_hblk_chgattr(sfmmup,
4837 				    hmeblkp, addr, endaddr, &dmr, attr, mode);
4838 			}
4839 			SFMMU_HASH_UNLOCK(hmebp);
4840 			hashno = 1;
4841 			continue;
4842 		}
4843 		SFMMU_HASH_UNLOCK(hmebp);
4844 
4845 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
4846 			/*
4847 			 * We have traversed the whole list and rehashed
4848 			 * if necessary without finding the address to chgattr.
4849 			 * This is ok, so we increment the address by the
4850 			 * smallest hmeblk range for kernel mappings or for
4851 			 * user mappings with no large pages, and the largest
4852 			 * hmeblk range, to account for shadow hmeblks, for
4853 			 * user mappings with large pages and continue.
4854 			 */
4855 			if (sfmmup == ksfmmup)
4856 				addr = (caddr_t)P2END((uintptr_t)addr,
4857 				    TTEBYTES(1));
4858 			else
4859 				addr = (caddr_t)P2END((uintptr_t)addr,
4860 				    TTEBYTES(hashno));
4861 			hashno = 1;
4862 		} else {
4863 			hashno++;
4864 		}
4865 	}
4866 
4867 	sfmmu_hblks_list_purge(&list, 0);
4868 	DEMAP_RANGE_FLUSH(&dmr);
4869 	cpuset = sfmmup->sfmmu_cpusran;
4870 	xt_sync(cpuset);
4871 }
4872 
4873 /*
4874  * This function chgattr on a range of addresses in an hmeblk.  It returns the
4875  * next addres that needs to be chgattr.
4876  * It should be called with the hash lock held.
4877  * XXX It should be possible to optimize chgattr by not flushing every time but
4878  * on the other hand:
4879  * 1. do one flush crosscall.
4880  * 2. only flush if we are increasing permissions (make sure this will work)
4881  */
4882 static caddr_t
4883 sfmmu_hblk_chgattr(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
4884 	caddr_t endaddr, demap_range_t *dmrp, uint_t attr, int mode)
4885 {
4886 	tte_t tte, tteattr, tteflags, ttemod;
4887 	struct sf_hment *sfhmep;
4888 	int ttesz;
4889 	struct page *pp = NULL;
4890 	kmutex_t *pml, *pmtx;
4891 	int ret;
4892 	int use_demap_range;
4893 #if defined(SF_ERRATA_57)
4894 	int check_exec;
4895 #endif
4896 
4897 	ASSERT(in_hblk_range(hmeblkp, addr));
4898 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4899 	ASSERT(!hmeblkp->hblk_shared);
4900 
4901 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4902 	ttesz = get_hblk_ttesz(hmeblkp);
4903 
4904 	/*
4905 	 * Flush the current demap region if addresses have been
4906 	 * skipped or the page size doesn't match.
4907 	 */
4908 	use_demap_range = (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp));
4909 	if (use_demap_range) {
4910 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
4911 	} else {
4912 		DEMAP_RANGE_FLUSH(dmrp);
4913 	}
4914 
4915 	tteattr.ll = sfmmu_vtop_attr(attr, mode, &tteflags);
4916 #if defined(SF_ERRATA_57)
4917 	check_exec = (sfmmup != ksfmmup) &&
4918 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
4919 	    TTE_IS_EXECUTABLE(&tteattr);
4920 #endif
4921 	HBLKTOHME(sfhmep, hmeblkp, addr);
4922 	while (addr < endaddr) {
4923 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
4924 		if (TTE_IS_VALID(&tte)) {
4925 			if ((tte.ll & tteflags.ll) == tteattr.ll) {
4926 				/*
4927 				 * if the new attr is the same as old
4928 				 * continue
4929 				 */
4930 				goto next_addr;
4931 			}
4932 			if (!TTE_IS_WRITABLE(&tteattr)) {
4933 				/*
4934 				 * make sure we clear hw modify bit if we
4935 				 * removing write protections
4936 				 */
4937 				tteflags.tte_intlo |= TTE_HWWR_INT;
4938 			}
4939 
4940 			pml = NULL;
4941 			pp = sfhmep->hme_page;
4942 			if (pp) {
4943 				pml = sfmmu_mlist_enter(pp);
4944 			}
4945 
4946 			if (pp != sfhmep->hme_page) {
4947 				/*
4948 				 * tte must have been unloaded.
4949 				 */
4950 				ASSERT(pml);
4951 				sfmmu_mlist_exit(pml);
4952 				continue;
4953 			}
4954 
4955 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
4956 
4957 			ttemod = tte;
4958 			ttemod.ll = (ttemod.ll & ~tteflags.ll) | tteattr.ll;
4959 			ASSERT(TTE_TO_TTEPFN(&ttemod) == TTE_TO_TTEPFN(&tte));
4960 
4961 #if defined(SF_ERRATA_57)
4962 			if (check_exec && addr < errata57_limit)
4963 				ttemod.tte_exec_perm = 0;
4964 #endif
4965 			ret = sfmmu_modifytte_try(&tte, &ttemod,
4966 			    &sfhmep->hme_tte);
4967 
4968 			if (ret < 0) {
4969 				/* tte changed underneath us */
4970 				if (pml) {
4971 					sfmmu_mlist_exit(pml);
4972 				}
4973 				continue;
4974 			}
4975 
4976 			if (tteflags.tte_intlo & TTE_HWWR_INT) {
4977 				/*
4978 				 * need to sync if we are clearing modify bit.
4979 				 */
4980 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
4981 			}
4982 
4983 			if (pp && PP_ISRO(pp)) {
4984 				if (tteattr.tte_intlo & TTE_WRPRM_INT) {
4985 					pmtx = sfmmu_page_enter(pp);
4986 					PP_CLRRO(pp);
4987 					sfmmu_page_exit(pmtx);
4988 				}
4989 			}
4990 
4991 			if (ret > 0 && use_demap_range) {
4992 				DEMAP_RANGE_MARKPG(dmrp, addr);
4993 			} else if (ret > 0) {
4994 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
4995 			}
4996 
4997 			if (pml) {
4998 				sfmmu_mlist_exit(pml);
4999 			}
5000 		}
5001 next_addr:
5002 		addr += TTEBYTES(ttesz);
5003 		sfhmep++;
5004 		DEMAP_RANGE_NEXTPG(dmrp);
5005 	}
5006 	return (addr);
5007 }
5008 
5009 /*
5010  * This routine converts virtual attributes to physical ones.  It will
5011  * update the tteflags field with the tte mask corresponding to the attributes
5012  * affected and it returns the new attributes.  It will also clear the modify
5013  * bit if we are taking away write permission.  This is necessary since the
5014  * modify bit is the hardware permission bit and we need to clear it in order
5015  * to detect write faults.
5016  */
5017 static uint64_t
5018 sfmmu_vtop_attr(uint_t attr, int mode, tte_t *ttemaskp)
5019 {
5020 	tte_t ttevalue;
5021 
5022 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
5023 
5024 	switch (mode) {
5025 	case SFMMU_CHGATTR:
5026 		/* all attributes specified */
5027 		ttevalue.tte_inthi = MAKE_TTEATTR_INTHI(attr);
5028 		ttevalue.tte_intlo = MAKE_TTEATTR_INTLO(attr);
5029 		ttemaskp->tte_inthi = TTEINTHI_ATTR;
5030 		ttemaskp->tte_intlo = TTEINTLO_ATTR;
5031 		break;
5032 	case SFMMU_SETATTR:
5033 		ASSERT(!(attr & ~HAT_PROT_MASK));
5034 		ttemaskp->ll = 0;
5035 		ttevalue.ll = 0;
5036 		/*
5037 		 * a valid tte implies exec and read for sfmmu
5038 		 * so no need to do anything about them.
5039 		 * since priviledged access implies user access
5040 		 * PROT_USER doesn't make sense either.
5041 		 */
5042 		if (attr & PROT_WRITE) {
5043 			ttemaskp->tte_intlo |= TTE_WRPRM_INT;
5044 			ttevalue.tte_intlo |= TTE_WRPRM_INT;
5045 		}
5046 		break;
5047 	case SFMMU_CLRATTR:
5048 		/* attributes will be nand with current ones */
5049 		if (attr & ~(PROT_WRITE | PROT_USER)) {
5050 			panic("sfmmu: attr %x not supported", attr);
5051 		}
5052 		ttemaskp->ll = 0;
5053 		ttevalue.ll = 0;
5054 		if (attr & PROT_WRITE) {
5055 			/* clear both writable and modify bit */
5056 			ttemaskp->tte_intlo |= TTE_WRPRM_INT | TTE_HWWR_INT;
5057 		}
5058 		if (attr & PROT_USER) {
5059 			ttemaskp->tte_intlo |= TTE_PRIV_INT;
5060 			ttevalue.tte_intlo |= TTE_PRIV_INT;
5061 		}
5062 		break;
5063 	default:
5064 		panic("sfmmu_vtop_attr: bad mode %x", mode);
5065 	}
5066 	ASSERT(TTE_TO_TTEPFN(&ttevalue) == 0);
5067 	return (ttevalue.ll);
5068 }
5069 
5070 static uint_t
5071 sfmmu_ptov_attr(tte_t *ttep)
5072 {
5073 	uint_t attr;
5074 
5075 	ASSERT(TTE_IS_VALID(ttep));
5076 
5077 	attr = PROT_READ;
5078 
5079 	if (TTE_IS_WRITABLE(ttep)) {
5080 		attr |= PROT_WRITE;
5081 	}
5082 	if (TTE_IS_EXECUTABLE(ttep)) {
5083 		attr |= PROT_EXEC;
5084 	}
5085 	if (!TTE_IS_PRIVILEGED(ttep)) {
5086 		attr |= PROT_USER;
5087 	}
5088 	if (TTE_IS_NFO(ttep)) {
5089 		attr |= HAT_NOFAULT;
5090 	}
5091 	if (TTE_IS_NOSYNC(ttep)) {
5092 		attr |= HAT_NOSYNC;
5093 	}
5094 	if (TTE_IS_SIDEFFECT(ttep)) {
5095 		attr |= SFMMU_SIDEFFECT;
5096 	}
5097 	if (!TTE_IS_VCACHEABLE(ttep)) {
5098 		attr |= SFMMU_UNCACHEVTTE;
5099 	}
5100 	if (!TTE_IS_PCACHEABLE(ttep)) {
5101 		attr |= SFMMU_UNCACHEPTTE;
5102 	}
5103 	return (attr);
5104 }
5105 
5106 /*
5107  * hat_chgprot is a deprecated hat call.  New segment drivers
5108  * should store all attributes and use hat_*attr calls.
5109  *
5110  * Change the protections in the virtual address range
5111  * given to the specified virtual protection.  If vprot is ~PROT_WRITE,
5112  * then remove write permission, leaving the other
5113  * permissions unchanged.  If vprot is ~PROT_USER, remove user permissions.
5114  *
5115  */
5116 void
5117 hat_chgprot(struct hat *sfmmup, caddr_t addr, size_t len, uint_t vprot)
5118 {
5119 	struct hmehash_bucket *hmebp;
5120 	hmeblk_tag hblktag;
5121 	int hmeshift, hashno = 1;
5122 	struct hme_blk *hmeblkp, *list = NULL;
5123 	caddr_t endaddr;
5124 	cpuset_t cpuset;
5125 	demap_range_t dmr;
5126 
5127 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5128 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
5129 
5130 	if (sfmmup->sfmmu_xhat_provider) {
5131 		XHAT_CHGPROT(sfmmup, addr, len, vprot);
5132 		return;
5133 	} else {
5134 		/*
5135 		 * This must be a CPU HAT. If the address space has
5136 		 * XHATs attached, change attributes for all of them,
5137 		 * just in case
5138 		 */
5139 		ASSERT(sfmmup->sfmmu_as != NULL);
5140 		if (sfmmup->sfmmu_as->a_xhat != NULL)
5141 			xhat_chgprot_all(sfmmup->sfmmu_as, addr, len, vprot);
5142 	}
5143 
5144 	CPUSET_ZERO(cpuset);
5145 
5146 	if ((vprot != (uint_t)~PROT_WRITE) && (vprot & PROT_USER) &&
5147 	    ((addr + len) > (caddr_t)USERLIMIT)) {
5148 		panic("user addr %p vprot %x in kernel space",
5149 		    (void *)addr, vprot);
5150 	}
5151 	endaddr = addr + len;
5152 	hblktag.htag_id = sfmmup;
5153 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5154 	DEMAP_RANGE_INIT(sfmmup, &dmr);
5155 
5156 	while (addr < endaddr) {
5157 		hmeshift = HME_HASH_SHIFT(hashno);
5158 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5159 		hblktag.htag_rehash = hashno;
5160 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5161 
5162 		SFMMU_HASH_LOCK(hmebp);
5163 
5164 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
5165 		if (hmeblkp != NULL) {
5166 			ASSERT(!hmeblkp->hblk_shared);
5167 			/*
5168 			 * We've encountered a shadow hmeblk so skip the range
5169 			 * of the next smaller mapping size.
5170 			 */
5171 			if (hmeblkp->hblk_shw_bit) {
5172 				ASSERT(sfmmup != ksfmmup);
5173 				ASSERT(hashno > 1);
5174 				addr = (caddr_t)P2END((uintptr_t)addr,
5175 				    TTEBYTES(hashno - 1));
5176 			} else {
5177 				addr = sfmmu_hblk_chgprot(sfmmup, hmeblkp,
5178 				    addr, endaddr, &dmr, vprot);
5179 			}
5180 			SFMMU_HASH_UNLOCK(hmebp);
5181 			hashno = 1;
5182 			continue;
5183 		}
5184 		SFMMU_HASH_UNLOCK(hmebp);
5185 
5186 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
5187 			/*
5188 			 * We have traversed the whole list and rehashed
5189 			 * if necessary without finding the address to chgprot.
5190 			 * This is ok so we increment the address by the
5191 			 * smallest hmeblk range for kernel mappings and the
5192 			 * largest hmeblk range, to account for shadow hmeblks,
5193 			 * for user mappings and continue.
5194 			 */
5195 			if (sfmmup == ksfmmup)
5196 				addr = (caddr_t)P2END((uintptr_t)addr,
5197 				    TTEBYTES(1));
5198 			else
5199 				addr = (caddr_t)P2END((uintptr_t)addr,
5200 				    TTEBYTES(hashno));
5201 			hashno = 1;
5202 		} else {
5203 			hashno++;
5204 		}
5205 	}
5206 
5207 	sfmmu_hblks_list_purge(&list, 0);
5208 	DEMAP_RANGE_FLUSH(&dmr);
5209 	cpuset = sfmmup->sfmmu_cpusran;
5210 	xt_sync(cpuset);
5211 }
5212 
5213 /*
5214  * This function chgprots a range of addresses in an hmeblk.  It returns the
5215  * next addres that needs to be chgprot.
5216  * It should be called with the hash lock held.
5217  * XXX It shold be possible to optimize chgprot by not flushing every time but
5218  * on the other hand:
5219  * 1. do one flush crosscall.
5220  * 2. only flush if we are increasing permissions (make sure this will work)
5221  */
5222 static caddr_t
5223 sfmmu_hblk_chgprot(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5224 	caddr_t endaddr, demap_range_t *dmrp, uint_t vprot)
5225 {
5226 	uint_t pprot;
5227 	tte_t tte, ttemod;
5228 	struct sf_hment *sfhmep;
5229 	uint_t tteflags;
5230 	int ttesz;
5231 	struct page *pp = NULL;
5232 	kmutex_t *pml, *pmtx;
5233 	int ret;
5234 	int use_demap_range;
5235 #if defined(SF_ERRATA_57)
5236 	int check_exec;
5237 #endif
5238 
5239 	ASSERT(in_hblk_range(hmeblkp, addr));
5240 	ASSERT(hmeblkp->hblk_shw_bit == 0);
5241 	ASSERT(!hmeblkp->hblk_shared);
5242 
5243 #ifdef DEBUG
5244 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5245 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5246 		panic("sfmmu_hblk_chgprot: partial chgprot of large page");
5247 	}
5248 #endif /* DEBUG */
5249 
5250 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5251 	ttesz = get_hblk_ttesz(hmeblkp);
5252 
5253 	pprot = sfmmu_vtop_prot(vprot, &tteflags);
5254 #if defined(SF_ERRATA_57)
5255 	check_exec = (sfmmup != ksfmmup) &&
5256 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
5257 	    ((vprot & PROT_EXEC) == PROT_EXEC);
5258 #endif
5259 	HBLKTOHME(sfhmep, hmeblkp, addr);
5260 
5261 	/*
5262 	 * Flush the current demap region if addresses have been
5263 	 * skipped or the page size doesn't match.
5264 	 */
5265 	use_demap_range = (TTEBYTES(ttesz) == MMU_PAGESIZE);
5266 	if (use_demap_range) {
5267 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5268 	} else {
5269 		DEMAP_RANGE_FLUSH(dmrp);
5270 	}
5271 
5272 	while (addr < endaddr) {
5273 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5274 		if (TTE_IS_VALID(&tte)) {
5275 			if (TTE_GET_LOFLAGS(&tte, tteflags) == pprot) {
5276 				/*
5277 				 * if the new protection is the same as old
5278 				 * continue
5279 				 */
5280 				goto next_addr;
5281 			}
5282 			pml = NULL;
5283 			pp = sfhmep->hme_page;
5284 			if (pp) {
5285 				pml = sfmmu_mlist_enter(pp);
5286 			}
5287 			if (pp != sfhmep->hme_page) {
5288 				/*
5289 				 * tte most have been unloaded
5290 				 * underneath us.  Recheck
5291 				 */
5292 				ASSERT(pml);
5293 				sfmmu_mlist_exit(pml);
5294 				continue;
5295 			}
5296 
5297 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5298 
5299 			ttemod = tte;
5300 			TTE_SET_LOFLAGS(&ttemod, tteflags, pprot);
5301 #if defined(SF_ERRATA_57)
5302 			if (check_exec && addr < errata57_limit)
5303 				ttemod.tte_exec_perm = 0;
5304 #endif
5305 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5306 			    &sfhmep->hme_tte);
5307 
5308 			if (ret < 0) {
5309 				/* tte changed underneath us */
5310 				if (pml) {
5311 					sfmmu_mlist_exit(pml);
5312 				}
5313 				continue;
5314 			}
5315 
5316 			if (tteflags & TTE_HWWR_INT) {
5317 				/*
5318 				 * need to sync if we are clearing modify bit.
5319 				 */
5320 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5321 			}
5322 
5323 			if (pp && PP_ISRO(pp)) {
5324 				if (pprot & TTE_WRPRM_INT) {
5325 					pmtx = sfmmu_page_enter(pp);
5326 					PP_CLRRO(pp);
5327 					sfmmu_page_exit(pmtx);
5328 				}
5329 			}
5330 
5331 			if (ret > 0 && use_demap_range) {
5332 				DEMAP_RANGE_MARKPG(dmrp, addr);
5333 			} else if (ret > 0) {
5334 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5335 			}
5336 
5337 			if (pml) {
5338 				sfmmu_mlist_exit(pml);
5339 			}
5340 		}
5341 next_addr:
5342 		addr += TTEBYTES(ttesz);
5343 		sfhmep++;
5344 		DEMAP_RANGE_NEXTPG(dmrp);
5345 	}
5346 	return (addr);
5347 }
5348 
5349 /*
5350  * This routine is deprecated and should only be used by hat_chgprot.
5351  * The correct routine is sfmmu_vtop_attr.
5352  * This routine converts virtual page protections to physical ones.  It will
5353  * update the tteflags field with the tte mask corresponding to the protections
5354  * affected and it returns the new protections.  It will also clear the modify
5355  * bit if we are taking away write permission.  This is necessary since the
5356  * modify bit is the hardware permission bit and we need to clear it in order
5357  * to detect write faults.
5358  * It accepts the following special protections:
5359  * ~PROT_WRITE = remove write permissions.
5360  * ~PROT_USER = remove user permissions.
5361  */
5362 static uint_t
5363 sfmmu_vtop_prot(uint_t vprot, uint_t *tteflagsp)
5364 {
5365 	if (vprot == (uint_t)~PROT_WRITE) {
5366 		*tteflagsp = TTE_WRPRM_INT | TTE_HWWR_INT;
5367 		return (0);		/* will cause wrprm to be cleared */
5368 	}
5369 	if (vprot == (uint_t)~PROT_USER) {
5370 		*tteflagsp = TTE_PRIV_INT;
5371 		return (0);		/* will cause privprm to be cleared */
5372 	}
5373 	if ((vprot == 0) || (vprot == PROT_USER) ||
5374 	    ((vprot & PROT_ALL) != vprot)) {
5375 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5376 	}
5377 
5378 	switch (vprot) {
5379 	case (PROT_READ):
5380 	case (PROT_EXEC):
5381 	case (PROT_EXEC | PROT_READ):
5382 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5383 		return (TTE_PRIV_INT); 		/* set prv and clr wrt */
5384 	case (PROT_WRITE):
5385 	case (PROT_WRITE | PROT_READ):
5386 	case (PROT_EXEC | PROT_WRITE):
5387 	case (PROT_EXEC | PROT_WRITE | PROT_READ):
5388 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5389 		return (TTE_PRIV_INT | TTE_WRPRM_INT); 	/* set prv and wrt */
5390 	case (PROT_USER | PROT_READ):
5391 	case (PROT_USER | PROT_EXEC):
5392 	case (PROT_USER | PROT_EXEC | PROT_READ):
5393 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5394 		return (0); 			/* clr prv and wrt */
5395 	case (PROT_USER | PROT_WRITE):
5396 	case (PROT_USER | PROT_WRITE | PROT_READ):
5397 	case (PROT_USER | PROT_EXEC | PROT_WRITE):
5398 	case (PROT_USER | PROT_EXEC | PROT_WRITE | PROT_READ):
5399 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5400 		return (TTE_WRPRM_INT); 	/* clr prv and set wrt */
5401 	default:
5402 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5403 	}
5404 	return (0);
5405 }
5406 
5407 /*
5408  * Alternate unload for very large virtual ranges. With a true 64 bit VA,
5409  * the normal algorithm would take too long for a very large VA range with
5410  * few real mappings. This routine just walks thru all HMEs in the global
5411  * hash table to find and remove mappings.
5412  */
5413 static void
5414 hat_unload_large_virtual(
5415 	struct hat		*sfmmup,
5416 	caddr_t			startaddr,
5417 	size_t			len,
5418 	uint_t			flags,
5419 	hat_callback_t		*callback)
5420 {
5421 	struct hmehash_bucket *hmebp;
5422 	struct hme_blk *hmeblkp;
5423 	struct hme_blk *pr_hblk = NULL;
5424 	struct hme_blk *nx_hblk;
5425 	struct hme_blk *list = NULL;
5426 	int i;
5427 	demap_range_t dmr, *dmrp;
5428 	cpuset_t cpuset;
5429 	caddr_t	endaddr = startaddr + len;
5430 	caddr_t	sa;
5431 	caddr_t	ea;
5432 	caddr_t	cb_sa[MAX_CB_ADDR];
5433 	caddr_t	cb_ea[MAX_CB_ADDR];
5434 	int	addr_cnt = 0;
5435 	int	a = 0;
5436 
5437 	if (sfmmup->sfmmu_free) {
5438 		dmrp = NULL;
5439 	} else {
5440 		dmrp = &dmr;
5441 		DEMAP_RANGE_INIT(sfmmup, dmrp);
5442 	}
5443 
5444 	/*
5445 	 * Loop through all the hash buckets of HME blocks looking for matches.
5446 	 */
5447 	for (i = 0; i <= UHMEHASH_SZ; i++) {
5448 		hmebp = &uhme_hash[i];
5449 		SFMMU_HASH_LOCK(hmebp);
5450 		hmeblkp = hmebp->hmeblkp;
5451 		pr_hblk = NULL;
5452 		while (hmeblkp) {
5453 			nx_hblk = hmeblkp->hblk_next;
5454 
5455 			/*
5456 			 * skip if not this context, if a shadow block or
5457 			 * if the mapping is not in the requested range
5458 			 */
5459 			if (hmeblkp->hblk_tag.htag_id != sfmmup ||
5460 			    hmeblkp->hblk_shw_bit ||
5461 			    (sa = (caddr_t)get_hblk_base(hmeblkp)) >= endaddr ||
5462 			    (ea = get_hblk_endaddr(hmeblkp)) <= startaddr) {
5463 				pr_hblk = hmeblkp;
5464 				goto next_block;
5465 			}
5466 
5467 			ASSERT(!hmeblkp->hblk_shared);
5468 			/*
5469 			 * unload if there are any current valid mappings
5470 			 */
5471 			if (hmeblkp->hblk_vcnt != 0 ||
5472 			    hmeblkp->hblk_hmecnt != 0)
5473 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
5474 				    sa, ea, dmrp, flags);
5475 
5476 			/*
5477 			 * on unmap we also release the HME block itself, once
5478 			 * all mappings are gone.
5479 			 */
5480 			if ((flags & HAT_UNLOAD_UNMAP) != 0 &&
5481 			    !hmeblkp->hblk_vcnt &&
5482 			    !hmeblkp->hblk_hmecnt) {
5483 				ASSERT(!hmeblkp->hblk_lckcnt);
5484 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
5485 				    &list, 0);
5486 			} else {
5487 				pr_hblk = hmeblkp;
5488 			}
5489 
5490 			if (callback == NULL)
5491 				goto next_block;
5492 
5493 			/*
5494 			 * HME blocks may span more than one page, but we may be
5495 			 * unmapping only one page, so check for a smaller range
5496 			 * for the callback
5497 			 */
5498 			if (sa < startaddr)
5499 				sa = startaddr;
5500 			if (--ea > endaddr)
5501 				ea = endaddr - 1;
5502 
5503 			cb_sa[addr_cnt] = sa;
5504 			cb_ea[addr_cnt] = ea;
5505 			if (++addr_cnt == MAX_CB_ADDR) {
5506 				if (dmrp != NULL) {
5507 					DEMAP_RANGE_FLUSH(dmrp);
5508 					cpuset = sfmmup->sfmmu_cpusran;
5509 					xt_sync(cpuset);
5510 				}
5511 
5512 				for (a = 0; a < MAX_CB_ADDR; ++a) {
5513 					callback->hcb_start_addr = cb_sa[a];
5514 					callback->hcb_end_addr = cb_ea[a];
5515 					callback->hcb_function(callback);
5516 				}
5517 				addr_cnt = 0;
5518 			}
5519 
5520 next_block:
5521 			hmeblkp = nx_hblk;
5522 		}
5523 		SFMMU_HASH_UNLOCK(hmebp);
5524 	}
5525 
5526 	sfmmu_hblks_list_purge(&list, 0);
5527 	if (dmrp != NULL) {
5528 		DEMAP_RANGE_FLUSH(dmrp);
5529 		cpuset = sfmmup->sfmmu_cpusran;
5530 		xt_sync(cpuset);
5531 	}
5532 
5533 	for (a = 0; a < addr_cnt; ++a) {
5534 		callback->hcb_start_addr = cb_sa[a];
5535 		callback->hcb_end_addr = cb_ea[a];
5536 		callback->hcb_function(callback);
5537 	}
5538 
5539 	/*
5540 	 * Check TSB and TLB page sizes if the process isn't exiting.
5541 	 */
5542 	if (!sfmmup->sfmmu_free)
5543 		sfmmu_check_page_sizes(sfmmup, 0);
5544 }
5545 
5546 /*
5547  * Unload all the mappings in the range [addr..addr+len). addr and len must
5548  * be MMU_PAGESIZE aligned.
5549  */
5550 
5551 extern struct seg *segkmap;
5552 #define	ISSEGKMAP(sfmmup, addr) (sfmmup == ksfmmup && \
5553 segkmap->s_base <= (addr) && (addr) < (segkmap->s_base + segkmap->s_size))
5554 
5555 
5556 void
5557 hat_unload_callback(
5558 	struct hat *sfmmup,
5559 	caddr_t addr,
5560 	size_t len,
5561 	uint_t flags,
5562 	hat_callback_t *callback)
5563 {
5564 	struct hmehash_bucket *hmebp;
5565 	hmeblk_tag hblktag;
5566 	int hmeshift, hashno, iskernel;
5567 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
5568 	caddr_t endaddr;
5569 	cpuset_t cpuset;
5570 	int addr_count = 0;
5571 	int a;
5572 	caddr_t cb_start_addr[MAX_CB_ADDR];
5573 	caddr_t cb_end_addr[MAX_CB_ADDR];
5574 	int issegkmap = ISSEGKMAP(sfmmup, addr);
5575 	demap_range_t dmr, *dmrp;
5576 
5577 	if (sfmmup->sfmmu_xhat_provider) {
5578 		XHAT_UNLOAD_CALLBACK(sfmmup, addr, len, flags, callback);
5579 		return;
5580 	} else {
5581 		/*
5582 		 * This must be a CPU HAT. If the address space has
5583 		 * XHATs attached, unload the mappings for all of them,
5584 		 * just in case
5585 		 */
5586 		ASSERT(sfmmup->sfmmu_as != NULL);
5587 		if (sfmmup->sfmmu_as->a_xhat != NULL)
5588 			xhat_unload_callback_all(sfmmup->sfmmu_as, addr,
5589 			    len, flags, callback);
5590 	}
5591 
5592 	ASSERT((sfmmup == ksfmmup) || (flags & HAT_UNLOAD_OTHER) || \
5593 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
5594 
5595 	ASSERT(sfmmup != NULL);
5596 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5597 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
5598 
5599 	/*
5600 	 * Probing through a large VA range (say 63 bits) will be slow, even
5601 	 * at 4 Meg steps between the probes. So, when the virtual address range
5602 	 * is very large, search the HME entries for what to unload.
5603 	 *
5604 	 *	len >> TTE_PAGE_SHIFT(TTE4M) is the # of 4Meg probes we'd need
5605 	 *
5606 	 *	UHMEHASH_SZ is number of hash buckets to examine
5607 	 *
5608 	 */
5609 	if (sfmmup != KHATID && (len >> TTE_PAGE_SHIFT(TTE4M)) > UHMEHASH_SZ) {
5610 		hat_unload_large_virtual(sfmmup, addr, len, flags, callback);
5611 		return;
5612 	}
5613 
5614 	CPUSET_ZERO(cpuset);
5615 
5616 	/*
5617 	 * If the process is exiting, we can save a lot of fuss since
5618 	 * we'll flush the TLB when we free the ctx anyway.
5619 	 */
5620 	if (sfmmup->sfmmu_free)
5621 		dmrp = NULL;
5622 	else
5623 		dmrp = &dmr;
5624 
5625 	DEMAP_RANGE_INIT(sfmmup, dmrp);
5626 	endaddr = addr + len;
5627 	hblktag.htag_id = sfmmup;
5628 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5629 
5630 	/*
5631 	 * It is likely for the vm to call unload over a wide range of
5632 	 * addresses that are actually very sparsely populated by
5633 	 * translations.  In order to speed this up the sfmmu hat supports
5634 	 * the concept of shadow hmeblks. Dummy large page hmeblks that
5635 	 * correspond to actual small translations are allocated at tteload
5636 	 * time and are referred to as shadow hmeblks.  Now, during unload
5637 	 * time, we first check if we have a shadow hmeblk for that
5638 	 * translation.  The absence of one means the corresponding address
5639 	 * range is empty and can be skipped.
5640 	 *
5641 	 * The kernel is an exception to above statement and that is why
5642 	 * we don't use shadow hmeblks and hash starting from the smallest
5643 	 * page size.
5644 	 */
5645 	if (sfmmup == KHATID) {
5646 		iskernel = 1;
5647 		hashno = TTE64K;
5648 	} else {
5649 		iskernel = 0;
5650 		if (mmu_page_sizes == max_mmu_page_sizes) {
5651 			hashno = TTE256M;
5652 		} else {
5653 			hashno = TTE4M;
5654 		}
5655 	}
5656 	while (addr < endaddr) {
5657 		hmeshift = HME_HASH_SHIFT(hashno);
5658 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5659 		hblktag.htag_rehash = hashno;
5660 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5661 
5662 		SFMMU_HASH_LOCK(hmebp);
5663 
5664 		HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
5665 		if (hmeblkp == NULL) {
5666 			/*
5667 			 * didn't find an hmeblk. skip the appropiate
5668 			 * address range.
5669 			 */
5670 			SFMMU_HASH_UNLOCK(hmebp);
5671 			if (iskernel) {
5672 				if (hashno < mmu_hashcnt) {
5673 					hashno++;
5674 					continue;
5675 				} else {
5676 					hashno = TTE64K;
5677 					addr = (caddr_t)roundup((uintptr_t)addr
5678 					    + 1, MMU_PAGESIZE64K);
5679 					continue;
5680 				}
5681 			}
5682 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5683 			    (1 << hmeshift));
5684 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5685 				ASSERT(hashno == TTE64K);
5686 				continue;
5687 			}
5688 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5689 				hashno = TTE512K;
5690 				continue;
5691 			}
5692 			if (mmu_page_sizes == max_mmu_page_sizes) {
5693 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5694 					hashno = TTE4M;
5695 					continue;
5696 				}
5697 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5698 					hashno = TTE32M;
5699 					continue;
5700 				}
5701 				hashno = TTE256M;
5702 				continue;
5703 			} else {
5704 				hashno = TTE4M;
5705 				continue;
5706 			}
5707 		}
5708 		ASSERT(hmeblkp);
5709 		ASSERT(!hmeblkp->hblk_shared);
5710 		if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5711 			/*
5712 			 * If the valid count is zero we can skip the range
5713 			 * mapped by this hmeblk.
5714 			 * We free hblks in the case of HAT_UNMAP.  HAT_UNMAP
5715 			 * is used by segment drivers as a hint
5716 			 * that the mapping resource won't be used any longer.
5717 			 * The best example of this is during exit().
5718 			 */
5719 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5720 			    get_hblk_span(hmeblkp));
5721 			if ((flags & HAT_UNLOAD_UNMAP) ||
5722 			    (iskernel && !issegkmap)) {
5723 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
5724 				    &list, 0);
5725 			}
5726 			SFMMU_HASH_UNLOCK(hmebp);
5727 
5728 			if (iskernel) {
5729 				hashno = TTE64K;
5730 				continue;
5731 			}
5732 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5733 				ASSERT(hashno == TTE64K);
5734 				continue;
5735 			}
5736 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5737 				hashno = TTE512K;
5738 				continue;
5739 			}
5740 			if (mmu_page_sizes == max_mmu_page_sizes) {
5741 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5742 					hashno = TTE4M;
5743 					continue;
5744 				}
5745 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5746 					hashno = TTE32M;
5747 					continue;
5748 				}
5749 				hashno = TTE256M;
5750 				continue;
5751 			} else {
5752 				hashno = TTE4M;
5753 				continue;
5754 			}
5755 		}
5756 		if (hmeblkp->hblk_shw_bit) {
5757 			/*
5758 			 * If we encounter a shadow hmeblk we know there is
5759 			 * smaller sized hmeblks mapping the same address space.
5760 			 * Decrement the hash size and rehash.
5761 			 */
5762 			ASSERT(sfmmup != KHATID);
5763 			hashno--;
5764 			SFMMU_HASH_UNLOCK(hmebp);
5765 			continue;
5766 		}
5767 
5768 		/*
5769 		 * track callback address ranges.
5770 		 * only start a new range when it's not contiguous
5771 		 */
5772 		if (callback != NULL) {
5773 			if (addr_count > 0 &&
5774 			    addr == cb_end_addr[addr_count - 1])
5775 				--addr_count;
5776 			else
5777 				cb_start_addr[addr_count] = addr;
5778 		}
5779 
5780 		addr = sfmmu_hblk_unload(sfmmup, hmeblkp, addr, endaddr,
5781 		    dmrp, flags);
5782 
5783 		if (callback != NULL)
5784 			cb_end_addr[addr_count++] = addr;
5785 
5786 		if (((flags & HAT_UNLOAD_UNMAP) || (iskernel && !issegkmap)) &&
5787 		    !hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5788 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 0);
5789 		}
5790 		SFMMU_HASH_UNLOCK(hmebp);
5791 
5792 		/*
5793 		 * Notify our caller as to exactly which pages
5794 		 * have been unloaded. We do these in clumps,
5795 		 * to minimize the number of xt_sync()s that need to occur.
5796 		 */
5797 		if (callback != NULL && addr_count == MAX_CB_ADDR) {
5798 			DEMAP_RANGE_FLUSH(dmrp);
5799 			if (dmrp != NULL) {
5800 				cpuset = sfmmup->sfmmu_cpusran;
5801 				xt_sync(cpuset);
5802 			}
5803 
5804 			for (a = 0; a < MAX_CB_ADDR; ++a) {
5805 				callback->hcb_start_addr = cb_start_addr[a];
5806 				callback->hcb_end_addr = cb_end_addr[a];
5807 				callback->hcb_function(callback);
5808 			}
5809 			addr_count = 0;
5810 		}
5811 		if (iskernel) {
5812 			hashno = TTE64K;
5813 			continue;
5814 		}
5815 		if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5816 			ASSERT(hashno == TTE64K);
5817 			continue;
5818 		}
5819 		if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5820 			hashno = TTE512K;
5821 			continue;
5822 		}
5823 		if (mmu_page_sizes == max_mmu_page_sizes) {
5824 			if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5825 				hashno = TTE4M;
5826 				continue;
5827 			}
5828 			if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5829 				hashno = TTE32M;
5830 				continue;
5831 			}
5832 			hashno = TTE256M;
5833 		} else {
5834 			hashno = TTE4M;
5835 		}
5836 	}
5837 
5838 	sfmmu_hblks_list_purge(&list, 0);
5839 	DEMAP_RANGE_FLUSH(dmrp);
5840 	if (dmrp != NULL) {
5841 		cpuset = sfmmup->sfmmu_cpusran;
5842 		xt_sync(cpuset);
5843 	}
5844 	if (callback && addr_count != 0) {
5845 		for (a = 0; a < addr_count; ++a) {
5846 			callback->hcb_start_addr = cb_start_addr[a];
5847 			callback->hcb_end_addr = cb_end_addr[a];
5848 			callback->hcb_function(callback);
5849 		}
5850 	}
5851 
5852 	/*
5853 	 * Check TSB and TLB page sizes if the process isn't exiting.
5854 	 */
5855 	if (!sfmmup->sfmmu_free)
5856 		sfmmu_check_page_sizes(sfmmup, 0);
5857 }
5858 
5859 /*
5860  * Unload all the mappings in the range [addr..addr+len). addr and len must
5861  * be MMU_PAGESIZE aligned.
5862  */
5863 void
5864 hat_unload(struct hat *sfmmup, caddr_t addr, size_t len, uint_t flags)
5865 {
5866 	if (sfmmup->sfmmu_xhat_provider) {
5867 		XHAT_UNLOAD(sfmmup, addr, len, flags);
5868 		return;
5869 	}
5870 	hat_unload_callback(sfmmup, addr, len, flags, NULL);
5871 }
5872 
5873 
5874 /*
5875  * Find the largest mapping size for this page.
5876  */
5877 int
5878 fnd_mapping_sz(page_t *pp)
5879 {
5880 	int sz;
5881 	int p_index;
5882 
5883 	p_index = PP_MAPINDEX(pp);
5884 
5885 	sz = 0;
5886 	p_index >>= 1;	/* don't care about 8K bit */
5887 	for (; p_index; p_index >>= 1) {
5888 		sz++;
5889 	}
5890 
5891 	return (sz);
5892 }
5893 
5894 /*
5895  * This function unloads a range of addresses for an hmeblk.
5896  * It returns the next address to be unloaded.
5897  * It should be called with the hash lock held.
5898  */
5899 static caddr_t
5900 sfmmu_hblk_unload(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5901 	caddr_t endaddr, demap_range_t *dmrp, uint_t flags)
5902 {
5903 	tte_t	tte, ttemod;
5904 	struct	sf_hment *sfhmep;
5905 	int	ttesz;
5906 	long	ttecnt;
5907 	page_t *pp;
5908 	kmutex_t *pml;
5909 	int ret;
5910 	int use_demap_range;
5911 
5912 	ASSERT(in_hblk_range(hmeblkp, addr));
5913 	ASSERT(!hmeblkp->hblk_shw_bit);
5914 	ASSERT(sfmmup != NULL || hmeblkp->hblk_shared);
5915 	ASSERT(sfmmup == NULL || !hmeblkp->hblk_shared);
5916 	ASSERT(dmrp == NULL || !hmeblkp->hblk_shared);
5917 
5918 #ifdef DEBUG
5919 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5920 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5921 		panic("sfmmu_hblk_unload: partial unload of large page");
5922 	}
5923 #endif /* DEBUG */
5924 
5925 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5926 	ttesz = get_hblk_ttesz(hmeblkp);
5927 
5928 	use_demap_range = ((dmrp == NULL) ||
5929 	    (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp)));
5930 
5931 	if (use_demap_range) {
5932 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5933 	} else {
5934 		DEMAP_RANGE_FLUSH(dmrp);
5935 	}
5936 	ttecnt = 0;
5937 	HBLKTOHME(sfhmep, hmeblkp, addr);
5938 
5939 	while (addr < endaddr) {
5940 		pml = NULL;
5941 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5942 		if (TTE_IS_VALID(&tte)) {
5943 			pp = sfhmep->hme_page;
5944 			if (pp != NULL) {
5945 				pml = sfmmu_mlist_enter(pp);
5946 			}
5947 
5948 			/*
5949 			 * Verify if hme still points to 'pp' now that
5950 			 * we have p_mapping lock.
5951 			 */
5952 			if (sfhmep->hme_page != pp) {
5953 				if (pp != NULL && sfhmep->hme_page != NULL) {
5954 					ASSERT(pml != NULL);
5955 					sfmmu_mlist_exit(pml);
5956 					/* Re-start this iteration. */
5957 					continue;
5958 				}
5959 				ASSERT((pp != NULL) &&
5960 				    (sfhmep->hme_page == NULL));
5961 				goto tte_unloaded;
5962 			}
5963 
5964 			/*
5965 			 * This point on we have both HASH and p_mapping
5966 			 * lock.
5967 			 */
5968 			ASSERT(pp == sfhmep->hme_page);
5969 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5970 
5971 			/*
5972 			 * We need to loop on modify tte because it is
5973 			 * possible for pagesync to come along and
5974 			 * change the software bits beneath us.
5975 			 *
5976 			 * Page_unload can also invalidate the tte after
5977 			 * we read tte outside of p_mapping lock.
5978 			 */
5979 again:
5980 			ttemod = tte;
5981 
5982 			TTE_SET_INVALID(&ttemod);
5983 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5984 			    &sfhmep->hme_tte);
5985 
5986 			if (ret <= 0) {
5987 				if (TTE_IS_VALID(&tte)) {
5988 					ASSERT(ret < 0);
5989 					goto again;
5990 				}
5991 				if (pp != NULL) {
5992 					panic("sfmmu_hblk_unload: pp = 0x%p "
5993 					    "tte became invalid under mlist"
5994 					    " lock = 0x%p", (void *)pp,
5995 					    (void *)pml);
5996 				}
5997 				continue;
5998 			}
5999 
6000 			if (!(flags & HAT_UNLOAD_NOSYNC)) {
6001 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
6002 			}
6003 
6004 			/*
6005 			 * Ok- we invalidated the tte. Do the rest of the job.
6006 			 */
6007 			ttecnt++;
6008 
6009 			if (flags & HAT_UNLOAD_UNLOCK) {
6010 				ASSERT(hmeblkp->hblk_lckcnt > 0);
6011 				atomic_add_32(&hmeblkp->hblk_lckcnt, -1);
6012 				HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
6013 			}
6014 
6015 			/*
6016 			 * Normally we would need to flush the page
6017 			 * from the virtual cache at this point in
6018 			 * order to prevent a potential cache alias
6019 			 * inconsistency.
6020 			 * The particular scenario we need to worry
6021 			 * about is:
6022 			 * Given:  va1 and va2 are two virtual address
6023 			 * that alias and map the same physical
6024 			 * address.
6025 			 * 1.   mapping exists from va1 to pa and data
6026 			 * has been read into the cache.
6027 			 * 2.   unload va1.
6028 			 * 3.   load va2 and modify data using va2.
6029 			 * 4    unload va2.
6030 			 * 5.   load va1 and reference data.  Unless we
6031 			 * flush the data cache when we unload we will
6032 			 * get stale data.
6033 			 * Fortunately, page coloring eliminates the
6034 			 * above scenario by remembering the color a
6035 			 * physical page was last or is currently
6036 			 * mapped to.  Now, we delay the flush until
6037 			 * the loading of translations.  Only when the
6038 			 * new translation is of a different color
6039 			 * are we forced to flush.
6040 			 */
6041 			if (use_demap_range) {
6042 				/*
6043 				 * Mark this page as needing a demap.
6044 				 */
6045 				DEMAP_RANGE_MARKPG(dmrp, addr);
6046 			} else {
6047 				ASSERT(sfmmup != NULL);
6048 				ASSERT(!hmeblkp->hblk_shared);
6049 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
6050 				    sfmmup->sfmmu_free, 0);
6051 			}
6052 
6053 			if (pp) {
6054 				/*
6055 				 * Remove the hment from the mapping list
6056 				 */
6057 				ASSERT(hmeblkp->hblk_hmecnt > 0);
6058 
6059 				/*
6060 				 * Again, we cannot
6061 				 * ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS);
6062 				 */
6063 				HME_SUB(sfhmep, pp);
6064 				membar_stst();
6065 				atomic_add_16(&hmeblkp->hblk_hmecnt, -1);
6066 			}
6067 
6068 			ASSERT(hmeblkp->hblk_vcnt > 0);
6069 			atomic_add_16(&hmeblkp->hblk_vcnt, -1);
6070 
6071 			ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
6072 			    !hmeblkp->hblk_lckcnt);
6073 
6074 #ifdef VAC
6075 			if (pp && (pp->p_nrm & (P_KPMC | P_KPMS | P_TNC))) {
6076 				if (PP_ISTNC(pp)) {
6077 					/*
6078 					 * If page was temporary
6079 					 * uncached, try to recache
6080 					 * it. Note that HME_SUB() was
6081 					 * called above so p_index and
6082 					 * mlist had been updated.
6083 					 */
6084 					conv_tnc(pp, ttesz);
6085 				} else if (pp->p_mapping == NULL) {
6086 					ASSERT(kpm_enable);
6087 					/*
6088 					 * Page is marked to be in VAC conflict
6089 					 * to an existing kpm mapping and/or is
6090 					 * kpm mapped using only the regular
6091 					 * pagesize.
6092 					 */
6093 					sfmmu_kpm_hme_unload(pp);
6094 				}
6095 			}
6096 #endif	/* VAC */
6097 		} else if ((pp = sfhmep->hme_page) != NULL) {
6098 				/*
6099 				 * TTE is invalid but the hme
6100 				 * still exists. let pageunload
6101 				 * complete its job.
6102 				 */
6103 				ASSERT(pml == NULL);
6104 				pml = sfmmu_mlist_enter(pp);
6105 				if (sfhmep->hme_page != NULL) {
6106 					sfmmu_mlist_exit(pml);
6107 					continue;
6108 				}
6109 				ASSERT(sfhmep->hme_page == NULL);
6110 		} else if (hmeblkp->hblk_hmecnt != 0) {
6111 			/*
6112 			 * pageunload may have not finished decrementing
6113 			 * hblk_vcnt and hblk_hmecnt. Find page_t if any and
6114 			 * wait for pageunload to finish. Rely on pageunload
6115 			 * to decrement hblk_hmecnt after hblk_vcnt.
6116 			 */
6117 			pfn_t pfn = TTE_TO_TTEPFN(&tte);
6118 			ASSERT(pml == NULL);
6119 			if (pf_is_memory(pfn)) {
6120 				pp = page_numtopp_nolock(pfn);
6121 				if (pp != NULL) {
6122 					pml = sfmmu_mlist_enter(pp);
6123 					sfmmu_mlist_exit(pml);
6124 					pml = NULL;
6125 				}
6126 			}
6127 		}
6128 
6129 tte_unloaded:
6130 		/*
6131 		 * At this point, the tte we are looking at
6132 		 * should be unloaded, and hme has been unlinked
6133 		 * from page too. This is important because in
6134 		 * pageunload, it does ttesync() then HME_SUB.
6135 		 * We need to make sure HME_SUB has been completed
6136 		 * so we know ttesync() has been completed. Otherwise,
6137 		 * at exit time, after return from hat layer, VM will
6138 		 * release as structure which hat_setstat() (called
6139 		 * by ttesync()) needs.
6140 		 */
6141 #ifdef DEBUG
6142 		{
6143 			tte_t	dtte;
6144 
6145 			ASSERT(sfhmep->hme_page == NULL);
6146 
6147 			sfmmu_copytte(&sfhmep->hme_tte, &dtte);
6148 			ASSERT(!TTE_IS_VALID(&dtte));
6149 		}
6150 #endif
6151 
6152 		if (pml) {
6153 			sfmmu_mlist_exit(pml);
6154 		}
6155 
6156 		addr += TTEBYTES(ttesz);
6157 		sfhmep++;
6158 		DEMAP_RANGE_NEXTPG(dmrp);
6159 	}
6160 	/*
6161 	 * For shared hmeblks this routine is only called when region is freed
6162 	 * and no longer referenced.  So no need to decrement ttecnt
6163 	 * in the region structure here.
6164 	 */
6165 	if (ttecnt > 0 && sfmmup != NULL) {
6166 		atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -ttecnt);
6167 	}
6168 	return (addr);
6169 }
6170 
6171 /*
6172  * Invalidate a virtual address range for the local CPU.
6173  * For best performance ensure that the va range is completely
6174  * mapped, otherwise the entire TLB will be flushed.
6175  */
6176 void
6177 hat_flush_range(struct hat *sfmmup, caddr_t va, size_t size)
6178 {
6179 	ssize_t sz;
6180 	caddr_t endva = va + size;
6181 
6182 	while (va < endva) {
6183 		sz = hat_getpagesize(sfmmup, va);
6184 		if (sz < 0) {
6185 			vtag_flushall();
6186 			break;
6187 		}
6188 		vtag_flushpage(va, (uint64_t)sfmmup);
6189 		va += sz;
6190 	}
6191 }
6192 
6193 /*
6194  * Synchronize all the mappings in the range [addr..addr+len).
6195  * Can be called with clearflag having two states:
6196  * HAT_SYNC_DONTZERO means just return the rm stats
6197  * HAT_SYNC_ZERORM means zero rm bits in the tte and return the stats
6198  */
6199 void
6200 hat_sync(struct hat *sfmmup, caddr_t addr, size_t len, uint_t clearflag)
6201 {
6202 	struct hmehash_bucket *hmebp;
6203 	hmeblk_tag hblktag;
6204 	int hmeshift, hashno = 1;
6205 	struct hme_blk *hmeblkp, *list = NULL;
6206 	caddr_t endaddr;
6207 	cpuset_t cpuset;
6208 
6209 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
6210 	ASSERT((sfmmup == ksfmmup) ||
6211 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
6212 	ASSERT((len & MMU_PAGEOFFSET) == 0);
6213 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
6214 	    (clearflag == HAT_SYNC_ZERORM));
6215 
6216 	CPUSET_ZERO(cpuset);
6217 
6218 	endaddr = addr + len;
6219 	hblktag.htag_id = sfmmup;
6220 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
6221 
6222 	/*
6223 	 * Spitfire supports 4 page sizes.
6224 	 * Most pages are expected to be of the smallest page
6225 	 * size (8K) and these will not need to be rehashed. 64K
6226 	 * pages also don't need to be rehashed because the an hmeblk
6227 	 * spans 64K of address space. 512K pages might need 1 rehash and
6228 	 * and 4M pages 2 rehashes.
6229 	 */
6230 	while (addr < endaddr) {
6231 		hmeshift = HME_HASH_SHIFT(hashno);
6232 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
6233 		hblktag.htag_rehash = hashno;
6234 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
6235 
6236 		SFMMU_HASH_LOCK(hmebp);
6237 
6238 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
6239 		if (hmeblkp != NULL) {
6240 			ASSERT(!hmeblkp->hblk_shared);
6241 			/*
6242 			 * We've encountered a shadow hmeblk so skip the range
6243 			 * of the next smaller mapping size.
6244 			 */
6245 			if (hmeblkp->hblk_shw_bit) {
6246 				ASSERT(sfmmup != ksfmmup);
6247 				ASSERT(hashno > 1);
6248 				addr = (caddr_t)P2END((uintptr_t)addr,
6249 				    TTEBYTES(hashno - 1));
6250 			} else {
6251 				addr = sfmmu_hblk_sync(sfmmup, hmeblkp,
6252 				    addr, endaddr, clearflag);
6253 			}
6254 			SFMMU_HASH_UNLOCK(hmebp);
6255 			hashno = 1;
6256 			continue;
6257 		}
6258 		SFMMU_HASH_UNLOCK(hmebp);
6259 
6260 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
6261 			/*
6262 			 * We have traversed the whole list and rehashed
6263 			 * if necessary without finding the address to sync.
6264 			 * This is ok so we increment the address by the
6265 			 * smallest hmeblk range for kernel mappings and the
6266 			 * largest hmeblk range, to account for shadow hmeblks,
6267 			 * for user mappings and continue.
6268 			 */
6269 			if (sfmmup == ksfmmup)
6270 				addr = (caddr_t)P2END((uintptr_t)addr,
6271 				    TTEBYTES(1));
6272 			else
6273 				addr = (caddr_t)P2END((uintptr_t)addr,
6274 				    TTEBYTES(hashno));
6275 			hashno = 1;
6276 		} else {
6277 			hashno++;
6278 		}
6279 	}
6280 	sfmmu_hblks_list_purge(&list, 0);
6281 	cpuset = sfmmup->sfmmu_cpusran;
6282 	xt_sync(cpuset);
6283 }
6284 
6285 static caddr_t
6286 sfmmu_hblk_sync(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
6287 	caddr_t endaddr, int clearflag)
6288 {
6289 	tte_t	tte, ttemod;
6290 	struct sf_hment *sfhmep;
6291 	int ttesz;
6292 	struct page *pp;
6293 	kmutex_t *pml;
6294 	int ret;
6295 
6296 	ASSERT(hmeblkp->hblk_shw_bit == 0);
6297 	ASSERT(!hmeblkp->hblk_shared);
6298 
6299 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
6300 
6301 	ttesz = get_hblk_ttesz(hmeblkp);
6302 	HBLKTOHME(sfhmep, hmeblkp, addr);
6303 
6304 	while (addr < endaddr) {
6305 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6306 		if (TTE_IS_VALID(&tte)) {
6307 			pml = NULL;
6308 			pp = sfhmep->hme_page;
6309 			if (pp) {
6310 				pml = sfmmu_mlist_enter(pp);
6311 			}
6312 			if (pp != sfhmep->hme_page) {
6313 				/*
6314 				 * tte most have been unloaded
6315 				 * underneath us.  Recheck
6316 				 */
6317 				ASSERT(pml);
6318 				sfmmu_mlist_exit(pml);
6319 				continue;
6320 			}
6321 
6322 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
6323 
6324 			if (clearflag == HAT_SYNC_ZERORM) {
6325 				ttemod = tte;
6326 				TTE_CLR_RM(&ttemod);
6327 				ret = sfmmu_modifytte_try(&tte, &ttemod,
6328 				    &sfhmep->hme_tte);
6329 				if (ret < 0) {
6330 					if (pml) {
6331 						sfmmu_mlist_exit(pml);
6332 					}
6333 					continue;
6334 				}
6335 
6336 				if (ret > 0) {
6337 					sfmmu_tlb_demap(addr, sfmmup,
6338 					    hmeblkp, 0, 0);
6339 				}
6340 			}
6341 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
6342 			if (pml) {
6343 				sfmmu_mlist_exit(pml);
6344 			}
6345 		}
6346 		addr += TTEBYTES(ttesz);
6347 		sfhmep++;
6348 	}
6349 	return (addr);
6350 }
6351 
6352 /*
6353  * This function will sync a tte to the page struct and it will
6354  * update the hat stats. Currently it allows us to pass a NULL pp
6355  * and we will simply update the stats.  We may want to change this
6356  * so we only keep stats for pages backed by pp's.
6357  */
6358 static void
6359 sfmmu_ttesync(struct hat *sfmmup, caddr_t addr, tte_t *ttep, page_t *pp)
6360 {
6361 	uint_t rm = 0;
6362 	int   	sz;
6363 	pgcnt_t	npgs;
6364 
6365 	ASSERT(TTE_IS_VALID(ttep));
6366 
6367 	if (TTE_IS_NOSYNC(ttep)) {
6368 		return;
6369 	}
6370 
6371 	if (TTE_IS_REF(ttep))  {
6372 		rm = P_REF;
6373 	}
6374 	if (TTE_IS_MOD(ttep))  {
6375 		rm |= P_MOD;
6376 	}
6377 
6378 	if (rm == 0) {
6379 		return;
6380 	}
6381 
6382 	sz = TTE_CSZ(ttep);
6383 	if (sfmmup != NULL && sfmmup->sfmmu_rmstat) {
6384 		int i;
6385 		caddr_t	vaddr = addr;
6386 
6387 		for (i = 0; i < TTEPAGES(sz); i++, vaddr += MMU_PAGESIZE) {
6388 			hat_setstat(sfmmup->sfmmu_as, vaddr, MMU_PAGESIZE, rm);
6389 		}
6390 
6391 	}
6392 
6393 	/*
6394 	 * XXX I want to use cas to update nrm bits but they
6395 	 * currently belong in common/vm and not in hat where
6396 	 * they should be.
6397 	 * The nrm bits are protected by the same mutex as
6398 	 * the one that protects the page's mapping list.
6399 	 */
6400 	if (!pp)
6401 		return;
6402 	ASSERT(sfmmu_mlist_held(pp));
6403 	/*
6404 	 * If the tte is for a large page, we need to sync all the
6405 	 * pages covered by the tte.
6406 	 */
6407 	if (sz != TTE8K) {
6408 		ASSERT(pp->p_szc != 0);
6409 		pp = PP_GROUPLEADER(pp, sz);
6410 		ASSERT(sfmmu_mlist_held(pp));
6411 	}
6412 
6413 	/* Get number of pages from tte size. */
6414 	npgs = TTEPAGES(sz);
6415 
6416 	do {
6417 		ASSERT(pp);
6418 		ASSERT(sfmmu_mlist_held(pp));
6419 		if (((rm & P_REF) != 0 && !PP_ISREF(pp)) ||
6420 		    ((rm & P_MOD) != 0 && !PP_ISMOD(pp)))
6421 			hat_page_setattr(pp, rm);
6422 
6423 		/*
6424 		 * Are we done? If not, we must have a large mapping.
6425 		 * For large mappings we need to sync the rest of the pages
6426 		 * covered by this tte; goto the next page.
6427 		 */
6428 	} while (--npgs > 0 && (pp = PP_PAGENEXT(pp)));
6429 }
6430 
6431 /*
6432  * Execute pre-callback handler of each pa_hment linked to pp
6433  *
6434  * Inputs:
6435  *   flag: either HAT_PRESUSPEND or HAT_SUSPEND.
6436  *   capture_cpus: pointer to return value (below)
6437  *
6438  * Returns:
6439  *   Propagates the subsystem callback return values back to the caller;
6440  *   returns 0 on success.  If capture_cpus is non-NULL, the value returned
6441  *   is zero if all of the pa_hments are of a type that do not require
6442  *   capturing CPUs prior to suspending the mapping, else it is 1.
6443  */
6444 static int
6445 hat_pageprocess_precallbacks(struct page *pp, uint_t flag, int *capture_cpus)
6446 {
6447 	struct sf_hment	*sfhmep;
6448 	struct pa_hment *pahmep;
6449 	int (*f)(caddr_t, uint_t, uint_t, void *);
6450 	int		ret;
6451 	id_t		id;
6452 	int		locked = 0;
6453 	kmutex_t	*pml;
6454 
6455 	ASSERT(PAGE_EXCL(pp));
6456 	if (!sfmmu_mlist_held(pp)) {
6457 		pml = sfmmu_mlist_enter(pp);
6458 		locked = 1;
6459 	}
6460 
6461 	if (capture_cpus)
6462 		*capture_cpus = 0;
6463 
6464 top:
6465 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6466 		/*
6467 		 * skip sf_hments corresponding to VA<->PA mappings;
6468 		 * for pa_hment's, hme_tte.ll is zero
6469 		 */
6470 		if (!IS_PAHME(sfhmep))
6471 			continue;
6472 
6473 		pahmep = sfhmep->hme_data;
6474 		ASSERT(pahmep != NULL);
6475 
6476 		/*
6477 		 * skip if pre-handler has been called earlier in this loop
6478 		 */
6479 		if (pahmep->flags & flag)
6480 			continue;
6481 
6482 		id = pahmep->cb_id;
6483 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6484 		if (capture_cpus && sfmmu_cb_table[id].capture_cpus != 0)
6485 			*capture_cpus = 1;
6486 		if ((f = sfmmu_cb_table[id].prehandler) == NULL) {
6487 			pahmep->flags |= flag;
6488 			continue;
6489 		}
6490 
6491 		/*
6492 		 * Drop the mapping list lock to avoid locking order issues.
6493 		 */
6494 		if (locked)
6495 			sfmmu_mlist_exit(pml);
6496 
6497 		ret = f(pahmep->addr, pahmep->len, flag, pahmep->pvt);
6498 		if (ret != 0)
6499 			return (ret);	/* caller must do the cleanup */
6500 
6501 		if (locked) {
6502 			pml = sfmmu_mlist_enter(pp);
6503 			pahmep->flags |= flag;
6504 			goto top;
6505 		}
6506 
6507 		pahmep->flags |= flag;
6508 	}
6509 
6510 	if (locked)
6511 		sfmmu_mlist_exit(pml);
6512 
6513 	return (0);
6514 }
6515 
6516 /*
6517  * Execute post-callback handler of each pa_hment linked to pp
6518  *
6519  * Same overall assumptions and restrictions apply as for
6520  * hat_pageprocess_precallbacks().
6521  */
6522 static void
6523 hat_pageprocess_postcallbacks(struct page *pp, uint_t flag)
6524 {
6525 	pfn_t pgpfn = pp->p_pagenum;
6526 	pfn_t pgmask = btop(page_get_pagesize(pp->p_szc)) - 1;
6527 	pfn_t newpfn;
6528 	struct sf_hment *sfhmep;
6529 	struct pa_hment *pahmep;
6530 	int (*f)(caddr_t, uint_t, uint_t, void *, pfn_t);
6531 	id_t	id;
6532 	int	locked = 0;
6533 	kmutex_t *pml;
6534 
6535 	ASSERT(PAGE_EXCL(pp));
6536 	if (!sfmmu_mlist_held(pp)) {
6537 		pml = sfmmu_mlist_enter(pp);
6538 		locked = 1;
6539 	}
6540 
6541 top:
6542 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6543 		/*
6544 		 * skip sf_hments corresponding to VA<->PA mappings;
6545 		 * for pa_hment's, hme_tte.ll is zero
6546 		 */
6547 		if (!IS_PAHME(sfhmep))
6548 			continue;
6549 
6550 		pahmep = sfhmep->hme_data;
6551 		ASSERT(pahmep != NULL);
6552 
6553 		if ((pahmep->flags & flag) == 0)
6554 			continue;
6555 
6556 		pahmep->flags &= ~flag;
6557 
6558 		id = pahmep->cb_id;
6559 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6560 		if ((f = sfmmu_cb_table[id].posthandler) == NULL)
6561 			continue;
6562 
6563 		/*
6564 		 * Convert the base page PFN into the constituent PFN
6565 		 * which is needed by the callback handler.
6566 		 */
6567 		newpfn = pgpfn | (btop((uintptr_t)pahmep->addr) & pgmask);
6568 
6569 		/*
6570 		 * Drop the mapping list lock to avoid locking order issues.
6571 		 */
6572 		if (locked)
6573 			sfmmu_mlist_exit(pml);
6574 
6575 		if (f(pahmep->addr, pahmep->len, flag, pahmep->pvt, newpfn)
6576 		    != 0)
6577 			panic("sfmmu: posthandler failed");
6578 
6579 		if (locked) {
6580 			pml = sfmmu_mlist_enter(pp);
6581 			goto top;
6582 		}
6583 	}
6584 
6585 	if (locked)
6586 		sfmmu_mlist_exit(pml);
6587 }
6588 
6589 /*
6590  * Suspend locked kernel mapping
6591  */
6592 void
6593 hat_pagesuspend(struct page *pp)
6594 {
6595 	struct sf_hment *sfhmep;
6596 	sfmmu_t *sfmmup;
6597 	tte_t tte, ttemod;
6598 	struct hme_blk *hmeblkp;
6599 	caddr_t addr;
6600 	int index, cons;
6601 	cpuset_t cpuset;
6602 
6603 	ASSERT(PAGE_EXCL(pp));
6604 	ASSERT(sfmmu_mlist_held(pp));
6605 
6606 	mutex_enter(&kpr_suspendlock);
6607 
6608 	/*
6609 	 * We're about to suspend a kernel mapping so mark this thread as
6610 	 * non-traceable by DTrace. This prevents us from running into issues
6611 	 * with probe context trying to touch a suspended page
6612 	 * in the relocation codepath itself.
6613 	 */
6614 	curthread->t_flag |= T_DONTDTRACE;
6615 
6616 	index = PP_MAPINDEX(pp);
6617 	cons = TTE8K;
6618 
6619 retry:
6620 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6621 
6622 		if (IS_PAHME(sfhmep))
6623 			continue;
6624 
6625 		if (get_hblk_ttesz(sfmmu_hmetohblk(sfhmep)) != cons)
6626 			continue;
6627 
6628 		/*
6629 		 * Loop until we successfully set the suspend bit in
6630 		 * the TTE.
6631 		 */
6632 again:
6633 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6634 		ASSERT(TTE_IS_VALID(&tte));
6635 
6636 		ttemod = tte;
6637 		TTE_SET_SUSPEND(&ttemod);
6638 		if (sfmmu_modifytte_try(&tte, &ttemod,
6639 		    &sfhmep->hme_tte) < 0)
6640 			goto again;
6641 
6642 		/*
6643 		 * Invalidate TSB entry
6644 		 */
6645 		hmeblkp = sfmmu_hmetohblk(sfhmep);
6646 
6647 		sfmmup = hblktosfmmu(hmeblkp);
6648 		ASSERT(sfmmup == ksfmmup);
6649 		ASSERT(!hmeblkp->hblk_shared);
6650 
6651 		addr = tte_to_vaddr(hmeblkp, tte);
6652 
6653 		/*
6654 		 * No need to make sure that the TSB for this sfmmu is
6655 		 * not being relocated since it is ksfmmup and thus it
6656 		 * will never be relocated.
6657 		 */
6658 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
6659 
6660 		/*
6661 		 * Update xcall stats
6662 		 */
6663 		cpuset = cpu_ready_set;
6664 		CPUSET_DEL(cpuset, CPU->cpu_id);
6665 
6666 		/* LINTED: constant in conditional context */
6667 		SFMMU_XCALL_STATS(ksfmmup);
6668 
6669 		/*
6670 		 * Flush TLB entry on remote CPU's
6671 		 */
6672 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
6673 		    (uint64_t)ksfmmup);
6674 		xt_sync(cpuset);
6675 
6676 		/*
6677 		 * Flush TLB entry on local CPU
6678 		 */
6679 		vtag_flushpage(addr, (uint64_t)ksfmmup);
6680 	}
6681 
6682 	while (index != 0) {
6683 		index = index >> 1;
6684 		if (index != 0)
6685 			cons++;
6686 		if (index & 0x1) {
6687 			pp = PP_GROUPLEADER(pp, cons);
6688 			goto retry;
6689 		}
6690 	}
6691 }
6692 
6693 #ifdef	DEBUG
6694 
6695 #define	N_PRLE	1024
6696 struct prle {
6697 	page_t *targ;
6698 	page_t *repl;
6699 	int status;
6700 	int pausecpus;
6701 	hrtime_t whence;
6702 };
6703 
6704 static struct prle page_relocate_log[N_PRLE];
6705 static int prl_entry;
6706 static kmutex_t prl_mutex;
6707 
6708 #define	PAGE_RELOCATE_LOG(t, r, s, p)					\
6709 	mutex_enter(&prl_mutex);					\
6710 	page_relocate_log[prl_entry].targ = *(t);			\
6711 	page_relocate_log[prl_entry].repl = *(r);			\
6712 	page_relocate_log[prl_entry].status = (s);			\
6713 	page_relocate_log[prl_entry].pausecpus = (p);			\
6714 	page_relocate_log[prl_entry].whence = gethrtime();		\
6715 	prl_entry = (prl_entry == (N_PRLE - 1))? 0 : prl_entry + 1;	\
6716 	mutex_exit(&prl_mutex);
6717 
6718 #else	/* !DEBUG */
6719 #define	PAGE_RELOCATE_LOG(t, r, s, p)
6720 #endif
6721 
6722 /*
6723  * Core Kernel Page Relocation Algorithm
6724  *
6725  * Input:
6726  *
6727  * target : 	constituent pages are SE_EXCL locked.
6728  * replacement:	constituent pages are SE_EXCL locked.
6729  *
6730  * Output:
6731  *
6732  * nrelocp:	number of pages relocated
6733  */
6734 int
6735 hat_page_relocate(page_t **target, page_t **replacement, spgcnt_t *nrelocp)
6736 {
6737 	page_t		*targ, *repl;
6738 	page_t		*tpp, *rpp;
6739 	kmutex_t	*low, *high;
6740 	spgcnt_t	npages, i;
6741 	page_t		*pl = NULL;
6742 	int		old_pil;
6743 	cpuset_t	cpuset;
6744 	int		cap_cpus;
6745 	int		ret;
6746 #ifdef VAC
6747 	int		cflags = 0;
6748 #endif
6749 
6750 	if (hat_kpr_enabled == 0 || !kcage_on || PP_ISNORELOC(*target)) {
6751 		PAGE_RELOCATE_LOG(target, replacement, EAGAIN, -1);
6752 		return (EAGAIN);
6753 	}
6754 
6755 	mutex_enter(&kpr_mutex);
6756 	kreloc_thread = curthread;
6757 
6758 	targ = *target;
6759 	repl = *replacement;
6760 	ASSERT(repl != NULL);
6761 	ASSERT(targ->p_szc == repl->p_szc);
6762 
6763 	npages = page_get_pagecnt(targ->p_szc);
6764 
6765 	/*
6766 	 * unload VA<->PA mappings that are not locked
6767 	 */
6768 	tpp = targ;
6769 	for (i = 0; i < npages; i++) {
6770 		(void) hat_pageunload(tpp, SFMMU_KERNEL_RELOC);
6771 		tpp++;
6772 	}
6773 
6774 	/*
6775 	 * Do "presuspend" callbacks, in a context from which we can still
6776 	 * block as needed. Note that we don't hold the mapping list lock
6777 	 * of "targ" at this point due to potential locking order issues;
6778 	 * we assume that between the hat_pageunload() above and holding
6779 	 * the SE_EXCL lock that the mapping list *cannot* change at this
6780 	 * point.
6781 	 */
6782 	ret = hat_pageprocess_precallbacks(targ, HAT_PRESUSPEND, &cap_cpus);
6783 	if (ret != 0) {
6784 		/*
6785 		 * EIO translates to fatal error, for all others cleanup
6786 		 * and return EAGAIN.
6787 		 */
6788 		ASSERT(ret != EIO);
6789 		hat_pageprocess_postcallbacks(targ, HAT_POSTUNSUSPEND);
6790 		PAGE_RELOCATE_LOG(target, replacement, ret, -1);
6791 		kreloc_thread = NULL;
6792 		mutex_exit(&kpr_mutex);
6793 		return (EAGAIN);
6794 	}
6795 
6796 	/*
6797 	 * acquire p_mapping list lock for both the target and replacement
6798 	 * root pages.
6799 	 *
6800 	 * low and high refer to the need to grab the mlist locks in a
6801 	 * specific order in order to prevent race conditions.  Thus the
6802 	 * lower lock must be grabbed before the higher lock.
6803 	 *
6804 	 * This will block hat_unload's accessing p_mapping list.  Since
6805 	 * we have SE_EXCL lock, hat_memload and hat_pageunload will be
6806 	 * blocked.  Thus, no one else will be accessing the p_mapping list
6807 	 * while we suspend and reload the locked mapping below.
6808 	 */
6809 	tpp = targ;
6810 	rpp = repl;
6811 	sfmmu_mlist_reloc_enter(tpp, rpp, &low, &high);
6812 
6813 	kpreempt_disable();
6814 
6815 	/*
6816 	 * We raise our PIL to 13 so that we don't get captured by
6817 	 * another CPU or pinned by an interrupt thread.  We can't go to
6818 	 * PIL 14 since the nexus driver(s) may need to interrupt at
6819 	 * that level in the case of IOMMU pseudo mappings.
6820 	 */
6821 	cpuset = cpu_ready_set;
6822 	CPUSET_DEL(cpuset, CPU->cpu_id);
6823 	if (!cap_cpus || CPUSET_ISNULL(cpuset)) {
6824 		old_pil = splr(XCALL_PIL);
6825 	} else {
6826 		old_pil = -1;
6827 		xc_attention(cpuset);
6828 	}
6829 	ASSERT(getpil() == XCALL_PIL);
6830 
6831 	/*
6832 	 * Now do suspend callbacks. In the case of an IOMMU mapping
6833 	 * this will suspend all DMA activity to the page while it is
6834 	 * being relocated. Since we are well above LOCK_LEVEL and CPUs
6835 	 * may be captured at this point we should have acquired any needed
6836 	 * locks in the presuspend callback.
6837 	 */
6838 	ret = hat_pageprocess_precallbacks(targ, HAT_SUSPEND, NULL);
6839 	if (ret != 0) {
6840 		repl = targ;
6841 		goto suspend_fail;
6842 	}
6843 
6844 	/*
6845 	 * Raise the PIL yet again, this time to block all high-level
6846 	 * interrupts on this CPU. This is necessary to prevent an
6847 	 * interrupt routine from pinning the thread which holds the
6848 	 * mapping suspended and then touching the suspended page.
6849 	 *
6850 	 * Once the page is suspended we also need to be careful to
6851 	 * avoid calling any functions which touch any seg_kmem memory
6852 	 * since that memory may be backed by the very page we are
6853 	 * relocating in here!
6854 	 */
6855 	hat_pagesuspend(targ);
6856 
6857 	/*
6858 	 * Now that we are confident everybody has stopped using this page,
6859 	 * copy the page contents.  Note we use a physical copy to prevent
6860 	 * locking issues and to avoid fpRAS because we can't handle it in
6861 	 * this context.
6862 	 */
6863 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6864 #ifdef VAC
6865 		/*
6866 		 * If the replacement has a different vcolor than
6867 		 * the one being replacd, we need to handle VAC
6868 		 * consistency for it just as we were setting up
6869 		 * a new mapping to it.
6870 		 */
6871 		if ((PP_GET_VCOLOR(rpp) != NO_VCOLOR) &&
6872 		    (tpp->p_vcolor != rpp->p_vcolor) &&
6873 		    !CacheColor_IsFlushed(cflags, PP_GET_VCOLOR(rpp))) {
6874 			CacheColor_SetFlushed(cflags, PP_GET_VCOLOR(rpp));
6875 			sfmmu_cache_flushcolor(PP_GET_VCOLOR(rpp),
6876 			    rpp->p_pagenum);
6877 		}
6878 #endif
6879 		/*
6880 		 * Copy the contents of the page.
6881 		 */
6882 		ppcopy_kernel(tpp, rpp);
6883 	}
6884 
6885 	tpp = targ;
6886 	rpp = repl;
6887 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6888 		/*
6889 		 * Copy attributes.  VAC consistency was handled above,
6890 		 * if required.
6891 		 */
6892 		rpp->p_nrm = tpp->p_nrm;
6893 		tpp->p_nrm = 0;
6894 		rpp->p_index = tpp->p_index;
6895 		tpp->p_index = 0;
6896 #ifdef VAC
6897 		rpp->p_vcolor = tpp->p_vcolor;
6898 #endif
6899 	}
6900 
6901 	/*
6902 	 * First, unsuspend the page, if we set the suspend bit, and transfer
6903 	 * the mapping list from the target page to the replacement page.
6904 	 * Next process postcallbacks; since pa_hment's are linked only to the
6905 	 * p_mapping list of root page, we don't iterate over the constituent
6906 	 * pages.
6907 	 */
6908 	hat_pagereload(targ, repl);
6909 
6910 suspend_fail:
6911 	hat_pageprocess_postcallbacks(repl, HAT_UNSUSPEND);
6912 
6913 	/*
6914 	 * Now lower our PIL and release any captured CPUs since we
6915 	 * are out of the "danger zone".  After this it will again be
6916 	 * safe to acquire adaptive mutex locks, or to drop them...
6917 	 */
6918 	if (old_pil != -1) {
6919 		splx(old_pil);
6920 	} else {
6921 		xc_dismissed(cpuset);
6922 	}
6923 
6924 	kpreempt_enable();
6925 
6926 	sfmmu_mlist_reloc_exit(low, high);
6927 
6928 	/*
6929 	 * Postsuspend callbacks should drop any locks held across
6930 	 * the suspend callbacks.  As before, we don't hold the mapping
6931 	 * list lock at this point.. our assumption is that the mapping
6932 	 * list still can't change due to our holding SE_EXCL lock and
6933 	 * there being no unlocked mappings left. Hence the restriction
6934 	 * on calling context to hat_delete_callback()
6935 	 */
6936 	hat_pageprocess_postcallbacks(repl, HAT_POSTUNSUSPEND);
6937 	if (ret != 0) {
6938 		/*
6939 		 * The second presuspend call failed: we got here through
6940 		 * the suspend_fail label above.
6941 		 */
6942 		ASSERT(ret != EIO);
6943 		PAGE_RELOCATE_LOG(target, replacement, ret, cap_cpus);
6944 		kreloc_thread = NULL;
6945 		mutex_exit(&kpr_mutex);
6946 		return (EAGAIN);
6947 	}
6948 
6949 	/*
6950 	 * Now that we're out of the performance critical section we can
6951 	 * take care of updating the hash table, since we still
6952 	 * hold all the pages locked SE_EXCL at this point we
6953 	 * needn't worry about things changing out from under us.
6954 	 */
6955 	tpp = targ;
6956 	rpp = repl;
6957 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6958 
6959 		/*
6960 		 * replace targ with replacement in page_hash table
6961 		 */
6962 		targ = tpp;
6963 		page_relocate_hash(rpp, targ);
6964 
6965 		/*
6966 		 * concatenate target; caller of platform_page_relocate()
6967 		 * expects target to be concatenated after returning.
6968 		 */
6969 		ASSERT(targ->p_next == targ);
6970 		ASSERT(targ->p_prev == targ);
6971 		page_list_concat(&pl, &targ);
6972 	}
6973 
6974 	ASSERT(*target == pl);
6975 	*nrelocp = npages;
6976 	PAGE_RELOCATE_LOG(target, replacement, 0, cap_cpus);
6977 	kreloc_thread = NULL;
6978 	mutex_exit(&kpr_mutex);
6979 	return (0);
6980 }
6981 
6982 /*
6983  * Called when stray pa_hments are found attached to a page which is
6984  * being freed.  Notify the subsystem which attached the pa_hment of
6985  * the error if it registered a suitable handler, else panic.
6986  */
6987 static void
6988 sfmmu_pahment_leaked(struct pa_hment *pahmep)
6989 {
6990 	id_t cb_id = pahmep->cb_id;
6991 
6992 	ASSERT(cb_id >= (id_t)0 && cb_id < sfmmu_cb_nextid);
6993 	if (sfmmu_cb_table[cb_id].errhandler != NULL) {
6994 		if (sfmmu_cb_table[cb_id].errhandler(pahmep->addr, pahmep->len,
6995 		    HAT_CB_ERR_LEAKED, pahmep->pvt) == 0)
6996 			return;		/* non-fatal */
6997 	}
6998 	panic("pa_hment leaked: 0x%p", (void *)pahmep);
6999 }
7000 
7001 /*
7002  * Remove all mappings to page 'pp'.
7003  */
7004 int
7005 hat_pageunload(struct page *pp, uint_t forceflag)
7006 {
7007 	struct page *origpp = pp;
7008 	struct sf_hment *sfhme, *tmphme;
7009 	struct hme_blk *hmeblkp;
7010 	kmutex_t *pml;
7011 #ifdef VAC
7012 	kmutex_t *pmtx;
7013 #endif
7014 	cpuset_t cpuset, tset;
7015 	int index, cons;
7016 	int xhme_blks;
7017 	int pa_hments;
7018 
7019 	ASSERT(PAGE_EXCL(pp));
7020 
7021 retry_xhat:
7022 	tmphme = NULL;
7023 	xhme_blks = 0;
7024 	pa_hments = 0;
7025 	CPUSET_ZERO(cpuset);
7026 
7027 	pml = sfmmu_mlist_enter(pp);
7028 
7029 #ifdef VAC
7030 	if (pp->p_kpmref)
7031 		sfmmu_kpm_pageunload(pp);
7032 	ASSERT(!PP_ISMAPPED_KPM(pp));
7033 #endif
7034 	/*
7035 	 * Clear vpm reference. Since the page is exclusively locked
7036 	 * vpm cannot be referencing it.
7037 	 */
7038 	if (vpm_enable) {
7039 		pp->p_vpmref = 0;
7040 	}
7041 
7042 	index = PP_MAPINDEX(pp);
7043 	cons = TTE8K;
7044 retry:
7045 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7046 		tmphme = sfhme->hme_next;
7047 
7048 		if (IS_PAHME(sfhme)) {
7049 			ASSERT(sfhme->hme_data != NULL);
7050 			pa_hments++;
7051 			continue;
7052 		}
7053 
7054 		hmeblkp = sfmmu_hmetohblk(sfhme);
7055 		if (hmeblkp->hblk_xhat_bit) {
7056 			struct xhat_hme_blk *xblk =
7057 			    (struct xhat_hme_blk *)hmeblkp;
7058 
7059 			(void) XHAT_PAGEUNLOAD(xblk->xhat_hme_blk_hat,
7060 			    pp, forceflag, XBLK2PROVBLK(xblk));
7061 
7062 			xhme_blks = 1;
7063 			continue;
7064 		}
7065 
7066 		/*
7067 		 * If there are kernel mappings don't unload them, they will
7068 		 * be suspended.
7069 		 */
7070 		if (forceflag == SFMMU_KERNEL_RELOC && hmeblkp->hblk_lckcnt &&
7071 		    hmeblkp->hblk_tag.htag_id == ksfmmup)
7072 			continue;
7073 
7074 		tset = sfmmu_pageunload(pp, sfhme, cons);
7075 		CPUSET_OR(cpuset, tset);
7076 	}
7077 
7078 	while (index != 0) {
7079 		index = index >> 1;
7080 		if (index != 0)
7081 			cons++;
7082 		if (index & 0x1) {
7083 			/* Go to leading page */
7084 			pp = PP_GROUPLEADER(pp, cons);
7085 			ASSERT(sfmmu_mlist_held(pp));
7086 			goto retry;
7087 		}
7088 	}
7089 
7090 	/*
7091 	 * cpuset may be empty if the page was only mapped by segkpm,
7092 	 * in which case we won't actually cross-trap.
7093 	 */
7094 	xt_sync(cpuset);
7095 
7096 	/*
7097 	 * The page should have no mappings at this point, unless
7098 	 * we were called from hat_page_relocate() in which case we
7099 	 * leave the locked mappings which will be suspended later.
7100 	 */
7101 	ASSERT(!PP_ISMAPPED(origpp) || xhme_blks || pa_hments ||
7102 	    (forceflag == SFMMU_KERNEL_RELOC));
7103 
7104 #ifdef VAC
7105 	if (PP_ISTNC(pp)) {
7106 		if (cons == TTE8K) {
7107 			pmtx = sfmmu_page_enter(pp);
7108 			PP_CLRTNC(pp);
7109 			sfmmu_page_exit(pmtx);
7110 		} else {
7111 			conv_tnc(pp, cons);
7112 		}
7113 	}
7114 #endif	/* VAC */
7115 
7116 	if (pa_hments && forceflag != SFMMU_KERNEL_RELOC) {
7117 		/*
7118 		 * Unlink any pa_hments and free them, calling back
7119 		 * the responsible subsystem to notify it of the error.
7120 		 * This can occur in situations such as drivers leaking
7121 		 * DMA handles: naughty, but common enough that we'd like
7122 		 * to keep the system running rather than bringing it
7123 		 * down with an obscure error like "pa_hment leaked"
7124 		 * which doesn't aid the user in debugging their driver.
7125 		 */
7126 		for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7127 			tmphme = sfhme->hme_next;
7128 			if (IS_PAHME(sfhme)) {
7129 				struct pa_hment *pahmep = sfhme->hme_data;
7130 				sfmmu_pahment_leaked(pahmep);
7131 				HME_SUB(sfhme, pp);
7132 				kmem_cache_free(pa_hment_cache, pahmep);
7133 			}
7134 		}
7135 
7136 		ASSERT(!PP_ISMAPPED(origpp) || xhme_blks);
7137 	}
7138 
7139 	sfmmu_mlist_exit(pml);
7140 
7141 	/*
7142 	 * XHAT may not have finished unloading pages
7143 	 * because some other thread was waiting for
7144 	 * mlist lock and XHAT_PAGEUNLOAD let it do
7145 	 * the job.
7146 	 */
7147 	if (xhme_blks) {
7148 		pp = origpp;
7149 		goto retry_xhat;
7150 	}
7151 
7152 	return (0);
7153 }
7154 
7155 cpuset_t
7156 sfmmu_pageunload(page_t *pp, struct sf_hment *sfhme, int cons)
7157 {
7158 	struct hme_blk *hmeblkp;
7159 	sfmmu_t *sfmmup;
7160 	tte_t tte, ttemod;
7161 #ifdef DEBUG
7162 	tte_t orig_old;
7163 #endif /* DEBUG */
7164 	caddr_t addr;
7165 	int ttesz;
7166 	int ret;
7167 	cpuset_t cpuset;
7168 
7169 	ASSERT(pp != NULL);
7170 	ASSERT(sfmmu_mlist_held(pp));
7171 	ASSERT(!PP_ISKAS(pp));
7172 
7173 	CPUSET_ZERO(cpuset);
7174 
7175 	hmeblkp = sfmmu_hmetohblk(sfhme);
7176 
7177 readtte:
7178 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7179 	if (TTE_IS_VALID(&tte)) {
7180 		sfmmup = hblktosfmmu(hmeblkp);
7181 		ttesz = get_hblk_ttesz(hmeblkp);
7182 		/*
7183 		 * Only unload mappings of 'cons' size.
7184 		 */
7185 		if (ttesz != cons)
7186 			return (cpuset);
7187 
7188 		/*
7189 		 * Note that we have p_mapping lock, but no hash lock here.
7190 		 * hblk_unload() has to have both hash lock AND p_mapping
7191 		 * lock before it tries to modify tte. So, the tte could
7192 		 * not become invalid in the sfmmu_modifytte_try() below.
7193 		 */
7194 		ttemod = tte;
7195 #ifdef DEBUG
7196 		orig_old = tte;
7197 #endif /* DEBUG */
7198 
7199 		TTE_SET_INVALID(&ttemod);
7200 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7201 		if (ret < 0) {
7202 #ifdef DEBUG
7203 			/* only R/M bits can change. */
7204 			chk_tte(&orig_old, &tte, &ttemod, hmeblkp);
7205 #endif /* DEBUG */
7206 			goto readtte;
7207 		}
7208 
7209 		if (ret == 0) {
7210 			panic("pageunload: cas failed?");
7211 		}
7212 
7213 		addr = tte_to_vaddr(hmeblkp, tte);
7214 
7215 		if (hmeblkp->hblk_shared) {
7216 			sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7217 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
7218 			sf_region_t *rgnp;
7219 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7220 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7221 			ASSERT(srdp != NULL);
7222 			rgnp = srdp->srd_hmergnp[rid];
7223 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
7224 			cpuset = sfmmu_rgntlb_demap(addr, rgnp, hmeblkp, 1);
7225 			sfmmu_ttesync(NULL, addr, &tte, pp);
7226 			ASSERT(rgnp->rgn_ttecnt[ttesz] > 0);
7227 			atomic_add_long(&rgnp->rgn_ttecnt[ttesz], -1);
7228 		} else {
7229 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
7230 			atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -1);
7231 
7232 			/*
7233 			 * We need to flush the page from the virtual cache
7234 			 * in order to prevent a virtual cache alias
7235 			 * inconsistency. The particular scenario we need
7236 			 * to worry about is:
7237 			 * Given:  va1 and va2 are two virtual address that
7238 			 * alias and will map the same physical address.
7239 			 * 1.   mapping exists from va1 to pa and data has
7240 			 *	been read into the cache.
7241 			 * 2.   unload va1.
7242 			 * 3.   load va2 and modify data using va2.
7243 			 * 4    unload va2.
7244 			 * 5.   load va1 and reference data.  Unless we flush
7245 			 *	the data cache when we unload we will get
7246 			 *	stale data.
7247 			 * This scenario is taken care of by using virtual
7248 			 * page coloring.
7249 			 */
7250 			if (sfmmup->sfmmu_ismhat) {
7251 				/*
7252 				 * Flush TSBs, TLBs and caches
7253 				 * of every process
7254 				 * sharing this ism segment.
7255 				 */
7256 				sfmmu_hat_lock_all();
7257 				mutex_enter(&ism_mlist_lock);
7258 				kpreempt_disable();
7259 				sfmmu_ismtlbcache_demap(addr, sfmmup, hmeblkp,
7260 				    pp->p_pagenum, CACHE_NO_FLUSH);
7261 				kpreempt_enable();
7262 				mutex_exit(&ism_mlist_lock);
7263 				sfmmu_hat_unlock_all();
7264 				cpuset = cpu_ready_set;
7265 			} else {
7266 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7267 				cpuset = sfmmup->sfmmu_cpusran;
7268 			}
7269 		}
7270 
7271 		/*
7272 		 * Hme_sub has to run after ttesync() and a_rss update.
7273 		 * See hblk_unload().
7274 		 */
7275 		HME_SUB(sfhme, pp);
7276 		membar_stst();
7277 
7278 		/*
7279 		 * We can not make ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
7280 		 * since pteload may have done a HME_ADD() right after
7281 		 * we did the HME_SUB() above. Hmecnt is now maintained
7282 		 * by cas only. no lock guranteed its value. The only
7283 		 * gurantee we have is the hmecnt should not be less than
7284 		 * what it should be so the hblk will not be taken away.
7285 		 * It's also important that we decremented the hmecnt after
7286 		 * we are done with hmeblkp so that this hmeblk won't be
7287 		 * stolen.
7288 		 */
7289 		ASSERT(hmeblkp->hblk_hmecnt > 0);
7290 		ASSERT(hmeblkp->hblk_vcnt > 0);
7291 		atomic_add_16(&hmeblkp->hblk_vcnt, -1);
7292 		atomic_add_16(&hmeblkp->hblk_hmecnt, -1);
7293 		/*
7294 		 * This is bug 4063182.
7295 		 * XXX: fixme
7296 		 * ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
7297 		 *	!hmeblkp->hblk_lckcnt);
7298 		 */
7299 	} else {
7300 		panic("invalid tte? pp %p &tte %p",
7301 		    (void *)pp, (void *)&tte);
7302 	}
7303 
7304 	return (cpuset);
7305 }
7306 
7307 /*
7308  * While relocating a kernel page, this function will move the mappings
7309  * from tpp to dpp and modify any associated data with these mappings.
7310  * It also unsuspends the suspended kernel mapping.
7311  */
7312 static void
7313 hat_pagereload(struct page *tpp, struct page *dpp)
7314 {
7315 	struct sf_hment *sfhme;
7316 	tte_t tte, ttemod;
7317 	int index, cons;
7318 
7319 	ASSERT(getpil() == PIL_MAX);
7320 	ASSERT(sfmmu_mlist_held(tpp));
7321 	ASSERT(sfmmu_mlist_held(dpp));
7322 
7323 	index = PP_MAPINDEX(tpp);
7324 	cons = TTE8K;
7325 
7326 	/* Update real mappings to the page */
7327 retry:
7328 	for (sfhme = tpp->p_mapping; sfhme != NULL; sfhme = sfhme->hme_next) {
7329 		if (IS_PAHME(sfhme))
7330 			continue;
7331 		sfmmu_copytte(&sfhme->hme_tte, &tte);
7332 		ttemod = tte;
7333 
7334 		/*
7335 		 * replace old pfn with new pfn in TTE
7336 		 */
7337 		PFN_TO_TTE(ttemod, dpp->p_pagenum);
7338 
7339 		/*
7340 		 * clear suspend bit
7341 		 */
7342 		ASSERT(TTE_IS_SUSPEND(&ttemod));
7343 		TTE_CLR_SUSPEND(&ttemod);
7344 
7345 		if (sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte) < 0)
7346 			panic("hat_pagereload(): sfmmu_modifytte_try() failed");
7347 
7348 		/*
7349 		 * set hme_page point to new page
7350 		 */
7351 		sfhme->hme_page = dpp;
7352 	}
7353 
7354 	/*
7355 	 * move p_mapping list from old page to new page
7356 	 */
7357 	dpp->p_mapping = tpp->p_mapping;
7358 	tpp->p_mapping = NULL;
7359 	dpp->p_share = tpp->p_share;
7360 	tpp->p_share = 0;
7361 
7362 	while (index != 0) {
7363 		index = index >> 1;
7364 		if (index != 0)
7365 			cons++;
7366 		if (index & 0x1) {
7367 			tpp = PP_GROUPLEADER(tpp, cons);
7368 			dpp = PP_GROUPLEADER(dpp, cons);
7369 			goto retry;
7370 		}
7371 	}
7372 
7373 	curthread->t_flag &= ~T_DONTDTRACE;
7374 	mutex_exit(&kpr_suspendlock);
7375 }
7376 
7377 uint_t
7378 hat_pagesync(struct page *pp, uint_t clearflag)
7379 {
7380 	struct sf_hment *sfhme, *tmphme = NULL;
7381 	struct hme_blk *hmeblkp;
7382 	kmutex_t *pml;
7383 	cpuset_t cpuset, tset;
7384 	int	index, cons;
7385 	extern	ulong_t po_share;
7386 	page_t	*save_pp = pp;
7387 	int	stop_on_sh = 0;
7388 	uint_t	shcnt;
7389 
7390 	CPUSET_ZERO(cpuset);
7391 
7392 	if (PP_ISRO(pp) && (clearflag & HAT_SYNC_STOPON_MOD)) {
7393 		return (PP_GENERIC_ATTR(pp));
7394 	}
7395 
7396 	if ((clearflag & HAT_SYNC_ZERORM) == 0) {
7397 		if ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(pp)) {
7398 			return (PP_GENERIC_ATTR(pp));
7399 		}
7400 		if ((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(pp)) {
7401 			return (PP_GENERIC_ATTR(pp));
7402 		}
7403 		if (clearflag & HAT_SYNC_STOPON_SHARED) {
7404 			if (pp->p_share > po_share) {
7405 				hat_page_setattr(pp, P_REF);
7406 				return (PP_GENERIC_ATTR(pp));
7407 			}
7408 			stop_on_sh = 1;
7409 			shcnt = 0;
7410 		}
7411 	}
7412 
7413 	clearflag &= ~HAT_SYNC_STOPON_SHARED;
7414 	pml = sfmmu_mlist_enter(pp);
7415 	index = PP_MAPINDEX(pp);
7416 	cons = TTE8K;
7417 retry:
7418 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7419 		/*
7420 		 * We need to save the next hment on the list since
7421 		 * it is possible for pagesync to remove an invalid hment
7422 		 * from the list.
7423 		 */
7424 		tmphme = sfhme->hme_next;
7425 		if (IS_PAHME(sfhme))
7426 			continue;
7427 		/*
7428 		 * If we are looking for large mappings and this hme doesn't
7429 		 * reach the range we are seeking, just ignore it.
7430 		 */
7431 		hmeblkp = sfmmu_hmetohblk(sfhme);
7432 		if (hmeblkp->hblk_xhat_bit)
7433 			continue;
7434 
7435 		if (hme_size(sfhme) < cons)
7436 			continue;
7437 
7438 		if (stop_on_sh) {
7439 			if (hmeblkp->hblk_shared) {
7440 				sf_srd_t *srdp = hblktosrd(hmeblkp);
7441 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7442 				sf_region_t *rgnp;
7443 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7444 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7445 				ASSERT(srdp != NULL);
7446 				rgnp = srdp->srd_hmergnp[rid];
7447 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
7448 				    rgnp, rid);
7449 				shcnt += rgnp->rgn_refcnt;
7450 			} else {
7451 				shcnt++;
7452 			}
7453 			if (shcnt > po_share) {
7454 				/*
7455 				 * tell the pager to spare the page this time
7456 				 * around.
7457 				 */
7458 				hat_page_setattr(save_pp, P_REF);
7459 				index = 0;
7460 				break;
7461 			}
7462 		}
7463 		tset = sfmmu_pagesync(pp, sfhme,
7464 		    clearflag & ~HAT_SYNC_STOPON_RM);
7465 		CPUSET_OR(cpuset, tset);
7466 
7467 		/*
7468 		 * If clearflag is HAT_SYNC_DONTZERO, break out as soon
7469 		 * as the "ref" or "mod" is set or share cnt exceeds po_share.
7470 		 */
7471 		if ((clearflag & ~HAT_SYNC_STOPON_RM) == HAT_SYNC_DONTZERO &&
7472 		    (((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(save_pp)) ||
7473 		    ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(save_pp)))) {
7474 			index = 0;
7475 			break;
7476 		}
7477 	}
7478 
7479 	while (index) {
7480 		index = index >> 1;
7481 		cons++;
7482 		if (index & 0x1) {
7483 			/* Go to leading page */
7484 			pp = PP_GROUPLEADER(pp, cons);
7485 			goto retry;
7486 		}
7487 	}
7488 
7489 	xt_sync(cpuset);
7490 	sfmmu_mlist_exit(pml);
7491 	return (PP_GENERIC_ATTR(save_pp));
7492 }
7493 
7494 /*
7495  * Get all the hardware dependent attributes for a page struct
7496  */
7497 static cpuset_t
7498 sfmmu_pagesync(struct page *pp, struct sf_hment *sfhme,
7499 	uint_t clearflag)
7500 {
7501 	caddr_t addr;
7502 	tte_t tte, ttemod;
7503 	struct hme_blk *hmeblkp;
7504 	int ret;
7505 	sfmmu_t *sfmmup;
7506 	cpuset_t cpuset;
7507 
7508 	ASSERT(pp != NULL);
7509 	ASSERT(sfmmu_mlist_held(pp));
7510 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
7511 	    (clearflag == HAT_SYNC_ZERORM));
7512 
7513 	SFMMU_STAT(sf_pagesync);
7514 
7515 	CPUSET_ZERO(cpuset);
7516 
7517 sfmmu_pagesync_retry:
7518 
7519 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7520 	if (TTE_IS_VALID(&tte)) {
7521 		hmeblkp = sfmmu_hmetohblk(sfhme);
7522 		sfmmup = hblktosfmmu(hmeblkp);
7523 		addr = tte_to_vaddr(hmeblkp, tte);
7524 		if (clearflag == HAT_SYNC_ZERORM) {
7525 			ttemod = tte;
7526 			TTE_CLR_RM(&ttemod);
7527 			ret = sfmmu_modifytte_try(&tte, &ttemod,
7528 			    &sfhme->hme_tte);
7529 			if (ret < 0) {
7530 				/*
7531 				 * cas failed and the new value is not what
7532 				 * we want.
7533 				 */
7534 				goto sfmmu_pagesync_retry;
7535 			}
7536 
7537 			if (ret > 0) {
7538 				/* we win the cas */
7539 				if (hmeblkp->hblk_shared) {
7540 					sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7541 					uint_t rid =
7542 					    hmeblkp->hblk_tag.htag_rid;
7543 					sf_region_t *rgnp;
7544 					ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7545 					ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7546 					ASSERT(srdp != NULL);
7547 					rgnp = srdp->srd_hmergnp[rid];
7548 					SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7549 					    srdp, rgnp, rid);
7550 					cpuset = sfmmu_rgntlb_demap(addr,
7551 					    rgnp, hmeblkp, 1);
7552 				} else {
7553 					sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
7554 					    0, 0);
7555 					cpuset = sfmmup->sfmmu_cpusran;
7556 				}
7557 			}
7558 		}
7559 		sfmmu_ttesync(hmeblkp->hblk_shared ? NULL : sfmmup, addr,
7560 		    &tte, pp);
7561 	}
7562 	return (cpuset);
7563 }
7564 
7565 /*
7566  * Remove write permission from a mappings to a page, so that
7567  * we can detect the next modification of it. This requires modifying
7568  * the TTE then invalidating (demap) any TLB entry using that TTE.
7569  * This code is similar to sfmmu_pagesync().
7570  */
7571 static cpuset_t
7572 sfmmu_pageclrwrt(struct page *pp, struct sf_hment *sfhme)
7573 {
7574 	caddr_t addr;
7575 	tte_t tte;
7576 	tte_t ttemod;
7577 	struct hme_blk *hmeblkp;
7578 	int ret;
7579 	sfmmu_t *sfmmup;
7580 	cpuset_t cpuset;
7581 
7582 	ASSERT(pp != NULL);
7583 	ASSERT(sfmmu_mlist_held(pp));
7584 
7585 	CPUSET_ZERO(cpuset);
7586 	SFMMU_STAT(sf_clrwrt);
7587 
7588 retry:
7589 
7590 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7591 	if (TTE_IS_VALID(&tte) && TTE_IS_WRITABLE(&tte)) {
7592 		hmeblkp = sfmmu_hmetohblk(sfhme);
7593 
7594 		/*
7595 		 * xhat mappings should never be to a VMODSORT page.
7596 		 */
7597 		ASSERT(hmeblkp->hblk_xhat_bit == 0);
7598 
7599 		sfmmup = hblktosfmmu(hmeblkp);
7600 		addr = tte_to_vaddr(hmeblkp, tte);
7601 
7602 		ttemod = tte;
7603 		TTE_CLR_WRT(&ttemod);
7604 		TTE_CLR_MOD(&ttemod);
7605 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7606 
7607 		/*
7608 		 * if cas failed and the new value is not what
7609 		 * we want retry
7610 		 */
7611 		if (ret < 0)
7612 			goto retry;
7613 
7614 		/* we win the cas */
7615 		if (ret > 0) {
7616 			if (hmeblkp->hblk_shared) {
7617 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7618 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7619 				sf_region_t *rgnp;
7620 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7621 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7622 				ASSERT(srdp != NULL);
7623 				rgnp = srdp->srd_hmergnp[rid];
7624 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7625 				    srdp, rgnp, rid);
7626 				cpuset = sfmmu_rgntlb_demap(addr,
7627 				    rgnp, hmeblkp, 1);
7628 			} else {
7629 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7630 				cpuset = sfmmup->sfmmu_cpusran;
7631 			}
7632 		}
7633 	}
7634 
7635 	return (cpuset);
7636 }
7637 
7638 /*
7639  * Walk all mappings of a page, removing write permission and clearing the
7640  * ref/mod bits. This code is similar to hat_pagesync()
7641  */
7642 static void
7643 hat_page_clrwrt(page_t *pp)
7644 {
7645 	struct sf_hment *sfhme;
7646 	struct sf_hment *tmphme = NULL;
7647 	kmutex_t *pml;
7648 	cpuset_t cpuset;
7649 	cpuset_t tset;
7650 	int	index;
7651 	int	 cons;
7652 
7653 	CPUSET_ZERO(cpuset);
7654 
7655 	pml = sfmmu_mlist_enter(pp);
7656 	index = PP_MAPINDEX(pp);
7657 	cons = TTE8K;
7658 retry:
7659 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7660 		tmphme = sfhme->hme_next;
7661 
7662 		/*
7663 		 * If we are looking for large mappings and this hme doesn't
7664 		 * reach the range we are seeking, just ignore its.
7665 		 */
7666 
7667 		if (hme_size(sfhme) < cons)
7668 			continue;
7669 
7670 		tset = sfmmu_pageclrwrt(pp, sfhme);
7671 		CPUSET_OR(cpuset, tset);
7672 	}
7673 
7674 	while (index) {
7675 		index = index >> 1;
7676 		cons++;
7677 		if (index & 0x1) {
7678 			/* Go to leading page */
7679 			pp = PP_GROUPLEADER(pp, cons);
7680 			goto retry;
7681 		}
7682 	}
7683 
7684 	xt_sync(cpuset);
7685 	sfmmu_mlist_exit(pml);
7686 }
7687 
7688 /*
7689  * Set the given REF/MOD/RO bits for the given page.
7690  * For a vnode with a sorted v_pages list, we need to change
7691  * the attributes and the v_pages list together under page_vnode_mutex.
7692  */
7693 void
7694 hat_page_setattr(page_t *pp, uint_t flag)
7695 {
7696 	vnode_t		*vp = pp->p_vnode;
7697 	page_t		**listp;
7698 	kmutex_t	*pmtx;
7699 	kmutex_t	*vphm = NULL;
7700 	int		noshuffle;
7701 
7702 	noshuffle = flag & P_NSH;
7703 	flag &= ~P_NSH;
7704 
7705 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7706 
7707 	/*
7708 	 * nothing to do if attribute already set
7709 	 */
7710 	if ((pp->p_nrm & flag) == flag)
7711 		return;
7712 
7713 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp) &&
7714 	    !noshuffle) {
7715 		vphm = page_vnode_mutex(vp);
7716 		mutex_enter(vphm);
7717 	}
7718 
7719 	pmtx = sfmmu_page_enter(pp);
7720 	pp->p_nrm |= flag;
7721 	sfmmu_page_exit(pmtx);
7722 
7723 	if (vphm != NULL) {
7724 		/*
7725 		 * Some File Systems examine v_pages for NULL w/o
7726 		 * grabbing the vphm mutex. Must not let it become NULL when
7727 		 * pp is the only page on the list.
7728 		 */
7729 		if (pp->p_vpnext != pp) {
7730 			page_vpsub(&vp->v_pages, pp);
7731 			if (vp->v_pages != NULL)
7732 				listp = &vp->v_pages->p_vpprev->p_vpnext;
7733 			else
7734 				listp = &vp->v_pages;
7735 			page_vpadd(listp, pp);
7736 		}
7737 		mutex_exit(vphm);
7738 	}
7739 }
7740 
7741 void
7742 hat_page_clrattr(page_t *pp, uint_t flag)
7743 {
7744 	vnode_t		*vp = pp->p_vnode;
7745 	kmutex_t	*pmtx;
7746 
7747 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7748 
7749 	pmtx = sfmmu_page_enter(pp);
7750 
7751 	/*
7752 	 * Caller is expected to hold page's io lock for VMODSORT to work
7753 	 * correctly with pvn_vplist_dirty() and pvn_getdirty() when mod
7754 	 * bit is cleared.
7755 	 * We don't have assert to avoid tripping some existing third party
7756 	 * code. The dirty page is moved back to top of the v_page list
7757 	 * after IO is done in pvn_write_done().
7758 	 */
7759 	pp->p_nrm &= ~flag;
7760 	sfmmu_page_exit(pmtx);
7761 
7762 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) {
7763 
7764 		/*
7765 		 * VMODSORT works by removing write permissions and getting
7766 		 * a fault when a page is made dirty. At this point
7767 		 * we need to remove write permission from all mappings
7768 		 * to this page.
7769 		 */
7770 		hat_page_clrwrt(pp);
7771 	}
7772 }
7773 
7774 uint_t
7775 hat_page_getattr(page_t *pp, uint_t flag)
7776 {
7777 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7778 	return ((uint_t)(pp->p_nrm & flag));
7779 }
7780 
7781 /*
7782  * DEBUG kernels: verify that a kernel va<->pa translation
7783  * is safe by checking the underlying page_t is in a page
7784  * relocation-safe state.
7785  */
7786 #ifdef	DEBUG
7787 void
7788 sfmmu_check_kpfn(pfn_t pfn)
7789 {
7790 	page_t *pp;
7791 	int index, cons;
7792 
7793 	if (hat_check_vtop == 0)
7794 		return;
7795 
7796 	if (hat_kpr_enabled == 0 || kvseg.s_base == NULL || panicstr)
7797 		return;
7798 
7799 	pp = page_numtopp_nolock(pfn);
7800 	if (!pp)
7801 		return;
7802 
7803 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7804 		return;
7805 
7806 	/*
7807 	 * Handed a large kernel page, we dig up the root page since we
7808 	 * know the root page might have the lock also.
7809 	 */
7810 	if (pp->p_szc != 0) {
7811 		index = PP_MAPINDEX(pp);
7812 		cons = TTE8K;
7813 again:
7814 		while (index != 0) {
7815 			index >>= 1;
7816 			if (index != 0)
7817 				cons++;
7818 			if (index & 0x1) {
7819 				pp = PP_GROUPLEADER(pp, cons);
7820 				goto again;
7821 			}
7822 		}
7823 	}
7824 
7825 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7826 		return;
7827 
7828 	/*
7829 	 * Pages need to be locked or allocated "permanent" (either from
7830 	 * static_arena arena or explicitly setting PG_NORELOC when calling
7831 	 * page_create_va()) for VA->PA translations to be valid.
7832 	 */
7833 	if (!PP_ISNORELOC(pp))
7834 		panic("Illegal VA->PA translation, pp 0x%p not permanent",
7835 		    (void *)pp);
7836 	else
7837 		panic("Illegal VA->PA translation, pp 0x%p not locked",
7838 		    (void *)pp);
7839 }
7840 #endif	/* DEBUG */
7841 
7842 /*
7843  * Returns a page frame number for a given virtual address.
7844  * Returns PFN_INVALID to indicate an invalid mapping
7845  */
7846 pfn_t
7847 hat_getpfnum(struct hat *hat, caddr_t addr)
7848 {
7849 	pfn_t pfn;
7850 	tte_t tte;
7851 
7852 	/*
7853 	 * We would like to
7854 	 * ASSERT(AS_LOCK_HELD(as, &as->a_lock));
7855 	 * but we can't because the iommu driver will call this
7856 	 * routine at interrupt time and it can't grab the as lock
7857 	 * or it will deadlock: A thread could have the as lock
7858 	 * and be waiting for io.  The io can't complete
7859 	 * because the interrupt thread is blocked trying to grab
7860 	 * the as lock.
7861 	 */
7862 
7863 	ASSERT(hat->sfmmu_xhat_provider == NULL);
7864 
7865 	if (hat == ksfmmup) {
7866 		if (IS_KMEM_VA_LARGEPAGE(addr)) {
7867 			ASSERT(segkmem_lpszc > 0);
7868 			pfn = sfmmu_kvaszc2pfn(addr, segkmem_lpszc);
7869 			if (pfn != PFN_INVALID) {
7870 				sfmmu_check_kpfn(pfn);
7871 				return (pfn);
7872 			}
7873 		} else if (segkpm && IS_KPM_ADDR(addr)) {
7874 			return (sfmmu_kpm_vatopfn(addr));
7875 		}
7876 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
7877 		    == PFN_SUSPENDED) {
7878 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
7879 		}
7880 		sfmmu_check_kpfn(pfn);
7881 		return (pfn);
7882 	} else {
7883 		return (sfmmu_uvatopfn(addr, hat, NULL));
7884 	}
7885 }
7886 
7887 /*
7888  * hat_getkpfnum() is an obsolete DDI routine, and its use is discouraged.
7889  * Use hat_getpfnum(kas.a_hat, ...) instead.
7890  *
7891  * We'd like to return PFN_INVALID if the mappings have underlying page_t's
7892  * but can't right now due to the fact that some software has grown to use
7893  * this interface incorrectly. So for now when the interface is misused,
7894  * return a warning to the user that in the future it won't work in the
7895  * way they're abusing it, and carry on (after disabling page relocation).
7896  */
7897 pfn_t
7898 hat_getkpfnum(caddr_t addr)
7899 {
7900 	pfn_t pfn;
7901 	tte_t tte;
7902 	int badcaller = 0;
7903 	extern int segkmem_reloc;
7904 
7905 	if (segkpm && IS_KPM_ADDR(addr)) {
7906 		badcaller = 1;
7907 		pfn = sfmmu_kpm_vatopfn(addr);
7908 	} else {
7909 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
7910 		    == PFN_SUSPENDED) {
7911 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
7912 		}
7913 		badcaller = pf_is_memory(pfn);
7914 	}
7915 
7916 	if (badcaller) {
7917 		/*
7918 		 * We can't return PFN_INVALID or the caller may panic
7919 		 * or corrupt the system.  The only alternative is to
7920 		 * disable page relocation at this point for all kernel
7921 		 * memory.  This will impact any callers of page_relocate()
7922 		 * such as FMA or DR.
7923 		 *
7924 		 * RFE: Add junk here to spit out an ereport so the sysadmin
7925 		 * can be advised that he should upgrade his device driver
7926 		 * so that this doesn't happen.
7927 		 */
7928 		hat_getkpfnum_badcall(caller());
7929 		if (hat_kpr_enabled && segkmem_reloc) {
7930 			hat_kpr_enabled = 0;
7931 			segkmem_reloc = 0;
7932 			cmn_err(CE_WARN, "Kernel Page Relocation is DISABLED");
7933 		}
7934 	}
7935 	return (pfn);
7936 }
7937 
7938 /*
7939  * This routine will return both pfn and tte for the vaddr.
7940  */
7941 static pfn_t
7942 sfmmu_uvatopfn(caddr_t vaddr, struct hat *sfmmup, tte_t *ttep)
7943 {
7944 	struct hmehash_bucket *hmebp;
7945 	hmeblk_tag hblktag;
7946 	int hmeshift, hashno = 1;
7947 	struct hme_blk *hmeblkp = NULL;
7948 	tte_t tte;
7949 
7950 	struct sf_hment *sfhmep;
7951 	pfn_t pfn;
7952 
7953 	/* support for ISM */
7954 	ism_map_t	*ism_map;
7955 	ism_blk_t	*ism_blkp;
7956 	int		i;
7957 	sfmmu_t *ism_hatid = NULL;
7958 	sfmmu_t *locked_hatid = NULL;
7959 	sfmmu_t	*sv_sfmmup = sfmmup;
7960 	caddr_t	sv_vaddr = vaddr;
7961 	sf_srd_t *srdp;
7962 
7963 	if (ttep == NULL) {
7964 		ttep = &tte;
7965 	} else {
7966 		ttep->ll = 0;
7967 	}
7968 
7969 	ASSERT(sfmmup != ksfmmup);
7970 	SFMMU_STAT(sf_user_vtop);
7971 	/*
7972 	 * Set ism_hatid if vaddr falls in a ISM segment.
7973 	 */
7974 	ism_blkp = sfmmup->sfmmu_iblk;
7975 	if (ism_blkp != NULL) {
7976 		sfmmu_ismhat_enter(sfmmup, 0);
7977 		locked_hatid = sfmmup;
7978 	}
7979 	while (ism_blkp != NULL && ism_hatid == NULL) {
7980 		ism_map = ism_blkp->iblk_maps;
7981 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
7982 			if (vaddr >= ism_start(ism_map[i]) &&
7983 			    vaddr < ism_end(ism_map[i])) {
7984 				sfmmup = ism_hatid = ism_map[i].imap_ismhat;
7985 				vaddr = (caddr_t)(vaddr -
7986 				    ism_start(ism_map[i]));
7987 				break;
7988 			}
7989 		}
7990 		ism_blkp = ism_blkp->iblk_next;
7991 	}
7992 	if (locked_hatid) {
7993 		sfmmu_ismhat_exit(locked_hatid, 0);
7994 	}
7995 
7996 	hblktag.htag_id = sfmmup;
7997 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
7998 	do {
7999 		hmeshift = HME_HASH_SHIFT(hashno);
8000 		hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
8001 		hblktag.htag_rehash = hashno;
8002 		hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
8003 
8004 		SFMMU_HASH_LOCK(hmebp);
8005 
8006 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
8007 		if (hmeblkp != NULL) {
8008 			ASSERT(!hmeblkp->hblk_shared);
8009 			HBLKTOHME(sfhmep, hmeblkp, vaddr);
8010 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
8011 			SFMMU_HASH_UNLOCK(hmebp);
8012 			if (TTE_IS_VALID(ttep)) {
8013 				pfn = TTE_TO_PFN(vaddr, ttep);
8014 				return (pfn);
8015 			}
8016 			break;
8017 		}
8018 		SFMMU_HASH_UNLOCK(hmebp);
8019 		hashno++;
8020 	} while (HME_REHASH(sfmmup) && (hashno <= mmu_hashcnt));
8021 
8022 	if (SF_HMERGNMAP_ISNULL(sv_sfmmup)) {
8023 		return (PFN_INVALID);
8024 	}
8025 	srdp = sv_sfmmup->sfmmu_srdp;
8026 	ASSERT(srdp != NULL);
8027 	ASSERT(srdp->srd_refcnt != 0);
8028 	hblktag.htag_id = srdp;
8029 	hashno = 1;
8030 	do {
8031 		hmeshift = HME_HASH_SHIFT(hashno);
8032 		hblktag.htag_bspage = HME_HASH_BSPAGE(sv_vaddr, hmeshift);
8033 		hblktag.htag_rehash = hashno;
8034 		hmebp = HME_HASH_FUNCTION(srdp, sv_vaddr, hmeshift);
8035 
8036 		SFMMU_HASH_LOCK(hmebp);
8037 		for (hmeblkp = hmebp->hmeblkp; hmeblkp != NULL;
8038 		    hmeblkp = hmeblkp->hblk_next) {
8039 			uint_t rid;
8040 			sf_region_t *rgnp;
8041 			caddr_t rsaddr;
8042 			caddr_t readdr;
8043 
8044 			if (!HTAGS_EQ_SHME(hmeblkp->hblk_tag, hblktag,
8045 			    sv_sfmmup->sfmmu_hmeregion_map)) {
8046 				continue;
8047 			}
8048 			ASSERT(hmeblkp->hblk_shared);
8049 			rid = hmeblkp->hblk_tag.htag_rid;
8050 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8051 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8052 			rgnp = srdp->srd_hmergnp[rid];
8053 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
8054 			HBLKTOHME(sfhmep, hmeblkp, sv_vaddr);
8055 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
8056 			rsaddr = rgnp->rgn_saddr;
8057 			readdr = rsaddr + rgnp->rgn_size;
8058 #ifdef DEBUG
8059 			if (TTE_IS_VALID(ttep) ||
8060 			    get_hblk_ttesz(hmeblkp) > TTE8K) {
8061 				caddr_t eva = tte_to_evaddr(hmeblkp, ttep);
8062 				ASSERT(eva > sv_vaddr);
8063 				ASSERT(sv_vaddr >= rsaddr);
8064 				ASSERT(sv_vaddr < readdr);
8065 				ASSERT(eva <= readdr);
8066 			}
8067 #endif /* DEBUG */
8068 			/*
8069 			 * Continue the search if we
8070 			 * found an invalid 8K tte outside of the area
8071 			 * covered by this hmeblk's region.
8072 			 */
8073 			if (TTE_IS_VALID(ttep)) {
8074 				SFMMU_HASH_UNLOCK(hmebp);
8075 				pfn = TTE_TO_PFN(sv_vaddr, ttep);
8076 				return (pfn);
8077 			} else if (get_hblk_ttesz(hmeblkp) > TTE8K ||
8078 			    (sv_vaddr >= rsaddr && sv_vaddr < readdr)) {
8079 				SFMMU_HASH_UNLOCK(hmebp);
8080 				pfn = PFN_INVALID;
8081 				return (pfn);
8082 			}
8083 		}
8084 		SFMMU_HASH_UNLOCK(hmebp);
8085 		hashno++;
8086 	} while (hashno <= mmu_hashcnt);
8087 	return (PFN_INVALID);
8088 }
8089 
8090 
8091 /*
8092  * For compatability with AT&T and later optimizations
8093  */
8094 /* ARGSUSED */
8095 void
8096 hat_map(struct hat *hat, caddr_t addr, size_t len, uint_t flags)
8097 {
8098 	ASSERT(hat != NULL);
8099 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8100 }
8101 
8102 /*
8103  * Return the number of mappings to a particular page.  This number is an
8104  * approximation of the number of people sharing the page.
8105  *
8106  * shared hmeblks or ism hmeblks are counted as 1 mapping here.
8107  * hat_page_checkshare() can be used to compare threshold to share
8108  * count that reflects the number of region sharers albeit at higher cost.
8109  */
8110 ulong_t
8111 hat_page_getshare(page_t *pp)
8112 {
8113 	page_t *spp = pp;	/* start page */
8114 	kmutex_t *pml;
8115 	ulong_t	cnt;
8116 	int index, sz = TTE64K;
8117 
8118 	/*
8119 	 * We need to grab the mlist lock to make sure any outstanding
8120 	 * load/unloads complete.  Otherwise we could return zero
8121 	 * even though the unload(s) hasn't finished yet.
8122 	 */
8123 	pml = sfmmu_mlist_enter(spp);
8124 	cnt = spp->p_share;
8125 
8126 #ifdef VAC
8127 	if (kpm_enable)
8128 		cnt += spp->p_kpmref;
8129 #endif
8130 	if (vpm_enable && pp->p_vpmref) {
8131 		cnt += 1;
8132 	}
8133 
8134 	/*
8135 	 * If we have any large mappings, we count the number of
8136 	 * mappings that this large page is part of.
8137 	 */
8138 	index = PP_MAPINDEX(spp);
8139 	index >>= 1;
8140 	while (index) {
8141 		pp = PP_GROUPLEADER(spp, sz);
8142 		if ((index & 0x1) && pp != spp) {
8143 			cnt += pp->p_share;
8144 			spp = pp;
8145 		}
8146 		index >>= 1;
8147 		sz++;
8148 	}
8149 	sfmmu_mlist_exit(pml);
8150 	return (cnt);
8151 }
8152 
8153 /*
8154  * Return 1 if the number of mappings exceeds sh_thresh. Return 0
8155  * otherwise. Count shared hmeblks by region's refcnt.
8156  */
8157 int
8158 hat_page_checkshare(page_t *pp, ulong_t sh_thresh)
8159 {
8160 	kmutex_t *pml;
8161 	ulong_t	cnt = 0;
8162 	int index, sz = TTE8K;
8163 	struct sf_hment *sfhme, *tmphme = NULL;
8164 	struct hme_blk *hmeblkp;
8165 
8166 	pml = sfmmu_mlist_enter(pp);
8167 
8168 #ifdef VAC
8169 	if (kpm_enable)
8170 		cnt = pp->p_kpmref;
8171 #endif
8172 
8173 	if (vpm_enable && pp->p_vpmref) {
8174 		cnt += 1;
8175 	}
8176 
8177 	if (pp->p_share + cnt > sh_thresh) {
8178 		sfmmu_mlist_exit(pml);
8179 		return (1);
8180 	}
8181 
8182 	index = PP_MAPINDEX(pp);
8183 
8184 again:
8185 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
8186 		tmphme = sfhme->hme_next;
8187 		if (IS_PAHME(sfhme)) {
8188 			continue;
8189 		}
8190 
8191 		hmeblkp = sfmmu_hmetohblk(sfhme);
8192 		if (hmeblkp->hblk_xhat_bit) {
8193 			cnt++;
8194 			if (cnt > sh_thresh) {
8195 				sfmmu_mlist_exit(pml);
8196 				return (1);
8197 			}
8198 			continue;
8199 		}
8200 		if (hme_size(sfhme) != sz) {
8201 			continue;
8202 		}
8203 
8204 		if (hmeblkp->hblk_shared) {
8205 			sf_srd_t *srdp = hblktosrd(hmeblkp);
8206 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
8207 			sf_region_t *rgnp;
8208 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8209 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8210 			ASSERT(srdp != NULL);
8211 			rgnp = srdp->srd_hmergnp[rid];
8212 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
8213 			    rgnp, rid);
8214 			cnt += rgnp->rgn_refcnt;
8215 		} else {
8216 			cnt++;
8217 		}
8218 		if (cnt > sh_thresh) {
8219 			sfmmu_mlist_exit(pml);
8220 			return (1);
8221 		}
8222 	}
8223 
8224 	index >>= 1;
8225 	sz++;
8226 	while (index) {
8227 		pp = PP_GROUPLEADER(pp, sz);
8228 		ASSERT(sfmmu_mlist_held(pp));
8229 		if (index & 0x1) {
8230 			goto again;
8231 		}
8232 		index >>= 1;
8233 		sz++;
8234 	}
8235 	sfmmu_mlist_exit(pml);
8236 	return (0);
8237 }
8238 
8239 /*
8240  * Unload all large mappings to the pp and reset the p_szc field of every
8241  * constituent page according to the remaining mappings.
8242  *
8243  * pp must be locked SE_EXCL. Even though no other constituent pages are
8244  * locked it's legal to unload the large mappings to the pp because all
8245  * constituent pages of large locked mappings have to be locked SE_SHARED.
8246  * This means if we have SE_EXCL lock on one of constituent pages none of the
8247  * large mappings to pp are locked.
8248  *
8249  * Decrease p_szc field starting from the last constituent page and ending
8250  * with the root page. This method is used because other threads rely on the
8251  * root's p_szc to find the lock to syncronize on. After a root page_t's p_szc
8252  * is demoted then other threads will succeed in sfmmu_mlspl_enter(). This
8253  * ensures that p_szc changes of the constituent pages appears atomic for all
8254  * threads that use sfmmu_mlspl_enter() to examine p_szc field.
8255  *
8256  * This mechanism is only used for file system pages where it's not always
8257  * possible to get SE_EXCL locks on all constituent pages to demote the size
8258  * code (as is done for anonymous or kernel large pages).
8259  *
8260  * See more comments in front of sfmmu_mlspl_enter().
8261  */
8262 void
8263 hat_page_demote(page_t *pp)
8264 {
8265 	int index;
8266 	int sz;
8267 	cpuset_t cpuset;
8268 	int sync = 0;
8269 	page_t *rootpp;
8270 	struct sf_hment *sfhme;
8271 	struct sf_hment *tmphme = NULL;
8272 	struct hme_blk *hmeblkp;
8273 	uint_t pszc;
8274 	page_t *lastpp;
8275 	cpuset_t tset;
8276 	pgcnt_t npgs;
8277 	kmutex_t *pml;
8278 	kmutex_t *pmtx = NULL;
8279 
8280 	ASSERT(PAGE_EXCL(pp));
8281 	ASSERT(!PP_ISFREE(pp));
8282 	ASSERT(!PP_ISKAS(pp));
8283 	ASSERT(page_szc_lock_assert(pp));
8284 	pml = sfmmu_mlist_enter(pp);
8285 
8286 	pszc = pp->p_szc;
8287 	if (pszc == 0) {
8288 		goto out;
8289 	}
8290 
8291 	index = PP_MAPINDEX(pp) >> 1;
8292 
8293 	if (index) {
8294 		CPUSET_ZERO(cpuset);
8295 		sz = TTE64K;
8296 		sync = 1;
8297 	}
8298 
8299 	while (index) {
8300 		if (!(index & 0x1)) {
8301 			index >>= 1;
8302 			sz++;
8303 			continue;
8304 		}
8305 		ASSERT(sz <= pszc);
8306 		rootpp = PP_GROUPLEADER(pp, sz);
8307 		for (sfhme = rootpp->p_mapping; sfhme; sfhme = tmphme) {
8308 			tmphme = sfhme->hme_next;
8309 			ASSERT(!IS_PAHME(sfhme));
8310 			hmeblkp = sfmmu_hmetohblk(sfhme);
8311 			if (hme_size(sfhme) != sz) {
8312 				continue;
8313 			}
8314 			if (hmeblkp->hblk_xhat_bit) {
8315 				cmn_err(CE_PANIC,
8316 				    "hat_page_demote: xhat hmeblk");
8317 			}
8318 			tset = sfmmu_pageunload(rootpp, sfhme, sz);
8319 			CPUSET_OR(cpuset, tset);
8320 		}
8321 		if (index >>= 1) {
8322 			sz++;
8323 		}
8324 	}
8325 
8326 	ASSERT(!PP_ISMAPPED_LARGE(pp));
8327 
8328 	if (sync) {
8329 		xt_sync(cpuset);
8330 #ifdef VAC
8331 		if (PP_ISTNC(pp)) {
8332 			conv_tnc(rootpp, sz);
8333 		}
8334 #endif	/* VAC */
8335 	}
8336 
8337 	pmtx = sfmmu_page_enter(pp);
8338 
8339 	ASSERT(pp->p_szc == pszc);
8340 	rootpp = PP_PAGEROOT(pp);
8341 	ASSERT(rootpp->p_szc == pszc);
8342 	lastpp = PP_PAGENEXT_N(rootpp, TTEPAGES(pszc) - 1);
8343 
8344 	while (lastpp != rootpp) {
8345 		sz = PP_MAPINDEX(lastpp) ? fnd_mapping_sz(lastpp) : 0;
8346 		ASSERT(sz < pszc);
8347 		npgs = (sz == 0) ? 1 : TTEPAGES(sz);
8348 		ASSERT(P2PHASE(lastpp->p_pagenum, npgs) == npgs - 1);
8349 		while (--npgs > 0) {
8350 			lastpp->p_szc = (uchar_t)sz;
8351 			lastpp = PP_PAGEPREV(lastpp);
8352 		}
8353 		if (sz) {
8354 			/*
8355 			 * make sure before current root's pszc
8356 			 * is updated all updates to constituent pages pszc
8357 			 * fields are globally visible.
8358 			 */
8359 			membar_producer();
8360 		}
8361 		lastpp->p_szc = sz;
8362 		ASSERT(IS_P2ALIGNED(lastpp->p_pagenum, TTEPAGES(sz)));
8363 		if (lastpp != rootpp) {
8364 			lastpp = PP_PAGEPREV(lastpp);
8365 		}
8366 	}
8367 	if (sz == 0) {
8368 		/* the loop above doesn't cover this case */
8369 		rootpp->p_szc = 0;
8370 	}
8371 out:
8372 	ASSERT(pp->p_szc == 0);
8373 	if (pmtx != NULL) {
8374 		sfmmu_page_exit(pmtx);
8375 	}
8376 	sfmmu_mlist_exit(pml);
8377 }
8378 
8379 /*
8380  * Refresh the HAT ismttecnt[] element for size szc.
8381  * Caller must have set ISM busy flag to prevent mapping
8382  * lists from changing while we're traversing them.
8383  */
8384 pgcnt_t
8385 ism_tsb_entries(sfmmu_t *sfmmup, int szc)
8386 {
8387 	ism_blk_t	*ism_blkp = sfmmup->sfmmu_iblk;
8388 	ism_map_t	*ism_map;
8389 	pgcnt_t		npgs = 0;
8390 	pgcnt_t		npgs_scd = 0;
8391 	int		j;
8392 	sf_scd_t	*scdp;
8393 	uchar_t		rid;
8394 
8395 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
8396 	scdp = sfmmup->sfmmu_scdp;
8397 
8398 	for (; ism_blkp != NULL; ism_blkp = ism_blkp->iblk_next) {
8399 		ism_map = ism_blkp->iblk_maps;
8400 		for (j = 0; ism_map[j].imap_ismhat && j < ISM_MAP_SLOTS; j++) {
8401 			rid = ism_map[j].imap_rid;
8402 			ASSERT(rid == SFMMU_INVALID_ISMRID ||
8403 			    rid < sfmmup->sfmmu_srdp->srd_next_ismrid);
8404 
8405 			if (scdp != NULL && rid != SFMMU_INVALID_ISMRID &&
8406 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
8407 				/* ISM is in sfmmup's SCD */
8408 				npgs_scd +=
8409 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8410 			} else {
8411 				/* ISMs is not in SCD */
8412 				npgs +=
8413 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8414 			}
8415 		}
8416 	}
8417 	sfmmup->sfmmu_ismttecnt[szc] = npgs;
8418 	sfmmup->sfmmu_scdismttecnt[szc] = npgs_scd;
8419 	return (npgs);
8420 }
8421 
8422 /*
8423  * Yield the memory claim requirement for an address space.
8424  *
8425  * This is currently implemented as the number of bytes that have active
8426  * hardware translations that have page structures.  Therefore, it can
8427  * underestimate the traditional resident set size, eg, if the
8428  * physical page is present and the hardware translation is missing;
8429  * and it can overestimate the rss, eg, if there are active
8430  * translations to a frame buffer with page structs.
8431  * Also, it does not take sharing into account.
8432  *
8433  * Note that we don't acquire locks here since this function is most often
8434  * called from the clock thread.
8435  */
8436 size_t
8437 hat_get_mapped_size(struct hat *hat)
8438 {
8439 	size_t		assize = 0;
8440 	int 		i;
8441 
8442 	if (hat == NULL)
8443 		return (0);
8444 
8445 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8446 
8447 	for (i = 0; i < mmu_page_sizes; i++)
8448 		assize += ((pgcnt_t)hat->sfmmu_ttecnt[i] +
8449 		    (pgcnt_t)hat->sfmmu_scdrttecnt[i]) * TTEBYTES(i);
8450 
8451 	if (hat->sfmmu_iblk == NULL)
8452 		return (assize);
8453 
8454 	for (i = 0; i < mmu_page_sizes; i++)
8455 		assize += ((pgcnt_t)hat->sfmmu_ismttecnt[i] +
8456 		    (pgcnt_t)hat->sfmmu_scdismttecnt[i]) * TTEBYTES(i);
8457 
8458 	return (assize);
8459 }
8460 
8461 int
8462 hat_stats_enable(struct hat *hat)
8463 {
8464 	hatlock_t	*hatlockp;
8465 
8466 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8467 
8468 	hatlockp = sfmmu_hat_enter(hat);
8469 	hat->sfmmu_rmstat++;
8470 	sfmmu_hat_exit(hatlockp);
8471 	return (1);
8472 }
8473 
8474 void
8475 hat_stats_disable(struct hat *hat)
8476 {
8477 	hatlock_t	*hatlockp;
8478 
8479 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8480 
8481 	hatlockp = sfmmu_hat_enter(hat);
8482 	hat->sfmmu_rmstat--;
8483 	sfmmu_hat_exit(hatlockp);
8484 }
8485 
8486 /*
8487  * Routines for entering or removing  ourselves from the
8488  * ism_hat's mapping list. This is used for both private and
8489  * SCD hats.
8490  */
8491 static void
8492 iment_add(struct ism_ment *iment,  struct hat *ism_hat)
8493 {
8494 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8495 
8496 	iment->iment_prev = NULL;
8497 	iment->iment_next = ism_hat->sfmmu_iment;
8498 	if (ism_hat->sfmmu_iment) {
8499 		ism_hat->sfmmu_iment->iment_prev = iment;
8500 	}
8501 	ism_hat->sfmmu_iment = iment;
8502 }
8503 
8504 static void
8505 iment_sub(struct ism_ment *iment, struct hat *ism_hat)
8506 {
8507 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8508 
8509 	if (ism_hat->sfmmu_iment == NULL) {
8510 		panic("ism map entry remove - no entries");
8511 	}
8512 
8513 	if (iment->iment_prev) {
8514 		ASSERT(ism_hat->sfmmu_iment != iment);
8515 		iment->iment_prev->iment_next = iment->iment_next;
8516 	} else {
8517 		ASSERT(ism_hat->sfmmu_iment == iment);
8518 		ism_hat->sfmmu_iment = iment->iment_next;
8519 	}
8520 
8521 	if (iment->iment_next) {
8522 		iment->iment_next->iment_prev = iment->iment_prev;
8523 	}
8524 
8525 	/*
8526 	 * zero out the entry
8527 	 */
8528 	iment->iment_next = NULL;
8529 	iment->iment_prev = NULL;
8530 	iment->iment_hat =  NULL;
8531 	iment->iment_base_va = 0;
8532 }
8533 
8534 /*
8535  * Hat_share()/unshare() return an (non-zero) error
8536  * when saddr and daddr are not properly aligned.
8537  *
8538  * The top level mapping element determines the alignment
8539  * requirement for saddr and daddr, depending on different
8540  * architectures.
8541  *
8542  * When hat_share()/unshare() are not supported,
8543  * HATOP_SHARE()/UNSHARE() return 0
8544  */
8545 int
8546 hat_share(struct hat *sfmmup, caddr_t addr,
8547 	struct hat *ism_hatid, caddr_t sptaddr, size_t len, uint_t ismszc)
8548 {
8549 	ism_blk_t	*ism_blkp;
8550 	ism_blk_t	*new_iblk;
8551 	ism_map_t 	*ism_map;
8552 	ism_ment_t	*ism_ment;
8553 	int		i, added;
8554 	hatlock_t	*hatlockp;
8555 	int		reload_mmu = 0;
8556 	uint_t		ismshift = page_get_shift(ismszc);
8557 	size_t		ismpgsz = page_get_pagesize(ismszc);
8558 	uint_t		ismmask = (uint_t)ismpgsz - 1;
8559 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8560 	ushort_t	ismhatflag;
8561 	hat_region_cookie_t rcookie;
8562 	sf_scd_t	*old_scdp;
8563 
8564 #ifdef DEBUG
8565 	caddr_t		eaddr = addr + len;
8566 #endif /* DEBUG */
8567 
8568 	ASSERT(ism_hatid != NULL && sfmmup != NULL);
8569 	ASSERT(sptaddr == ISMID_STARTADDR);
8570 	/*
8571 	 * Check the alignment.
8572 	 */
8573 	if (!ISM_ALIGNED(ismshift, addr) || !ISM_ALIGNED(ismshift, sptaddr))
8574 		return (EINVAL);
8575 
8576 	/*
8577 	 * Check size alignment.
8578 	 */
8579 	if (!ISM_ALIGNED(ismshift, len))
8580 		return (EINVAL);
8581 
8582 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
8583 
8584 	/*
8585 	 * Allocate ism_ment for the ism_hat's mapping list, and an
8586 	 * ism map blk in case we need one.  We must do our
8587 	 * allocations before acquiring locks to prevent a deadlock
8588 	 * in the kmem allocator on the mapping list lock.
8589 	 */
8590 	new_iblk = kmem_cache_alloc(ism_blk_cache, KM_SLEEP);
8591 	ism_ment = kmem_cache_alloc(ism_ment_cache, KM_SLEEP);
8592 
8593 	/*
8594 	 * Serialize ISM mappings with the ISM busy flag, and also the
8595 	 * trap handlers.
8596 	 */
8597 	sfmmu_ismhat_enter(sfmmup, 0);
8598 
8599 	/*
8600 	 * Allocate an ism map blk if necessary.
8601 	 */
8602 	if (sfmmup->sfmmu_iblk == NULL) {
8603 		sfmmup->sfmmu_iblk = new_iblk;
8604 		bzero(new_iblk, sizeof (*new_iblk));
8605 		new_iblk->iblk_nextpa = (uint64_t)-1;
8606 		membar_stst();	/* make sure next ptr visible to all CPUs */
8607 		sfmmup->sfmmu_ismblkpa = va_to_pa((caddr_t)new_iblk);
8608 		reload_mmu = 1;
8609 		new_iblk = NULL;
8610 	}
8611 
8612 #ifdef DEBUG
8613 	/*
8614 	 * Make sure mapping does not already exist.
8615 	 */
8616 	ism_blkp = sfmmup->sfmmu_iblk;
8617 	while (ism_blkp != NULL) {
8618 		ism_map = ism_blkp->iblk_maps;
8619 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
8620 			if ((addr >= ism_start(ism_map[i]) &&
8621 			    addr < ism_end(ism_map[i])) ||
8622 			    eaddr > ism_start(ism_map[i]) &&
8623 			    eaddr <= ism_end(ism_map[i])) {
8624 				panic("sfmmu_share: Already mapped!");
8625 			}
8626 		}
8627 		ism_blkp = ism_blkp->iblk_next;
8628 	}
8629 #endif /* DEBUG */
8630 
8631 	ASSERT(ismszc >= TTE4M);
8632 	if (ismszc == TTE4M) {
8633 		ismhatflag = HAT_4M_FLAG;
8634 	} else if (ismszc == TTE32M) {
8635 		ismhatflag = HAT_32M_FLAG;
8636 	} else if (ismszc == TTE256M) {
8637 		ismhatflag = HAT_256M_FLAG;
8638 	}
8639 	/*
8640 	 * Add mapping to first available mapping slot.
8641 	 */
8642 	ism_blkp = sfmmup->sfmmu_iblk;
8643 	added = 0;
8644 	while (!added) {
8645 		ism_map = ism_blkp->iblk_maps;
8646 		for (i = 0; i < ISM_MAP_SLOTS; i++)  {
8647 			if (ism_map[i].imap_ismhat == NULL) {
8648 
8649 				ism_map[i].imap_ismhat = ism_hatid;
8650 				ism_map[i].imap_vb_shift = (uchar_t)ismshift;
8651 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8652 				ism_map[i].imap_hatflags = ismhatflag;
8653 				ism_map[i].imap_sz_mask = ismmask;
8654 				/*
8655 				 * imap_seg is checked in ISM_CHECK to see if
8656 				 * non-NULL, then other info assumed valid.
8657 				 */
8658 				membar_stst();
8659 				ism_map[i].imap_seg = (uintptr_t)addr | sh_size;
8660 				ism_map[i].imap_ment = ism_ment;
8661 
8662 				/*
8663 				 * Now add ourselves to the ism_hat's
8664 				 * mapping list.
8665 				 */
8666 				ism_ment->iment_hat = sfmmup;
8667 				ism_ment->iment_base_va = addr;
8668 				ism_hatid->sfmmu_ismhat = 1;
8669 				mutex_enter(&ism_mlist_lock);
8670 				iment_add(ism_ment, ism_hatid);
8671 				mutex_exit(&ism_mlist_lock);
8672 				added = 1;
8673 				break;
8674 			}
8675 		}
8676 		if (!added && ism_blkp->iblk_next == NULL) {
8677 			ism_blkp->iblk_next = new_iblk;
8678 			new_iblk = NULL;
8679 			bzero(ism_blkp->iblk_next,
8680 			    sizeof (*ism_blkp->iblk_next));
8681 			ism_blkp->iblk_next->iblk_nextpa = (uint64_t)-1;
8682 			membar_stst();
8683 			ism_blkp->iblk_nextpa =
8684 			    va_to_pa((caddr_t)ism_blkp->iblk_next);
8685 		}
8686 		ism_blkp = ism_blkp->iblk_next;
8687 	}
8688 
8689 	/*
8690 	 * After calling hat_join_region, sfmmup may join a new SCD or
8691 	 * move from the old scd to a new scd, in which case, we want to
8692 	 * shrink the sfmmup's private tsb size, i.e., pass shrink to
8693 	 * sfmmu_check_page_sizes at the end of this routine.
8694 	 */
8695 	old_scdp = sfmmup->sfmmu_scdp;
8696 
8697 	rcookie = hat_join_region(sfmmup, addr, len, (void *)ism_hatid, 0,
8698 	    PROT_ALL, ismszc, NULL, HAT_REGION_ISM);
8699 	if (rcookie != HAT_INVALID_REGION_COOKIE) {
8700 		ism_map[i].imap_rid = (uchar_t)((uint64_t)rcookie);
8701 	}
8702 	/*
8703 	 * Update our counters for this sfmmup's ism mappings.
8704 	 */
8705 	for (i = 0; i <= ismszc; i++) {
8706 		if (!(disable_ism_large_pages & (1 << i)))
8707 			(void) ism_tsb_entries(sfmmup, i);
8708 	}
8709 
8710 	/*
8711 	 * For ISM and DISM we do not support 512K pages, so we only only
8712 	 * search the 4M and 8K/64K hashes for 4 pagesize cpus, and search the
8713 	 * 256M or 32M, and 4M and 8K/64K hashes for 6 pagesize cpus.
8714 	 *
8715 	 * Need to set 32M/256M ISM flags to make sure
8716 	 * sfmmu_check_page_sizes() enables them on Panther.
8717 	 */
8718 	ASSERT((disable_ism_large_pages & (1 << TTE512K)) != 0);
8719 
8720 	switch (ismszc) {
8721 	case TTE256M:
8722 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_256M_ISM)) {
8723 			hatlockp = sfmmu_hat_enter(sfmmup);
8724 			SFMMU_FLAGS_SET(sfmmup, HAT_256M_ISM);
8725 			sfmmu_hat_exit(hatlockp);
8726 		}
8727 		break;
8728 	case TTE32M:
8729 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_32M_ISM)) {
8730 			hatlockp = sfmmu_hat_enter(sfmmup);
8731 			SFMMU_FLAGS_SET(sfmmup, HAT_32M_ISM);
8732 			sfmmu_hat_exit(hatlockp);
8733 		}
8734 		break;
8735 	default:
8736 		break;
8737 	}
8738 
8739 	/*
8740 	 * If we updated the ismblkpa for this HAT we must make
8741 	 * sure all CPUs running this process reload their tsbmiss area.
8742 	 * Otherwise they will fail to load the mappings in the tsbmiss
8743 	 * handler and will loop calling pagefault().
8744 	 */
8745 	if (reload_mmu) {
8746 		hatlockp = sfmmu_hat_enter(sfmmup);
8747 		sfmmu_sync_mmustate(sfmmup);
8748 		sfmmu_hat_exit(hatlockp);
8749 	}
8750 
8751 	sfmmu_ismhat_exit(sfmmup, 0);
8752 
8753 	/*
8754 	 * Free up ismblk if we didn't use it.
8755 	 */
8756 	if (new_iblk != NULL)
8757 		kmem_cache_free(ism_blk_cache, new_iblk);
8758 
8759 	/*
8760 	 * Check TSB and TLB page sizes.
8761 	 */
8762 	if (sfmmup->sfmmu_scdp != NULL && old_scdp != sfmmup->sfmmu_scdp) {
8763 		sfmmu_check_page_sizes(sfmmup, 0);
8764 	} else {
8765 		sfmmu_check_page_sizes(sfmmup, 1);
8766 	}
8767 	return (0);
8768 }
8769 
8770 /*
8771  * hat_unshare removes exactly one ism_map from
8772  * this process's as.  It expects multiple calls
8773  * to hat_unshare for multiple shm segments.
8774  */
8775 void
8776 hat_unshare(struct hat *sfmmup, caddr_t addr, size_t len, uint_t ismszc)
8777 {
8778 	ism_map_t 	*ism_map;
8779 	ism_ment_t	*free_ment = NULL;
8780 	ism_blk_t	*ism_blkp;
8781 	struct hat	*ism_hatid;
8782 	int 		found, i;
8783 	hatlock_t	*hatlockp;
8784 	struct tsb_info	*tsbinfo;
8785 	uint_t		ismshift = page_get_shift(ismszc);
8786 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8787 	uchar_t		ism_rid;
8788 	sf_scd_t	*old_scdp;
8789 
8790 	ASSERT(ISM_ALIGNED(ismshift, addr));
8791 	ASSERT(ISM_ALIGNED(ismshift, len));
8792 	ASSERT(sfmmup != NULL);
8793 	ASSERT(sfmmup != ksfmmup);
8794 
8795 	if (sfmmup->sfmmu_xhat_provider) {
8796 		XHAT_UNSHARE(sfmmup, addr, len);
8797 		return;
8798 	} else {
8799 		/*
8800 		 * This must be a CPU HAT. If the address space has
8801 		 * XHATs attached, inform all XHATs that ISM segment
8802 		 * is going away
8803 		 */
8804 		ASSERT(sfmmup->sfmmu_as != NULL);
8805 		if (sfmmup->sfmmu_as->a_xhat != NULL)
8806 			xhat_unshare_all(sfmmup->sfmmu_as, addr, len);
8807 	}
8808 
8809 	/*
8810 	 * Make sure that during the entire time ISM mappings are removed,
8811 	 * the trap handlers serialize behind us, and that no one else
8812 	 * can be mucking with ISM mappings.  This also lets us get away
8813 	 * with not doing expensive cross calls to flush the TLB -- we
8814 	 * just discard the context, flush the entire TSB, and call it
8815 	 * a day.
8816 	 */
8817 	sfmmu_ismhat_enter(sfmmup, 0);
8818 
8819 	/*
8820 	 * Remove the mapping.
8821 	 *
8822 	 * We can't have any holes in the ism map.
8823 	 * The tsb miss code while searching the ism map will
8824 	 * stop on an empty map slot.  So we must move
8825 	 * everyone past the hole up 1 if any.
8826 	 *
8827 	 * Also empty ism map blks are not freed until the
8828 	 * process exits. This is to prevent a MT race condition
8829 	 * between sfmmu_unshare() and sfmmu_tsbmiss_exception().
8830 	 */
8831 	found = 0;
8832 	ism_blkp = sfmmup->sfmmu_iblk;
8833 	while (!found && ism_blkp != NULL) {
8834 		ism_map = ism_blkp->iblk_maps;
8835 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
8836 			if (addr == ism_start(ism_map[i]) &&
8837 			    sh_size == (size_t)(ism_size(ism_map[i]))) {
8838 				found = 1;
8839 				break;
8840 			}
8841 		}
8842 		if (!found)
8843 			ism_blkp = ism_blkp->iblk_next;
8844 	}
8845 
8846 	if (found) {
8847 		ism_hatid = ism_map[i].imap_ismhat;
8848 		ism_rid = ism_map[i].imap_rid;
8849 		ASSERT(ism_hatid != NULL);
8850 		ASSERT(ism_hatid->sfmmu_ismhat == 1);
8851 
8852 		/*
8853 		 * After hat_leave_region, the sfmmup may leave SCD,
8854 		 * in which case, we want to grow the private tsb size when
8855 		 * calling sfmmu_check_page_sizes at the end of the routine.
8856 		 */
8857 		old_scdp = sfmmup->sfmmu_scdp;
8858 		/*
8859 		 * Then remove ourselves from the region.
8860 		 */
8861 		if (ism_rid != SFMMU_INVALID_ISMRID) {
8862 			hat_leave_region(sfmmup, (void *)((uint64_t)ism_rid),
8863 			    HAT_REGION_ISM);
8864 		}
8865 
8866 		/*
8867 		 * And now guarantee that any other cpu
8868 		 * that tries to process an ISM miss
8869 		 * will go to tl=0.
8870 		 */
8871 		hatlockp = sfmmu_hat_enter(sfmmup);
8872 		sfmmu_invalidate_ctx(sfmmup);
8873 		sfmmu_hat_exit(hatlockp);
8874 
8875 		/*
8876 		 * Remove ourselves from the ism mapping list.
8877 		 */
8878 		mutex_enter(&ism_mlist_lock);
8879 		iment_sub(ism_map[i].imap_ment, ism_hatid);
8880 		mutex_exit(&ism_mlist_lock);
8881 		free_ment = ism_map[i].imap_ment;
8882 
8883 		/*
8884 		 * We delete the ism map by copying
8885 		 * the next map over the current one.
8886 		 * We will take the next one in the maps
8887 		 * array or from the next ism_blk.
8888 		 */
8889 		while (ism_blkp != NULL) {
8890 			ism_map = ism_blkp->iblk_maps;
8891 			while (i < (ISM_MAP_SLOTS - 1)) {
8892 				ism_map[i] = ism_map[i + 1];
8893 				i++;
8894 			}
8895 			/* i == (ISM_MAP_SLOTS - 1) */
8896 			ism_blkp = ism_blkp->iblk_next;
8897 			if (ism_blkp != NULL) {
8898 				ism_map[i] = ism_blkp->iblk_maps[0];
8899 				i = 0;
8900 			} else {
8901 				ism_map[i].imap_seg = 0;
8902 				ism_map[i].imap_vb_shift = 0;
8903 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8904 				ism_map[i].imap_hatflags = 0;
8905 				ism_map[i].imap_sz_mask = 0;
8906 				ism_map[i].imap_ismhat = NULL;
8907 				ism_map[i].imap_ment = NULL;
8908 			}
8909 		}
8910 
8911 		/*
8912 		 * Now flush entire TSB for the process, since
8913 		 * demapping page by page can be too expensive.
8914 		 * We don't have to flush the TLB here anymore
8915 		 * since we switch to a new TLB ctx instead.
8916 		 * Also, there is no need to flush if the process
8917 		 * is exiting since the TSB will be freed later.
8918 		 */
8919 		if (!sfmmup->sfmmu_free) {
8920 			hatlockp = sfmmu_hat_enter(sfmmup);
8921 			for (tsbinfo = sfmmup->sfmmu_tsb; tsbinfo != NULL;
8922 			    tsbinfo = tsbinfo->tsb_next) {
8923 				if (tsbinfo->tsb_flags & TSB_SWAPPED)
8924 					continue;
8925 				if (tsbinfo->tsb_flags & TSB_RELOC_FLAG) {
8926 					tsbinfo->tsb_flags |=
8927 					    TSB_FLUSH_NEEDED;
8928 					continue;
8929 				}
8930 
8931 				sfmmu_inv_tsb(tsbinfo->tsb_va,
8932 				    TSB_BYTES(tsbinfo->tsb_szc));
8933 			}
8934 			sfmmu_hat_exit(hatlockp);
8935 		}
8936 	}
8937 
8938 	/*
8939 	 * Update our counters for this sfmmup's ism mappings.
8940 	 */
8941 	for (i = 0; i <= ismszc; i++) {
8942 		if (!(disable_ism_large_pages & (1 << i)))
8943 			(void) ism_tsb_entries(sfmmup, i);
8944 	}
8945 
8946 	sfmmu_ismhat_exit(sfmmup, 0);
8947 
8948 	/*
8949 	 * We must do our freeing here after dropping locks
8950 	 * to prevent a deadlock in the kmem allocator on the
8951 	 * mapping list lock.
8952 	 */
8953 	if (free_ment != NULL)
8954 		kmem_cache_free(ism_ment_cache, free_ment);
8955 
8956 	/*
8957 	 * Check TSB and TLB page sizes if the process isn't exiting.
8958 	 */
8959 	if (!sfmmup->sfmmu_free) {
8960 		if (found && old_scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
8961 			sfmmu_check_page_sizes(sfmmup, 1);
8962 		} else {
8963 			sfmmu_check_page_sizes(sfmmup, 0);
8964 		}
8965 	}
8966 }
8967 
8968 /* ARGSUSED */
8969 static int
8970 sfmmu_idcache_constructor(void *buf, void *cdrarg, int kmflags)
8971 {
8972 	/* void *buf is sfmmu_t pointer */
8973 	bzero(buf, sizeof (sfmmu_t));
8974 
8975 	return (0);
8976 }
8977 
8978 /* ARGSUSED */
8979 static void
8980 sfmmu_idcache_destructor(void *buf, void *cdrarg)
8981 {
8982 	/* void *buf is sfmmu_t pointer */
8983 }
8984 
8985 /*
8986  * setup kmem hmeblks by bzeroing all members and initializing the nextpa
8987  * field to be the pa of this hmeblk
8988  */
8989 /* ARGSUSED */
8990 static int
8991 sfmmu_hblkcache_constructor(void *buf, void *cdrarg, int kmflags)
8992 {
8993 	struct hme_blk *hmeblkp;
8994 
8995 	bzero(buf, (size_t)cdrarg);
8996 	hmeblkp = (struct hme_blk *)buf;
8997 	hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
8998 
8999 #ifdef	HBLK_TRACE
9000 	mutex_init(&hmeblkp->hblk_audit_lock, NULL, MUTEX_DEFAULT, NULL);
9001 #endif	/* HBLK_TRACE */
9002 
9003 	return (0);
9004 }
9005 
9006 /* ARGSUSED */
9007 static void
9008 sfmmu_hblkcache_destructor(void *buf, void *cdrarg)
9009 {
9010 
9011 #ifdef	HBLK_TRACE
9012 
9013 	struct hme_blk *hmeblkp;
9014 
9015 	hmeblkp = (struct hme_blk *)buf;
9016 	mutex_destroy(&hmeblkp->hblk_audit_lock);
9017 
9018 #endif	/* HBLK_TRACE */
9019 }
9020 
9021 #define	SFMMU_CACHE_RECLAIM_SCAN_RATIO 8
9022 static int sfmmu_cache_reclaim_scan_ratio = SFMMU_CACHE_RECLAIM_SCAN_RATIO;
9023 /*
9024  * The kmem allocator will callback into our reclaim routine when the system
9025  * is running low in memory.  We traverse the hash and free up all unused but
9026  * still cached hme_blks.  We also traverse the free list and free them up
9027  * as well.
9028  */
9029 /*ARGSUSED*/
9030 static void
9031 sfmmu_hblkcache_reclaim(void *cdrarg)
9032 {
9033 	int i;
9034 	struct hmehash_bucket *hmebp;
9035 	struct hme_blk *hmeblkp, *nx_hblk, *pr_hblk = NULL;
9036 	static struct hmehash_bucket *uhmehash_reclaim_hand;
9037 	static struct hmehash_bucket *khmehash_reclaim_hand;
9038 	struct hme_blk *list = NULL, *last_hmeblkp;
9039 	cpuset_t cpuset = cpu_ready_set;
9040 	cpu_hme_pend_t *cpuhp;
9041 
9042 	/* Free up hmeblks on the cpu pending lists */
9043 	for (i = 0; i < NCPU; i++) {
9044 		cpuhp = &cpu_hme_pend[i];
9045 		if (cpuhp->chp_listp != NULL)  {
9046 			mutex_enter(&cpuhp->chp_mutex);
9047 			if (cpuhp->chp_listp == NULL) {
9048 				mutex_exit(&cpuhp->chp_mutex);
9049 				continue;
9050 			}
9051 			for (last_hmeblkp = cpuhp->chp_listp;
9052 			    last_hmeblkp->hblk_next != NULL;
9053 			    last_hmeblkp = last_hmeblkp->hblk_next)
9054 				;
9055 			last_hmeblkp->hblk_next = list;
9056 			list = cpuhp->chp_listp;
9057 			cpuhp->chp_listp = NULL;
9058 			cpuhp->chp_count = 0;
9059 			mutex_exit(&cpuhp->chp_mutex);
9060 		}
9061 
9062 	}
9063 
9064 	if (list != NULL) {
9065 		kpreempt_disable();
9066 		CPUSET_DEL(cpuset, CPU->cpu_id);
9067 		xt_sync(cpuset);
9068 		xt_sync(cpuset);
9069 		kpreempt_enable();
9070 		sfmmu_hblk_free(&list);
9071 		list = NULL;
9072 	}
9073 
9074 	hmebp = uhmehash_reclaim_hand;
9075 	if (hmebp == NULL || hmebp > &uhme_hash[UHMEHASH_SZ])
9076 		uhmehash_reclaim_hand = hmebp = uhme_hash;
9077 	uhmehash_reclaim_hand += UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
9078 
9079 	for (i = UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
9080 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
9081 			hmeblkp = hmebp->hmeblkp;
9082 			pr_hblk = NULL;
9083 			while (hmeblkp) {
9084 				nx_hblk = hmeblkp->hblk_next;
9085 				if (!hmeblkp->hblk_vcnt &&
9086 				    !hmeblkp->hblk_hmecnt) {
9087 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9088 					    pr_hblk, &list, 0);
9089 				} else {
9090 					pr_hblk = hmeblkp;
9091 				}
9092 				hmeblkp = nx_hblk;
9093 			}
9094 			SFMMU_HASH_UNLOCK(hmebp);
9095 		}
9096 		if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
9097 			hmebp = uhme_hash;
9098 	}
9099 
9100 	hmebp = khmehash_reclaim_hand;
9101 	if (hmebp == NULL || hmebp > &khme_hash[KHMEHASH_SZ])
9102 		khmehash_reclaim_hand = hmebp = khme_hash;
9103 	khmehash_reclaim_hand += KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
9104 
9105 	for (i = KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
9106 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
9107 			hmeblkp = hmebp->hmeblkp;
9108 			pr_hblk = NULL;
9109 			while (hmeblkp) {
9110 				nx_hblk = hmeblkp->hblk_next;
9111 				if (!hmeblkp->hblk_vcnt &&
9112 				    !hmeblkp->hblk_hmecnt) {
9113 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9114 					    pr_hblk, &list, 0);
9115 				} else {
9116 					pr_hblk = hmeblkp;
9117 				}
9118 				hmeblkp = nx_hblk;
9119 			}
9120 			SFMMU_HASH_UNLOCK(hmebp);
9121 		}
9122 		if (hmebp++ == &khme_hash[KHMEHASH_SZ])
9123 			hmebp = khme_hash;
9124 	}
9125 	sfmmu_hblks_list_purge(&list, 0);
9126 }
9127 
9128 /*
9129  * sfmmu_get_ppvcolor should become a vm_machdep or hatop interface.
9130  * same goes for sfmmu_get_addrvcolor().
9131  *
9132  * This function will return the virtual color for the specified page. The
9133  * virtual color corresponds to this page current mapping or its last mapping.
9134  * It is used by memory allocators to choose addresses with the correct
9135  * alignment so vac consistency is automatically maintained.  If the page
9136  * has no color it returns -1.
9137  */
9138 /*ARGSUSED*/
9139 int
9140 sfmmu_get_ppvcolor(struct page *pp)
9141 {
9142 #ifdef VAC
9143 	int color;
9144 
9145 	if (!(cache & CACHE_VAC) || PP_NEWPAGE(pp)) {
9146 		return (-1);
9147 	}
9148 	color = PP_GET_VCOLOR(pp);
9149 	ASSERT(color < mmu_btop(shm_alignment));
9150 	return (color);
9151 #else
9152 	return (-1);
9153 #endif	/* VAC */
9154 }
9155 
9156 /*
9157  * This function will return the desired alignment for vac consistency
9158  * (vac color) given a virtual address.  If no vac is present it returns -1.
9159  */
9160 /*ARGSUSED*/
9161 int
9162 sfmmu_get_addrvcolor(caddr_t vaddr)
9163 {
9164 #ifdef VAC
9165 	if (cache & CACHE_VAC) {
9166 		return (addr_to_vcolor(vaddr));
9167 	} else {
9168 		return (-1);
9169 	}
9170 #else
9171 	return (-1);
9172 #endif	/* VAC */
9173 }
9174 
9175 #ifdef VAC
9176 /*
9177  * Check for conflicts.
9178  * A conflict exists if the new and existent mappings do not match in
9179  * their "shm_alignment fields. If conflicts exist, the existant mappings
9180  * are flushed unless one of them is locked. If one of them is locked, then
9181  * the mappings are flushed and converted to non-cacheable mappings.
9182  */
9183 static void
9184 sfmmu_vac_conflict(struct hat *hat, caddr_t addr, page_t *pp)
9185 {
9186 	struct hat *tmphat;
9187 	struct sf_hment *sfhmep, *tmphme = NULL;
9188 	struct hme_blk *hmeblkp;
9189 	int vcolor;
9190 	tte_t tte;
9191 
9192 	ASSERT(sfmmu_mlist_held(pp));
9193 	ASSERT(!PP_ISNC(pp));		/* page better be cacheable */
9194 
9195 	vcolor = addr_to_vcolor(addr);
9196 	if (PP_NEWPAGE(pp)) {
9197 		PP_SET_VCOLOR(pp, vcolor);
9198 		return;
9199 	}
9200 
9201 	if (PP_GET_VCOLOR(pp) == vcolor) {
9202 		return;
9203 	}
9204 
9205 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
9206 		/*
9207 		 * Previous user of page had a different color
9208 		 * but since there are no current users
9209 		 * we just flush the cache and change the color.
9210 		 */
9211 		SFMMU_STAT(sf_pgcolor_conflict);
9212 		sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9213 		PP_SET_VCOLOR(pp, vcolor);
9214 		return;
9215 	}
9216 
9217 	/*
9218 	 * If we get here we have a vac conflict with a current
9219 	 * mapping.  VAC conflict policy is as follows.
9220 	 * - The default is to unload the other mappings unless:
9221 	 * - If we have a large mapping we uncache the page.
9222 	 * We need to uncache the rest of the large page too.
9223 	 * - If any of the mappings are locked we uncache the page.
9224 	 * - If the requested mapping is inconsistent
9225 	 * with another mapping and that mapping
9226 	 * is in the same address space we have to
9227 	 * make it non-cached.  The default thing
9228 	 * to do is unload the inconsistent mapping
9229 	 * but if they are in the same address space
9230 	 * we run the risk of unmapping the pc or the
9231 	 * stack which we will use as we return to the user,
9232 	 * in which case we can then fault on the thing
9233 	 * we just unloaded and get into an infinite loop.
9234 	 */
9235 	if (PP_ISMAPPED_LARGE(pp)) {
9236 		int sz;
9237 
9238 		/*
9239 		 * Existing mapping is for big pages. We don't unload
9240 		 * existing big mappings to satisfy new mappings.
9241 		 * Always convert all mappings to TNC.
9242 		 */
9243 		sz = fnd_mapping_sz(pp);
9244 		pp = PP_GROUPLEADER(pp, sz);
9245 		SFMMU_STAT_ADD(sf_uncache_conflict, TTEPAGES(sz));
9246 		sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH,
9247 		    TTEPAGES(sz));
9248 
9249 		return;
9250 	}
9251 
9252 	/*
9253 	 * check if any mapping is in same as or if it is locked
9254 	 * since in that case we need to uncache.
9255 	 */
9256 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9257 		tmphme = sfhmep->hme_next;
9258 		if (IS_PAHME(sfhmep))
9259 			continue;
9260 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9261 		if (hmeblkp->hblk_xhat_bit)
9262 			continue;
9263 		tmphat = hblktosfmmu(hmeblkp);
9264 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
9265 		ASSERT(TTE_IS_VALID(&tte));
9266 		if (hmeblkp->hblk_shared || tmphat == hat ||
9267 		    hmeblkp->hblk_lckcnt) {
9268 			/*
9269 			 * We have an uncache conflict
9270 			 */
9271 			SFMMU_STAT(sf_uncache_conflict);
9272 			sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH, 1);
9273 			return;
9274 		}
9275 	}
9276 
9277 	/*
9278 	 * We have an unload conflict
9279 	 * We have already checked for LARGE mappings, therefore
9280 	 * the remaining mapping(s) must be TTE8K.
9281 	 */
9282 	SFMMU_STAT(sf_unload_conflict);
9283 
9284 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9285 		tmphme = sfhmep->hme_next;
9286 		if (IS_PAHME(sfhmep))
9287 			continue;
9288 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9289 		if (hmeblkp->hblk_xhat_bit)
9290 			continue;
9291 		ASSERT(!hmeblkp->hblk_shared);
9292 		(void) sfmmu_pageunload(pp, sfhmep, TTE8K);
9293 	}
9294 
9295 	if (PP_ISMAPPED_KPM(pp))
9296 		sfmmu_kpm_vac_unload(pp, addr);
9297 
9298 	/*
9299 	 * Unloads only do TLB flushes so we need to flush the
9300 	 * cache here.
9301 	 */
9302 	sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9303 	PP_SET_VCOLOR(pp, vcolor);
9304 }
9305 
9306 /*
9307  * Whenever a mapping is unloaded and the page is in TNC state,
9308  * we see if the page can be made cacheable again. 'pp' is
9309  * the page that we just unloaded a mapping from, the size
9310  * of mapping that was unloaded is 'ottesz'.
9311  * Remark:
9312  * The recache policy for mpss pages can leave a performance problem
9313  * under the following circumstances:
9314  * . A large page in uncached mode has just been unmapped.
9315  * . All constituent pages are TNC due to a conflicting small mapping.
9316  * . There are many other, non conflicting, small mappings around for
9317  *   a lot of the constituent pages.
9318  * . We're called w/ the "old" groupleader page and the old ottesz,
9319  *   but this is irrelevant, since we're no more "PP_ISMAPPED_LARGE", so
9320  *   we end up w/ TTE8K or npages == 1.
9321  * . We call tst_tnc w/ the old groupleader only, and if there is no
9322  *   conflict, we re-cache only this page.
9323  * . All other small mappings are not checked and will be left in TNC mode.
9324  * The problem is not very serious because:
9325  * . mpss is actually only defined for heap and stack, so the probability
9326  *   is not very high that a large page mapping exists in parallel to a small
9327  *   one (this is possible, but seems to be bad programming style in the
9328  *   appl).
9329  * . The problem gets a little bit more serious, when those TNC pages
9330  *   have to be mapped into kernel space, e.g. for networking.
9331  * . When VAC alias conflicts occur in applications, this is regarded
9332  *   as an application bug. So if kstat's show them, the appl should
9333  *   be changed anyway.
9334  */
9335 void
9336 conv_tnc(page_t *pp, int ottesz)
9337 {
9338 	int cursz, dosz;
9339 	pgcnt_t curnpgs, dopgs;
9340 	pgcnt_t pg64k;
9341 	page_t *pp2;
9342 
9343 	/*
9344 	 * Determine how big a range we check for TNC and find
9345 	 * leader page. cursz is the size of the biggest
9346 	 * mapping that still exist on 'pp'.
9347 	 */
9348 	if (PP_ISMAPPED_LARGE(pp)) {
9349 		cursz = fnd_mapping_sz(pp);
9350 	} else {
9351 		cursz = TTE8K;
9352 	}
9353 
9354 	if (ottesz >= cursz) {
9355 		dosz = ottesz;
9356 		pp2 = pp;
9357 	} else {
9358 		dosz = cursz;
9359 		pp2 = PP_GROUPLEADER(pp, dosz);
9360 	}
9361 
9362 	pg64k = TTEPAGES(TTE64K);
9363 	dopgs = TTEPAGES(dosz);
9364 
9365 	ASSERT(dopgs == 1 || ((dopgs & (pg64k - 1)) == 0));
9366 
9367 	while (dopgs != 0) {
9368 		curnpgs = TTEPAGES(cursz);
9369 		if (tst_tnc(pp2, curnpgs)) {
9370 			SFMMU_STAT_ADD(sf_recache, curnpgs);
9371 			sfmmu_page_cache_array(pp2, HAT_CACHE, CACHE_NO_FLUSH,
9372 			    curnpgs);
9373 		}
9374 
9375 		ASSERT(dopgs >= curnpgs);
9376 		dopgs -= curnpgs;
9377 
9378 		if (dopgs == 0) {
9379 			break;
9380 		}
9381 
9382 		pp2 = PP_PAGENEXT_N(pp2, curnpgs);
9383 		if (((dopgs & (pg64k - 1)) == 0) && PP_ISMAPPED_LARGE(pp2)) {
9384 			cursz = fnd_mapping_sz(pp2);
9385 		} else {
9386 			cursz = TTE8K;
9387 		}
9388 	}
9389 }
9390 
9391 /*
9392  * Returns 1 if page(s) can be converted from TNC to cacheable setting,
9393  * returns 0 otherwise. Note that oaddr argument is valid for only
9394  * 8k pages.
9395  */
9396 int
9397 tst_tnc(page_t *pp, pgcnt_t npages)
9398 {
9399 	struct	sf_hment *sfhme;
9400 	struct	hme_blk *hmeblkp;
9401 	tte_t	tte;
9402 	caddr_t	vaddr;
9403 	int	clr_valid = 0;
9404 	int 	color, color1, bcolor;
9405 	int	i, ncolors;
9406 
9407 	ASSERT(pp != NULL);
9408 	ASSERT(!(cache & CACHE_WRITEBACK));
9409 
9410 	if (npages > 1) {
9411 		ncolors = CACHE_NUM_COLOR;
9412 	}
9413 
9414 	for (i = 0; i < npages; i++) {
9415 		ASSERT(sfmmu_mlist_held(pp));
9416 		ASSERT(PP_ISTNC(pp));
9417 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
9418 
9419 		if (PP_ISPNC(pp)) {
9420 			return (0);
9421 		}
9422 
9423 		clr_valid = 0;
9424 		if (PP_ISMAPPED_KPM(pp)) {
9425 			caddr_t kpmvaddr;
9426 
9427 			ASSERT(kpm_enable);
9428 			kpmvaddr = hat_kpm_page2va(pp, 1);
9429 			ASSERT(!(npages > 1 && IS_KPM_ALIAS_RANGE(kpmvaddr)));
9430 			color1 = addr_to_vcolor(kpmvaddr);
9431 			clr_valid = 1;
9432 		}
9433 
9434 		for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9435 			if (IS_PAHME(sfhme))
9436 				continue;
9437 			hmeblkp = sfmmu_hmetohblk(sfhme);
9438 			if (hmeblkp->hblk_xhat_bit)
9439 				continue;
9440 
9441 			sfmmu_copytte(&sfhme->hme_tte, &tte);
9442 			ASSERT(TTE_IS_VALID(&tte));
9443 
9444 			vaddr = tte_to_vaddr(hmeblkp, tte);
9445 			color = addr_to_vcolor(vaddr);
9446 
9447 			if (npages > 1) {
9448 				/*
9449 				 * If there is a big mapping, make sure
9450 				 * 8K mapping is consistent with the big
9451 				 * mapping.
9452 				 */
9453 				bcolor = i % ncolors;
9454 				if (color != bcolor) {
9455 					return (0);
9456 				}
9457 			}
9458 			if (!clr_valid) {
9459 				clr_valid = 1;
9460 				color1 = color;
9461 			}
9462 
9463 			if (color1 != color) {
9464 				return (0);
9465 			}
9466 		}
9467 
9468 		pp = PP_PAGENEXT(pp);
9469 	}
9470 
9471 	return (1);
9472 }
9473 
9474 void
9475 sfmmu_page_cache_array(page_t *pp, int flags, int cache_flush_flag,
9476 	pgcnt_t npages)
9477 {
9478 	kmutex_t *pmtx;
9479 	int i, ncolors, bcolor;
9480 	kpm_hlk_t *kpmp;
9481 	cpuset_t cpuset;
9482 
9483 	ASSERT(pp != NULL);
9484 	ASSERT(!(cache & CACHE_WRITEBACK));
9485 
9486 	kpmp = sfmmu_kpm_kpmp_enter(pp, npages);
9487 	pmtx = sfmmu_page_enter(pp);
9488 
9489 	/*
9490 	 * Fast path caching single unmapped page
9491 	 */
9492 	if (npages == 1 && !PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp) &&
9493 	    flags == HAT_CACHE) {
9494 		PP_CLRTNC(pp);
9495 		PP_CLRPNC(pp);
9496 		sfmmu_page_exit(pmtx);
9497 		sfmmu_kpm_kpmp_exit(kpmp);
9498 		return;
9499 	}
9500 
9501 	/*
9502 	 * We need to capture all cpus in order to change cacheability
9503 	 * because we can't allow one cpu to access the same physical
9504 	 * page using a cacheable and a non-cachebale mapping at the same
9505 	 * time. Since we may end up walking the ism mapping list
9506 	 * have to grab it's lock now since we can't after all the
9507 	 * cpus have been captured.
9508 	 */
9509 	sfmmu_hat_lock_all();
9510 	mutex_enter(&ism_mlist_lock);
9511 	kpreempt_disable();
9512 	cpuset = cpu_ready_set;
9513 	xc_attention(cpuset);
9514 
9515 	if (npages > 1) {
9516 		/*
9517 		 * Make sure all colors are flushed since the
9518 		 * sfmmu_page_cache() only flushes one color-
9519 		 * it does not know big pages.
9520 		 */
9521 		ncolors = CACHE_NUM_COLOR;
9522 		if (flags & HAT_TMPNC) {
9523 			for (i = 0; i < ncolors; i++) {
9524 				sfmmu_cache_flushcolor(i, pp->p_pagenum);
9525 			}
9526 			cache_flush_flag = CACHE_NO_FLUSH;
9527 		}
9528 	}
9529 
9530 	for (i = 0; i < npages; i++) {
9531 
9532 		ASSERT(sfmmu_mlist_held(pp));
9533 
9534 		if (!(flags == HAT_TMPNC && PP_ISTNC(pp))) {
9535 
9536 			if (npages > 1) {
9537 				bcolor = i % ncolors;
9538 			} else {
9539 				bcolor = NO_VCOLOR;
9540 			}
9541 
9542 			sfmmu_page_cache(pp, flags, cache_flush_flag,
9543 			    bcolor);
9544 		}
9545 
9546 		pp = PP_PAGENEXT(pp);
9547 	}
9548 
9549 	xt_sync(cpuset);
9550 	xc_dismissed(cpuset);
9551 	mutex_exit(&ism_mlist_lock);
9552 	sfmmu_hat_unlock_all();
9553 	sfmmu_page_exit(pmtx);
9554 	sfmmu_kpm_kpmp_exit(kpmp);
9555 	kpreempt_enable();
9556 }
9557 
9558 /*
9559  * This function changes the virtual cacheability of all mappings to a
9560  * particular page.  When changing from uncache to cacheable the mappings will
9561  * only be changed if all of them have the same virtual color.
9562  * We need to flush the cache in all cpus.  It is possible that
9563  * a process referenced a page as cacheable but has sinced exited
9564  * and cleared the mapping list.  We still to flush it but have no
9565  * state so all cpus is the only alternative.
9566  */
9567 static void
9568 sfmmu_page_cache(page_t *pp, int flags, int cache_flush_flag, int bcolor)
9569 {
9570 	struct	sf_hment *sfhme;
9571 	struct	hme_blk *hmeblkp;
9572 	sfmmu_t *sfmmup;
9573 	tte_t	tte, ttemod;
9574 	caddr_t	vaddr;
9575 	int	ret, color;
9576 	pfn_t	pfn;
9577 
9578 	color = bcolor;
9579 	pfn = pp->p_pagenum;
9580 
9581 	for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9582 
9583 		if (IS_PAHME(sfhme))
9584 			continue;
9585 		hmeblkp = sfmmu_hmetohblk(sfhme);
9586 
9587 		if (hmeblkp->hblk_xhat_bit)
9588 			continue;
9589 
9590 		sfmmu_copytte(&sfhme->hme_tte, &tte);
9591 		ASSERT(TTE_IS_VALID(&tte));
9592 		vaddr = tte_to_vaddr(hmeblkp, tte);
9593 		color = addr_to_vcolor(vaddr);
9594 
9595 #ifdef DEBUG
9596 		if ((flags & HAT_CACHE) && bcolor != NO_VCOLOR) {
9597 			ASSERT(color == bcolor);
9598 		}
9599 #endif
9600 
9601 		ASSERT(flags != HAT_TMPNC || color == PP_GET_VCOLOR(pp));
9602 
9603 		ttemod = tte;
9604 		if (flags & (HAT_UNCACHE | HAT_TMPNC)) {
9605 			TTE_CLR_VCACHEABLE(&ttemod);
9606 		} else {	/* flags & HAT_CACHE */
9607 			TTE_SET_VCACHEABLE(&ttemod);
9608 		}
9609 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
9610 		if (ret < 0) {
9611 			/*
9612 			 * Since all cpus are captured modifytte should not
9613 			 * fail.
9614 			 */
9615 			panic("sfmmu_page_cache: write to tte failed");
9616 		}
9617 
9618 		sfmmup = hblktosfmmu(hmeblkp);
9619 		if (cache_flush_flag == CACHE_FLUSH) {
9620 			/*
9621 			 * Flush TSBs, TLBs and caches
9622 			 */
9623 			if (hmeblkp->hblk_shared) {
9624 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9625 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9626 				sf_region_t *rgnp;
9627 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9628 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9629 				ASSERT(srdp != NULL);
9630 				rgnp = srdp->srd_hmergnp[rid];
9631 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9632 				    srdp, rgnp, rid);
9633 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9634 				    hmeblkp, 0);
9635 				sfmmu_cache_flush(pfn, addr_to_vcolor(vaddr));
9636 			} else if (sfmmup->sfmmu_ismhat) {
9637 				if (flags & HAT_CACHE) {
9638 					SFMMU_STAT(sf_ism_recache);
9639 				} else {
9640 					SFMMU_STAT(sf_ism_uncache);
9641 				}
9642 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9643 				    pfn, CACHE_FLUSH);
9644 			} else {
9645 				sfmmu_tlbcache_demap(vaddr, sfmmup, hmeblkp,
9646 				    pfn, 0, FLUSH_ALL_CPUS, CACHE_FLUSH, 1);
9647 			}
9648 
9649 			/*
9650 			 * all cache entries belonging to this pfn are
9651 			 * now flushed.
9652 			 */
9653 			cache_flush_flag = CACHE_NO_FLUSH;
9654 		} else {
9655 			/*
9656 			 * Flush only TSBs and TLBs.
9657 			 */
9658 			if (hmeblkp->hblk_shared) {
9659 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9660 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9661 				sf_region_t *rgnp;
9662 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9663 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9664 				ASSERT(srdp != NULL);
9665 				rgnp = srdp->srd_hmergnp[rid];
9666 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9667 				    srdp, rgnp, rid);
9668 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9669 				    hmeblkp, 0);
9670 			} else if (sfmmup->sfmmu_ismhat) {
9671 				if (flags & HAT_CACHE) {
9672 					SFMMU_STAT(sf_ism_recache);
9673 				} else {
9674 					SFMMU_STAT(sf_ism_uncache);
9675 				}
9676 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9677 				    pfn, CACHE_NO_FLUSH);
9678 			} else {
9679 				sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 1);
9680 			}
9681 		}
9682 	}
9683 
9684 	if (PP_ISMAPPED_KPM(pp))
9685 		sfmmu_kpm_page_cache(pp, flags, cache_flush_flag);
9686 
9687 	switch (flags) {
9688 
9689 		default:
9690 			panic("sfmmu_pagecache: unknown flags");
9691 			break;
9692 
9693 		case HAT_CACHE:
9694 			PP_CLRTNC(pp);
9695 			PP_CLRPNC(pp);
9696 			PP_SET_VCOLOR(pp, color);
9697 			break;
9698 
9699 		case HAT_TMPNC:
9700 			PP_SETTNC(pp);
9701 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9702 			break;
9703 
9704 		case HAT_UNCACHE:
9705 			PP_SETPNC(pp);
9706 			PP_CLRTNC(pp);
9707 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9708 			break;
9709 	}
9710 }
9711 #endif	/* VAC */
9712 
9713 
9714 /*
9715  * Wrapper routine used to return a context.
9716  *
9717  * It's the responsibility of the caller to guarantee that the
9718  * process serializes on calls here by taking the HAT lock for
9719  * the hat.
9720  *
9721  */
9722 static void
9723 sfmmu_get_ctx(sfmmu_t *sfmmup)
9724 {
9725 	mmu_ctx_t *mmu_ctxp;
9726 	uint_t pstate_save;
9727 	int ret;
9728 
9729 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9730 	ASSERT(sfmmup != ksfmmup);
9731 
9732 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID)) {
9733 		sfmmu_setup_tsbinfo(sfmmup);
9734 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ALLCTX_INVALID);
9735 	}
9736 
9737 	kpreempt_disable();
9738 
9739 	mmu_ctxp = CPU_MMU_CTXP(CPU);
9740 	ASSERT(mmu_ctxp);
9741 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
9742 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
9743 
9744 	/*
9745 	 * Do a wrap-around if cnum reaches the max # cnum supported by a MMU.
9746 	 */
9747 	if (mmu_ctxp->mmu_cnum == mmu_ctxp->mmu_nctxs)
9748 		sfmmu_ctx_wrap_around(mmu_ctxp);
9749 
9750 	/*
9751 	 * Let the MMU set up the page sizes to use for
9752 	 * this context in the TLB. Don't program 2nd dtlb for ism hat.
9753 	 */
9754 	if ((&mmu_set_ctx_page_sizes) && (sfmmup->sfmmu_ismhat == 0)) {
9755 		mmu_set_ctx_page_sizes(sfmmup);
9756 	}
9757 
9758 	/*
9759 	 * sfmmu_alloc_ctx and sfmmu_load_mmustate will be performed with
9760 	 * interrupts disabled to prevent race condition with wrap-around
9761 	 * ctx invalidatation. In sun4v, ctx invalidation also involves
9762 	 * a HV call to set the number of TSBs to 0. If interrupts are not
9763 	 * disabled until after sfmmu_load_mmustate is complete TSBs may
9764 	 * become assigned to INVALID_CONTEXT. This is not allowed.
9765 	 */
9766 	pstate_save = sfmmu_disable_intrs();
9767 
9768 	if (sfmmu_alloc_ctx(sfmmup, 1, CPU, SFMMU_PRIVATE) &&
9769 	    sfmmup->sfmmu_scdp != NULL) {
9770 		sf_scd_t *scdp = sfmmup->sfmmu_scdp;
9771 		sfmmu_t *scsfmmup = scdp->scd_sfmmup;
9772 		ret = sfmmu_alloc_ctx(scsfmmup, 1, CPU, SFMMU_SHARED);
9773 		/* debug purpose only */
9774 		ASSERT(!ret || scsfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
9775 		    != INVALID_CONTEXT);
9776 	}
9777 	sfmmu_load_mmustate(sfmmup);
9778 
9779 	sfmmu_enable_intrs(pstate_save);
9780 
9781 	kpreempt_enable();
9782 }
9783 
9784 /*
9785  * When all cnums are used up in a MMU, cnum will wrap around to the
9786  * next generation and start from 2.
9787  */
9788 static void
9789 sfmmu_ctx_wrap_around(mmu_ctx_t *mmu_ctxp)
9790 {
9791 
9792 	/* caller must have disabled the preemption */
9793 	ASSERT(curthread->t_preempt >= 1);
9794 	ASSERT(mmu_ctxp != NULL);
9795 
9796 	/* acquire Per-MMU (PM) spin lock */
9797 	mutex_enter(&mmu_ctxp->mmu_lock);
9798 
9799 	/* re-check to see if wrap-around is needed */
9800 	if (mmu_ctxp->mmu_cnum < mmu_ctxp->mmu_nctxs)
9801 		goto done;
9802 
9803 	SFMMU_MMU_STAT(mmu_wrap_around);
9804 
9805 	/* update gnum */
9806 	ASSERT(mmu_ctxp->mmu_gnum != 0);
9807 	mmu_ctxp->mmu_gnum++;
9808 	if (mmu_ctxp->mmu_gnum == 0 ||
9809 	    mmu_ctxp->mmu_gnum > MAX_SFMMU_GNUM_VAL) {
9810 		cmn_err(CE_PANIC, "mmu_gnum of mmu_ctx 0x%p is out of bound.",
9811 		    (void *)mmu_ctxp);
9812 	}
9813 
9814 	if (mmu_ctxp->mmu_ncpus > 1) {
9815 		cpuset_t cpuset;
9816 
9817 		membar_enter(); /* make sure updated gnum visible */
9818 
9819 		SFMMU_XCALL_STATS(NULL);
9820 
9821 		/* xcall to others on the same MMU to invalidate ctx */
9822 		cpuset = mmu_ctxp->mmu_cpuset;
9823 		ASSERT(CPU_IN_SET(cpuset, CPU->cpu_id));
9824 		CPUSET_DEL(cpuset, CPU->cpu_id);
9825 		CPUSET_AND(cpuset, cpu_ready_set);
9826 
9827 		/*
9828 		 * Pass in INVALID_CONTEXT as the first parameter to
9829 		 * sfmmu_raise_tsb_exception, which invalidates the context
9830 		 * of any process running on the CPUs in the MMU.
9831 		 */
9832 		xt_some(cpuset, sfmmu_raise_tsb_exception,
9833 		    INVALID_CONTEXT, INVALID_CONTEXT);
9834 		xt_sync(cpuset);
9835 
9836 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
9837 	}
9838 
9839 	if (sfmmu_getctx_sec() != INVALID_CONTEXT) {
9840 		sfmmu_setctx_sec(INVALID_CONTEXT);
9841 		sfmmu_clear_utsbinfo();
9842 	}
9843 
9844 	/*
9845 	 * No xcall is needed here. For sun4u systems all CPUs in context
9846 	 * domain share a single physical MMU therefore it's enough to flush
9847 	 * TLB on local CPU. On sun4v systems we use 1 global context
9848 	 * domain and flush all remote TLBs in sfmmu_raise_tsb_exception
9849 	 * handler. Note that vtag_flushall_uctxs() is called
9850 	 * for Ultra II machine, where the equivalent flushall functionality
9851 	 * is implemented in SW, and only user ctx TLB entries are flushed.
9852 	 */
9853 	if (&vtag_flushall_uctxs != NULL) {
9854 		vtag_flushall_uctxs();
9855 	} else {
9856 		vtag_flushall();
9857 	}
9858 
9859 	/* reset mmu cnum, skips cnum 0 and 1 */
9860 	mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
9861 
9862 done:
9863 	mutex_exit(&mmu_ctxp->mmu_lock);
9864 }
9865 
9866 
9867 /*
9868  * For multi-threaded process, set the process context to INVALID_CONTEXT
9869  * so that it faults and reloads the MMU state from TL=0. For single-threaded
9870  * process, we can just load the MMU state directly without having to
9871  * set context invalid. Caller must hold the hat lock since we don't
9872  * acquire it here.
9873  */
9874 static void
9875 sfmmu_sync_mmustate(sfmmu_t *sfmmup)
9876 {
9877 	uint_t cnum;
9878 	uint_t pstate_save;
9879 
9880 	ASSERT(sfmmup != ksfmmup);
9881 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9882 
9883 	kpreempt_disable();
9884 
9885 	/*
9886 	 * We check whether the pass'ed-in sfmmup is the same as the
9887 	 * current running proc. This is to makes sure the current proc
9888 	 * stays single-threaded if it already is.
9889 	 */
9890 	if ((sfmmup == curthread->t_procp->p_as->a_hat) &&
9891 	    (curthread->t_procp->p_lwpcnt == 1)) {
9892 		/* single-thread */
9893 		cnum = sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum;
9894 		if (cnum != INVALID_CONTEXT) {
9895 			uint_t curcnum;
9896 			/*
9897 			 * Disable interrupts to prevent race condition
9898 			 * with sfmmu_ctx_wrap_around ctx invalidation.
9899 			 * In sun4v, ctx invalidation involves setting
9900 			 * TSB to NULL, hence, interrupts should be disabled
9901 			 * untill after sfmmu_load_mmustate is completed.
9902 			 */
9903 			pstate_save = sfmmu_disable_intrs();
9904 			curcnum = sfmmu_getctx_sec();
9905 			if (curcnum == cnum)
9906 				sfmmu_load_mmustate(sfmmup);
9907 			sfmmu_enable_intrs(pstate_save);
9908 			ASSERT(curcnum == cnum || curcnum == INVALID_CONTEXT);
9909 		}
9910 	} else {
9911 		/*
9912 		 * multi-thread
9913 		 * or when sfmmup is not the same as the curproc.
9914 		 */
9915 		sfmmu_invalidate_ctx(sfmmup);
9916 	}
9917 
9918 	kpreempt_enable();
9919 }
9920 
9921 
9922 /*
9923  * Replace the specified TSB with a new TSB.  This function gets called when
9924  * we grow, shrink or swapin a TSB.  When swapping in a TSB (TSB_SWAPIN), the
9925  * TSB_FORCEALLOC flag may be used to force allocation of a minimum-sized TSB
9926  * (8K).
9927  *
9928  * Caller must hold the HAT lock, but should assume any tsb_info
9929  * pointers it has are no longer valid after calling this function.
9930  *
9931  * Return values:
9932  *	TSB_ALLOCFAIL	Failed to allocate a TSB, due to memory constraints
9933  *	TSB_LOSTRACE	HAT is busy, i.e. another thread is already doing
9934  *			something to this tsbinfo/TSB
9935  *	TSB_SUCCESS	Operation succeeded
9936  */
9937 static tsb_replace_rc_t
9938 sfmmu_replace_tsb(sfmmu_t *sfmmup, struct tsb_info *old_tsbinfo, uint_t szc,
9939     hatlock_t *hatlockp, uint_t flags)
9940 {
9941 	struct tsb_info *new_tsbinfo = NULL;
9942 	struct tsb_info *curtsb, *prevtsb;
9943 	uint_t tte_sz_mask;
9944 	int i;
9945 
9946 	ASSERT(sfmmup != ksfmmup);
9947 	ASSERT(sfmmup->sfmmu_ismhat == 0);
9948 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9949 	ASSERT(szc <= tsb_max_growsize);
9950 
9951 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_BUSY))
9952 		return (TSB_LOSTRACE);
9953 
9954 	/*
9955 	 * Find the tsb_info ahead of this one in the list, and
9956 	 * also make sure that the tsb_info passed in really
9957 	 * exists!
9958 	 */
9959 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
9960 	    curtsb != old_tsbinfo && curtsb != NULL;
9961 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
9962 		;
9963 	ASSERT(curtsb != NULL);
9964 
9965 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
9966 		/*
9967 		 * The process is swapped out, so just set the new size
9968 		 * code.  When it swaps back in, we'll allocate a new one
9969 		 * of the new chosen size.
9970 		 */
9971 		curtsb->tsb_szc = szc;
9972 		return (TSB_SUCCESS);
9973 	}
9974 	SFMMU_FLAGS_SET(sfmmup, HAT_BUSY);
9975 
9976 	tte_sz_mask = old_tsbinfo->tsb_ttesz_mask;
9977 
9978 	/*
9979 	 * All initialization is done inside of sfmmu_tsbinfo_alloc().
9980 	 * If we fail to allocate a TSB, exit.
9981 	 *
9982 	 * If tsb grows with new tsb size > 4M and old tsb size < 4M,
9983 	 * then try 4M slab after the initial alloc fails.
9984 	 *
9985 	 * If tsb swapin with tsb size > 4M, then try 4M after the
9986 	 * initial alloc fails.
9987 	 */
9988 	sfmmu_hat_exit(hatlockp);
9989 	if (sfmmu_tsbinfo_alloc(&new_tsbinfo, szc,
9990 	    tte_sz_mask, flags, sfmmup) &&
9991 	    (!(flags & (TSB_GROW | TSB_SWAPIN)) || (szc <= TSB_4M_SZCODE) ||
9992 	    (!(flags & TSB_SWAPIN) &&
9993 	    (old_tsbinfo->tsb_szc >= TSB_4M_SZCODE)) ||
9994 	    sfmmu_tsbinfo_alloc(&new_tsbinfo, TSB_4M_SZCODE,
9995 	    tte_sz_mask, flags, sfmmup))) {
9996 		(void) sfmmu_hat_enter(sfmmup);
9997 		if (!(flags & TSB_SWAPIN))
9998 			SFMMU_STAT(sf_tsb_resize_failures);
9999 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10000 		return (TSB_ALLOCFAIL);
10001 	}
10002 	(void) sfmmu_hat_enter(sfmmup);
10003 
10004 	/*
10005 	 * Re-check to make sure somebody else didn't muck with us while we
10006 	 * didn't hold the HAT lock.  If the process swapped out, fine, just
10007 	 * exit; this can happen if we try to shrink the TSB from the context
10008 	 * of another process (such as on an ISM unmap), though it is rare.
10009 	 */
10010 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
10011 		SFMMU_STAT(sf_tsb_resize_failures);
10012 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10013 		sfmmu_hat_exit(hatlockp);
10014 		sfmmu_tsbinfo_free(new_tsbinfo);
10015 		(void) sfmmu_hat_enter(sfmmup);
10016 		return (TSB_LOSTRACE);
10017 	}
10018 
10019 #ifdef	DEBUG
10020 	/* Reverify that the tsb_info still exists.. for debugging only */
10021 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
10022 	    curtsb != old_tsbinfo && curtsb != NULL;
10023 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
10024 		;
10025 	ASSERT(curtsb != NULL);
10026 #endif	/* DEBUG */
10027 
10028 	/*
10029 	 * Quiesce any CPUs running this process on their next TLB miss
10030 	 * so they atomically see the new tsb_info.  We temporarily set the
10031 	 * context to invalid context so new threads that come on processor
10032 	 * after we do the xcall to cpusran will also serialize behind the
10033 	 * HAT lock on TLB miss and will see the new TSB.  Since this short
10034 	 * race with a new thread coming on processor is relatively rare,
10035 	 * this synchronization mechanism should be cheaper than always
10036 	 * pausing all CPUs for the duration of the setup, which is what
10037 	 * the old implementation did.  This is particuarly true if we are
10038 	 * copying a huge chunk of memory around during that window.
10039 	 *
10040 	 * The memory barriers are to make sure things stay consistent
10041 	 * with resume() since it does not hold the HAT lock while
10042 	 * walking the list of tsb_info structures.
10043 	 */
10044 	if ((flags & TSB_SWAPIN) != TSB_SWAPIN) {
10045 		/* The TSB is either growing or shrinking. */
10046 		sfmmu_invalidate_ctx(sfmmup);
10047 	} else {
10048 		/*
10049 		 * It is illegal to swap in TSBs from a process other
10050 		 * than a process being swapped in.  This in turn
10051 		 * implies we do not have a valid MMU context here
10052 		 * since a process needs one to resolve translation
10053 		 * misses.
10054 		 */
10055 		ASSERT(curthread->t_procp->p_as->a_hat == sfmmup);
10056 	}
10057 
10058 #ifdef DEBUG
10059 	ASSERT(max_mmu_ctxdoms > 0);
10060 
10061 	/*
10062 	 * Process should have INVALID_CONTEXT on all MMUs
10063 	 */
10064 	for (i = 0; i < max_mmu_ctxdoms; i++) {
10065 
10066 		ASSERT(sfmmup->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
10067 	}
10068 #endif
10069 
10070 	new_tsbinfo->tsb_next = old_tsbinfo->tsb_next;
10071 	membar_stst();	/* strict ordering required */
10072 	if (prevtsb)
10073 		prevtsb->tsb_next = new_tsbinfo;
10074 	else
10075 		sfmmup->sfmmu_tsb = new_tsbinfo;
10076 	membar_enter();	/* make sure new TSB globally visible */
10077 
10078 	/*
10079 	 * We need to migrate TSB entries from the old TSB to the new TSB
10080 	 * if tsb_remap_ttes is set and the TSB is growing.
10081 	 */
10082 	if (tsb_remap_ttes && ((flags & TSB_GROW) == TSB_GROW))
10083 		sfmmu_copy_tsb(old_tsbinfo, new_tsbinfo);
10084 
10085 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10086 
10087 	/*
10088 	 * Drop the HAT lock to free our old tsb_info.
10089 	 */
10090 	sfmmu_hat_exit(hatlockp);
10091 
10092 	if ((flags & TSB_GROW) == TSB_GROW) {
10093 		SFMMU_STAT(sf_tsb_grow);
10094 	} else if ((flags & TSB_SHRINK) == TSB_SHRINK) {
10095 		SFMMU_STAT(sf_tsb_shrink);
10096 	}
10097 
10098 	sfmmu_tsbinfo_free(old_tsbinfo);
10099 
10100 	(void) sfmmu_hat_enter(sfmmup);
10101 	return (TSB_SUCCESS);
10102 }
10103 
10104 /*
10105  * This function will re-program hat pgsz array, and invalidate the
10106  * process' context, forcing the process to switch to another
10107  * context on the next TLB miss, and therefore start using the
10108  * TLB that is reprogrammed for the new page sizes.
10109  */
10110 void
10111 sfmmu_reprog_pgsz_arr(sfmmu_t *sfmmup, uint8_t *tmp_pgsz)
10112 {
10113 	int i;
10114 	hatlock_t *hatlockp = NULL;
10115 
10116 	hatlockp = sfmmu_hat_enter(sfmmup);
10117 	/* USIII+-IV+ optimization, requires hat lock */
10118 	if (tmp_pgsz) {
10119 		for (i = 0; i < mmu_page_sizes; i++)
10120 			sfmmup->sfmmu_pgsz[i] = tmp_pgsz[i];
10121 	}
10122 	SFMMU_STAT(sf_tlb_reprog_pgsz);
10123 
10124 	sfmmu_invalidate_ctx(sfmmup);
10125 
10126 	sfmmu_hat_exit(hatlockp);
10127 }
10128 
10129 /*
10130  * The scd_rttecnt field in the SCD must be updated to take account of the
10131  * regions which it contains.
10132  */
10133 static void
10134 sfmmu_set_scd_rttecnt(sf_srd_t *srdp, sf_scd_t *scdp)
10135 {
10136 	uint_t rid;
10137 	uint_t i, j;
10138 	ulong_t w;
10139 	sf_region_t *rgnp;
10140 
10141 	ASSERT(srdp != NULL);
10142 
10143 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
10144 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
10145 			continue;
10146 		}
10147 
10148 		j = 0;
10149 		while (w) {
10150 			if (!(w & 0x1)) {
10151 				j++;
10152 				w >>= 1;
10153 				continue;
10154 			}
10155 			rid = (i << BT_ULSHIFT) | j;
10156 			j++;
10157 			w >>= 1;
10158 
10159 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
10160 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
10161 			rgnp = srdp->srd_hmergnp[rid];
10162 			ASSERT(rgnp->rgn_refcnt > 0);
10163 			ASSERT(rgnp->rgn_id == rid);
10164 
10165 			scdp->scd_rttecnt[rgnp->rgn_pgszc] +=
10166 			    rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
10167 
10168 			/*
10169 			 * Maintain the tsb0 inflation cnt for the regions
10170 			 * in the SCD.
10171 			 */
10172 			if (rgnp->rgn_pgszc >= TTE4M) {
10173 				scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt +=
10174 				    rgnp->rgn_size >>
10175 				    (TTE_PAGE_SHIFT(TTE8K) + 2);
10176 			}
10177 		}
10178 	}
10179 }
10180 
10181 /*
10182  * This function assumes that there are either four or six supported page
10183  * sizes and at most two programmable TLBs, so we need to decide which
10184  * page sizes are most important and then tell the MMU layer so it
10185  * can adjust the TLB page sizes accordingly (if supported).
10186  *
10187  * If these assumptions change, this function will need to be
10188  * updated to support whatever the new limits are.
10189  *
10190  * The growing flag is nonzero if we are growing the address space,
10191  * and zero if it is shrinking.  This allows us to decide whether
10192  * to grow or shrink our TSB, depending upon available memory
10193  * conditions.
10194  */
10195 static void
10196 sfmmu_check_page_sizes(sfmmu_t *sfmmup, int growing)
10197 {
10198 	uint64_t ttecnt[MMU_PAGE_SIZES];
10199 	uint64_t tte8k_cnt, tte4m_cnt;
10200 	uint8_t i;
10201 	int sectsb_thresh;
10202 
10203 	/*
10204 	 * Kernel threads, processes with small address spaces not using
10205 	 * large pages, and dummy ISM HATs need not apply.
10206 	 */
10207 	if (sfmmup == ksfmmup || sfmmup->sfmmu_ismhat != NULL)
10208 		return;
10209 
10210 	if (!SFMMU_LGPGS_INUSE(sfmmup) &&
10211 	    sfmmup->sfmmu_ttecnt[TTE8K] <= tsb_rss_factor)
10212 		return;
10213 
10214 	for (i = 0; i < mmu_page_sizes; i++) {
10215 		ttecnt[i] = sfmmup->sfmmu_ttecnt[i] +
10216 		    sfmmup->sfmmu_ismttecnt[i];
10217 	}
10218 
10219 	/* Check pagesizes in use, and possibly reprogram DTLB. */
10220 	if (&mmu_check_page_sizes)
10221 		mmu_check_page_sizes(sfmmup, ttecnt);
10222 
10223 	/*
10224 	 * Calculate the number of 8k ttes to represent the span of these
10225 	 * pages.
10226 	 */
10227 	tte8k_cnt = ttecnt[TTE8K] +
10228 	    (ttecnt[TTE64K] << (MMU_PAGESHIFT64K - MMU_PAGESHIFT)) +
10229 	    (ttecnt[TTE512K] << (MMU_PAGESHIFT512K - MMU_PAGESHIFT));
10230 	if (mmu_page_sizes == max_mmu_page_sizes) {
10231 		tte4m_cnt = ttecnt[TTE4M] +
10232 		    (ttecnt[TTE32M] << (MMU_PAGESHIFT32M - MMU_PAGESHIFT4M)) +
10233 		    (ttecnt[TTE256M] << (MMU_PAGESHIFT256M - MMU_PAGESHIFT4M));
10234 	} else {
10235 		tte4m_cnt = ttecnt[TTE4M];
10236 	}
10237 
10238 	/*
10239 	 * Inflate tte8k_cnt to allow for region large page allocation failure.
10240 	 */
10241 	tte8k_cnt += sfmmup->sfmmu_tsb0_4minflcnt;
10242 
10243 	/*
10244 	 * Inflate TSB sizes by a factor of 2 if this process
10245 	 * uses 4M text pages to minimize extra conflict misses
10246 	 * in the first TSB since without counting text pages
10247 	 * 8K TSB may become too small.
10248 	 *
10249 	 * Also double the size of the second TSB to minimize
10250 	 * extra conflict misses due to competition between 4M text pages
10251 	 * and data pages.
10252 	 *
10253 	 * We need to adjust the second TSB allocation threshold by the
10254 	 * inflation factor, since there is no point in creating a second
10255 	 * TSB when we know all the mappings can fit in the I/D TLBs.
10256 	 */
10257 	sectsb_thresh = tsb_sectsb_threshold;
10258 	if (sfmmup->sfmmu_flags & HAT_4MTEXT_FLAG) {
10259 		tte8k_cnt <<= 1;
10260 		tte4m_cnt <<= 1;
10261 		sectsb_thresh <<= 1;
10262 	}
10263 
10264 	/*
10265 	 * Check to see if our TSB is the right size; we may need to
10266 	 * grow or shrink it.  If the process is small, our work is
10267 	 * finished at this point.
10268 	 */
10269 	if (tte8k_cnt <= tsb_rss_factor && tte4m_cnt <= sectsb_thresh) {
10270 		return;
10271 	}
10272 	sfmmu_size_tsb(sfmmup, growing, tte8k_cnt, tte4m_cnt, sectsb_thresh);
10273 }
10274 
10275 static void
10276 sfmmu_size_tsb(sfmmu_t *sfmmup, int growing, uint64_t tte8k_cnt,
10277 	uint64_t tte4m_cnt, int sectsb_thresh)
10278 {
10279 	int tsb_bits;
10280 	uint_t tsb_szc;
10281 	struct tsb_info *tsbinfop;
10282 	hatlock_t *hatlockp = NULL;
10283 
10284 	hatlockp = sfmmu_hat_enter(sfmmup);
10285 	ASSERT(hatlockp != NULL);
10286 	tsbinfop = sfmmup->sfmmu_tsb;
10287 	ASSERT(tsbinfop != NULL);
10288 
10289 	/*
10290 	 * If we're growing, select the size based on RSS.  If we're
10291 	 * shrinking, leave some room so we don't have to turn around and
10292 	 * grow again immediately.
10293 	 */
10294 	if (growing)
10295 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
10296 	else
10297 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt << 1);
10298 
10299 	if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10300 	    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10301 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10302 		    hatlockp, TSB_SHRINK);
10303 	} else if (growing && tsb_szc > tsbinfop->tsb_szc && TSB_OK_GROW()) {
10304 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10305 		    hatlockp, TSB_GROW);
10306 	}
10307 	tsbinfop = sfmmup->sfmmu_tsb;
10308 
10309 	/*
10310 	 * With the TLB and first TSB out of the way, we need to see if
10311 	 * we need a second TSB for 4M pages.  If we managed to reprogram
10312 	 * the TLB page sizes above, the process will start using this new
10313 	 * TSB right away; otherwise, it will start using it on the next
10314 	 * context switch.  Either way, it's no big deal so there's no
10315 	 * synchronization with the trap handlers here unless we grow the
10316 	 * TSB (in which case it's required to prevent using the old one
10317 	 * after it's freed). Note: second tsb is required for 32M/256M
10318 	 * page sizes.
10319 	 */
10320 	if (tte4m_cnt > sectsb_thresh) {
10321 		/*
10322 		 * If we're growing, select the size based on RSS.  If we're
10323 		 * shrinking, leave some room so we don't have to turn
10324 		 * around and grow again immediately.
10325 		 */
10326 		if (growing)
10327 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
10328 		else
10329 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt << 1);
10330 		if (tsbinfop->tsb_next == NULL) {
10331 			struct tsb_info *newtsb;
10332 			int allocflags = SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)?
10333 			    0 : TSB_ALLOC;
10334 
10335 			sfmmu_hat_exit(hatlockp);
10336 
10337 			/*
10338 			 * Try to allocate a TSB for 4[32|256]M pages.  If we
10339 			 * can't get the size we want, retry w/a minimum sized
10340 			 * TSB.  If that still didn't work, give up; we can
10341 			 * still run without one.
10342 			 */
10343 			tsb_bits = (mmu_page_sizes == max_mmu_page_sizes)?
10344 			    TSB4M|TSB32M|TSB256M:TSB4M;
10345 			if ((sfmmu_tsbinfo_alloc(&newtsb, tsb_szc, tsb_bits,
10346 			    allocflags, sfmmup)) &&
10347 			    (tsb_szc <= TSB_4M_SZCODE ||
10348 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
10349 			    tsb_bits, allocflags, sfmmup)) &&
10350 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_MIN_SZCODE,
10351 			    tsb_bits, allocflags, sfmmup)) {
10352 				return;
10353 			}
10354 
10355 			hatlockp = sfmmu_hat_enter(sfmmup);
10356 
10357 			sfmmu_invalidate_ctx(sfmmup);
10358 
10359 			if (sfmmup->sfmmu_tsb->tsb_next == NULL) {
10360 				sfmmup->sfmmu_tsb->tsb_next = newtsb;
10361 				SFMMU_STAT(sf_tsb_sectsb_create);
10362 				sfmmu_hat_exit(hatlockp);
10363 				return;
10364 			} else {
10365 				/*
10366 				 * It's annoying, but possible for us
10367 				 * to get here.. we dropped the HAT lock
10368 				 * because of locking order in the kmem
10369 				 * allocator, and while we were off getting
10370 				 * our memory, some other thread decided to
10371 				 * do us a favor and won the race to get a
10372 				 * second TSB for this process.  Sigh.
10373 				 */
10374 				sfmmu_hat_exit(hatlockp);
10375 				sfmmu_tsbinfo_free(newtsb);
10376 				return;
10377 			}
10378 		}
10379 
10380 		/*
10381 		 * We have a second TSB, see if it's big enough.
10382 		 */
10383 		tsbinfop = tsbinfop->tsb_next;
10384 
10385 		/*
10386 		 * Check to see if our second TSB is the right size;
10387 		 * we may need to grow or shrink it.
10388 		 * To prevent thrashing (e.g. growing the TSB on a
10389 		 * subsequent map operation), only try to shrink if
10390 		 * the TSB reach exceeds twice the virtual address
10391 		 * space size.
10392 		 */
10393 		if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10394 		    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10395 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10396 			    tsb_szc, hatlockp, TSB_SHRINK);
10397 		} else if (growing && tsb_szc > tsbinfop->tsb_szc &&
10398 		    TSB_OK_GROW()) {
10399 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10400 			    tsb_szc, hatlockp, TSB_GROW);
10401 		}
10402 	}
10403 
10404 	sfmmu_hat_exit(hatlockp);
10405 }
10406 
10407 /*
10408  * Free up a sfmmu
10409  * Since the sfmmu is currently embedded in the hat struct we simply zero
10410  * out our fields and free up the ism map blk list if any.
10411  */
10412 static void
10413 sfmmu_free_sfmmu(sfmmu_t *sfmmup)
10414 {
10415 	ism_blk_t	*blkp, *nx_blkp;
10416 #ifdef	DEBUG
10417 	ism_map_t	*map;
10418 	int 		i;
10419 #endif
10420 
10421 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
10422 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
10423 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
10424 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
10425 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
10426 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
10427 	ASSERT(SF_RGNMAP_ISNULL(sfmmup));
10428 
10429 	sfmmup->sfmmu_free = 0;
10430 	sfmmup->sfmmu_ismhat = 0;
10431 
10432 	blkp = sfmmup->sfmmu_iblk;
10433 	sfmmup->sfmmu_iblk = NULL;
10434 
10435 	while (blkp) {
10436 #ifdef	DEBUG
10437 		map = blkp->iblk_maps;
10438 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
10439 			ASSERT(map[i].imap_seg == 0);
10440 			ASSERT(map[i].imap_ismhat == NULL);
10441 			ASSERT(map[i].imap_ment == NULL);
10442 		}
10443 #endif
10444 		nx_blkp = blkp->iblk_next;
10445 		blkp->iblk_next = NULL;
10446 		blkp->iblk_nextpa = (uint64_t)-1;
10447 		kmem_cache_free(ism_blk_cache, blkp);
10448 		blkp = nx_blkp;
10449 	}
10450 }
10451 
10452 /*
10453  * Locking primitves accessed by HATLOCK macros
10454  */
10455 
10456 #define	SFMMU_SPL_MTX	(0x0)
10457 #define	SFMMU_ML_MTX	(0x1)
10458 
10459 #define	SFMMU_MLSPL_MTX(type, pg)	(((type) == SFMMU_SPL_MTX) ? \
10460 					    SPL_HASH(pg) : MLIST_HASH(pg))
10461 
10462 kmutex_t *
10463 sfmmu_page_enter(struct page *pp)
10464 {
10465 	return (sfmmu_mlspl_enter(pp, SFMMU_SPL_MTX));
10466 }
10467 
10468 void
10469 sfmmu_page_exit(kmutex_t *spl)
10470 {
10471 	mutex_exit(spl);
10472 }
10473 
10474 int
10475 sfmmu_page_spl_held(struct page *pp)
10476 {
10477 	return (sfmmu_mlspl_held(pp, SFMMU_SPL_MTX));
10478 }
10479 
10480 kmutex_t *
10481 sfmmu_mlist_enter(struct page *pp)
10482 {
10483 	return (sfmmu_mlspl_enter(pp, SFMMU_ML_MTX));
10484 }
10485 
10486 void
10487 sfmmu_mlist_exit(kmutex_t *mml)
10488 {
10489 	mutex_exit(mml);
10490 }
10491 
10492 int
10493 sfmmu_mlist_held(struct page *pp)
10494 {
10495 
10496 	return (sfmmu_mlspl_held(pp, SFMMU_ML_MTX));
10497 }
10498 
10499 /*
10500  * Common code for sfmmu_mlist_enter() and sfmmu_page_enter().  For
10501  * sfmmu_mlist_enter() case mml_table lock array is used and for
10502  * sfmmu_page_enter() sfmmu_page_lock lock array is used.
10503  *
10504  * The lock is taken on a root page so that it protects an operation on all
10505  * constituent pages of a large page pp belongs to.
10506  *
10507  * The routine takes a lock from the appropriate array. The lock is determined
10508  * by hashing the root page. After taking the lock this routine checks if the
10509  * root page has the same size code that was used to determine the root (i.e
10510  * that root hasn't changed).  If root page has the expected p_szc field we
10511  * have the right lock and it's returned to the caller. If root's p_szc
10512  * decreased we release the lock and retry from the beginning.  This case can
10513  * happen due to hat_page_demote() decreasing p_szc between our load of p_szc
10514  * value and taking the lock. The number of retries due to p_szc decrease is
10515  * limited by the maximum p_szc value. If p_szc is 0 we return the lock
10516  * determined by hashing pp itself.
10517  *
10518  * If our caller doesn't hold a SE_SHARED or SE_EXCL lock on pp it's also
10519  * possible that p_szc can increase. To increase p_szc a thread has to lock
10520  * all constituent pages EXCL and do hat_pageunload() on all of them. All the
10521  * callers that don't hold a page locked recheck if hmeblk through which pp
10522  * was found still maps this pp.  If it doesn't map it anymore returned lock
10523  * is immediately dropped. Therefore if sfmmu_mlspl_enter() hits the case of
10524  * p_szc increase after taking the lock it returns this lock without further
10525  * retries because in this case the caller doesn't care about which lock was
10526  * taken. The caller will drop it right away.
10527  *
10528  * After the routine returns it's guaranteed that hat_page_demote() can't
10529  * change p_szc field of any of constituent pages of a large page pp belongs
10530  * to as long as pp was either locked at least SHARED prior to this call or
10531  * the caller finds that hment that pointed to this pp still references this
10532  * pp (this also assumes that the caller holds hme hash bucket lock so that
10533  * the same pp can't be remapped into the same hmeblk after it was unmapped by
10534  * hat_pageunload()).
10535  */
10536 static kmutex_t *
10537 sfmmu_mlspl_enter(struct page *pp, int type)
10538 {
10539 	kmutex_t	*mtx;
10540 	uint_t		prev_rszc = UINT_MAX;
10541 	page_t		*rootpp;
10542 	uint_t		szc;
10543 	uint_t		rszc;
10544 	uint_t		pszc = pp->p_szc;
10545 
10546 	ASSERT(pp != NULL);
10547 
10548 again:
10549 	if (pszc == 0) {
10550 		mtx = SFMMU_MLSPL_MTX(type, pp);
10551 		mutex_enter(mtx);
10552 		return (mtx);
10553 	}
10554 
10555 	/* The lock lives in the root page */
10556 	rootpp = PP_GROUPLEADER(pp, pszc);
10557 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10558 	mutex_enter(mtx);
10559 
10560 	/*
10561 	 * Return mml in the following 3 cases:
10562 	 *
10563 	 * 1) If pp itself is root since if its p_szc decreased before we took
10564 	 * the lock pp is still the root of smaller szc page. And if its p_szc
10565 	 * increased it doesn't matter what lock we return (see comment in
10566 	 * front of this routine).
10567 	 *
10568 	 * 2) If pp's not root but rootpp is the root of a rootpp->p_szc size
10569 	 * large page we have the right lock since any previous potential
10570 	 * hat_page_demote() is done demoting from greater than current root's
10571 	 * p_szc because hat_page_demote() changes root's p_szc last. No
10572 	 * further hat_page_demote() can start or be in progress since it
10573 	 * would need the same lock we currently hold.
10574 	 *
10575 	 * 3) If rootpp's p_szc increased since previous iteration it doesn't
10576 	 * matter what lock we return (see comment in front of this routine).
10577 	 */
10578 	if (pp == rootpp || (rszc = rootpp->p_szc) == pszc ||
10579 	    rszc >= prev_rszc) {
10580 		return (mtx);
10581 	}
10582 
10583 	/*
10584 	 * hat_page_demote() could have decreased root's p_szc.
10585 	 * In this case pp's p_szc must also be smaller than pszc.
10586 	 * Retry.
10587 	 */
10588 	if (rszc < pszc) {
10589 		szc = pp->p_szc;
10590 		if (szc < pszc) {
10591 			mutex_exit(mtx);
10592 			pszc = szc;
10593 			goto again;
10594 		}
10595 		/*
10596 		 * pp's p_szc increased after it was decreased.
10597 		 * page cannot be mapped. Return current lock. The caller
10598 		 * will drop it right away.
10599 		 */
10600 		return (mtx);
10601 	}
10602 
10603 	/*
10604 	 * root's p_szc is greater than pp's p_szc.
10605 	 * hat_page_demote() is not done with all pages
10606 	 * yet. Wait for it to complete.
10607 	 */
10608 	mutex_exit(mtx);
10609 	rootpp = PP_GROUPLEADER(rootpp, rszc);
10610 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10611 	mutex_enter(mtx);
10612 	mutex_exit(mtx);
10613 	prev_rszc = rszc;
10614 	goto again;
10615 }
10616 
10617 static int
10618 sfmmu_mlspl_held(struct page *pp, int type)
10619 {
10620 	kmutex_t	*mtx;
10621 
10622 	ASSERT(pp != NULL);
10623 	/* The lock lives in the root page */
10624 	pp = PP_PAGEROOT(pp);
10625 	ASSERT(pp != NULL);
10626 
10627 	mtx = SFMMU_MLSPL_MTX(type, pp);
10628 	return (MUTEX_HELD(mtx));
10629 }
10630 
10631 static uint_t
10632 sfmmu_get_free_hblk(struct hme_blk **hmeblkpp, uint_t critical)
10633 {
10634 	struct  hme_blk *hblkp;
10635 
10636 
10637 	if (freehblkp != NULL) {
10638 		mutex_enter(&freehblkp_lock);
10639 		if (freehblkp != NULL) {
10640 			/*
10641 			 * If the current thread is owning hblk_reserve OR
10642 			 * critical request from sfmmu_hblk_steal()
10643 			 * let it succeed even if freehblkcnt is really low.
10644 			 */
10645 			if (freehblkcnt <= HBLK_RESERVE_MIN && !critical) {
10646 				SFMMU_STAT(sf_get_free_throttle);
10647 				mutex_exit(&freehblkp_lock);
10648 				return (0);
10649 			}
10650 			freehblkcnt--;
10651 			*hmeblkpp = freehblkp;
10652 			hblkp = *hmeblkpp;
10653 			freehblkp = hblkp->hblk_next;
10654 			mutex_exit(&freehblkp_lock);
10655 			hblkp->hblk_next = NULL;
10656 			SFMMU_STAT(sf_get_free_success);
10657 
10658 			ASSERT(hblkp->hblk_hmecnt == 0);
10659 			ASSERT(hblkp->hblk_vcnt == 0);
10660 			ASSERT(hblkp->hblk_nextpa == va_to_pa((caddr_t)hblkp));
10661 
10662 			return (1);
10663 		}
10664 		mutex_exit(&freehblkp_lock);
10665 	}
10666 
10667 	/* Check cpu hblk pending queues */
10668 	if ((*hmeblkpp = sfmmu_check_pending_hblks(TTE8K)) != NULL) {
10669 		hblkp = *hmeblkpp;
10670 		hblkp->hblk_next = NULL;
10671 		hblkp->hblk_nextpa = va_to_pa((caddr_t)hblkp);
10672 
10673 		ASSERT(hblkp->hblk_hmecnt == 0);
10674 		ASSERT(hblkp->hblk_vcnt == 0);
10675 
10676 		return (1);
10677 	}
10678 
10679 	SFMMU_STAT(sf_get_free_fail);
10680 	return (0);
10681 }
10682 
10683 static uint_t
10684 sfmmu_put_free_hblk(struct hme_blk *hmeblkp, uint_t critical)
10685 {
10686 	struct  hme_blk *hblkp;
10687 
10688 	ASSERT(hmeblkp->hblk_hmecnt == 0);
10689 	ASSERT(hmeblkp->hblk_vcnt == 0);
10690 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
10691 
10692 	/*
10693 	 * If the current thread is mapping into kernel space,
10694 	 * let it succede even if freehblkcnt is max
10695 	 * so that it will avoid freeing it to kmem.
10696 	 * This will prevent stack overflow due to
10697 	 * possible recursion since kmem_cache_free()
10698 	 * might require creation of a slab which
10699 	 * in turn needs an hmeblk to map that slab;
10700 	 * let's break this vicious chain at the first
10701 	 * opportunity.
10702 	 */
10703 	if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10704 		mutex_enter(&freehblkp_lock);
10705 		if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10706 			SFMMU_STAT(sf_put_free_success);
10707 			freehblkcnt++;
10708 			hmeblkp->hblk_next = freehblkp;
10709 			freehblkp = hmeblkp;
10710 			mutex_exit(&freehblkp_lock);
10711 			return (1);
10712 		}
10713 		mutex_exit(&freehblkp_lock);
10714 	}
10715 
10716 	/*
10717 	 * Bring down freehblkcnt to HBLK_RESERVE_CNT. We are here
10718 	 * only if freehblkcnt is at least HBLK_RESERVE_CNT *and*
10719 	 * we are not in the process of mapping into kernel space.
10720 	 */
10721 	ASSERT(!critical);
10722 	while (freehblkcnt > HBLK_RESERVE_CNT) {
10723 		mutex_enter(&freehblkp_lock);
10724 		if (freehblkcnt > HBLK_RESERVE_CNT) {
10725 			freehblkcnt--;
10726 			hblkp = freehblkp;
10727 			freehblkp = hblkp->hblk_next;
10728 			mutex_exit(&freehblkp_lock);
10729 			ASSERT(get_hblk_cache(hblkp) == sfmmu8_cache);
10730 			kmem_cache_free(sfmmu8_cache, hblkp);
10731 			continue;
10732 		}
10733 		mutex_exit(&freehblkp_lock);
10734 	}
10735 	SFMMU_STAT(sf_put_free_fail);
10736 	return (0);
10737 }
10738 
10739 static void
10740 sfmmu_hblk_swap(struct hme_blk *new)
10741 {
10742 	struct hme_blk *old, *hblkp, *prev;
10743 	uint64_t newpa;
10744 	caddr_t	base, vaddr, endaddr;
10745 	struct hmehash_bucket *hmebp;
10746 	struct sf_hment *osfhme, *nsfhme;
10747 	page_t *pp;
10748 	kmutex_t *pml;
10749 	tte_t tte;
10750 	struct hme_blk *list = NULL;
10751 
10752 #ifdef	DEBUG
10753 	hmeblk_tag		hblktag;
10754 	struct hme_blk		*found;
10755 #endif
10756 	old = HBLK_RESERVE;
10757 	ASSERT(!old->hblk_shared);
10758 
10759 	/*
10760 	 * save pa before bcopy clobbers it
10761 	 */
10762 	newpa = new->hblk_nextpa;
10763 
10764 	base = (caddr_t)get_hblk_base(old);
10765 	endaddr = base + get_hblk_span(old);
10766 
10767 	/*
10768 	 * acquire hash bucket lock.
10769 	 */
10770 	hmebp = sfmmu_tteload_acquire_hashbucket(ksfmmup, base, TTE8K,
10771 	    SFMMU_INVALID_SHMERID);
10772 
10773 	/*
10774 	 * copy contents from old to new
10775 	 */
10776 	bcopy((void *)old, (void *)new, HME8BLK_SZ);
10777 
10778 	/*
10779 	 * add new to hash chain
10780 	 */
10781 	sfmmu_hblk_hash_add(hmebp, new, newpa);
10782 
10783 	/*
10784 	 * search hash chain for hblk_reserve; this needs to be performed
10785 	 * after adding new, otherwise prev won't correspond to the hblk which
10786 	 * is prior to old in hash chain when we call sfmmu_hblk_hash_rm to
10787 	 * remove old later.
10788 	 */
10789 	for (prev = NULL,
10790 	    hblkp = hmebp->hmeblkp; hblkp != NULL && hblkp != old;
10791 	    prev = hblkp, hblkp = hblkp->hblk_next)
10792 		;
10793 
10794 	if (hblkp != old)
10795 		panic("sfmmu_hblk_swap: hblk_reserve not found");
10796 
10797 	/*
10798 	 * p_mapping list is still pointing to hments in hblk_reserve;
10799 	 * fix up p_mapping list so that they point to hments in new.
10800 	 *
10801 	 * Since all these mappings are created by hblk_reserve_thread
10802 	 * on the way and it's using at least one of the buffers from each of
10803 	 * the newly minted slabs, there is no danger of any of these
10804 	 * mappings getting unloaded by another thread.
10805 	 *
10806 	 * tsbmiss could only modify ref/mod bits of hments in old/new.
10807 	 * Since all of these hments hold mappings established by segkmem
10808 	 * and mappings in segkmem are setup with HAT_NOSYNC, ref/mod bits
10809 	 * have no meaning for the mappings in hblk_reserve.  hments in
10810 	 * old and new are identical except for ref/mod bits.
10811 	 */
10812 	for (vaddr = base; vaddr < endaddr; vaddr += TTEBYTES(TTE8K)) {
10813 
10814 		HBLKTOHME(osfhme, old, vaddr);
10815 		sfmmu_copytte(&osfhme->hme_tte, &tte);
10816 
10817 		if (TTE_IS_VALID(&tte)) {
10818 			if ((pp = osfhme->hme_page) == NULL)
10819 				panic("sfmmu_hblk_swap: page not mapped");
10820 
10821 			pml = sfmmu_mlist_enter(pp);
10822 
10823 			if (pp != osfhme->hme_page)
10824 				panic("sfmmu_hblk_swap: mapping changed");
10825 
10826 			HBLKTOHME(nsfhme, new, vaddr);
10827 
10828 			HME_ADD(nsfhme, pp);
10829 			HME_SUB(osfhme, pp);
10830 
10831 			sfmmu_mlist_exit(pml);
10832 		}
10833 	}
10834 
10835 	/*
10836 	 * remove old from hash chain
10837 	 */
10838 	sfmmu_hblk_hash_rm(hmebp, old, prev, &list, 1);
10839 
10840 #ifdef	DEBUG
10841 
10842 	hblktag.htag_id = ksfmmup;
10843 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
10844 	hblktag.htag_bspage = HME_HASH_BSPAGE(base, HME_HASH_SHIFT(TTE8K));
10845 	hblktag.htag_rehash = HME_HASH_REHASH(TTE8K);
10846 	HME_HASH_FAST_SEARCH(hmebp, hblktag, found);
10847 
10848 	if (found != new)
10849 		panic("sfmmu_hblk_swap: new hblk not found");
10850 #endif
10851 
10852 	SFMMU_HASH_UNLOCK(hmebp);
10853 
10854 	/*
10855 	 * Reset hblk_reserve
10856 	 */
10857 	bzero((void *)old, HME8BLK_SZ);
10858 	old->hblk_nextpa = va_to_pa((caddr_t)old);
10859 }
10860 
10861 /*
10862  * Grab the mlist mutex for both pages passed in.
10863  *
10864  * low and high will be returned as pointers to the mutexes for these pages.
10865  * low refers to the mutex residing in the lower bin of the mlist hash, while
10866  * high refers to the mutex residing in the higher bin of the mlist hash.  This
10867  * is due to the locking order restrictions on the same thread grabbing
10868  * multiple mlist mutexes.  The low lock must be acquired before the high lock.
10869  *
10870  * If both pages hash to the same mutex, only grab that single mutex, and
10871  * high will be returned as NULL
10872  * If the pages hash to different bins in the hash, grab the lower addressed
10873  * lock first and then the higher addressed lock in order to follow the locking
10874  * rules involved with the same thread grabbing multiple mlist mutexes.
10875  * low and high will both have non-NULL values.
10876  */
10877 static void
10878 sfmmu_mlist_reloc_enter(struct page *targ, struct page *repl,
10879     kmutex_t **low, kmutex_t **high)
10880 {
10881 	kmutex_t	*mml_targ, *mml_repl;
10882 
10883 	/*
10884 	 * no need to do the dance around szc as in sfmmu_mlist_enter()
10885 	 * because this routine is only called by hat_page_relocate() and all
10886 	 * targ and repl pages are already locked EXCL so szc can't change.
10887 	 */
10888 
10889 	mml_targ = MLIST_HASH(PP_PAGEROOT(targ));
10890 	mml_repl = MLIST_HASH(PP_PAGEROOT(repl));
10891 
10892 	if (mml_targ == mml_repl) {
10893 		*low = mml_targ;
10894 		*high = NULL;
10895 	} else {
10896 		if (mml_targ < mml_repl) {
10897 			*low = mml_targ;
10898 			*high = mml_repl;
10899 		} else {
10900 			*low = mml_repl;
10901 			*high = mml_targ;
10902 		}
10903 	}
10904 
10905 	mutex_enter(*low);
10906 	if (*high)
10907 		mutex_enter(*high);
10908 }
10909 
10910 static void
10911 sfmmu_mlist_reloc_exit(kmutex_t *low, kmutex_t *high)
10912 {
10913 	if (high)
10914 		mutex_exit(high);
10915 	mutex_exit(low);
10916 }
10917 
10918 static hatlock_t *
10919 sfmmu_hat_enter(sfmmu_t *sfmmup)
10920 {
10921 	hatlock_t	*hatlockp;
10922 
10923 	if (sfmmup != ksfmmup) {
10924 		hatlockp = TSB_HASH(sfmmup);
10925 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
10926 		return (hatlockp);
10927 	}
10928 	return (NULL);
10929 }
10930 
10931 static hatlock_t *
10932 sfmmu_hat_tryenter(sfmmu_t *sfmmup)
10933 {
10934 	hatlock_t	*hatlockp;
10935 
10936 	if (sfmmup != ksfmmup) {
10937 		hatlockp = TSB_HASH(sfmmup);
10938 		if (mutex_tryenter(HATLOCK_MUTEXP(hatlockp)) == 0)
10939 			return (NULL);
10940 		return (hatlockp);
10941 	}
10942 	return (NULL);
10943 }
10944 
10945 static void
10946 sfmmu_hat_exit(hatlock_t *hatlockp)
10947 {
10948 	if (hatlockp != NULL)
10949 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
10950 }
10951 
10952 static void
10953 sfmmu_hat_lock_all(void)
10954 {
10955 	int i;
10956 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
10957 		mutex_enter(HATLOCK_MUTEXP(&hat_lock[i]));
10958 }
10959 
10960 static void
10961 sfmmu_hat_unlock_all(void)
10962 {
10963 	int i;
10964 	for (i = SFMMU_NUM_LOCK - 1; i >= 0; i--)
10965 		mutex_exit(HATLOCK_MUTEXP(&hat_lock[i]));
10966 }
10967 
10968 int
10969 sfmmu_hat_lock_held(sfmmu_t *sfmmup)
10970 {
10971 	ASSERT(sfmmup != ksfmmup);
10972 	return (MUTEX_HELD(HATLOCK_MUTEXP(TSB_HASH(sfmmup))));
10973 }
10974 
10975 /*
10976  * Locking primitives to provide consistency between ISM unmap
10977  * and other operations.  Since ISM unmap can take a long time, we
10978  * use HAT_ISMBUSY flag (protected by the hatlock) to avoid creating
10979  * contention on the hatlock buckets while ISM segments are being
10980  * unmapped.  The tradeoff is that the flags don't prevent priority
10981  * inversion from occurring, so we must request kernel priority in
10982  * case we have to sleep to keep from getting buried while holding
10983  * the HAT_ISMBUSY flag set, which in turn could block other kernel
10984  * threads from running (for example, in sfmmu_uvatopfn()).
10985  */
10986 static void
10987 sfmmu_ismhat_enter(sfmmu_t *sfmmup, int hatlock_held)
10988 {
10989 	hatlock_t *hatlockp;
10990 
10991 	THREAD_KPRI_REQUEST();
10992 	if (!hatlock_held)
10993 		hatlockp = sfmmu_hat_enter(sfmmup);
10994 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY))
10995 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
10996 	SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
10997 	if (!hatlock_held)
10998 		sfmmu_hat_exit(hatlockp);
10999 }
11000 
11001 static void
11002 sfmmu_ismhat_exit(sfmmu_t *sfmmup, int hatlock_held)
11003 {
11004 	hatlock_t *hatlockp;
11005 
11006 	if (!hatlock_held)
11007 		hatlockp = sfmmu_hat_enter(sfmmup);
11008 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
11009 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
11010 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11011 	if (!hatlock_held)
11012 		sfmmu_hat_exit(hatlockp);
11013 	THREAD_KPRI_RELEASE();
11014 }
11015 
11016 /*
11017  *
11018  * Algorithm:
11019  *
11020  * (1) if segkmem is not ready, allocate hblk from an array of pre-alloc'ed
11021  *	hblks.
11022  *
11023  * (2) if we are allocating an hblk for mapping a slab in sfmmu_cache,
11024  *
11025  * 		(a) try to return an hblk from reserve pool of free hblks;
11026  *		(b) if the reserve pool is empty, acquire hblk_reserve_lock
11027  *		    and return hblk_reserve.
11028  *
11029  * (3) call kmem_cache_alloc() to allocate hblk;
11030  *
11031  *		(a) if hblk_reserve_lock is held by the current thread,
11032  *		    atomically replace hblk_reserve by the hblk that is
11033  *		    returned by kmem_cache_alloc; release hblk_reserve_lock
11034  *		    and call kmem_cache_alloc() again.
11035  *		(b) if reserve pool is not full, add the hblk that is
11036  *		    returned by kmem_cache_alloc to reserve pool and
11037  *		    call kmem_cache_alloc again.
11038  *
11039  */
11040 static struct hme_blk *
11041 sfmmu_hblk_alloc(sfmmu_t *sfmmup, caddr_t vaddr,
11042 	struct hmehash_bucket *hmebp, uint_t size, hmeblk_tag hblktag,
11043 	uint_t flags, uint_t rid)
11044 {
11045 	struct hme_blk *hmeblkp = NULL;
11046 	struct hme_blk *newhblkp;
11047 	struct hme_blk *shw_hblkp = NULL;
11048 	struct kmem_cache *sfmmu_cache = NULL;
11049 	uint64_t hblkpa;
11050 	ulong_t index;
11051 	uint_t owner;		/* set to 1 if using hblk_reserve */
11052 	uint_t forcefree;
11053 	int sleep;
11054 	sf_srd_t *srdp;
11055 	sf_region_t *rgnp;
11056 
11057 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11058 	ASSERT(hblktag.htag_rid == rid);
11059 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
11060 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
11061 	    IS_P2ALIGNED(vaddr, TTEBYTES(size)));
11062 
11063 	/*
11064 	 * If segkmem is not created yet, allocate from static hmeblks
11065 	 * created at the end of startup_modules().  See the block comment
11066 	 * in startup_modules() describing how we estimate the number of
11067 	 * static hmeblks that will be needed during re-map.
11068 	 */
11069 	if (!hblk_alloc_dynamic) {
11070 
11071 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11072 
11073 		if (size == TTE8K) {
11074 			index = nucleus_hblk8.index;
11075 			if (index >= nucleus_hblk8.len) {
11076 				/*
11077 				 * If we panic here, see startup_modules() to
11078 				 * make sure that we are calculating the
11079 				 * number of hblk8's that we need correctly.
11080 				 */
11081 				prom_panic("no nucleus hblk8 to allocate");
11082 			}
11083 			hmeblkp =
11084 			    (struct hme_blk *)&nucleus_hblk8.list[index];
11085 			nucleus_hblk8.index++;
11086 			SFMMU_STAT(sf_hblk8_nalloc);
11087 		} else {
11088 			index = nucleus_hblk1.index;
11089 			if (nucleus_hblk1.index >= nucleus_hblk1.len) {
11090 				/*
11091 				 * If we panic here, see startup_modules().
11092 				 * Most likely you need to update the
11093 				 * calculation of the number of hblk1 elements
11094 				 * that the kernel needs to boot.
11095 				 */
11096 				prom_panic("no nucleus hblk1 to allocate");
11097 			}
11098 			hmeblkp =
11099 			    (struct hme_blk *)&nucleus_hblk1.list[index];
11100 			nucleus_hblk1.index++;
11101 			SFMMU_STAT(sf_hblk1_nalloc);
11102 		}
11103 
11104 		goto hblk_init;
11105 	}
11106 
11107 	SFMMU_HASH_UNLOCK(hmebp);
11108 
11109 	if (sfmmup != KHATID && !SFMMU_IS_SHMERID_VALID(rid)) {
11110 		if (mmu_page_sizes == max_mmu_page_sizes) {
11111 			if (size < TTE256M)
11112 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
11113 				    size, flags);
11114 		} else {
11115 			if (size < TTE4M)
11116 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
11117 				    size, flags);
11118 		}
11119 	} else if (SFMMU_IS_SHMERID_VALID(rid)) {
11120 		/*
11121 		 * Shared hmes use per region bitmaps in rgn_hmeflag
11122 		 * rather than shadow hmeblks to keep track of the
11123 		 * mapping sizes which have been allocated for the region.
11124 		 * Here we cleanup old invalid hmeblks with this rid,
11125 		 * which may be left around by pageunload().
11126 		 */
11127 		int ttesz;
11128 		caddr_t va;
11129 		caddr_t	eva = vaddr + TTEBYTES(size);
11130 
11131 		ASSERT(sfmmup != KHATID);
11132 
11133 		srdp = sfmmup->sfmmu_srdp;
11134 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11135 		rgnp = srdp->srd_hmergnp[rid];
11136 		ASSERT(rgnp != NULL && rgnp->rgn_id == rid);
11137 		ASSERT(rgnp->rgn_refcnt != 0);
11138 		ASSERT(size <= rgnp->rgn_pgszc);
11139 
11140 		ttesz = HBLK_MIN_TTESZ;
11141 		do {
11142 			if (!(rgnp->rgn_hmeflags & (0x1 << ttesz))) {
11143 				continue;
11144 			}
11145 
11146 			if (ttesz > size && ttesz != HBLK_MIN_TTESZ) {
11147 				sfmmu_cleanup_rhblk(srdp, vaddr, rid, ttesz);
11148 			} else if (ttesz < size) {
11149 				for (va = vaddr; va < eva;
11150 				    va += TTEBYTES(ttesz)) {
11151 					sfmmu_cleanup_rhblk(srdp, va, rid,
11152 					    ttesz);
11153 				}
11154 			}
11155 		} while (++ttesz <= rgnp->rgn_pgszc);
11156 	}
11157 
11158 fill_hblk:
11159 	owner = (hblk_reserve_thread == curthread) ? 1 : 0;
11160 
11161 	if (owner && size == TTE8K) {
11162 
11163 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11164 		/*
11165 		 * We are really in a tight spot. We already own
11166 		 * hblk_reserve and we need another hblk.  In anticipation
11167 		 * of this kind of scenario, we specifically set aside
11168 		 * HBLK_RESERVE_MIN number of hblks to be used exclusively
11169 		 * by owner of hblk_reserve.
11170 		 */
11171 		SFMMU_STAT(sf_hblk_recurse_cnt);
11172 
11173 		if (!sfmmu_get_free_hblk(&hmeblkp, 1))
11174 			panic("sfmmu_hblk_alloc: reserve list is empty");
11175 
11176 		goto hblk_verify;
11177 	}
11178 
11179 	ASSERT(!owner);
11180 
11181 	if ((flags & HAT_NO_KALLOC) == 0) {
11182 
11183 		sfmmu_cache = ((size == TTE8K) ? sfmmu8_cache : sfmmu1_cache);
11184 		sleep = ((sfmmup == KHATID) ? KM_NOSLEEP : KM_SLEEP);
11185 
11186 		if ((hmeblkp = kmem_cache_alloc(sfmmu_cache, sleep)) == NULL) {
11187 			hmeblkp = sfmmu_hblk_steal(size);
11188 		} else {
11189 			/*
11190 			 * if we are the owner of hblk_reserve,
11191 			 * swap hblk_reserve with hmeblkp and
11192 			 * start a fresh life.  Hope things go
11193 			 * better this time.
11194 			 */
11195 			if (hblk_reserve_thread == curthread) {
11196 				ASSERT(sfmmu_cache == sfmmu8_cache);
11197 				sfmmu_hblk_swap(hmeblkp);
11198 				hblk_reserve_thread = NULL;
11199 				mutex_exit(&hblk_reserve_lock);
11200 				goto fill_hblk;
11201 			}
11202 			/*
11203 			 * let's donate this hblk to our reserve list if
11204 			 * we are not mapping kernel range
11205 			 */
11206 			if (size == TTE8K && sfmmup != KHATID) {
11207 				if (sfmmu_put_free_hblk(hmeblkp, 0))
11208 					goto fill_hblk;
11209 			}
11210 		}
11211 	} else {
11212 		/*
11213 		 * We are here to map the slab in sfmmu8_cache; let's
11214 		 * check if we could tap our reserve list; if successful,
11215 		 * this will avoid the pain of going thru sfmmu_hblk_swap
11216 		 */
11217 		SFMMU_STAT(sf_hblk_slab_cnt);
11218 		if (!sfmmu_get_free_hblk(&hmeblkp, 0)) {
11219 			/*
11220 			 * let's start hblk_reserve dance
11221 			 */
11222 			SFMMU_STAT(sf_hblk_reserve_cnt);
11223 			owner = 1;
11224 			mutex_enter(&hblk_reserve_lock);
11225 			hmeblkp = HBLK_RESERVE;
11226 			hblk_reserve_thread = curthread;
11227 		}
11228 	}
11229 
11230 hblk_verify:
11231 	ASSERT(hmeblkp != NULL);
11232 	set_hblk_sz(hmeblkp, size);
11233 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
11234 	SFMMU_HASH_LOCK(hmebp);
11235 	HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11236 	if (newhblkp != NULL) {
11237 		SFMMU_HASH_UNLOCK(hmebp);
11238 		if (hmeblkp != HBLK_RESERVE) {
11239 			/*
11240 			 * This is really tricky!
11241 			 *
11242 			 * vmem_alloc(vmem_seg_arena)
11243 			 *  vmem_alloc(vmem_internal_arena)
11244 			 *   segkmem_alloc(heap_arena)
11245 			 *    vmem_alloc(heap_arena)
11246 			 *    page_create()
11247 			 *    hat_memload()
11248 			 *	kmem_cache_free()
11249 			 *	 kmem_cache_alloc()
11250 			 *	  kmem_slab_create()
11251 			 *	   vmem_alloc(kmem_internal_arena)
11252 			 *	    segkmem_alloc(heap_arena)
11253 			 *		vmem_alloc(heap_arena)
11254 			 *		page_create()
11255 			 *		hat_memload()
11256 			 *		  kmem_cache_free()
11257 			 *		...
11258 			 *
11259 			 * Thus, hat_memload() could call kmem_cache_free
11260 			 * for enough number of times that we could easily
11261 			 * hit the bottom of the stack or run out of reserve
11262 			 * list of vmem_seg structs.  So, we must donate
11263 			 * this hblk to reserve list if it's allocated
11264 			 * from sfmmu8_cache *and* mapping kernel range.
11265 			 * We don't need to worry about freeing hmeblk1's
11266 			 * to kmem since they don't map any kmem slabs.
11267 			 *
11268 			 * Note: When segkmem supports largepages, we must
11269 			 * free hmeblk1's to reserve list as well.
11270 			 */
11271 			forcefree = (sfmmup == KHATID) ? 1 : 0;
11272 			if (size == TTE8K &&
11273 			    sfmmu_put_free_hblk(hmeblkp, forcefree)) {
11274 				goto re_verify;
11275 			}
11276 			ASSERT(sfmmup != KHATID);
11277 			kmem_cache_free(get_hblk_cache(hmeblkp), hmeblkp);
11278 		} else {
11279 			/*
11280 			 * Hey! we don't need hblk_reserve any more.
11281 			 */
11282 			ASSERT(owner);
11283 			hblk_reserve_thread = NULL;
11284 			mutex_exit(&hblk_reserve_lock);
11285 			owner = 0;
11286 		}
11287 re_verify:
11288 		/*
11289 		 * let's check if the goodies are still present
11290 		 */
11291 		SFMMU_HASH_LOCK(hmebp);
11292 		HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11293 		if (newhblkp != NULL) {
11294 			/*
11295 			 * return newhblkp if it's not hblk_reserve;
11296 			 * if newhblkp is hblk_reserve, return it
11297 			 * _only if_ we are the owner of hblk_reserve.
11298 			 */
11299 			if (newhblkp != HBLK_RESERVE || owner) {
11300 				ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
11301 				    newhblkp->hblk_shared);
11302 				ASSERT(SFMMU_IS_SHMERID_VALID(rid) ||
11303 				    !newhblkp->hblk_shared);
11304 				return (newhblkp);
11305 			} else {
11306 				/*
11307 				 * we just hit hblk_reserve in the hash and
11308 				 * we are not the owner of that;
11309 				 *
11310 				 * block until hblk_reserve_thread completes
11311 				 * swapping hblk_reserve and try the dance
11312 				 * once again.
11313 				 */
11314 				SFMMU_HASH_UNLOCK(hmebp);
11315 				mutex_enter(&hblk_reserve_lock);
11316 				mutex_exit(&hblk_reserve_lock);
11317 				SFMMU_STAT(sf_hblk_reserve_hit);
11318 				goto fill_hblk;
11319 			}
11320 		} else {
11321 			/*
11322 			 * it's no more! try the dance once again.
11323 			 */
11324 			SFMMU_HASH_UNLOCK(hmebp);
11325 			goto fill_hblk;
11326 		}
11327 	}
11328 
11329 hblk_init:
11330 	if (SFMMU_IS_SHMERID_VALID(rid)) {
11331 		uint16_t tteflag = 0x1 <<
11332 		    ((size < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ : size);
11333 
11334 		if (!(rgnp->rgn_hmeflags & tteflag)) {
11335 			atomic_or_16(&rgnp->rgn_hmeflags, tteflag);
11336 		}
11337 		hmeblkp->hblk_shared = 1;
11338 	} else {
11339 		hmeblkp->hblk_shared = 0;
11340 	}
11341 	set_hblk_sz(hmeblkp, size);
11342 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11343 	hmeblkp->hblk_next = (struct hme_blk *)NULL;
11344 	hmeblkp->hblk_tag = hblktag;
11345 	hmeblkp->hblk_shadow = shw_hblkp;
11346 	hblkpa = hmeblkp->hblk_nextpa;
11347 	hmeblkp->hblk_nextpa = HMEBLK_ENDPA;
11348 
11349 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
11350 	ASSERT(get_hblk_span(hmeblkp) == HMEBLK_SPAN(size));
11351 	ASSERT(hmeblkp->hblk_hmecnt == 0);
11352 	ASSERT(hmeblkp->hblk_vcnt == 0);
11353 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11354 	ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
11355 	sfmmu_hblk_hash_add(hmebp, hmeblkp, hblkpa);
11356 	return (hmeblkp);
11357 }
11358 
11359 /*
11360  * This function cleans up the hme_blk and returns it to the free list.
11361  */
11362 /* ARGSUSED */
11363 static void
11364 sfmmu_hblk_free(struct hme_blk **listp)
11365 {
11366 	struct hme_blk *hmeblkp, *next_hmeblkp;
11367 	int		size;
11368 	uint_t		critical;
11369 	uint64_t	hblkpa;
11370 
11371 	ASSERT(*listp != NULL);
11372 
11373 	hmeblkp = *listp;
11374 	while (hmeblkp != NULL) {
11375 		next_hmeblkp = hmeblkp->hblk_next;
11376 		ASSERT(!hmeblkp->hblk_hmecnt);
11377 		ASSERT(!hmeblkp->hblk_vcnt);
11378 		ASSERT(!hmeblkp->hblk_lckcnt);
11379 		ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
11380 		ASSERT(hmeblkp->hblk_shared == 0);
11381 		ASSERT(hmeblkp->hblk_shw_bit == 0);
11382 		ASSERT(hmeblkp->hblk_shadow == NULL);
11383 
11384 		hblkpa = va_to_pa((caddr_t)hmeblkp);
11385 		ASSERT(hblkpa != (uint64_t)-1);
11386 		critical = (hblktosfmmu(hmeblkp) == KHATID) ? 1 : 0;
11387 
11388 		size = get_hblk_ttesz(hmeblkp);
11389 		hmeblkp->hblk_next = NULL;
11390 		hmeblkp->hblk_nextpa = hblkpa;
11391 
11392 		if (hmeblkp->hblk_nuc_bit == 0) {
11393 
11394 			if (size != TTE8K ||
11395 			    !sfmmu_put_free_hblk(hmeblkp, critical))
11396 				kmem_cache_free(get_hblk_cache(hmeblkp),
11397 				    hmeblkp);
11398 		}
11399 		hmeblkp = next_hmeblkp;
11400 	}
11401 }
11402 
11403 #define	BUCKETS_TO_SEARCH_BEFORE_UNLOAD	30
11404 #define	SFMMU_HBLK_STEAL_THRESHOLD 5
11405 
11406 static uint_t sfmmu_hblk_steal_twice;
11407 static uint_t sfmmu_hblk_steal_count, sfmmu_hblk_steal_unload_count;
11408 
11409 /*
11410  * Steal a hmeblk from user or kernel hme hash lists.
11411  * For 8K tte grab one from reserve pool (freehblkp) before proceeding to
11412  * steal and if we fail to steal after SFMMU_HBLK_STEAL_THRESHOLD attempts
11413  * tap into critical reserve of freehblkp.
11414  * Note: We remain looping in this routine until we find one.
11415  */
11416 static struct hme_blk *
11417 sfmmu_hblk_steal(int size)
11418 {
11419 	static struct hmehash_bucket *uhmehash_steal_hand = NULL;
11420 	struct hmehash_bucket *hmebp;
11421 	struct hme_blk *hmeblkp = NULL, *pr_hblk;
11422 	uint64_t hblkpa;
11423 	int i;
11424 	uint_t loop_cnt = 0, critical;
11425 
11426 	for (;;) {
11427 		/* Check cpu hblk pending queues */
11428 		if ((hmeblkp = sfmmu_check_pending_hblks(size)) != NULL) {
11429 			hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
11430 			ASSERT(hmeblkp->hblk_hmecnt == 0);
11431 			ASSERT(hmeblkp->hblk_vcnt == 0);
11432 			return (hmeblkp);
11433 		}
11434 
11435 		if (size == TTE8K) {
11436 			critical =
11437 			    (++loop_cnt > SFMMU_HBLK_STEAL_THRESHOLD) ? 1 : 0;
11438 			if (sfmmu_get_free_hblk(&hmeblkp, critical))
11439 				return (hmeblkp);
11440 		}
11441 
11442 		hmebp = (uhmehash_steal_hand == NULL) ? uhme_hash :
11443 		    uhmehash_steal_hand;
11444 		ASSERT(hmebp >= uhme_hash && hmebp <= &uhme_hash[UHMEHASH_SZ]);
11445 
11446 		for (i = 0; hmeblkp == NULL && i <= UHMEHASH_SZ +
11447 		    BUCKETS_TO_SEARCH_BEFORE_UNLOAD; i++) {
11448 			SFMMU_HASH_LOCK(hmebp);
11449 			hmeblkp = hmebp->hmeblkp;
11450 			hblkpa = hmebp->hmeh_nextpa;
11451 			pr_hblk = NULL;
11452 			while (hmeblkp) {
11453 				/*
11454 				 * check if it is a hmeblk that is not locked
11455 				 * and not shared. skip shadow hmeblks with
11456 				 * shadow_mask set i.e valid count non zero.
11457 				 */
11458 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11459 				    (hmeblkp->hblk_shw_bit == 0 ||
11460 				    hmeblkp->hblk_vcnt == 0) &&
11461 				    (hmeblkp->hblk_lckcnt == 0)) {
11462 					/*
11463 					 * there is a high probability that we
11464 					 * will find a free one. search some
11465 					 * buckets for a free hmeblk initially
11466 					 * before unloading a valid hmeblk.
11467 					 */
11468 					if ((hmeblkp->hblk_vcnt == 0 &&
11469 					    hmeblkp->hblk_hmecnt == 0) || (i >=
11470 					    BUCKETS_TO_SEARCH_BEFORE_UNLOAD)) {
11471 						if (sfmmu_steal_this_hblk(hmebp,
11472 						    hmeblkp, hblkpa, pr_hblk)) {
11473 							/*
11474 							 * Hblk is unloaded
11475 							 * successfully
11476 							 */
11477 							break;
11478 						}
11479 					}
11480 				}
11481 				pr_hblk = hmeblkp;
11482 				hblkpa = hmeblkp->hblk_nextpa;
11483 				hmeblkp = hmeblkp->hblk_next;
11484 			}
11485 
11486 			SFMMU_HASH_UNLOCK(hmebp);
11487 			if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
11488 				hmebp = uhme_hash;
11489 		}
11490 		uhmehash_steal_hand = hmebp;
11491 
11492 		if (hmeblkp != NULL)
11493 			break;
11494 
11495 		/*
11496 		 * in the worst case, look for a free one in the kernel
11497 		 * hash table.
11498 		 */
11499 		for (i = 0, hmebp = khme_hash; i <= KHMEHASH_SZ; i++) {
11500 			SFMMU_HASH_LOCK(hmebp);
11501 			hmeblkp = hmebp->hmeblkp;
11502 			hblkpa = hmebp->hmeh_nextpa;
11503 			pr_hblk = NULL;
11504 			while (hmeblkp) {
11505 				/*
11506 				 * check if it is free hmeblk
11507 				 */
11508 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11509 				    (hmeblkp->hblk_lckcnt == 0) &&
11510 				    (hmeblkp->hblk_vcnt == 0) &&
11511 				    (hmeblkp->hblk_hmecnt == 0)) {
11512 					if (sfmmu_steal_this_hblk(hmebp,
11513 					    hmeblkp, hblkpa, pr_hblk)) {
11514 						break;
11515 					} else {
11516 						/*
11517 						 * Cannot fail since we have
11518 						 * hash lock.
11519 						 */
11520 						panic("fail to steal?");
11521 					}
11522 				}
11523 
11524 				pr_hblk = hmeblkp;
11525 				hblkpa = hmeblkp->hblk_nextpa;
11526 				hmeblkp = hmeblkp->hblk_next;
11527 			}
11528 
11529 			SFMMU_HASH_UNLOCK(hmebp);
11530 			if (hmebp++ == &khme_hash[KHMEHASH_SZ])
11531 				hmebp = khme_hash;
11532 		}
11533 
11534 		if (hmeblkp != NULL)
11535 			break;
11536 		sfmmu_hblk_steal_twice++;
11537 	}
11538 	return (hmeblkp);
11539 }
11540 
11541 /*
11542  * This routine does real work to prepare a hblk to be "stolen" by
11543  * unloading the mappings, updating shadow counts ....
11544  * It returns 1 if the block is ready to be reused (stolen), or 0
11545  * means the block cannot be stolen yet- pageunload is still working
11546  * on this hblk.
11547  */
11548 static int
11549 sfmmu_steal_this_hblk(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
11550 	uint64_t hblkpa, struct hme_blk *pr_hblk)
11551 {
11552 	int shw_size, vshift;
11553 	struct hme_blk *shw_hblkp;
11554 	caddr_t vaddr;
11555 	uint_t shw_mask, newshw_mask;
11556 	struct hme_blk *list = NULL;
11557 
11558 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11559 
11560 	/*
11561 	 * check if the hmeblk is free, unload if necessary
11562 	 */
11563 	if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11564 		sfmmu_t *sfmmup;
11565 		demap_range_t dmr;
11566 
11567 		sfmmup = hblktosfmmu(hmeblkp);
11568 		if (hmeblkp->hblk_shared || sfmmup->sfmmu_ismhat) {
11569 			return (0);
11570 		}
11571 		DEMAP_RANGE_INIT(sfmmup, &dmr);
11572 		(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
11573 		    (caddr_t)get_hblk_base(hmeblkp),
11574 		    get_hblk_endaddr(hmeblkp), &dmr, HAT_UNLOAD);
11575 		DEMAP_RANGE_FLUSH(&dmr);
11576 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11577 			/*
11578 			 * Pageunload is working on the same hblk.
11579 			 */
11580 			return (0);
11581 		}
11582 
11583 		sfmmu_hblk_steal_unload_count++;
11584 	}
11585 
11586 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11587 	ASSERT(hmeblkp->hblk_vcnt == 0 && hmeblkp->hblk_hmecnt == 0);
11588 
11589 	sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 1);
11590 	hmeblkp->hblk_nextpa = hblkpa;
11591 
11592 	shw_hblkp = hmeblkp->hblk_shadow;
11593 	if (shw_hblkp) {
11594 		ASSERT(!hmeblkp->hblk_shared);
11595 		shw_size = get_hblk_ttesz(shw_hblkp);
11596 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
11597 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
11598 		ASSERT(vshift < 8);
11599 		/*
11600 		 * Atomically clear shadow mask bit
11601 		 */
11602 		do {
11603 			shw_mask = shw_hblkp->hblk_shw_mask;
11604 			ASSERT(shw_mask & (1 << vshift));
11605 			newshw_mask = shw_mask & ~(1 << vshift);
11606 			newshw_mask = cas32(&shw_hblkp->hblk_shw_mask,
11607 			    shw_mask, newshw_mask);
11608 		} while (newshw_mask != shw_mask);
11609 		hmeblkp->hblk_shadow = NULL;
11610 	}
11611 
11612 	/*
11613 	 * remove shadow bit if we are stealing an unused shadow hmeblk.
11614 	 * sfmmu_hblk_alloc needs it that way, will set shadow bit later if
11615 	 * we are indeed allocating a shadow hmeblk.
11616 	 */
11617 	hmeblkp->hblk_shw_bit = 0;
11618 
11619 	if (hmeblkp->hblk_shared) {
11620 		sf_srd_t	*srdp;
11621 		sf_region_t	*rgnp;
11622 		uint_t		rid;
11623 
11624 		srdp = hblktosrd(hmeblkp);
11625 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11626 		rid = hmeblkp->hblk_tag.htag_rid;
11627 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
11628 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
11629 		rgnp = srdp->srd_hmergnp[rid];
11630 		ASSERT(rgnp != NULL);
11631 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
11632 		hmeblkp->hblk_shared = 0;
11633 	}
11634 
11635 	sfmmu_hblk_steal_count++;
11636 	SFMMU_STAT(sf_steal_count);
11637 
11638 	return (1);
11639 }
11640 
11641 struct hme_blk *
11642 sfmmu_hmetohblk(struct sf_hment *sfhme)
11643 {
11644 	struct hme_blk *hmeblkp;
11645 	struct sf_hment *sfhme0;
11646 	struct hme_blk *hblk_dummy = 0;
11647 
11648 	/*
11649 	 * No dummy sf_hments, please.
11650 	 */
11651 	ASSERT(sfhme->hme_tte.ll != 0);
11652 
11653 	sfhme0 = sfhme - sfhme->hme_tte.tte_hmenum;
11654 	hmeblkp = (struct hme_blk *)((uintptr_t)sfhme0 -
11655 	    (uintptr_t)&hblk_dummy->hblk_hme[0]);
11656 
11657 	return (hmeblkp);
11658 }
11659 
11660 /*
11661  * On swapin, get appropriately sized TSB(s) and clear the HAT_SWAPPED flag.
11662  * If we can't get appropriately sized TSB(s), try for 8K TSB(s) using
11663  * KM_SLEEP allocation.
11664  *
11665  * Return 0 on success, -1 otherwise.
11666  */
11667 static void
11668 sfmmu_tsb_swapin(sfmmu_t *sfmmup, hatlock_t *hatlockp)
11669 {
11670 	struct tsb_info *tsbinfop, *next;
11671 	tsb_replace_rc_t rc;
11672 	boolean_t gotfirst = B_FALSE;
11673 
11674 	ASSERT(sfmmup != ksfmmup);
11675 	ASSERT(sfmmu_hat_lock_held(sfmmup));
11676 
11677 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPIN)) {
11678 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
11679 	}
11680 
11681 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
11682 		SFMMU_FLAGS_SET(sfmmup, HAT_SWAPIN);
11683 	} else {
11684 		return;
11685 	}
11686 
11687 	ASSERT(sfmmup->sfmmu_tsb != NULL);
11688 
11689 	/*
11690 	 * Loop over all tsbinfo's replacing them with ones that actually have
11691 	 * a TSB.  If any of the replacements ever fail, bail out of the loop.
11692 	 */
11693 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; tsbinfop = next) {
11694 		ASSERT(tsbinfop->tsb_flags & TSB_SWAPPED);
11695 		next = tsbinfop->tsb_next;
11696 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, tsbinfop->tsb_szc,
11697 		    hatlockp, TSB_SWAPIN);
11698 		if (rc != TSB_SUCCESS) {
11699 			break;
11700 		}
11701 		gotfirst = B_TRUE;
11702 	}
11703 
11704 	switch (rc) {
11705 	case TSB_SUCCESS:
11706 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11707 		cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11708 		return;
11709 	case TSB_LOSTRACE:
11710 		break;
11711 	case TSB_ALLOCFAIL:
11712 		break;
11713 	default:
11714 		panic("sfmmu_replace_tsb returned unrecognized failure code "
11715 		    "%d", rc);
11716 	}
11717 
11718 	/*
11719 	 * In this case, we failed to get one of our TSBs.  If we failed to
11720 	 * get the first TSB, get one of minimum size (8KB).  Walk the list
11721 	 * and throw away the tsbinfos, starting where the allocation failed;
11722 	 * we can get by with just one TSB as long as we don't leave the
11723 	 * SWAPPED tsbinfo structures lying around.
11724 	 */
11725 	tsbinfop = sfmmup->sfmmu_tsb;
11726 	next = tsbinfop->tsb_next;
11727 	tsbinfop->tsb_next = NULL;
11728 
11729 	sfmmu_hat_exit(hatlockp);
11730 	for (tsbinfop = next; tsbinfop != NULL; tsbinfop = next) {
11731 		next = tsbinfop->tsb_next;
11732 		sfmmu_tsbinfo_free(tsbinfop);
11733 	}
11734 	hatlockp = sfmmu_hat_enter(sfmmup);
11735 
11736 	/*
11737 	 * If we don't have any TSBs, get a single 8K TSB for 8K, 64K and 512K
11738 	 * pages.
11739 	 */
11740 	if (!gotfirst) {
11741 		tsbinfop = sfmmup->sfmmu_tsb;
11742 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, TSB_MIN_SZCODE,
11743 		    hatlockp, TSB_SWAPIN | TSB_FORCEALLOC);
11744 		ASSERT(rc == TSB_SUCCESS);
11745 	}
11746 
11747 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11748 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11749 }
11750 
11751 static int
11752 sfmmu_is_rgnva(sf_srd_t *srdp, caddr_t addr, ulong_t w, ulong_t bmw)
11753 {
11754 	ulong_t bix = 0;
11755 	uint_t rid;
11756 	sf_region_t *rgnp;
11757 
11758 	ASSERT(srdp != NULL);
11759 	ASSERT(srdp->srd_refcnt != 0);
11760 
11761 	w <<= BT_ULSHIFT;
11762 	while (bmw) {
11763 		if (!(bmw & 0x1)) {
11764 			bix++;
11765 			bmw >>= 1;
11766 			continue;
11767 		}
11768 		rid = w | bix;
11769 		rgnp = srdp->srd_hmergnp[rid];
11770 		ASSERT(rgnp->rgn_refcnt > 0);
11771 		ASSERT(rgnp->rgn_id == rid);
11772 		if (addr < rgnp->rgn_saddr ||
11773 		    addr >= (rgnp->rgn_saddr + rgnp->rgn_size)) {
11774 			bix++;
11775 			bmw >>= 1;
11776 		} else {
11777 			return (1);
11778 		}
11779 	}
11780 	return (0);
11781 }
11782 
11783 /*
11784  * Handle exceptions for low level tsb_handler.
11785  *
11786  * There are many scenarios that could land us here:
11787  *
11788  * If the context is invalid we land here. The context can be invalid
11789  * for 3 reasons: 1) we couldn't allocate a new context and now need to
11790  * perform a wrap around operation in order to allocate a new context.
11791  * 2) Context was invalidated to change pagesize programming 3) ISMs or
11792  * TSBs configuration is changeing for this process and we are forced into
11793  * here to do a syncronization operation. If the context is valid we can
11794  * be here from window trap hanlder. In this case just call trap to handle
11795  * the fault.
11796  *
11797  * Note that the process will run in INVALID_CONTEXT before
11798  * faulting into here and subsequently loading the MMU registers
11799  * (including the TSB base register) associated with this process.
11800  * For this reason, the trap handlers must all test for
11801  * INVALID_CONTEXT before attempting to access any registers other
11802  * than the context registers.
11803  */
11804 void
11805 sfmmu_tsbmiss_exception(struct regs *rp, uintptr_t tagaccess, uint_t traptype)
11806 {
11807 	sfmmu_t *sfmmup, *shsfmmup;
11808 	uint_t ctxtype;
11809 	klwp_id_t lwp;
11810 	char lwp_save_state;
11811 	hatlock_t *hatlockp, *shatlockp;
11812 	struct tsb_info *tsbinfop;
11813 	struct tsbmiss *tsbmp;
11814 	sf_scd_t *scdp;
11815 
11816 	SFMMU_STAT(sf_tsb_exceptions);
11817 	SFMMU_MMU_STAT(mmu_tsb_exceptions);
11818 	sfmmup = astosfmmu(curthread->t_procp->p_as);
11819 	/*
11820 	 * note that in sun4u, tagacces register contains ctxnum
11821 	 * while sun4v passes ctxtype in the tagaccess register.
11822 	 */
11823 	ctxtype = tagaccess & TAGACC_CTX_MASK;
11824 
11825 	ASSERT(sfmmup != ksfmmup && ctxtype != KCONTEXT);
11826 	ASSERT(sfmmup->sfmmu_ismhat == 0);
11827 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED) ||
11828 	    ctxtype == INVALID_CONTEXT);
11829 
11830 	if (ctxtype != INVALID_CONTEXT && traptype != T_DATA_PROT) {
11831 		/*
11832 		 * We may land here because shme bitmap and pagesize
11833 		 * flags are updated lazily in tsbmiss area on other cpus.
11834 		 * If we detect here that tsbmiss area is out of sync with
11835 		 * sfmmu update it and retry the trapped instruction.
11836 		 * Otherwise call trap().
11837 		 */
11838 		int ret = 0;
11839 		uchar_t tteflag_mask = (1 << TTE64K) | (1 << TTE8K);
11840 		caddr_t addr = (caddr_t)(tagaccess & TAGACC_VADDR_MASK);
11841 
11842 		/*
11843 		 * Must set lwp state to LWP_SYS before
11844 		 * trying to acquire any adaptive lock
11845 		 */
11846 		lwp = ttolwp(curthread);
11847 		ASSERT(lwp);
11848 		lwp_save_state = lwp->lwp_state;
11849 		lwp->lwp_state = LWP_SYS;
11850 
11851 		hatlockp = sfmmu_hat_enter(sfmmup);
11852 		kpreempt_disable();
11853 		tsbmp = &tsbmiss_area[CPU->cpu_id];
11854 		ASSERT(sfmmup == tsbmp->usfmmup);
11855 		if (((tsbmp->uhat_tteflags ^ sfmmup->sfmmu_tteflags) &
11856 		    ~tteflag_mask) ||
11857 		    ((tsbmp->uhat_rtteflags ^  sfmmup->sfmmu_rtteflags) &
11858 		    ~tteflag_mask)) {
11859 			tsbmp->uhat_tteflags = sfmmup->sfmmu_tteflags;
11860 			tsbmp->uhat_rtteflags = sfmmup->sfmmu_rtteflags;
11861 			ret = 1;
11862 		}
11863 		if (sfmmup->sfmmu_srdp != NULL) {
11864 			ulong_t *sm = sfmmup->sfmmu_hmeregion_map.bitmap;
11865 			ulong_t *tm = tsbmp->shmermap;
11866 			ulong_t i;
11867 			for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
11868 				ulong_t d = tm[i] ^ sm[i];
11869 				if (d) {
11870 					if (d & sm[i]) {
11871 						if (!ret && sfmmu_is_rgnva(
11872 						    sfmmup->sfmmu_srdp,
11873 						    addr, i, d & sm[i])) {
11874 							ret = 1;
11875 						}
11876 					}
11877 					tm[i] = sm[i];
11878 				}
11879 			}
11880 		}
11881 		kpreempt_enable();
11882 		sfmmu_hat_exit(hatlockp);
11883 		lwp->lwp_state = lwp_save_state;
11884 		if (ret) {
11885 			return;
11886 		}
11887 	} else if (ctxtype == INVALID_CONTEXT) {
11888 		/*
11889 		 * First, make sure we come out of here with a valid ctx,
11890 		 * since if we don't get one we'll simply loop on the
11891 		 * faulting instruction.
11892 		 *
11893 		 * If the ISM mappings are changing, the TSB is relocated,
11894 		 * the process is swapped, the process is joining SCD or
11895 		 * leaving SCD or shared regions we serialize behind the
11896 		 * controlling thread with hat lock, sfmmu_flags and
11897 		 * sfmmu_tsb_cv condition variable.
11898 		 */
11899 
11900 		/*
11901 		 * Must set lwp state to LWP_SYS before
11902 		 * trying to acquire any adaptive lock
11903 		 */
11904 		lwp = ttolwp(curthread);
11905 		ASSERT(lwp);
11906 		lwp_save_state = lwp->lwp_state;
11907 		lwp->lwp_state = LWP_SYS;
11908 
11909 		hatlockp = sfmmu_hat_enter(sfmmup);
11910 retry:
11911 		if ((scdp = sfmmup->sfmmu_scdp) != NULL) {
11912 			shsfmmup = scdp->scd_sfmmup;
11913 			ASSERT(shsfmmup != NULL);
11914 
11915 			for (tsbinfop = shsfmmup->sfmmu_tsb; tsbinfop != NULL;
11916 			    tsbinfop = tsbinfop->tsb_next) {
11917 				if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
11918 					/* drop the private hat lock */
11919 					sfmmu_hat_exit(hatlockp);
11920 					/* acquire the shared hat lock */
11921 					shatlockp = sfmmu_hat_enter(shsfmmup);
11922 					/*
11923 					 * recheck to see if anything changed
11924 					 * after we drop the private hat lock.
11925 					 */
11926 					if (sfmmup->sfmmu_scdp == scdp &&
11927 					    shsfmmup == scdp->scd_sfmmup) {
11928 						sfmmu_tsb_chk_reloc(shsfmmup,
11929 						    shatlockp);
11930 					}
11931 					sfmmu_hat_exit(shatlockp);
11932 					hatlockp = sfmmu_hat_enter(sfmmup);
11933 					goto retry;
11934 				}
11935 			}
11936 		}
11937 
11938 		for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
11939 		    tsbinfop = tsbinfop->tsb_next) {
11940 			if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
11941 				cv_wait(&sfmmup->sfmmu_tsb_cv,
11942 				    HATLOCK_MUTEXP(hatlockp));
11943 				goto retry;
11944 			}
11945 		}
11946 
11947 		/*
11948 		 * Wait for ISM maps to be updated.
11949 		 */
11950 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
11951 			cv_wait(&sfmmup->sfmmu_tsb_cv,
11952 			    HATLOCK_MUTEXP(hatlockp));
11953 			goto retry;
11954 		}
11955 
11956 		/* Is this process joining an SCD? */
11957 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
11958 			/*
11959 			 * Flush private TSB and setup shared TSB.
11960 			 * sfmmu_finish_join_scd() does not drop the
11961 			 * hat lock.
11962 			 */
11963 			sfmmu_finish_join_scd(sfmmup);
11964 			SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
11965 		}
11966 
11967 		/*
11968 		 * If we're swapping in, get TSB(s).  Note that we must do
11969 		 * this before we get a ctx or load the MMU state.  Once
11970 		 * we swap in we have to recheck to make sure the TSB(s) and
11971 		 * ISM mappings didn't change while we slept.
11972 		 */
11973 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
11974 			sfmmu_tsb_swapin(sfmmup, hatlockp);
11975 			goto retry;
11976 		}
11977 
11978 		sfmmu_get_ctx(sfmmup);
11979 
11980 		sfmmu_hat_exit(hatlockp);
11981 		/*
11982 		 * Must restore lwp_state if not calling
11983 		 * trap() for further processing. Restore
11984 		 * it anyway.
11985 		 */
11986 		lwp->lwp_state = lwp_save_state;
11987 		return;
11988 	}
11989 	trap(rp, (caddr_t)tagaccess, traptype, 0);
11990 }
11991 
11992 static void
11993 sfmmu_tsb_chk_reloc(sfmmu_t *sfmmup, hatlock_t *hatlockp)
11994 {
11995 	struct tsb_info *tp;
11996 
11997 	ASSERT(sfmmu_hat_lock_held(sfmmup));
11998 
11999 	for (tp = sfmmup->sfmmu_tsb; tp != NULL; tp = tp->tsb_next) {
12000 		if (tp->tsb_flags & TSB_RELOC_FLAG) {
12001 			cv_wait(&sfmmup->sfmmu_tsb_cv,
12002 			    HATLOCK_MUTEXP(hatlockp));
12003 			break;
12004 		}
12005 	}
12006 }
12007 
12008 /*
12009  * sfmmu_vatopfn_suspended is called from GET_TTE when TL=0 and
12010  * TTE_SUSPENDED bit set in tte we block on aquiring a page lock
12011  * rather than spinning to avoid send mondo timeouts with
12012  * interrupts enabled. When the lock is acquired it is immediately
12013  * released and we return back to sfmmu_vatopfn just after
12014  * the GET_TTE call.
12015  */
12016 void
12017 sfmmu_vatopfn_suspended(caddr_t vaddr, sfmmu_t *sfmmu, tte_t *ttep)
12018 {
12019 	struct page	**pp;
12020 
12021 	(void) as_pagelock(sfmmu->sfmmu_as, &pp, vaddr, TTE_CSZ(ttep), S_WRITE);
12022 	as_pageunlock(sfmmu->sfmmu_as, pp, vaddr, TTE_CSZ(ttep), S_WRITE);
12023 }
12024 
12025 /*
12026  * sfmmu_tsbmiss_suspended is called from GET_TTE when TL>0 and
12027  * TTE_SUSPENDED bit set in tte. We do this so that we can handle
12028  * cross traps which cannot be handled while spinning in the
12029  * trap handlers. Simply enter and exit the kpr_suspendlock spin
12030  * mutex, which is held by the holder of the suspend bit, and then
12031  * retry the trapped instruction after unwinding.
12032  */
12033 /*ARGSUSED*/
12034 void
12035 sfmmu_tsbmiss_suspended(struct regs *rp, uintptr_t tagacc, uint_t traptype)
12036 {
12037 	ASSERT(curthread != kreloc_thread);
12038 	mutex_enter(&kpr_suspendlock);
12039 	mutex_exit(&kpr_suspendlock);
12040 }
12041 
12042 /*
12043  * This routine could be optimized to reduce the number of xcalls by flushing
12044  * the entire TLBs if region reference count is above some threshold but the
12045  * tradeoff will depend on the size of the TLB. So for now flush the specific
12046  * page a context at a time.
12047  *
12048  * If uselocks is 0 then it's called after all cpus were captured and all the
12049  * hat locks were taken. In this case don't take the region lock by relying on
12050  * the order of list region update operations in hat_join_region(),
12051  * hat_leave_region() and hat_dup_region(). The ordering in those routines
12052  * guarantees that list is always forward walkable and reaches active sfmmus
12053  * regardless of where xc_attention() captures a cpu.
12054  */
12055 cpuset_t
12056 sfmmu_rgntlb_demap(caddr_t addr, sf_region_t *rgnp,
12057     struct hme_blk *hmeblkp, int uselocks)
12058 {
12059 	sfmmu_t	*sfmmup;
12060 	cpuset_t cpuset;
12061 	cpuset_t rcpuset;
12062 	hatlock_t *hatlockp;
12063 	uint_t rid = rgnp->rgn_id;
12064 	sf_rgn_link_t *rlink;
12065 	sf_scd_t *scdp;
12066 
12067 	ASSERT(hmeblkp->hblk_shared);
12068 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
12069 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
12070 
12071 	CPUSET_ZERO(rcpuset);
12072 	if (uselocks) {
12073 		mutex_enter(&rgnp->rgn_mutex);
12074 	}
12075 	sfmmup = rgnp->rgn_sfmmu_head;
12076 	while (sfmmup != NULL) {
12077 		if (uselocks) {
12078 			hatlockp = sfmmu_hat_enter(sfmmup);
12079 		}
12080 
12081 		/*
12082 		 * When an SCD is created the SCD hat is linked on the sfmmu
12083 		 * region lists for each hme region which is part of the
12084 		 * SCD. If we find an SCD hat, when walking these lists,
12085 		 * then we flush the shared TSBs, if we find a private hat,
12086 		 * which is part of an SCD, but where the region
12087 		 * is not part of the SCD then we flush the private TSBs.
12088 		 */
12089 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12090 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
12091 			scdp = sfmmup->sfmmu_scdp;
12092 			if (SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
12093 				if (uselocks) {
12094 					sfmmu_hat_exit(hatlockp);
12095 				}
12096 				goto next;
12097 			}
12098 		}
12099 
12100 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12101 
12102 		kpreempt_disable();
12103 		cpuset = sfmmup->sfmmu_cpusran;
12104 		CPUSET_AND(cpuset, cpu_ready_set);
12105 		CPUSET_DEL(cpuset, CPU->cpu_id);
12106 		SFMMU_XCALL_STATS(sfmmup);
12107 		xt_some(cpuset, vtag_flushpage_tl1,
12108 		    (uint64_t)addr, (uint64_t)sfmmup);
12109 		vtag_flushpage(addr, (uint64_t)sfmmup);
12110 		if (uselocks) {
12111 			sfmmu_hat_exit(hatlockp);
12112 		}
12113 		kpreempt_enable();
12114 		CPUSET_OR(rcpuset, cpuset);
12115 
12116 next:
12117 		/* LINTED: constant in conditional context */
12118 		SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
12119 		ASSERT(rlink != NULL);
12120 		sfmmup = rlink->next;
12121 	}
12122 	if (uselocks) {
12123 		mutex_exit(&rgnp->rgn_mutex);
12124 	}
12125 	return (rcpuset);
12126 }
12127 
12128 /*
12129  * This routine takes an sfmmu pointer and the va for an adddress in an
12130  * ISM region as input and returns the corresponding region id in ism_rid.
12131  * The return value of 1 indicates that a region has been found and ism_rid
12132  * is valid, otherwise 0 is returned.
12133  */
12134 static int
12135 find_ism_rid(sfmmu_t *sfmmup, sfmmu_t *ism_sfmmup, caddr_t va, uint_t *ism_rid)
12136 {
12137 	ism_blk_t	*ism_blkp;
12138 	int		i;
12139 	ism_map_t	*ism_map;
12140 #ifdef DEBUG
12141 	struct hat	*ism_hatid;
12142 #endif
12143 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12144 
12145 	ism_blkp = sfmmup->sfmmu_iblk;
12146 	while (ism_blkp != NULL) {
12147 		ism_map = ism_blkp->iblk_maps;
12148 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
12149 			if ((va >= ism_start(ism_map[i])) &&
12150 			    (va < ism_end(ism_map[i]))) {
12151 
12152 				*ism_rid = ism_map[i].imap_rid;
12153 #ifdef DEBUG
12154 				ism_hatid = ism_map[i].imap_ismhat;
12155 				ASSERT(ism_hatid == ism_sfmmup);
12156 				ASSERT(ism_hatid->sfmmu_ismhat);
12157 #endif
12158 				return (1);
12159 			}
12160 		}
12161 		ism_blkp = ism_blkp->iblk_next;
12162 	}
12163 	return (0);
12164 }
12165 
12166 /*
12167  * Special routine to flush out ism mappings- TSBs, TLBs and D-caches.
12168  * This routine may be called with all cpu's captured. Therefore, the
12169  * caller is responsible for holding all locks and disabling kernel
12170  * preemption.
12171  */
12172 /* ARGSUSED */
12173 static void
12174 sfmmu_ismtlbcache_demap(caddr_t addr, sfmmu_t *ism_sfmmup,
12175 	struct hme_blk *hmeblkp, pfn_t pfnum, int cache_flush_flag)
12176 {
12177 	cpuset_t 	cpuset;
12178 	caddr_t 	va;
12179 	ism_ment_t	*ment;
12180 	sfmmu_t		*sfmmup;
12181 #ifdef VAC
12182 	int 		vcolor;
12183 #endif
12184 
12185 	sf_scd_t	*scdp;
12186 	uint_t		ism_rid;
12187 
12188 	ASSERT(!hmeblkp->hblk_shared);
12189 	/*
12190 	 * Walk the ism_hat's mapping list and flush the page
12191 	 * from every hat sharing this ism_hat. This routine
12192 	 * may be called while all cpu's have been captured.
12193 	 * Therefore we can't attempt to grab any locks. For now
12194 	 * this means we will protect the ism mapping list under
12195 	 * a single lock which will be grabbed by the caller.
12196 	 * If hat_share/unshare scalibility becomes a performance
12197 	 * problem then we may need to re-think ism mapping list locking.
12198 	 */
12199 	ASSERT(ism_sfmmup->sfmmu_ismhat);
12200 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
12201 	addr = addr - ISMID_STARTADDR;
12202 
12203 	for (ment = ism_sfmmup->sfmmu_iment; ment; ment = ment->iment_next) {
12204 
12205 		sfmmup = ment->iment_hat;
12206 
12207 		va = ment->iment_base_va;
12208 		va = (caddr_t)((uintptr_t)va  + (uintptr_t)addr);
12209 
12210 		/*
12211 		 * When an SCD is created the SCD hat is linked on the ism
12212 		 * mapping lists for each ISM segment which is part of the
12213 		 * SCD. If we find an SCD hat, when walking these lists,
12214 		 * then we flush the shared TSBs, if we find a private hat,
12215 		 * which is part of an SCD, but where the region
12216 		 * corresponding to this va is not part of the SCD then we
12217 		 * flush the private TSBs.
12218 		 */
12219 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12220 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD) &&
12221 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
12222 			if (!find_ism_rid(sfmmup, ism_sfmmup, va,
12223 			    &ism_rid)) {
12224 				cmn_err(CE_PANIC,
12225 				    "can't find matching ISM rid!");
12226 			}
12227 
12228 			scdp = sfmmup->sfmmu_scdp;
12229 			if (SFMMU_IS_ISMRID_VALID(ism_rid) &&
12230 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map,
12231 			    ism_rid)) {
12232 				continue;
12233 			}
12234 		}
12235 		SFMMU_UNLOAD_TSB(va, sfmmup, hmeblkp, 1);
12236 
12237 		cpuset = sfmmup->sfmmu_cpusran;
12238 		CPUSET_AND(cpuset, cpu_ready_set);
12239 		CPUSET_DEL(cpuset, CPU->cpu_id);
12240 		SFMMU_XCALL_STATS(sfmmup);
12241 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)va,
12242 		    (uint64_t)sfmmup);
12243 		vtag_flushpage(va, (uint64_t)sfmmup);
12244 
12245 #ifdef VAC
12246 		/*
12247 		 * Flush D$
12248 		 * When flushing D$ we must flush all
12249 		 * cpu's. See sfmmu_cache_flush().
12250 		 */
12251 		if (cache_flush_flag == CACHE_FLUSH) {
12252 			cpuset = cpu_ready_set;
12253 			CPUSET_DEL(cpuset, CPU->cpu_id);
12254 
12255 			SFMMU_XCALL_STATS(sfmmup);
12256 			vcolor = addr_to_vcolor(va);
12257 			xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12258 			vac_flushpage(pfnum, vcolor);
12259 		}
12260 #endif	/* VAC */
12261 	}
12262 }
12263 
12264 /*
12265  * Demaps the TSB, CPU caches, and flushes all TLBs on all CPUs of
12266  * a particular virtual address and ctx.  If noflush is set we do not
12267  * flush the TLB/TSB.  This function may or may not be called with the
12268  * HAT lock held.
12269  */
12270 static void
12271 sfmmu_tlbcache_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12272 	pfn_t pfnum, int tlb_noflush, int cpu_flag, int cache_flush_flag,
12273 	int hat_lock_held)
12274 {
12275 #ifdef VAC
12276 	int vcolor;
12277 #endif
12278 	cpuset_t cpuset;
12279 	hatlock_t *hatlockp;
12280 
12281 	ASSERT(!hmeblkp->hblk_shared);
12282 
12283 #if defined(lint) && !defined(VAC)
12284 	pfnum = pfnum;
12285 	cpu_flag = cpu_flag;
12286 	cache_flush_flag = cache_flush_flag;
12287 #endif
12288 
12289 	/*
12290 	 * There is no longer a need to protect against ctx being
12291 	 * stolen here since we don't store the ctx in the TSB anymore.
12292 	 */
12293 #ifdef VAC
12294 	vcolor = addr_to_vcolor(addr);
12295 #endif
12296 
12297 	/*
12298 	 * We must hold the hat lock during the flush of TLB,
12299 	 * to avoid a race with sfmmu_invalidate_ctx(), where
12300 	 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12301 	 * causing TLB demap routine to skip flush on that MMU.
12302 	 * If the context on a MMU has already been set to
12303 	 * INVALID_CONTEXT, we just get an extra flush on
12304 	 * that MMU.
12305 	 */
12306 	if (!hat_lock_held && !tlb_noflush)
12307 		hatlockp = sfmmu_hat_enter(sfmmup);
12308 
12309 	kpreempt_disable();
12310 	if (!tlb_noflush) {
12311 		/*
12312 		 * Flush the TSB and TLB.
12313 		 */
12314 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12315 
12316 		cpuset = sfmmup->sfmmu_cpusran;
12317 		CPUSET_AND(cpuset, cpu_ready_set);
12318 		CPUSET_DEL(cpuset, CPU->cpu_id);
12319 
12320 		SFMMU_XCALL_STATS(sfmmup);
12321 
12322 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
12323 		    (uint64_t)sfmmup);
12324 
12325 		vtag_flushpage(addr, (uint64_t)sfmmup);
12326 	}
12327 
12328 	if (!hat_lock_held && !tlb_noflush)
12329 		sfmmu_hat_exit(hatlockp);
12330 
12331 #ifdef VAC
12332 	/*
12333 	 * Flush the D$
12334 	 *
12335 	 * Even if the ctx is stolen, we need to flush the
12336 	 * cache. Our ctx stealer only flushes the TLBs.
12337 	 */
12338 	if (cache_flush_flag == CACHE_FLUSH) {
12339 		if (cpu_flag & FLUSH_ALL_CPUS) {
12340 			cpuset = cpu_ready_set;
12341 		} else {
12342 			cpuset = sfmmup->sfmmu_cpusran;
12343 			CPUSET_AND(cpuset, cpu_ready_set);
12344 		}
12345 		CPUSET_DEL(cpuset, CPU->cpu_id);
12346 		SFMMU_XCALL_STATS(sfmmup);
12347 		xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12348 		vac_flushpage(pfnum, vcolor);
12349 	}
12350 #endif	/* VAC */
12351 	kpreempt_enable();
12352 }
12353 
12354 /*
12355  * Demaps the TSB and flushes all TLBs on all cpus for a particular virtual
12356  * address and ctx.  If noflush is set we do not currently do anything.
12357  * This function may or may not be called with the HAT lock held.
12358  */
12359 static void
12360 sfmmu_tlb_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12361 	int tlb_noflush, int hat_lock_held)
12362 {
12363 	cpuset_t cpuset;
12364 	hatlock_t *hatlockp;
12365 
12366 	ASSERT(!hmeblkp->hblk_shared);
12367 
12368 	/*
12369 	 * If the process is exiting we have nothing to do.
12370 	 */
12371 	if (tlb_noflush)
12372 		return;
12373 
12374 	/*
12375 	 * Flush TSB.
12376 	 */
12377 	if (!hat_lock_held)
12378 		hatlockp = sfmmu_hat_enter(sfmmup);
12379 	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12380 
12381 	kpreempt_disable();
12382 
12383 	cpuset = sfmmup->sfmmu_cpusran;
12384 	CPUSET_AND(cpuset, cpu_ready_set);
12385 	CPUSET_DEL(cpuset, CPU->cpu_id);
12386 
12387 	SFMMU_XCALL_STATS(sfmmup);
12388 	xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, (uint64_t)sfmmup);
12389 
12390 	vtag_flushpage(addr, (uint64_t)sfmmup);
12391 
12392 	if (!hat_lock_held)
12393 		sfmmu_hat_exit(hatlockp);
12394 
12395 	kpreempt_enable();
12396 
12397 }
12398 
12399 /*
12400  * Special case of sfmmu_tlb_demap for MMU_PAGESIZE hblks. Use the xcall
12401  * call handler that can flush a range of pages to save on xcalls.
12402  */
12403 static int sfmmu_xcall_save;
12404 
12405 /*
12406  * this routine is never used for demaping addresses backed by SRD hmeblks.
12407  */
12408 static void
12409 sfmmu_tlb_range_demap(demap_range_t *dmrp)
12410 {
12411 	sfmmu_t *sfmmup = dmrp->dmr_sfmmup;
12412 	hatlock_t *hatlockp;
12413 	cpuset_t cpuset;
12414 	uint64_t sfmmu_pgcnt;
12415 	pgcnt_t pgcnt = 0;
12416 	int pgunload = 0;
12417 	int dirtypg = 0;
12418 	caddr_t addr = dmrp->dmr_addr;
12419 	caddr_t eaddr;
12420 	uint64_t bitvec = dmrp->dmr_bitvec;
12421 
12422 	ASSERT(bitvec & 1);
12423 
12424 	/*
12425 	 * Flush TSB and calculate number of pages to flush.
12426 	 */
12427 	while (bitvec != 0) {
12428 		dirtypg = 0;
12429 		/*
12430 		 * Find the first page to flush and then count how many
12431 		 * pages there are after it that also need to be flushed.
12432 		 * This way the number of TSB flushes is minimized.
12433 		 */
12434 		while ((bitvec & 1) == 0) {
12435 			pgcnt++;
12436 			addr += MMU_PAGESIZE;
12437 			bitvec >>= 1;
12438 		}
12439 		while (bitvec & 1) {
12440 			dirtypg++;
12441 			bitvec >>= 1;
12442 		}
12443 		eaddr = addr + ptob(dirtypg);
12444 		hatlockp = sfmmu_hat_enter(sfmmup);
12445 		sfmmu_unload_tsb_range(sfmmup, addr, eaddr, TTE8K);
12446 		sfmmu_hat_exit(hatlockp);
12447 		pgunload += dirtypg;
12448 		addr = eaddr;
12449 		pgcnt += dirtypg;
12450 	}
12451 
12452 	ASSERT((pgcnt<<MMU_PAGESHIFT) <= dmrp->dmr_endaddr - dmrp->dmr_addr);
12453 	if (sfmmup->sfmmu_free == 0) {
12454 		addr = dmrp->dmr_addr;
12455 		bitvec = dmrp->dmr_bitvec;
12456 
12457 		/*
12458 		 * make sure it has SFMMU_PGCNT_SHIFT bits only,
12459 		 * as it will be used to pack argument for xt_some
12460 		 */
12461 		ASSERT((pgcnt > 0) &&
12462 		    (pgcnt <= (1 << SFMMU_PGCNT_SHIFT)));
12463 
12464 		/*
12465 		 * Encode pgcnt as (pgcnt -1 ), and pass (pgcnt - 1) in
12466 		 * the low 6 bits of sfmmup. This is doable since pgcnt
12467 		 * always >= 1.
12468 		 */
12469 		ASSERT(!((uint64_t)sfmmup & SFMMU_PGCNT_MASK));
12470 		sfmmu_pgcnt = (uint64_t)sfmmup |
12471 		    ((pgcnt - 1) & SFMMU_PGCNT_MASK);
12472 
12473 		/*
12474 		 * We must hold the hat lock during the flush of TLB,
12475 		 * to avoid a race with sfmmu_invalidate_ctx(), where
12476 		 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12477 		 * causing TLB demap routine to skip flush on that MMU.
12478 		 * If the context on a MMU has already been set to
12479 		 * INVALID_CONTEXT, we just get an extra flush on
12480 		 * that MMU.
12481 		 */
12482 		hatlockp = sfmmu_hat_enter(sfmmup);
12483 		kpreempt_disable();
12484 
12485 		cpuset = sfmmup->sfmmu_cpusran;
12486 		CPUSET_AND(cpuset, cpu_ready_set);
12487 		CPUSET_DEL(cpuset, CPU->cpu_id);
12488 
12489 		SFMMU_XCALL_STATS(sfmmup);
12490 		xt_some(cpuset, vtag_flush_pgcnt_tl1, (uint64_t)addr,
12491 		    sfmmu_pgcnt);
12492 
12493 		for (; bitvec != 0; bitvec >>= 1) {
12494 			if (bitvec & 1)
12495 				vtag_flushpage(addr, (uint64_t)sfmmup);
12496 			addr += MMU_PAGESIZE;
12497 		}
12498 		kpreempt_enable();
12499 		sfmmu_hat_exit(hatlockp);
12500 
12501 		sfmmu_xcall_save += (pgunload-1);
12502 	}
12503 	dmrp->dmr_bitvec = 0;
12504 }
12505 
12506 /*
12507  * In cases where we need to synchronize with TLB/TSB miss trap
12508  * handlers, _and_ need to flush the TLB, it's a lot easier to
12509  * throw away the context from the process than to do a
12510  * special song and dance to keep things consistent for the
12511  * handlers.
12512  *
12513  * Since the process suddenly ends up without a context and our caller
12514  * holds the hat lock, threads that fault after this function is called
12515  * will pile up on the lock.  We can then do whatever we need to
12516  * atomically from the context of the caller.  The first blocked thread
12517  * to resume executing will get the process a new context, and the
12518  * process will resume executing.
12519  *
12520  * One added advantage of this approach is that on MMUs that
12521  * support a "flush all" operation, we will delay the flush until
12522  * cnum wrap-around, and then flush the TLB one time.  This
12523  * is rather rare, so it's a lot less expensive than making 8000
12524  * x-calls to flush the TLB 8000 times.
12525  *
12526  * A per-process (PP) lock is used to synchronize ctx allocations in
12527  * resume() and ctx invalidations here.
12528  */
12529 static void
12530 sfmmu_invalidate_ctx(sfmmu_t *sfmmup)
12531 {
12532 	cpuset_t cpuset;
12533 	int cnum, currcnum;
12534 	mmu_ctx_t *mmu_ctxp;
12535 	int i;
12536 	uint_t pstate_save;
12537 
12538 	SFMMU_STAT(sf_ctx_inv);
12539 
12540 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12541 	ASSERT(sfmmup != ksfmmup);
12542 
12543 	kpreempt_disable();
12544 
12545 	mmu_ctxp = CPU_MMU_CTXP(CPU);
12546 	ASSERT(mmu_ctxp);
12547 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
12548 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
12549 
12550 	currcnum = sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum;
12551 
12552 	pstate_save = sfmmu_disable_intrs();
12553 
12554 	lock_set(&sfmmup->sfmmu_ctx_lock);	/* acquire PP lock */
12555 	/* set HAT cnum invalid across all context domains. */
12556 	for (i = 0; i < max_mmu_ctxdoms; i++) {
12557 
12558 		cnum = 	sfmmup->sfmmu_ctxs[i].cnum;
12559 		if (cnum == INVALID_CONTEXT) {
12560 			continue;
12561 		}
12562 
12563 		sfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
12564 	}
12565 	membar_enter();	/* make sure globally visible to all CPUs */
12566 	lock_clear(&sfmmup->sfmmu_ctx_lock);	/* release PP lock */
12567 
12568 	sfmmu_enable_intrs(pstate_save);
12569 
12570 	cpuset = sfmmup->sfmmu_cpusran;
12571 	CPUSET_DEL(cpuset, CPU->cpu_id);
12572 	CPUSET_AND(cpuset, cpu_ready_set);
12573 	if (!CPUSET_ISNULL(cpuset)) {
12574 		SFMMU_XCALL_STATS(sfmmup);
12575 		xt_some(cpuset, sfmmu_raise_tsb_exception,
12576 		    (uint64_t)sfmmup, INVALID_CONTEXT);
12577 		xt_sync(cpuset);
12578 		SFMMU_STAT(sf_tsb_raise_exception);
12579 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
12580 	}
12581 
12582 	/*
12583 	 * If the hat to-be-invalidated is the same as the current
12584 	 * process on local CPU we need to invalidate
12585 	 * this CPU context as well.
12586 	 */
12587 	if ((sfmmu_getctx_sec() == currcnum) &&
12588 	    (currcnum != INVALID_CONTEXT)) {
12589 		/* sets shared context to INVALID too */
12590 		sfmmu_setctx_sec(INVALID_CONTEXT);
12591 		sfmmu_clear_utsbinfo();
12592 	}
12593 
12594 	SFMMU_FLAGS_SET(sfmmup, HAT_ALLCTX_INVALID);
12595 
12596 	kpreempt_enable();
12597 
12598 	/*
12599 	 * we hold the hat lock, so nobody should allocate a context
12600 	 * for us yet
12601 	 */
12602 	ASSERT(sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum == INVALID_CONTEXT);
12603 }
12604 
12605 #ifdef VAC
12606 /*
12607  * We need to flush the cache in all cpus.  It is possible that
12608  * a process referenced a page as cacheable but has sinced exited
12609  * and cleared the mapping list.  We still to flush it but have no
12610  * state so all cpus is the only alternative.
12611  */
12612 void
12613 sfmmu_cache_flush(pfn_t pfnum, int vcolor)
12614 {
12615 	cpuset_t cpuset;
12616 
12617 	kpreempt_disable();
12618 	cpuset = cpu_ready_set;
12619 	CPUSET_DEL(cpuset, CPU->cpu_id);
12620 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12621 	xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12622 	xt_sync(cpuset);
12623 	vac_flushpage(pfnum, vcolor);
12624 	kpreempt_enable();
12625 }
12626 
12627 void
12628 sfmmu_cache_flushcolor(int vcolor, pfn_t pfnum)
12629 {
12630 	cpuset_t cpuset;
12631 
12632 	ASSERT(vcolor >= 0);
12633 
12634 	kpreempt_disable();
12635 	cpuset = cpu_ready_set;
12636 	CPUSET_DEL(cpuset, CPU->cpu_id);
12637 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12638 	xt_some(cpuset, vac_flushcolor_tl1, vcolor, pfnum);
12639 	xt_sync(cpuset);
12640 	vac_flushcolor(vcolor, pfnum);
12641 	kpreempt_enable();
12642 }
12643 #endif	/* VAC */
12644 
12645 /*
12646  * We need to prevent processes from accessing the TSB using a cached physical
12647  * address.  It's alright if they try to access the TSB via virtual address
12648  * since they will just fault on that virtual address once the mapping has
12649  * been suspended.
12650  */
12651 #pragma weak sendmondo_in_recover
12652 
12653 /* ARGSUSED */
12654 static int
12655 sfmmu_tsb_pre_relocator(caddr_t va, uint_t tsbsz, uint_t flags, void *tsbinfo)
12656 {
12657 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12658 	sfmmu_t *sfmmup = tsbinfop->tsb_sfmmu;
12659 	hatlock_t *hatlockp;
12660 	sf_scd_t *scdp;
12661 
12662 	if (flags != HAT_PRESUSPEND)
12663 		return (0);
12664 
12665 	/*
12666 	 * If tsb is a shared TSB with TSB_SHAREDCTX set, sfmmup must
12667 	 * be a shared hat, then set SCD's tsbinfo's flag.
12668 	 * If tsb is not shared, sfmmup is a private hat, then set
12669 	 * its private tsbinfo's flag.
12670 	 */
12671 	hatlockp = sfmmu_hat_enter(sfmmup);
12672 	tsbinfop->tsb_flags |= TSB_RELOC_FLAG;
12673 
12674 	if (!(tsbinfop->tsb_flags & TSB_SHAREDCTX)) {
12675 		sfmmu_tsb_inv_ctx(sfmmup);
12676 		sfmmu_hat_exit(hatlockp);
12677 	} else {
12678 		/* release lock on the shared hat */
12679 		sfmmu_hat_exit(hatlockp);
12680 		/* sfmmup is a shared hat */
12681 		ASSERT(sfmmup->sfmmu_scdhat);
12682 		scdp = sfmmup->sfmmu_scdp;
12683 		ASSERT(scdp != NULL);
12684 		/* get private hat from the scd list */
12685 		mutex_enter(&scdp->scd_mutex);
12686 		sfmmup = scdp->scd_sf_list;
12687 		while (sfmmup != NULL) {
12688 			hatlockp = sfmmu_hat_enter(sfmmup);
12689 			/*
12690 			 * We do not call sfmmu_tsb_inv_ctx here because
12691 			 * sendmondo_in_recover check is only needed for
12692 			 * sun4u.
12693 			 */
12694 			sfmmu_invalidate_ctx(sfmmup);
12695 			sfmmu_hat_exit(hatlockp);
12696 			sfmmup = sfmmup->sfmmu_scd_link.next;
12697 
12698 		}
12699 		mutex_exit(&scdp->scd_mutex);
12700 	}
12701 	return (0);
12702 }
12703 
12704 static void
12705 sfmmu_tsb_inv_ctx(sfmmu_t *sfmmup)
12706 {
12707 	extern uint32_t sendmondo_in_recover;
12708 
12709 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12710 
12711 	/*
12712 	 * For Cheetah+ Erratum 25:
12713 	 * Wait for any active recovery to finish.  We can't risk
12714 	 * relocating the TSB of the thread running mondo_recover_proc()
12715 	 * since, if we did that, we would deadlock.  The scenario we are
12716 	 * trying to avoid is as follows:
12717 	 *
12718 	 * THIS CPU			RECOVER CPU
12719 	 * --------			-----------
12720 	 *				Begins recovery, walking through TSB
12721 	 * hat_pagesuspend() TSB TTE
12722 	 *				TLB miss on TSB TTE, spins at TL1
12723 	 * xt_sync()
12724 	 *	send_mondo_timeout()
12725 	 *	mondo_recover_proc()
12726 	 *	((deadlocked))
12727 	 *
12728 	 * The second half of the workaround is that mondo_recover_proc()
12729 	 * checks to see if the tsb_info has the RELOC flag set, and if it
12730 	 * does, it skips over that TSB without ever touching tsbinfop->tsb_va
12731 	 * and hence avoiding the TLB miss that could result in a deadlock.
12732 	 */
12733 	if (&sendmondo_in_recover) {
12734 		membar_enter();	/* make sure RELOC flag visible */
12735 		while (sendmondo_in_recover) {
12736 			drv_usecwait(1);
12737 			membar_consumer();
12738 		}
12739 	}
12740 
12741 	sfmmu_invalidate_ctx(sfmmup);
12742 }
12743 
12744 /* ARGSUSED */
12745 static int
12746 sfmmu_tsb_post_relocator(caddr_t va, uint_t tsbsz, uint_t flags,
12747 	void *tsbinfo, pfn_t newpfn)
12748 {
12749 	hatlock_t *hatlockp;
12750 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12751 	sfmmu_t	*sfmmup = tsbinfop->tsb_sfmmu;
12752 
12753 	if (flags != HAT_POSTUNSUSPEND)
12754 		return (0);
12755 
12756 	hatlockp = sfmmu_hat_enter(sfmmup);
12757 
12758 	SFMMU_STAT(sf_tsb_reloc);
12759 
12760 	/*
12761 	 * The process may have swapped out while we were relocating one
12762 	 * of its TSBs.  If so, don't bother doing the setup since the
12763 	 * process can't be using the memory anymore.
12764 	 */
12765 	if ((tsbinfop->tsb_flags & TSB_SWAPPED) == 0) {
12766 		ASSERT(va == tsbinfop->tsb_va);
12767 		sfmmu_tsbinfo_setup_phys(tsbinfop, newpfn);
12768 
12769 		if (tsbinfop->tsb_flags & TSB_FLUSH_NEEDED) {
12770 			sfmmu_inv_tsb(tsbinfop->tsb_va,
12771 			    TSB_BYTES(tsbinfop->tsb_szc));
12772 			tsbinfop->tsb_flags &= ~TSB_FLUSH_NEEDED;
12773 		}
12774 	}
12775 
12776 	membar_exit();
12777 	tsbinfop->tsb_flags &= ~TSB_RELOC_FLAG;
12778 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
12779 
12780 	sfmmu_hat_exit(hatlockp);
12781 
12782 	return (0);
12783 }
12784 
12785 /*
12786  * Allocate and initialize a tsb_info structure.  Note that we may or may not
12787  * allocate a TSB here, depending on the flags passed in.
12788  */
12789 static int
12790 sfmmu_tsbinfo_alloc(struct tsb_info **tsbinfopp, int tsb_szc, int tte_sz_mask,
12791 	uint_t flags, sfmmu_t *sfmmup)
12792 {
12793 	int err;
12794 
12795 	*tsbinfopp = (struct tsb_info *)kmem_cache_alloc(
12796 	    sfmmu_tsbinfo_cache, KM_SLEEP);
12797 
12798 	if ((err = sfmmu_init_tsbinfo(*tsbinfopp, tte_sz_mask,
12799 	    tsb_szc, flags, sfmmup)) != 0) {
12800 		kmem_cache_free(sfmmu_tsbinfo_cache, *tsbinfopp);
12801 		SFMMU_STAT(sf_tsb_allocfail);
12802 		*tsbinfopp = NULL;
12803 		return (err);
12804 	}
12805 	SFMMU_STAT(sf_tsb_alloc);
12806 
12807 	/*
12808 	 * Bump the TSB size counters for this TSB size.
12809 	 */
12810 	(*(((int *)&sfmmu_tsbsize_stat) + tsb_szc))++;
12811 	return (0);
12812 }
12813 
12814 static void
12815 sfmmu_tsb_free(struct tsb_info *tsbinfo)
12816 {
12817 	caddr_t tsbva = tsbinfo->tsb_va;
12818 	uint_t tsb_size = TSB_BYTES(tsbinfo->tsb_szc);
12819 	struct kmem_cache *kmem_cachep = tsbinfo->tsb_cache;
12820 	vmem_t	*vmp = tsbinfo->tsb_vmp;
12821 
12822 	/*
12823 	 * If we allocated this TSB from relocatable kernel memory, then we
12824 	 * need to uninstall the callback handler.
12825 	 */
12826 	if (tsbinfo->tsb_cache != sfmmu_tsb8k_cache) {
12827 		uintptr_t slab_mask;
12828 		caddr_t slab_vaddr;
12829 		page_t **ppl;
12830 		int ret;
12831 
12832 		ASSERT(tsb_size <= MMU_PAGESIZE4M || use_bigtsb_arena);
12833 		if (tsb_size > MMU_PAGESIZE4M)
12834 			slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
12835 		else
12836 			slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
12837 		slab_vaddr = (caddr_t)((uintptr_t)tsbva & slab_mask);
12838 
12839 		ret = as_pagelock(&kas, &ppl, slab_vaddr, PAGESIZE, S_WRITE);
12840 		ASSERT(ret == 0);
12841 		hat_delete_callback(tsbva, (uint_t)tsb_size, (void *)tsbinfo,
12842 		    0, NULL);
12843 		as_pageunlock(&kas, ppl, slab_vaddr, PAGESIZE, S_WRITE);
12844 	}
12845 
12846 	if (kmem_cachep != NULL) {
12847 		kmem_cache_free(kmem_cachep, tsbva);
12848 	} else {
12849 		vmem_xfree(vmp, (void *)tsbva, tsb_size);
12850 	}
12851 	tsbinfo->tsb_va = (caddr_t)0xbad00bad;
12852 	atomic_add_64(&tsb_alloc_bytes, -(int64_t)tsb_size);
12853 }
12854 
12855 static void
12856 sfmmu_tsbinfo_free(struct tsb_info *tsbinfo)
12857 {
12858 	if ((tsbinfo->tsb_flags & TSB_SWAPPED) == 0) {
12859 		sfmmu_tsb_free(tsbinfo);
12860 	}
12861 	kmem_cache_free(sfmmu_tsbinfo_cache, tsbinfo);
12862 
12863 }
12864 
12865 /*
12866  * Setup all the references to physical memory for this tsbinfo.
12867  * The underlying page(s) must be locked.
12868  */
12869 static void
12870 sfmmu_tsbinfo_setup_phys(struct tsb_info *tsbinfo, pfn_t pfn)
12871 {
12872 	ASSERT(pfn != PFN_INVALID);
12873 	ASSERT(pfn == va_to_pfn(tsbinfo->tsb_va));
12874 
12875 #ifndef sun4v
12876 	if (tsbinfo->tsb_szc == 0) {
12877 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn,
12878 		    PROT_WRITE|PROT_READ, TTE8K);
12879 	} else {
12880 		/*
12881 		 * Round down PA and use a large mapping; the handlers will
12882 		 * compute the TSB pointer at the correct offset into the
12883 		 * big virtual page.  NOTE: this assumes all TSBs larger
12884 		 * than 8K must come from physically contiguous slabs of
12885 		 * size tsb_slab_size.
12886 		 */
12887 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn & ~tsb_slab_mask,
12888 		    PROT_WRITE|PROT_READ, tsb_slab_ttesz);
12889 	}
12890 	tsbinfo->tsb_pa = ptob(pfn);
12891 
12892 	TTE_SET_LOCKED(&tsbinfo->tsb_tte); /* lock the tte into dtlb */
12893 	TTE_SET_MOD(&tsbinfo->tsb_tte);    /* enable writes */
12894 
12895 	ASSERT(TTE_IS_PRIVILEGED(&tsbinfo->tsb_tte));
12896 	ASSERT(TTE_IS_LOCKED(&tsbinfo->tsb_tte));
12897 #else /* sun4v */
12898 	tsbinfo->tsb_pa = ptob(pfn);
12899 #endif /* sun4v */
12900 }
12901 
12902 
12903 /*
12904  * Returns zero on success, ENOMEM if over the high water mark,
12905  * or EAGAIN if the caller needs to retry with a smaller TSB
12906  * size (or specify TSB_FORCEALLOC if the allocation can't fail).
12907  *
12908  * This call cannot fail to allocate a TSB if TSB_FORCEALLOC
12909  * is specified and the TSB requested is PAGESIZE, though it
12910  * may sleep waiting for memory if sufficient memory is not
12911  * available.
12912  */
12913 static int
12914 sfmmu_init_tsbinfo(struct tsb_info *tsbinfo, int tteszmask,
12915     int tsbcode, uint_t flags, sfmmu_t *sfmmup)
12916 {
12917 	caddr_t vaddr = NULL;
12918 	caddr_t slab_vaddr;
12919 	uintptr_t slab_mask;
12920 	int tsbbytes = TSB_BYTES(tsbcode);
12921 	int lowmem = 0;
12922 	struct kmem_cache *kmem_cachep = NULL;
12923 	vmem_t *vmp = NULL;
12924 	lgrp_id_t lgrpid = LGRP_NONE;
12925 	pfn_t pfn;
12926 	uint_t cbflags = HAC_SLEEP;
12927 	page_t **pplist;
12928 	int ret;
12929 
12930 	ASSERT(tsbbytes <= MMU_PAGESIZE4M || use_bigtsb_arena);
12931 	if (tsbbytes > MMU_PAGESIZE4M)
12932 		slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
12933 	else
12934 		slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
12935 
12936 	if (flags & (TSB_FORCEALLOC | TSB_SWAPIN | TSB_GROW | TSB_SHRINK))
12937 		flags |= TSB_ALLOC;
12938 
12939 	ASSERT((flags & TSB_FORCEALLOC) == 0 || tsbcode == TSB_MIN_SZCODE);
12940 
12941 	tsbinfo->tsb_sfmmu = sfmmup;
12942 
12943 	/*
12944 	 * If not allocating a TSB, set up the tsbinfo, set TSB_SWAPPED, and
12945 	 * return.
12946 	 */
12947 	if ((flags & TSB_ALLOC) == 0) {
12948 		tsbinfo->tsb_szc = tsbcode;
12949 		tsbinfo->tsb_ttesz_mask = tteszmask;
12950 		tsbinfo->tsb_va = (caddr_t)0xbadbadbeef;
12951 		tsbinfo->tsb_pa = -1;
12952 		tsbinfo->tsb_tte.ll = 0;
12953 		tsbinfo->tsb_next = NULL;
12954 		tsbinfo->tsb_flags = TSB_SWAPPED;
12955 		tsbinfo->tsb_cache = NULL;
12956 		tsbinfo->tsb_vmp = NULL;
12957 		return (0);
12958 	}
12959 
12960 #ifdef DEBUG
12961 	/*
12962 	 * For debugging:
12963 	 * Randomly force allocation failures every tsb_alloc_mtbf
12964 	 * tries if TSB_FORCEALLOC is not specified.  This will
12965 	 * return ENOMEM if tsb_alloc_mtbf is odd, or EAGAIN if
12966 	 * it is even, to allow testing of both failure paths...
12967 	 */
12968 	if (tsb_alloc_mtbf && ((flags & TSB_FORCEALLOC) == 0) &&
12969 	    (tsb_alloc_count++ == tsb_alloc_mtbf)) {
12970 		tsb_alloc_count = 0;
12971 		tsb_alloc_fail_mtbf++;
12972 		return ((tsb_alloc_mtbf & 1)? ENOMEM : EAGAIN);
12973 	}
12974 #endif	/* DEBUG */
12975 
12976 	/*
12977 	 * Enforce high water mark if we are not doing a forced allocation
12978 	 * and are not shrinking a process' TSB.
12979 	 */
12980 	if ((flags & TSB_SHRINK) == 0 &&
12981 	    (tsbbytes + tsb_alloc_bytes) > tsb_alloc_hiwater) {
12982 		if ((flags & TSB_FORCEALLOC) == 0)
12983 			return (ENOMEM);
12984 		lowmem = 1;
12985 	}
12986 
12987 	/*
12988 	 * Allocate from the correct location based upon the size of the TSB
12989 	 * compared to the base page size, and what memory conditions dictate.
12990 	 * Note we always do nonblocking allocations from the TSB arena since
12991 	 * we don't want memory fragmentation to cause processes to block
12992 	 * indefinitely waiting for memory; until the kernel algorithms that
12993 	 * coalesce large pages are improved this is our best option.
12994 	 *
12995 	 * Algorithm:
12996 	 *	If allocating a "large" TSB (>8K), allocate from the
12997 	 *		appropriate kmem_tsb_default_arena vmem arena
12998 	 *	else if low on memory or the TSB_FORCEALLOC flag is set or
12999 	 *	tsb_forceheap is set
13000 	 *		Allocate from kernel heap via sfmmu_tsb8k_cache with
13001 	 *		KM_SLEEP (never fails)
13002 	 *	else
13003 	 *		Allocate from appropriate sfmmu_tsb_cache with
13004 	 *		KM_NOSLEEP
13005 	 *	endif
13006 	 */
13007 	if (tsb_lgrp_affinity)
13008 		lgrpid = lgrp_home_id(curthread);
13009 	if (lgrpid == LGRP_NONE)
13010 		lgrpid = 0;	/* use lgrp of boot CPU */
13011 
13012 	if (tsbbytes > MMU_PAGESIZE) {
13013 		if (tsbbytes > MMU_PAGESIZE4M) {
13014 			vmp = kmem_bigtsb_default_arena[lgrpid];
13015 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
13016 			    0, 0, NULL, NULL, VM_NOSLEEP);
13017 		} else {
13018 			vmp = kmem_tsb_default_arena[lgrpid];
13019 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
13020 			    0, 0, NULL, NULL, VM_NOSLEEP);
13021 		}
13022 #ifdef	DEBUG
13023 	} else if (lowmem || (flags & TSB_FORCEALLOC) || tsb_forceheap) {
13024 #else	/* !DEBUG */
13025 	} else if (lowmem || (flags & TSB_FORCEALLOC)) {
13026 #endif	/* DEBUG */
13027 		kmem_cachep = sfmmu_tsb8k_cache;
13028 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_SLEEP);
13029 		ASSERT(vaddr != NULL);
13030 	} else {
13031 		kmem_cachep = sfmmu_tsb_cache[lgrpid];
13032 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_NOSLEEP);
13033 	}
13034 
13035 	tsbinfo->tsb_cache = kmem_cachep;
13036 	tsbinfo->tsb_vmp = vmp;
13037 
13038 	if (vaddr == NULL) {
13039 		return (EAGAIN);
13040 	}
13041 
13042 	atomic_add_64(&tsb_alloc_bytes, (int64_t)tsbbytes);
13043 	kmem_cachep = tsbinfo->tsb_cache;
13044 
13045 	/*
13046 	 * If we are allocating from outside the cage, then we need to
13047 	 * register a relocation callback handler.  Note that for now
13048 	 * since pseudo mappings always hang off of the slab's root page,
13049 	 * we need only lock the first 8K of the TSB slab.  This is a bit
13050 	 * hacky but it is good for performance.
13051 	 */
13052 	if (kmem_cachep != sfmmu_tsb8k_cache) {
13053 		slab_vaddr = (caddr_t)((uintptr_t)vaddr & slab_mask);
13054 		ret = as_pagelock(&kas, &pplist, slab_vaddr, PAGESIZE, S_WRITE);
13055 		ASSERT(ret == 0);
13056 		ret = hat_add_callback(sfmmu_tsb_cb_id, vaddr, (uint_t)tsbbytes,
13057 		    cbflags, (void *)tsbinfo, &pfn, NULL);
13058 
13059 		/*
13060 		 * Need to free up resources if we could not successfully
13061 		 * add the callback function and return an error condition.
13062 		 */
13063 		if (ret != 0) {
13064 			if (kmem_cachep) {
13065 				kmem_cache_free(kmem_cachep, vaddr);
13066 			} else {
13067 				vmem_xfree(vmp, (void *)vaddr, tsbbytes);
13068 			}
13069 			as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE,
13070 			    S_WRITE);
13071 			return (EAGAIN);
13072 		}
13073 	} else {
13074 		/*
13075 		 * Since allocation of 8K TSBs from heap is rare and occurs
13076 		 * during memory pressure we allocate them from permanent
13077 		 * memory rather than using callbacks to get the PFN.
13078 		 */
13079 		pfn = hat_getpfnum(kas.a_hat, vaddr);
13080 	}
13081 
13082 	tsbinfo->tsb_va = vaddr;
13083 	tsbinfo->tsb_szc = tsbcode;
13084 	tsbinfo->tsb_ttesz_mask = tteszmask;
13085 	tsbinfo->tsb_next = NULL;
13086 	tsbinfo->tsb_flags = 0;
13087 
13088 	sfmmu_tsbinfo_setup_phys(tsbinfo, pfn);
13089 
13090 	sfmmu_inv_tsb(vaddr, tsbbytes);
13091 
13092 	if (kmem_cachep != sfmmu_tsb8k_cache) {
13093 		as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE, S_WRITE);
13094 	}
13095 
13096 	return (0);
13097 }
13098 
13099 /*
13100  * Initialize per cpu tsb and per cpu tsbmiss_area
13101  */
13102 void
13103 sfmmu_init_tsbs(void)
13104 {
13105 	int i;
13106 	struct tsbmiss	*tsbmissp;
13107 	struct kpmtsbm	*kpmtsbmp;
13108 #ifndef sun4v
13109 	extern int	dcache_line_mask;
13110 #endif /* sun4v */
13111 	extern uint_t	vac_colors;
13112 
13113 	/*
13114 	 * Init. tsb miss area.
13115 	 */
13116 	tsbmissp = tsbmiss_area;
13117 
13118 	for (i = 0; i < NCPU; tsbmissp++, i++) {
13119 		/*
13120 		 * initialize the tsbmiss area.
13121 		 * Do this for all possible CPUs as some may be added
13122 		 * while the system is running. There is no cost to this.
13123 		 */
13124 		tsbmissp->ksfmmup = ksfmmup;
13125 #ifndef sun4v
13126 		tsbmissp->dcache_line_mask = (uint16_t)dcache_line_mask;
13127 #endif /* sun4v */
13128 		tsbmissp->khashstart =
13129 		    (struct hmehash_bucket *)va_to_pa((caddr_t)khme_hash);
13130 		tsbmissp->uhashstart =
13131 		    (struct hmehash_bucket *)va_to_pa((caddr_t)uhme_hash);
13132 		tsbmissp->khashsz = khmehash_num;
13133 		tsbmissp->uhashsz = uhmehash_num;
13134 	}
13135 
13136 	sfmmu_tsb_cb_id = hat_register_callback('T'<<16 | 'S' << 8 | 'B',
13137 	    sfmmu_tsb_pre_relocator, sfmmu_tsb_post_relocator, NULL, 0);
13138 
13139 	if (kpm_enable == 0)
13140 		return;
13141 
13142 	/* -- Begin KPM specific init -- */
13143 
13144 	if (kpm_smallpages) {
13145 		/*
13146 		 * If we're using base pagesize pages for seg_kpm
13147 		 * mappings, we use the kernel TSB since we can't afford
13148 		 * to allocate a second huge TSB for these mappings.
13149 		 */
13150 		kpm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13151 		kpm_tsbsz = ktsb_szcode;
13152 		kpmsm_tsbbase = kpm_tsbbase;
13153 		kpmsm_tsbsz = kpm_tsbsz;
13154 	} else {
13155 		/*
13156 		 * In VAC conflict case, just put the entries in the
13157 		 * kernel 8K indexed TSB for now so we can find them.
13158 		 * This could really be changed in the future if we feel
13159 		 * the need...
13160 		 */
13161 		kpmsm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13162 		kpmsm_tsbsz = ktsb_szcode;
13163 		kpm_tsbbase = ktsb_phys? ktsb4m_pbase : (uint64_t)ktsb4m_base;
13164 		kpm_tsbsz = ktsb4m_szcode;
13165 	}
13166 
13167 	kpmtsbmp = kpmtsbm_area;
13168 	for (i = 0; i < NCPU; kpmtsbmp++, i++) {
13169 		/*
13170 		 * Initialize the kpmtsbm area.
13171 		 * Do this for all possible CPUs as some may be added
13172 		 * while the system is running. There is no cost to this.
13173 		 */
13174 		kpmtsbmp->vbase = kpm_vbase;
13175 		kpmtsbmp->vend = kpm_vbase + kpm_size * vac_colors;
13176 		kpmtsbmp->sz_shift = kpm_size_shift;
13177 		kpmtsbmp->kpmp_shift = kpmp_shift;
13178 		kpmtsbmp->kpmp2pshft = (uchar_t)kpmp2pshft;
13179 		if (kpm_smallpages == 0) {
13180 			kpmtsbmp->kpmp_table_sz = kpmp_table_sz;
13181 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_table);
13182 		} else {
13183 			kpmtsbmp->kpmp_table_sz = kpmp_stable_sz;
13184 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_stable);
13185 		}
13186 		kpmtsbmp->msegphashpa = va_to_pa(memseg_phash);
13187 		kpmtsbmp->flags = KPMTSBM_ENABLE_FLAG;
13188 #ifdef	DEBUG
13189 		kpmtsbmp->flags |= (kpm_tsbmtl) ?  KPMTSBM_TLTSBM_FLAG : 0;
13190 #endif	/* DEBUG */
13191 		if (ktsb_phys)
13192 			kpmtsbmp->flags |= KPMTSBM_TSBPHYS_FLAG;
13193 	}
13194 
13195 	/* -- End KPM specific init -- */
13196 }
13197 
13198 /* Avoid using sfmmu_tsbinfo_alloc() to avoid kmem_alloc - no real reason */
13199 struct tsb_info ktsb_info[2];
13200 
13201 /*
13202  * Called from hat_kern_setup() to setup the tsb_info for ksfmmup.
13203  */
13204 void
13205 sfmmu_init_ktsbinfo()
13206 {
13207 	ASSERT(ksfmmup != NULL);
13208 	ASSERT(ksfmmup->sfmmu_tsb == NULL);
13209 	/*
13210 	 * Allocate tsbinfos for kernel and copy in data
13211 	 * to make debug easier and sun4v setup easier.
13212 	 */
13213 	ktsb_info[0].tsb_sfmmu = ksfmmup;
13214 	ktsb_info[0].tsb_szc = ktsb_szcode;
13215 	ktsb_info[0].tsb_ttesz_mask = TSB8K|TSB64K|TSB512K;
13216 	ktsb_info[0].tsb_va = ktsb_base;
13217 	ktsb_info[0].tsb_pa = ktsb_pbase;
13218 	ktsb_info[0].tsb_flags = 0;
13219 	ktsb_info[0].tsb_tte.ll = 0;
13220 	ktsb_info[0].tsb_cache = NULL;
13221 
13222 	ktsb_info[1].tsb_sfmmu = ksfmmup;
13223 	ktsb_info[1].tsb_szc = ktsb4m_szcode;
13224 	ktsb_info[1].tsb_ttesz_mask = TSB4M;
13225 	ktsb_info[1].tsb_va = ktsb4m_base;
13226 	ktsb_info[1].tsb_pa = ktsb4m_pbase;
13227 	ktsb_info[1].tsb_flags = 0;
13228 	ktsb_info[1].tsb_tte.ll = 0;
13229 	ktsb_info[1].tsb_cache = NULL;
13230 
13231 	/* Link them into ksfmmup. */
13232 	ktsb_info[0].tsb_next = &ktsb_info[1];
13233 	ktsb_info[1].tsb_next = NULL;
13234 	ksfmmup->sfmmu_tsb = &ktsb_info[0];
13235 
13236 	sfmmu_setup_tsbinfo(ksfmmup);
13237 }
13238 
13239 /*
13240  * Cache the last value returned from va_to_pa().  If the VA specified
13241  * in the current call to cached_va_to_pa() maps to the same Page (as the
13242  * previous call to cached_va_to_pa()), then compute the PA using
13243  * cached info, else call va_to_pa().
13244  *
13245  * Note: this function is neither MT-safe nor consistent in the presence
13246  * of multiple, interleaved threads.  This function was created to enable
13247  * an optimization used during boot (at a point when there's only one thread
13248  * executing on the "boot CPU", and before startup_vm() has been called).
13249  */
13250 static uint64_t
13251 cached_va_to_pa(void *vaddr)
13252 {
13253 	static uint64_t prev_vaddr_base = 0;
13254 	static uint64_t prev_pfn = 0;
13255 
13256 	if ((((uint64_t)vaddr) & MMU_PAGEMASK) == prev_vaddr_base) {
13257 		return (prev_pfn | ((uint64_t)vaddr & MMU_PAGEOFFSET));
13258 	} else {
13259 		uint64_t pa = va_to_pa(vaddr);
13260 
13261 		if (pa != ((uint64_t)-1)) {
13262 			/*
13263 			 * Computed physical address is valid.  Cache its
13264 			 * related info for the next cached_va_to_pa() call.
13265 			 */
13266 			prev_pfn = pa & MMU_PAGEMASK;
13267 			prev_vaddr_base = ((uint64_t)vaddr) & MMU_PAGEMASK;
13268 		}
13269 
13270 		return (pa);
13271 	}
13272 }
13273 
13274 /*
13275  * Carve up our nucleus hblk region.  We may allocate more hblks than
13276  * asked due to rounding errors but we are guaranteed to have at least
13277  * enough space to allocate the requested number of hblk8's and hblk1's.
13278  */
13279 void
13280 sfmmu_init_nucleus_hblks(caddr_t addr, size_t size, int nhblk8, int nhblk1)
13281 {
13282 	struct hme_blk *hmeblkp;
13283 	size_t hme8blk_sz, hme1blk_sz;
13284 	size_t i;
13285 	size_t hblk8_bound;
13286 	ulong_t j = 0, k = 0;
13287 
13288 	ASSERT(addr != NULL && size != 0);
13289 
13290 	/* Need to use proper structure alignment */
13291 	hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t));
13292 	hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t));
13293 
13294 	nucleus_hblk8.list = (void *)addr;
13295 	nucleus_hblk8.index = 0;
13296 
13297 	/*
13298 	 * Use as much memory as possible for hblk8's since we
13299 	 * expect all bop_alloc'ed memory to be allocated in 8k chunks.
13300 	 * We need to hold back enough space for the hblk1's which
13301 	 * we'll allocate next.
13302 	 */
13303 	hblk8_bound = size - (nhblk1 * hme1blk_sz) - hme8blk_sz;
13304 	for (i = 0; i <= hblk8_bound; i += hme8blk_sz, j++) {
13305 		hmeblkp = (struct hme_blk *)addr;
13306 		addr += hme8blk_sz;
13307 		hmeblkp->hblk_nuc_bit = 1;
13308 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13309 	}
13310 	nucleus_hblk8.len = j;
13311 	ASSERT(j >= nhblk8);
13312 	SFMMU_STAT_ADD(sf_hblk8_ncreate, j);
13313 
13314 	nucleus_hblk1.list = (void *)addr;
13315 	nucleus_hblk1.index = 0;
13316 	for (; i <= (size - hme1blk_sz); i += hme1blk_sz, k++) {
13317 		hmeblkp = (struct hme_blk *)addr;
13318 		addr += hme1blk_sz;
13319 		hmeblkp->hblk_nuc_bit = 1;
13320 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13321 	}
13322 	ASSERT(k >= nhblk1);
13323 	nucleus_hblk1.len = k;
13324 	SFMMU_STAT_ADD(sf_hblk1_ncreate, k);
13325 }
13326 
13327 /*
13328  * This function is currently not supported on this platform. For what
13329  * it's supposed to do, see hat.c and hat_srmmu.c
13330  */
13331 /* ARGSUSED */
13332 faultcode_t
13333 hat_softlock(struct hat *hat, caddr_t addr, size_t *lenp, page_t **ppp,
13334     uint_t flags)
13335 {
13336 	ASSERT(hat->sfmmu_xhat_provider == NULL);
13337 	return (FC_NOSUPPORT);
13338 }
13339 
13340 /*
13341  * Searchs the mapping list of the page for a mapping of the same size. If not
13342  * found the corresponding bit is cleared in the p_index field. When large
13343  * pages are more prevalent in the system, we can maintain the mapping list
13344  * in order and we don't have to traverse the list each time. Just check the
13345  * next and prev entries, and if both are of different size, we clear the bit.
13346  */
13347 static void
13348 sfmmu_rm_large_mappings(page_t *pp, int ttesz)
13349 {
13350 	struct sf_hment *sfhmep;
13351 	struct hme_blk *hmeblkp;
13352 	int	index;
13353 	pgcnt_t	npgs;
13354 
13355 	ASSERT(ttesz > TTE8K);
13356 
13357 	ASSERT(sfmmu_mlist_held(pp));
13358 
13359 	ASSERT(PP_ISMAPPED_LARGE(pp));
13360 
13361 	/*
13362 	 * Traverse mapping list looking for another mapping of same size.
13363 	 * since we only want to clear index field if all mappings of
13364 	 * that size are gone.
13365 	 */
13366 
13367 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
13368 		if (IS_PAHME(sfhmep))
13369 			continue;
13370 		hmeblkp = sfmmu_hmetohblk(sfhmep);
13371 		if (hmeblkp->hblk_xhat_bit)
13372 			continue;
13373 		if (hme_size(sfhmep) == ttesz) {
13374 			/*
13375 			 * another mapping of the same size. don't clear index.
13376 			 */
13377 			return;
13378 		}
13379 	}
13380 
13381 	/*
13382 	 * Clear the p_index bit for large page.
13383 	 */
13384 	index = PAGESZ_TO_INDEX(ttesz);
13385 	npgs = TTEPAGES(ttesz);
13386 	while (npgs-- > 0) {
13387 		ASSERT(pp->p_index & index);
13388 		pp->p_index &= ~index;
13389 		pp = PP_PAGENEXT(pp);
13390 	}
13391 }
13392 
13393 /*
13394  * return supported features
13395  */
13396 /* ARGSUSED */
13397 int
13398 hat_supported(enum hat_features feature, void *arg)
13399 {
13400 	switch (feature) {
13401 	case    HAT_SHARED_PT:
13402 	case	HAT_DYNAMIC_ISM_UNMAP:
13403 	case	HAT_VMODSORT:
13404 		return (1);
13405 	case	HAT_SHARED_REGIONS:
13406 		if (shctx_on)
13407 			return (1);
13408 		else
13409 			return (0);
13410 	default:
13411 		return (0);
13412 	}
13413 }
13414 
13415 void
13416 hat_enter(struct hat *hat)
13417 {
13418 	hatlock_t	*hatlockp;
13419 
13420 	if (hat != ksfmmup) {
13421 		hatlockp = TSB_HASH(hat);
13422 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
13423 	}
13424 }
13425 
13426 void
13427 hat_exit(struct hat *hat)
13428 {
13429 	hatlock_t	*hatlockp;
13430 
13431 	if (hat != ksfmmup) {
13432 		hatlockp = TSB_HASH(hat);
13433 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
13434 	}
13435 }
13436 
13437 /*ARGSUSED*/
13438 void
13439 hat_reserve(struct as *as, caddr_t addr, size_t len)
13440 {
13441 }
13442 
13443 static void
13444 hat_kstat_init(void)
13445 {
13446 	kstat_t *ksp;
13447 
13448 	ksp = kstat_create("unix", 0, "sfmmu_global_stat", "hat",
13449 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_global_stat),
13450 	    KSTAT_FLAG_VIRTUAL);
13451 	if (ksp) {
13452 		ksp->ks_data = (void *) &sfmmu_global_stat;
13453 		kstat_install(ksp);
13454 	}
13455 	ksp = kstat_create("unix", 0, "sfmmu_tsbsize_stat", "hat",
13456 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_tsbsize_stat),
13457 	    KSTAT_FLAG_VIRTUAL);
13458 	if (ksp) {
13459 		ksp->ks_data = (void *) &sfmmu_tsbsize_stat;
13460 		kstat_install(ksp);
13461 	}
13462 	ksp = kstat_create("unix", 0, "sfmmu_percpu_stat", "hat",
13463 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_percpu_stat) * NCPU,
13464 	    KSTAT_FLAG_WRITABLE);
13465 	if (ksp) {
13466 		ksp->ks_update = sfmmu_kstat_percpu_update;
13467 		kstat_install(ksp);
13468 	}
13469 }
13470 
13471 /* ARGSUSED */
13472 static int
13473 sfmmu_kstat_percpu_update(kstat_t *ksp, int rw)
13474 {
13475 	struct sfmmu_percpu_stat *cpu_kstat = ksp->ks_data;
13476 	struct tsbmiss *tsbm = tsbmiss_area;
13477 	struct kpmtsbm *kpmtsbm = kpmtsbm_area;
13478 	int i;
13479 
13480 	ASSERT(cpu_kstat);
13481 	if (rw == KSTAT_READ) {
13482 		for (i = 0; i < NCPU; cpu_kstat++, tsbm++, kpmtsbm++, i++) {
13483 			cpu_kstat->sf_itlb_misses = 0;
13484 			cpu_kstat->sf_dtlb_misses = 0;
13485 			cpu_kstat->sf_utsb_misses = tsbm->utsb_misses -
13486 			    tsbm->uprot_traps;
13487 			cpu_kstat->sf_ktsb_misses = tsbm->ktsb_misses +
13488 			    kpmtsbm->kpm_tsb_misses - tsbm->kprot_traps;
13489 			cpu_kstat->sf_tsb_hits = 0;
13490 			cpu_kstat->sf_umod_faults = tsbm->uprot_traps;
13491 			cpu_kstat->sf_kmod_faults = tsbm->kprot_traps;
13492 		}
13493 	} else {
13494 		/* KSTAT_WRITE is used to clear stats */
13495 		for (i = 0; i < NCPU; tsbm++, kpmtsbm++, i++) {
13496 			tsbm->utsb_misses = 0;
13497 			tsbm->ktsb_misses = 0;
13498 			tsbm->uprot_traps = 0;
13499 			tsbm->kprot_traps = 0;
13500 			kpmtsbm->kpm_dtlb_misses = 0;
13501 			kpmtsbm->kpm_tsb_misses = 0;
13502 		}
13503 	}
13504 	return (0);
13505 }
13506 
13507 #ifdef	DEBUG
13508 
13509 tte_t  *gorig[NCPU], *gcur[NCPU], *gnew[NCPU];
13510 
13511 /*
13512  * A tte checker. *orig_old is the value we read before cas.
13513  *	*cur is the value returned by cas.
13514  *	*new is the desired value when we do the cas.
13515  *
13516  *	*hmeblkp is currently unused.
13517  */
13518 
13519 /* ARGSUSED */
13520 void
13521 chk_tte(tte_t *orig_old, tte_t *cur, tte_t *new, struct hme_blk *hmeblkp)
13522 {
13523 	pfn_t i, j, k;
13524 	int cpuid = CPU->cpu_id;
13525 
13526 	gorig[cpuid] = orig_old;
13527 	gcur[cpuid] = cur;
13528 	gnew[cpuid] = new;
13529 
13530 #ifdef lint
13531 	hmeblkp = hmeblkp;
13532 #endif
13533 
13534 	if (TTE_IS_VALID(orig_old)) {
13535 		if (TTE_IS_VALID(cur)) {
13536 			i = TTE_TO_TTEPFN(orig_old);
13537 			j = TTE_TO_TTEPFN(cur);
13538 			k = TTE_TO_TTEPFN(new);
13539 			if (i != j) {
13540 				/* remap error? */
13541 				panic("chk_tte: bad pfn, 0x%lx, 0x%lx", i, j);
13542 			}
13543 
13544 			if (i != k) {
13545 				/* remap error? */
13546 				panic("chk_tte: bad pfn2, 0x%lx, 0x%lx", i, k);
13547 			}
13548 		} else {
13549 			if (TTE_IS_VALID(new)) {
13550 				panic("chk_tte: invalid cur? ");
13551 			}
13552 
13553 			i = TTE_TO_TTEPFN(orig_old);
13554 			k = TTE_TO_TTEPFN(new);
13555 			if (i != k) {
13556 				panic("chk_tte: bad pfn3, 0x%lx, 0x%lx", i, k);
13557 			}
13558 		}
13559 	} else {
13560 		if (TTE_IS_VALID(cur)) {
13561 			j = TTE_TO_TTEPFN(cur);
13562 			if (TTE_IS_VALID(new)) {
13563 				k = TTE_TO_TTEPFN(new);
13564 				if (j != k) {
13565 					panic("chk_tte: bad pfn4, 0x%lx, 0x%lx",
13566 					    j, k);
13567 				}
13568 			} else {
13569 				panic("chk_tte: why here?");
13570 			}
13571 		} else {
13572 			if (!TTE_IS_VALID(new)) {
13573 				panic("chk_tte: why here2 ?");
13574 			}
13575 		}
13576 	}
13577 }
13578 
13579 #endif /* DEBUG */
13580 
13581 extern void prefetch_tsbe_read(struct tsbe *);
13582 extern void prefetch_tsbe_write(struct tsbe *);
13583 
13584 
13585 /*
13586  * We want to prefetch 7 cache lines ahead for our read prefetch.  This gives
13587  * us optimal performance on Cheetah+.  You can only have 8 outstanding
13588  * prefetches at any one time, so we opted for 7 read prefetches and 1 write
13589  * prefetch to make the most utilization of the prefetch capability.
13590  */
13591 #define	TSBE_PREFETCH_STRIDE (7)
13592 
13593 void
13594 sfmmu_copy_tsb(struct tsb_info *old_tsbinfo, struct tsb_info *new_tsbinfo)
13595 {
13596 	int old_bytes = TSB_BYTES(old_tsbinfo->tsb_szc);
13597 	int new_bytes = TSB_BYTES(new_tsbinfo->tsb_szc);
13598 	int old_entries = TSB_ENTRIES(old_tsbinfo->tsb_szc);
13599 	int new_entries = TSB_ENTRIES(new_tsbinfo->tsb_szc);
13600 	struct tsbe *old;
13601 	struct tsbe *new;
13602 	struct tsbe *new_base = (struct tsbe *)new_tsbinfo->tsb_va;
13603 	uint64_t va;
13604 	int new_offset;
13605 	int i;
13606 	int vpshift;
13607 	int last_prefetch;
13608 
13609 	if (old_bytes == new_bytes) {
13610 		bcopy(old_tsbinfo->tsb_va, new_tsbinfo->tsb_va, new_bytes);
13611 	} else {
13612 
13613 		/*
13614 		 * A TSBE is 16 bytes which means there are four TSBE's per
13615 		 * P$ line (64 bytes), thus every 4 TSBE's we prefetch.
13616 		 */
13617 		old = (struct tsbe *)old_tsbinfo->tsb_va;
13618 		last_prefetch = old_entries - (4*(TSBE_PREFETCH_STRIDE+1));
13619 		for (i = 0; i < old_entries; i++, old++) {
13620 			if (((i & (4-1)) == 0) && (i < last_prefetch))
13621 				prefetch_tsbe_read(old);
13622 			if (!old->tte_tag.tag_invalid) {
13623 				/*
13624 				 * We have a valid TTE to remap.  Check the
13625 				 * size.  We won't remap 64K or 512K TTEs
13626 				 * because they span more than one TSB entry
13627 				 * and are indexed using an 8K virt. page.
13628 				 * Ditto for 32M and 256M TTEs.
13629 				 */
13630 				if (TTE_CSZ(&old->tte_data) == TTE64K ||
13631 				    TTE_CSZ(&old->tte_data) == TTE512K)
13632 					continue;
13633 				if (mmu_page_sizes == max_mmu_page_sizes) {
13634 					if (TTE_CSZ(&old->tte_data) == TTE32M ||
13635 					    TTE_CSZ(&old->tte_data) == TTE256M)
13636 						continue;
13637 				}
13638 
13639 				/* clear the lower 22 bits of the va */
13640 				va = *(uint64_t *)old << 22;
13641 				/* turn va into a virtual pfn */
13642 				va >>= 22 - TSB_START_SIZE;
13643 				/*
13644 				 * or in bits from the offset in the tsb
13645 				 * to get the real virtual pfn. These
13646 				 * correspond to bits [21:13] in the va
13647 				 */
13648 				vpshift =
13649 				    TTE_BSZS_SHIFT(TTE_CSZ(&old->tte_data)) &
13650 				    0x1ff;
13651 				va |= (i << vpshift);
13652 				va >>= vpshift;
13653 				new_offset = va & (new_entries - 1);
13654 				new = new_base + new_offset;
13655 				prefetch_tsbe_write(new);
13656 				*new = *old;
13657 			}
13658 		}
13659 	}
13660 }
13661 
13662 /*
13663  * unused in sfmmu
13664  */
13665 void
13666 hat_dump(void)
13667 {
13668 }
13669 
13670 /*
13671  * Called when a thread is exiting and we have switched to the kernel address
13672  * space.  Perform the same VM initialization resume() uses when switching
13673  * processes.
13674  *
13675  * Note that sfmmu_load_mmustate() is currently a no-op for kernel threads, but
13676  * we call it anyway in case the semantics change in the future.
13677  */
13678 /*ARGSUSED*/
13679 void
13680 hat_thread_exit(kthread_t *thd)
13681 {
13682 	uint_t pgsz_cnum;
13683 	uint_t pstate_save;
13684 
13685 	ASSERT(thd->t_procp->p_as == &kas);
13686 
13687 	pgsz_cnum = KCONTEXT;
13688 #ifdef sun4u
13689 	pgsz_cnum |= (ksfmmup->sfmmu_cext << CTXREG_EXT_SHIFT);
13690 #endif
13691 
13692 	/*
13693 	 * Note that sfmmu_load_mmustate() is currently a no-op for
13694 	 * kernel threads. We need to disable interrupts here,
13695 	 * simply because otherwise sfmmu_load_mmustate() would panic
13696 	 * if the caller does not disable interrupts.
13697 	 */
13698 	pstate_save = sfmmu_disable_intrs();
13699 
13700 	/* Compatibility Note: hw takes care of MMU_SCONTEXT1 */
13701 	sfmmu_setctx_sec(pgsz_cnum);
13702 	sfmmu_load_mmustate(ksfmmup);
13703 	sfmmu_enable_intrs(pstate_save);
13704 }
13705 
13706 
13707 /*
13708  * SRD support
13709  */
13710 #define	SRD_HASH_FUNCTION(vp)	(((((uintptr_t)(vp)) >> 4) ^ \
13711 				    (((uintptr_t)(vp)) >> 11)) & \
13712 				    srd_hashmask)
13713 
13714 /*
13715  * Attach the process to the srd struct associated with the exec vnode
13716  * from which the process is started.
13717  */
13718 void
13719 hat_join_srd(struct hat *sfmmup, vnode_t *evp)
13720 {
13721 	uint_t hash = SRD_HASH_FUNCTION(evp);
13722 	sf_srd_t *srdp;
13723 	sf_srd_t *newsrdp;
13724 
13725 	ASSERT(sfmmup != ksfmmup);
13726 	ASSERT(sfmmup->sfmmu_srdp == NULL);
13727 
13728 	if (!shctx_on) {
13729 		return;
13730 	}
13731 
13732 	VN_HOLD(evp);
13733 
13734 	if (srd_buckets[hash].srdb_srdp != NULL) {
13735 		mutex_enter(&srd_buckets[hash].srdb_lock);
13736 		for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13737 		    srdp = srdp->srd_hash) {
13738 			if (srdp->srd_evp == evp) {
13739 				ASSERT(srdp->srd_refcnt >= 0);
13740 				sfmmup->sfmmu_srdp = srdp;
13741 				atomic_add_32(
13742 				    (volatile uint_t *)&srdp->srd_refcnt, 1);
13743 				mutex_exit(&srd_buckets[hash].srdb_lock);
13744 				return;
13745 			}
13746 		}
13747 		mutex_exit(&srd_buckets[hash].srdb_lock);
13748 	}
13749 	newsrdp = kmem_cache_alloc(srd_cache, KM_SLEEP);
13750 	ASSERT(newsrdp->srd_next_ismrid == 0 && newsrdp->srd_next_hmerid == 0);
13751 
13752 	newsrdp->srd_evp = evp;
13753 	newsrdp->srd_refcnt = 1;
13754 	newsrdp->srd_hmergnfree = NULL;
13755 	newsrdp->srd_ismrgnfree = NULL;
13756 
13757 	mutex_enter(&srd_buckets[hash].srdb_lock);
13758 	for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13759 	    srdp = srdp->srd_hash) {
13760 		if (srdp->srd_evp == evp) {
13761 			ASSERT(srdp->srd_refcnt >= 0);
13762 			sfmmup->sfmmu_srdp = srdp;
13763 			atomic_add_32((volatile uint_t *)&srdp->srd_refcnt, 1);
13764 			mutex_exit(&srd_buckets[hash].srdb_lock);
13765 			kmem_cache_free(srd_cache, newsrdp);
13766 			return;
13767 		}
13768 	}
13769 	newsrdp->srd_hash = srd_buckets[hash].srdb_srdp;
13770 	srd_buckets[hash].srdb_srdp = newsrdp;
13771 	sfmmup->sfmmu_srdp = newsrdp;
13772 
13773 	mutex_exit(&srd_buckets[hash].srdb_lock);
13774 
13775 }
13776 
13777 static void
13778 sfmmu_leave_srd(sfmmu_t *sfmmup)
13779 {
13780 	vnode_t *evp;
13781 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
13782 	uint_t hash;
13783 	sf_srd_t **prev_srdpp;
13784 	sf_region_t *rgnp;
13785 	sf_region_t *nrgnp;
13786 #ifdef DEBUG
13787 	int rgns = 0;
13788 #endif
13789 	int i;
13790 
13791 	ASSERT(sfmmup != ksfmmup);
13792 	ASSERT(srdp != NULL);
13793 	ASSERT(srdp->srd_refcnt > 0);
13794 	ASSERT(sfmmup->sfmmu_scdp == NULL);
13795 	ASSERT(sfmmup->sfmmu_free == 1);
13796 
13797 	sfmmup->sfmmu_srdp = NULL;
13798 	evp = srdp->srd_evp;
13799 	ASSERT(evp != NULL);
13800 	if (atomic_add_32_nv(
13801 	    (volatile uint_t *)&srdp->srd_refcnt, -1)) {
13802 		VN_RELE(evp);
13803 		return;
13804 	}
13805 
13806 	hash = SRD_HASH_FUNCTION(evp);
13807 	mutex_enter(&srd_buckets[hash].srdb_lock);
13808 	for (prev_srdpp = &srd_buckets[hash].srdb_srdp;
13809 	    (srdp = *prev_srdpp) != NULL; prev_srdpp = &srdp->srd_hash) {
13810 		if (srdp->srd_evp == evp) {
13811 			break;
13812 		}
13813 	}
13814 	if (srdp == NULL || srdp->srd_refcnt) {
13815 		mutex_exit(&srd_buckets[hash].srdb_lock);
13816 		VN_RELE(evp);
13817 		return;
13818 	}
13819 	*prev_srdpp = srdp->srd_hash;
13820 	mutex_exit(&srd_buckets[hash].srdb_lock);
13821 
13822 	ASSERT(srdp->srd_refcnt == 0);
13823 	VN_RELE(evp);
13824 
13825 #ifdef DEBUG
13826 	for (i = 0; i < SFMMU_MAX_REGION_BUCKETS; i++) {
13827 		ASSERT(srdp->srd_rgnhash[i] == NULL);
13828 	}
13829 #endif /* DEBUG */
13830 
13831 	/* free each hme regions in the srd */
13832 	for (rgnp = srdp->srd_hmergnfree; rgnp != NULL; rgnp = nrgnp) {
13833 		nrgnp = rgnp->rgn_next;
13834 		ASSERT(rgnp->rgn_id < srdp->srd_next_hmerid);
13835 		ASSERT(rgnp->rgn_refcnt == 0);
13836 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13837 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13838 		ASSERT(rgnp->rgn_hmeflags == 0);
13839 		ASSERT(srdp->srd_hmergnp[rgnp->rgn_id] == rgnp);
13840 #ifdef DEBUG
13841 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13842 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13843 		}
13844 		rgns++;
13845 #endif /* DEBUG */
13846 		kmem_cache_free(region_cache, rgnp);
13847 	}
13848 	ASSERT(rgns == srdp->srd_next_hmerid);
13849 
13850 #ifdef DEBUG
13851 	rgns = 0;
13852 #endif
13853 	/* free each ism rgns in the srd */
13854 	for (rgnp = srdp->srd_ismrgnfree; rgnp != NULL; rgnp = nrgnp) {
13855 		nrgnp = rgnp->rgn_next;
13856 		ASSERT(rgnp->rgn_id < srdp->srd_next_ismrid);
13857 		ASSERT(rgnp->rgn_refcnt == 0);
13858 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13859 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13860 		ASSERT(srdp->srd_ismrgnp[rgnp->rgn_id] == rgnp);
13861 #ifdef DEBUG
13862 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13863 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13864 		}
13865 		rgns++;
13866 #endif /* DEBUG */
13867 		kmem_cache_free(region_cache, rgnp);
13868 	}
13869 	ASSERT(rgns == srdp->srd_next_ismrid);
13870 	ASSERT(srdp->srd_ismbusyrgns == 0);
13871 	ASSERT(srdp->srd_hmebusyrgns == 0);
13872 
13873 	srdp->srd_next_ismrid = 0;
13874 	srdp->srd_next_hmerid = 0;
13875 
13876 	bzero((void *)srdp->srd_ismrgnp,
13877 	    sizeof (sf_region_t *) * SFMMU_MAX_ISM_REGIONS);
13878 	bzero((void *)srdp->srd_hmergnp,
13879 	    sizeof (sf_region_t *) * SFMMU_MAX_HME_REGIONS);
13880 
13881 	ASSERT(srdp->srd_scdp == NULL);
13882 	kmem_cache_free(srd_cache, srdp);
13883 }
13884 
13885 /* ARGSUSED */
13886 static int
13887 sfmmu_srdcache_constructor(void *buf, void *cdrarg, int kmflags)
13888 {
13889 	sf_srd_t *srdp = (sf_srd_t *)buf;
13890 	bzero(buf, sizeof (*srdp));
13891 
13892 	mutex_init(&srdp->srd_mutex, NULL, MUTEX_DEFAULT, NULL);
13893 	mutex_init(&srdp->srd_scd_mutex, NULL, MUTEX_DEFAULT, NULL);
13894 	return (0);
13895 }
13896 
13897 /* ARGSUSED */
13898 static void
13899 sfmmu_srdcache_destructor(void *buf, void *cdrarg)
13900 {
13901 	sf_srd_t *srdp = (sf_srd_t *)buf;
13902 
13903 	mutex_destroy(&srdp->srd_mutex);
13904 	mutex_destroy(&srdp->srd_scd_mutex);
13905 }
13906 
13907 /*
13908  * The caller makes sure hat_join_region()/hat_leave_region() can't be called
13909  * at the same time for the same process and address range. This is ensured by
13910  * the fact that address space is locked as writer when a process joins the
13911  * regions. Therefore there's no need to hold an srd lock during the entire
13912  * execution of hat_join_region()/hat_leave_region().
13913  */
13914 
13915 #define	RGN_HASH_FUNCTION(obj)	(((((uintptr_t)(obj)) >> 4) ^ \
13916 				    (((uintptr_t)(obj)) >> 11)) & \
13917 					srd_rgn_hashmask)
13918 /*
13919  * This routine implements the shared context functionality required when
13920  * attaching a segment to an address space. It must be called from
13921  * hat_share() for D(ISM) segments and from segvn_create() for segments
13922  * with the MAP_PRIVATE and MAP_TEXT flags set. It returns a region_cookie
13923  * which is saved in the private segment data for hme segments and
13924  * the ism_map structure for ism segments.
13925  */
13926 hat_region_cookie_t
13927 hat_join_region(struct hat *sfmmup,
13928 	caddr_t r_saddr,
13929 	size_t r_size,
13930 	void *r_obj,
13931 	u_offset_t r_objoff,
13932 	uchar_t r_perm,
13933 	uchar_t r_pgszc,
13934 	hat_rgn_cb_func_t r_cb_function,
13935 	uint_t flags)
13936 {
13937 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
13938 	uint_t rhash;
13939 	uint_t rid;
13940 	hatlock_t *hatlockp;
13941 	sf_region_t *rgnp;
13942 	sf_region_t *new_rgnp = NULL;
13943 	int i;
13944 	uint16_t *nextidp;
13945 	sf_region_t **freelistp;
13946 	int maxids;
13947 	sf_region_t **rarrp;
13948 	uint16_t *busyrgnsp;
13949 	ulong_t rttecnt;
13950 	uchar_t tteflag;
13951 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
13952 	int text = (r_type == HAT_REGION_TEXT);
13953 
13954 	if (srdp == NULL || r_size == 0) {
13955 		return (HAT_INVALID_REGION_COOKIE);
13956 	}
13957 
13958 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
13959 	ASSERT(sfmmup != ksfmmup);
13960 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
13961 	ASSERT(srdp->srd_refcnt > 0);
13962 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
13963 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
13964 	ASSERT(r_pgszc < mmu_page_sizes);
13965 	if (!IS_P2ALIGNED(r_saddr, TTEBYTES(r_pgszc)) ||
13966 	    !IS_P2ALIGNED(r_size, TTEBYTES(r_pgszc))) {
13967 		panic("hat_join_region: region addr or size is not aligned\n");
13968 	}
13969 
13970 
13971 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
13972 	    SFMMU_REGION_HME;
13973 	/*
13974 	 * Currently only support shared hmes for the read only main text
13975 	 * region.
13976 	 */
13977 	if (r_type == SFMMU_REGION_HME && ((r_obj != srdp->srd_evp) ||
13978 	    (r_perm & PROT_WRITE))) {
13979 		return (HAT_INVALID_REGION_COOKIE);
13980 	}
13981 
13982 	rhash = RGN_HASH_FUNCTION(r_obj);
13983 
13984 	if (r_type == SFMMU_REGION_ISM) {
13985 		nextidp = &srdp->srd_next_ismrid;
13986 		freelistp = &srdp->srd_ismrgnfree;
13987 		maxids = SFMMU_MAX_ISM_REGIONS;
13988 		rarrp = srdp->srd_ismrgnp;
13989 		busyrgnsp = &srdp->srd_ismbusyrgns;
13990 	} else {
13991 		nextidp = &srdp->srd_next_hmerid;
13992 		freelistp = &srdp->srd_hmergnfree;
13993 		maxids = SFMMU_MAX_HME_REGIONS;
13994 		rarrp = srdp->srd_hmergnp;
13995 		busyrgnsp = &srdp->srd_hmebusyrgns;
13996 	}
13997 
13998 	mutex_enter(&srdp->srd_mutex);
13999 
14000 	for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
14001 	    rgnp = rgnp->rgn_hash) {
14002 		if (rgnp->rgn_saddr == r_saddr && rgnp->rgn_size == r_size &&
14003 		    rgnp->rgn_obj == r_obj && rgnp->rgn_objoff == r_objoff &&
14004 		    rgnp->rgn_perm == r_perm && rgnp->rgn_pgszc == r_pgszc) {
14005 			break;
14006 		}
14007 	}
14008 
14009 rfound:
14010 	if (rgnp != NULL) {
14011 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14012 		ASSERT(rgnp->rgn_cb_function == r_cb_function);
14013 		ASSERT(rgnp->rgn_refcnt >= 0);
14014 		rid = rgnp->rgn_id;
14015 		ASSERT(rid < maxids);
14016 		ASSERT(rarrp[rid] == rgnp);
14017 		ASSERT(rid < *nextidp);
14018 		atomic_add_32((volatile uint_t *)&rgnp->rgn_refcnt, 1);
14019 		mutex_exit(&srdp->srd_mutex);
14020 		if (new_rgnp != NULL) {
14021 			kmem_cache_free(region_cache, new_rgnp);
14022 		}
14023 		if (r_type == SFMMU_REGION_HME) {
14024 			int myjoin =
14025 			    (sfmmup == astosfmmu(curthread->t_procp->p_as));
14026 
14027 			sfmmu_link_to_hmeregion(sfmmup, rgnp);
14028 			/*
14029 			 * bitmap should be updated after linking sfmmu on
14030 			 * region list so that pageunload() doesn't skip
14031 			 * TSB/TLB flush. As soon as bitmap is updated another
14032 			 * thread in this process can already start accessing
14033 			 * this region.
14034 			 */
14035 			/*
14036 			 * Normally ttecnt accounting is done as part of
14037 			 * pagefault handling. But a process may not take any
14038 			 * pagefaults on shared hmeblks created by some other
14039 			 * process. To compensate for this assume that the
14040 			 * entire region will end up faulted in using
14041 			 * the region's pagesize.
14042 			 *
14043 			 */
14044 			if (r_pgszc > TTE8K) {
14045 				tteflag = 1 << r_pgszc;
14046 				if (disable_large_pages & tteflag) {
14047 					tteflag = 0;
14048 				}
14049 			} else {
14050 				tteflag = 0;
14051 			}
14052 			if (tteflag && !(sfmmup->sfmmu_rtteflags & tteflag)) {
14053 				hatlockp = sfmmu_hat_enter(sfmmup);
14054 				sfmmup->sfmmu_rtteflags |= tteflag;
14055 				sfmmu_hat_exit(hatlockp);
14056 			}
14057 			hatlockp = sfmmu_hat_enter(sfmmup);
14058 
14059 			/*
14060 			 * Preallocate 1/4 of ttecnt's in 8K TSB for >= 4M
14061 			 * region to allow for large page allocation failure.
14062 			 */
14063 			if (r_pgszc >= TTE4M) {
14064 				sfmmup->sfmmu_tsb0_4minflcnt +=
14065 				    r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14066 			}
14067 
14068 			/* update sfmmu_ttecnt with the shme rgn ttecnt */
14069 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14070 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14071 			    rttecnt);
14072 
14073 			if (text && r_pgszc >= TTE4M &&
14074 			    (tteflag || ((disable_large_pages >> TTE4M) &
14075 			    ((1 << (r_pgszc - TTE4M + 1)) - 1))) &&
14076 			    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
14077 				SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
14078 			}
14079 
14080 			sfmmu_hat_exit(hatlockp);
14081 			/*
14082 			 * On Panther we need to make sure TLB is programmed
14083 			 * to accept 32M/256M pages.  Call
14084 			 * sfmmu_check_page_sizes() now to make sure TLB is
14085 			 * setup before making hmeregions visible to other
14086 			 * threads.
14087 			 */
14088 			sfmmu_check_page_sizes(sfmmup, 1);
14089 			hatlockp = sfmmu_hat_enter(sfmmup);
14090 			SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14091 
14092 			/*
14093 			 * if context is invalid tsb miss exception code will
14094 			 * call sfmmu_check_page_sizes() and update tsbmiss
14095 			 * area later.
14096 			 */
14097 			kpreempt_disable();
14098 			if (myjoin &&
14099 			    (sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
14100 			    != INVALID_CONTEXT)) {
14101 				struct tsbmiss *tsbmp;
14102 
14103 				tsbmp = &tsbmiss_area[CPU->cpu_id];
14104 				ASSERT(sfmmup == tsbmp->usfmmup);
14105 				BT_SET(tsbmp->shmermap, rid);
14106 				if (r_pgszc > TTE64K) {
14107 					tsbmp->uhat_rtteflags |= tteflag;
14108 				}
14109 
14110 			}
14111 			kpreempt_enable();
14112 
14113 			sfmmu_hat_exit(hatlockp);
14114 			ASSERT((hat_region_cookie_t)((uint64_t)rid) !=
14115 			    HAT_INVALID_REGION_COOKIE);
14116 		} else {
14117 			hatlockp = sfmmu_hat_enter(sfmmup);
14118 			SF_RGNMAP_ADD(sfmmup->sfmmu_ismregion_map, rid);
14119 			sfmmu_hat_exit(hatlockp);
14120 		}
14121 		ASSERT(rid < maxids);
14122 
14123 		if (r_type == SFMMU_REGION_ISM) {
14124 			sfmmu_find_scd(sfmmup);
14125 		}
14126 		return ((hat_region_cookie_t)((uint64_t)rid));
14127 	}
14128 
14129 	ASSERT(new_rgnp == NULL);
14130 
14131 	if (*busyrgnsp >= maxids) {
14132 		mutex_exit(&srdp->srd_mutex);
14133 		return (HAT_INVALID_REGION_COOKIE);
14134 	}
14135 
14136 	ASSERT(MUTEX_HELD(&srdp->srd_mutex));
14137 	if (*freelistp != NULL) {
14138 		rgnp = *freelistp;
14139 		*freelistp = rgnp->rgn_next;
14140 		ASSERT(rgnp->rgn_id < *nextidp);
14141 		ASSERT(rgnp->rgn_id < maxids);
14142 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
14143 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK)
14144 		    == r_type);
14145 		ASSERT(rarrp[rgnp->rgn_id] == rgnp);
14146 		ASSERT(rgnp->rgn_hmeflags == 0);
14147 	} else {
14148 		/*
14149 		 * release local locks before memory allocation.
14150 		 */
14151 		mutex_exit(&srdp->srd_mutex);
14152 
14153 		new_rgnp = kmem_cache_alloc(region_cache, KM_SLEEP);
14154 
14155 		mutex_enter(&srdp->srd_mutex);
14156 		for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
14157 		    rgnp = rgnp->rgn_hash) {
14158 			if (rgnp->rgn_saddr == r_saddr &&
14159 			    rgnp->rgn_size == r_size &&
14160 			    rgnp->rgn_obj == r_obj &&
14161 			    rgnp->rgn_objoff == r_objoff &&
14162 			    rgnp->rgn_perm == r_perm &&
14163 			    rgnp->rgn_pgszc == r_pgszc) {
14164 				break;
14165 			}
14166 		}
14167 		if (rgnp != NULL) {
14168 			goto rfound;
14169 		}
14170 
14171 		if (*nextidp >= maxids) {
14172 			mutex_exit(&srdp->srd_mutex);
14173 			goto fail;
14174 		}
14175 		rgnp = new_rgnp;
14176 		new_rgnp = NULL;
14177 		rgnp->rgn_id = (*nextidp)++;
14178 		ASSERT(rgnp->rgn_id < maxids);
14179 		ASSERT(rarrp[rgnp->rgn_id] == NULL);
14180 		rarrp[rgnp->rgn_id] = rgnp;
14181 	}
14182 
14183 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14184 	ASSERT(rgnp->rgn_hmeflags == 0);
14185 #ifdef DEBUG
14186 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14187 		ASSERT(rgnp->rgn_ttecnt[i] == 0);
14188 	}
14189 #endif
14190 	rgnp->rgn_saddr = r_saddr;
14191 	rgnp->rgn_size = r_size;
14192 	rgnp->rgn_obj = r_obj;
14193 	rgnp->rgn_objoff = r_objoff;
14194 	rgnp->rgn_perm = r_perm;
14195 	rgnp->rgn_pgszc = r_pgszc;
14196 	rgnp->rgn_flags = r_type;
14197 	rgnp->rgn_refcnt = 0;
14198 	rgnp->rgn_cb_function = r_cb_function;
14199 	rgnp->rgn_hash = srdp->srd_rgnhash[rhash];
14200 	srdp->srd_rgnhash[rhash] = rgnp;
14201 	(*busyrgnsp)++;
14202 	ASSERT(*busyrgnsp <= maxids);
14203 	goto rfound;
14204 
14205 fail:
14206 	ASSERT(new_rgnp != NULL);
14207 	kmem_cache_free(region_cache, new_rgnp);
14208 	return (HAT_INVALID_REGION_COOKIE);
14209 }
14210 
14211 /*
14212  * This function implements the shared context functionality required
14213  * when detaching a segment from an address space. It must be called
14214  * from hat_unshare() for all D(ISM) segments and from segvn_unmap(),
14215  * for segments with a valid region_cookie.
14216  * It will also be called from all seg_vn routines which change a
14217  * segment's attributes such as segvn_setprot(), segvn_setpagesize(),
14218  * segvn_clrszc() & segvn_advise(), as well as in the case of COW fault
14219  * from segvn_fault().
14220  */
14221 void
14222 hat_leave_region(struct hat *sfmmup, hat_region_cookie_t rcookie, uint_t flags)
14223 {
14224 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14225 	sf_scd_t *scdp;
14226 	uint_t rhash;
14227 	uint_t rid = (uint_t)((uint64_t)rcookie);
14228 	hatlock_t *hatlockp = NULL;
14229 	sf_region_t *rgnp;
14230 	sf_region_t **prev_rgnpp;
14231 	sf_region_t *cur_rgnp;
14232 	void *r_obj;
14233 	int i;
14234 	caddr_t	r_saddr;
14235 	caddr_t r_eaddr;
14236 	size_t	r_size;
14237 	uchar_t	r_pgszc;
14238 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
14239 
14240 	ASSERT(sfmmup != ksfmmup);
14241 	ASSERT(srdp != NULL);
14242 	ASSERT(srdp->srd_refcnt > 0);
14243 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
14244 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
14245 	ASSERT(!sfmmup->sfmmu_free || sfmmup->sfmmu_scdp == NULL);
14246 
14247 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
14248 	    SFMMU_REGION_HME;
14249 
14250 	if (r_type == SFMMU_REGION_ISM) {
14251 		ASSERT(SFMMU_IS_ISMRID_VALID(rid));
14252 		ASSERT(rid < SFMMU_MAX_ISM_REGIONS);
14253 		rgnp = srdp->srd_ismrgnp[rid];
14254 	} else {
14255 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14256 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14257 		rgnp = srdp->srd_hmergnp[rid];
14258 	}
14259 	ASSERT(rgnp != NULL);
14260 	ASSERT(rgnp->rgn_id == rid);
14261 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14262 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14263 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
14264 
14265 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
14266 	if (r_type == SFMMU_REGION_HME && sfmmup->sfmmu_as->a_xhat != NULL) {
14267 		xhat_unload_callback_all(sfmmup->sfmmu_as, rgnp->rgn_saddr,
14268 		    rgnp->rgn_size, 0, NULL);
14269 	}
14270 
14271 	if (sfmmup->sfmmu_free) {
14272 		ulong_t rttecnt;
14273 		r_pgszc = rgnp->rgn_pgszc;
14274 		r_size = rgnp->rgn_size;
14275 
14276 		ASSERT(sfmmup->sfmmu_scdp == NULL);
14277 		if (r_type == SFMMU_REGION_ISM) {
14278 			SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14279 		} else {
14280 			/* update shme rgns ttecnt in sfmmu_ttecnt */
14281 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14282 			ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14283 
14284 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14285 			    -rttecnt);
14286 
14287 			SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14288 		}
14289 	} else if (r_type == SFMMU_REGION_ISM) {
14290 		hatlockp = sfmmu_hat_enter(sfmmup);
14291 		ASSERT(rid < srdp->srd_next_ismrid);
14292 		SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14293 		scdp = sfmmup->sfmmu_scdp;
14294 		if (scdp != NULL &&
14295 		    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
14296 			sfmmu_leave_scd(sfmmup, r_type);
14297 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14298 		}
14299 		sfmmu_hat_exit(hatlockp);
14300 	} else {
14301 		ulong_t rttecnt;
14302 		r_pgszc = rgnp->rgn_pgszc;
14303 		r_saddr = rgnp->rgn_saddr;
14304 		r_size = rgnp->rgn_size;
14305 		r_eaddr = r_saddr + r_size;
14306 
14307 		ASSERT(r_type == SFMMU_REGION_HME);
14308 		hatlockp = sfmmu_hat_enter(sfmmup);
14309 		ASSERT(rid < srdp->srd_next_hmerid);
14310 		SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14311 
14312 		/*
14313 		 * If region is part of an SCD call sfmmu_leave_scd().
14314 		 * Otherwise if process is not exiting and has valid context
14315 		 * just drop the context on the floor to lose stale TLB
14316 		 * entries and force the update of tsb miss area to reflect
14317 		 * the new region map. After that clean our TSB entries.
14318 		 */
14319 		scdp = sfmmup->sfmmu_scdp;
14320 		if (scdp != NULL &&
14321 		    SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
14322 			sfmmu_leave_scd(sfmmup, r_type);
14323 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14324 		}
14325 		sfmmu_invalidate_ctx(sfmmup);
14326 
14327 		i = TTE8K;
14328 		while (i < mmu_page_sizes) {
14329 			if (rgnp->rgn_ttecnt[i] != 0) {
14330 				sfmmu_unload_tsb_range(sfmmup, r_saddr,
14331 				    r_eaddr, i);
14332 				if (i < TTE4M) {
14333 					i = TTE4M;
14334 					continue;
14335 				} else {
14336 					break;
14337 				}
14338 			}
14339 			i++;
14340 		}
14341 		/* Remove the preallocated 1/4 8k ttecnt for 4M regions. */
14342 		if (r_pgszc >= TTE4M) {
14343 			rttecnt = r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14344 			ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
14345 			    rttecnt);
14346 			sfmmup->sfmmu_tsb0_4minflcnt -= rttecnt;
14347 		}
14348 
14349 		/* update shme rgns ttecnt in sfmmu_ttecnt */
14350 		rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14351 		ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14352 		atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc], -rttecnt);
14353 
14354 		sfmmu_hat_exit(hatlockp);
14355 		if (scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
14356 			/* sfmmup left the scd, grow private tsb */
14357 			sfmmu_check_page_sizes(sfmmup, 1);
14358 		} else {
14359 			sfmmu_check_page_sizes(sfmmup, 0);
14360 		}
14361 	}
14362 
14363 	if (r_type == SFMMU_REGION_HME) {
14364 		sfmmu_unlink_from_hmeregion(sfmmup, rgnp);
14365 	}
14366 
14367 	r_obj = rgnp->rgn_obj;
14368 	if (atomic_add_32_nv((volatile uint_t *)&rgnp->rgn_refcnt, -1)) {
14369 		return;
14370 	}
14371 
14372 	/*
14373 	 * looks like nobody uses this region anymore. Free it.
14374 	 */
14375 	rhash = RGN_HASH_FUNCTION(r_obj);
14376 	mutex_enter(&srdp->srd_mutex);
14377 	for (prev_rgnpp = &srdp->srd_rgnhash[rhash];
14378 	    (cur_rgnp = *prev_rgnpp) != NULL;
14379 	    prev_rgnpp = &cur_rgnp->rgn_hash) {
14380 		if (cur_rgnp == rgnp && cur_rgnp->rgn_refcnt == 0) {
14381 			break;
14382 		}
14383 	}
14384 
14385 	if (cur_rgnp == NULL) {
14386 		mutex_exit(&srdp->srd_mutex);
14387 		return;
14388 	}
14389 
14390 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14391 	*prev_rgnpp = rgnp->rgn_hash;
14392 	if (r_type == SFMMU_REGION_ISM) {
14393 		rgnp->rgn_flags |= SFMMU_REGION_FREE;
14394 		ASSERT(rid < srdp->srd_next_ismrid);
14395 		rgnp->rgn_next = srdp->srd_ismrgnfree;
14396 		srdp->srd_ismrgnfree = rgnp;
14397 		ASSERT(srdp->srd_ismbusyrgns > 0);
14398 		srdp->srd_ismbusyrgns--;
14399 		mutex_exit(&srdp->srd_mutex);
14400 		return;
14401 	}
14402 	mutex_exit(&srdp->srd_mutex);
14403 
14404 	/*
14405 	 * Destroy region's hmeblks.
14406 	 */
14407 	sfmmu_unload_hmeregion(srdp, rgnp);
14408 
14409 	rgnp->rgn_hmeflags = 0;
14410 
14411 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14412 	ASSERT(rgnp->rgn_id == rid);
14413 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14414 		rgnp->rgn_ttecnt[i] = 0;
14415 	}
14416 	rgnp->rgn_flags |= SFMMU_REGION_FREE;
14417 	mutex_enter(&srdp->srd_mutex);
14418 	ASSERT(rid < srdp->srd_next_hmerid);
14419 	rgnp->rgn_next = srdp->srd_hmergnfree;
14420 	srdp->srd_hmergnfree = rgnp;
14421 	ASSERT(srdp->srd_hmebusyrgns > 0);
14422 	srdp->srd_hmebusyrgns--;
14423 	mutex_exit(&srdp->srd_mutex);
14424 }
14425 
14426 /*
14427  * For now only called for hmeblk regions and not for ISM regions.
14428  */
14429 void
14430 hat_dup_region(struct hat *sfmmup, hat_region_cookie_t rcookie)
14431 {
14432 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14433 	uint_t rid = (uint_t)((uint64_t)rcookie);
14434 	sf_region_t *rgnp;
14435 	sf_rgn_link_t *rlink;
14436 	sf_rgn_link_t *hrlink;
14437 	ulong_t	rttecnt;
14438 
14439 	ASSERT(sfmmup != ksfmmup);
14440 	ASSERT(srdp != NULL);
14441 	ASSERT(srdp->srd_refcnt > 0);
14442 
14443 	ASSERT(rid < srdp->srd_next_hmerid);
14444 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14445 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14446 
14447 	rgnp = srdp->srd_hmergnp[rid];
14448 	ASSERT(rgnp->rgn_refcnt > 0);
14449 	ASSERT(rgnp->rgn_id == rid);
14450 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == SFMMU_REGION_HME);
14451 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14452 
14453 	atomic_add_32((volatile uint_t *)&rgnp->rgn_refcnt, 1);
14454 
14455 	/* LINTED: constant in conditional context */
14456 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 0);
14457 	ASSERT(rlink != NULL);
14458 	mutex_enter(&rgnp->rgn_mutex);
14459 	ASSERT(rgnp->rgn_sfmmu_head != NULL);
14460 	/* LINTED: constant in conditional context */
14461 	SFMMU_HMERID2RLINKP(rgnp->rgn_sfmmu_head, rid, hrlink, 0, 0);
14462 	ASSERT(hrlink != NULL);
14463 	ASSERT(hrlink->prev == NULL);
14464 	rlink->next = rgnp->rgn_sfmmu_head;
14465 	rlink->prev = NULL;
14466 	hrlink->prev = sfmmup;
14467 	/*
14468 	 * make sure rlink's next field is correct
14469 	 * before making this link visible.
14470 	 */
14471 	membar_stst();
14472 	rgnp->rgn_sfmmu_head = sfmmup;
14473 	mutex_exit(&rgnp->rgn_mutex);
14474 
14475 	/* update sfmmu_ttecnt with the shme rgn ttecnt */
14476 	rttecnt = rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
14477 	atomic_add_long(&sfmmup->sfmmu_ttecnt[rgnp->rgn_pgszc], rttecnt);
14478 	/* update tsb0 inflation count */
14479 	if (rgnp->rgn_pgszc >= TTE4M) {
14480 		sfmmup->sfmmu_tsb0_4minflcnt +=
14481 		    rgnp->rgn_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14482 	}
14483 	/*
14484 	 * Update regionid bitmask without hat lock since no other thread
14485 	 * can update this region bitmask right now.
14486 	 */
14487 	SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14488 }
14489 
14490 /* ARGSUSED */
14491 static int
14492 sfmmu_rgncache_constructor(void *buf, void *cdrarg, int kmflags)
14493 {
14494 	sf_region_t *rgnp = (sf_region_t *)buf;
14495 	bzero(buf, sizeof (*rgnp));
14496 
14497 	mutex_init(&rgnp->rgn_mutex, NULL, MUTEX_DEFAULT, NULL);
14498 
14499 	return (0);
14500 }
14501 
14502 /* ARGSUSED */
14503 static void
14504 sfmmu_rgncache_destructor(void *buf, void *cdrarg)
14505 {
14506 	sf_region_t *rgnp = (sf_region_t *)buf;
14507 	mutex_destroy(&rgnp->rgn_mutex);
14508 }
14509 
14510 static int
14511 sfrgnmap_isnull(sf_region_map_t *map)
14512 {
14513 	int i;
14514 
14515 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14516 		if (map->bitmap[i] != 0) {
14517 			return (0);
14518 		}
14519 	}
14520 	return (1);
14521 }
14522 
14523 static int
14524 sfhmergnmap_isnull(sf_hmeregion_map_t *map)
14525 {
14526 	int i;
14527 
14528 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
14529 		if (map->bitmap[i] != 0) {
14530 			return (0);
14531 		}
14532 	}
14533 	return (1);
14534 }
14535 
14536 #ifdef DEBUG
14537 static void
14538 check_scd_sfmmu_list(sfmmu_t **headp, sfmmu_t *sfmmup, int onlist)
14539 {
14540 	sfmmu_t *sp;
14541 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14542 
14543 	for (sp = *headp; sp != NULL; sp = sp->sfmmu_scd_link.next) {
14544 		ASSERT(srdp == sp->sfmmu_srdp);
14545 		if (sp == sfmmup) {
14546 			if (onlist) {
14547 				return;
14548 			} else {
14549 				panic("shctx: sfmmu 0x%p found on scd"
14550 				    "list 0x%p", (void *)sfmmup,
14551 				    (void *)*headp);
14552 			}
14553 		}
14554 	}
14555 	if (onlist) {
14556 		panic("shctx: sfmmu 0x%p not found on scd list 0x%p",
14557 		    (void *)sfmmup, (void *)*headp);
14558 	} else {
14559 		return;
14560 	}
14561 }
14562 #else /* DEBUG */
14563 #define	check_scd_sfmmu_list(headp, sfmmup, onlist)
14564 #endif /* DEBUG */
14565 
14566 /*
14567  * Removes an sfmmu from the SCD sfmmu list.
14568  */
14569 static void
14570 sfmmu_from_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14571 {
14572 	ASSERT(sfmmup->sfmmu_srdp != NULL);
14573 	check_scd_sfmmu_list(headp, sfmmup, 1);
14574 	if (sfmmup->sfmmu_scd_link.prev != NULL) {
14575 		ASSERT(*headp != sfmmup);
14576 		sfmmup->sfmmu_scd_link.prev->sfmmu_scd_link.next =
14577 		    sfmmup->sfmmu_scd_link.next;
14578 	} else {
14579 		ASSERT(*headp == sfmmup);
14580 		*headp = sfmmup->sfmmu_scd_link.next;
14581 	}
14582 	if (sfmmup->sfmmu_scd_link.next != NULL) {
14583 		sfmmup->sfmmu_scd_link.next->sfmmu_scd_link.prev =
14584 		    sfmmup->sfmmu_scd_link.prev;
14585 	}
14586 }
14587 
14588 
14589 /*
14590  * Adds an sfmmu to the start of the queue.
14591  */
14592 static void
14593 sfmmu_to_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14594 {
14595 	check_scd_sfmmu_list(headp, sfmmup, 0);
14596 	sfmmup->sfmmu_scd_link.prev = NULL;
14597 	sfmmup->sfmmu_scd_link.next = *headp;
14598 	if (*headp != NULL)
14599 		(*headp)->sfmmu_scd_link.prev = sfmmup;
14600 	*headp = sfmmup;
14601 }
14602 
14603 /*
14604  * Remove an scd from the start of the queue.
14605  */
14606 static void
14607 sfmmu_remove_scd(sf_scd_t **headp, sf_scd_t *scdp)
14608 {
14609 	if (scdp->scd_prev != NULL) {
14610 		ASSERT(*headp != scdp);
14611 		scdp->scd_prev->scd_next = scdp->scd_next;
14612 	} else {
14613 		ASSERT(*headp == scdp);
14614 		*headp = scdp->scd_next;
14615 	}
14616 
14617 	if (scdp->scd_next != NULL) {
14618 		scdp->scd_next->scd_prev = scdp->scd_prev;
14619 	}
14620 }
14621 
14622 /*
14623  * Add an scd to the start of the queue.
14624  */
14625 static void
14626 sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *scdp)
14627 {
14628 	scdp->scd_prev = NULL;
14629 	scdp->scd_next = *headp;
14630 	if (*headp != NULL) {
14631 		(*headp)->scd_prev = scdp;
14632 	}
14633 	*headp = scdp;
14634 }
14635 
14636 static int
14637 sfmmu_alloc_scd_tsbs(sf_srd_t *srdp, sf_scd_t *scdp)
14638 {
14639 	uint_t rid;
14640 	uint_t i;
14641 	uint_t j;
14642 	ulong_t w;
14643 	sf_region_t *rgnp;
14644 	ulong_t tte8k_cnt = 0;
14645 	ulong_t tte4m_cnt = 0;
14646 	uint_t tsb_szc;
14647 	sfmmu_t *scsfmmup = scdp->scd_sfmmup;
14648 	sfmmu_t	*ism_hatid;
14649 	struct tsb_info *newtsb;
14650 	int szc;
14651 
14652 	ASSERT(srdp != NULL);
14653 
14654 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14655 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14656 			continue;
14657 		}
14658 		j = 0;
14659 		while (w) {
14660 			if (!(w & 0x1)) {
14661 				j++;
14662 				w >>= 1;
14663 				continue;
14664 			}
14665 			rid = (i << BT_ULSHIFT) | j;
14666 			j++;
14667 			w >>= 1;
14668 
14669 			if (rid < SFMMU_MAX_HME_REGIONS) {
14670 				rgnp = srdp->srd_hmergnp[rid];
14671 				ASSERT(rgnp->rgn_id == rid);
14672 				ASSERT(rgnp->rgn_refcnt > 0);
14673 
14674 				if (rgnp->rgn_pgszc < TTE4M) {
14675 					tte8k_cnt += rgnp->rgn_size >>
14676 					    TTE_PAGE_SHIFT(TTE8K);
14677 				} else {
14678 					ASSERT(rgnp->rgn_pgszc >= TTE4M);
14679 					tte4m_cnt += rgnp->rgn_size >>
14680 					    TTE_PAGE_SHIFT(TTE4M);
14681 					/*
14682 					 * Inflate SCD tsb0 by preallocating
14683 					 * 1/4 8k ttecnt for 4M regions to
14684 					 * allow for lgpg alloc failure.
14685 					 */
14686 					tte8k_cnt += rgnp->rgn_size >>
14687 					    (TTE_PAGE_SHIFT(TTE8K) + 2);
14688 				}
14689 			} else {
14690 				rid -= SFMMU_MAX_HME_REGIONS;
14691 				rgnp = srdp->srd_ismrgnp[rid];
14692 				ASSERT(rgnp->rgn_id == rid);
14693 				ASSERT(rgnp->rgn_refcnt > 0);
14694 
14695 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14696 				ASSERT(ism_hatid->sfmmu_ismhat);
14697 
14698 				for (szc = 0; szc < TTE4M; szc++) {
14699 					tte8k_cnt +=
14700 					    ism_hatid->sfmmu_ttecnt[szc] <<
14701 					    TTE_BSZS_SHIFT(szc);
14702 				}
14703 
14704 				ASSERT(rgnp->rgn_pgszc >= TTE4M);
14705 				if (rgnp->rgn_pgszc >= TTE4M) {
14706 					tte4m_cnt += rgnp->rgn_size >>
14707 					    TTE_PAGE_SHIFT(TTE4M);
14708 				}
14709 			}
14710 		}
14711 	}
14712 
14713 	tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
14714 
14715 	/* Allocate both the SCD TSBs here. */
14716 	if (sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14717 	    tsb_szc, TSB8K|TSB64K|TSB512K, TSB_ALLOC, scsfmmup) &&
14718 	    (tsb_szc <= TSB_4M_SZCODE ||
14719 	    sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14720 	    TSB_4M_SZCODE, TSB8K|TSB64K|TSB512K,
14721 	    TSB_ALLOC, scsfmmup))) {
14722 
14723 		SFMMU_STAT(sf_scd_1sttsb_allocfail);
14724 		return (TSB_ALLOCFAIL);
14725 	} else {
14726 		scsfmmup->sfmmu_tsb->tsb_flags |= TSB_SHAREDCTX;
14727 
14728 		if (tte4m_cnt) {
14729 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
14730 			if (sfmmu_tsbinfo_alloc(&newtsb, tsb_szc,
14731 			    TSB4M|TSB32M|TSB256M, TSB_ALLOC, scsfmmup) &&
14732 			    (tsb_szc <= TSB_4M_SZCODE ||
14733 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
14734 			    TSB4M|TSB32M|TSB256M,
14735 			    TSB_ALLOC, scsfmmup))) {
14736 				/*
14737 				 * If we fail to allocate the 2nd shared tsb,
14738 				 * just free the 1st tsb, return failure.
14739 				 */
14740 				sfmmu_tsbinfo_free(scsfmmup->sfmmu_tsb);
14741 				SFMMU_STAT(sf_scd_2ndtsb_allocfail);
14742 				return (TSB_ALLOCFAIL);
14743 			} else {
14744 				ASSERT(scsfmmup->sfmmu_tsb->tsb_next == NULL);
14745 				newtsb->tsb_flags |= TSB_SHAREDCTX;
14746 				scsfmmup->sfmmu_tsb->tsb_next = newtsb;
14747 				SFMMU_STAT(sf_scd_2ndtsb_alloc);
14748 			}
14749 		}
14750 		SFMMU_STAT(sf_scd_1sttsb_alloc);
14751 	}
14752 	return (TSB_SUCCESS);
14753 }
14754 
14755 static void
14756 sfmmu_free_scd_tsbs(sfmmu_t *scd_sfmmu)
14757 {
14758 	while (scd_sfmmu->sfmmu_tsb != NULL) {
14759 		struct tsb_info *next = scd_sfmmu->sfmmu_tsb->tsb_next;
14760 		sfmmu_tsbinfo_free(scd_sfmmu->sfmmu_tsb);
14761 		scd_sfmmu->sfmmu_tsb = next;
14762 	}
14763 }
14764 
14765 /*
14766  * Link the sfmmu onto the hme region list.
14767  */
14768 void
14769 sfmmu_link_to_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14770 {
14771 	uint_t rid;
14772 	sf_rgn_link_t *rlink;
14773 	sfmmu_t *head;
14774 	sf_rgn_link_t *hrlink;
14775 
14776 	rid = rgnp->rgn_id;
14777 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14778 
14779 	/* LINTED: constant in conditional context */
14780 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 1);
14781 	ASSERT(rlink != NULL);
14782 	mutex_enter(&rgnp->rgn_mutex);
14783 	if ((head = rgnp->rgn_sfmmu_head) == NULL) {
14784 		rlink->next = NULL;
14785 		rlink->prev = NULL;
14786 		/*
14787 		 * make sure rlink's next field is NULL
14788 		 * before making this link visible.
14789 		 */
14790 		membar_stst();
14791 		rgnp->rgn_sfmmu_head = sfmmup;
14792 	} else {
14793 		/* LINTED: constant in conditional context */
14794 		SFMMU_HMERID2RLINKP(head, rid, hrlink, 0, 0);
14795 		ASSERT(hrlink != NULL);
14796 		ASSERT(hrlink->prev == NULL);
14797 		rlink->next = head;
14798 		rlink->prev = NULL;
14799 		hrlink->prev = sfmmup;
14800 		/*
14801 		 * make sure rlink's next field is correct
14802 		 * before making this link visible.
14803 		 */
14804 		membar_stst();
14805 		rgnp->rgn_sfmmu_head = sfmmup;
14806 	}
14807 	mutex_exit(&rgnp->rgn_mutex);
14808 }
14809 
14810 /*
14811  * Unlink the sfmmu from the hme region list.
14812  */
14813 void
14814 sfmmu_unlink_from_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14815 {
14816 	uint_t rid;
14817 	sf_rgn_link_t *rlink;
14818 
14819 	rid = rgnp->rgn_id;
14820 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14821 
14822 	/* LINTED: constant in conditional context */
14823 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
14824 	ASSERT(rlink != NULL);
14825 	mutex_enter(&rgnp->rgn_mutex);
14826 	if (rgnp->rgn_sfmmu_head == sfmmup) {
14827 		sfmmu_t *next = rlink->next;
14828 		rgnp->rgn_sfmmu_head = next;
14829 		/*
14830 		 * if we are stopped by xc_attention() after this
14831 		 * point the forward link walking in
14832 		 * sfmmu_rgntlb_demap() will work correctly since the
14833 		 * head correctly points to the next element.
14834 		 */
14835 		membar_stst();
14836 		rlink->next = NULL;
14837 		ASSERT(rlink->prev == NULL);
14838 		if (next != NULL) {
14839 			sf_rgn_link_t *nrlink;
14840 			/* LINTED: constant in conditional context */
14841 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14842 			ASSERT(nrlink != NULL);
14843 			ASSERT(nrlink->prev == sfmmup);
14844 			nrlink->prev = NULL;
14845 		}
14846 	} else {
14847 		sfmmu_t *next = rlink->next;
14848 		sfmmu_t *prev = rlink->prev;
14849 		sf_rgn_link_t *prlink;
14850 
14851 		ASSERT(prev != NULL);
14852 		/* LINTED: constant in conditional context */
14853 		SFMMU_HMERID2RLINKP(prev, rid, prlink, 0, 0);
14854 		ASSERT(prlink != NULL);
14855 		ASSERT(prlink->next == sfmmup);
14856 		prlink->next = next;
14857 		/*
14858 		 * if we are stopped by xc_attention()
14859 		 * after this point the forward link walking
14860 		 * will work correctly since the prev element
14861 		 * correctly points to the next element.
14862 		 */
14863 		membar_stst();
14864 		rlink->next = NULL;
14865 		rlink->prev = NULL;
14866 		if (next != NULL) {
14867 			sf_rgn_link_t *nrlink;
14868 			/* LINTED: constant in conditional context */
14869 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14870 			ASSERT(nrlink != NULL);
14871 			ASSERT(nrlink->prev == sfmmup);
14872 			nrlink->prev = prev;
14873 		}
14874 	}
14875 	mutex_exit(&rgnp->rgn_mutex);
14876 }
14877 
14878 /*
14879  * Link scd sfmmu onto ism or hme region list for each region in the
14880  * scd region map.
14881  */
14882 void
14883 sfmmu_link_scd_to_regions(sf_srd_t *srdp, sf_scd_t *scdp)
14884 {
14885 	uint_t rid;
14886 	uint_t i;
14887 	uint_t j;
14888 	ulong_t w;
14889 	sf_region_t *rgnp;
14890 	sfmmu_t *scsfmmup;
14891 
14892 	scsfmmup = scdp->scd_sfmmup;
14893 	ASSERT(scsfmmup->sfmmu_scdhat);
14894 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14895 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14896 			continue;
14897 		}
14898 		j = 0;
14899 		while (w) {
14900 			if (!(w & 0x1)) {
14901 				j++;
14902 				w >>= 1;
14903 				continue;
14904 			}
14905 			rid = (i << BT_ULSHIFT) | j;
14906 			j++;
14907 			w >>= 1;
14908 
14909 			if (rid < SFMMU_MAX_HME_REGIONS) {
14910 				rgnp = srdp->srd_hmergnp[rid];
14911 				ASSERT(rgnp->rgn_id == rid);
14912 				ASSERT(rgnp->rgn_refcnt > 0);
14913 				sfmmu_link_to_hmeregion(scsfmmup, rgnp);
14914 			} else {
14915 				sfmmu_t *ism_hatid = NULL;
14916 				ism_ment_t *ism_ment;
14917 				rid -= SFMMU_MAX_HME_REGIONS;
14918 				rgnp = srdp->srd_ismrgnp[rid];
14919 				ASSERT(rgnp->rgn_id == rid);
14920 				ASSERT(rgnp->rgn_refcnt > 0);
14921 
14922 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14923 				ASSERT(ism_hatid->sfmmu_ismhat);
14924 				ism_ment = &scdp->scd_ism_links[rid];
14925 				ism_ment->iment_hat = scsfmmup;
14926 				ism_ment->iment_base_va = rgnp->rgn_saddr;
14927 				mutex_enter(&ism_mlist_lock);
14928 				iment_add(ism_ment, ism_hatid);
14929 				mutex_exit(&ism_mlist_lock);
14930 
14931 			}
14932 		}
14933 	}
14934 }
14935 /*
14936  * Unlink scd sfmmu from ism or hme region list for each region in the
14937  * scd region map.
14938  */
14939 void
14940 sfmmu_unlink_scd_from_regions(sf_srd_t *srdp, sf_scd_t *scdp)
14941 {
14942 	uint_t rid;
14943 	uint_t i;
14944 	uint_t j;
14945 	ulong_t w;
14946 	sf_region_t *rgnp;
14947 	sfmmu_t *scsfmmup;
14948 
14949 	scsfmmup = scdp->scd_sfmmup;
14950 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14951 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14952 			continue;
14953 		}
14954 		j = 0;
14955 		while (w) {
14956 			if (!(w & 0x1)) {
14957 				j++;
14958 				w >>= 1;
14959 				continue;
14960 			}
14961 			rid = (i << BT_ULSHIFT) | j;
14962 			j++;
14963 			w >>= 1;
14964 
14965 			if (rid < SFMMU_MAX_HME_REGIONS) {
14966 				rgnp = srdp->srd_hmergnp[rid];
14967 				ASSERT(rgnp->rgn_id == rid);
14968 				ASSERT(rgnp->rgn_refcnt > 0);
14969 				sfmmu_unlink_from_hmeregion(scsfmmup,
14970 				    rgnp);
14971 
14972 			} else {
14973 				sfmmu_t *ism_hatid = NULL;
14974 				ism_ment_t *ism_ment;
14975 				rid -= SFMMU_MAX_HME_REGIONS;
14976 				rgnp = srdp->srd_ismrgnp[rid];
14977 				ASSERT(rgnp->rgn_id == rid);
14978 				ASSERT(rgnp->rgn_refcnt > 0);
14979 
14980 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14981 				ASSERT(ism_hatid->sfmmu_ismhat);
14982 				ism_ment = &scdp->scd_ism_links[rid];
14983 				ASSERT(ism_ment->iment_hat == scdp->scd_sfmmup);
14984 				ASSERT(ism_ment->iment_base_va ==
14985 				    rgnp->rgn_saddr);
14986 				mutex_enter(&ism_mlist_lock);
14987 				iment_sub(ism_ment, ism_hatid);
14988 				mutex_exit(&ism_mlist_lock);
14989 
14990 			}
14991 		}
14992 	}
14993 }
14994 /*
14995  * Allocates and initialises a new SCD structure, this is called with
14996  * the srd_scd_mutex held and returns with the reference count
14997  * initialised to 1.
14998  */
14999 static sf_scd_t *
15000 sfmmu_alloc_scd(sf_srd_t *srdp, sf_region_map_t *new_map)
15001 {
15002 	sf_scd_t *new_scdp;
15003 	sfmmu_t *scsfmmup;
15004 	int i;
15005 
15006 	ASSERT(MUTEX_HELD(&srdp->srd_scd_mutex));
15007 	new_scdp = kmem_cache_alloc(scd_cache, KM_SLEEP);
15008 
15009 	scsfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
15010 	new_scdp->scd_sfmmup = scsfmmup;
15011 	scsfmmup->sfmmu_srdp = srdp;
15012 	scsfmmup->sfmmu_scdp = new_scdp;
15013 	scsfmmup->sfmmu_tsb0_4minflcnt = 0;
15014 	scsfmmup->sfmmu_scdhat = 1;
15015 	CPUSET_ALL(scsfmmup->sfmmu_cpusran);
15016 	bzero(scsfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
15017 
15018 	ASSERT(max_mmu_ctxdoms > 0);
15019 	for (i = 0; i < max_mmu_ctxdoms; i++) {
15020 		scsfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
15021 		scsfmmup->sfmmu_ctxs[i].gnum = 0;
15022 	}
15023 
15024 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
15025 		new_scdp->scd_rttecnt[i] = 0;
15026 	}
15027 
15028 	new_scdp->scd_region_map = *new_map;
15029 	new_scdp->scd_refcnt = 1;
15030 	if (sfmmu_alloc_scd_tsbs(srdp, new_scdp) != TSB_SUCCESS) {
15031 		kmem_cache_free(scd_cache, new_scdp);
15032 		kmem_cache_free(sfmmuid_cache, scsfmmup);
15033 		return (NULL);
15034 	}
15035 	if (&mmu_init_scd) {
15036 		mmu_init_scd(new_scdp);
15037 	}
15038 	return (new_scdp);
15039 }
15040 
15041 /*
15042  * The first phase of a process joining an SCD. The hat structure is
15043  * linked to the SCD queue and then the HAT_JOIN_SCD sfmmu flag is set
15044  * and a cross-call with context invalidation is used to cause the
15045  * remaining work to be carried out in the sfmmu_tsbmiss_exception()
15046  * routine.
15047  */
15048 static void
15049 sfmmu_join_scd(sf_scd_t *scdp, sfmmu_t *sfmmup)
15050 {
15051 	hatlock_t *hatlockp;
15052 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15053 	int i;
15054 	sf_scd_t *old_scdp;
15055 
15056 	ASSERT(srdp != NULL);
15057 	ASSERT(scdp != NULL);
15058 	ASSERT(scdp->scd_refcnt > 0);
15059 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15060 
15061 	if ((old_scdp = sfmmup->sfmmu_scdp) != NULL) {
15062 		ASSERT(old_scdp != scdp);
15063 
15064 		mutex_enter(&old_scdp->scd_mutex);
15065 		sfmmu_from_scd_list(&old_scdp->scd_sf_list, sfmmup);
15066 		mutex_exit(&old_scdp->scd_mutex);
15067 		/*
15068 		 * sfmmup leaves the old scd. Update sfmmu_ttecnt to
15069 		 * include the shme rgn ttecnt for rgns that
15070 		 * were in the old SCD
15071 		 */
15072 		for (i = 0; i < mmu_page_sizes; i++) {
15073 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15074 			    old_scdp->scd_rttecnt[i]);
15075 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15076 			    sfmmup->sfmmu_scdrttecnt[i]);
15077 		}
15078 	}
15079 
15080 	/*
15081 	 * Move sfmmu to the scd lists.
15082 	 */
15083 	mutex_enter(&scdp->scd_mutex);
15084 	sfmmu_to_scd_list(&scdp->scd_sf_list, sfmmup);
15085 	mutex_exit(&scdp->scd_mutex);
15086 	SF_SCD_INCR_REF(scdp);
15087 
15088 	hatlockp = sfmmu_hat_enter(sfmmup);
15089 	/*
15090 	 * For a multi-thread process, we must stop
15091 	 * all the other threads before joining the scd.
15092 	 */
15093 
15094 	SFMMU_FLAGS_SET(sfmmup, HAT_JOIN_SCD);
15095 
15096 	sfmmu_invalidate_ctx(sfmmup);
15097 	sfmmup->sfmmu_scdp = scdp;
15098 
15099 	/*
15100 	 * Copy scd_rttecnt into sfmmup's sfmmu_scdrttecnt, and update
15101 	 * sfmmu_ttecnt to not include the rgn ttecnt just joined in SCD.
15102 	 */
15103 	for (i = 0; i < mmu_page_sizes; i++) {
15104 		sfmmup->sfmmu_scdrttecnt[i] = scdp->scd_rttecnt[i];
15105 		ASSERT(sfmmup->sfmmu_ttecnt[i] >= scdp->scd_rttecnt[i]);
15106 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15107 		    -sfmmup->sfmmu_scdrttecnt[i]);
15108 	}
15109 	/* update tsb0 inflation count */
15110 	if (old_scdp != NULL) {
15111 		sfmmup->sfmmu_tsb0_4minflcnt +=
15112 		    old_scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15113 	}
15114 	ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
15115 	    scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt);
15116 	sfmmup->sfmmu_tsb0_4minflcnt -= scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15117 
15118 	sfmmu_hat_exit(hatlockp);
15119 
15120 	if (old_scdp != NULL) {
15121 		SF_SCD_DECR_REF(srdp, old_scdp);
15122 	}
15123 
15124 }
15125 
15126 /*
15127  * This routine is called by a process to become part of an SCD. It is called
15128  * from sfmmu_tsbmiss_exception() once most of the initial work has been
15129  * done by sfmmu_join_scd(). This routine must not drop the hat lock.
15130  */
15131 static void
15132 sfmmu_finish_join_scd(sfmmu_t *sfmmup)
15133 {
15134 	struct tsb_info	*tsbinfop;
15135 
15136 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15137 	ASSERT(sfmmup->sfmmu_scdp != NULL);
15138 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD));
15139 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15140 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID));
15141 
15142 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
15143 	    tsbinfop = tsbinfop->tsb_next) {
15144 		if (tsbinfop->tsb_flags & TSB_SWAPPED) {
15145 			continue;
15146 		}
15147 		ASSERT(!(tsbinfop->tsb_flags & TSB_RELOC_FLAG));
15148 
15149 		sfmmu_inv_tsb(tsbinfop->tsb_va,
15150 		    TSB_BYTES(tsbinfop->tsb_szc));
15151 	}
15152 
15153 	/* Set HAT_CTX1_FLAG for all SCD ISMs */
15154 	sfmmu_ism_hatflags(sfmmup, 1);
15155 
15156 	SFMMU_STAT(sf_join_scd);
15157 }
15158 
15159 /*
15160  * This routine is called in order to check if there is an SCD which matches
15161  * the process's region map if not then a new SCD may be created.
15162  */
15163 static void
15164 sfmmu_find_scd(sfmmu_t *sfmmup)
15165 {
15166 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15167 	sf_scd_t *scdp, *new_scdp;
15168 	int ret;
15169 
15170 	ASSERT(srdp != NULL);
15171 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15172 
15173 	mutex_enter(&srdp->srd_scd_mutex);
15174 	for (scdp = srdp->srd_scdp; scdp != NULL;
15175 	    scdp = scdp->scd_next) {
15176 		SF_RGNMAP_EQUAL(&scdp->scd_region_map,
15177 		    &sfmmup->sfmmu_region_map, ret);
15178 		if (ret == 1) {
15179 			SF_SCD_INCR_REF(scdp);
15180 			mutex_exit(&srdp->srd_scd_mutex);
15181 			sfmmu_join_scd(scdp, sfmmup);
15182 			ASSERT(scdp->scd_refcnt >= 2);
15183 			atomic_add_32((volatile uint32_t *)
15184 			    &scdp->scd_refcnt, -1);
15185 			return;
15186 		} else {
15187 			/*
15188 			 * If the sfmmu region map is a subset of the scd
15189 			 * region map, then the assumption is that this process
15190 			 * will continue attaching to ISM segments until the
15191 			 * region maps are equal.
15192 			 */
15193 			SF_RGNMAP_IS_SUBSET(&scdp->scd_region_map,
15194 			    &sfmmup->sfmmu_region_map, ret);
15195 			if (ret == 1) {
15196 				mutex_exit(&srdp->srd_scd_mutex);
15197 				return;
15198 			}
15199 		}
15200 	}
15201 
15202 	ASSERT(scdp == NULL);
15203 	/*
15204 	 * No matching SCD has been found, create a new one.
15205 	 */
15206 	if ((new_scdp = sfmmu_alloc_scd(srdp, &sfmmup->sfmmu_region_map)) ==
15207 	    NULL) {
15208 		mutex_exit(&srdp->srd_scd_mutex);
15209 		return;
15210 	}
15211 
15212 	/*
15213 	 * sfmmu_alloc_scd() returns with a ref count of 1 on the scd.
15214 	 */
15215 
15216 	/* Set scd_rttecnt for shme rgns in SCD */
15217 	sfmmu_set_scd_rttecnt(srdp, new_scdp);
15218 
15219 	/*
15220 	 * Link scd onto srd_scdp list and scd sfmmu onto region/iment lists.
15221 	 */
15222 	sfmmu_link_scd_to_regions(srdp, new_scdp);
15223 	sfmmu_add_scd(&srdp->srd_scdp, new_scdp);
15224 	SFMMU_STAT_ADD(sf_create_scd, 1);
15225 
15226 	mutex_exit(&srdp->srd_scd_mutex);
15227 	sfmmu_join_scd(new_scdp, sfmmup);
15228 	ASSERT(new_scdp->scd_refcnt >= 2);
15229 	atomic_add_32((volatile uint32_t *)&new_scdp->scd_refcnt, -1);
15230 }
15231 
15232 /*
15233  * This routine is called by a process to remove itself from an SCD. It is
15234  * either called when the processes has detached from a segment or from
15235  * hat_free_start() as a result of calling exit.
15236  */
15237 static void
15238 sfmmu_leave_scd(sfmmu_t *sfmmup, uchar_t r_type)
15239 {
15240 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15241 	sf_srd_t *srdp =  sfmmup->sfmmu_srdp;
15242 	hatlock_t *hatlockp = TSB_HASH(sfmmup);
15243 	int i;
15244 
15245 	ASSERT(scdp != NULL);
15246 	ASSERT(srdp != NULL);
15247 
15248 	if (sfmmup->sfmmu_free) {
15249 		/*
15250 		 * If the process is part of an SCD the sfmmu is unlinked
15251 		 * from scd_sf_list.
15252 		 */
15253 		mutex_enter(&scdp->scd_mutex);
15254 		sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15255 		mutex_exit(&scdp->scd_mutex);
15256 		/*
15257 		 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15258 		 * are about to leave the SCD
15259 		 */
15260 		for (i = 0; i < mmu_page_sizes; i++) {
15261 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15262 			    scdp->scd_rttecnt[i]);
15263 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15264 			    sfmmup->sfmmu_scdrttecnt[i]);
15265 			sfmmup->sfmmu_scdrttecnt[i] = 0;
15266 		}
15267 		sfmmup->sfmmu_scdp = NULL;
15268 
15269 		SF_SCD_DECR_REF(srdp, scdp);
15270 		return;
15271 	}
15272 
15273 	ASSERT(r_type != SFMMU_REGION_ISM ||
15274 	    SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15275 	ASSERT(scdp->scd_refcnt);
15276 	ASSERT(!sfmmup->sfmmu_free);
15277 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15278 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15279 
15280 	/*
15281 	 * Wait for ISM maps to be updated.
15282 	 */
15283 	if (r_type != SFMMU_REGION_ISM) {
15284 		while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY) &&
15285 		    sfmmup->sfmmu_scdp != NULL) {
15286 			cv_wait(&sfmmup->sfmmu_tsb_cv,
15287 			    HATLOCK_MUTEXP(hatlockp));
15288 		}
15289 
15290 		if (sfmmup->sfmmu_scdp == NULL) {
15291 			sfmmu_hat_exit(hatlockp);
15292 			return;
15293 		}
15294 		SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
15295 	}
15296 
15297 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
15298 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
15299 		/*
15300 		 * Since HAT_JOIN_SCD was set our context
15301 		 * is still invalid.
15302 		 */
15303 	} else {
15304 		/*
15305 		 * For a multi-thread process, we must stop
15306 		 * all the other threads before leaving the scd.
15307 		 */
15308 
15309 		sfmmu_invalidate_ctx(sfmmup);
15310 	}
15311 
15312 	/* Clear all the rid's for ISM, delete flags, etc */
15313 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15314 	sfmmu_ism_hatflags(sfmmup, 0);
15315 
15316 	/*
15317 	 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15318 	 * are in SCD before this sfmmup leaves the SCD.
15319 	 */
15320 	for (i = 0; i < mmu_page_sizes; i++) {
15321 		ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15322 		    scdp->scd_rttecnt[i]);
15323 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15324 		    sfmmup->sfmmu_scdrttecnt[i]);
15325 		sfmmup->sfmmu_scdrttecnt[i] = 0;
15326 		/* update ismttecnt to include SCD ism before hat leaves SCD */
15327 		sfmmup->sfmmu_ismttecnt[i] += sfmmup->sfmmu_scdismttecnt[i];
15328 		sfmmup->sfmmu_scdismttecnt[i] = 0;
15329 	}
15330 	/* update tsb0 inflation count */
15331 	sfmmup->sfmmu_tsb0_4minflcnt += scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15332 
15333 	if (r_type != SFMMU_REGION_ISM) {
15334 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
15335 	}
15336 	sfmmup->sfmmu_scdp = NULL;
15337 
15338 	sfmmu_hat_exit(hatlockp);
15339 
15340 	/*
15341 	 * Unlink sfmmu from scd_sf_list this can be done without holding
15342 	 * the hat lock as we hold the sfmmu_as lock which prevents
15343 	 * hat_join_region from adding this thread to the scd again. Other
15344 	 * threads check if sfmmu_scdp is NULL under hat lock and if it's NULL
15345 	 * they won't get here, since sfmmu_leave_scd() clears sfmmu_scdp
15346 	 * while holding the hat lock.
15347 	 */
15348 	mutex_enter(&scdp->scd_mutex);
15349 	sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15350 	mutex_exit(&scdp->scd_mutex);
15351 	SFMMU_STAT(sf_leave_scd);
15352 
15353 	SF_SCD_DECR_REF(srdp, scdp);
15354 	hatlockp = sfmmu_hat_enter(sfmmup);
15355 
15356 }
15357 
15358 /*
15359  * Unlink and free up an SCD structure with a reference count of 0.
15360  */
15361 static void
15362 sfmmu_destroy_scd(sf_srd_t *srdp, sf_scd_t *scdp, sf_region_map_t *scd_rmap)
15363 {
15364 	sfmmu_t *scsfmmup;
15365 	sf_scd_t *sp;
15366 	hatlock_t *shatlockp;
15367 	int i, ret;
15368 
15369 	mutex_enter(&srdp->srd_scd_mutex);
15370 	for (sp = srdp->srd_scdp; sp != NULL; sp = sp->scd_next) {
15371 		if (sp == scdp)
15372 			break;
15373 	}
15374 	if (sp == NULL || sp->scd_refcnt) {
15375 		mutex_exit(&srdp->srd_scd_mutex);
15376 		return;
15377 	}
15378 
15379 	/*
15380 	 * It is possible that the scd has been freed and reallocated with a
15381 	 * different region map while we've been waiting for the srd_scd_mutex.
15382 	 */
15383 	SF_RGNMAP_EQUAL(scd_rmap, &sp->scd_region_map, ret);
15384 	if (ret != 1) {
15385 		mutex_exit(&srdp->srd_scd_mutex);
15386 		return;
15387 	}
15388 
15389 	ASSERT(scdp->scd_sf_list == NULL);
15390 	/*
15391 	 * Unlink scd from srd_scdp list.
15392 	 */
15393 	sfmmu_remove_scd(&srdp->srd_scdp, scdp);
15394 	mutex_exit(&srdp->srd_scd_mutex);
15395 
15396 	sfmmu_unlink_scd_from_regions(srdp, scdp);
15397 
15398 	/* Clear shared context tsb and release ctx */
15399 	scsfmmup = scdp->scd_sfmmup;
15400 
15401 	/*
15402 	 * create a barrier so that scd will not be destroyed
15403 	 * if other thread still holds the same shared hat lock.
15404 	 * E.g., sfmmu_tsbmiss_exception() needs to acquire the
15405 	 * shared hat lock before checking the shared tsb reloc flag.
15406 	 */
15407 	shatlockp = sfmmu_hat_enter(scsfmmup);
15408 	sfmmu_hat_exit(shatlockp);
15409 
15410 	sfmmu_free_scd_tsbs(scsfmmup);
15411 
15412 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
15413 		if (scsfmmup->sfmmu_hmeregion_links[i] != NULL) {
15414 			kmem_free(scsfmmup->sfmmu_hmeregion_links[i],
15415 			    SFMMU_L2_HMERLINKS_SIZE);
15416 			scsfmmup->sfmmu_hmeregion_links[i] = NULL;
15417 		}
15418 	}
15419 	kmem_cache_free(sfmmuid_cache, scsfmmup);
15420 	kmem_cache_free(scd_cache, scdp);
15421 	SFMMU_STAT(sf_destroy_scd);
15422 }
15423 
15424 /*
15425  * Modifies the HAT_CTX1_FLAG for each of the ISM segments which correspond to
15426  * bits which are set in the ism_region_map parameter. This flag indicates to
15427  * the tsbmiss handler that mapping for these segments should be loaded using
15428  * the shared context.
15429  */
15430 static void
15431 sfmmu_ism_hatflags(sfmmu_t *sfmmup, int addflag)
15432 {
15433 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15434 	ism_blk_t *ism_blkp;
15435 	ism_map_t *ism_map;
15436 	int i, rid;
15437 
15438 	ASSERT(sfmmup->sfmmu_iblk != NULL);
15439 	ASSERT(scdp != NULL);
15440 	/*
15441 	 * Note that the caller either set HAT_ISMBUSY flag or checked
15442 	 * under hat lock that HAT_ISMBUSY was not set by another thread.
15443 	 */
15444 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15445 
15446 	ism_blkp = sfmmup->sfmmu_iblk;
15447 	while (ism_blkp != NULL) {
15448 		ism_map = ism_blkp->iblk_maps;
15449 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
15450 			rid = ism_map[i].imap_rid;
15451 			if (rid == SFMMU_INVALID_ISMRID) {
15452 				continue;
15453 			}
15454 			ASSERT(rid >= 0 && rid < SFMMU_MAX_ISM_REGIONS);
15455 			if (SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid) &&
15456 			    addflag) {
15457 				ism_map[i].imap_hatflags |=
15458 				    HAT_CTX1_FLAG;
15459 			} else {
15460 				ism_map[i].imap_hatflags &=
15461 				    ~HAT_CTX1_FLAG;
15462 			}
15463 		}
15464 		ism_blkp = ism_blkp->iblk_next;
15465 	}
15466 }
15467 
15468 static int
15469 sfmmu_srd_lock_held(sf_srd_t *srdp)
15470 {
15471 	return (MUTEX_HELD(&srdp->srd_mutex));
15472 }
15473 
15474 /* ARGSUSED */
15475 static int
15476 sfmmu_scdcache_constructor(void *buf, void *cdrarg, int kmflags)
15477 {
15478 	sf_scd_t *scdp = (sf_scd_t *)buf;
15479 
15480 	bzero(buf, sizeof (sf_scd_t));
15481 	mutex_init(&scdp->scd_mutex, NULL, MUTEX_DEFAULT, NULL);
15482 	return (0);
15483 }
15484 
15485 /* ARGSUSED */
15486 static void
15487 sfmmu_scdcache_destructor(void *buf, void *cdrarg)
15488 {
15489 	sf_scd_t *scdp = (sf_scd_t *)buf;
15490 
15491 	mutex_destroy(&scdp->scd_mutex);
15492 }
15493 
15494 /*
15495  * The listp parameter is a pointer to a list of hmeblks which are partially
15496  * freed as result of calling sfmmu_hblk_hash_rm(), the last phase of the
15497  * freeing process is to cross-call all cpus to ensure that there are no
15498  * remaining cached references.
15499  *
15500  * If the local generation number is less than the global then we can free
15501  * hmeblks which are already on the pending queue as another cpu has completed
15502  * the cross-call.
15503  *
15504  * We cross-call to make sure that there are no threads on other cpus accessing
15505  * these hmblks and then complete the process of freeing them under the
15506  * following conditions:
15507  * 	The total number of pending hmeblks is greater than the threshold
15508  *	The reserve list has fewer than HBLK_RESERVE_CNT hmeblks
15509  *	It is at least 1 second since the last time we cross-called
15510  *
15511  * Otherwise, we add the hmeblks to the per-cpu pending queue.
15512  */
15513 static void
15514 sfmmu_hblks_list_purge(struct hme_blk **listp, int dontfree)
15515 {
15516 	struct hme_blk *hblkp, *pr_hblkp = NULL;
15517 	int		count = 0;
15518 	cpuset_t	cpuset = cpu_ready_set;
15519 	cpu_hme_pend_t	*cpuhp;
15520 	timestruc_t	now;
15521 	int		one_second_expired = 0;
15522 
15523 	gethrestime_lasttick(&now);
15524 
15525 	for (hblkp = *listp; hblkp != NULL; hblkp = hblkp->hblk_next) {
15526 		ASSERT(hblkp->hblk_shw_bit == 0);
15527 		ASSERT(hblkp->hblk_shared == 0);
15528 		count++;
15529 		pr_hblkp = hblkp;
15530 	}
15531 
15532 	cpuhp = &cpu_hme_pend[CPU->cpu_seqid];
15533 	mutex_enter(&cpuhp->chp_mutex);
15534 
15535 	if ((cpuhp->chp_count + count) == 0) {
15536 		mutex_exit(&cpuhp->chp_mutex);
15537 		return;
15538 	}
15539 
15540 	if ((now.tv_sec - cpuhp->chp_timestamp) > 1) {
15541 		one_second_expired  = 1;
15542 	}
15543 
15544 	if (!dontfree && (freehblkcnt < HBLK_RESERVE_CNT ||
15545 	    (cpuhp->chp_count + count) > cpu_hme_pend_thresh ||
15546 	    one_second_expired)) {
15547 		/* Append global list to local */
15548 		if (pr_hblkp == NULL) {
15549 			*listp = cpuhp->chp_listp;
15550 		} else {
15551 			pr_hblkp->hblk_next = cpuhp->chp_listp;
15552 		}
15553 		cpuhp->chp_listp = NULL;
15554 		cpuhp->chp_count = 0;
15555 		cpuhp->chp_timestamp = now.tv_sec;
15556 		mutex_exit(&cpuhp->chp_mutex);
15557 
15558 		kpreempt_disable();
15559 		CPUSET_DEL(cpuset, CPU->cpu_id);
15560 		xt_sync(cpuset);
15561 		xt_sync(cpuset);
15562 		kpreempt_enable();
15563 
15564 		/*
15565 		 * At this stage we know that no trap handlers on other
15566 		 * cpus can have references to hmeblks on the list.
15567 		 */
15568 		sfmmu_hblk_free(listp);
15569 	} else if (*listp != NULL) {
15570 		pr_hblkp->hblk_next = cpuhp->chp_listp;
15571 		cpuhp->chp_listp = *listp;
15572 		cpuhp->chp_count += count;
15573 		*listp = NULL;
15574 		mutex_exit(&cpuhp->chp_mutex);
15575 	} else {
15576 		mutex_exit(&cpuhp->chp_mutex);
15577 	}
15578 }
15579 
15580 /*
15581  * Add an hmeblk to the the hash list.
15582  */
15583 void
15584 sfmmu_hblk_hash_add(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
15585 	uint64_t hblkpa)
15586 {
15587 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
15588 #ifdef	DEBUG
15589 	if (hmebp->hmeblkp == NULL) {
15590 		ASSERT(hmebp->hmeh_nextpa == HMEBLK_ENDPA);
15591 	}
15592 #endif /* DEBUG */
15593 
15594 	hmeblkp->hblk_nextpa = hmebp->hmeh_nextpa;
15595 	/*
15596 	 * Since the TSB miss handler now does not lock the hash chain before
15597 	 * walking it, make sure that the hmeblks nextpa is globally visible
15598 	 * before we make the hmeblk globally visible by updating the chain root
15599 	 * pointer in the hash bucket.
15600 	 */
15601 	membar_producer();
15602 	hmebp->hmeh_nextpa = hblkpa;
15603 	hmeblkp->hblk_next = hmebp->hmeblkp;
15604 	hmebp->hmeblkp = hmeblkp;
15605 
15606 }
15607 
15608 /*
15609  * This function is the first part of a 2 part process to remove an hmeblk
15610  * from the hash chain. In this phase we unlink the hmeblk from the hash chain
15611  * but leave the next physical pointer unchanged. The hmeblk is then linked onto
15612  * a per-cpu pending list using the virtual address pointer.
15613  *
15614  * TSB miss trap handlers that start after this phase will no longer see
15615  * this hmeblk. TSB miss handlers that still cache this hmeblk in a register
15616  * can still use it for further chain traversal because we haven't yet modifed
15617  * the next physical pointer or freed it.
15618  *
15619  * In the second phase of hmeblk removal we'll issue a barrier xcall before
15620  * we reuse or free this hmeblk. This will make sure all lingering references to
15621  * the hmeblk after first phase disappear before we finally reclaim it.
15622  * This scheme eliminates the need for TSB miss handlers to lock hmeblk chains
15623  * during their traversal.
15624  *
15625  * The hmehash_mutex must be held when calling this function.
15626  *
15627  * Input:
15628  *	 hmebp - hme hash bucket pointer
15629  *	 hmeblkp - address of hmeblk to be removed
15630  *	 pr_hblk - virtual address of previous hmeblkp
15631  *	 listp - pointer to list of hmeblks linked by virtual address
15632  *	 free_now flag - indicates that a complete removal from the hash chains
15633  *			 is necessary.
15634  *
15635  * It is inefficient to use the free_now flag as a cross-call is required to
15636  * remove a single hmeblk from the hash chain but is necessary when hmeblks are
15637  * in short supply.
15638  */
15639 void
15640 sfmmu_hblk_hash_rm(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
15641     struct hme_blk *pr_hblk, struct hme_blk **listp,
15642     int free_now)
15643 {
15644 	int shw_size, vshift;
15645 	struct hme_blk *shw_hblkp;
15646 	uint_t		shw_mask, newshw_mask;
15647 	caddr_t		vaddr;
15648 	int		size;
15649 	cpuset_t cpuset = cpu_ready_set;
15650 
15651 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
15652 
15653 	if (hmebp->hmeblkp == hmeblkp) {
15654 		hmebp->hmeh_nextpa = hmeblkp->hblk_nextpa;
15655 		hmebp->hmeblkp = hmeblkp->hblk_next;
15656 	} else {
15657 		pr_hblk->hblk_nextpa = hmeblkp->hblk_nextpa;
15658 		pr_hblk->hblk_next = hmeblkp->hblk_next;
15659 	}
15660 
15661 	size = get_hblk_ttesz(hmeblkp);
15662 	shw_hblkp = hmeblkp->hblk_shadow;
15663 	if (shw_hblkp) {
15664 		ASSERT(hblktosfmmu(hmeblkp) != KHATID);
15665 		ASSERT(!hmeblkp->hblk_shared);
15666 #ifdef	DEBUG
15667 		if (mmu_page_sizes == max_mmu_page_sizes) {
15668 			ASSERT(size < TTE256M);
15669 		} else {
15670 			ASSERT(size < TTE4M);
15671 		}
15672 #endif /* DEBUG */
15673 
15674 		shw_size = get_hblk_ttesz(shw_hblkp);
15675 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
15676 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
15677 		ASSERT(vshift < 8);
15678 		/*
15679 		 * Atomically clear shadow mask bit
15680 		 */
15681 		do {
15682 			shw_mask = shw_hblkp->hblk_shw_mask;
15683 			ASSERT(shw_mask & (1 << vshift));
15684 			newshw_mask = shw_mask & ~(1 << vshift);
15685 			newshw_mask = cas32(&shw_hblkp->hblk_shw_mask,
15686 			    shw_mask, newshw_mask);
15687 		} while (newshw_mask != shw_mask);
15688 		hmeblkp->hblk_shadow = NULL;
15689 	}
15690 	hmeblkp->hblk_shw_bit = 0;
15691 
15692 	if (hmeblkp->hblk_shared) {
15693 #ifdef	DEBUG
15694 		sf_srd_t	*srdp;
15695 		sf_region_t	*rgnp;
15696 		uint_t		rid;
15697 
15698 		srdp = hblktosrd(hmeblkp);
15699 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
15700 		rid = hmeblkp->hblk_tag.htag_rid;
15701 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
15702 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
15703 		rgnp = srdp->srd_hmergnp[rid];
15704 		ASSERT(rgnp != NULL);
15705 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
15706 #endif /* DEBUG */
15707 		hmeblkp->hblk_shared = 0;
15708 	}
15709 	if (free_now) {
15710 		kpreempt_disable();
15711 		CPUSET_DEL(cpuset, CPU->cpu_id);
15712 		xt_sync(cpuset);
15713 		xt_sync(cpuset);
15714 		kpreempt_enable();
15715 
15716 		hmeblkp->hblk_nextpa = HMEBLK_ENDPA;
15717 		hmeblkp->hblk_next = NULL;
15718 	} else {
15719 		/* Append hmeblkp to listp for processing later. */
15720 		hmeblkp->hblk_next = *listp;
15721 		*listp = hmeblkp;
15722 	}
15723 }
15724 
15725 /*
15726  * This routine is called when memory is in short supply and returns a free
15727  * hmeblk of the requested size from the cpu pending lists.
15728  */
15729 static struct hme_blk *
15730 sfmmu_check_pending_hblks(int size)
15731 {
15732 	int i;
15733 	struct hme_blk *hmeblkp = NULL, *last_hmeblkp;
15734 	int found_hmeblk;
15735 	cpuset_t cpuset = cpu_ready_set;
15736 	cpu_hme_pend_t *cpuhp;
15737 
15738 	/* Flush cpu hblk pending queues */
15739 	for (i = 0; i < NCPU; i++) {
15740 		cpuhp = &cpu_hme_pend[i];
15741 		if (cpuhp->chp_listp != NULL)  {
15742 			mutex_enter(&cpuhp->chp_mutex);
15743 			if (cpuhp->chp_listp == NULL)  {
15744 				mutex_exit(&cpuhp->chp_mutex);
15745 				continue;
15746 			}
15747 			found_hmeblk = 0;
15748 			last_hmeblkp = NULL;
15749 			for (hmeblkp = cpuhp->chp_listp; hmeblkp != NULL;
15750 			    hmeblkp = hmeblkp->hblk_next) {
15751 				if (get_hblk_ttesz(hmeblkp) == size) {
15752 					if (last_hmeblkp == NULL) {
15753 						cpuhp->chp_listp =
15754 						    hmeblkp->hblk_next;
15755 					} else {
15756 						last_hmeblkp->hblk_next =
15757 						    hmeblkp->hblk_next;
15758 					}
15759 					ASSERT(cpuhp->chp_count > 0);
15760 					cpuhp->chp_count--;
15761 					found_hmeblk = 1;
15762 					break;
15763 				} else {
15764 					last_hmeblkp = hmeblkp;
15765 				}
15766 			}
15767 			mutex_exit(&cpuhp->chp_mutex);
15768 
15769 			if (found_hmeblk) {
15770 				kpreempt_disable();
15771 				CPUSET_DEL(cpuset, CPU->cpu_id);
15772 				xt_sync(cpuset);
15773 				xt_sync(cpuset);
15774 				kpreempt_enable();
15775 				return (hmeblkp);
15776 			}
15777 		}
15778 	}
15779 	return (NULL);
15780 }
15781