xref: /illumos-gate/usr/src/uts/sfmmu/vm/hat_sfmmu.c (revision 47842382d52f28aa3173aa6b511781c322ccb6a2)
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  * This flag is set to 0 via the MD in platforms that do not support
188  * I-cache coherency in hardware. Used to enable "soft exec" mode.
189  * The MD "coherency" property is optional, and defaults to 1 (because
190  * coherent I-cache is the norm.)
191  */
192 uint_t	icache_is_coherent = 1;
193 
194 /*
195  * Flag to allow the creation of non-cacheable translations
196  * to system memory. It is off by default. At the moment this
197  * flag is used by the ecache error injector. The error injector
198  * will turn it on when creating such a translation then shut it
199  * off when it's finished.
200  */
201 
202 int	sfmmu_allow_nc_trans = 0;
203 
204 /*
205  * Flag to disable large page support.
206  * 	value of 1 => disable all large pages.
207  *	bits 1, 2, and 3 are to disable 64K, 512K and 4M pages respectively.
208  *
209  * For example, use the value 0x4 to disable 512K pages.
210  *
211  */
212 #define	LARGE_PAGES_OFF		0x1
213 
214 /*
215  * The disable_large_pages and disable_ism_large_pages variables control
216  * hat_memload_array and the page sizes to be used by ISM and the kernel.
217  *
218  * The disable_auto_data_large_pages and disable_auto_text_large_pages variables
219  * are only used to control which OOB pages to use at upper VM segment creation
220  * time, and are set in hat_init_pagesizes and used in the map_pgsz* routines.
221  * Their values may come from platform or CPU specific code to disable page
222  * sizes that should not be used.
223  *
224  * WARNING: 512K pages are currently not supported for ISM/DISM.
225  */
226 uint_t	disable_large_pages = 0;
227 uint_t	disable_ism_large_pages = (1 << TTE512K);
228 uint_t	disable_auto_data_large_pages = 0;
229 uint_t	disable_auto_text_large_pages = 0;
230 
231 /*
232  * Private sfmmu data structures for hat management
233  */
234 static struct kmem_cache *sfmmuid_cache;
235 static struct kmem_cache *mmuctxdom_cache;
236 
237 /*
238  * Private sfmmu data structures for tsb management
239  */
240 static struct kmem_cache *sfmmu_tsbinfo_cache;
241 static struct kmem_cache *sfmmu_tsb8k_cache;
242 static struct kmem_cache *sfmmu_tsb_cache[NLGRPS_MAX];
243 static vmem_t *kmem_bigtsb_arena;
244 static vmem_t *kmem_tsb_arena;
245 
246 /*
247  * sfmmu static variables for hmeblk resource management.
248  */
249 static vmem_t *hat_memload1_arena; /* HAT translation arena for sfmmu1_cache */
250 static struct kmem_cache *sfmmu8_cache;
251 static struct kmem_cache *sfmmu1_cache;
252 static struct kmem_cache *pa_hment_cache;
253 
254 static kmutex_t 	ism_mlist_lock;	/* mutex for ism mapping list */
255 /*
256  * private data for ism
257  */
258 static struct kmem_cache *ism_blk_cache;
259 static struct kmem_cache *ism_ment_cache;
260 #define	ISMID_STARTADDR	NULL
261 
262 /*
263  * Region management data structures and function declarations.
264  */
265 
266 static void	sfmmu_leave_srd(sfmmu_t *);
267 static int	sfmmu_srdcache_constructor(void *, void *, int);
268 static void	sfmmu_srdcache_destructor(void *, void *);
269 static int	sfmmu_rgncache_constructor(void *, void *, int);
270 static void	sfmmu_rgncache_destructor(void *, void *);
271 static int	sfrgnmap_isnull(sf_region_map_t *);
272 static int	sfhmergnmap_isnull(sf_hmeregion_map_t *);
273 static int	sfmmu_scdcache_constructor(void *, void *, int);
274 static void	sfmmu_scdcache_destructor(void *, void *);
275 static void	sfmmu_rgn_cb_noop(caddr_t, caddr_t, caddr_t,
276     size_t, void *, u_offset_t);
277 
278 static uint_t srd_hashmask = SFMMU_MAX_SRD_BUCKETS - 1;
279 static sf_srd_bucket_t *srd_buckets;
280 static struct kmem_cache *srd_cache;
281 static uint_t srd_rgn_hashmask = SFMMU_MAX_REGION_BUCKETS - 1;
282 static struct kmem_cache *region_cache;
283 static struct kmem_cache *scd_cache;
284 
285 #ifdef sun4v
286 int use_bigtsb_arena = 1;
287 #else
288 int use_bigtsb_arena = 0;
289 #endif
290 
291 /* External /etc/system tunable, for turning on&off the shctx support */
292 int disable_shctx = 0;
293 /* Internal variable, set by MD if the HW supports shctx feature */
294 int shctx_on = 0;
295 
296 #ifdef DEBUG
297 static void check_scd_sfmmu_list(sfmmu_t **, sfmmu_t *, int);
298 #endif
299 static void sfmmu_to_scd_list(sfmmu_t **, sfmmu_t *);
300 static void sfmmu_from_scd_list(sfmmu_t **, sfmmu_t *);
301 
302 static sf_scd_t *sfmmu_alloc_scd(sf_srd_t *, sf_region_map_t *);
303 static void sfmmu_find_scd(sfmmu_t *);
304 static void sfmmu_join_scd(sf_scd_t *, sfmmu_t *);
305 static void sfmmu_finish_join_scd(sfmmu_t *);
306 static void sfmmu_leave_scd(sfmmu_t *, uchar_t);
307 static void sfmmu_destroy_scd(sf_srd_t *, sf_scd_t *, sf_region_map_t *);
308 static int sfmmu_alloc_scd_tsbs(sf_srd_t *, sf_scd_t *);
309 static void sfmmu_free_scd_tsbs(sfmmu_t *);
310 static void sfmmu_tsb_inv_ctx(sfmmu_t *);
311 static int find_ism_rid(sfmmu_t *, sfmmu_t *, caddr_t, uint_t *);
312 static void sfmmu_ism_hatflags(sfmmu_t *, int);
313 static int sfmmu_srd_lock_held(sf_srd_t *);
314 static void sfmmu_remove_scd(sf_scd_t **, sf_scd_t *);
315 static void sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *);
316 static void sfmmu_link_scd_to_regions(sf_srd_t *, sf_scd_t *);
317 static void sfmmu_unlink_scd_from_regions(sf_srd_t *, sf_scd_t *);
318 static void sfmmu_link_to_hmeregion(sfmmu_t *, sf_region_t *);
319 static void sfmmu_unlink_from_hmeregion(sfmmu_t *, sf_region_t *);
320 
321 /*
322  * ``hat_lock'' is a hashed mutex lock for protecting sfmmu TSB lists,
323  * HAT flags, synchronizing TLB/TSB coherency, and context management.
324  * The lock is hashed on the sfmmup since the case where we need to lock
325  * all processes is rare but does occur (e.g. we need to unload a shared
326  * mapping from all processes using the mapping).  We have a lot of buckets,
327  * and each slab of sfmmu_t's can use about a quarter of them, giving us
328  * a fairly good distribution without wasting too much space and overhead
329  * when we have to grab them all.
330  */
331 #define	SFMMU_NUM_LOCK	128		/* must be power of two */
332 hatlock_t	hat_lock[SFMMU_NUM_LOCK];
333 
334 /*
335  * Hash algorithm optimized for a small number of slabs.
336  *  7 is (highbit((sizeof sfmmu_t)) - 1)
337  * This hash algorithm is based upon the knowledge that sfmmu_t's come from a
338  * kmem_cache, and thus they will be sequential within that cache.  In
339  * addition, each new slab will have a different "color" up to cache_maxcolor
340  * which will skew the hashing for each successive slab which is allocated.
341  * If the size of sfmmu_t changed to a larger size, this algorithm may need
342  * to be revisited.
343  */
344 #define	TSB_HASH_SHIFT_BITS (7)
345 #define	PTR_HASH(x) ((uintptr_t)x >> TSB_HASH_SHIFT_BITS)
346 
347 #ifdef DEBUG
348 int tsb_hash_debug = 0;
349 #define	TSB_HASH(sfmmup)	\
350 	(tsb_hash_debug ? &hat_lock[0] : \
351 	&hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)])
352 #else	/* DEBUG */
353 #define	TSB_HASH(sfmmup)	&hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)]
354 #endif	/* DEBUG */
355 
356 
357 /* sfmmu_replace_tsb() return codes. */
358 typedef enum tsb_replace_rc {
359 	TSB_SUCCESS,
360 	TSB_ALLOCFAIL,
361 	TSB_LOSTRACE,
362 	TSB_ALREADY_SWAPPED,
363 	TSB_CANTGROW
364 } tsb_replace_rc_t;
365 
366 /*
367  * Flags for TSB allocation routines.
368  */
369 #define	TSB_ALLOC	0x01
370 #define	TSB_FORCEALLOC	0x02
371 #define	TSB_GROW	0x04
372 #define	TSB_SHRINK	0x08
373 #define	TSB_SWAPIN	0x10
374 
375 /*
376  * Support for HAT callbacks.
377  */
378 #define	SFMMU_MAX_RELOC_CALLBACKS	10
379 int sfmmu_max_cb_id = SFMMU_MAX_RELOC_CALLBACKS;
380 static id_t sfmmu_cb_nextid = 0;
381 static id_t sfmmu_tsb_cb_id;
382 struct sfmmu_callback *sfmmu_cb_table;
383 
384 /*
385  * Kernel page relocation is enabled by default for non-caged
386  * kernel pages.  This has little effect unless segkmem_reloc is
387  * set, since by default kernel memory comes from inside the
388  * kernel cage.
389  */
390 int hat_kpr_enabled = 1;
391 
392 kmutex_t	kpr_mutex;
393 kmutex_t	kpr_suspendlock;
394 kthread_t	*kreloc_thread;
395 
396 /*
397  * Enable VA->PA translation sanity checking on DEBUG kernels.
398  * Disabled by default.  This is incompatible with some
399  * drivers (error injector, RSM) so if it breaks you get
400  * to keep both pieces.
401  */
402 int hat_check_vtop = 0;
403 
404 /*
405  * Private sfmmu routines (prototypes)
406  */
407 static struct hme_blk *sfmmu_shadow_hcreate(sfmmu_t *, caddr_t, int, uint_t);
408 static struct 	hme_blk *sfmmu_hblk_alloc(sfmmu_t *, caddr_t,
409 			struct hmehash_bucket *, uint_t, hmeblk_tag, uint_t,
410 			uint_t);
411 static caddr_t	sfmmu_hblk_unload(struct hat *, struct hme_blk *, caddr_t,
412 			caddr_t, demap_range_t *, uint_t);
413 static caddr_t	sfmmu_hblk_sync(struct hat *, struct hme_blk *, caddr_t,
414 			caddr_t, int);
415 static void	sfmmu_hblk_free(struct hme_blk **);
416 static void	sfmmu_hblks_list_purge(struct hme_blk **, int);
417 static uint_t	sfmmu_get_free_hblk(struct hme_blk **, uint_t);
418 static uint_t	sfmmu_put_free_hblk(struct hme_blk *, uint_t);
419 static struct hme_blk *sfmmu_hblk_steal(int);
420 static int	sfmmu_steal_this_hblk(struct hmehash_bucket *,
421 			struct hme_blk *, uint64_t, struct hme_blk *);
422 static caddr_t	sfmmu_hblk_unlock(struct hme_blk *, caddr_t, caddr_t);
423 
424 static void	hat_do_memload_array(struct hat *, caddr_t, size_t,
425 		    struct page **, uint_t, uint_t, uint_t);
426 static void	hat_do_memload(struct hat *, caddr_t, struct page *,
427 		    uint_t, uint_t, uint_t);
428 static void	sfmmu_memload_batchsmall(struct hat *, caddr_t, page_t **,
429 		    uint_t, uint_t, pgcnt_t, uint_t);
430 void		sfmmu_tteload(struct hat *, tte_t *, caddr_t, page_t *,
431 			uint_t);
432 static int	sfmmu_tteload_array(sfmmu_t *, tte_t *, caddr_t, page_t **,
433 			uint_t, uint_t);
434 static struct hmehash_bucket *sfmmu_tteload_acquire_hashbucket(sfmmu_t *,
435 					caddr_t, int, uint_t);
436 static struct hme_blk *sfmmu_tteload_find_hmeblk(sfmmu_t *,
437 			struct hmehash_bucket *, caddr_t, uint_t, uint_t,
438 			uint_t);
439 static int	sfmmu_tteload_addentry(sfmmu_t *, struct hme_blk *, tte_t *,
440 			caddr_t, page_t **, uint_t, uint_t);
441 static void	sfmmu_tteload_release_hashbucket(struct hmehash_bucket *);
442 
443 static int	sfmmu_pagearray_setup(caddr_t, page_t **, tte_t *, int);
444 static pfn_t	sfmmu_uvatopfn(caddr_t, sfmmu_t *, tte_t *);
445 void		sfmmu_memtte(tte_t *, pfn_t, uint_t, int);
446 #ifdef VAC
447 static void	sfmmu_vac_conflict(struct hat *, caddr_t, page_t *);
448 static int	sfmmu_vacconflict_array(caddr_t, page_t *, int *);
449 int	tst_tnc(page_t *pp, pgcnt_t);
450 void	conv_tnc(page_t *pp, int);
451 #endif
452 
453 static void	sfmmu_get_ctx(sfmmu_t *);
454 static void	sfmmu_free_sfmmu(sfmmu_t *);
455 
456 static void	sfmmu_ttesync(struct hat *, caddr_t, tte_t *, page_t *);
457 static void	sfmmu_chgattr(struct hat *, caddr_t, size_t, uint_t, int);
458 
459 cpuset_t	sfmmu_pageunload(page_t *, struct sf_hment *, int);
460 static void	hat_pagereload(struct page *, struct page *);
461 static cpuset_t	sfmmu_pagesync(page_t *, struct sf_hment *, uint_t);
462 #ifdef VAC
463 void	sfmmu_page_cache_array(page_t *, int, int, pgcnt_t);
464 static void	sfmmu_page_cache(page_t *, int, int, int);
465 #endif
466 
467 cpuset_t	sfmmu_rgntlb_demap(caddr_t, sf_region_t *,
468     struct hme_blk *, int);
469 static void	sfmmu_tlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *,
470 			pfn_t, int, int, int, int);
471 static void	sfmmu_ismtlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *,
472 			pfn_t, int);
473 static void	sfmmu_tlb_demap(caddr_t, sfmmu_t *, struct hme_blk *, int, int);
474 static void	sfmmu_tlb_range_demap(demap_range_t *);
475 static void	sfmmu_invalidate_ctx(sfmmu_t *);
476 static void	sfmmu_sync_mmustate(sfmmu_t *);
477 
478 static void 	sfmmu_tsbinfo_setup_phys(struct tsb_info *, pfn_t);
479 static int	sfmmu_tsbinfo_alloc(struct tsb_info **, int, int, uint_t,
480 			sfmmu_t *);
481 static void	sfmmu_tsb_free(struct tsb_info *);
482 static void	sfmmu_tsbinfo_free(struct tsb_info *);
483 static int	sfmmu_init_tsbinfo(struct tsb_info *, int, int, uint_t,
484 			sfmmu_t *);
485 static void	sfmmu_tsb_chk_reloc(sfmmu_t *, hatlock_t *);
486 static void	sfmmu_tsb_swapin(sfmmu_t *, hatlock_t *);
487 static int	sfmmu_select_tsb_szc(pgcnt_t);
488 static void	sfmmu_mod_tsb(sfmmu_t *, caddr_t, tte_t *, int);
489 #define		sfmmu_load_tsb(sfmmup, vaddr, tte, szc) \
490 	sfmmu_mod_tsb(sfmmup, vaddr, tte, szc)
491 #define		sfmmu_unload_tsb(sfmmup, vaddr, szc)    \
492 	sfmmu_mod_tsb(sfmmup, vaddr, NULL, szc)
493 static void	sfmmu_copy_tsb(struct tsb_info *, struct tsb_info *);
494 static tsb_replace_rc_t sfmmu_replace_tsb(sfmmu_t *, struct tsb_info *, uint_t,
495     hatlock_t *, uint_t);
496 static void	sfmmu_size_tsb(sfmmu_t *, int, uint64_t, uint64_t, int);
497 
498 #ifdef VAC
499 void	sfmmu_cache_flush(pfn_t, int);
500 void	sfmmu_cache_flushcolor(int, pfn_t);
501 #endif
502 static caddr_t	sfmmu_hblk_chgattr(sfmmu_t *, struct hme_blk *, caddr_t,
503 			caddr_t, demap_range_t *, uint_t, int);
504 
505 static uint64_t	sfmmu_vtop_attr(uint_t, int mode, tte_t *);
506 static uint_t	sfmmu_ptov_attr(tte_t *);
507 static caddr_t	sfmmu_hblk_chgprot(sfmmu_t *, struct hme_blk *, caddr_t,
508 			caddr_t, demap_range_t *, uint_t);
509 static uint_t	sfmmu_vtop_prot(uint_t, uint_t *);
510 static int	sfmmu_idcache_constructor(void *, void *, int);
511 static void	sfmmu_idcache_destructor(void *, void *);
512 static int	sfmmu_hblkcache_constructor(void *, void *, int);
513 static void	sfmmu_hblkcache_destructor(void *, void *);
514 static void	sfmmu_hblkcache_reclaim(void *);
515 static void	sfmmu_shadow_hcleanup(sfmmu_t *, struct hme_blk *,
516 			struct hmehash_bucket *);
517 static void	sfmmu_hblk_hash_rm(struct hmehash_bucket *, struct hme_blk *,
518 			struct hme_blk *, struct hme_blk **, int);
519 static void	sfmmu_hblk_hash_add(struct hmehash_bucket *, struct hme_blk *,
520 			uint64_t);
521 static struct hme_blk *sfmmu_check_pending_hblks(int);
522 static void	sfmmu_free_hblks(sfmmu_t *, caddr_t, caddr_t, int);
523 static void	sfmmu_cleanup_rhblk(sf_srd_t *, caddr_t, uint_t, int);
524 static void	sfmmu_unload_hmeregion_va(sf_srd_t *, uint_t, caddr_t, caddr_t,
525 			int, caddr_t *);
526 static void	sfmmu_unload_hmeregion(sf_srd_t *, sf_region_t *);
527 
528 static void	sfmmu_rm_large_mappings(page_t *, int);
529 
530 static void	hat_lock_init(void);
531 static void	hat_kstat_init(void);
532 static int	sfmmu_kstat_percpu_update(kstat_t *ksp, int rw);
533 static void	sfmmu_set_scd_rttecnt(sf_srd_t *, sf_scd_t *);
534 static	int	sfmmu_is_rgnva(sf_srd_t *, caddr_t, ulong_t, ulong_t);
535 static void	sfmmu_check_page_sizes(sfmmu_t *, int);
536 int	fnd_mapping_sz(page_t *);
537 static void	iment_add(struct ism_ment *,  struct hat *);
538 static void	iment_sub(struct ism_ment *, struct hat *);
539 static pgcnt_t	ism_tsb_entries(sfmmu_t *, int szc);
540 extern void	sfmmu_setup_tsbinfo(sfmmu_t *);
541 extern void	sfmmu_clear_utsbinfo(void);
542 
543 static void	sfmmu_ctx_wrap_around(mmu_ctx_t *);
544 
545 /* kpm globals */
546 #ifdef	DEBUG
547 /*
548  * Enable trap level tsbmiss handling
549  */
550 int	kpm_tsbmtl = 1;
551 
552 /*
553  * Flush the TLB on kpm mapout. Note: Xcalls are used (again) for the
554  * required TLB shootdowns in this case, so handle w/ care. Off by default.
555  */
556 int	kpm_tlb_flush;
557 #endif	/* DEBUG */
558 
559 static void	*sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *, size_t, int);
560 
561 #ifdef DEBUG
562 static void	sfmmu_check_hblk_flist();
563 #endif
564 
565 /*
566  * Semi-private sfmmu data structures.  Some of them are initialize in
567  * startup or in hat_init. Some of them are private but accessed by
568  * assembly code or mach_sfmmu.c
569  */
570 struct hmehash_bucket *uhme_hash;	/* user hmeblk hash table */
571 struct hmehash_bucket *khme_hash;	/* kernel hmeblk hash table */
572 uint64_t	uhme_hash_pa;		/* PA of uhme_hash */
573 uint64_t	khme_hash_pa;		/* PA of khme_hash */
574 int 		uhmehash_num;		/* # of buckets in user hash table */
575 int 		khmehash_num;		/* # of buckets in kernel hash table */
576 
577 uint_t		max_mmu_ctxdoms = 0;	/* max context domains in the system */
578 mmu_ctx_t	**mmu_ctxs_tbl;		/* global array of context domains */
579 uint64_t	mmu_saved_gnum = 0;	/* to init incoming MMUs' gnums */
580 
581 #define	DEFAULT_NUM_CTXS_PER_MMU 8192
582 uint_t	nctxs = DEFAULT_NUM_CTXS_PER_MMU;
583 
584 int		cache;			/* describes system cache */
585 
586 caddr_t		ktsb_base;		/* kernel 8k-indexed tsb base address */
587 uint64_t	ktsb_pbase;		/* kernel 8k-indexed tsb phys address */
588 int		ktsb_szcode;		/* kernel 8k-indexed tsb size code */
589 int		ktsb_sz;		/* kernel 8k-indexed tsb size */
590 
591 caddr_t		ktsb4m_base;		/* kernel 4m-indexed tsb base address */
592 uint64_t	ktsb4m_pbase;		/* kernel 4m-indexed tsb phys address */
593 int		ktsb4m_szcode;		/* kernel 4m-indexed tsb size code */
594 int		ktsb4m_sz;		/* kernel 4m-indexed tsb size */
595 
596 uint64_t	kpm_tsbbase;		/* kernel seg_kpm 4M TSB base address */
597 int		kpm_tsbsz;		/* kernel seg_kpm 4M TSB size code */
598 uint64_t	kpmsm_tsbbase;		/* kernel seg_kpm 8K TSB base address */
599 int		kpmsm_tsbsz;		/* kernel seg_kpm 8K TSB size code */
600 
601 #ifndef sun4v
602 int		utsb_dtlb_ttenum = -1;	/* index in TLB for utsb locked TTE */
603 int		utsb4m_dtlb_ttenum = -1; /* index in TLB for 4M TSB TTE */
604 int		dtlb_resv_ttenum;	/* index in TLB of first reserved TTE */
605 caddr_t		utsb_vabase;		/* reserved kernel virtual memory */
606 caddr_t		utsb4m_vabase;		/* for trap handler TSB accesses */
607 #endif /* sun4v */
608 uint64_t	tsb_alloc_bytes = 0;	/* bytes allocated to TSBs */
609 vmem_t		*kmem_tsb_default_arena[NLGRPS_MAX];	/* For dynamic TSBs */
610 vmem_t		*kmem_bigtsb_default_arena[NLGRPS_MAX]; /* dynamic 256M TSBs */
611 
612 /*
613  * Size to use for TSB slabs.  Future platforms that support page sizes
614  * larger than 4M may wish to change these values, and provide their own
615  * assembly macros for building and decoding the TSB base register contents.
616  * Note disable_large_pages will override the value set here.
617  */
618 static	uint_t tsb_slab_ttesz = TTE4M;
619 size_t	tsb_slab_size = MMU_PAGESIZE4M;
620 uint_t	tsb_slab_shift = MMU_PAGESHIFT4M;
621 /* PFN mask for TTE */
622 size_t	tsb_slab_mask = MMU_PAGEOFFSET4M >> MMU_PAGESHIFT;
623 
624 /*
625  * Size to use for TSB slabs.  These are used only when 256M tsb arenas
626  * exist.
627  */
628 static uint_t	bigtsb_slab_ttesz = TTE256M;
629 static size_t	bigtsb_slab_size = MMU_PAGESIZE256M;
630 static uint_t	bigtsb_slab_shift = MMU_PAGESHIFT256M;
631 /* 256M page alignment for 8K pfn */
632 static size_t	bigtsb_slab_mask = MMU_PAGEOFFSET256M >> MMU_PAGESHIFT;
633 
634 /* largest TSB size to grow to, will be smaller on smaller memory systems */
635 static int	tsb_max_growsize = 0;
636 
637 /*
638  * Tunable parameters dealing with TSB policies.
639  */
640 
641 /*
642  * This undocumented tunable forces all 8K TSBs to be allocated from
643  * the kernel heap rather than from the kmem_tsb_default_arena arenas.
644  */
645 #ifdef	DEBUG
646 int	tsb_forceheap = 0;
647 #endif	/* DEBUG */
648 
649 /*
650  * Decide whether to use per-lgroup arenas, or one global set of
651  * TSB arenas.  The default is not to break up per-lgroup, since
652  * most platforms don't recognize any tangible benefit from it.
653  */
654 int	tsb_lgrp_affinity = 0;
655 
656 /*
657  * Used for growing the TSB based on the process RSS.
658  * tsb_rss_factor is based on the smallest TSB, and is
659  * shifted by the TSB size to determine if we need to grow.
660  * The default will grow the TSB if the number of TTEs for
661  * this page size exceeds 75% of the number of TSB entries,
662  * which should _almost_ eliminate all conflict misses
663  * (at the expense of using up lots and lots of memory).
664  */
665 #define	TSB_RSS_FACTOR		(TSB_ENTRIES(TSB_MIN_SZCODE) * 0.75)
666 #define	SFMMU_RSS_TSBSIZE(tsbszc)	(tsb_rss_factor << tsbszc)
667 #define	SELECT_TSB_SIZECODE(pgcnt) ( \
668 	(enable_tsb_rss_sizing)? sfmmu_select_tsb_szc(pgcnt) : \
669 	default_tsb_size)
670 #define	TSB_OK_SHRINK()	\
671 	(tsb_alloc_bytes > tsb_alloc_hiwater || freemem < desfree)
672 #define	TSB_OK_GROW()	\
673 	(tsb_alloc_bytes < tsb_alloc_hiwater && freemem > desfree)
674 
675 int	enable_tsb_rss_sizing = 1;
676 int	tsb_rss_factor	= (int)TSB_RSS_FACTOR;
677 
678 /* which TSB size code to use for new address spaces or if rss sizing off */
679 int default_tsb_size = TSB_8K_SZCODE;
680 
681 static uint64_t tsb_alloc_hiwater; /* limit TSB reserved memory */
682 uint64_t tsb_alloc_hiwater_factor; /* tsb_alloc_hiwater = physmem / this */
683 #define	TSB_ALLOC_HIWATER_FACTOR_DEFAULT	32
684 
685 #ifdef DEBUG
686 static int tsb_random_size = 0;	/* set to 1 to test random tsb sizes on alloc */
687 static int tsb_grow_stress = 0;	/* if set to 1, keep replacing TSB w/ random */
688 static int tsb_alloc_mtbf = 0;	/* fail allocation every n attempts */
689 static int tsb_alloc_fail_mtbf = 0;
690 static int tsb_alloc_count = 0;
691 #endif /* DEBUG */
692 
693 /* if set to 1, will remap valid TTEs when growing TSB. */
694 int tsb_remap_ttes = 1;
695 
696 /*
697  * If we have more than this many mappings, allocate a second TSB.
698  * This default is chosen because the I/D fully associative TLBs are
699  * assumed to have at least 8 available entries. Platforms with a
700  * larger fully-associative TLB could probably override the default.
701  */
702 
703 #ifdef sun4v
704 int tsb_sectsb_threshold = 0;
705 #else
706 int tsb_sectsb_threshold = 8;
707 #endif
708 
709 /*
710  * kstat data
711  */
712 struct sfmmu_global_stat sfmmu_global_stat;
713 struct sfmmu_tsbsize_stat sfmmu_tsbsize_stat;
714 
715 /*
716  * Global data
717  */
718 sfmmu_t 	*ksfmmup;		/* kernel's hat id */
719 
720 #ifdef DEBUG
721 static void	chk_tte(tte_t *, tte_t *, tte_t *, struct hme_blk *);
722 #endif
723 
724 /* sfmmu locking operations */
725 static kmutex_t *sfmmu_mlspl_enter(struct page *, int);
726 static int	sfmmu_mlspl_held(struct page *, int);
727 
728 kmutex_t *sfmmu_page_enter(page_t *);
729 void	sfmmu_page_exit(kmutex_t *);
730 int	sfmmu_page_spl_held(struct page *);
731 
732 /* sfmmu internal locking operations - accessed directly */
733 static void	sfmmu_mlist_reloc_enter(page_t *, page_t *,
734 				kmutex_t **, kmutex_t **);
735 static void	sfmmu_mlist_reloc_exit(kmutex_t *, kmutex_t *);
736 static hatlock_t *
737 		sfmmu_hat_enter(sfmmu_t *);
738 static hatlock_t *
739 		sfmmu_hat_tryenter(sfmmu_t *);
740 static void	sfmmu_hat_exit(hatlock_t *);
741 static void	sfmmu_hat_lock_all(void);
742 static void	sfmmu_hat_unlock_all(void);
743 static void	sfmmu_ismhat_enter(sfmmu_t *, int);
744 static void	sfmmu_ismhat_exit(sfmmu_t *, int);
745 
746 /*
747  * Array of mutexes protecting a page's mapping list and p_nrm field.
748  *
749  * The hash function looks complicated, but is made up so that:
750  *
751  * "pp" not shifted, so adjacent pp values will hash to different cache lines
752  *  (8 byte alignment * 8 bytes/mutes == 64 byte coherency subblock)
753  *
754  * "pp" >> mml_shift, incorporates more source bits into the hash result
755  *
756  *  "& (mml_table_size - 1), should be faster than using remainder "%"
757  *
758  * Hopefully, mml_table, mml_table_size and mml_shift are all in the same
759  * cacheline, since they get declared next to each other below. We'll trust
760  * ld not to do something random.
761  */
762 #ifdef	DEBUG
763 int mlist_hash_debug = 0;
764 #define	MLIST_HASH(pp)	(mlist_hash_debug ? &mml_table[0] : \
765 	&mml_table[((uintptr_t)(pp) + \
766 	((uintptr_t)(pp) >> mml_shift)) & (mml_table_sz - 1)])
767 #else	/* !DEBUG */
768 #define	MLIST_HASH(pp)   &mml_table[ \
769 	((uintptr_t)(pp) + ((uintptr_t)(pp) >> mml_shift)) & (mml_table_sz - 1)]
770 #endif	/* !DEBUG */
771 
772 kmutex_t		*mml_table;
773 uint_t			mml_table_sz;	/* must be a power of 2 */
774 uint_t			mml_shift;	/* log2(mml_table_sz) + 3 for align */
775 
776 kpm_hlk_t	*kpmp_table;
777 uint_t		kpmp_table_sz;	/* must be a power of 2 */
778 uchar_t		kpmp_shift;
779 
780 kpm_shlk_t	*kpmp_stable;
781 uint_t		kpmp_stable_sz;	/* must be a power of 2 */
782 
783 /*
784  * SPL_HASH was improved to avoid false cache line sharing
785  */
786 #define	SPL_TABLE_SIZE	128
787 #define	SPL_MASK	(SPL_TABLE_SIZE - 1)
788 #define	SPL_SHIFT	7		/* log2(SPL_TABLE_SIZE) */
789 
790 #define	SPL_INDEX(pp) \
791 	((((uintptr_t)(pp) >> SPL_SHIFT) ^ \
792 	((uintptr_t)(pp) >> (SPL_SHIFT << 1))) & \
793 	(SPL_TABLE_SIZE - 1))
794 
795 #define	SPL_HASH(pp)    \
796 	(&sfmmu_page_lock[SPL_INDEX(pp) & SPL_MASK].pad_mutex)
797 
798 static	pad_mutex_t	sfmmu_page_lock[SPL_TABLE_SIZE];
799 
800 
801 /*
802  * hat_unload_callback() will group together callbacks in order
803  * to avoid xt_sync() calls.  This is the maximum size of the group.
804  */
805 #define	MAX_CB_ADDR	32
806 
807 tte_t	hw_tte;
808 static ulong_t sfmmu_dmr_maxbit = DMR_MAXBIT;
809 
810 static char	*mmu_ctx_kstat_names[] = {
811 	"mmu_ctx_tsb_exceptions",
812 	"mmu_ctx_tsb_raise_exception",
813 	"mmu_ctx_wrap_around",
814 };
815 
816 /*
817  * Wrapper for vmem_xalloc since vmem_create only allows limited
818  * parameters for vm_source_alloc functions.  This function allows us
819  * to specify alignment consistent with the size of the object being
820  * allocated.
821  */
822 static void *
823 sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag)
824 {
825 	return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag));
826 }
827 
828 /* Common code for setting tsb_alloc_hiwater. */
829 #define	SFMMU_SET_TSB_ALLOC_HIWATER(pages)	tsb_alloc_hiwater = \
830 		ptob(pages) / tsb_alloc_hiwater_factor
831 
832 /*
833  * Set tsb_max_growsize to allow at most all of physical memory to be mapped by
834  * a single TSB.  physmem is the number of physical pages so we need physmem 8K
835  * TTEs to represent all those physical pages.  We round this up by using
836  * 1<<highbit().  To figure out which size code to use, remember that the size
837  * code is just an amount to shift the smallest TSB size to get the size of
838  * this TSB.  So we subtract that size, TSB_START_SIZE, from highbit() (or
839  * highbit() - 1) to get the size code for the smallest TSB that can represent
840  * all of physical memory, while erring on the side of too much.
841  *
842  * Restrict tsb_max_growsize to make sure that:
843  *	1) TSBs can't grow larger than the TSB slab size
844  *	2) TSBs can't grow larger than UTSB_MAX_SZCODE.
845  */
846 #define	SFMMU_SET_TSB_MAX_GROWSIZE(pages) {				\
847 	int	_i, _szc, _slabszc, _tsbszc;				\
848 									\
849 	_i = highbit(pages);						\
850 	if ((1 << (_i - 1)) == (pages))					\
851 		_i--;		/* 2^n case, round down */              \
852 	_szc = _i - TSB_START_SIZE;					\
853 	_slabszc = bigtsb_slab_shift - (TSB_START_SIZE + TSB_ENTRY_SHIFT); \
854 	_tsbszc = MIN(_szc, _slabszc);                                  \
855 	tsb_max_growsize = MIN(_tsbszc, UTSB_MAX_SZCODE);               \
856 }
857 
858 /*
859  * Given a pointer to an sfmmu and a TTE size code, return a pointer to the
860  * tsb_info which handles that TTE size.
861  */
862 #define	SFMMU_GET_TSBINFO(tsbinfop, sfmmup, tte_szc) {			\
863 	(tsbinfop) = (sfmmup)->sfmmu_tsb;				\
864 	ASSERT(((tsbinfop)->tsb_flags & TSB_SHAREDCTX) ||		\
865 	    sfmmu_hat_lock_held(sfmmup));				\
866 	if ((tte_szc) >= TTE4M)	{					\
867 		ASSERT((tsbinfop) != NULL);				\
868 		(tsbinfop) = (tsbinfop)->tsb_next;			\
869 	}								\
870 }
871 
872 /*
873  * Macro to use to unload entries from the TSB.
874  * It has knowledge of which page sizes get replicated in the TSB
875  * and will call the appropriate unload routine for the appropriate size.
876  */
877 #define	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, ismhat)		\
878 {									\
879 	int ttesz = get_hblk_ttesz(hmeblkp);				\
880 	if (ttesz == TTE8K || ttesz == TTE4M) {				\
881 		sfmmu_unload_tsb(sfmmup, addr, ttesz);			\
882 	} else {							\
883 		caddr_t sva = ismhat ? addr : 				\
884 		    (caddr_t)get_hblk_base(hmeblkp);			\
885 		caddr_t eva = sva + get_hblk_span(hmeblkp);		\
886 		ASSERT(addr >= sva && addr < eva);			\
887 		sfmmu_unload_tsb_range(sfmmup, sva, eva, ttesz);	\
888 	}								\
889 }
890 
891 
892 /* Update tsb_alloc_hiwater after memory is configured. */
893 /*ARGSUSED*/
894 static void
895 sfmmu_update_post_add(void *arg, pgcnt_t delta_pages)
896 {
897 	/* Assumes physmem has already been updated. */
898 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
899 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
900 }
901 
902 /*
903  * Update tsb_alloc_hiwater before memory is deleted.  We'll do nothing here
904  * and update tsb_alloc_hiwater and tsb_max_growsize after the memory is
905  * deleted.
906  */
907 /*ARGSUSED*/
908 static int
909 sfmmu_update_pre_del(void *arg, pgcnt_t delta_pages)
910 {
911 	return (0);
912 }
913 
914 /* Update tsb_alloc_hiwater after memory fails to be unconfigured. */
915 /*ARGSUSED*/
916 static void
917 sfmmu_update_post_del(void *arg, pgcnt_t delta_pages, int cancelled)
918 {
919 	/*
920 	 * Whether the delete was cancelled or not, just go ahead and update
921 	 * tsb_alloc_hiwater and tsb_max_growsize.
922 	 */
923 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
924 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
925 }
926 
927 static kphysm_setup_vector_t sfmmu_update_vec = {
928 	KPHYSM_SETUP_VECTOR_VERSION,	/* version */
929 	sfmmu_update_post_add,		/* post_add */
930 	sfmmu_update_pre_del,		/* pre_del */
931 	sfmmu_update_post_del		/* post_del */
932 };
933 
934 
935 /*
936  * HME_BLK HASH PRIMITIVES
937  */
938 
939 /*
940  * Enter a hme on the mapping list for page pp.
941  * When large pages are more prevalent in the system we might want to
942  * keep the mapping list in ascending order by the hment size. For now,
943  * small pages are more frequent, so don't slow it down.
944  */
945 #define	HME_ADD(hme, pp)					\
946 {								\
947 	ASSERT(sfmmu_mlist_held(pp));				\
948 								\
949 	hme->hme_prev = NULL;					\
950 	hme->hme_next = pp->p_mapping;				\
951 	hme->hme_page = pp;					\
952 	if (pp->p_mapping) {					\
953 		((struct sf_hment *)(pp->p_mapping))->hme_prev = hme;\
954 		ASSERT(pp->p_share > 0);			\
955 	} else  {						\
956 		/* EMPTY */					\
957 		ASSERT(pp->p_share == 0);			\
958 	}							\
959 	pp->p_mapping = hme;					\
960 	pp->p_share++;						\
961 }
962 
963 /*
964  * Enter a hme on the mapping list for page pp.
965  * If we are unmapping a large translation, we need to make sure that the
966  * change is reflect in the corresponding bit of the p_index field.
967  */
968 #define	HME_SUB(hme, pp)					\
969 {								\
970 	ASSERT(sfmmu_mlist_held(pp));				\
971 	ASSERT(hme->hme_page == pp || IS_PAHME(hme));		\
972 								\
973 	if (pp->p_mapping == NULL) {				\
974 		panic("hme_remove - no mappings");		\
975 	}							\
976 								\
977 	membar_stst();	/* ensure previous stores finish */	\
978 								\
979 	ASSERT(pp->p_share > 0);				\
980 	pp->p_share--;						\
981 								\
982 	if (hme->hme_prev) {					\
983 		ASSERT(pp->p_mapping != hme);			\
984 		ASSERT(hme->hme_prev->hme_page == pp ||		\
985 			IS_PAHME(hme->hme_prev));		\
986 		hme->hme_prev->hme_next = hme->hme_next;	\
987 	} else {						\
988 		ASSERT(pp->p_mapping == hme);			\
989 		pp->p_mapping = hme->hme_next;			\
990 		ASSERT((pp->p_mapping == NULL) ?		\
991 			(pp->p_share == 0) : 1);		\
992 	}							\
993 								\
994 	if (hme->hme_next) {					\
995 		ASSERT(hme->hme_next->hme_page == pp ||		\
996 			IS_PAHME(hme->hme_next));		\
997 		hme->hme_next->hme_prev = hme->hme_prev;	\
998 	}							\
999 								\
1000 	/* zero out the entry */				\
1001 	hme->hme_next = NULL;					\
1002 	hme->hme_prev = NULL;					\
1003 	hme->hme_page = NULL;					\
1004 								\
1005 	if (hme_size(hme) > TTE8K) {				\
1006 		/* remove mappings for remainder of large pg */	\
1007 		sfmmu_rm_large_mappings(pp, hme_size(hme));	\
1008 	}							\
1009 }
1010 
1011 /*
1012  * This function returns the hment given the hme_blk and a vaddr.
1013  * It assumes addr has already been checked to belong to hme_blk's
1014  * range.
1015  */
1016 #define	HBLKTOHME(hment, hmeblkp, addr)					\
1017 {									\
1018 	int index;							\
1019 	HBLKTOHME_IDX(hment, hmeblkp, addr, index)			\
1020 }
1021 
1022 /*
1023  * Version of HBLKTOHME that also returns the index in hmeblkp
1024  * of the hment.
1025  */
1026 #define	HBLKTOHME_IDX(hment, hmeblkp, addr, idx)			\
1027 {									\
1028 	ASSERT(in_hblk_range((hmeblkp), (addr)));			\
1029 									\
1030 	if (get_hblk_ttesz(hmeblkp) == TTE8K) {				\
1031 		idx = (((uintptr_t)(addr) >> MMU_PAGESHIFT) & (NHMENTS-1)); \
1032 	} else								\
1033 		idx = 0;						\
1034 									\
1035 	(hment) = &(hmeblkp)->hblk_hme[idx];				\
1036 }
1037 
1038 /*
1039  * Disable any page sizes not supported by the CPU
1040  */
1041 void
1042 hat_init_pagesizes()
1043 {
1044 	int 		i;
1045 
1046 	mmu_exported_page_sizes = 0;
1047 	for (i = TTE8K; i < max_mmu_page_sizes; i++) {
1048 
1049 		szc_2_userszc[i] = (uint_t)-1;
1050 		userszc_2_szc[i] = (uint_t)-1;
1051 
1052 		if ((mmu_exported_pagesize_mask & (1 << i)) == 0) {
1053 			disable_large_pages |= (1 << i);
1054 		} else {
1055 			szc_2_userszc[i] = mmu_exported_page_sizes;
1056 			userszc_2_szc[mmu_exported_page_sizes] = i;
1057 			mmu_exported_page_sizes++;
1058 		}
1059 	}
1060 
1061 	disable_ism_large_pages |= disable_large_pages;
1062 	disable_auto_data_large_pages = disable_large_pages;
1063 	disable_auto_text_large_pages = disable_large_pages;
1064 
1065 	/*
1066 	 * Initialize mmu-specific large page sizes.
1067 	 */
1068 	if (&mmu_large_pages_disabled) {
1069 		disable_large_pages |= mmu_large_pages_disabled(HAT_LOAD);
1070 		disable_ism_large_pages |=
1071 		    mmu_large_pages_disabled(HAT_LOAD_SHARE);
1072 		disable_auto_data_large_pages |=
1073 		    mmu_large_pages_disabled(HAT_AUTO_DATA);
1074 		disable_auto_text_large_pages |=
1075 		    mmu_large_pages_disabled(HAT_AUTO_TEXT);
1076 	}
1077 }
1078 
1079 /*
1080  * Initialize the hardware address translation structures.
1081  */
1082 void
1083 hat_init(void)
1084 {
1085 	int 		i;
1086 	uint_t		sz;
1087 	size_t		size;
1088 
1089 	hat_lock_init();
1090 	hat_kstat_init();
1091 
1092 	/*
1093 	 * Hardware-only bits in a TTE
1094 	 */
1095 	MAKE_TTE_MASK(&hw_tte);
1096 
1097 	hat_init_pagesizes();
1098 
1099 	/* Initialize the hash locks */
1100 	for (i = 0; i < khmehash_num; i++) {
1101 		mutex_init(&khme_hash[i].hmehash_mutex, NULL,
1102 		    MUTEX_DEFAULT, NULL);
1103 		khme_hash[i].hmeh_nextpa = HMEBLK_ENDPA;
1104 	}
1105 	for (i = 0; i < uhmehash_num; i++) {
1106 		mutex_init(&uhme_hash[i].hmehash_mutex, NULL,
1107 		    MUTEX_DEFAULT, NULL);
1108 		uhme_hash[i].hmeh_nextpa = HMEBLK_ENDPA;
1109 	}
1110 	khmehash_num--;		/* make sure counter starts from 0 */
1111 	uhmehash_num--;		/* make sure counter starts from 0 */
1112 
1113 	/*
1114 	 * Allocate context domain structures.
1115 	 *
1116 	 * A platform may choose to modify max_mmu_ctxdoms in
1117 	 * set_platform_defaults(). If a platform does not define
1118 	 * a set_platform_defaults() or does not choose to modify
1119 	 * max_mmu_ctxdoms, it gets one MMU context domain for every CPU.
1120 	 *
1121 	 * For sun4v, there will be one global context domain, this is to
1122 	 * avoid the ldom cpu substitution problem.
1123 	 *
1124 	 * For all platforms that have CPUs sharing MMUs, this
1125 	 * value must be defined.
1126 	 */
1127 	if (max_mmu_ctxdoms == 0) {
1128 #ifndef sun4v
1129 		max_mmu_ctxdoms = max_ncpus;
1130 #else /* sun4v */
1131 		max_mmu_ctxdoms = 1;
1132 #endif /* sun4v */
1133 	}
1134 
1135 	size = max_mmu_ctxdoms * sizeof (mmu_ctx_t *);
1136 	mmu_ctxs_tbl = kmem_zalloc(size, KM_SLEEP);
1137 
1138 	/* mmu_ctx_t is 64 bytes aligned */
1139 	mmuctxdom_cache = kmem_cache_create("mmuctxdom_cache",
1140 	    sizeof (mmu_ctx_t), 64, NULL, NULL, NULL, NULL, NULL, 0);
1141 	/*
1142 	 * MMU context domain initialization for the Boot CPU.
1143 	 * This needs the context domains array allocated above.
1144 	 */
1145 	mutex_enter(&cpu_lock);
1146 	sfmmu_cpu_init(CPU);
1147 	mutex_exit(&cpu_lock);
1148 
1149 	/*
1150 	 * Intialize ism mapping list lock.
1151 	 */
1152 
1153 	mutex_init(&ism_mlist_lock, NULL, MUTEX_DEFAULT, NULL);
1154 
1155 	/*
1156 	 * Each sfmmu structure carries an array of MMU context info
1157 	 * structures, one per context domain. The size of this array depends
1158 	 * on the maximum number of context domains. So, the size of the
1159 	 * sfmmu structure varies per platform.
1160 	 *
1161 	 * sfmmu is allocated from static arena, because trap
1162 	 * handler at TL > 0 is not allowed to touch kernel relocatable
1163 	 * memory. sfmmu's alignment is changed to 64 bytes from
1164 	 * default 8 bytes, as the lower 6 bits will be used to pass
1165 	 * pgcnt to vtag_flush_pgcnt_tl1.
1166 	 */
1167 	size = sizeof (sfmmu_t) + sizeof (sfmmu_ctx_t) * (max_mmu_ctxdoms - 1);
1168 
1169 	sfmmuid_cache = kmem_cache_create("sfmmuid_cache", size,
1170 	    64, sfmmu_idcache_constructor, sfmmu_idcache_destructor,
1171 	    NULL, NULL, static_arena, 0);
1172 
1173 	sfmmu_tsbinfo_cache = kmem_cache_create("sfmmu_tsbinfo_cache",
1174 	    sizeof (struct tsb_info), 0, NULL, NULL, NULL, NULL, NULL, 0);
1175 
1176 	/*
1177 	 * Since we only use the tsb8k cache to "borrow" pages for TSBs
1178 	 * from the heap when low on memory or when TSB_FORCEALLOC is
1179 	 * specified, don't use magazines to cache them--we want to return
1180 	 * them to the system as quickly as possible.
1181 	 */
1182 	sfmmu_tsb8k_cache = kmem_cache_create("sfmmu_tsb8k_cache",
1183 	    MMU_PAGESIZE, MMU_PAGESIZE, NULL, NULL, NULL, NULL,
1184 	    static_arena, KMC_NOMAGAZINE);
1185 
1186 	/*
1187 	 * Set tsb_alloc_hiwater to 1/tsb_alloc_hiwater_factor of physical
1188 	 * memory, which corresponds to the old static reserve for TSBs.
1189 	 * tsb_alloc_hiwater_factor defaults to 32.  This caps the amount of
1190 	 * memory we'll allocate for TSB slabs; beyond this point TSB
1191 	 * allocations will be taken from the kernel heap (via
1192 	 * sfmmu_tsb8k_cache) and will be throttled as would any other kmem
1193 	 * consumer.
1194 	 */
1195 	if (tsb_alloc_hiwater_factor == 0) {
1196 		tsb_alloc_hiwater_factor = TSB_ALLOC_HIWATER_FACTOR_DEFAULT;
1197 	}
1198 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
1199 
1200 	for (sz = tsb_slab_ttesz; sz > 0; sz--) {
1201 		if (!(disable_large_pages & (1 << sz)))
1202 			break;
1203 	}
1204 
1205 	if (sz < tsb_slab_ttesz) {
1206 		tsb_slab_ttesz = sz;
1207 		tsb_slab_shift = MMU_PAGESHIFT + (sz << 1) + sz;
1208 		tsb_slab_size = 1 << tsb_slab_shift;
1209 		tsb_slab_mask = (1 << (tsb_slab_shift - MMU_PAGESHIFT)) - 1;
1210 		use_bigtsb_arena = 0;
1211 	} else if (use_bigtsb_arena &&
1212 	    (disable_large_pages & (1 << bigtsb_slab_ttesz))) {
1213 		use_bigtsb_arena = 0;
1214 	}
1215 
1216 	if (!use_bigtsb_arena) {
1217 		bigtsb_slab_shift = tsb_slab_shift;
1218 	}
1219 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
1220 
1221 	/*
1222 	 * On smaller memory systems, allocate TSB memory in smaller chunks
1223 	 * than the default 4M slab size. We also honor disable_large_pages
1224 	 * here.
1225 	 *
1226 	 * The trap handlers need to be patched with the final slab shift,
1227 	 * since they need to be able to construct the TSB pointer at runtime.
1228 	 */
1229 	if ((tsb_max_growsize <= TSB_512K_SZCODE) &&
1230 	    !(disable_large_pages & (1 << TTE512K))) {
1231 		tsb_slab_ttesz = TTE512K;
1232 		tsb_slab_shift = MMU_PAGESHIFT512K;
1233 		tsb_slab_size = MMU_PAGESIZE512K;
1234 		tsb_slab_mask = MMU_PAGEOFFSET512K >> MMU_PAGESHIFT;
1235 		use_bigtsb_arena = 0;
1236 	}
1237 
1238 	if (!use_bigtsb_arena) {
1239 		bigtsb_slab_ttesz = tsb_slab_ttesz;
1240 		bigtsb_slab_shift = tsb_slab_shift;
1241 		bigtsb_slab_size = tsb_slab_size;
1242 		bigtsb_slab_mask = tsb_slab_mask;
1243 	}
1244 
1245 
1246 	/*
1247 	 * Set up memory callback to update tsb_alloc_hiwater and
1248 	 * tsb_max_growsize.
1249 	 */
1250 	i = kphysm_setup_func_register(&sfmmu_update_vec, (void *) 0);
1251 	ASSERT(i == 0);
1252 
1253 	/*
1254 	 * kmem_tsb_arena is the source from which large TSB slabs are
1255 	 * drawn.  The quantum of this arena corresponds to the largest
1256 	 * TSB size we can dynamically allocate for user processes.
1257 	 * Currently it must also be a supported page size since we
1258 	 * use exactly one translation entry to map each slab page.
1259 	 *
1260 	 * The per-lgroup kmem_tsb_default_arena arenas are the arenas from
1261 	 * which most TSBs are allocated.  Since most TSB allocations are
1262 	 * typically 8K we have a kmem cache we stack on top of each
1263 	 * kmem_tsb_default_arena to speed up those allocations.
1264 	 *
1265 	 * Note the two-level scheme of arenas is required only
1266 	 * because vmem_create doesn't allow us to specify alignment
1267 	 * requirements.  If this ever changes the code could be
1268 	 * simplified to use only one level of arenas.
1269 	 *
1270 	 * If 256M page support exists on sun4v, 256MB kmem_bigtsb_arena
1271 	 * will be provided in addition to the 4M kmem_tsb_arena.
1272 	 */
1273 	if (use_bigtsb_arena) {
1274 		kmem_bigtsb_arena = vmem_create("kmem_bigtsb", NULL, 0,
1275 		    bigtsb_slab_size, sfmmu_vmem_xalloc_aligned_wrapper,
1276 		    vmem_xfree, heap_arena, 0, VM_SLEEP);
1277 	}
1278 
1279 	kmem_tsb_arena = vmem_create("kmem_tsb", NULL, 0, tsb_slab_size,
1280 	    sfmmu_vmem_xalloc_aligned_wrapper,
1281 	    vmem_xfree, heap_arena, 0, VM_SLEEP);
1282 
1283 	if (tsb_lgrp_affinity) {
1284 		char s[50];
1285 		for (i = 0; i < NLGRPS_MAX; i++) {
1286 			if (use_bigtsb_arena) {
1287 				(void) sprintf(s, "kmem_bigtsb_lgrp%d", i);
1288 				kmem_bigtsb_default_arena[i] = vmem_create(s,
1289 				    NULL, 0, 2 * tsb_slab_size,
1290 				    sfmmu_tsb_segkmem_alloc,
1291 				    sfmmu_tsb_segkmem_free, kmem_bigtsb_arena,
1292 				    0, VM_SLEEP | VM_BESTFIT);
1293 			}
1294 
1295 			(void) sprintf(s, "kmem_tsb_lgrp%d", i);
1296 			kmem_tsb_default_arena[i] = vmem_create(s,
1297 			    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1298 			    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1299 			    VM_SLEEP | VM_BESTFIT);
1300 
1301 			(void) sprintf(s, "sfmmu_tsb_lgrp%d_cache", i);
1302 			sfmmu_tsb_cache[i] = kmem_cache_create(s,
1303 			    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1304 			    kmem_tsb_default_arena[i], 0);
1305 		}
1306 	} else {
1307 		if (use_bigtsb_arena) {
1308 			kmem_bigtsb_default_arena[0] =
1309 			    vmem_create("kmem_bigtsb_default", NULL, 0,
1310 			    2 * tsb_slab_size, sfmmu_tsb_segkmem_alloc,
1311 			    sfmmu_tsb_segkmem_free, kmem_bigtsb_arena, 0,
1312 			    VM_SLEEP | VM_BESTFIT);
1313 		}
1314 
1315 		kmem_tsb_default_arena[0] = vmem_create("kmem_tsb_default",
1316 		    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1317 		    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1318 		    VM_SLEEP | VM_BESTFIT);
1319 		sfmmu_tsb_cache[0] = kmem_cache_create("sfmmu_tsb_cache",
1320 		    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1321 		    kmem_tsb_default_arena[0], 0);
1322 	}
1323 
1324 	sfmmu8_cache = kmem_cache_create("sfmmu8_cache", HME8BLK_SZ,
1325 	    HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1326 	    sfmmu_hblkcache_destructor,
1327 	    sfmmu_hblkcache_reclaim, (void *)HME8BLK_SZ,
1328 	    hat_memload_arena, KMC_NOHASH);
1329 
1330 	hat_memload1_arena = vmem_create("hat_memload1", NULL, 0, PAGESIZE,
1331 	    segkmem_alloc_permanent, segkmem_free, heap_arena, 0, VM_SLEEP);
1332 
1333 	sfmmu1_cache = kmem_cache_create("sfmmu1_cache", HME1BLK_SZ,
1334 	    HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1335 	    sfmmu_hblkcache_destructor,
1336 	    NULL, (void *)HME1BLK_SZ,
1337 	    hat_memload1_arena, KMC_NOHASH);
1338 
1339 	pa_hment_cache = kmem_cache_create("pa_hment_cache", PAHME_SZ,
1340 	    0, NULL, NULL, NULL, NULL, static_arena, KMC_NOHASH);
1341 
1342 	ism_blk_cache = kmem_cache_create("ism_blk_cache",
1343 	    sizeof (ism_blk_t), ecache_alignsize, NULL, NULL,
1344 	    NULL, NULL, static_arena, KMC_NOHASH);
1345 
1346 	ism_ment_cache = kmem_cache_create("ism_ment_cache",
1347 	    sizeof (ism_ment_t), 0, NULL, NULL,
1348 	    NULL, NULL, NULL, 0);
1349 
1350 	/*
1351 	 * We grab the first hat for the kernel,
1352 	 */
1353 	AS_LOCK_ENTER(&kas, &kas.a_lock, RW_WRITER);
1354 	kas.a_hat = hat_alloc(&kas);
1355 	AS_LOCK_EXIT(&kas, &kas.a_lock);
1356 
1357 	/*
1358 	 * Initialize hblk_reserve.
1359 	 */
1360 	((struct hme_blk *)hblk_reserve)->hblk_nextpa =
1361 	    va_to_pa((caddr_t)hblk_reserve);
1362 
1363 #ifndef UTSB_PHYS
1364 	/*
1365 	 * Reserve some kernel virtual address space for the locked TTEs
1366 	 * that allow us to probe the TSB from TL>0.
1367 	 */
1368 	utsb_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1369 	    0, 0, NULL, NULL, VM_SLEEP);
1370 	utsb4m_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1371 	    0, 0, NULL, NULL, VM_SLEEP);
1372 #endif
1373 
1374 #ifdef VAC
1375 	/*
1376 	 * The big page VAC handling code assumes VAC
1377 	 * will not be bigger than the smallest big
1378 	 * page- which is 64K.
1379 	 */
1380 	if (TTEPAGES(TTE64K) < CACHE_NUM_COLOR) {
1381 		cmn_err(CE_PANIC, "VAC too big!");
1382 	}
1383 #endif
1384 
1385 	(void) xhat_init();
1386 
1387 	uhme_hash_pa = va_to_pa(uhme_hash);
1388 	khme_hash_pa = va_to_pa(khme_hash);
1389 
1390 	/*
1391 	 * Initialize relocation locks. kpr_suspendlock is held
1392 	 * at PIL_MAX to prevent interrupts from pinning the holder
1393 	 * of a suspended TTE which may access it leading to a
1394 	 * deadlock condition.
1395 	 */
1396 	mutex_init(&kpr_mutex, NULL, MUTEX_DEFAULT, NULL);
1397 	mutex_init(&kpr_suspendlock, NULL, MUTEX_SPIN, (void *)PIL_MAX);
1398 
1399 	/*
1400 	 * If Shared context support is disabled via /etc/system
1401 	 * set shctx_on to 0 here if it was set to 1 earlier in boot
1402 	 * sequence by cpu module initialization code.
1403 	 */
1404 	if (shctx_on && disable_shctx) {
1405 		shctx_on = 0;
1406 	}
1407 
1408 	if (shctx_on) {
1409 		srd_buckets = kmem_zalloc(SFMMU_MAX_SRD_BUCKETS *
1410 		    sizeof (srd_buckets[0]), KM_SLEEP);
1411 		for (i = 0; i < SFMMU_MAX_SRD_BUCKETS; i++) {
1412 			mutex_init(&srd_buckets[i].srdb_lock, NULL,
1413 			    MUTEX_DEFAULT, NULL);
1414 		}
1415 
1416 		srd_cache = kmem_cache_create("srd_cache", sizeof (sf_srd_t),
1417 		    0, sfmmu_srdcache_constructor, sfmmu_srdcache_destructor,
1418 		    NULL, NULL, NULL, 0);
1419 		region_cache = kmem_cache_create("region_cache",
1420 		    sizeof (sf_region_t), 0, sfmmu_rgncache_constructor,
1421 		    sfmmu_rgncache_destructor, NULL, NULL, NULL, 0);
1422 		scd_cache = kmem_cache_create("scd_cache", sizeof (sf_scd_t),
1423 		    0, sfmmu_scdcache_constructor,  sfmmu_scdcache_destructor,
1424 		    NULL, NULL, NULL, 0);
1425 	}
1426 
1427 	/*
1428 	 * Pre-allocate hrm_hashtab before enabling the collection of
1429 	 * refmod statistics.  Allocating on the fly would mean us
1430 	 * running the risk of suffering recursive mutex enters or
1431 	 * deadlocks.
1432 	 */
1433 	hrm_hashtab = kmem_zalloc(HRM_HASHSIZE * sizeof (struct hrmstat *),
1434 	    KM_SLEEP);
1435 
1436 	/* Allocate per-cpu pending freelist of hmeblks */
1437 	cpu_hme_pend = kmem_zalloc((NCPU * sizeof (cpu_hme_pend_t)) + 64,
1438 	    KM_SLEEP);
1439 	cpu_hme_pend = (cpu_hme_pend_t *)P2ROUNDUP(
1440 	    (uintptr_t)cpu_hme_pend, 64);
1441 
1442 	for (i = 0; i < NCPU; i++) {
1443 		mutex_init(&cpu_hme_pend[i].chp_mutex, NULL, MUTEX_DEFAULT,
1444 		    NULL);
1445 	}
1446 
1447 	if (cpu_hme_pend_thresh == 0) {
1448 		cpu_hme_pend_thresh = CPU_HME_PEND_THRESH;
1449 	}
1450 }
1451 
1452 /*
1453  * Initialize locking for the hat layer, called early during boot.
1454  */
1455 static void
1456 hat_lock_init()
1457 {
1458 	int i;
1459 
1460 	/*
1461 	 * initialize the array of mutexes protecting a page's mapping
1462 	 * list and p_nrm field.
1463 	 */
1464 	for (i = 0; i < mml_table_sz; i++)
1465 		mutex_init(&mml_table[i], NULL, MUTEX_DEFAULT, NULL);
1466 
1467 	if (kpm_enable) {
1468 		for (i = 0; i < kpmp_table_sz; i++) {
1469 			mutex_init(&kpmp_table[i].khl_mutex, NULL,
1470 			    MUTEX_DEFAULT, NULL);
1471 		}
1472 	}
1473 
1474 	/*
1475 	 * Initialize array of mutex locks that protects sfmmu fields and
1476 	 * TSB lists.
1477 	 */
1478 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
1479 		mutex_init(HATLOCK_MUTEXP(&hat_lock[i]), NULL, MUTEX_DEFAULT,
1480 		    NULL);
1481 }
1482 
1483 #define	SFMMU_KERNEL_MAXVA \
1484 	(kmem64_base ? (uintptr_t)kmem64_end : (SYSLIMIT))
1485 
1486 /*
1487  * Allocate a hat structure.
1488  * Called when an address space first uses a hat.
1489  */
1490 struct hat *
1491 hat_alloc(struct as *as)
1492 {
1493 	sfmmu_t *sfmmup;
1494 	int i;
1495 	uint64_t cnum;
1496 	extern uint_t get_color_start(struct as *);
1497 
1498 	ASSERT(AS_WRITE_HELD(as, &as->a_lock));
1499 	sfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
1500 	sfmmup->sfmmu_as = as;
1501 	sfmmup->sfmmu_flags = 0;
1502 	sfmmup->sfmmu_tteflags = 0;
1503 	sfmmup->sfmmu_rtteflags = 0;
1504 	LOCK_INIT_CLEAR(&sfmmup->sfmmu_ctx_lock);
1505 
1506 	if (as == &kas) {
1507 		ksfmmup = sfmmup;
1508 		sfmmup->sfmmu_cext = 0;
1509 		cnum = KCONTEXT;
1510 
1511 		sfmmup->sfmmu_clrstart = 0;
1512 		sfmmup->sfmmu_tsb = NULL;
1513 		/*
1514 		 * hat_kern_setup() will call sfmmu_init_ktsbinfo()
1515 		 * to setup tsb_info for ksfmmup.
1516 		 */
1517 	} else {
1518 
1519 		/*
1520 		 * Just set to invalid ctx. When it faults, it will
1521 		 * get a valid ctx. This would avoid the situation
1522 		 * where we get a ctx, but it gets stolen and then
1523 		 * we fault when we try to run and so have to get
1524 		 * another ctx.
1525 		 */
1526 		sfmmup->sfmmu_cext = 0;
1527 		cnum = INVALID_CONTEXT;
1528 
1529 		/* initialize original physical page coloring bin */
1530 		sfmmup->sfmmu_clrstart = get_color_start(as);
1531 #ifdef DEBUG
1532 		if (tsb_random_size) {
1533 			uint32_t randval = (uint32_t)gettick() >> 4;
1534 			int size = randval % (tsb_max_growsize + 1);
1535 
1536 			/* chose a random tsb size for stress testing */
1537 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb, size,
1538 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1539 		} else
1540 #endif /* DEBUG */
1541 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb,
1542 			    default_tsb_size,
1543 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1544 		sfmmup->sfmmu_flags = HAT_SWAPPED | HAT_ALLCTX_INVALID;
1545 		ASSERT(sfmmup->sfmmu_tsb != NULL);
1546 	}
1547 
1548 	ASSERT(max_mmu_ctxdoms > 0);
1549 	for (i = 0; i < max_mmu_ctxdoms; i++) {
1550 		sfmmup->sfmmu_ctxs[i].cnum = cnum;
1551 		sfmmup->sfmmu_ctxs[i].gnum = 0;
1552 	}
1553 
1554 	for (i = 0; i < max_mmu_page_sizes; i++) {
1555 		sfmmup->sfmmu_ttecnt[i] = 0;
1556 		sfmmup->sfmmu_scdrttecnt[i] = 0;
1557 		sfmmup->sfmmu_ismttecnt[i] = 0;
1558 		sfmmup->sfmmu_scdismttecnt[i] = 0;
1559 		sfmmup->sfmmu_pgsz[i] = TTE8K;
1560 	}
1561 	sfmmup->sfmmu_tsb0_4minflcnt = 0;
1562 	sfmmup->sfmmu_iblk = NULL;
1563 	sfmmup->sfmmu_ismhat = 0;
1564 	sfmmup->sfmmu_scdhat = 0;
1565 	sfmmup->sfmmu_ismblkpa = (uint64_t)-1;
1566 	if (sfmmup == ksfmmup) {
1567 		CPUSET_ALL(sfmmup->sfmmu_cpusran);
1568 	} else {
1569 		CPUSET_ZERO(sfmmup->sfmmu_cpusran);
1570 	}
1571 	sfmmup->sfmmu_free = 0;
1572 	sfmmup->sfmmu_rmstat = 0;
1573 	sfmmup->sfmmu_clrbin = sfmmup->sfmmu_clrstart;
1574 	sfmmup->sfmmu_xhat_provider = NULL;
1575 	cv_init(&sfmmup->sfmmu_tsb_cv, NULL, CV_DEFAULT, NULL);
1576 	sfmmup->sfmmu_srdp = NULL;
1577 	SF_RGNMAP_ZERO(sfmmup->sfmmu_region_map);
1578 	bzero(sfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
1579 	sfmmup->sfmmu_scdp = NULL;
1580 	sfmmup->sfmmu_scd_link.next = NULL;
1581 	sfmmup->sfmmu_scd_link.prev = NULL;
1582 	return (sfmmup);
1583 }
1584 
1585 /*
1586  * Create per-MMU context domain kstats for a given MMU ctx.
1587  */
1588 static void
1589 sfmmu_mmu_kstat_create(mmu_ctx_t *mmu_ctxp)
1590 {
1591 	mmu_ctx_stat_t	stat;
1592 	kstat_t		*mmu_kstat;
1593 
1594 	ASSERT(MUTEX_HELD(&cpu_lock));
1595 	ASSERT(mmu_ctxp->mmu_kstat == NULL);
1596 
1597 	mmu_kstat = kstat_create("unix", mmu_ctxp->mmu_idx, "mmu_ctx",
1598 	    "hat", KSTAT_TYPE_NAMED, MMU_CTX_NUM_STATS, KSTAT_FLAG_VIRTUAL);
1599 
1600 	if (mmu_kstat == NULL) {
1601 		cmn_err(CE_WARN, "kstat_create for MMU %d failed",
1602 		    mmu_ctxp->mmu_idx);
1603 	} else {
1604 		mmu_kstat->ks_data = mmu_ctxp->mmu_kstat_data;
1605 		for (stat = 0; stat < MMU_CTX_NUM_STATS; stat++)
1606 			kstat_named_init(&mmu_ctxp->mmu_kstat_data[stat],
1607 			    mmu_ctx_kstat_names[stat], KSTAT_DATA_INT64);
1608 		mmu_ctxp->mmu_kstat = mmu_kstat;
1609 		kstat_install(mmu_kstat);
1610 	}
1611 }
1612 
1613 /*
1614  * plat_cpuid_to_mmu_ctx_info() is a platform interface that returns MMU
1615  * context domain information for a given CPU. If a platform does not
1616  * specify that interface, then the function below is used instead to return
1617  * default information. The defaults are as follows:
1618  *
1619  *	- For sun4u systems there's one MMU context domain per CPU.
1620  *	  This default is used by all sun4u systems except OPL. OPL systems
1621  *	  provide platform specific interface to map CPU ids to MMU ids
1622  *	  because on OPL more than 1 CPU shares a single MMU.
1623  *        Note that on sun4v, there is one global context domain for
1624  *	  the entire system. This is to avoid running into potential problem
1625  *	  with ldom physical cpu substitution feature.
1626  *	- The number of MMU context IDs supported on any CPU in the
1627  *	  system is 8K.
1628  */
1629 /*ARGSUSED*/
1630 static void
1631 sfmmu_cpuid_to_mmu_ctx_info(processorid_t cpuid, mmu_ctx_info_t *infop)
1632 {
1633 	infop->mmu_nctxs = nctxs;
1634 #ifndef sun4v
1635 	infop->mmu_idx = cpu[cpuid]->cpu_seqid;
1636 #else /* sun4v */
1637 	infop->mmu_idx = 0;
1638 #endif /* sun4v */
1639 }
1640 
1641 /*
1642  * Called during CPU initialization to set the MMU context-related information
1643  * for a CPU.
1644  *
1645  * cpu_lock serializes accesses to mmu_ctxs and mmu_saved_gnum.
1646  */
1647 void
1648 sfmmu_cpu_init(cpu_t *cp)
1649 {
1650 	mmu_ctx_info_t	info;
1651 	mmu_ctx_t	*mmu_ctxp;
1652 
1653 	ASSERT(MUTEX_HELD(&cpu_lock));
1654 
1655 	if (&plat_cpuid_to_mmu_ctx_info == NULL)
1656 		sfmmu_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1657 	else
1658 		plat_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1659 
1660 	ASSERT(info.mmu_idx < max_mmu_ctxdoms);
1661 
1662 	if ((mmu_ctxp = mmu_ctxs_tbl[info.mmu_idx]) == NULL) {
1663 		/* Each mmu_ctx is cacheline aligned. */
1664 		mmu_ctxp = kmem_cache_alloc(mmuctxdom_cache, KM_SLEEP);
1665 		bzero(mmu_ctxp, sizeof (mmu_ctx_t));
1666 
1667 		mutex_init(&mmu_ctxp->mmu_lock, NULL, MUTEX_SPIN,
1668 		    (void *)ipltospl(DISP_LEVEL));
1669 		mmu_ctxp->mmu_idx = info.mmu_idx;
1670 		mmu_ctxp->mmu_nctxs = info.mmu_nctxs;
1671 		/*
1672 		 * Globally for lifetime of a system,
1673 		 * gnum must always increase.
1674 		 * mmu_saved_gnum is protected by the cpu_lock.
1675 		 */
1676 		mmu_ctxp->mmu_gnum = mmu_saved_gnum + 1;
1677 		mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
1678 
1679 		sfmmu_mmu_kstat_create(mmu_ctxp);
1680 
1681 		mmu_ctxs_tbl[info.mmu_idx] = mmu_ctxp;
1682 	} else {
1683 		ASSERT(mmu_ctxp->mmu_idx == info.mmu_idx);
1684 	}
1685 
1686 	/*
1687 	 * The mmu_lock is acquired here to prevent races with
1688 	 * the wrap-around code.
1689 	 */
1690 	mutex_enter(&mmu_ctxp->mmu_lock);
1691 
1692 
1693 	mmu_ctxp->mmu_ncpus++;
1694 	CPUSET_ADD(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1695 	CPU_MMU_IDX(cp) = info.mmu_idx;
1696 	CPU_MMU_CTXP(cp) = mmu_ctxp;
1697 
1698 	mutex_exit(&mmu_ctxp->mmu_lock);
1699 }
1700 
1701 /*
1702  * Called to perform MMU context-related cleanup for a CPU.
1703  */
1704 void
1705 sfmmu_cpu_cleanup(cpu_t *cp)
1706 {
1707 	mmu_ctx_t	*mmu_ctxp;
1708 
1709 	ASSERT(MUTEX_HELD(&cpu_lock));
1710 
1711 	mmu_ctxp = CPU_MMU_CTXP(cp);
1712 	ASSERT(mmu_ctxp != NULL);
1713 
1714 	/*
1715 	 * The mmu_lock is acquired here to prevent races with
1716 	 * the wrap-around code.
1717 	 */
1718 	mutex_enter(&mmu_ctxp->mmu_lock);
1719 
1720 	CPU_MMU_CTXP(cp) = NULL;
1721 
1722 	CPUSET_DEL(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1723 	if (--mmu_ctxp->mmu_ncpus == 0) {
1724 		mmu_ctxs_tbl[mmu_ctxp->mmu_idx] = NULL;
1725 		mutex_exit(&mmu_ctxp->mmu_lock);
1726 		mutex_destroy(&mmu_ctxp->mmu_lock);
1727 
1728 		if (mmu_ctxp->mmu_kstat)
1729 			kstat_delete(mmu_ctxp->mmu_kstat);
1730 
1731 		/* mmu_saved_gnum is protected by the cpu_lock. */
1732 		if (mmu_saved_gnum < mmu_ctxp->mmu_gnum)
1733 			mmu_saved_gnum = mmu_ctxp->mmu_gnum;
1734 
1735 		kmem_cache_free(mmuctxdom_cache, mmu_ctxp);
1736 
1737 		return;
1738 	}
1739 
1740 	mutex_exit(&mmu_ctxp->mmu_lock);
1741 }
1742 
1743 /*
1744  * Hat_setup, makes an address space context the current active one.
1745  * In sfmmu this translates to setting the secondary context with the
1746  * corresponding context.
1747  */
1748 void
1749 hat_setup(struct hat *sfmmup, int allocflag)
1750 {
1751 	hatlock_t *hatlockp;
1752 
1753 	/* Init needs some special treatment. */
1754 	if (allocflag == HAT_INIT) {
1755 		/*
1756 		 * Make sure that we have
1757 		 * 1. a TSB
1758 		 * 2. a valid ctx that doesn't get stolen after this point.
1759 		 */
1760 		hatlockp = sfmmu_hat_enter(sfmmup);
1761 
1762 		/*
1763 		 * Swap in the TSB.  hat_init() allocates tsbinfos without
1764 		 * TSBs, but we need one for init, since the kernel does some
1765 		 * special things to set up its stack and needs the TSB to
1766 		 * resolve page faults.
1767 		 */
1768 		sfmmu_tsb_swapin(sfmmup, hatlockp);
1769 
1770 		sfmmu_get_ctx(sfmmup);
1771 
1772 		sfmmu_hat_exit(hatlockp);
1773 	} else {
1774 		ASSERT(allocflag == HAT_ALLOC);
1775 
1776 		hatlockp = sfmmu_hat_enter(sfmmup);
1777 		kpreempt_disable();
1778 
1779 		CPUSET_ADD(sfmmup->sfmmu_cpusran, CPU->cpu_id);
1780 		/*
1781 		 * sfmmu_setctx_sec takes <pgsz|cnum> as a parameter,
1782 		 * pagesize bits don't matter in this case since we are passing
1783 		 * INVALID_CONTEXT to it.
1784 		 * Compatibility Note: hw takes care of MMU_SCONTEXT1
1785 		 */
1786 		sfmmu_setctx_sec(INVALID_CONTEXT);
1787 		sfmmu_clear_utsbinfo();
1788 
1789 		kpreempt_enable();
1790 		sfmmu_hat_exit(hatlockp);
1791 	}
1792 }
1793 
1794 /*
1795  * Free all the translation resources for the specified address space.
1796  * Called from as_free when an address space is being destroyed.
1797  */
1798 void
1799 hat_free_start(struct hat *sfmmup)
1800 {
1801 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
1802 	ASSERT(sfmmup != ksfmmup);
1803 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1804 
1805 	sfmmup->sfmmu_free = 1;
1806 	if (sfmmup->sfmmu_scdp != NULL) {
1807 		sfmmu_leave_scd(sfmmup, 0);
1808 	}
1809 
1810 	ASSERT(sfmmup->sfmmu_scdp == NULL);
1811 }
1812 
1813 void
1814 hat_free_end(struct hat *sfmmup)
1815 {
1816 	int i;
1817 
1818 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1819 	ASSERT(sfmmup->sfmmu_free == 1);
1820 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
1821 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
1822 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
1823 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
1824 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
1825 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
1826 
1827 	if (sfmmup->sfmmu_rmstat) {
1828 		hat_freestat(sfmmup->sfmmu_as, NULL);
1829 	}
1830 
1831 	while (sfmmup->sfmmu_tsb != NULL) {
1832 		struct tsb_info *next = sfmmup->sfmmu_tsb->tsb_next;
1833 		sfmmu_tsbinfo_free(sfmmup->sfmmu_tsb);
1834 		sfmmup->sfmmu_tsb = next;
1835 	}
1836 
1837 	if (sfmmup->sfmmu_srdp != NULL) {
1838 		sfmmu_leave_srd(sfmmup);
1839 		ASSERT(sfmmup->sfmmu_srdp == NULL);
1840 		for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1841 			if (sfmmup->sfmmu_hmeregion_links[i] != NULL) {
1842 				kmem_free(sfmmup->sfmmu_hmeregion_links[i],
1843 				    SFMMU_L2_HMERLINKS_SIZE);
1844 				sfmmup->sfmmu_hmeregion_links[i] = NULL;
1845 			}
1846 		}
1847 	}
1848 	sfmmu_free_sfmmu(sfmmup);
1849 
1850 #ifdef DEBUG
1851 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1852 		ASSERT(sfmmup->sfmmu_hmeregion_links[i] == NULL);
1853 	}
1854 #endif
1855 
1856 	kmem_cache_free(sfmmuid_cache, sfmmup);
1857 }
1858 
1859 /*
1860  * Set up any translation structures, for the specified address space,
1861  * that are needed or preferred when the process is being swapped in.
1862  */
1863 /* ARGSUSED */
1864 void
1865 hat_swapin(struct hat *hat)
1866 {
1867 	ASSERT(hat->sfmmu_xhat_provider == NULL);
1868 }
1869 
1870 /*
1871  * Free all of the translation resources, for the specified address space,
1872  * that can be freed while the process is swapped out. Called from as_swapout.
1873  * Also, free up the ctx that this process was using.
1874  */
1875 void
1876 hat_swapout(struct hat *sfmmup)
1877 {
1878 	struct hmehash_bucket *hmebp;
1879 	struct hme_blk *hmeblkp;
1880 	struct hme_blk *pr_hblk = NULL;
1881 	struct hme_blk *nx_hblk;
1882 	int i;
1883 	struct hme_blk *list = NULL;
1884 	hatlock_t *hatlockp;
1885 	struct tsb_info *tsbinfop;
1886 	struct free_tsb {
1887 		struct free_tsb *next;
1888 		struct tsb_info *tsbinfop;
1889 	};			/* free list of TSBs */
1890 	struct free_tsb *freelist, *last, *next;
1891 
1892 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1893 	SFMMU_STAT(sf_swapout);
1894 
1895 	/*
1896 	 * There is no way to go from an as to all its translations in sfmmu.
1897 	 * Here is one of the times when we take the big hit and traverse
1898 	 * the hash looking for hme_blks to free up.  Not only do we free up
1899 	 * this as hme_blks but all those that are free.  We are obviously
1900 	 * swapping because we need memory so let's free up as much
1901 	 * as we can.
1902 	 *
1903 	 * Note that we don't flush TLB/TSB here -- it's not necessary
1904 	 * because:
1905 	 *  1) we free the ctx we're using and throw away the TSB(s);
1906 	 *  2) processes aren't runnable while being swapped out.
1907 	 */
1908 	ASSERT(sfmmup != KHATID);
1909 	for (i = 0; i <= UHMEHASH_SZ; i++) {
1910 		hmebp = &uhme_hash[i];
1911 		SFMMU_HASH_LOCK(hmebp);
1912 		hmeblkp = hmebp->hmeblkp;
1913 		pr_hblk = NULL;
1914 		while (hmeblkp) {
1915 
1916 			ASSERT(!hmeblkp->hblk_xhat_bit);
1917 
1918 			if ((hmeblkp->hblk_tag.htag_id == sfmmup) &&
1919 			    !hmeblkp->hblk_shw_bit && !hmeblkp->hblk_lckcnt) {
1920 				ASSERT(!hmeblkp->hblk_shared);
1921 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
1922 				    (caddr_t)get_hblk_base(hmeblkp),
1923 				    get_hblk_endaddr(hmeblkp),
1924 				    NULL, HAT_UNLOAD);
1925 			}
1926 			nx_hblk = hmeblkp->hblk_next;
1927 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
1928 				ASSERT(!hmeblkp->hblk_lckcnt);
1929 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
1930 				    &list, 0);
1931 			} else {
1932 				pr_hblk = hmeblkp;
1933 			}
1934 			hmeblkp = nx_hblk;
1935 		}
1936 		SFMMU_HASH_UNLOCK(hmebp);
1937 	}
1938 
1939 	sfmmu_hblks_list_purge(&list, 0);
1940 
1941 	/*
1942 	 * Now free up the ctx so that others can reuse it.
1943 	 */
1944 	hatlockp = sfmmu_hat_enter(sfmmup);
1945 
1946 	sfmmu_invalidate_ctx(sfmmup);
1947 
1948 	/*
1949 	 * Free TSBs, but not tsbinfos, and set SWAPPED flag.
1950 	 * If TSBs were never swapped in, just return.
1951 	 * This implies that we don't support partial swapping
1952 	 * of TSBs -- either all are swapped out, or none are.
1953 	 *
1954 	 * We must hold the HAT lock here to prevent racing with another
1955 	 * thread trying to unmap TTEs from the TSB or running the post-
1956 	 * relocator after relocating the TSB's memory.  Unfortunately, we
1957 	 * can't free memory while holding the HAT lock or we could
1958 	 * deadlock, so we build a list of TSBs to be freed after marking
1959 	 * the tsbinfos as swapped out and free them after dropping the
1960 	 * lock.
1961 	 */
1962 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
1963 		sfmmu_hat_exit(hatlockp);
1964 		return;
1965 	}
1966 
1967 	SFMMU_FLAGS_SET(sfmmup, HAT_SWAPPED);
1968 	last = freelist = NULL;
1969 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
1970 	    tsbinfop = tsbinfop->tsb_next) {
1971 		ASSERT((tsbinfop->tsb_flags & TSB_SWAPPED) == 0);
1972 
1973 		/*
1974 		 * Cast the TSB into a struct free_tsb and put it on the free
1975 		 * list.
1976 		 */
1977 		if (freelist == NULL) {
1978 			last = freelist = (struct free_tsb *)tsbinfop->tsb_va;
1979 		} else {
1980 			last->next = (struct free_tsb *)tsbinfop->tsb_va;
1981 			last = last->next;
1982 		}
1983 		last->next = NULL;
1984 		last->tsbinfop = tsbinfop;
1985 		tsbinfop->tsb_flags |= TSB_SWAPPED;
1986 		/*
1987 		 * Zero out the TTE to clear the valid bit.
1988 		 * Note we can't use a value like 0xbad because we want to
1989 		 * ensure diagnostic bits are NEVER set on TTEs that might
1990 		 * be loaded.  The intent is to catch any invalid access
1991 		 * to the swapped TSB, such as a thread running with a valid
1992 		 * context without first calling sfmmu_tsb_swapin() to
1993 		 * allocate TSB memory.
1994 		 */
1995 		tsbinfop->tsb_tte.ll = 0;
1996 	}
1997 
1998 	/* Now we can drop the lock and free the TSB memory. */
1999 	sfmmu_hat_exit(hatlockp);
2000 	for (; freelist != NULL; freelist = next) {
2001 		next = freelist->next;
2002 		sfmmu_tsb_free(freelist->tsbinfop);
2003 	}
2004 }
2005 
2006 /*
2007  * Duplicate the translations of an as into another newas
2008  */
2009 /* ARGSUSED */
2010 int
2011 hat_dup(struct hat *hat, struct hat *newhat, caddr_t addr, size_t len,
2012 	uint_t flag)
2013 {
2014 	sf_srd_t *srdp;
2015 	sf_scd_t *scdp;
2016 	int i;
2017 	extern uint_t get_color_start(struct as *);
2018 
2019 	ASSERT(hat->sfmmu_xhat_provider == NULL);
2020 	ASSERT((flag == 0) || (flag == HAT_DUP_ALL) || (flag == HAT_DUP_COW) ||
2021 	    (flag == HAT_DUP_SRD));
2022 	ASSERT(hat != ksfmmup);
2023 	ASSERT(newhat != ksfmmup);
2024 	ASSERT(flag != HAT_DUP_ALL || hat->sfmmu_srdp == newhat->sfmmu_srdp);
2025 
2026 	if (flag == HAT_DUP_COW) {
2027 		panic("hat_dup: HAT_DUP_COW not supported");
2028 	}
2029 
2030 	if (flag == HAT_DUP_SRD && ((srdp = hat->sfmmu_srdp) != NULL)) {
2031 		ASSERT(srdp->srd_evp != NULL);
2032 		VN_HOLD(srdp->srd_evp);
2033 		ASSERT(srdp->srd_refcnt > 0);
2034 		newhat->sfmmu_srdp = srdp;
2035 		atomic_add_32((volatile uint_t *)&srdp->srd_refcnt, 1);
2036 	}
2037 
2038 	/*
2039 	 * HAT_DUP_ALL flag is used after as duplication is done.
2040 	 */
2041 	if (flag == HAT_DUP_ALL && ((srdp = newhat->sfmmu_srdp) != NULL)) {
2042 		ASSERT(newhat->sfmmu_srdp->srd_refcnt >= 2);
2043 		newhat->sfmmu_rtteflags = hat->sfmmu_rtteflags;
2044 		if (hat->sfmmu_flags & HAT_4MTEXT_FLAG) {
2045 			newhat->sfmmu_flags |= HAT_4MTEXT_FLAG;
2046 		}
2047 
2048 		/* check if need to join scd */
2049 		if ((scdp = hat->sfmmu_scdp) != NULL &&
2050 		    newhat->sfmmu_scdp != scdp) {
2051 			int ret;
2052 			SF_RGNMAP_IS_SUBSET(&newhat->sfmmu_region_map,
2053 			    &scdp->scd_region_map, ret);
2054 			ASSERT(ret);
2055 			sfmmu_join_scd(scdp, newhat);
2056 			ASSERT(newhat->sfmmu_scdp == scdp &&
2057 			    scdp->scd_refcnt >= 2);
2058 			for (i = 0; i < max_mmu_page_sizes; i++) {
2059 				newhat->sfmmu_ismttecnt[i] =
2060 				    hat->sfmmu_ismttecnt[i];
2061 				newhat->sfmmu_scdismttecnt[i] =
2062 				    hat->sfmmu_scdismttecnt[i];
2063 			}
2064 		}
2065 
2066 		sfmmu_check_page_sizes(newhat, 1);
2067 	}
2068 
2069 	if (flag == HAT_DUP_ALL && consistent_coloring == 0 &&
2070 	    update_proc_pgcolorbase_after_fork != 0) {
2071 		hat->sfmmu_clrbin = get_color_start(hat->sfmmu_as);
2072 	}
2073 	return (0);
2074 }
2075 
2076 void
2077 hat_memload(struct hat *hat, caddr_t addr, struct page *pp,
2078 	uint_t attr, uint_t flags)
2079 {
2080 	hat_do_memload(hat, addr, pp, attr, flags,
2081 	    SFMMU_INVALID_SHMERID);
2082 }
2083 
2084 void
2085 hat_memload_region(struct hat *hat, caddr_t addr, struct page *pp,
2086 	uint_t attr, uint_t flags, hat_region_cookie_t rcookie)
2087 {
2088 	uint_t rid;
2089 	if (rcookie == HAT_INVALID_REGION_COOKIE ||
2090 	    hat->sfmmu_xhat_provider != NULL) {
2091 		hat_do_memload(hat, addr, pp, attr, flags,
2092 		    SFMMU_INVALID_SHMERID);
2093 		return;
2094 	}
2095 	rid = (uint_t)((uint64_t)rcookie);
2096 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2097 	hat_do_memload(hat, addr, pp, attr, flags, rid);
2098 }
2099 
2100 /*
2101  * Set up addr to map to page pp with protection prot.
2102  * As an optimization we also load the TSB with the
2103  * corresponding tte but it is no big deal if  the tte gets kicked out.
2104  */
2105 static void
2106 hat_do_memload(struct hat *hat, caddr_t addr, struct page *pp,
2107 	uint_t attr, uint_t flags, uint_t rid)
2108 {
2109 	tte_t tte;
2110 
2111 
2112 	ASSERT(hat != NULL);
2113 	ASSERT(PAGE_LOCKED(pp));
2114 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2115 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2116 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2117 	SFMMU_VALIDATE_HMERID(hat, rid, addr, MMU_PAGESIZE);
2118 
2119 	if (PP_ISFREE(pp)) {
2120 		panic("hat_memload: loading a mapping to free page %p",
2121 		    (void *)pp);
2122 	}
2123 
2124 	if (hat->sfmmu_xhat_provider) {
2125 		/* no regions for xhats */
2126 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
2127 		XHAT_MEMLOAD(hat, addr, pp, attr, flags);
2128 		return;
2129 	}
2130 
2131 	ASSERT((hat == ksfmmup) ||
2132 	    AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock));
2133 
2134 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2135 		cmn_err(CE_NOTE, "hat_memload: unsupported flags %d",
2136 		    flags & ~SFMMU_LOAD_ALLFLAG);
2137 
2138 	if (hat->sfmmu_rmstat)
2139 		hat_resvstat(MMU_PAGESIZE, hat->sfmmu_as, addr);
2140 
2141 #if defined(SF_ERRATA_57)
2142 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2143 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2144 	    !(flags & HAT_LOAD_SHARE)) {
2145 		cmn_err(CE_WARN, "hat_memload: illegal attempt to make user "
2146 		    " page executable");
2147 		attr &= ~PROT_EXEC;
2148 	}
2149 #endif
2150 
2151 	sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2152 	(void) sfmmu_tteload_array(hat, &tte, addr, &pp, flags, rid);
2153 
2154 	/*
2155 	 * Check TSB and TLB page sizes.
2156 	 */
2157 	if ((flags & HAT_LOAD_SHARE) == 0) {
2158 		sfmmu_check_page_sizes(hat, 1);
2159 	}
2160 }
2161 
2162 /*
2163  * hat_devload can be called to map real memory (e.g.
2164  * /dev/kmem) and even though hat_devload will determine pf is
2165  * for memory, it will be unable to get a shared lock on the
2166  * page (because someone else has it exclusively) and will
2167  * pass dp = NULL.  If tteload doesn't get a non-NULL
2168  * page pointer it can't cache memory.
2169  */
2170 void
2171 hat_devload(struct hat *hat, caddr_t addr, size_t len, pfn_t pfn,
2172 	uint_t attr, int flags)
2173 {
2174 	tte_t tte;
2175 	struct page *pp = NULL;
2176 	int use_lgpg = 0;
2177 
2178 	ASSERT(hat != NULL);
2179 
2180 	if (hat->sfmmu_xhat_provider) {
2181 		XHAT_DEVLOAD(hat, addr, len, pfn, attr, flags);
2182 		return;
2183 	}
2184 
2185 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2186 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2187 	ASSERT((hat == ksfmmup) ||
2188 	    AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock));
2189 	if (len == 0)
2190 		panic("hat_devload: zero len");
2191 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2192 		cmn_err(CE_NOTE, "hat_devload: unsupported flags %d",
2193 		    flags & ~SFMMU_LOAD_ALLFLAG);
2194 
2195 #if defined(SF_ERRATA_57)
2196 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2197 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2198 	    !(flags & HAT_LOAD_SHARE)) {
2199 		cmn_err(CE_WARN, "hat_devload: illegal attempt to make user "
2200 		    " page executable");
2201 		attr &= ~PROT_EXEC;
2202 	}
2203 #endif
2204 
2205 	/*
2206 	 * If it's a memory page find its pp
2207 	 */
2208 	if (!(flags & HAT_LOAD_NOCONSIST) && pf_is_memory(pfn)) {
2209 		pp = page_numtopp_nolock(pfn);
2210 		if (pp == NULL) {
2211 			flags |= HAT_LOAD_NOCONSIST;
2212 		} else {
2213 			if (PP_ISFREE(pp)) {
2214 				panic("hat_memload: loading "
2215 				    "a mapping to free page %p",
2216 				    (void *)pp);
2217 			}
2218 			if (!PAGE_LOCKED(pp) && !PP_ISNORELOC(pp)) {
2219 				panic("hat_memload: loading a mapping "
2220 				    "to unlocked relocatable page %p",
2221 				    (void *)pp);
2222 			}
2223 			ASSERT(len == MMU_PAGESIZE);
2224 		}
2225 	}
2226 
2227 	if (hat->sfmmu_rmstat)
2228 		hat_resvstat(len, hat->sfmmu_as, addr);
2229 
2230 	if (flags & HAT_LOAD_NOCONSIST) {
2231 		attr |= SFMMU_UNCACHEVTTE;
2232 		use_lgpg = 1;
2233 	}
2234 	if (!pf_is_memory(pfn)) {
2235 		attr |= SFMMU_UNCACHEPTTE | HAT_NOSYNC;
2236 		use_lgpg = 1;
2237 		switch (attr & HAT_ORDER_MASK) {
2238 			case HAT_STRICTORDER:
2239 			case HAT_UNORDERED_OK:
2240 				/*
2241 				 * we set the side effect bit for all non
2242 				 * memory mappings unless merging is ok
2243 				 */
2244 				attr |= SFMMU_SIDEFFECT;
2245 				break;
2246 			case HAT_MERGING_OK:
2247 			case HAT_LOADCACHING_OK:
2248 			case HAT_STORECACHING_OK:
2249 				break;
2250 			default:
2251 				panic("hat_devload: bad attr");
2252 				break;
2253 		}
2254 	}
2255 	while (len) {
2256 		if (!use_lgpg) {
2257 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2258 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2259 			    flags, SFMMU_INVALID_SHMERID);
2260 			len -= MMU_PAGESIZE;
2261 			addr += MMU_PAGESIZE;
2262 			pfn++;
2263 			continue;
2264 		}
2265 		/*
2266 		 *  try to use large pages, check va/pa alignments
2267 		 *  Note that 32M/256M page sizes are not (yet) supported.
2268 		 */
2269 		if ((len >= MMU_PAGESIZE4M) &&
2270 		    !((uintptr_t)addr & MMU_PAGEOFFSET4M) &&
2271 		    !(disable_large_pages & (1 << TTE4M)) &&
2272 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET4M)) {
2273 			sfmmu_memtte(&tte, pfn, attr, TTE4M);
2274 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2275 			    flags, SFMMU_INVALID_SHMERID);
2276 			len -= MMU_PAGESIZE4M;
2277 			addr += MMU_PAGESIZE4M;
2278 			pfn += MMU_PAGESIZE4M / MMU_PAGESIZE;
2279 		} else if ((len >= MMU_PAGESIZE512K) &&
2280 		    !((uintptr_t)addr & MMU_PAGEOFFSET512K) &&
2281 		    !(disable_large_pages & (1 << TTE512K)) &&
2282 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET512K)) {
2283 			sfmmu_memtte(&tte, pfn, attr, TTE512K);
2284 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2285 			    flags, SFMMU_INVALID_SHMERID);
2286 			len -= MMU_PAGESIZE512K;
2287 			addr += MMU_PAGESIZE512K;
2288 			pfn += MMU_PAGESIZE512K / MMU_PAGESIZE;
2289 		} else if ((len >= MMU_PAGESIZE64K) &&
2290 		    !((uintptr_t)addr & MMU_PAGEOFFSET64K) &&
2291 		    !(disable_large_pages & (1 << TTE64K)) &&
2292 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET64K)) {
2293 			sfmmu_memtte(&tte, pfn, attr, TTE64K);
2294 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2295 			    flags, SFMMU_INVALID_SHMERID);
2296 			len -= MMU_PAGESIZE64K;
2297 			addr += MMU_PAGESIZE64K;
2298 			pfn += MMU_PAGESIZE64K / MMU_PAGESIZE;
2299 		} else {
2300 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2301 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2302 			    flags, SFMMU_INVALID_SHMERID);
2303 			len -= MMU_PAGESIZE;
2304 			addr += MMU_PAGESIZE;
2305 			pfn++;
2306 		}
2307 	}
2308 
2309 	/*
2310 	 * Check TSB and TLB page sizes.
2311 	 */
2312 	if ((flags & HAT_LOAD_SHARE) == 0) {
2313 		sfmmu_check_page_sizes(hat, 1);
2314 	}
2315 }
2316 
2317 void
2318 hat_memload_array(struct hat *hat, caddr_t addr, size_t len,
2319 	struct page **pps, uint_t attr, uint_t flags)
2320 {
2321 	hat_do_memload_array(hat, addr, len, pps, attr, flags,
2322 	    SFMMU_INVALID_SHMERID);
2323 }
2324 
2325 void
2326 hat_memload_array_region(struct hat *hat, caddr_t addr, size_t len,
2327 	struct page **pps, uint_t attr, uint_t flags,
2328 	hat_region_cookie_t rcookie)
2329 {
2330 	uint_t rid;
2331 	if (rcookie == HAT_INVALID_REGION_COOKIE ||
2332 	    hat->sfmmu_xhat_provider != NULL) {
2333 		hat_do_memload_array(hat, addr, len, pps, attr, flags,
2334 		    SFMMU_INVALID_SHMERID);
2335 		return;
2336 	}
2337 	rid = (uint_t)((uint64_t)rcookie);
2338 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2339 	hat_do_memload_array(hat, addr, len, pps, attr, flags, rid);
2340 }
2341 
2342 /*
2343  * Map the largest extend possible out of the page array. The array may NOT
2344  * be in order.  The largest possible mapping a page can have
2345  * is specified in the p_szc field.  The p_szc field
2346  * cannot change as long as there any mappings (large or small)
2347  * to any of the pages that make up the large page. (ie. any
2348  * promotion/demotion of page size is not up to the hat but up to
2349  * the page free list manager).  The array
2350  * should consist of properly aligned contigous pages that are
2351  * part of a big page for a large mapping to be created.
2352  */
2353 static void
2354 hat_do_memload_array(struct hat *hat, caddr_t addr, size_t len,
2355 	struct page **pps, uint_t attr, uint_t flags, uint_t rid)
2356 {
2357 	int  ttesz;
2358 	size_t mapsz;
2359 	pgcnt_t	numpg, npgs;
2360 	tte_t tte;
2361 	page_t *pp;
2362 	uint_t large_pages_disable;
2363 
2364 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2365 	SFMMU_VALIDATE_HMERID(hat, rid, addr, len);
2366 
2367 	if (hat->sfmmu_xhat_provider) {
2368 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
2369 		XHAT_MEMLOAD_ARRAY(hat, addr, len, pps, attr, flags);
2370 		return;
2371 	}
2372 
2373 	if (hat->sfmmu_rmstat)
2374 		hat_resvstat(len, hat->sfmmu_as, addr);
2375 
2376 #if defined(SF_ERRATA_57)
2377 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2378 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2379 	    !(flags & HAT_LOAD_SHARE)) {
2380 		cmn_err(CE_WARN, "hat_memload_array: illegal attempt to make "
2381 		    "user page executable");
2382 		attr &= ~PROT_EXEC;
2383 	}
2384 #endif
2385 
2386 	/* Get number of pages */
2387 	npgs = len >> MMU_PAGESHIFT;
2388 
2389 	if (flags & HAT_LOAD_SHARE) {
2390 		large_pages_disable = disable_ism_large_pages;
2391 	} else {
2392 		large_pages_disable = disable_large_pages;
2393 	}
2394 
2395 	if (npgs < NHMENTS || large_pages_disable == LARGE_PAGES_OFF) {
2396 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2397 		    rid);
2398 		return;
2399 	}
2400 
2401 	while (npgs >= NHMENTS) {
2402 		pp = *pps;
2403 		for (ttesz = pp->p_szc; ttesz != TTE8K; ttesz--) {
2404 			/*
2405 			 * Check if this page size is disabled.
2406 			 */
2407 			if (large_pages_disable & (1 << ttesz))
2408 				continue;
2409 
2410 			numpg = TTEPAGES(ttesz);
2411 			mapsz = numpg << MMU_PAGESHIFT;
2412 			if ((npgs >= numpg) &&
2413 			    IS_P2ALIGNED(addr, mapsz) &&
2414 			    IS_P2ALIGNED(pp->p_pagenum, numpg)) {
2415 				/*
2416 				 * At this point we have enough pages and
2417 				 * we know the virtual address and the pfn
2418 				 * are properly aligned.  We still need
2419 				 * to check for physical contiguity but since
2420 				 * it is very likely that this is the case
2421 				 * we will assume they are so and undo
2422 				 * the request if necessary.  It would
2423 				 * be great if we could get a hint flag
2424 				 * like HAT_CONTIG which would tell us
2425 				 * the pages are contigous for sure.
2426 				 */
2427 				sfmmu_memtte(&tte, (*pps)->p_pagenum,
2428 				    attr, ttesz);
2429 				if (!sfmmu_tteload_array(hat, &tte, addr,
2430 				    pps, flags, rid)) {
2431 					break;
2432 				}
2433 			}
2434 		}
2435 		if (ttesz == TTE8K) {
2436 			/*
2437 			 * We were not able to map array using a large page
2438 			 * batch a hmeblk or fraction at a time.
2439 			 */
2440 			numpg = ((uintptr_t)addr >> MMU_PAGESHIFT)
2441 			    & (NHMENTS-1);
2442 			numpg = NHMENTS - numpg;
2443 			ASSERT(numpg <= npgs);
2444 			mapsz = numpg * MMU_PAGESIZE;
2445 			sfmmu_memload_batchsmall(hat, addr, pps, attr, flags,
2446 			    numpg, rid);
2447 		}
2448 		addr += mapsz;
2449 		npgs -= numpg;
2450 		pps += numpg;
2451 	}
2452 
2453 	if (npgs) {
2454 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2455 		    rid);
2456 	}
2457 
2458 	/*
2459 	 * Check TSB and TLB page sizes.
2460 	 */
2461 	if ((flags & HAT_LOAD_SHARE) == 0) {
2462 		sfmmu_check_page_sizes(hat, 1);
2463 	}
2464 }
2465 
2466 /*
2467  * Function tries to batch 8K pages into the same hme blk.
2468  */
2469 static void
2470 sfmmu_memload_batchsmall(struct hat *hat, caddr_t vaddr, page_t **pps,
2471 		    uint_t attr, uint_t flags, pgcnt_t npgs, uint_t rid)
2472 {
2473 	tte_t	tte;
2474 	page_t *pp;
2475 	struct hmehash_bucket *hmebp;
2476 	struct hme_blk *hmeblkp;
2477 	int	index;
2478 
2479 	while (npgs) {
2480 		/*
2481 		 * Acquire the hash bucket.
2482 		 */
2483 		hmebp = sfmmu_tteload_acquire_hashbucket(hat, vaddr, TTE8K,
2484 		    rid);
2485 		ASSERT(hmebp);
2486 
2487 		/*
2488 		 * Find the hment block.
2489 		 */
2490 		hmeblkp = sfmmu_tteload_find_hmeblk(hat, hmebp, vaddr,
2491 		    TTE8K, flags, rid);
2492 		ASSERT(hmeblkp);
2493 
2494 		do {
2495 			/*
2496 			 * Make the tte.
2497 			 */
2498 			pp = *pps;
2499 			sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2500 
2501 			/*
2502 			 * Add the translation.
2503 			 */
2504 			(void) sfmmu_tteload_addentry(hat, hmeblkp, &tte,
2505 			    vaddr, pps, flags, rid);
2506 
2507 			/*
2508 			 * Goto next page.
2509 			 */
2510 			pps++;
2511 			npgs--;
2512 
2513 			/*
2514 			 * Goto next address.
2515 			 */
2516 			vaddr += MMU_PAGESIZE;
2517 
2518 			/*
2519 			 * Don't crossover into a different hmentblk.
2520 			 */
2521 			index = (int)(((uintptr_t)vaddr >> MMU_PAGESHIFT) &
2522 			    (NHMENTS-1));
2523 
2524 		} while (index != 0 && npgs != 0);
2525 
2526 		/*
2527 		 * Release the hash bucket.
2528 		 */
2529 
2530 		sfmmu_tteload_release_hashbucket(hmebp);
2531 	}
2532 }
2533 
2534 /*
2535  * Construct a tte for a page:
2536  *
2537  * tte_valid = 1
2538  * tte_size2 = size & TTE_SZ2_BITS (Panther and Olympus-C only)
2539  * tte_size = size
2540  * tte_nfo = attr & HAT_NOFAULT
2541  * tte_ie = attr & HAT_STRUCTURE_LE
2542  * tte_hmenum = hmenum
2543  * tte_pahi = pp->p_pagenum >> TTE_PASHIFT;
2544  * tte_palo = pp->p_pagenum & TTE_PALOMASK;
2545  * tte_ref = 1 (optimization)
2546  * tte_wr_perm = attr & PROT_WRITE;
2547  * tte_no_sync = attr & HAT_NOSYNC
2548  * tte_lock = attr & SFMMU_LOCKTTE
2549  * tte_cp = !(attr & SFMMU_UNCACHEPTTE)
2550  * tte_cv = !(attr & SFMMU_UNCACHEVTTE)
2551  * tte_e = attr & SFMMU_SIDEFFECT
2552  * tte_priv = !(attr & PROT_USER)
2553  * tte_hwwr = if nosync is set and it is writable we set the mod bit (opt)
2554  * tte_glb = 0
2555  */
2556 void
2557 sfmmu_memtte(tte_t *ttep, pfn_t pfn, uint_t attr, int tte_sz)
2558 {
2559 	ASSERT((attr & ~(SFMMU_LOAD_ALLATTR | HAT_ATTR_NOSOFTEXEC)) == 0);
2560 
2561 	ttep->tte_inthi = MAKE_TTE_INTHI(pfn, attr, tte_sz, 0 /* hmenum */);
2562 	ttep->tte_intlo = MAKE_TTE_INTLO(pfn, attr, tte_sz, 0 /* hmenum */);
2563 
2564 	if (TTE_IS_NOSYNC(ttep)) {
2565 		TTE_SET_REF(ttep);
2566 		if (TTE_IS_WRITABLE(ttep)) {
2567 			TTE_SET_MOD(ttep);
2568 		}
2569 	}
2570 	if (TTE_IS_NFO(ttep) && TTE_IS_EXECUTABLE(ttep)) {
2571 		panic("sfmmu_memtte: can't set both NFO and EXEC bits");
2572 	}
2573 
2574 	/*
2575 	 * Disable hardware execute permission to force a fault if
2576 	 * this page is executed, so we can detect the execution.  Set
2577 	 * the soft exec bit to remember that this TTE has execute
2578 	 * permission.
2579 	 */
2580 	if (TTE_IS_EXECUTABLE(ttep) && (attr & HAT_ATTR_NOSOFTEXEC) == 0 &&
2581 	    icache_is_coherent == 0) {
2582 		TTE_CLR_EXEC(ttep);
2583 		TTE_SET_SOFTEXEC(ttep);
2584 	}
2585 }
2586 
2587 /*
2588  * This function will add a translation to the hme_blk and allocate the
2589  * hme_blk if one does not exist.
2590  * If a page structure is specified then it will add the
2591  * corresponding hment to the mapping list.
2592  * It will also update the hmenum field for the tte.
2593  *
2594  * Currently this function is only used for kernel mappings.
2595  * So pass invalid region to sfmmu_tteload_array().
2596  */
2597 void
2598 sfmmu_tteload(struct hat *sfmmup, tte_t *ttep, caddr_t vaddr, page_t *pp,
2599 	uint_t flags)
2600 {
2601 	ASSERT(sfmmup == ksfmmup);
2602 	(void) sfmmu_tteload_array(sfmmup, ttep, vaddr, &pp, flags,
2603 	    SFMMU_INVALID_SHMERID);
2604 }
2605 
2606 /*
2607  * Load (ttep != NULL) or unload (ttep == NULL) one entry in the TSB.
2608  * Assumes that a particular page size may only be resident in one TSB.
2609  */
2610 static void
2611 sfmmu_mod_tsb(sfmmu_t *sfmmup, caddr_t vaddr, tte_t *ttep, int ttesz)
2612 {
2613 	struct tsb_info *tsbinfop = NULL;
2614 	uint64_t tag;
2615 	struct tsbe *tsbe_addr;
2616 	uint64_t tsb_base;
2617 	uint_t tsb_size;
2618 	int vpshift = MMU_PAGESHIFT;
2619 	int phys = 0;
2620 
2621 	if (sfmmup == ksfmmup) { /* No support for 32/256M ksfmmu pages */
2622 		phys = ktsb_phys;
2623 		if (ttesz >= TTE4M) {
2624 #ifndef sun4v
2625 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2626 #endif
2627 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2628 			tsb_size = ktsb4m_szcode;
2629 		} else {
2630 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2631 			tsb_size = ktsb_szcode;
2632 		}
2633 	} else {
2634 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2635 
2636 		/*
2637 		 * If there isn't a TSB for this page size, or the TSB is
2638 		 * swapped out, there is nothing to do.  Note that the latter
2639 		 * case seems impossible but can occur if hat_pageunload()
2640 		 * is called on an ISM mapping while the process is swapped
2641 		 * out.
2642 		 */
2643 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2644 			return;
2645 
2646 		/*
2647 		 * If another thread is in the middle of relocating a TSB
2648 		 * we can't unload the entry so set a flag so that the
2649 		 * TSB will be flushed before it can be accessed by the
2650 		 * process.
2651 		 */
2652 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2653 			if (ttep == NULL)
2654 				tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2655 			return;
2656 		}
2657 #if defined(UTSB_PHYS)
2658 		phys = 1;
2659 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2660 #else
2661 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2662 #endif
2663 		tsb_size = tsbinfop->tsb_szc;
2664 	}
2665 	if (ttesz >= TTE4M)
2666 		vpshift = MMU_PAGESHIFT4M;
2667 
2668 	tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2669 	tag = sfmmu_make_tsbtag(vaddr);
2670 
2671 	if (ttep == NULL) {
2672 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2673 	} else {
2674 		if (ttesz >= TTE4M) {
2675 			SFMMU_STAT(sf_tsb_load4m);
2676 		} else {
2677 			SFMMU_STAT(sf_tsb_load8k);
2678 		}
2679 
2680 		sfmmu_load_tsbe(tsbe_addr, tag, ttep, phys);
2681 	}
2682 }
2683 
2684 /*
2685  * Unmap all entries from [start, end) matching the given page size.
2686  *
2687  * This function is used primarily to unmap replicated 64K or 512K entries
2688  * from the TSB that are inserted using the base page size TSB pointer, but
2689  * it may also be called to unmap a range of addresses from the TSB.
2690  */
2691 void
2692 sfmmu_unload_tsb_range(sfmmu_t *sfmmup, caddr_t start, caddr_t end, int ttesz)
2693 {
2694 	struct tsb_info *tsbinfop;
2695 	uint64_t tag;
2696 	struct tsbe *tsbe_addr;
2697 	caddr_t vaddr;
2698 	uint64_t tsb_base;
2699 	int vpshift, vpgsz;
2700 	uint_t tsb_size;
2701 	int phys = 0;
2702 
2703 	/*
2704 	 * Assumptions:
2705 	 *  If ttesz == 8K, 64K or 512K, we walk through the range 8K
2706 	 *  at a time shooting down any valid entries we encounter.
2707 	 *
2708 	 *  If ttesz >= 4M we walk the range 4M at a time shooting
2709 	 *  down any valid mappings we find.
2710 	 */
2711 	if (sfmmup == ksfmmup) {
2712 		phys = ktsb_phys;
2713 		if (ttesz >= TTE4M) {
2714 #ifndef sun4v
2715 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2716 #endif
2717 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2718 			tsb_size = ktsb4m_szcode;
2719 		} else {
2720 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2721 			tsb_size = ktsb_szcode;
2722 		}
2723 	} else {
2724 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2725 
2726 		/*
2727 		 * If there isn't a TSB for this page size, or the TSB is
2728 		 * swapped out, there is nothing to do.  Note that the latter
2729 		 * case seems impossible but can occur if hat_pageunload()
2730 		 * is called on an ISM mapping while the process is swapped
2731 		 * out.
2732 		 */
2733 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2734 			return;
2735 
2736 		/*
2737 		 * If another thread is in the middle of relocating a TSB
2738 		 * we can't unload the entry so set a flag so that the
2739 		 * TSB will be flushed before it can be accessed by the
2740 		 * process.
2741 		 */
2742 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2743 			tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2744 			return;
2745 		}
2746 #if defined(UTSB_PHYS)
2747 		phys = 1;
2748 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2749 #else
2750 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2751 #endif
2752 		tsb_size = tsbinfop->tsb_szc;
2753 	}
2754 	if (ttesz >= TTE4M) {
2755 		vpshift = MMU_PAGESHIFT4M;
2756 		vpgsz = MMU_PAGESIZE4M;
2757 	} else {
2758 		vpshift = MMU_PAGESHIFT;
2759 		vpgsz = MMU_PAGESIZE;
2760 	}
2761 
2762 	for (vaddr = start; vaddr < end; vaddr += vpgsz) {
2763 		tag = sfmmu_make_tsbtag(vaddr);
2764 		tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2765 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2766 	}
2767 }
2768 
2769 /*
2770  * Select the optimum TSB size given the number of mappings
2771  * that need to be cached.
2772  */
2773 static int
2774 sfmmu_select_tsb_szc(pgcnt_t pgcnt)
2775 {
2776 	int szc = 0;
2777 
2778 #ifdef DEBUG
2779 	if (tsb_grow_stress) {
2780 		uint32_t randval = (uint32_t)gettick() >> 4;
2781 		return (randval % (tsb_max_growsize + 1));
2782 	}
2783 #endif	/* DEBUG */
2784 
2785 	while ((szc < tsb_max_growsize) && (pgcnt > SFMMU_RSS_TSBSIZE(szc)))
2786 		szc++;
2787 	return (szc);
2788 }
2789 
2790 /*
2791  * This function will add a translation to the hme_blk and allocate the
2792  * hme_blk if one does not exist.
2793  * If a page structure is specified then it will add the
2794  * corresponding hment to the mapping list.
2795  * It will also update the hmenum field for the tte.
2796  * Furthermore, it attempts to create a large page translation
2797  * for <addr,hat> at page array pps.  It assumes addr and first
2798  * pp is correctly aligned.  It returns 0 if successful and 1 otherwise.
2799  */
2800 static int
2801 sfmmu_tteload_array(sfmmu_t *sfmmup, tte_t *ttep, caddr_t vaddr,
2802 	page_t **pps, uint_t flags, uint_t rid)
2803 {
2804 	struct hmehash_bucket *hmebp;
2805 	struct hme_blk *hmeblkp;
2806 	int 	ret;
2807 	uint_t	size;
2808 
2809 	/*
2810 	 * Get mapping size.
2811 	 */
2812 	size = TTE_CSZ(ttep);
2813 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
2814 
2815 	/*
2816 	 * Acquire the hash bucket.
2817 	 */
2818 	hmebp = sfmmu_tteload_acquire_hashbucket(sfmmup, vaddr, size, rid);
2819 	ASSERT(hmebp);
2820 
2821 	/*
2822 	 * Find the hment block.
2823 	 */
2824 	hmeblkp = sfmmu_tteload_find_hmeblk(sfmmup, hmebp, vaddr, size, flags,
2825 	    rid);
2826 	ASSERT(hmeblkp);
2827 
2828 	/*
2829 	 * Add the translation.
2830 	 */
2831 	ret = sfmmu_tteload_addentry(sfmmup, hmeblkp, ttep, vaddr, pps, flags,
2832 	    rid);
2833 
2834 	/*
2835 	 * Release the hash bucket.
2836 	 */
2837 	sfmmu_tteload_release_hashbucket(hmebp);
2838 
2839 	return (ret);
2840 }
2841 
2842 /*
2843  * Function locks and returns a pointer to the hash bucket for vaddr and size.
2844  */
2845 static struct hmehash_bucket *
2846 sfmmu_tteload_acquire_hashbucket(sfmmu_t *sfmmup, caddr_t vaddr, int size,
2847     uint_t rid)
2848 {
2849 	struct hmehash_bucket *hmebp;
2850 	int hmeshift;
2851 	void *htagid = sfmmutohtagid(sfmmup, rid);
2852 
2853 	ASSERT(htagid != NULL);
2854 
2855 	hmeshift = HME_HASH_SHIFT(size);
2856 
2857 	hmebp = HME_HASH_FUNCTION(htagid, vaddr, hmeshift);
2858 
2859 	SFMMU_HASH_LOCK(hmebp);
2860 
2861 	return (hmebp);
2862 }
2863 
2864 /*
2865  * Function returns a pointer to an hmeblk in the hash bucket, hmebp. If the
2866  * hmeblk doesn't exists for the [sfmmup, vaddr & size] signature, a hmeblk is
2867  * allocated.
2868  */
2869 static struct hme_blk *
2870 sfmmu_tteload_find_hmeblk(sfmmu_t *sfmmup, struct hmehash_bucket *hmebp,
2871 	caddr_t vaddr, uint_t size, uint_t flags, uint_t rid)
2872 {
2873 	hmeblk_tag hblktag;
2874 	int hmeshift;
2875 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
2876 
2877 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
2878 
2879 	hblktag.htag_id = sfmmutohtagid(sfmmup, rid);
2880 	ASSERT(hblktag.htag_id != NULL);
2881 	hmeshift = HME_HASH_SHIFT(size);
2882 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
2883 	hblktag.htag_rehash = HME_HASH_REHASH(size);
2884 	hblktag.htag_rid = rid;
2885 
2886 ttearray_realloc:
2887 
2888 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
2889 
2890 	/*
2891 	 * We block until hblk_reserve_lock is released; it's held by
2892 	 * the thread, temporarily using hblk_reserve, until hblk_reserve is
2893 	 * replaced by a hblk from sfmmu8_cache.
2894 	 */
2895 	if (hmeblkp == (struct hme_blk *)hblk_reserve &&
2896 	    hblk_reserve_thread != curthread) {
2897 		SFMMU_HASH_UNLOCK(hmebp);
2898 		mutex_enter(&hblk_reserve_lock);
2899 		mutex_exit(&hblk_reserve_lock);
2900 		SFMMU_STAT(sf_hblk_reserve_hit);
2901 		SFMMU_HASH_LOCK(hmebp);
2902 		goto ttearray_realloc;
2903 	}
2904 
2905 	if (hmeblkp == NULL) {
2906 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
2907 		    hblktag, flags, rid);
2908 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
2909 		ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
2910 	} else {
2911 		/*
2912 		 * It is possible for 8k and 64k hblks to collide since they
2913 		 * have the same rehash value. This is because we
2914 		 * lazily free hblks and 8K/64K blks could be lingering.
2915 		 * If we find size mismatch we free the block and & try again.
2916 		 */
2917 		if (get_hblk_ttesz(hmeblkp) != size) {
2918 			ASSERT(!hmeblkp->hblk_vcnt);
2919 			ASSERT(!hmeblkp->hblk_hmecnt);
2920 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
2921 			    &list, 0);
2922 			goto ttearray_realloc;
2923 		}
2924 		if (hmeblkp->hblk_shw_bit) {
2925 			/*
2926 			 * if the hblk was previously used as a shadow hblk then
2927 			 * we will change it to a normal hblk
2928 			 */
2929 			ASSERT(!hmeblkp->hblk_shared);
2930 			if (hmeblkp->hblk_shw_mask) {
2931 				sfmmu_shadow_hcleanup(sfmmup, hmeblkp, hmebp);
2932 				ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
2933 				goto ttearray_realloc;
2934 			} else {
2935 				hmeblkp->hblk_shw_bit = 0;
2936 			}
2937 		}
2938 		SFMMU_STAT(sf_hblk_hit);
2939 	}
2940 
2941 	/*
2942 	 * hat_memload() should never call kmem_cache_free() for kernel hmeblks;
2943 	 * see block comment showing the stacktrace in sfmmu_hblk_alloc();
2944 	 * set the flag parameter to 1 so that sfmmu_hblks_list_purge() will
2945 	 * just add these hmeblks to the per-cpu pending queue.
2946 	 */
2947 	sfmmu_hblks_list_purge(&list, 1);
2948 
2949 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
2950 	ASSERT(!hmeblkp->hblk_shw_bit);
2951 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
2952 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
2953 	ASSERT(hmeblkp->hblk_tag.htag_rid == rid);
2954 
2955 	return (hmeblkp);
2956 }
2957 
2958 /*
2959  * Function adds a tte entry into the hmeblk. It returns 0 if successful and 1
2960  * otherwise.
2961  */
2962 static int
2963 sfmmu_tteload_addentry(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, tte_t *ttep,
2964 	caddr_t vaddr, page_t **pps, uint_t flags, uint_t rid)
2965 {
2966 	page_t *pp = *pps;
2967 	int hmenum, size, remap;
2968 	tte_t tteold, flush_tte;
2969 #ifdef DEBUG
2970 	tte_t orig_old;
2971 #endif /* DEBUG */
2972 	struct sf_hment *sfhme;
2973 	kmutex_t *pml, *pmtx;
2974 	hatlock_t *hatlockp;
2975 	int myflt;
2976 
2977 	/*
2978 	 * remove this panic when we decide to let user virtual address
2979 	 * space be >= USERLIMIT.
2980 	 */
2981 	if (!TTE_IS_PRIVILEGED(ttep) && vaddr >= (caddr_t)USERLIMIT)
2982 		panic("user addr %p in kernel space", (void *)vaddr);
2983 #if defined(TTE_IS_GLOBAL)
2984 	if (TTE_IS_GLOBAL(ttep))
2985 		panic("sfmmu_tteload: creating global tte");
2986 #endif
2987 
2988 #ifdef DEBUG
2989 	if (pf_is_memory(sfmmu_ttetopfn(ttep, vaddr)) &&
2990 	    !TTE_IS_PCACHEABLE(ttep) && !sfmmu_allow_nc_trans)
2991 		panic("sfmmu_tteload: non cacheable memory tte");
2992 #endif /* DEBUG */
2993 
2994 	/* don't simulate dirty bit for writeable ISM/DISM mappings */
2995 	if ((flags & HAT_LOAD_SHARE) && TTE_IS_WRITABLE(ttep)) {
2996 		TTE_SET_REF(ttep);
2997 		TTE_SET_MOD(ttep);
2998 	}
2999 
3000 	if ((flags & HAT_LOAD_SHARE) || !TTE_IS_REF(ttep) ||
3001 	    !TTE_IS_MOD(ttep)) {
3002 		/*
3003 		 * Don't load TSB for dummy as in ISM.  Also don't preload
3004 		 * the TSB if the TTE isn't writable since we're likely to
3005 		 * fault on it again -- preloading can be fairly expensive.
3006 		 */
3007 		flags |= SFMMU_NO_TSBLOAD;
3008 	}
3009 
3010 	size = TTE_CSZ(ttep);
3011 	switch (size) {
3012 	case TTE8K:
3013 		SFMMU_STAT(sf_tteload8k);
3014 		break;
3015 	case TTE64K:
3016 		SFMMU_STAT(sf_tteload64k);
3017 		break;
3018 	case TTE512K:
3019 		SFMMU_STAT(sf_tteload512k);
3020 		break;
3021 	case TTE4M:
3022 		SFMMU_STAT(sf_tteload4m);
3023 		break;
3024 	case (TTE32M):
3025 		SFMMU_STAT(sf_tteload32m);
3026 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3027 		break;
3028 	case (TTE256M):
3029 		SFMMU_STAT(sf_tteload256m);
3030 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3031 		break;
3032 	}
3033 
3034 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
3035 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
3036 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3037 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3038 
3039 	HBLKTOHME_IDX(sfhme, hmeblkp, vaddr, hmenum);
3040 
3041 	/*
3042 	 * Need to grab mlist lock here so that pageunload
3043 	 * will not change tte behind us.
3044 	 */
3045 	if (pp) {
3046 		pml = sfmmu_mlist_enter(pp);
3047 	}
3048 
3049 	sfmmu_copytte(&sfhme->hme_tte, &tteold);
3050 	/*
3051 	 * Look for corresponding hment and if valid verify
3052 	 * pfns are equal.
3053 	 */
3054 	remap = TTE_IS_VALID(&tteold);
3055 	if (remap) {
3056 		pfn_t	new_pfn, old_pfn;
3057 
3058 		old_pfn = TTE_TO_PFN(vaddr, &tteold);
3059 		new_pfn = TTE_TO_PFN(vaddr, ttep);
3060 
3061 		if (flags & HAT_LOAD_REMAP) {
3062 			/* make sure we are remapping same type of pages */
3063 			if (pf_is_memory(old_pfn) != pf_is_memory(new_pfn)) {
3064 				panic("sfmmu_tteload - tte remap io<->memory");
3065 			}
3066 			if (old_pfn != new_pfn &&
3067 			    (pp != NULL || sfhme->hme_page != NULL)) {
3068 				panic("sfmmu_tteload - tte remap pp != NULL");
3069 			}
3070 		} else if (old_pfn != new_pfn) {
3071 			panic("sfmmu_tteload - tte remap, hmeblkp 0x%p",
3072 			    (void *)hmeblkp);
3073 		}
3074 		ASSERT(TTE_CSZ(&tteold) == TTE_CSZ(ttep));
3075 
3076 		if (TTE_IS_EXECUTABLE(&tteold) && TTE_IS_SOFTEXEC(ttep)) {
3077 			TTE_SET_EXEC(ttep);
3078 		}
3079 	}
3080 
3081 	if (pp) {
3082 		/*
3083 		 * If we know that this page will be executed, because
3084 		 * it was in the past (PP_ISEXEC is already true), or
3085 		 * if the caller says it will likely be executed
3086 		 * (HAT_LOAD_TEXT is true), then there is no need to
3087 		 * dynamically detect execution with a soft exec
3088 		 * fault. Enable hardware execute permission now.
3089 		 */
3090 		if ((PP_ISEXEC(pp) || (flags & HAT_LOAD_TEXT)) &&
3091 		    TTE_IS_SOFTEXEC(ttep)) {
3092 			TTE_SET_EXEC(ttep);
3093 		}
3094 
3095 		if (size == TTE8K) {
3096 #ifdef VAC
3097 			/*
3098 			 * Handle VAC consistency
3099 			 */
3100 			if (!remap && (cache & CACHE_VAC) && !PP_ISNC(pp)) {
3101 				sfmmu_vac_conflict(sfmmup, vaddr, pp);
3102 			}
3103 #endif
3104 
3105 			if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3106 				pmtx = sfmmu_page_enter(pp);
3107 				PP_CLRRO(pp);
3108 				sfmmu_page_exit(pmtx);
3109 			} else if (!PP_ISMAPPED(pp) &&
3110 			    (!TTE_IS_WRITABLE(ttep)) && !(PP_ISMOD(pp))) {
3111 				pmtx = sfmmu_page_enter(pp);
3112 				if (!(PP_ISMOD(pp))) {
3113 					PP_SETRO(pp);
3114 				}
3115 				sfmmu_page_exit(pmtx);
3116 			}
3117 
3118 			if (TTE_EXECUTED(ttep)) {
3119 				pmtx = sfmmu_page_enter(pp);
3120 				PP_SETEXEC(pp);
3121 				sfmmu_page_exit(pmtx);
3122 			}
3123 
3124 		} else if (sfmmu_pagearray_setup(vaddr, pps, ttep, remap)) {
3125 			/*
3126 			 * sfmmu_pagearray_setup failed so return
3127 			 */
3128 			sfmmu_mlist_exit(pml);
3129 			return (1);
3130 		}
3131 
3132 	} else if (TTE_IS_SOFTEXEC(ttep)) {
3133 		TTE_SET_EXEC(ttep);
3134 	}
3135 
3136 	/*
3137 	 * Make sure hment is not on a mapping list.
3138 	 */
3139 	ASSERT(remap || (sfhme->hme_page == NULL));
3140 
3141 	/* if it is not a remap then hme->next better be NULL */
3142 	ASSERT((!remap) ? sfhme->hme_next == NULL : 1);
3143 
3144 	if (flags & HAT_LOAD_LOCK) {
3145 		if ((hmeblkp->hblk_lckcnt + 1) >= MAX_HBLK_LCKCNT) {
3146 			panic("too high lckcnt-hmeblk %p",
3147 			    (void *)hmeblkp);
3148 		}
3149 		atomic_add_32(&hmeblkp->hblk_lckcnt, 1);
3150 
3151 		HBLK_STACK_TRACE(hmeblkp, HBLK_LOCK);
3152 	}
3153 
3154 #ifdef VAC
3155 	if (pp && PP_ISNC(pp)) {
3156 		/*
3157 		 * If the physical page is marked to be uncacheable, like
3158 		 * by a vac conflict, make sure the new mapping is also
3159 		 * uncacheable.
3160 		 */
3161 		TTE_CLR_VCACHEABLE(ttep);
3162 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
3163 	}
3164 #endif
3165 	ttep->tte_hmenum = hmenum;
3166 
3167 #ifdef DEBUG
3168 	orig_old = tteold;
3169 #endif /* DEBUG */
3170 
3171 	while (sfmmu_modifytte_try(&tteold, ttep, &sfhme->hme_tte) < 0) {
3172 		if ((sfmmup == KHATID) &&
3173 		    (flags & (HAT_LOAD_LOCK | HAT_LOAD_REMAP))) {
3174 			sfmmu_copytte(&sfhme->hme_tte, &tteold);
3175 		}
3176 #ifdef DEBUG
3177 		chk_tte(&orig_old, &tteold, ttep, hmeblkp);
3178 #endif /* DEBUG */
3179 	}
3180 	ASSERT(TTE_IS_VALID(&sfhme->hme_tte));
3181 
3182 	if (!TTE_IS_VALID(&tteold)) {
3183 
3184 		atomic_add_16(&hmeblkp->hblk_vcnt, 1);
3185 		if (rid == SFMMU_INVALID_SHMERID) {
3186 			atomic_add_long(&sfmmup->sfmmu_ttecnt[size], 1);
3187 		} else {
3188 			sf_srd_t *srdp = sfmmup->sfmmu_srdp;
3189 			sf_region_t *rgnp = srdp->srd_hmergnp[rid];
3190 			/*
3191 			 * We already accounted for region ttecnt's in sfmmu
3192 			 * during hat_join_region() processing. Here we
3193 			 * only update ttecnt's in region struture.
3194 			 */
3195 			atomic_add_long(&rgnp->rgn_ttecnt[size], 1);
3196 		}
3197 	}
3198 
3199 	myflt = (astosfmmu(curthread->t_procp->p_as) == sfmmup);
3200 	if (size > TTE8K && (flags & HAT_LOAD_SHARE) == 0 &&
3201 	    sfmmup != ksfmmup) {
3202 		uchar_t tteflag = 1 << size;
3203 		if (rid == SFMMU_INVALID_SHMERID) {
3204 			if (!(sfmmup->sfmmu_tteflags & tteflag)) {
3205 				hatlockp = sfmmu_hat_enter(sfmmup);
3206 				sfmmup->sfmmu_tteflags |= tteflag;
3207 				sfmmu_hat_exit(hatlockp);
3208 			}
3209 		} else if (!(sfmmup->sfmmu_rtteflags & tteflag)) {
3210 			hatlockp = sfmmu_hat_enter(sfmmup);
3211 			sfmmup->sfmmu_rtteflags |= tteflag;
3212 			sfmmu_hat_exit(hatlockp);
3213 		}
3214 		/*
3215 		 * Update the current CPU tsbmiss area, so the current thread
3216 		 * won't need to take the tsbmiss for the new pagesize.
3217 		 * The other threads in the process will update their tsb
3218 		 * miss area lazily in sfmmu_tsbmiss_exception() when they
3219 		 * fail to find the translation for a newly added pagesize.
3220 		 */
3221 		if (size > TTE64K && myflt) {
3222 			struct tsbmiss *tsbmp;
3223 			kpreempt_disable();
3224 			tsbmp = &tsbmiss_area[CPU->cpu_id];
3225 			if (rid == SFMMU_INVALID_SHMERID) {
3226 				if (!(tsbmp->uhat_tteflags & tteflag)) {
3227 					tsbmp->uhat_tteflags |= tteflag;
3228 				}
3229 			} else {
3230 				if (!(tsbmp->uhat_rtteflags & tteflag)) {
3231 					tsbmp->uhat_rtteflags |= tteflag;
3232 				}
3233 			}
3234 			kpreempt_enable();
3235 		}
3236 	}
3237 
3238 	if (size >= TTE4M && (flags & HAT_LOAD_TEXT) &&
3239 	    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
3240 		hatlockp = sfmmu_hat_enter(sfmmup);
3241 		SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
3242 		sfmmu_hat_exit(hatlockp);
3243 	}
3244 
3245 	flush_tte.tte_intlo = (tteold.tte_intlo ^ ttep->tte_intlo) &
3246 	    hw_tte.tte_intlo;
3247 	flush_tte.tte_inthi = (tteold.tte_inthi ^ ttep->tte_inthi) &
3248 	    hw_tte.tte_inthi;
3249 
3250 	if (remap && (flush_tte.tte_inthi || flush_tte.tte_intlo)) {
3251 		/*
3252 		 * If remap and new tte differs from old tte we need
3253 		 * to sync the mod bit and flush TLB/TSB.  We don't
3254 		 * need to sync ref bit because we currently always set
3255 		 * ref bit in tteload.
3256 		 */
3257 		ASSERT(TTE_IS_REF(ttep));
3258 		if (TTE_IS_MOD(&tteold) || (TTE_EXECUTED(&tteold) &&
3259 		    !TTE_IS_EXECUTABLE(ttep))) {
3260 			sfmmu_ttesync(sfmmup, vaddr, &tteold, pp);
3261 		}
3262 		/*
3263 		 * hwtte bits shouldn't change for SRD hmeblks as long as SRD
3264 		 * hmes are only used for read only text. Adding this code for
3265 		 * completeness and future use of shared hmeblks with writable
3266 		 * mappings of VMODSORT vnodes.
3267 		 */
3268 		if (hmeblkp->hblk_shared) {
3269 			cpuset_t cpuset = sfmmu_rgntlb_demap(vaddr,
3270 			    sfmmup->sfmmu_srdp->srd_hmergnp[rid], hmeblkp, 1);
3271 			xt_sync(cpuset);
3272 			SFMMU_STAT_ADD(sf_region_remap_demap, 1);
3273 		} else {
3274 			sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 0);
3275 			xt_sync(sfmmup->sfmmu_cpusran);
3276 		}
3277 	}
3278 
3279 	if ((flags & SFMMU_NO_TSBLOAD) == 0) {
3280 		/*
3281 		 * We only preload 8K and 4M mappings into the TSB, since
3282 		 * 64K and 512K mappings are replicated and hence don't
3283 		 * have a single, unique TSB entry. Ditto for 32M/256M.
3284 		 */
3285 		if (size == TTE8K || size == TTE4M) {
3286 			sf_scd_t *scdp;
3287 			hatlockp = sfmmu_hat_enter(sfmmup);
3288 			/*
3289 			 * Don't preload private TSB if the mapping is used
3290 			 * by the shctx in the SCD.
3291 			 */
3292 			scdp = sfmmup->sfmmu_scdp;
3293 			if (rid == SFMMU_INVALID_SHMERID || scdp == NULL ||
3294 			    !SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
3295 				sfmmu_load_tsb(sfmmup, vaddr, &sfhme->hme_tte,
3296 				    size);
3297 			}
3298 			sfmmu_hat_exit(hatlockp);
3299 		}
3300 	}
3301 	if (pp) {
3302 		if (!remap) {
3303 			HME_ADD(sfhme, pp);
3304 			atomic_add_16(&hmeblkp->hblk_hmecnt, 1);
3305 			ASSERT(hmeblkp->hblk_hmecnt > 0);
3306 
3307 			/*
3308 			 * Cannot ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
3309 			 * see pageunload() for comment.
3310 			 */
3311 		}
3312 		sfmmu_mlist_exit(pml);
3313 	}
3314 
3315 	return (0);
3316 }
3317 /*
3318  * Function unlocks hash bucket.
3319  */
3320 static void
3321 sfmmu_tteload_release_hashbucket(struct hmehash_bucket *hmebp)
3322 {
3323 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3324 	SFMMU_HASH_UNLOCK(hmebp);
3325 }
3326 
3327 /*
3328  * function which checks and sets up page array for a large
3329  * translation.  Will set p_vcolor, p_index, p_ro fields.
3330  * Assumes addr and pfnum of first page are properly aligned.
3331  * Will check for physical contiguity. If check fails it return
3332  * non null.
3333  */
3334 static int
3335 sfmmu_pagearray_setup(caddr_t addr, page_t **pps, tte_t *ttep, int remap)
3336 {
3337 	int 	i, index, ttesz;
3338 	pfn_t	pfnum;
3339 	pgcnt_t	npgs;
3340 	page_t *pp, *pp1;
3341 	kmutex_t *pmtx;
3342 #ifdef VAC
3343 	int osz;
3344 	int cflags = 0;
3345 	int vac_err = 0;
3346 #endif
3347 	int newidx = 0;
3348 
3349 	ttesz = TTE_CSZ(ttep);
3350 
3351 	ASSERT(ttesz > TTE8K);
3352 
3353 	npgs = TTEPAGES(ttesz);
3354 	index = PAGESZ_TO_INDEX(ttesz);
3355 
3356 	pfnum = (*pps)->p_pagenum;
3357 	ASSERT(IS_P2ALIGNED(pfnum, npgs));
3358 
3359 	/*
3360 	 * Save the first pp so we can do HAT_TMPNC at the end.
3361 	 */
3362 	pp1 = *pps;
3363 #ifdef VAC
3364 	osz = fnd_mapping_sz(pp1);
3365 #endif
3366 
3367 	for (i = 0; i < npgs; i++, pps++) {
3368 		pp = *pps;
3369 		ASSERT(PAGE_LOCKED(pp));
3370 		ASSERT(pp->p_szc >= ttesz);
3371 		ASSERT(pp->p_szc == pp1->p_szc);
3372 		ASSERT(sfmmu_mlist_held(pp));
3373 
3374 		/*
3375 		 * XXX is it possible to maintain P_RO on the root only?
3376 		 */
3377 		if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3378 			pmtx = sfmmu_page_enter(pp);
3379 			PP_CLRRO(pp);
3380 			sfmmu_page_exit(pmtx);
3381 		} else if (!PP_ISMAPPED(pp) && !TTE_IS_WRITABLE(ttep) &&
3382 		    !PP_ISMOD(pp)) {
3383 			pmtx = sfmmu_page_enter(pp);
3384 			if (!(PP_ISMOD(pp))) {
3385 				PP_SETRO(pp);
3386 			}
3387 			sfmmu_page_exit(pmtx);
3388 		}
3389 
3390 		if (TTE_EXECUTED(ttep)) {
3391 			pmtx = sfmmu_page_enter(pp);
3392 			PP_SETEXEC(pp);
3393 			sfmmu_page_exit(pmtx);
3394 		}
3395 
3396 		/*
3397 		 * If this is a remap we skip vac & contiguity checks.
3398 		 */
3399 		if (remap)
3400 			continue;
3401 
3402 		/*
3403 		 * set p_vcolor and detect any vac conflicts.
3404 		 */
3405 #ifdef VAC
3406 		if (vac_err == 0) {
3407 			vac_err = sfmmu_vacconflict_array(addr, pp, &cflags);
3408 
3409 		}
3410 #endif
3411 
3412 		/*
3413 		 * Save current index in case we need to undo it.
3414 		 * Note: "PAGESZ_TO_INDEX(sz)	(1 << (sz))"
3415 		 *	"SFMMU_INDEX_SHIFT	6"
3416 		 *	 "SFMMU_INDEX_MASK	((1 << SFMMU_INDEX_SHIFT) - 1)"
3417 		 *	 "PP_MAPINDEX(p_index)	(p_index & SFMMU_INDEX_MASK)"
3418 		 *
3419 		 * So:	index = PAGESZ_TO_INDEX(ttesz);
3420 		 *	if ttesz == 1 then index = 0x2
3421 		 *		    2 then index = 0x4
3422 		 *		    3 then index = 0x8
3423 		 *		    4 then index = 0x10
3424 		 *		    5 then index = 0x20
3425 		 * The code below checks if it's a new pagesize (ie, newidx)
3426 		 * in case we need to take it back out of p_index,
3427 		 * and then or's the new index into the existing index.
3428 		 */
3429 		if ((PP_MAPINDEX(pp) & index) == 0)
3430 			newidx = 1;
3431 		pp->p_index = (PP_MAPINDEX(pp) | index);
3432 
3433 		/*
3434 		 * contiguity check
3435 		 */
3436 		if (pp->p_pagenum != pfnum) {
3437 			/*
3438 			 * If we fail the contiguity test then
3439 			 * the only thing we need to fix is the p_index field.
3440 			 * We might get a few extra flushes but since this
3441 			 * path is rare that is ok.  The p_ro field will
3442 			 * get automatically fixed on the next tteload to
3443 			 * the page.  NO TNC bit is set yet.
3444 			 */
3445 			while (i >= 0) {
3446 				pp = *pps;
3447 				if (newidx)
3448 					pp->p_index = (PP_MAPINDEX(pp) &
3449 					    ~index);
3450 				pps--;
3451 				i--;
3452 			}
3453 			return (1);
3454 		}
3455 		pfnum++;
3456 		addr += MMU_PAGESIZE;
3457 	}
3458 
3459 #ifdef VAC
3460 	if (vac_err) {
3461 		if (ttesz > osz) {
3462 			/*
3463 			 * There are some smaller mappings that causes vac
3464 			 * conflicts. Convert all existing small mappings to
3465 			 * TNC.
3466 			 */
3467 			SFMMU_STAT_ADD(sf_uncache_conflict, npgs);
3468 			sfmmu_page_cache_array(pp1, HAT_TMPNC, CACHE_FLUSH,
3469 			    npgs);
3470 		} else {
3471 			/* EMPTY */
3472 			/*
3473 			 * If there exists an big page mapping,
3474 			 * that means the whole existing big page
3475 			 * has TNC setting already. No need to covert to
3476 			 * TNC again.
3477 			 */
3478 			ASSERT(PP_ISTNC(pp1));
3479 		}
3480 	}
3481 #endif	/* VAC */
3482 
3483 	return (0);
3484 }
3485 
3486 #ifdef VAC
3487 /*
3488  * Routine that detects vac consistency for a large page. It also
3489  * sets virtual color for all pp's for this big mapping.
3490  */
3491 static int
3492 sfmmu_vacconflict_array(caddr_t addr, page_t *pp, int *cflags)
3493 {
3494 	int vcolor, ocolor;
3495 
3496 	ASSERT(sfmmu_mlist_held(pp));
3497 
3498 	if (PP_ISNC(pp)) {
3499 		return (HAT_TMPNC);
3500 	}
3501 
3502 	vcolor = addr_to_vcolor(addr);
3503 	if (PP_NEWPAGE(pp)) {
3504 		PP_SET_VCOLOR(pp, vcolor);
3505 		return (0);
3506 	}
3507 
3508 	ocolor = PP_GET_VCOLOR(pp);
3509 	if (ocolor == vcolor) {
3510 		return (0);
3511 	}
3512 
3513 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
3514 		/*
3515 		 * Previous user of page had a differnet color
3516 		 * but since there are no current users
3517 		 * we just flush the cache and change the color.
3518 		 * As an optimization for large pages we flush the
3519 		 * entire cache of that color and set a flag.
3520 		 */
3521 		SFMMU_STAT(sf_pgcolor_conflict);
3522 		if (!CacheColor_IsFlushed(*cflags, ocolor)) {
3523 			CacheColor_SetFlushed(*cflags, ocolor);
3524 			sfmmu_cache_flushcolor(ocolor, pp->p_pagenum);
3525 		}
3526 		PP_SET_VCOLOR(pp, vcolor);
3527 		return (0);
3528 	}
3529 
3530 	/*
3531 	 * We got a real conflict with a current mapping.
3532 	 * set flags to start unencaching all mappings
3533 	 * and return failure so we restart looping
3534 	 * the pp array from the beginning.
3535 	 */
3536 	return (HAT_TMPNC);
3537 }
3538 #endif	/* VAC */
3539 
3540 /*
3541  * creates a large page shadow hmeblk for a tte.
3542  * The purpose of this routine is to allow us to do quick unloads because
3543  * the vm layer can easily pass a very large but sparsely populated range.
3544  */
3545 static struct hme_blk *
3546 sfmmu_shadow_hcreate(sfmmu_t *sfmmup, caddr_t vaddr, int ttesz, uint_t flags)
3547 {
3548 	struct hmehash_bucket *hmebp;
3549 	hmeblk_tag hblktag;
3550 	int hmeshift, size, vshift;
3551 	uint_t shw_mask, newshw_mask;
3552 	struct hme_blk *hmeblkp;
3553 
3554 	ASSERT(sfmmup != KHATID);
3555 	if (mmu_page_sizes == max_mmu_page_sizes) {
3556 		ASSERT(ttesz < TTE256M);
3557 	} else {
3558 		ASSERT(ttesz < TTE4M);
3559 		ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
3560 		ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
3561 	}
3562 
3563 	if (ttesz == TTE8K) {
3564 		size = TTE512K;
3565 	} else {
3566 		size = ++ttesz;
3567 	}
3568 
3569 	hblktag.htag_id = sfmmup;
3570 	hmeshift = HME_HASH_SHIFT(size);
3571 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
3572 	hblktag.htag_rehash = HME_HASH_REHASH(size);
3573 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3574 	hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
3575 
3576 	SFMMU_HASH_LOCK(hmebp);
3577 
3578 	HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
3579 	ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
3580 	if (hmeblkp == NULL) {
3581 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
3582 		    hblktag, flags, SFMMU_INVALID_SHMERID);
3583 	}
3584 	ASSERT(hmeblkp);
3585 	if (!hmeblkp->hblk_shw_mask) {
3586 		/*
3587 		 * if this is a unused hblk it was just allocated or could
3588 		 * potentially be a previous large page hblk so we need to
3589 		 * set the shadow bit.
3590 		 */
3591 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3592 		hmeblkp->hblk_shw_bit = 1;
3593 	} else if (hmeblkp->hblk_shw_bit == 0) {
3594 		panic("sfmmu_shadow_hcreate: shw bit not set in hmeblkp 0x%p",
3595 		    (void *)hmeblkp);
3596 	}
3597 	ASSERT(hmeblkp->hblk_shw_bit == 1);
3598 	ASSERT(!hmeblkp->hblk_shared);
3599 	vshift = vaddr_to_vshift(hblktag, vaddr, size);
3600 	ASSERT(vshift < 8);
3601 	/*
3602 	 * Atomically set shw mask bit
3603 	 */
3604 	do {
3605 		shw_mask = hmeblkp->hblk_shw_mask;
3606 		newshw_mask = shw_mask | (1 << vshift);
3607 		newshw_mask = cas32(&hmeblkp->hblk_shw_mask, shw_mask,
3608 		    newshw_mask);
3609 	} while (newshw_mask != shw_mask);
3610 
3611 	SFMMU_HASH_UNLOCK(hmebp);
3612 
3613 	return (hmeblkp);
3614 }
3615 
3616 /*
3617  * This routine cleanup a previous shadow hmeblk and changes it to
3618  * a regular hblk.  This happens rarely but it is possible
3619  * when a process wants to use large pages and there are hblks still
3620  * lying around from the previous as that used these hmeblks.
3621  * The alternative was to cleanup the shadow hblks at unload time
3622  * but since so few user processes actually use large pages, it is
3623  * better to be lazy and cleanup at this time.
3624  */
3625 static void
3626 sfmmu_shadow_hcleanup(sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
3627 	struct hmehash_bucket *hmebp)
3628 {
3629 	caddr_t addr, endaddr;
3630 	int hashno, size;
3631 
3632 	ASSERT(hmeblkp->hblk_shw_bit);
3633 	ASSERT(!hmeblkp->hblk_shared);
3634 
3635 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3636 
3637 	if (!hmeblkp->hblk_shw_mask) {
3638 		hmeblkp->hblk_shw_bit = 0;
3639 		return;
3640 	}
3641 	addr = (caddr_t)get_hblk_base(hmeblkp);
3642 	endaddr = get_hblk_endaddr(hmeblkp);
3643 	size = get_hblk_ttesz(hmeblkp);
3644 	hashno = size - 1;
3645 	ASSERT(hashno > 0);
3646 	SFMMU_HASH_UNLOCK(hmebp);
3647 
3648 	sfmmu_free_hblks(sfmmup, addr, endaddr, hashno);
3649 
3650 	SFMMU_HASH_LOCK(hmebp);
3651 }
3652 
3653 static void
3654 sfmmu_free_hblks(sfmmu_t *sfmmup, caddr_t addr, caddr_t endaddr,
3655 	int hashno)
3656 {
3657 	int hmeshift, shadow = 0;
3658 	hmeblk_tag hblktag;
3659 	struct hmehash_bucket *hmebp;
3660 	struct hme_blk *hmeblkp;
3661 	struct hme_blk *nx_hblk, *pr_hblk, *list = NULL;
3662 
3663 	ASSERT(hashno > 0);
3664 	hblktag.htag_id = sfmmup;
3665 	hblktag.htag_rehash = hashno;
3666 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3667 
3668 	hmeshift = HME_HASH_SHIFT(hashno);
3669 
3670 	while (addr < endaddr) {
3671 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3672 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3673 		SFMMU_HASH_LOCK(hmebp);
3674 		/* inline HME_HASH_SEARCH */
3675 		hmeblkp = hmebp->hmeblkp;
3676 		pr_hblk = NULL;
3677 		while (hmeblkp) {
3678 			if (HTAGS_EQ(hmeblkp->hblk_tag, hblktag)) {
3679 				/* found hme_blk */
3680 				ASSERT(!hmeblkp->hblk_shared);
3681 				if (hmeblkp->hblk_shw_bit) {
3682 					if (hmeblkp->hblk_shw_mask) {
3683 						shadow = 1;
3684 						sfmmu_shadow_hcleanup(sfmmup,
3685 						    hmeblkp, hmebp);
3686 						break;
3687 					} else {
3688 						hmeblkp->hblk_shw_bit = 0;
3689 					}
3690 				}
3691 
3692 				/*
3693 				 * Hblk_hmecnt and hblk_vcnt could be non zero
3694 				 * since hblk_unload() does not gurantee that.
3695 				 *
3696 				 * XXX - this could cause tteload() to spin
3697 				 * where sfmmu_shadow_hcleanup() is called.
3698 				 */
3699 			}
3700 
3701 			nx_hblk = hmeblkp->hblk_next;
3702 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
3703 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3704 				    &list, 0);
3705 			} else {
3706 				pr_hblk = hmeblkp;
3707 			}
3708 			hmeblkp = nx_hblk;
3709 		}
3710 
3711 		SFMMU_HASH_UNLOCK(hmebp);
3712 
3713 		if (shadow) {
3714 			/*
3715 			 * We found another shadow hblk so cleaned its
3716 			 * children.  We need to go back and cleanup
3717 			 * the original hblk so we don't change the
3718 			 * addr.
3719 			 */
3720 			shadow = 0;
3721 		} else {
3722 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
3723 			    (1 << hmeshift));
3724 		}
3725 	}
3726 	sfmmu_hblks_list_purge(&list, 0);
3727 }
3728 
3729 /*
3730  * This routine's job is to delete stale invalid shared hmeregions hmeblks that
3731  * may still linger on after pageunload.
3732  */
3733 static void
3734 sfmmu_cleanup_rhblk(sf_srd_t *srdp, caddr_t addr, uint_t rid, int ttesz)
3735 {
3736 	int hmeshift;
3737 	hmeblk_tag hblktag;
3738 	struct hmehash_bucket *hmebp;
3739 	struct hme_blk *hmeblkp;
3740 	struct hme_blk *pr_hblk;
3741 	struct hme_blk *list = NULL;
3742 
3743 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3744 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3745 
3746 	hmeshift = HME_HASH_SHIFT(ttesz);
3747 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3748 	hblktag.htag_rehash = ttesz;
3749 	hblktag.htag_rid = rid;
3750 	hblktag.htag_id = srdp;
3751 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3752 
3753 	SFMMU_HASH_LOCK(hmebp);
3754 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3755 	if (hmeblkp != NULL) {
3756 		ASSERT(hmeblkp->hblk_shared);
3757 		ASSERT(!hmeblkp->hblk_shw_bit);
3758 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3759 			panic("sfmmu_cleanup_rhblk: valid hmeblk");
3760 		}
3761 		ASSERT(!hmeblkp->hblk_lckcnt);
3762 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3763 		    &list, 0);
3764 	}
3765 	SFMMU_HASH_UNLOCK(hmebp);
3766 	sfmmu_hblks_list_purge(&list, 0);
3767 }
3768 
3769 /* ARGSUSED */
3770 static void
3771 sfmmu_rgn_cb_noop(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
3772     size_t r_size, void *r_obj, u_offset_t r_objoff)
3773 {
3774 }
3775 
3776 /*
3777  * Searches for an hmeblk which maps addr, then unloads this mapping
3778  * and updates *eaddrp, if the hmeblk is found.
3779  */
3780 static void
3781 sfmmu_unload_hmeregion_va(sf_srd_t *srdp, uint_t rid, caddr_t addr,
3782     caddr_t eaddr, int ttesz, caddr_t *eaddrp)
3783 {
3784 	int hmeshift;
3785 	hmeblk_tag hblktag;
3786 	struct hmehash_bucket *hmebp;
3787 	struct hme_blk *hmeblkp;
3788 	struct hme_blk *pr_hblk;
3789 	struct hme_blk *list = NULL;
3790 
3791 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3792 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3793 	ASSERT(ttesz >= HBLK_MIN_TTESZ);
3794 
3795 	hmeshift = HME_HASH_SHIFT(ttesz);
3796 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3797 	hblktag.htag_rehash = ttesz;
3798 	hblktag.htag_rid = rid;
3799 	hblktag.htag_id = srdp;
3800 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3801 
3802 	SFMMU_HASH_LOCK(hmebp);
3803 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3804 	if (hmeblkp != NULL) {
3805 		ASSERT(hmeblkp->hblk_shared);
3806 		ASSERT(!hmeblkp->hblk_lckcnt);
3807 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3808 			*eaddrp = sfmmu_hblk_unload(NULL, hmeblkp, addr,
3809 			    eaddr, NULL, HAT_UNLOAD);
3810 			ASSERT(*eaddrp > addr);
3811 		}
3812 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3813 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3814 		    &list, 0);
3815 	}
3816 	SFMMU_HASH_UNLOCK(hmebp);
3817 	sfmmu_hblks_list_purge(&list, 0);
3818 }
3819 
3820 static void
3821 sfmmu_unload_hmeregion(sf_srd_t *srdp, sf_region_t *rgnp)
3822 {
3823 	int ttesz = rgnp->rgn_pgszc;
3824 	size_t rsz = rgnp->rgn_size;
3825 	caddr_t rsaddr = rgnp->rgn_saddr;
3826 	caddr_t readdr = rsaddr + rsz;
3827 	caddr_t rhsaddr;
3828 	caddr_t va;
3829 	uint_t rid = rgnp->rgn_id;
3830 	caddr_t cbsaddr;
3831 	caddr_t cbeaddr;
3832 	hat_rgn_cb_func_t rcbfunc;
3833 	ulong_t cnt;
3834 
3835 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3836 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3837 
3838 	ASSERT(IS_P2ALIGNED(rsaddr, TTEBYTES(ttesz)));
3839 	ASSERT(IS_P2ALIGNED(rsz, TTEBYTES(ttesz)));
3840 	if (ttesz < HBLK_MIN_TTESZ) {
3841 		ttesz = HBLK_MIN_TTESZ;
3842 		rhsaddr = (caddr_t)P2ALIGN((uintptr_t)rsaddr, HBLK_MIN_BYTES);
3843 	} else {
3844 		rhsaddr = rsaddr;
3845 	}
3846 
3847 	if ((rcbfunc = rgnp->rgn_cb_function) == NULL) {
3848 		rcbfunc = sfmmu_rgn_cb_noop;
3849 	}
3850 
3851 	while (ttesz >= HBLK_MIN_TTESZ) {
3852 		cbsaddr = rsaddr;
3853 		cbeaddr = rsaddr;
3854 		if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
3855 			ttesz--;
3856 			continue;
3857 		}
3858 		cnt = 0;
3859 		va = rsaddr;
3860 		while (va < readdr) {
3861 			ASSERT(va >= rhsaddr);
3862 			if (va != cbeaddr) {
3863 				if (cbeaddr != cbsaddr) {
3864 					ASSERT(cbeaddr > cbsaddr);
3865 					(*rcbfunc)(cbsaddr, cbeaddr,
3866 					    rsaddr, rsz, rgnp->rgn_obj,
3867 					    rgnp->rgn_objoff);
3868 				}
3869 				cbsaddr = va;
3870 				cbeaddr = va;
3871 			}
3872 			sfmmu_unload_hmeregion_va(srdp, rid, va, readdr,
3873 			    ttesz, &cbeaddr);
3874 			cnt++;
3875 			va = rhsaddr + (cnt << TTE_PAGE_SHIFT(ttesz));
3876 		}
3877 		if (cbeaddr != cbsaddr) {
3878 			ASSERT(cbeaddr > cbsaddr);
3879 			(*rcbfunc)(cbsaddr, cbeaddr, rsaddr,
3880 			    rsz, rgnp->rgn_obj,
3881 			    rgnp->rgn_objoff);
3882 		}
3883 		ttesz--;
3884 	}
3885 }
3886 
3887 /*
3888  * Release one hardware address translation lock on the given address range.
3889  */
3890 void
3891 hat_unlock(struct hat *sfmmup, caddr_t addr, size_t len)
3892 {
3893 	struct hmehash_bucket *hmebp;
3894 	hmeblk_tag hblktag;
3895 	int hmeshift, hashno = 1;
3896 	struct hme_blk *hmeblkp, *list = NULL;
3897 	caddr_t endaddr;
3898 
3899 	ASSERT(sfmmup != NULL);
3900 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3901 
3902 	ASSERT((sfmmup == ksfmmup) ||
3903 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
3904 	ASSERT((len & MMU_PAGEOFFSET) == 0);
3905 	endaddr = addr + len;
3906 	hblktag.htag_id = sfmmup;
3907 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3908 
3909 	/*
3910 	 * Spitfire supports 4 page sizes.
3911 	 * Most pages are expected to be of the smallest page size (8K) and
3912 	 * these will not need to be rehashed. 64K pages also don't need to be
3913 	 * rehashed because an hmeblk spans 64K of address space. 512K pages
3914 	 * might need 1 rehash and and 4M pages might need 2 rehashes.
3915 	 */
3916 	while (addr < endaddr) {
3917 		hmeshift = HME_HASH_SHIFT(hashno);
3918 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3919 		hblktag.htag_rehash = hashno;
3920 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3921 
3922 		SFMMU_HASH_LOCK(hmebp);
3923 
3924 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
3925 		if (hmeblkp != NULL) {
3926 			ASSERT(!hmeblkp->hblk_shared);
3927 			/*
3928 			 * If we encounter a shadow hmeblk then
3929 			 * we know there are no valid hmeblks mapping
3930 			 * this address at this size or larger.
3931 			 * Just increment address by the smallest
3932 			 * page size.
3933 			 */
3934 			if (hmeblkp->hblk_shw_bit) {
3935 				addr += MMU_PAGESIZE;
3936 			} else {
3937 				addr = sfmmu_hblk_unlock(hmeblkp, addr,
3938 				    endaddr);
3939 			}
3940 			SFMMU_HASH_UNLOCK(hmebp);
3941 			hashno = 1;
3942 			continue;
3943 		}
3944 		SFMMU_HASH_UNLOCK(hmebp);
3945 
3946 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
3947 			/*
3948 			 * We have traversed the whole list and rehashed
3949 			 * if necessary without finding the address to unlock
3950 			 * which should never happen.
3951 			 */
3952 			panic("sfmmu_unlock: addr not found. "
3953 			    "addr %p hat %p", (void *)addr, (void *)sfmmup);
3954 		} else {
3955 			hashno++;
3956 		}
3957 	}
3958 
3959 	sfmmu_hblks_list_purge(&list, 0);
3960 }
3961 
3962 void
3963 hat_unlock_region(struct hat *sfmmup, caddr_t addr, size_t len,
3964     hat_region_cookie_t rcookie)
3965 {
3966 	sf_srd_t *srdp;
3967 	sf_region_t *rgnp;
3968 	int ttesz;
3969 	uint_t rid;
3970 	caddr_t eaddr;
3971 	caddr_t va;
3972 	int hmeshift;
3973 	hmeblk_tag hblktag;
3974 	struct hmehash_bucket *hmebp;
3975 	struct hme_blk *hmeblkp;
3976 	struct hme_blk *pr_hblk;
3977 	struct hme_blk *list;
3978 
3979 	if (rcookie == HAT_INVALID_REGION_COOKIE) {
3980 		hat_unlock(sfmmup, addr, len);
3981 		return;
3982 	}
3983 
3984 	ASSERT(sfmmup != NULL);
3985 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3986 	ASSERT(sfmmup != ksfmmup);
3987 
3988 	srdp = sfmmup->sfmmu_srdp;
3989 	rid = (uint_t)((uint64_t)rcookie);
3990 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3991 	eaddr = addr + len;
3992 	va = addr;
3993 	list = NULL;
3994 	rgnp = srdp->srd_hmergnp[rid];
3995 	SFMMU_VALIDATE_HMERID(sfmmup, rid, addr, len);
3996 
3997 	ASSERT(IS_P2ALIGNED(addr, TTEBYTES(rgnp->rgn_pgszc)));
3998 	ASSERT(IS_P2ALIGNED(len, TTEBYTES(rgnp->rgn_pgszc)));
3999 	if (rgnp->rgn_pgszc < HBLK_MIN_TTESZ) {
4000 		ttesz = HBLK_MIN_TTESZ;
4001 	} else {
4002 		ttesz = rgnp->rgn_pgszc;
4003 	}
4004 	while (va < eaddr) {
4005 		while (ttesz < rgnp->rgn_pgszc &&
4006 		    IS_P2ALIGNED(va, TTEBYTES(ttesz + 1))) {
4007 			ttesz++;
4008 		}
4009 		while (ttesz >= HBLK_MIN_TTESZ) {
4010 			if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
4011 				ttesz--;
4012 				continue;
4013 			}
4014 			hmeshift = HME_HASH_SHIFT(ttesz);
4015 			hblktag.htag_bspage = HME_HASH_BSPAGE(va, hmeshift);
4016 			hblktag.htag_rehash = ttesz;
4017 			hblktag.htag_rid = rid;
4018 			hblktag.htag_id = srdp;
4019 			hmebp = HME_HASH_FUNCTION(srdp, va, hmeshift);
4020 			SFMMU_HASH_LOCK(hmebp);
4021 			HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk,
4022 			    &list);
4023 			if (hmeblkp == NULL) {
4024 				SFMMU_HASH_UNLOCK(hmebp);
4025 				ttesz--;
4026 				continue;
4027 			}
4028 			ASSERT(hmeblkp->hblk_shared);
4029 			va = sfmmu_hblk_unlock(hmeblkp, va, eaddr);
4030 			ASSERT(va >= eaddr ||
4031 			    IS_P2ALIGNED((uintptr_t)va, TTEBYTES(ttesz)));
4032 			SFMMU_HASH_UNLOCK(hmebp);
4033 			break;
4034 		}
4035 		if (ttesz < HBLK_MIN_TTESZ) {
4036 			panic("hat_unlock_region: addr not found "
4037 			    "addr %p hat %p", (void *)va, (void *)sfmmup);
4038 		}
4039 	}
4040 	sfmmu_hblks_list_purge(&list, 0);
4041 }
4042 
4043 /*
4044  * Function to unlock a range of addresses in an hmeblk.  It returns the
4045  * next address that needs to be unlocked.
4046  * Should be called with the hash lock held.
4047  */
4048 static caddr_t
4049 sfmmu_hblk_unlock(struct hme_blk *hmeblkp, caddr_t addr, caddr_t endaddr)
4050 {
4051 	struct sf_hment *sfhme;
4052 	tte_t tteold, ttemod;
4053 	int ttesz, ret;
4054 
4055 	ASSERT(in_hblk_range(hmeblkp, addr));
4056 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4057 
4058 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4059 	ttesz = get_hblk_ttesz(hmeblkp);
4060 
4061 	HBLKTOHME(sfhme, hmeblkp, addr);
4062 	while (addr < endaddr) {
4063 readtte:
4064 		sfmmu_copytte(&sfhme->hme_tte, &tteold);
4065 		if (TTE_IS_VALID(&tteold)) {
4066 
4067 			ttemod = tteold;
4068 
4069 			ret = sfmmu_modifytte_try(&tteold, &ttemod,
4070 			    &sfhme->hme_tte);
4071 
4072 			if (ret < 0)
4073 				goto readtte;
4074 
4075 			if (hmeblkp->hblk_lckcnt == 0)
4076 				panic("zero hblk lckcnt");
4077 
4078 			if (((uintptr_t)addr + TTEBYTES(ttesz)) >
4079 			    (uintptr_t)endaddr)
4080 				panic("can't unlock large tte");
4081 
4082 			ASSERT(hmeblkp->hblk_lckcnt > 0);
4083 			atomic_add_32(&hmeblkp->hblk_lckcnt, -1);
4084 			HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
4085 		} else {
4086 			panic("sfmmu_hblk_unlock: invalid tte");
4087 		}
4088 		addr += TTEBYTES(ttesz);
4089 		sfhme++;
4090 	}
4091 	return (addr);
4092 }
4093 
4094 /*
4095  * Physical Address Mapping Framework
4096  *
4097  * General rules:
4098  *
4099  * (1) Applies only to seg_kmem memory pages. To make things easier,
4100  *     seg_kpm addresses are also accepted by the routines, but nothing
4101  *     is done with them since by definition their PA mappings are static.
4102  * (2) hat_add_callback() may only be called while holding the page lock
4103  *     SE_SHARED or SE_EXCL of the underlying page (e.g., as_pagelock()),
4104  *     or passing HAC_PAGELOCK flag.
4105  * (3) prehandler() and posthandler() may not call hat_add_callback() or
4106  *     hat_delete_callback(), nor should they allocate memory. Post quiesce
4107  *     callbacks may not sleep or acquire adaptive mutex locks.
4108  * (4) Either prehandler() or posthandler() (but not both) may be specified
4109  *     as being NULL.  Specifying an errhandler() is optional.
4110  *
4111  * Details of using the framework:
4112  *
4113  * registering a callback (hat_register_callback())
4114  *
4115  *	Pass prehandler, posthandler, errhandler addresses
4116  *	as described below. If capture_cpus argument is nonzero,
4117  *	suspend callback to the prehandler will occur with CPUs
4118  *	captured and executing xc_loop() and CPUs will remain
4119  *	captured until after the posthandler suspend callback
4120  *	occurs.
4121  *
4122  * adding a callback (hat_add_callback())
4123  *
4124  *      as_pagelock();
4125  *	hat_add_callback();
4126  *      save returned pfn in private data structures or program registers;
4127  *      as_pageunlock();
4128  *
4129  * prehandler()
4130  *
4131  *	Stop all accesses by physical address to this memory page.
4132  *	Called twice: the first, PRESUSPEND, is a context safe to acquire
4133  *	adaptive locks. The second, SUSPEND, is called at high PIL with
4134  *	CPUs captured so adaptive locks may NOT be acquired (and all spin
4135  *	locks must be XCALL_PIL or higher locks).
4136  *
4137  *	May return the following errors:
4138  *		EIO:	A fatal error has occurred. This will result in panic.
4139  *		EAGAIN:	The page cannot be suspended. This will fail the
4140  *			relocation.
4141  *		0:	Success.
4142  *
4143  * posthandler()
4144  *
4145  *      Save new pfn in private data structures or program registers;
4146  *	not allowed to fail (non-zero return values will result in panic).
4147  *
4148  * errhandler()
4149  *
4150  *	called when an error occurs related to the callback.  Currently
4151  *	the only such error is HAT_CB_ERR_LEAKED which indicates that
4152  *	a page is being freed, but there are still outstanding callback(s)
4153  *	registered on the page.
4154  *
4155  * removing a callback (hat_delete_callback(); e.g., prior to freeing memory)
4156  *
4157  *	stop using physical address
4158  *	hat_delete_callback();
4159  *
4160  */
4161 
4162 /*
4163  * Register a callback class.  Each subsystem should do this once and
4164  * cache the id_t returned for use in setting up and tearing down callbacks.
4165  *
4166  * There is no facility for removing callback IDs once they are created;
4167  * the "key" should be unique for each module, so in case a module is unloaded
4168  * and subsequently re-loaded, we can recycle the module's previous entry.
4169  */
4170 id_t
4171 hat_register_callback(int key,
4172 	int (*prehandler)(caddr_t, uint_t, uint_t, void *),
4173 	int (*posthandler)(caddr_t, uint_t, uint_t, void *, pfn_t),
4174 	int (*errhandler)(caddr_t, uint_t, uint_t, void *),
4175 	int capture_cpus)
4176 {
4177 	id_t id;
4178 
4179 	/*
4180 	 * Search the table for a pre-existing callback associated with
4181 	 * the identifier "key".  If one exists, we re-use that entry in
4182 	 * the table for this instance, otherwise we assign the next
4183 	 * available table slot.
4184 	 */
4185 	for (id = 0; id < sfmmu_max_cb_id; id++) {
4186 		if (sfmmu_cb_table[id].key == key)
4187 			break;
4188 	}
4189 
4190 	if (id == sfmmu_max_cb_id) {
4191 		id = sfmmu_cb_nextid++;
4192 		if (id >= sfmmu_max_cb_id)
4193 			panic("hat_register_callback: out of callback IDs");
4194 	}
4195 
4196 	ASSERT(prehandler != NULL || posthandler != NULL);
4197 
4198 	sfmmu_cb_table[id].key = key;
4199 	sfmmu_cb_table[id].prehandler = prehandler;
4200 	sfmmu_cb_table[id].posthandler = posthandler;
4201 	sfmmu_cb_table[id].errhandler = errhandler;
4202 	sfmmu_cb_table[id].capture_cpus = capture_cpus;
4203 
4204 	return (id);
4205 }
4206 
4207 #define	HAC_COOKIE_NONE	(void *)-1
4208 
4209 /*
4210  * Add relocation callbacks to the specified addr/len which will be called
4211  * when relocating the associated page. See the description of pre and
4212  * posthandler above for more details.
4213  *
4214  * If HAC_PAGELOCK is included in flags, the underlying memory page is
4215  * locked internally so the caller must be able to deal with the callback
4216  * running even before this function has returned.  If HAC_PAGELOCK is not
4217  * set, it is assumed that the underlying memory pages are locked.
4218  *
4219  * Since the caller must track the individual page boundaries anyway,
4220  * we only allow a callback to be added to a single page (large
4221  * or small).  Thus [addr, addr + len) MUST be contained within a single
4222  * page.
4223  *
4224  * Registering multiple callbacks on the same [addr, addr+len) is supported,
4225  * _provided_that_ a unique parameter is specified for each callback.
4226  * If multiple callbacks are registered on the same range the callback will
4227  * be invoked with each unique parameter. Registering the same callback with
4228  * the same argument more than once will result in corrupted kernel state.
4229  *
4230  * Returns the pfn of the underlying kernel page in *rpfn
4231  * on success, or PFN_INVALID on failure.
4232  *
4233  * cookiep (if passed) provides storage space for an opaque cookie
4234  * to return later to hat_delete_callback(). This cookie makes the callback
4235  * deletion significantly quicker by avoiding a potentially lengthy hash
4236  * search.
4237  *
4238  * Returns values:
4239  *    0:      success
4240  *    ENOMEM: memory allocation failure (e.g. flags was passed as HAC_NOSLEEP)
4241  *    EINVAL: callback ID is not valid
4242  *    ENXIO:  ["vaddr", "vaddr" + len) is not mapped in the kernel's address
4243  *            space
4244  *    ERANGE: ["vaddr", "vaddr" + len) crosses a page boundary
4245  */
4246 int
4247 hat_add_callback(id_t callback_id, caddr_t vaddr, uint_t len, uint_t flags,
4248 	void *pvt, pfn_t *rpfn, void **cookiep)
4249 {
4250 	struct 		hmehash_bucket *hmebp;
4251 	hmeblk_tag 	hblktag;
4252 	struct hme_blk	*hmeblkp;
4253 	int 		hmeshift, hashno;
4254 	caddr_t 	saddr, eaddr, baseaddr;
4255 	struct pa_hment *pahmep;
4256 	struct sf_hment *sfhmep, *osfhmep;
4257 	kmutex_t	*pml;
4258 	tte_t   	tte;
4259 	page_t		*pp;
4260 	vnode_t		*vp;
4261 	u_offset_t	off;
4262 	pfn_t		pfn;
4263 	int		kmflags = (flags & HAC_SLEEP)? KM_SLEEP : KM_NOSLEEP;
4264 	int		locked = 0;
4265 
4266 	/*
4267 	 * For KPM mappings, just return the physical address since we
4268 	 * don't need to register any callbacks.
4269 	 */
4270 	if (IS_KPM_ADDR(vaddr)) {
4271 		uint64_t paddr;
4272 		SFMMU_KPM_VTOP(vaddr, paddr);
4273 		*rpfn = btop(paddr);
4274 		if (cookiep != NULL)
4275 			*cookiep = HAC_COOKIE_NONE;
4276 		return (0);
4277 	}
4278 
4279 	if (callback_id < (id_t)0 || callback_id >= sfmmu_cb_nextid) {
4280 		*rpfn = PFN_INVALID;
4281 		return (EINVAL);
4282 	}
4283 
4284 	if ((pahmep = kmem_cache_alloc(pa_hment_cache, kmflags)) == NULL) {
4285 		*rpfn = PFN_INVALID;
4286 		return (ENOMEM);
4287 	}
4288 
4289 	sfhmep = &pahmep->sfment;
4290 
4291 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4292 	eaddr = saddr + len;
4293 
4294 rehash:
4295 	/* Find the mapping(s) for this page */
4296 	for (hashno = TTE64K, hmeblkp = NULL;
4297 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4298 	    hashno++) {
4299 		hmeshift = HME_HASH_SHIFT(hashno);
4300 		hblktag.htag_id = ksfmmup;
4301 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4302 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4303 		hblktag.htag_rehash = hashno;
4304 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4305 
4306 		SFMMU_HASH_LOCK(hmebp);
4307 
4308 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4309 
4310 		if (hmeblkp == NULL)
4311 			SFMMU_HASH_UNLOCK(hmebp);
4312 	}
4313 
4314 	if (hmeblkp == NULL) {
4315 		kmem_cache_free(pa_hment_cache, pahmep);
4316 		*rpfn = PFN_INVALID;
4317 		return (ENXIO);
4318 	}
4319 
4320 	ASSERT(!hmeblkp->hblk_shared);
4321 
4322 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4323 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4324 
4325 	if (!TTE_IS_VALID(&tte)) {
4326 		SFMMU_HASH_UNLOCK(hmebp);
4327 		kmem_cache_free(pa_hment_cache, pahmep);
4328 		*rpfn = PFN_INVALID;
4329 		return (ENXIO);
4330 	}
4331 
4332 	/*
4333 	 * Make sure the boundaries for the callback fall within this
4334 	 * single mapping.
4335 	 */
4336 	baseaddr = (caddr_t)get_hblk_base(hmeblkp);
4337 	ASSERT(saddr >= baseaddr);
4338 	if (eaddr > saddr + TTEBYTES(TTE_CSZ(&tte))) {
4339 		SFMMU_HASH_UNLOCK(hmebp);
4340 		kmem_cache_free(pa_hment_cache, pahmep);
4341 		*rpfn = PFN_INVALID;
4342 		return (ERANGE);
4343 	}
4344 
4345 	pfn = sfmmu_ttetopfn(&tte, vaddr);
4346 
4347 	/*
4348 	 * The pfn may not have a page_t underneath in which case we
4349 	 * just return it. This can happen if we are doing I/O to a
4350 	 * static portion of the kernel's address space, for instance.
4351 	 */
4352 	pp = osfhmep->hme_page;
4353 	if (pp == NULL) {
4354 		SFMMU_HASH_UNLOCK(hmebp);
4355 		kmem_cache_free(pa_hment_cache, pahmep);
4356 		*rpfn = pfn;
4357 		if (cookiep)
4358 			*cookiep = HAC_COOKIE_NONE;
4359 		return (0);
4360 	}
4361 	ASSERT(pp == PP_PAGEROOT(pp));
4362 
4363 	vp = pp->p_vnode;
4364 	off = pp->p_offset;
4365 
4366 	pml = sfmmu_mlist_enter(pp);
4367 
4368 	if (flags & HAC_PAGELOCK) {
4369 		if (!page_trylock(pp, SE_SHARED)) {
4370 			/*
4371 			 * Somebody is holding SE_EXCL lock. Might
4372 			 * even be hat_page_relocate(). Drop all
4373 			 * our locks, lookup the page in &kvp, and
4374 			 * retry. If it doesn't exist in &kvp and &zvp,
4375 			 * then we must be dealing with a kernel mapped
4376 			 * page which doesn't actually belong to
4377 			 * segkmem so we punt.
4378 			 */
4379 			sfmmu_mlist_exit(pml);
4380 			SFMMU_HASH_UNLOCK(hmebp);
4381 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4382 
4383 			/* check zvp before giving up */
4384 			if (pp == NULL)
4385 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4386 				    SE_SHARED);
4387 
4388 			/* Okay, we didn't find it, give up */
4389 			if (pp == NULL) {
4390 				kmem_cache_free(pa_hment_cache, pahmep);
4391 				*rpfn = pfn;
4392 				if (cookiep)
4393 					*cookiep = HAC_COOKIE_NONE;
4394 				return (0);
4395 			}
4396 			page_unlock(pp);
4397 			goto rehash;
4398 		}
4399 		locked = 1;
4400 	}
4401 
4402 	if (!PAGE_LOCKED(pp) && !panicstr)
4403 		panic("hat_add_callback: page 0x%p not locked", (void *)pp);
4404 
4405 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4406 	    pp->p_offset != off) {
4407 		/*
4408 		 * The page moved before we got our hands on it.  Drop
4409 		 * all the locks and try again.
4410 		 */
4411 		ASSERT((flags & HAC_PAGELOCK) != 0);
4412 		sfmmu_mlist_exit(pml);
4413 		SFMMU_HASH_UNLOCK(hmebp);
4414 		page_unlock(pp);
4415 		locked = 0;
4416 		goto rehash;
4417 	}
4418 
4419 	if (!VN_ISKAS(vp)) {
4420 		/*
4421 		 * This is not a segkmem page but another page which
4422 		 * has been kernel mapped. It had better have at least
4423 		 * a share lock on it. Return the pfn.
4424 		 */
4425 		sfmmu_mlist_exit(pml);
4426 		SFMMU_HASH_UNLOCK(hmebp);
4427 		if (locked)
4428 			page_unlock(pp);
4429 		kmem_cache_free(pa_hment_cache, pahmep);
4430 		ASSERT(PAGE_LOCKED(pp));
4431 		*rpfn = pfn;
4432 		if (cookiep)
4433 			*cookiep = HAC_COOKIE_NONE;
4434 		return (0);
4435 	}
4436 
4437 	/*
4438 	 * Setup this pa_hment and link its embedded dummy sf_hment into
4439 	 * the mapping list.
4440 	 */
4441 	pp->p_share++;
4442 	pahmep->cb_id = callback_id;
4443 	pahmep->addr = vaddr;
4444 	pahmep->len = len;
4445 	pahmep->refcnt = 1;
4446 	pahmep->flags = 0;
4447 	pahmep->pvt = pvt;
4448 
4449 	sfhmep->hme_tte.ll = 0;
4450 	sfhmep->hme_data = pahmep;
4451 	sfhmep->hme_prev = osfhmep;
4452 	sfhmep->hme_next = osfhmep->hme_next;
4453 
4454 	if (osfhmep->hme_next)
4455 		osfhmep->hme_next->hme_prev = sfhmep;
4456 
4457 	osfhmep->hme_next = sfhmep;
4458 
4459 	sfmmu_mlist_exit(pml);
4460 	SFMMU_HASH_UNLOCK(hmebp);
4461 
4462 	if (locked)
4463 		page_unlock(pp);
4464 
4465 	*rpfn = pfn;
4466 	if (cookiep)
4467 		*cookiep = (void *)pahmep;
4468 
4469 	return (0);
4470 }
4471 
4472 /*
4473  * Remove the relocation callbacks from the specified addr/len.
4474  */
4475 void
4476 hat_delete_callback(caddr_t vaddr, uint_t len, void *pvt, uint_t flags,
4477 	void *cookie)
4478 {
4479 	struct		hmehash_bucket *hmebp;
4480 	hmeblk_tag	hblktag;
4481 	struct hme_blk	*hmeblkp;
4482 	int		hmeshift, hashno;
4483 	caddr_t		saddr;
4484 	struct pa_hment	*pahmep;
4485 	struct sf_hment	*sfhmep, *osfhmep;
4486 	kmutex_t	*pml;
4487 	tte_t		tte;
4488 	page_t		*pp;
4489 	vnode_t		*vp;
4490 	u_offset_t	off;
4491 	int		locked = 0;
4492 
4493 	/*
4494 	 * If the cookie is HAC_COOKIE_NONE then there is no pa_hment to
4495 	 * remove so just return.
4496 	 */
4497 	if (cookie == HAC_COOKIE_NONE || IS_KPM_ADDR(vaddr))
4498 		return;
4499 
4500 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4501 
4502 rehash:
4503 	/* Find the mapping(s) for this page */
4504 	for (hashno = TTE64K, hmeblkp = NULL;
4505 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4506 	    hashno++) {
4507 		hmeshift = HME_HASH_SHIFT(hashno);
4508 		hblktag.htag_id = ksfmmup;
4509 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4510 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4511 		hblktag.htag_rehash = hashno;
4512 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4513 
4514 		SFMMU_HASH_LOCK(hmebp);
4515 
4516 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4517 
4518 		if (hmeblkp == NULL)
4519 			SFMMU_HASH_UNLOCK(hmebp);
4520 	}
4521 
4522 	if (hmeblkp == NULL)
4523 		return;
4524 
4525 	ASSERT(!hmeblkp->hblk_shared);
4526 
4527 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4528 
4529 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4530 	if (!TTE_IS_VALID(&tte)) {
4531 		SFMMU_HASH_UNLOCK(hmebp);
4532 		return;
4533 	}
4534 
4535 	pp = osfhmep->hme_page;
4536 	if (pp == NULL) {
4537 		SFMMU_HASH_UNLOCK(hmebp);
4538 		ASSERT(cookie == NULL);
4539 		return;
4540 	}
4541 
4542 	vp = pp->p_vnode;
4543 	off = pp->p_offset;
4544 
4545 	pml = sfmmu_mlist_enter(pp);
4546 
4547 	if (flags & HAC_PAGELOCK) {
4548 		if (!page_trylock(pp, SE_SHARED)) {
4549 			/*
4550 			 * Somebody is holding SE_EXCL lock. Might
4551 			 * even be hat_page_relocate(). Drop all
4552 			 * our locks, lookup the page in &kvp, and
4553 			 * retry. If it doesn't exist in &kvp and &zvp,
4554 			 * then we must be dealing with a kernel mapped
4555 			 * page which doesn't actually belong to
4556 			 * segkmem so we punt.
4557 			 */
4558 			sfmmu_mlist_exit(pml);
4559 			SFMMU_HASH_UNLOCK(hmebp);
4560 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4561 			/* check zvp before giving up */
4562 			if (pp == NULL)
4563 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4564 				    SE_SHARED);
4565 
4566 			if (pp == NULL) {
4567 				ASSERT(cookie == NULL);
4568 				return;
4569 			}
4570 			page_unlock(pp);
4571 			goto rehash;
4572 		}
4573 		locked = 1;
4574 	}
4575 
4576 	ASSERT(PAGE_LOCKED(pp));
4577 
4578 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4579 	    pp->p_offset != off) {
4580 		/*
4581 		 * The page moved before we got our hands on it.  Drop
4582 		 * all the locks and try again.
4583 		 */
4584 		ASSERT((flags & HAC_PAGELOCK) != 0);
4585 		sfmmu_mlist_exit(pml);
4586 		SFMMU_HASH_UNLOCK(hmebp);
4587 		page_unlock(pp);
4588 		locked = 0;
4589 		goto rehash;
4590 	}
4591 
4592 	if (!VN_ISKAS(vp)) {
4593 		/*
4594 		 * This is not a segkmem page but another page which
4595 		 * has been kernel mapped.
4596 		 */
4597 		sfmmu_mlist_exit(pml);
4598 		SFMMU_HASH_UNLOCK(hmebp);
4599 		if (locked)
4600 			page_unlock(pp);
4601 		ASSERT(cookie == NULL);
4602 		return;
4603 	}
4604 
4605 	if (cookie != NULL) {
4606 		pahmep = (struct pa_hment *)cookie;
4607 		sfhmep = &pahmep->sfment;
4608 	} else {
4609 		for (sfhmep = pp->p_mapping; sfhmep != NULL;
4610 		    sfhmep = sfhmep->hme_next) {
4611 
4612 			/*
4613 			 * skip va<->pa mappings
4614 			 */
4615 			if (!IS_PAHME(sfhmep))
4616 				continue;
4617 
4618 			pahmep = sfhmep->hme_data;
4619 			ASSERT(pahmep != NULL);
4620 
4621 			/*
4622 			 * if pa_hment matches, remove it
4623 			 */
4624 			if ((pahmep->pvt == pvt) &&
4625 			    (pahmep->addr == vaddr) &&
4626 			    (pahmep->len == len)) {
4627 				break;
4628 			}
4629 		}
4630 	}
4631 
4632 	if (sfhmep == NULL) {
4633 		if (!panicstr) {
4634 			panic("hat_delete_callback: pa_hment not found, pp %p",
4635 			    (void *)pp);
4636 		}
4637 		return;
4638 	}
4639 
4640 	/*
4641 	 * Note: at this point a valid kernel mapping must still be
4642 	 * present on this page.
4643 	 */
4644 	pp->p_share--;
4645 	if (pp->p_share <= 0)
4646 		panic("hat_delete_callback: zero p_share");
4647 
4648 	if (--pahmep->refcnt == 0) {
4649 		if (pahmep->flags != 0)
4650 			panic("hat_delete_callback: pa_hment is busy");
4651 
4652 		/*
4653 		 * Remove sfhmep from the mapping list for the page.
4654 		 */
4655 		if (sfhmep->hme_prev) {
4656 			sfhmep->hme_prev->hme_next = sfhmep->hme_next;
4657 		} else {
4658 			pp->p_mapping = sfhmep->hme_next;
4659 		}
4660 
4661 		if (sfhmep->hme_next)
4662 			sfhmep->hme_next->hme_prev = sfhmep->hme_prev;
4663 
4664 		sfmmu_mlist_exit(pml);
4665 		SFMMU_HASH_UNLOCK(hmebp);
4666 
4667 		if (locked)
4668 			page_unlock(pp);
4669 
4670 		kmem_cache_free(pa_hment_cache, pahmep);
4671 		return;
4672 	}
4673 
4674 	sfmmu_mlist_exit(pml);
4675 	SFMMU_HASH_UNLOCK(hmebp);
4676 	if (locked)
4677 		page_unlock(pp);
4678 }
4679 
4680 /*
4681  * hat_probe returns 1 if the translation for the address 'addr' is
4682  * loaded, zero otherwise.
4683  *
4684  * hat_probe should be used only for advisorary purposes because it may
4685  * occasionally return the wrong value. The implementation must guarantee that
4686  * returning the wrong value is a very rare event. hat_probe is used
4687  * to implement optimizations in the segment drivers.
4688  *
4689  */
4690 int
4691 hat_probe(struct hat *sfmmup, caddr_t addr)
4692 {
4693 	pfn_t pfn;
4694 	tte_t tte;
4695 
4696 	ASSERT(sfmmup != NULL);
4697 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4698 
4699 	ASSERT((sfmmup == ksfmmup) ||
4700 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
4701 
4702 	if (sfmmup == ksfmmup) {
4703 		while ((pfn = sfmmu_vatopfn(addr, sfmmup, &tte))
4704 		    == PFN_SUSPENDED) {
4705 			sfmmu_vatopfn_suspended(addr, sfmmup, &tte);
4706 		}
4707 	} else {
4708 		pfn = sfmmu_uvatopfn(addr, sfmmup, NULL);
4709 	}
4710 
4711 	if (pfn != PFN_INVALID)
4712 		return (1);
4713 	else
4714 		return (0);
4715 }
4716 
4717 ssize_t
4718 hat_getpagesize(struct hat *sfmmup, caddr_t addr)
4719 {
4720 	tte_t tte;
4721 
4722 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4723 
4724 	if (sfmmup == ksfmmup) {
4725 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4726 			return (-1);
4727 		}
4728 	} else {
4729 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4730 			return (-1);
4731 		}
4732 	}
4733 
4734 	ASSERT(TTE_IS_VALID(&tte));
4735 	return (TTEBYTES(TTE_CSZ(&tte)));
4736 }
4737 
4738 uint_t
4739 hat_getattr(struct hat *sfmmup, caddr_t addr, uint_t *attr)
4740 {
4741 	tte_t tte;
4742 
4743 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4744 
4745 	if (sfmmup == ksfmmup) {
4746 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4747 			tte.ll = 0;
4748 		}
4749 	} else {
4750 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4751 			tte.ll = 0;
4752 		}
4753 	}
4754 	if (TTE_IS_VALID(&tte)) {
4755 		*attr = sfmmu_ptov_attr(&tte);
4756 		return (0);
4757 	}
4758 	*attr = 0;
4759 	return ((uint_t)0xffffffff);
4760 }
4761 
4762 /*
4763  * Enables more attributes on specified address range (ie. logical OR)
4764  */
4765 void
4766 hat_setattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4767 {
4768 	if (hat->sfmmu_xhat_provider) {
4769 		XHAT_SETATTR(hat, addr, len, attr);
4770 		return;
4771 	} else {
4772 		/*
4773 		 * This must be a CPU HAT. If the address space has
4774 		 * XHATs attached, change attributes for all of them,
4775 		 * just in case
4776 		 */
4777 		ASSERT(hat->sfmmu_as != NULL);
4778 		if (hat->sfmmu_as->a_xhat != NULL)
4779 			xhat_setattr_all(hat->sfmmu_as, addr, len, attr);
4780 	}
4781 
4782 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_SETATTR);
4783 }
4784 
4785 /*
4786  * Assigns attributes to the specified address range.  All the attributes
4787  * are specified.
4788  */
4789 void
4790 hat_chgattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4791 {
4792 	if (hat->sfmmu_xhat_provider) {
4793 		XHAT_CHGATTR(hat, addr, len, attr);
4794 		return;
4795 	} else {
4796 		/*
4797 		 * This must be a CPU HAT. If the address space has
4798 		 * XHATs attached, change attributes for all of them,
4799 		 * just in case
4800 		 */
4801 		ASSERT(hat->sfmmu_as != NULL);
4802 		if (hat->sfmmu_as->a_xhat != NULL)
4803 			xhat_chgattr_all(hat->sfmmu_as, addr, len, attr);
4804 	}
4805 
4806 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CHGATTR);
4807 }
4808 
4809 /*
4810  * Remove attributes on the specified address range (ie. loginal NAND)
4811  */
4812 void
4813 hat_clrattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4814 {
4815 	if (hat->sfmmu_xhat_provider) {
4816 		XHAT_CLRATTR(hat, addr, len, attr);
4817 		return;
4818 	} else {
4819 		/*
4820 		 * This must be a CPU HAT. If the address space has
4821 		 * XHATs attached, change attributes for all of them,
4822 		 * just in case
4823 		 */
4824 		ASSERT(hat->sfmmu_as != NULL);
4825 		if (hat->sfmmu_as->a_xhat != NULL)
4826 			xhat_clrattr_all(hat->sfmmu_as, addr, len, attr);
4827 	}
4828 
4829 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CLRATTR);
4830 }
4831 
4832 /*
4833  * Change attributes on an address range to that specified by attr and mode.
4834  */
4835 static void
4836 sfmmu_chgattr(struct hat *sfmmup, caddr_t addr, size_t len, uint_t attr,
4837 	int mode)
4838 {
4839 	struct hmehash_bucket *hmebp;
4840 	hmeblk_tag hblktag;
4841 	int hmeshift, hashno = 1;
4842 	struct hme_blk *hmeblkp, *list = NULL;
4843 	caddr_t endaddr;
4844 	cpuset_t cpuset;
4845 	demap_range_t dmr;
4846 
4847 	CPUSET_ZERO(cpuset);
4848 
4849 	ASSERT((sfmmup == ksfmmup) ||
4850 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
4851 	ASSERT((len & MMU_PAGEOFFSET) == 0);
4852 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
4853 
4854 	if ((attr & PROT_USER) && (mode != SFMMU_CLRATTR) &&
4855 	    ((addr + len) > (caddr_t)USERLIMIT)) {
4856 		panic("user addr %p in kernel space",
4857 		    (void *)addr);
4858 	}
4859 
4860 	endaddr = addr + len;
4861 	hblktag.htag_id = sfmmup;
4862 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4863 	DEMAP_RANGE_INIT(sfmmup, &dmr);
4864 
4865 	while (addr < endaddr) {
4866 		hmeshift = HME_HASH_SHIFT(hashno);
4867 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
4868 		hblktag.htag_rehash = hashno;
4869 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
4870 
4871 		SFMMU_HASH_LOCK(hmebp);
4872 
4873 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
4874 		if (hmeblkp != NULL) {
4875 			ASSERT(!hmeblkp->hblk_shared);
4876 			/*
4877 			 * We've encountered a shadow hmeblk so skip the range
4878 			 * of the next smaller mapping size.
4879 			 */
4880 			if (hmeblkp->hblk_shw_bit) {
4881 				ASSERT(sfmmup != ksfmmup);
4882 				ASSERT(hashno > 1);
4883 				addr = (caddr_t)P2END((uintptr_t)addr,
4884 				    TTEBYTES(hashno - 1));
4885 			} else {
4886 				addr = sfmmu_hblk_chgattr(sfmmup,
4887 				    hmeblkp, addr, endaddr, &dmr, attr, mode);
4888 			}
4889 			SFMMU_HASH_UNLOCK(hmebp);
4890 			hashno = 1;
4891 			continue;
4892 		}
4893 		SFMMU_HASH_UNLOCK(hmebp);
4894 
4895 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
4896 			/*
4897 			 * We have traversed the whole list and rehashed
4898 			 * if necessary without finding the address to chgattr.
4899 			 * This is ok, so we increment the address by the
4900 			 * smallest hmeblk range for kernel mappings or for
4901 			 * user mappings with no large pages, and the largest
4902 			 * hmeblk range, to account for shadow hmeblks, for
4903 			 * user mappings with large pages and continue.
4904 			 */
4905 			if (sfmmup == ksfmmup)
4906 				addr = (caddr_t)P2END((uintptr_t)addr,
4907 				    TTEBYTES(1));
4908 			else
4909 				addr = (caddr_t)P2END((uintptr_t)addr,
4910 				    TTEBYTES(hashno));
4911 			hashno = 1;
4912 		} else {
4913 			hashno++;
4914 		}
4915 	}
4916 
4917 	sfmmu_hblks_list_purge(&list, 0);
4918 	DEMAP_RANGE_FLUSH(&dmr);
4919 	cpuset = sfmmup->sfmmu_cpusran;
4920 	xt_sync(cpuset);
4921 }
4922 
4923 /*
4924  * This function chgattr on a range of addresses in an hmeblk.  It returns the
4925  * next addres that needs to be chgattr.
4926  * It should be called with the hash lock held.
4927  * XXX It should be possible to optimize chgattr by not flushing every time but
4928  * on the other hand:
4929  * 1. do one flush crosscall.
4930  * 2. only flush if we are increasing permissions (make sure this will work)
4931  */
4932 static caddr_t
4933 sfmmu_hblk_chgattr(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
4934 	caddr_t endaddr, demap_range_t *dmrp, uint_t attr, int mode)
4935 {
4936 	tte_t tte, tteattr, tteflags, ttemod;
4937 	struct sf_hment *sfhmep;
4938 	int ttesz;
4939 	struct page *pp = NULL;
4940 	kmutex_t *pml, *pmtx;
4941 	int ret;
4942 	int use_demap_range;
4943 #if defined(SF_ERRATA_57)
4944 	int check_exec;
4945 #endif
4946 
4947 	ASSERT(in_hblk_range(hmeblkp, addr));
4948 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4949 	ASSERT(!hmeblkp->hblk_shared);
4950 
4951 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4952 	ttesz = get_hblk_ttesz(hmeblkp);
4953 
4954 	/*
4955 	 * Flush the current demap region if addresses have been
4956 	 * skipped or the page size doesn't match.
4957 	 */
4958 	use_demap_range = (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp));
4959 	if (use_demap_range) {
4960 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
4961 	} else {
4962 		DEMAP_RANGE_FLUSH(dmrp);
4963 	}
4964 
4965 	tteattr.ll = sfmmu_vtop_attr(attr, mode, &tteflags);
4966 #if defined(SF_ERRATA_57)
4967 	check_exec = (sfmmup != ksfmmup) &&
4968 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
4969 	    TTE_IS_EXECUTABLE(&tteattr);
4970 #endif
4971 	HBLKTOHME(sfhmep, hmeblkp, addr);
4972 	while (addr < endaddr) {
4973 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
4974 		if (TTE_IS_VALID(&tte)) {
4975 			if ((tte.ll & tteflags.ll) == tteattr.ll) {
4976 				/*
4977 				 * if the new attr is the same as old
4978 				 * continue
4979 				 */
4980 				goto next_addr;
4981 			}
4982 			if (!TTE_IS_WRITABLE(&tteattr)) {
4983 				/*
4984 				 * make sure we clear hw modify bit if we
4985 				 * removing write protections
4986 				 */
4987 				tteflags.tte_intlo |= TTE_HWWR_INT;
4988 			}
4989 
4990 			pml = NULL;
4991 			pp = sfhmep->hme_page;
4992 			if (pp) {
4993 				pml = sfmmu_mlist_enter(pp);
4994 			}
4995 
4996 			if (pp != sfhmep->hme_page) {
4997 				/*
4998 				 * tte must have been unloaded.
4999 				 */
5000 				ASSERT(pml);
5001 				sfmmu_mlist_exit(pml);
5002 				continue;
5003 			}
5004 
5005 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5006 
5007 			ttemod = tte;
5008 			ttemod.ll = (ttemod.ll & ~tteflags.ll) | tteattr.ll;
5009 			ASSERT(TTE_TO_TTEPFN(&ttemod) == TTE_TO_TTEPFN(&tte));
5010 
5011 #if defined(SF_ERRATA_57)
5012 			if (check_exec && addr < errata57_limit)
5013 				ttemod.tte_exec_perm = 0;
5014 #endif
5015 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5016 			    &sfhmep->hme_tte);
5017 
5018 			if (ret < 0) {
5019 				/* tte changed underneath us */
5020 				if (pml) {
5021 					sfmmu_mlist_exit(pml);
5022 				}
5023 				continue;
5024 			}
5025 
5026 			if ((tteflags.tte_intlo & TTE_HWWR_INT) ||
5027 			    (TTE_EXECUTED(&tte) &&
5028 			    !TTE_IS_EXECUTABLE(&ttemod))) {
5029 				/*
5030 				 * need to sync if clearing modify/exec bit.
5031 				 */
5032 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5033 			}
5034 
5035 			if (pp && PP_ISRO(pp)) {
5036 				if (tteattr.tte_intlo & TTE_WRPRM_INT) {
5037 					pmtx = sfmmu_page_enter(pp);
5038 					PP_CLRRO(pp);
5039 					sfmmu_page_exit(pmtx);
5040 				}
5041 			}
5042 
5043 			if (ret > 0 && use_demap_range) {
5044 				DEMAP_RANGE_MARKPG(dmrp, addr);
5045 			} else if (ret > 0) {
5046 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5047 			}
5048 
5049 			if (pml) {
5050 				sfmmu_mlist_exit(pml);
5051 			}
5052 		}
5053 next_addr:
5054 		addr += TTEBYTES(ttesz);
5055 		sfhmep++;
5056 		DEMAP_RANGE_NEXTPG(dmrp);
5057 	}
5058 	return (addr);
5059 }
5060 
5061 /*
5062  * This routine converts virtual attributes to physical ones.  It will
5063  * update the tteflags field with the tte mask corresponding to the attributes
5064  * affected and it returns the new attributes.  It will also clear the modify
5065  * bit if we are taking away write permission.  This is necessary since the
5066  * modify bit is the hardware permission bit and we need to clear it in order
5067  * to detect write faults.
5068  */
5069 static uint64_t
5070 sfmmu_vtop_attr(uint_t attr, int mode, tte_t *ttemaskp)
5071 {
5072 	tte_t ttevalue;
5073 
5074 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
5075 
5076 	switch (mode) {
5077 	case SFMMU_CHGATTR:
5078 		/* all attributes specified */
5079 		ttevalue.tte_inthi = MAKE_TTEATTR_INTHI(attr);
5080 		ttevalue.tte_intlo = MAKE_TTEATTR_INTLO(attr);
5081 		ttemaskp->tte_inthi = TTEINTHI_ATTR;
5082 		ttemaskp->tte_intlo = TTEINTLO_ATTR;
5083 		if (!icache_is_coherent) {
5084 			if (!(attr & PROT_EXEC)) {
5085 				TTE_SET_SOFTEXEC(ttemaskp);
5086 			} else {
5087 				TTE_CLR_EXEC(ttemaskp);
5088 				TTE_SET_SOFTEXEC(&ttevalue);
5089 			}
5090 		}
5091 		break;
5092 	case SFMMU_SETATTR:
5093 		ASSERT(!(attr & ~HAT_PROT_MASK));
5094 		ttemaskp->ll = 0;
5095 		ttevalue.ll = 0;
5096 		/*
5097 		 * a valid tte implies exec and read for sfmmu
5098 		 * so no need to do anything about them.
5099 		 * since priviledged access implies user access
5100 		 * PROT_USER doesn't make sense either.
5101 		 */
5102 		if (attr & PROT_WRITE) {
5103 			ttemaskp->tte_intlo |= TTE_WRPRM_INT;
5104 			ttevalue.tte_intlo |= TTE_WRPRM_INT;
5105 		}
5106 		break;
5107 	case SFMMU_CLRATTR:
5108 		/* attributes will be nand with current ones */
5109 		if (attr & ~(PROT_WRITE | PROT_USER)) {
5110 			panic("sfmmu: attr %x not supported", attr);
5111 		}
5112 		ttemaskp->ll = 0;
5113 		ttevalue.ll = 0;
5114 		if (attr & PROT_WRITE) {
5115 			/* clear both writable and modify bit */
5116 			ttemaskp->tte_intlo |= TTE_WRPRM_INT | TTE_HWWR_INT;
5117 		}
5118 		if (attr & PROT_USER) {
5119 			ttemaskp->tte_intlo |= TTE_PRIV_INT;
5120 			ttevalue.tte_intlo |= TTE_PRIV_INT;
5121 		}
5122 		break;
5123 	default:
5124 		panic("sfmmu_vtop_attr: bad mode %x", mode);
5125 	}
5126 	ASSERT(TTE_TO_TTEPFN(&ttevalue) == 0);
5127 	return (ttevalue.ll);
5128 }
5129 
5130 static uint_t
5131 sfmmu_ptov_attr(tte_t *ttep)
5132 {
5133 	uint_t attr;
5134 
5135 	ASSERT(TTE_IS_VALID(ttep));
5136 
5137 	attr = PROT_READ;
5138 
5139 	if (TTE_IS_WRITABLE(ttep)) {
5140 		attr |= PROT_WRITE;
5141 	}
5142 	if (TTE_IS_EXECUTABLE(ttep)) {
5143 		attr |= PROT_EXEC;
5144 	}
5145 	if (TTE_IS_SOFTEXEC(ttep)) {
5146 		attr |= PROT_EXEC;
5147 	}
5148 	if (!TTE_IS_PRIVILEGED(ttep)) {
5149 		attr |= PROT_USER;
5150 	}
5151 	if (TTE_IS_NFO(ttep)) {
5152 		attr |= HAT_NOFAULT;
5153 	}
5154 	if (TTE_IS_NOSYNC(ttep)) {
5155 		attr |= HAT_NOSYNC;
5156 	}
5157 	if (TTE_IS_SIDEFFECT(ttep)) {
5158 		attr |= SFMMU_SIDEFFECT;
5159 	}
5160 	if (!TTE_IS_VCACHEABLE(ttep)) {
5161 		attr |= SFMMU_UNCACHEVTTE;
5162 	}
5163 	if (!TTE_IS_PCACHEABLE(ttep)) {
5164 		attr |= SFMMU_UNCACHEPTTE;
5165 	}
5166 	return (attr);
5167 }
5168 
5169 /*
5170  * hat_chgprot is a deprecated hat call.  New segment drivers
5171  * should store all attributes and use hat_*attr calls.
5172  *
5173  * Change the protections in the virtual address range
5174  * given to the specified virtual protection.  If vprot is ~PROT_WRITE,
5175  * then remove write permission, leaving the other
5176  * permissions unchanged.  If vprot is ~PROT_USER, remove user permissions.
5177  *
5178  */
5179 void
5180 hat_chgprot(struct hat *sfmmup, caddr_t addr, size_t len, uint_t vprot)
5181 {
5182 	struct hmehash_bucket *hmebp;
5183 	hmeblk_tag hblktag;
5184 	int hmeshift, hashno = 1;
5185 	struct hme_blk *hmeblkp, *list = NULL;
5186 	caddr_t endaddr;
5187 	cpuset_t cpuset;
5188 	demap_range_t dmr;
5189 
5190 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5191 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
5192 
5193 	if (sfmmup->sfmmu_xhat_provider) {
5194 		XHAT_CHGPROT(sfmmup, addr, len, vprot);
5195 		return;
5196 	} else {
5197 		/*
5198 		 * This must be a CPU HAT. If the address space has
5199 		 * XHATs attached, change attributes for all of them,
5200 		 * just in case
5201 		 */
5202 		ASSERT(sfmmup->sfmmu_as != NULL);
5203 		if (sfmmup->sfmmu_as->a_xhat != NULL)
5204 			xhat_chgprot_all(sfmmup->sfmmu_as, addr, len, vprot);
5205 	}
5206 
5207 	CPUSET_ZERO(cpuset);
5208 
5209 	if ((vprot != (uint_t)~PROT_WRITE) && (vprot & PROT_USER) &&
5210 	    ((addr + len) > (caddr_t)USERLIMIT)) {
5211 		panic("user addr %p vprot %x in kernel space",
5212 		    (void *)addr, vprot);
5213 	}
5214 	endaddr = addr + len;
5215 	hblktag.htag_id = sfmmup;
5216 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5217 	DEMAP_RANGE_INIT(sfmmup, &dmr);
5218 
5219 	while (addr < endaddr) {
5220 		hmeshift = HME_HASH_SHIFT(hashno);
5221 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5222 		hblktag.htag_rehash = hashno;
5223 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5224 
5225 		SFMMU_HASH_LOCK(hmebp);
5226 
5227 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
5228 		if (hmeblkp != NULL) {
5229 			ASSERT(!hmeblkp->hblk_shared);
5230 			/*
5231 			 * We've encountered a shadow hmeblk so skip the range
5232 			 * of the next smaller mapping size.
5233 			 */
5234 			if (hmeblkp->hblk_shw_bit) {
5235 				ASSERT(sfmmup != ksfmmup);
5236 				ASSERT(hashno > 1);
5237 				addr = (caddr_t)P2END((uintptr_t)addr,
5238 				    TTEBYTES(hashno - 1));
5239 			} else {
5240 				addr = sfmmu_hblk_chgprot(sfmmup, hmeblkp,
5241 				    addr, endaddr, &dmr, vprot);
5242 			}
5243 			SFMMU_HASH_UNLOCK(hmebp);
5244 			hashno = 1;
5245 			continue;
5246 		}
5247 		SFMMU_HASH_UNLOCK(hmebp);
5248 
5249 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
5250 			/*
5251 			 * We have traversed the whole list and rehashed
5252 			 * if necessary without finding the address to chgprot.
5253 			 * This is ok so we increment the address by the
5254 			 * smallest hmeblk range for kernel mappings and the
5255 			 * largest hmeblk range, to account for shadow hmeblks,
5256 			 * for user mappings and continue.
5257 			 */
5258 			if (sfmmup == ksfmmup)
5259 				addr = (caddr_t)P2END((uintptr_t)addr,
5260 				    TTEBYTES(1));
5261 			else
5262 				addr = (caddr_t)P2END((uintptr_t)addr,
5263 				    TTEBYTES(hashno));
5264 			hashno = 1;
5265 		} else {
5266 			hashno++;
5267 		}
5268 	}
5269 
5270 	sfmmu_hblks_list_purge(&list, 0);
5271 	DEMAP_RANGE_FLUSH(&dmr);
5272 	cpuset = sfmmup->sfmmu_cpusran;
5273 	xt_sync(cpuset);
5274 }
5275 
5276 /*
5277  * This function chgprots a range of addresses in an hmeblk.  It returns the
5278  * next addres that needs to be chgprot.
5279  * It should be called with the hash lock held.
5280  * XXX It shold be possible to optimize chgprot by not flushing every time but
5281  * on the other hand:
5282  * 1. do one flush crosscall.
5283  * 2. only flush if we are increasing permissions (make sure this will work)
5284  */
5285 static caddr_t
5286 sfmmu_hblk_chgprot(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5287 	caddr_t endaddr, demap_range_t *dmrp, uint_t vprot)
5288 {
5289 	uint_t pprot;
5290 	tte_t tte, ttemod;
5291 	struct sf_hment *sfhmep;
5292 	uint_t tteflags;
5293 	int ttesz;
5294 	struct page *pp = NULL;
5295 	kmutex_t *pml, *pmtx;
5296 	int ret;
5297 	int use_demap_range;
5298 #if defined(SF_ERRATA_57)
5299 	int check_exec;
5300 #endif
5301 
5302 	ASSERT(in_hblk_range(hmeblkp, addr));
5303 	ASSERT(hmeblkp->hblk_shw_bit == 0);
5304 	ASSERT(!hmeblkp->hblk_shared);
5305 
5306 #ifdef DEBUG
5307 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5308 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5309 		panic("sfmmu_hblk_chgprot: partial chgprot of large page");
5310 	}
5311 #endif /* DEBUG */
5312 
5313 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5314 	ttesz = get_hblk_ttesz(hmeblkp);
5315 
5316 	pprot = sfmmu_vtop_prot(vprot, &tteflags);
5317 #if defined(SF_ERRATA_57)
5318 	check_exec = (sfmmup != ksfmmup) &&
5319 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
5320 	    ((vprot & PROT_EXEC) == PROT_EXEC);
5321 #endif
5322 	HBLKTOHME(sfhmep, hmeblkp, addr);
5323 
5324 	/*
5325 	 * Flush the current demap region if addresses have been
5326 	 * skipped or the page size doesn't match.
5327 	 */
5328 	use_demap_range = (TTEBYTES(ttesz) == MMU_PAGESIZE);
5329 	if (use_demap_range) {
5330 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5331 	} else {
5332 		DEMAP_RANGE_FLUSH(dmrp);
5333 	}
5334 
5335 	while (addr < endaddr) {
5336 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5337 		if (TTE_IS_VALID(&tte)) {
5338 			if (TTE_GET_LOFLAGS(&tte, tteflags) == pprot) {
5339 				/*
5340 				 * if the new protection is the same as old
5341 				 * continue
5342 				 */
5343 				goto next_addr;
5344 			}
5345 			pml = NULL;
5346 			pp = sfhmep->hme_page;
5347 			if (pp) {
5348 				pml = sfmmu_mlist_enter(pp);
5349 			}
5350 			if (pp != sfhmep->hme_page) {
5351 				/*
5352 				 * tte most have been unloaded
5353 				 * underneath us.  Recheck
5354 				 */
5355 				ASSERT(pml);
5356 				sfmmu_mlist_exit(pml);
5357 				continue;
5358 			}
5359 
5360 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5361 
5362 			ttemod = tte;
5363 			TTE_SET_LOFLAGS(&ttemod, tteflags, pprot);
5364 			ASSERT(TTE_IS_SOFTEXEC(&tte) ==
5365 			    TTE_IS_SOFTEXEC(&ttemod));
5366 			ASSERT(TTE_IS_EXECUTABLE(&tte) ==
5367 			    TTE_IS_EXECUTABLE(&ttemod));
5368 
5369 #if defined(SF_ERRATA_57)
5370 			if (check_exec && addr < errata57_limit)
5371 				ttemod.tte_exec_perm = 0;
5372 #endif
5373 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5374 			    &sfhmep->hme_tte);
5375 
5376 			if (ret < 0) {
5377 				/* tte changed underneath us */
5378 				if (pml) {
5379 					sfmmu_mlist_exit(pml);
5380 				}
5381 				continue;
5382 			}
5383 
5384 			if (tteflags & TTE_HWWR_INT) {
5385 				/*
5386 				 * need to sync if we are clearing modify bit.
5387 				 */
5388 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5389 			}
5390 
5391 			if (pp && PP_ISRO(pp)) {
5392 				if (pprot & TTE_WRPRM_INT) {
5393 					pmtx = sfmmu_page_enter(pp);
5394 					PP_CLRRO(pp);
5395 					sfmmu_page_exit(pmtx);
5396 				}
5397 			}
5398 
5399 			if (ret > 0 && use_demap_range) {
5400 				DEMAP_RANGE_MARKPG(dmrp, addr);
5401 			} else if (ret > 0) {
5402 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5403 			}
5404 
5405 			if (pml) {
5406 				sfmmu_mlist_exit(pml);
5407 			}
5408 		}
5409 next_addr:
5410 		addr += TTEBYTES(ttesz);
5411 		sfhmep++;
5412 		DEMAP_RANGE_NEXTPG(dmrp);
5413 	}
5414 	return (addr);
5415 }
5416 
5417 /*
5418  * This routine is deprecated and should only be used by hat_chgprot.
5419  * The correct routine is sfmmu_vtop_attr.
5420  * This routine converts virtual page protections to physical ones.  It will
5421  * update the tteflags field with the tte mask corresponding to the protections
5422  * affected and it returns the new protections.  It will also clear the modify
5423  * bit if we are taking away write permission.  This is necessary since the
5424  * modify bit is the hardware permission bit and we need to clear it in order
5425  * to detect write faults.
5426  * It accepts the following special protections:
5427  * ~PROT_WRITE = remove write permissions.
5428  * ~PROT_USER = remove user permissions.
5429  */
5430 static uint_t
5431 sfmmu_vtop_prot(uint_t vprot, uint_t *tteflagsp)
5432 {
5433 	if (vprot == (uint_t)~PROT_WRITE) {
5434 		*tteflagsp = TTE_WRPRM_INT | TTE_HWWR_INT;
5435 		return (0);		/* will cause wrprm to be cleared */
5436 	}
5437 	if (vprot == (uint_t)~PROT_USER) {
5438 		*tteflagsp = TTE_PRIV_INT;
5439 		return (0);		/* will cause privprm to be cleared */
5440 	}
5441 	if ((vprot == 0) || (vprot == PROT_USER) ||
5442 	    ((vprot & PROT_ALL) != vprot)) {
5443 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5444 	}
5445 
5446 	switch (vprot) {
5447 	case (PROT_READ):
5448 	case (PROT_EXEC):
5449 	case (PROT_EXEC | PROT_READ):
5450 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5451 		return (TTE_PRIV_INT); 		/* set prv and clr wrt */
5452 	case (PROT_WRITE):
5453 	case (PROT_WRITE | PROT_READ):
5454 	case (PROT_EXEC | PROT_WRITE):
5455 	case (PROT_EXEC | PROT_WRITE | PROT_READ):
5456 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5457 		return (TTE_PRIV_INT | TTE_WRPRM_INT); 	/* set prv and wrt */
5458 	case (PROT_USER | PROT_READ):
5459 	case (PROT_USER | PROT_EXEC):
5460 	case (PROT_USER | PROT_EXEC | PROT_READ):
5461 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5462 		return (0); 			/* clr prv and wrt */
5463 	case (PROT_USER | PROT_WRITE):
5464 	case (PROT_USER | PROT_WRITE | PROT_READ):
5465 	case (PROT_USER | PROT_EXEC | PROT_WRITE):
5466 	case (PROT_USER | PROT_EXEC | PROT_WRITE | PROT_READ):
5467 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5468 		return (TTE_WRPRM_INT); 	/* clr prv and set wrt */
5469 	default:
5470 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5471 	}
5472 	return (0);
5473 }
5474 
5475 /*
5476  * Alternate unload for very large virtual ranges. With a true 64 bit VA,
5477  * the normal algorithm would take too long for a very large VA range with
5478  * few real mappings. This routine just walks thru all HMEs in the global
5479  * hash table to find and remove mappings.
5480  */
5481 static void
5482 hat_unload_large_virtual(
5483 	struct hat		*sfmmup,
5484 	caddr_t			startaddr,
5485 	size_t			len,
5486 	uint_t			flags,
5487 	hat_callback_t		*callback)
5488 {
5489 	struct hmehash_bucket *hmebp;
5490 	struct hme_blk *hmeblkp;
5491 	struct hme_blk *pr_hblk = NULL;
5492 	struct hme_blk *nx_hblk;
5493 	struct hme_blk *list = NULL;
5494 	int i;
5495 	demap_range_t dmr, *dmrp;
5496 	cpuset_t cpuset;
5497 	caddr_t	endaddr = startaddr + len;
5498 	caddr_t	sa;
5499 	caddr_t	ea;
5500 	caddr_t	cb_sa[MAX_CB_ADDR];
5501 	caddr_t	cb_ea[MAX_CB_ADDR];
5502 	int	addr_cnt = 0;
5503 	int	a = 0;
5504 
5505 	if (sfmmup->sfmmu_free) {
5506 		dmrp = NULL;
5507 	} else {
5508 		dmrp = &dmr;
5509 		DEMAP_RANGE_INIT(sfmmup, dmrp);
5510 	}
5511 
5512 	/*
5513 	 * Loop through all the hash buckets of HME blocks looking for matches.
5514 	 */
5515 	for (i = 0; i <= UHMEHASH_SZ; i++) {
5516 		hmebp = &uhme_hash[i];
5517 		SFMMU_HASH_LOCK(hmebp);
5518 		hmeblkp = hmebp->hmeblkp;
5519 		pr_hblk = NULL;
5520 		while (hmeblkp) {
5521 			nx_hblk = hmeblkp->hblk_next;
5522 
5523 			/*
5524 			 * skip if not this context, if a shadow block or
5525 			 * if the mapping is not in the requested range
5526 			 */
5527 			if (hmeblkp->hblk_tag.htag_id != sfmmup ||
5528 			    hmeblkp->hblk_shw_bit ||
5529 			    (sa = (caddr_t)get_hblk_base(hmeblkp)) >= endaddr ||
5530 			    (ea = get_hblk_endaddr(hmeblkp)) <= startaddr) {
5531 				pr_hblk = hmeblkp;
5532 				goto next_block;
5533 			}
5534 
5535 			ASSERT(!hmeblkp->hblk_shared);
5536 			/*
5537 			 * unload if there are any current valid mappings
5538 			 */
5539 			if (hmeblkp->hblk_vcnt != 0 ||
5540 			    hmeblkp->hblk_hmecnt != 0)
5541 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
5542 				    sa, ea, dmrp, flags);
5543 
5544 			/*
5545 			 * on unmap we also release the HME block itself, once
5546 			 * all mappings are gone.
5547 			 */
5548 			if ((flags & HAT_UNLOAD_UNMAP) != 0 &&
5549 			    !hmeblkp->hblk_vcnt &&
5550 			    !hmeblkp->hblk_hmecnt) {
5551 				ASSERT(!hmeblkp->hblk_lckcnt);
5552 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
5553 				    &list, 0);
5554 			} else {
5555 				pr_hblk = hmeblkp;
5556 			}
5557 
5558 			if (callback == NULL)
5559 				goto next_block;
5560 
5561 			/*
5562 			 * HME blocks may span more than one page, but we may be
5563 			 * unmapping only one page, so check for a smaller range
5564 			 * for the callback
5565 			 */
5566 			if (sa < startaddr)
5567 				sa = startaddr;
5568 			if (--ea > endaddr)
5569 				ea = endaddr - 1;
5570 
5571 			cb_sa[addr_cnt] = sa;
5572 			cb_ea[addr_cnt] = ea;
5573 			if (++addr_cnt == MAX_CB_ADDR) {
5574 				if (dmrp != NULL) {
5575 					DEMAP_RANGE_FLUSH(dmrp);
5576 					cpuset = sfmmup->sfmmu_cpusran;
5577 					xt_sync(cpuset);
5578 				}
5579 
5580 				for (a = 0; a < MAX_CB_ADDR; ++a) {
5581 					callback->hcb_start_addr = cb_sa[a];
5582 					callback->hcb_end_addr = cb_ea[a];
5583 					callback->hcb_function(callback);
5584 				}
5585 				addr_cnt = 0;
5586 			}
5587 
5588 next_block:
5589 			hmeblkp = nx_hblk;
5590 		}
5591 		SFMMU_HASH_UNLOCK(hmebp);
5592 	}
5593 
5594 	sfmmu_hblks_list_purge(&list, 0);
5595 	if (dmrp != NULL) {
5596 		DEMAP_RANGE_FLUSH(dmrp);
5597 		cpuset = sfmmup->sfmmu_cpusran;
5598 		xt_sync(cpuset);
5599 	}
5600 
5601 	for (a = 0; a < addr_cnt; ++a) {
5602 		callback->hcb_start_addr = cb_sa[a];
5603 		callback->hcb_end_addr = cb_ea[a];
5604 		callback->hcb_function(callback);
5605 	}
5606 
5607 	/*
5608 	 * Check TSB and TLB page sizes if the process isn't exiting.
5609 	 */
5610 	if (!sfmmup->sfmmu_free)
5611 		sfmmu_check_page_sizes(sfmmup, 0);
5612 }
5613 
5614 /*
5615  * Unload all the mappings in the range [addr..addr+len). addr and len must
5616  * be MMU_PAGESIZE aligned.
5617  */
5618 
5619 extern struct seg *segkmap;
5620 #define	ISSEGKMAP(sfmmup, addr) (sfmmup == ksfmmup && \
5621 segkmap->s_base <= (addr) && (addr) < (segkmap->s_base + segkmap->s_size))
5622 
5623 
5624 void
5625 hat_unload_callback(
5626 	struct hat *sfmmup,
5627 	caddr_t addr,
5628 	size_t len,
5629 	uint_t flags,
5630 	hat_callback_t *callback)
5631 {
5632 	struct hmehash_bucket *hmebp;
5633 	hmeblk_tag hblktag;
5634 	int hmeshift, hashno, iskernel;
5635 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
5636 	caddr_t endaddr;
5637 	cpuset_t cpuset;
5638 	int addr_count = 0;
5639 	int a;
5640 	caddr_t cb_start_addr[MAX_CB_ADDR];
5641 	caddr_t cb_end_addr[MAX_CB_ADDR];
5642 	int issegkmap = ISSEGKMAP(sfmmup, addr);
5643 	demap_range_t dmr, *dmrp;
5644 
5645 	if (sfmmup->sfmmu_xhat_provider) {
5646 		XHAT_UNLOAD_CALLBACK(sfmmup, addr, len, flags, callback);
5647 		return;
5648 	} else {
5649 		/*
5650 		 * This must be a CPU HAT. If the address space has
5651 		 * XHATs attached, unload the mappings for all of them,
5652 		 * just in case
5653 		 */
5654 		ASSERT(sfmmup->sfmmu_as != NULL);
5655 		if (sfmmup->sfmmu_as->a_xhat != NULL)
5656 			xhat_unload_callback_all(sfmmup->sfmmu_as, addr,
5657 			    len, flags, callback);
5658 	}
5659 
5660 	ASSERT((sfmmup == ksfmmup) || (flags & HAT_UNLOAD_OTHER) || \
5661 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
5662 
5663 	ASSERT(sfmmup != NULL);
5664 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5665 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
5666 
5667 	/*
5668 	 * Probing through a large VA range (say 63 bits) will be slow, even
5669 	 * at 4 Meg steps between the probes. So, when the virtual address range
5670 	 * is very large, search the HME entries for what to unload.
5671 	 *
5672 	 *	len >> TTE_PAGE_SHIFT(TTE4M) is the # of 4Meg probes we'd need
5673 	 *
5674 	 *	UHMEHASH_SZ is number of hash buckets to examine
5675 	 *
5676 	 */
5677 	if (sfmmup != KHATID && (len >> TTE_PAGE_SHIFT(TTE4M)) > UHMEHASH_SZ) {
5678 		hat_unload_large_virtual(sfmmup, addr, len, flags, callback);
5679 		return;
5680 	}
5681 
5682 	CPUSET_ZERO(cpuset);
5683 
5684 	/*
5685 	 * If the process is exiting, we can save a lot of fuss since
5686 	 * we'll flush the TLB when we free the ctx anyway.
5687 	 */
5688 	if (sfmmup->sfmmu_free)
5689 		dmrp = NULL;
5690 	else
5691 		dmrp = &dmr;
5692 
5693 	DEMAP_RANGE_INIT(sfmmup, dmrp);
5694 	endaddr = addr + len;
5695 	hblktag.htag_id = sfmmup;
5696 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5697 
5698 	/*
5699 	 * It is likely for the vm to call unload over a wide range of
5700 	 * addresses that are actually very sparsely populated by
5701 	 * translations.  In order to speed this up the sfmmu hat supports
5702 	 * the concept of shadow hmeblks. Dummy large page hmeblks that
5703 	 * correspond to actual small translations are allocated at tteload
5704 	 * time and are referred to as shadow hmeblks.  Now, during unload
5705 	 * time, we first check if we have a shadow hmeblk for that
5706 	 * translation.  The absence of one means the corresponding address
5707 	 * range is empty and can be skipped.
5708 	 *
5709 	 * The kernel is an exception to above statement and that is why
5710 	 * we don't use shadow hmeblks and hash starting from the smallest
5711 	 * page size.
5712 	 */
5713 	if (sfmmup == KHATID) {
5714 		iskernel = 1;
5715 		hashno = TTE64K;
5716 	} else {
5717 		iskernel = 0;
5718 		if (mmu_page_sizes == max_mmu_page_sizes) {
5719 			hashno = TTE256M;
5720 		} else {
5721 			hashno = TTE4M;
5722 		}
5723 	}
5724 	while (addr < endaddr) {
5725 		hmeshift = HME_HASH_SHIFT(hashno);
5726 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5727 		hblktag.htag_rehash = hashno;
5728 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5729 
5730 		SFMMU_HASH_LOCK(hmebp);
5731 
5732 		HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
5733 		if (hmeblkp == NULL) {
5734 			/*
5735 			 * didn't find an hmeblk. skip the appropiate
5736 			 * address range.
5737 			 */
5738 			SFMMU_HASH_UNLOCK(hmebp);
5739 			if (iskernel) {
5740 				if (hashno < mmu_hashcnt) {
5741 					hashno++;
5742 					continue;
5743 				} else {
5744 					hashno = TTE64K;
5745 					addr = (caddr_t)roundup((uintptr_t)addr
5746 					    + 1, MMU_PAGESIZE64K);
5747 					continue;
5748 				}
5749 			}
5750 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5751 			    (1 << hmeshift));
5752 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5753 				ASSERT(hashno == TTE64K);
5754 				continue;
5755 			}
5756 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5757 				hashno = TTE512K;
5758 				continue;
5759 			}
5760 			if (mmu_page_sizes == max_mmu_page_sizes) {
5761 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5762 					hashno = TTE4M;
5763 					continue;
5764 				}
5765 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5766 					hashno = TTE32M;
5767 					continue;
5768 				}
5769 				hashno = TTE256M;
5770 				continue;
5771 			} else {
5772 				hashno = TTE4M;
5773 				continue;
5774 			}
5775 		}
5776 		ASSERT(hmeblkp);
5777 		ASSERT(!hmeblkp->hblk_shared);
5778 		if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5779 			/*
5780 			 * If the valid count is zero we can skip the range
5781 			 * mapped by this hmeblk.
5782 			 * We free hblks in the case of HAT_UNMAP.  HAT_UNMAP
5783 			 * is used by segment drivers as a hint
5784 			 * that the mapping resource won't be used any longer.
5785 			 * The best example of this is during exit().
5786 			 */
5787 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5788 			    get_hblk_span(hmeblkp));
5789 			if ((flags & HAT_UNLOAD_UNMAP) ||
5790 			    (iskernel && !issegkmap)) {
5791 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
5792 				    &list, 0);
5793 			}
5794 			SFMMU_HASH_UNLOCK(hmebp);
5795 
5796 			if (iskernel) {
5797 				hashno = TTE64K;
5798 				continue;
5799 			}
5800 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5801 				ASSERT(hashno == TTE64K);
5802 				continue;
5803 			}
5804 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5805 				hashno = TTE512K;
5806 				continue;
5807 			}
5808 			if (mmu_page_sizes == max_mmu_page_sizes) {
5809 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5810 					hashno = TTE4M;
5811 					continue;
5812 				}
5813 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5814 					hashno = TTE32M;
5815 					continue;
5816 				}
5817 				hashno = TTE256M;
5818 				continue;
5819 			} else {
5820 				hashno = TTE4M;
5821 				continue;
5822 			}
5823 		}
5824 		if (hmeblkp->hblk_shw_bit) {
5825 			/*
5826 			 * If we encounter a shadow hmeblk we know there is
5827 			 * smaller sized hmeblks mapping the same address space.
5828 			 * Decrement the hash size and rehash.
5829 			 */
5830 			ASSERT(sfmmup != KHATID);
5831 			hashno--;
5832 			SFMMU_HASH_UNLOCK(hmebp);
5833 			continue;
5834 		}
5835 
5836 		/*
5837 		 * track callback address ranges.
5838 		 * only start a new range when it's not contiguous
5839 		 */
5840 		if (callback != NULL) {
5841 			if (addr_count > 0 &&
5842 			    addr == cb_end_addr[addr_count - 1])
5843 				--addr_count;
5844 			else
5845 				cb_start_addr[addr_count] = addr;
5846 		}
5847 
5848 		addr = sfmmu_hblk_unload(sfmmup, hmeblkp, addr, endaddr,
5849 		    dmrp, flags);
5850 
5851 		if (callback != NULL)
5852 			cb_end_addr[addr_count++] = addr;
5853 
5854 		if (((flags & HAT_UNLOAD_UNMAP) || (iskernel && !issegkmap)) &&
5855 		    !hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5856 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 0);
5857 		}
5858 		SFMMU_HASH_UNLOCK(hmebp);
5859 
5860 		/*
5861 		 * Notify our caller as to exactly which pages
5862 		 * have been unloaded. We do these in clumps,
5863 		 * to minimize the number of xt_sync()s that need to occur.
5864 		 */
5865 		if (callback != NULL && addr_count == MAX_CB_ADDR) {
5866 			DEMAP_RANGE_FLUSH(dmrp);
5867 			if (dmrp != NULL) {
5868 				cpuset = sfmmup->sfmmu_cpusran;
5869 				xt_sync(cpuset);
5870 			}
5871 
5872 			for (a = 0; a < MAX_CB_ADDR; ++a) {
5873 				callback->hcb_start_addr = cb_start_addr[a];
5874 				callback->hcb_end_addr = cb_end_addr[a];
5875 				callback->hcb_function(callback);
5876 			}
5877 			addr_count = 0;
5878 		}
5879 		if (iskernel) {
5880 			hashno = TTE64K;
5881 			continue;
5882 		}
5883 		if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5884 			ASSERT(hashno == TTE64K);
5885 			continue;
5886 		}
5887 		if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5888 			hashno = TTE512K;
5889 			continue;
5890 		}
5891 		if (mmu_page_sizes == max_mmu_page_sizes) {
5892 			if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5893 				hashno = TTE4M;
5894 				continue;
5895 			}
5896 			if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5897 				hashno = TTE32M;
5898 				continue;
5899 			}
5900 			hashno = TTE256M;
5901 		} else {
5902 			hashno = TTE4M;
5903 		}
5904 	}
5905 
5906 	sfmmu_hblks_list_purge(&list, 0);
5907 	DEMAP_RANGE_FLUSH(dmrp);
5908 	if (dmrp != NULL) {
5909 		cpuset = sfmmup->sfmmu_cpusran;
5910 		xt_sync(cpuset);
5911 	}
5912 	if (callback && addr_count != 0) {
5913 		for (a = 0; a < addr_count; ++a) {
5914 			callback->hcb_start_addr = cb_start_addr[a];
5915 			callback->hcb_end_addr = cb_end_addr[a];
5916 			callback->hcb_function(callback);
5917 		}
5918 	}
5919 
5920 	/*
5921 	 * Check TSB and TLB page sizes if the process isn't exiting.
5922 	 */
5923 	if (!sfmmup->sfmmu_free)
5924 		sfmmu_check_page_sizes(sfmmup, 0);
5925 }
5926 
5927 /*
5928  * Unload all the mappings in the range [addr..addr+len). addr and len must
5929  * be MMU_PAGESIZE aligned.
5930  */
5931 void
5932 hat_unload(struct hat *sfmmup, caddr_t addr, size_t len, uint_t flags)
5933 {
5934 	if (sfmmup->sfmmu_xhat_provider) {
5935 		XHAT_UNLOAD(sfmmup, addr, len, flags);
5936 		return;
5937 	}
5938 	hat_unload_callback(sfmmup, addr, len, flags, NULL);
5939 }
5940 
5941 
5942 /*
5943  * Find the largest mapping size for this page.
5944  */
5945 int
5946 fnd_mapping_sz(page_t *pp)
5947 {
5948 	int sz;
5949 	int p_index;
5950 
5951 	p_index = PP_MAPINDEX(pp);
5952 
5953 	sz = 0;
5954 	p_index >>= 1;	/* don't care about 8K bit */
5955 	for (; p_index; p_index >>= 1) {
5956 		sz++;
5957 	}
5958 
5959 	return (sz);
5960 }
5961 
5962 /*
5963  * This function unloads a range of addresses for an hmeblk.
5964  * It returns the next address to be unloaded.
5965  * It should be called with the hash lock held.
5966  */
5967 static caddr_t
5968 sfmmu_hblk_unload(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5969 	caddr_t endaddr, demap_range_t *dmrp, uint_t flags)
5970 {
5971 	tte_t	tte, ttemod;
5972 	struct	sf_hment *sfhmep;
5973 	int	ttesz;
5974 	long	ttecnt;
5975 	page_t *pp;
5976 	kmutex_t *pml;
5977 	int ret;
5978 	int use_demap_range;
5979 
5980 	ASSERT(in_hblk_range(hmeblkp, addr));
5981 	ASSERT(!hmeblkp->hblk_shw_bit);
5982 	ASSERT(sfmmup != NULL || hmeblkp->hblk_shared);
5983 	ASSERT(sfmmup == NULL || !hmeblkp->hblk_shared);
5984 	ASSERT(dmrp == NULL || !hmeblkp->hblk_shared);
5985 
5986 #ifdef DEBUG
5987 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5988 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5989 		panic("sfmmu_hblk_unload: partial unload of large page");
5990 	}
5991 #endif /* DEBUG */
5992 
5993 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5994 	ttesz = get_hblk_ttesz(hmeblkp);
5995 
5996 	use_demap_range = ((dmrp == NULL) ||
5997 	    (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp)));
5998 
5999 	if (use_demap_range) {
6000 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
6001 	} else {
6002 		DEMAP_RANGE_FLUSH(dmrp);
6003 	}
6004 	ttecnt = 0;
6005 	HBLKTOHME(sfhmep, hmeblkp, addr);
6006 
6007 	while (addr < endaddr) {
6008 		pml = NULL;
6009 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6010 		if (TTE_IS_VALID(&tte)) {
6011 			pp = sfhmep->hme_page;
6012 			if (pp != NULL) {
6013 				pml = sfmmu_mlist_enter(pp);
6014 			}
6015 
6016 			/*
6017 			 * Verify if hme still points to 'pp' now that
6018 			 * we have p_mapping lock.
6019 			 */
6020 			if (sfhmep->hme_page != pp) {
6021 				if (pp != NULL && sfhmep->hme_page != NULL) {
6022 					ASSERT(pml != NULL);
6023 					sfmmu_mlist_exit(pml);
6024 					/* Re-start this iteration. */
6025 					continue;
6026 				}
6027 				ASSERT((pp != NULL) &&
6028 				    (sfhmep->hme_page == NULL));
6029 				goto tte_unloaded;
6030 			}
6031 
6032 			/*
6033 			 * This point on we have both HASH and p_mapping
6034 			 * lock.
6035 			 */
6036 			ASSERT(pp == sfhmep->hme_page);
6037 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
6038 
6039 			/*
6040 			 * We need to loop on modify tte because it is
6041 			 * possible for pagesync to come along and
6042 			 * change the software bits beneath us.
6043 			 *
6044 			 * Page_unload can also invalidate the tte after
6045 			 * we read tte outside of p_mapping lock.
6046 			 */
6047 again:
6048 			ttemod = tte;
6049 
6050 			TTE_SET_INVALID(&ttemod);
6051 			ret = sfmmu_modifytte_try(&tte, &ttemod,
6052 			    &sfhmep->hme_tte);
6053 
6054 			if (ret <= 0) {
6055 				if (TTE_IS_VALID(&tte)) {
6056 					ASSERT(ret < 0);
6057 					goto again;
6058 				}
6059 				if (pp != NULL) {
6060 					panic("sfmmu_hblk_unload: pp = 0x%p "
6061 					    "tte became invalid under mlist"
6062 					    " lock = 0x%p", (void *)pp,
6063 					    (void *)pml);
6064 				}
6065 				continue;
6066 			}
6067 
6068 			if (!(flags & HAT_UNLOAD_NOSYNC) ||
6069 			    (pp != NULL && TTE_EXECUTED(&tte))) {
6070 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
6071 			}
6072 
6073 			/*
6074 			 * Ok- we invalidated the tte. Do the rest of the job.
6075 			 */
6076 			ttecnt++;
6077 
6078 			if (flags & HAT_UNLOAD_UNLOCK) {
6079 				ASSERT(hmeblkp->hblk_lckcnt > 0);
6080 				atomic_add_32(&hmeblkp->hblk_lckcnt, -1);
6081 				HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
6082 			}
6083 
6084 			/*
6085 			 * Normally we would need to flush the page
6086 			 * from the virtual cache at this point in
6087 			 * order to prevent a potential cache alias
6088 			 * inconsistency.
6089 			 * The particular scenario we need to worry
6090 			 * about is:
6091 			 * Given:  va1 and va2 are two virtual address
6092 			 * that alias and map the same physical
6093 			 * address.
6094 			 * 1.   mapping exists from va1 to pa and data
6095 			 * has been read into the cache.
6096 			 * 2.   unload va1.
6097 			 * 3.   load va2 and modify data using va2.
6098 			 * 4    unload va2.
6099 			 * 5.   load va1 and reference data.  Unless we
6100 			 * flush the data cache when we unload we will
6101 			 * get stale data.
6102 			 * Fortunately, page coloring eliminates the
6103 			 * above scenario by remembering the color a
6104 			 * physical page was last or is currently
6105 			 * mapped to.  Now, we delay the flush until
6106 			 * the loading of translations.  Only when the
6107 			 * new translation is of a different color
6108 			 * are we forced to flush.
6109 			 */
6110 			if (use_demap_range) {
6111 				/*
6112 				 * Mark this page as needing a demap.
6113 				 */
6114 				DEMAP_RANGE_MARKPG(dmrp, addr);
6115 			} else {
6116 				ASSERT(sfmmup != NULL);
6117 				ASSERT(!hmeblkp->hblk_shared);
6118 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
6119 				    sfmmup->sfmmu_free, 0);
6120 			}
6121 
6122 			if (pp) {
6123 				/*
6124 				 * Remove the hment from the mapping list
6125 				 */
6126 				ASSERT(hmeblkp->hblk_hmecnt > 0);
6127 
6128 				/*
6129 				 * Again, we cannot
6130 				 * ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS);
6131 				 */
6132 				HME_SUB(sfhmep, pp);
6133 				membar_stst();
6134 				atomic_add_16(&hmeblkp->hblk_hmecnt, -1);
6135 			}
6136 
6137 			ASSERT(hmeblkp->hblk_vcnt > 0);
6138 			atomic_add_16(&hmeblkp->hblk_vcnt, -1);
6139 
6140 			ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
6141 			    !hmeblkp->hblk_lckcnt);
6142 
6143 #ifdef VAC
6144 			if (pp && (pp->p_nrm & (P_KPMC | P_KPMS | P_TNC))) {
6145 				if (PP_ISTNC(pp)) {
6146 					/*
6147 					 * If page was temporary
6148 					 * uncached, try to recache
6149 					 * it. Note that HME_SUB() was
6150 					 * called above so p_index and
6151 					 * mlist had been updated.
6152 					 */
6153 					conv_tnc(pp, ttesz);
6154 				} else if (pp->p_mapping == NULL) {
6155 					ASSERT(kpm_enable);
6156 					/*
6157 					 * Page is marked to be in VAC conflict
6158 					 * to an existing kpm mapping and/or is
6159 					 * kpm mapped using only the regular
6160 					 * pagesize.
6161 					 */
6162 					sfmmu_kpm_hme_unload(pp);
6163 				}
6164 			}
6165 #endif	/* VAC */
6166 		} else if ((pp = sfhmep->hme_page) != NULL) {
6167 				/*
6168 				 * TTE is invalid but the hme
6169 				 * still exists. let pageunload
6170 				 * complete its job.
6171 				 */
6172 				ASSERT(pml == NULL);
6173 				pml = sfmmu_mlist_enter(pp);
6174 				if (sfhmep->hme_page != NULL) {
6175 					sfmmu_mlist_exit(pml);
6176 					continue;
6177 				}
6178 				ASSERT(sfhmep->hme_page == NULL);
6179 		} else if (hmeblkp->hblk_hmecnt != 0) {
6180 			/*
6181 			 * pageunload may have not finished decrementing
6182 			 * hblk_vcnt and hblk_hmecnt. Find page_t if any and
6183 			 * wait for pageunload to finish. Rely on pageunload
6184 			 * to decrement hblk_hmecnt after hblk_vcnt.
6185 			 */
6186 			pfn_t pfn = TTE_TO_TTEPFN(&tte);
6187 			ASSERT(pml == NULL);
6188 			if (pf_is_memory(pfn)) {
6189 				pp = page_numtopp_nolock(pfn);
6190 				if (pp != NULL) {
6191 					pml = sfmmu_mlist_enter(pp);
6192 					sfmmu_mlist_exit(pml);
6193 					pml = NULL;
6194 				}
6195 			}
6196 		}
6197 
6198 tte_unloaded:
6199 		/*
6200 		 * At this point, the tte we are looking at
6201 		 * should be unloaded, and hme has been unlinked
6202 		 * from page too. This is important because in
6203 		 * pageunload, it does ttesync() then HME_SUB.
6204 		 * We need to make sure HME_SUB has been completed
6205 		 * so we know ttesync() has been completed. Otherwise,
6206 		 * at exit time, after return from hat layer, VM will
6207 		 * release as structure which hat_setstat() (called
6208 		 * by ttesync()) needs.
6209 		 */
6210 #ifdef DEBUG
6211 		{
6212 			tte_t	dtte;
6213 
6214 			ASSERT(sfhmep->hme_page == NULL);
6215 
6216 			sfmmu_copytte(&sfhmep->hme_tte, &dtte);
6217 			ASSERT(!TTE_IS_VALID(&dtte));
6218 		}
6219 #endif
6220 
6221 		if (pml) {
6222 			sfmmu_mlist_exit(pml);
6223 		}
6224 
6225 		addr += TTEBYTES(ttesz);
6226 		sfhmep++;
6227 		DEMAP_RANGE_NEXTPG(dmrp);
6228 	}
6229 	/*
6230 	 * For shared hmeblks this routine is only called when region is freed
6231 	 * and no longer referenced.  So no need to decrement ttecnt
6232 	 * in the region structure here.
6233 	 */
6234 	if (ttecnt > 0 && sfmmup != NULL) {
6235 		atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -ttecnt);
6236 	}
6237 	return (addr);
6238 }
6239 
6240 /*
6241  * Synchronize all the mappings in the range [addr..addr+len).
6242  * Can be called with clearflag having two states:
6243  * HAT_SYNC_DONTZERO means just return the rm stats
6244  * HAT_SYNC_ZERORM means zero rm bits in the tte and return the stats
6245  */
6246 void
6247 hat_sync(struct hat *sfmmup, caddr_t addr, size_t len, uint_t clearflag)
6248 {
6249 	struct hmehash_bucket *hmebp;
6250 	hmeblk_tag hblktag;
6251 	int hmeshift, hashno = 1;
6252 	struct hme_blk *hmeblkp, *list = NULL;
6253 	caddr_t endaddr;
6254 	cpuset_t cpuset;
6255 
6256 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
6257 	ASSERT((sfmmup == ksfmmup) ||
6258 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
6259 	ASSERT((len & MMU_PAGEOFFSET) == 0);
6260 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
6261 	    (clearflag == HAT_SYNC_ZERORM));
6262 
6263 	CPUSET_ZERO(cpuset);
6264 
6265 	endaddr = addr + len;
6266 	hblktag.htag_id = sfmmup;
6267 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
6268 
6269 	/*
6270 	 * Spitfire supports 4 page sizes.
6271 	 * Most pages are expected to be of the smallest page
6272 	 * size (8K) and these will not need to be rehashed. 64K
6273 	 * pages also don't need to be rehashed because the an hmeblk
6274 	 * spans 64K of address space. 512K pages might need 1 rehash and
6275 	 * and 4M pages 2 rehashes.
6276 	 */
6277 	while (addr < endaddr) {
6278 		hmeshift = HME_HASH_SHIFT(hashno);
6279 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
6280 		hblktag.htag_rehash = hashno;
6281 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
6282 
6283 		SFMMU_HASH_LOCK(hmebp);
6284 
6285 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
6286 		if (hmeblkp != NULL) {
6287 			ASSERT(!hmeblkp->hblk_shared);
6288 			/*
6289 			 * We've encountered a shadow hmeblk so skip the range
6290 			 * of the next smaller mapping size.
6291 			 */
6292 			if (hmeblkp->hblk_shw_bit) {
6293 				ASSERT(sfmmup != ksfmmup);
6294 				ASSERT(hashno > 1);
6295 				addr = (caddr_t)P2END((uintptr_t)addr,
6296 				    TTEBYTES(hashno - 1));
6297 			} else {
6298 				addr = sfmmu_hblk_sync(sfmmup, hmeblkp,
6299 				    addr, endaddr, clearflag);
6300 			}
6301 			SFMMU_HASH_UNLOCK(hmebp);
6302 			hashno = 1;
6303 			continue;
6304 		}
6305 		SFMMU_HASH_UNLOCK(hmebp);
6306 
6307 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
6308 			/*
6309 			 * We have traversed the whole list and rehashed
6310 			 * if necessary without finding the address to sync.
6311 			 * This is ok so we increment the address by the
6312 			 * smallest hmeblk range for kernel mappings and the
6313 			 * largest hmeblk range, to account for shadow hmeblks,
6314 			 * for user mappings and continue.
6315 			 */
6316 			if (sfmmup == ksfmmup)
6317 				addr = (caddr_t)P2END((uintptr_t)addr,
6318 				    TTEBYTES(1));
6319 			else
6320 				addr = (caddr_t)P2END((uintptr_t)addr,
6321 				    TTEBYTES(hashno));
6322 			hashno = 1;
6323 		} else {
6324 			hashno++;
6325 		}
6326 	}
6327 	sfmmu_hblks_list_purge(&list, 0);
6328 	cpuset = sfmmup->sfmmu_cpusran;
6329 	xt_sync(cpuset);
6330 }
6331 
6332 static caddr_t
6333 sfmmu_hblk_sync(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
6334 	caddr_t endaddr, int clearflag)
6335 {
6336 	tte_t	tte, ttemod;
6337 	struct sf_hment *sfhmep;
6338 	int ttesz;
6339 	struct page *pp;
6340 	kmutex_t *pml;
6341 	int ret;
6342 
6343 	ASSERT(hmeblkp->hblk_shw_bit == 0);
6344 	ASSERT(!hmeblkp->hblk_shared);
6345 
6346 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
6347 
6348 	ttesz = get_hblk_ttesz(hmeblkp);
6349 	HBLKTOHME(sfhmep, hmeblkp, addr);
6350 
6351 	while (addr < endaddr) {
6352 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6353 		if (TTE_IS_VALID(&tte)) {
6354 			pml = NULL;
6355 			pp = sfhmep->hme_page;
6356 			if (pp) {
6357 				pml = sfmmu_mlist_enter(pp);
6358 			}
6359 			if (pp != sfhmep->hme_page) {
6360 				/*
6361 				 * tte most have been unloaded
6362 				 * underneath us.  Recheck
6363 				 */
6364 				ASSERT(pml);
6365 				sfmmu_mlist_exit(pml);
6366 				continue;
6367 			}
6368 
6369 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
6370 
6371 			if (clearflag == HAT_SYNC_ZERORM) {
6372 				ttemod = tte;
6373 				TTE_CLR_RM(&ttemod);
6374 				ret = sfmmu_modifytte_try(&tte, &ttemod,
6375 				    &sfhmep->hme_tte);
6376 				if (ret < 0) {
6377 					if (pml) {
6378 						sfmmu_mlist_exit(pml);
6379 					}
6380 					continue;
6381 				}
6382 
6383 				if (ret > 0) {
6384 					sfmmu_tlb_demap(addr, sfmmup,
6385 					    hmeblkp, 0, 0);
6386 				}
6387 			}
6388 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
6389 			if (pml) {
6390 				sfmmu_mlist_exit(pml);
6391 			}
6392 		}
6393 		addr += TTEBYTES(ttesz);
6394 		sfhmep++;
6395 	}
6396 	return (addr);
6397 }
6398 
6399 /*
6400  * This function will sync a tte to the page struct and it will
6401  * update the hat stats. Currently it allows us to pass a NULL pp
6402  * and we will simply update the stats.  We may want to change this
6403  * so we only keep stats for pages backed by pp's.
6404  */
6405 static void
6406 sfmmu_ttesync(struct hat *sfmmup, caddr_t addr, tte_t *ttep, page_t *pp)
6407 {
6408 	uint_t rm = 0;
6409 	int sz = TTE_CSZ(ttep);
6410 	pgcnt_t	npgs;
6411 
6412 	ASSERT(TTE_IS_VALID(ttep));
6413 
6414 	if (!TTE_IS_NOSYNC(ttep)) {
6415 
6416 		if (TTE_IS_REF(ttep))
6417 			rm |= P_REF;
6418 
6419 		if (TTE_IS_MOD(ttep))
6420 			rm |= P_MOD;
6421 
6422 		if (rm != 0) {
6423 			if (sfmmup != NULL && sfmmup->sfmmu_rmstat) {
6424 				int i;
6425 				caddr_t	vaddr = addr;
6426 
6427 				for (i = 0; i < TTEPAGES(sz); i++) {
6428 					hat_setstat(sfmmup->sfmmu_as, vaddr,
6429 					    MMU_PAGESIZE, rm);
6430 					vaddr += MMU_PAGESIZE;
6431 				}
6432 			}
6433 		}
6434 	}
6435 
6436 	if (!pp)
6437 		return;
6438 
6439 	/*
6440 	 * If software says this page is executable, and the page was
6441 	 * in fact executed (indicated by hardware exec permission
6442 	 * being enabled), then set P_EXEC on the page to remember
6443 	 * that it was executed. The I$ will be flushed when the page
6444 	 * is reassigned.
6445 	 */
6446 	if (TTE_EXECUTED(ttep)) {
6447 		rm |= P_EXEC;
6448 	} else if (rm == 0) {
6449 		return;
6450 	}
6451 
6452 	/*
6453 	 * XXX I want to use cas to update nrm bits but they
6454 	 * currently belong in common/vm and not in hat where
6455 	 * they should be.
6456 	 * The nrm bits are protected by the same mutex as
6457 	 * the one that protects the page's mapping list.
6458 	 */
6459 	ASSERT(sfmmu_mlist_held(pp));
6460 	/*
6461 	 * If the tte is for a large page, we need to sync all the
6462 	 * pages covered by the tte.
6463 	 */
6464 	if (sz != TTE8K) {
6465 		ASSERT(pp->p_szc != 0);
6466 		pp = PP_GROUPLEADER(pp, sz);
6467 		ASSERT(sfmmu_mlist_held(pp));
6468 	}
6469 
6470 	/* Get number of pages from tte size. */
6471 	npgs = TTEPAGES(sz);
6472 
6473 	do {
6474 		ASSERT(pp);
6475 		ASSERT(sfmmu_mlist_held(pp));
6476 		if (((rm & P_REF) != 0 && !PP_ISREF(pp)) ||
6477 		    ((rm & P_MOD) != 0 && !PP_ISMOD(pp)) ||
6478 		    ((rm & P_EXEC) != 0 && !PP_ISEXEC(pp)))
6479 			hat_page_setattr(pp, rm);
6480 
6481 		/*
6482 		 * Are we done? If not, we must have a large mapping.
6483 		 * For large mappings we need to sync the rest of the pages
6484 		 * covered by this tte; goto the next page.
6485 		 */
6486 	} while (--npgs > 0 && (pp = PP_PAGENEXT(pp)));
6487 }
6488 
6489 /*
6490  * Execute pre-callback handler of each pa_hment linked to pp
6491  *
6492  * Inputs:
6493  *   flag: either HAT_PRESUSPEND or HAT_SUSPEND.
6494  *   capture_cpus: pointer to return value (below)
6495  *
6496  * Returns:
6497  *   Propagates the subsystem callback return values back to the caller;
6498  *   returns 0 on success.  If capture_cpus is non-NULL, the value returned
6499  *   is zero if all of the pa_hments are of a type that do not require
6500  *   capturing CPUs prior to suspending the mapping, else it is 1.
6501  */
6502 static int
6503 hat_pageprocess_precallbacks(struct page *pp, uint_t flag, int *capture_cpus)
6504 {
6505 	struct sf_hment	*sfhmep;
6506 	struct pa_hment *pahmep;
6507 	int (*f)(caddr_t, uint_t, uint_t, void *);
6508 	int		ret;
6509 	id_t		id;
6510 	int		locked = 0;
6511 	kmutex_t	*pml;
6512 
6513 	ASSERT(PAGE_EXCL(pp));
6514 	if (!sfmmu_mlist_held(pp)) {
6515 		pml = sfmmu_mlist_enter(pp);
6516 		locked = 1;
6517 	}
6518 
6519 	if (capture_cpus)
6520 		*capture_cpus = 0;
6521 
6522 top:
6523 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6524 		/*
6525 		 * skip sf_hments corresponding to VA<->PA mappings;
6526 		 * for pa_hment's, hme_tte.ll is zero
6527 		 */
6528 		if (!IS_PAHME(sfhmep))
6529 			continue;
6530 
6531 		pahmep = sfhmep->hme_data;
6532 		ASSERT(pahmep != NULL);
6533 
6534 		/*
6535 		 * skip if pre-handler has been called earlier in this loop
6536 		 */
6537 		if (pahmep->flags & flag)
6538 			continue;
6539 
6540 		id = pahmep->cb_id;
6541 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6542 		if (capture_cpus && sfmmu_cb_table[id].capture_cpus != 0)
6543 			*capture_cpus = 1;
6544 		if ((f = sfmmu_cb_table[id].prehandler) == NULL) {
6545 			pahmep->flags |= flag;
6546 			continue;
6547 		}
6548 
6549 		/*
6550 		 * Drop the mapping list lock to avoid locking order issues.
6551 		 */
6552 		if (locked)
6553 			sfmmu_mlist_exit(pml);
6554 
6555 		ret = f(pahmep->addr, pahmep->len, flag, pahmep->pvt);
6556 		if (ret != 0)
6557 			return (ret);	/* caller must do the cleanup */
6558 
6559 		if (locked) {
6560 			pml = sfmmu_mlist_enter(pp);
6561 			pahmep->flags |= flag;
6562 			goto top;
6563 		}
6564 
6565 		pahmep->flags |= flag;
6566 	}
6567 
6568 	if (locked)
6569 		sfmmu_mlist_exit(pml);
6570 
6571 	return (0);
6572 }
6573 
6574 /*
6575  * Execute post-callback handler of each pa_hment linked to pp
6576  *
6577  * Same overall assumptions and restrictions apply as for
6578  * hat_pageprocess_precallbacks().
6579  */
6580 static void
6581 hat_pageprocess_postcallbacks(struct page *pp, uint_t flag)
6582 {
6583 	pfn_t pgpfn = pp->p_pagenum;
6584 	pfn_t pgmask = btop(page_get_pagesize(pp->p_szc)) - 1;
6585 	pfn_t newpfn;
6586 	struct sf_hment *sfhmep;
6587 	struct pa_hment *pahmep;
6588 	int (*f)(caddr_t, uint_t, uint_t, void *, pfn_t);
6589 	id_t	id;
6590 	int	locked = 0;
6591 	kmutex_t *pml;
6592 
6593 	ASSERT(PAGE_EXCL(pp));
6594 	if (!sfmmu_mlist_held(pp)) {
6595 		pml = sfmmu_mlist_enter(pp);
6596 		locked = 1;
6597 	}
6598 
6599 top:
6600 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6601 		/*
6602 		 * skip sf_hments corresponding to VA<->PA mappings;
6603 		 * for pa_hment's, hme_tte.ll is zero
6604 		 */
6605 		if (!IS_PAHME(sfhmep))
6606 			continue;
6607 
6608 		pahmep = sfhmep->hme_data;
6609 		ASSERT(pahmep != NULL);
6610 
6611 		if ((pahmep->flags & flag) == 0)
6612 			continue;
6613 
6614 		pahmep->flags &= ~flag;
6615 
6616 		id = pahmep->cb_id;
6617 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6618 		if ((f = sfmmu_cb_table[id].posthandler) == NULL)
6619 			continue;
6620 
6621 		/*
6622 		 * Convert the base page PFN into the constituent PFN
6623 		 * which is needed by the callback handler.
6624 		 */
6625 		newpfn = pgpfn | (btop((uintptr_t)pahmep->addr) & pgmask);
6626 
6627 		/*
6628 		 * Drop the mapping list lock to avoid locking order issues.
6629 		 */
6630 		if (locked)
6631 			sfmmu_mlist_exit(pml);
6632 
6633 		if (f(pahmep->addr, pahmep->len, flag, pahmep->pvt, newpfn)
6634 		    != 0)
6635 			panic("sfmmu: posthandler failed");
6636 
6637 		if (locked) {
6638 			pml = sfmmu_mlist_enter(pp);
6639 			goto top;
6640 		}
6641 	}
6642 
6643 	if (locked)
6644 		sfmmu_mlist_exit(pml);
6645 }
6646 
6647 /*
6648  * Suspend locked kernel mapping
6649  */
6650 void
6651 hat_pagesuspend(struct page *pp)
6652 {
6653 	struct sf_hment *sfhmep;
6654 	sfmmu_t *sfmmup;
6655 	tte_t tte, ttemod;
6656 	struct hme_blk *hmeblkp;
6657 	caddr_t addr;
6658 	int index, cons;
6659 	cpuset_t cpuset;
6660 
6661 	ASSERT(PAGE_EXCL(pp));
6662 	ASSERT(sfmmu_mlist_held(pp));
6663 
6664 	mutex_enter(&kpr_suspendlock);
6665 
6666 	/*
6667 	 * We're about to suspend a kernel mapping so mark this thread as
6668 	 * non-traceable by DTrace. This prevents us from running into issues
6669 	 * with probe context trying to touch a suspended page
6670 	 * in the relocation codepath itself.
6671 	 */
6672 	curthread->t_flag |= T_DONTDTRACE;
6673 
6674 	index = PP_MAPINDEX(pp);
6675 	cons = TTE8K;
6676 
6677 retry:
6678 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6679 
6680 		if (IS_PAHME(sfhmep))
6681 			continue;
6682 
6683 		if (get_hblk_ttesz(sfmmu_hmetohblk(sfhmep)) != cons)
6684 			continue;
6685 
6686 		/*
6687 		 * Loop until we successfully set the suspend bit in
6688 		 * the TTE.
6689 		 */
6690 again:
6691 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6692 		ASSERT(TTE_IS_VALID(&tte));
6693 
6694 		ttemod = tte;
6695 		TTE_SET_SUSPEND(&ttemod);
6696 		if (sfmmu_modifytte_try(&tte, &ttemod,
6697 		    &sfhmep->hme_tte) < 0)
6698 			goto again;
6699 
6700 		/*
6701 		 * Invalidate TSB entry
6702 		 */
6703 		hmeblkp = sfmmu_hmetohblk(sfhmep);
6704 
6705 		sfmmup = hblktosfmmu(hmeblkp);
6706 		ASSERT(sfmmup == ksfmmup);
6707 		ASSERT(!hmeblkp->hblk_shared);
6708 
6709 		addr = tte_to_vaddr(hmeblkp, tte);
6710 
6711 		/*
6712 		 * No need to make sure that the TSB for this sfmmu is
6713 		 * not being relocated since it is ksfmmup and thus it
6714 		 * will never be relocated.
6715 		 */
6716 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
6717 
6718 		/*
6719 		 * Update xcall stats
6720 		 */
6721 		cpuset = cpu_ready_set;
6722 		CPUSET_DEL(cpuset, CPU->cpu_id);
6723 
6724 		/* LINTED: constant in conditional context */
6725 		SFMMU_XCALL_STATS(ksfmmup);
6726 
6727 		/*
6728 		 * Flush TLB entry on remote CPU's
6729 		 */
6730 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
6731 		    (uint64_t)ksfmmup);
6732 		xt_sync(cpuset);
6733 
6734 		/*
6735 		 * Flush TLB entry on local CPU
6736 		 */
6737 		vtag_flushpage(addr, (uint64_t)ksfmmup);
6738 	}
6739 
6740 	while (index != 0) {
6741 		index = index >> 1;
6742 		if (index != 0)
6743 			cons++;
6744 		if (index & 0x1) {
6745 			pp = PP_GROUPLEADER(pp, cons);
6746 			goto retry;
6747 		}
6748 	}
6749 }
6750 
6751 #ifdef	DEBUG
6752 
6753 #define	N_PRLE	1024
6754 struct prle {
6755 	page_t *targ;
6756 	page_t *repl;
6757 	int status;
6758 	int pausecpus;
6759 	hrtime_t whence;
6760 };
6761 
6762 static struct prle page_relocate_log[N_PRLE];
6763 static int prl_entry;
6764 static kmutex_t prl_mutex;
6765 
6766 #define	PAGE_RELOCATE_LOG(t, r, s, p)					\
6767 	mutex_enter(&prl_mutex);					\
6768 	page_relocate_log[prl_entry].targ = *(t);			\
6769 	page_relocate_log[prl_entry].repl = *(r);			\
6770 	page_relocate_log[prl_entry].status = (s);			\
6771 	page_relocate_log[prl_entry].pausecpus = (p);			\
6772 	page_relocate_log[prl_entry].whence = gethrtime();		\
6773 	prl_entry = (prl_entry == (N_PRLE - 1))? 0 : prl_entry + 1;	\
6774 	mutex_exit(&prl_mutex);
6775 
6776 #else	/* !DEBUG */
6777 #define	PAGE_RELOCATE_LOG(t, r, s, p)
6778 #endif
6779 
6780 /*
6781  * Core Kernel Page Relocation Algorithm
6782  *
6783  * Input:
6784  *
6785  * target : 	constituent pages are SE_EXCL locked.
6786  * replacement:	constituent pages are SE_EXCL locked.
6787  *
6788  * Output:
6789  *
6790  * nrelocp:	number of pages relocated
6791  */
6792 int
6793 hat_page_relocate(page_t **target, page_t **replacement, spgcnt_t *nrelocp)
6794 {
6795 	page_t		*targ, *repl;
6796 	page_t		*tpp, *rpp;
6797 	kmutex_t	*low, *high;
6798 	spgcnt_t	npages, i;
6799 	page_t		*pl = NULL;
6800 	uint_t		ppattr;
6801 	int		old_pil;
6802 	cpuset_t	cpuset;
6803 	int		cap_cpus;
6804 	int		ret;
6805 #ifdef VAC
6806 	int		cflags = 0;
6807 #endif
6808 
6809 	if (hat_kpr_enabled == 0 || !kcage_on || PP_ISNORELOC(*target)) {
6810 		PAGE_RELOCATE_LOG(target, replacement, EAGAIN, -1);
6811 		return (EAGAIN);
6812 	}
6813 
6814 	mutex_enter(&kpr_mutex);
6815 	kreloc_thread = curthread;
6816 
6817 	targ = *target;
6818 	repl = *replacement;
6819 	ASSERT(repl != NULL);
6820 	ASSERT(targ->p_szc == repl->p_szc);
6821 
6822 	npages = page_get_pagecnt(targ->p_szc);
6823 
6824 	/*
6825 	 * unload VA<->PA mappings that are not locked
6826 	 */
6827 	tpp = targ;
6828 	for (i = 0; i < npages; i++) {
6829 		(void) hat_pageunload(tpp, SFMMU_KERNEL_RELOC);
6830 		tpp++;
6831 	}
6832 
6833 	/*
6834 	 * Do "presuspend" callbacks, in a context from which we can still
6835 	 * block as needed. Note that we don't hold the mapping list lock
6836 	 * of "targ" at this point due to potential locking order issues;
6837 	 * we assume that between the hat_pageunload() above and holding
6838 	 * the SE_EXCL lock that the mapping list *cannot* change at this
6839 	 * point.
6840 	 */
6841 	ret = hat_pageprocess_precallbacks(targ, HAT_PRESUSPEND, &cap_cpus);
6842 	if (ret != 0) {
6843 		/*
6844 		 * EIO translates to fatal error, for all others cleanup
6845 		 * and return EAGAIN.
6846 		 */
6847 		ASSERT(ret != EIO);
6848 		hat_pageprocess_postcallbacks(targ, HAT_POSTUNSUSPEND);
6849 		PAGE_RELOCATE_LOG(target, replacement, ret, -1);
6850 		kreloc_thread = NULL;
6851 		mutex_exit(&kpr_mutex);
6852 		return (EAGAIN);
6853 	}
6854 
6855 	/*
6856 	 * acquire p_mapping list lock for both the target and replacement
6857 	 * root pages.
6858 	 *
6859 	 * low and high refer to the need to grab the mlist locks in a
6860 	 * specific order in order to prevent race conditions.  Thus the
6861 	 * lower lock must be grabbed before the higher lock.
6862 	 *
6863 	 * This will block hat_unload's accessing p_mapping list.  Since
6864 	 * we have SE_EXCL lock, hat_memload and hat_pageunload will be
6865 	 * blocked.  Thus, no one else will be accessing the p_mapping list
6866 	 * while we suspend and reload the locked mapping below.
6867 	 */
6868 	tpp = targ;
6869 	rpp = repl;
6870 	sfmmu_mlist_reloc_enter(tpp, rpp, &low, &high);
6871 
6872 	kpreempt_disable();
6873 
6874 	/*
6875 	 * We raise our PIL to 13 so that we don't get captured by
6876 	 * another CPU or pinned by an interrupt thread.  We can't go to
6877 	 * PIL 14 since the nexus driver(s) may need to interrupt at
6878 	 * that level in the case of IOMMU pseudo mappings.
6879 	 */
6880 	cpuset = cpu_ready_set;
6881 	CPUSET_DEL(cpuset, CPU->cpu_id);
6882 	if (!cap_cpus || CPUSET_ISNULL(cpuset)) {
6883 		old_pil = splr(XCALL_PIL);
6884 	} else {
6885 		old_pil = -1;
6886 		xc_attention(cpuset);
6887 	}
6888 	ASSERT(getpil() == XCALL_PIL);
6889 
6890 	/*
6891 	 * Now do suspend callbacks. In the case of an IOMMU mapping
6892 	 * this will suspend all DMA activity to the page while it is
6893 	 * being relocated. Since we are well above LOCK_LEVEL and CPUs
6894 	 * may be captured at this point we should have acquired any needed
6895 	 * locks in the presuspend callback.
6896 	 */
6897 	ret = hat_pageprocess_precallbacks(targ, HAT_SUSPEND, NULL);
6898 	if (ret != 0) {
6899 		repl = targ;
6900 		goto suspend_fail;
6901 	}
6902 
6903 	/*
6904 	 * Raise the PIL yet again, this time to block all high-level
6905 	 * interrupts on this CPU. This is necessary to prevent an
6906 	 * interrupt routine from pinning the thread which holds the
6907 	 * mapping suspended and then touching the suspended page.
6908 	 *
6909 	 * Once the page is suspended we also need to be careful to
6910 	 * avoid calling any functions which touch any seg_kmem memory
6911 	 * since that memory may be backed by the very page we are
6912 	 * relocating in here!
6913 	 */
6914 	hat_pagesuspend(targ);
6915 
6916 	/*
6917 	 * Now that we are confident everybody has stopped using this page,
6918 	 * copy the page contents.  Note we use a physical copy to prevent
6919 	 * locking issues and to avoid fpRAS because we can't handle it in
6920 	 * this context.
6921 	 */
6922 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6923 #ifdef VAC
6924 		/*
6925 		 * If the replacement has a different vcolor than
6926 		 * the one being replacd, we need to handle VAC
6927 		 * consistency for it just as we were setting up
6928 		 * a new mapping to it.
6929 		 */
6930 		if ((PP_GET_VCOLOR(rpp) != NO_VCOLOR) &&
6931 		    (tpp->p_vcolor != rpp->p_vcolor) &&
6932 		    !CacheColor_IsFlushed(cflags, PP_GET_VCOLOR(rpp))) {
6933 			CacheColor_SetFlushed(cflags, PP_GET_VCOLOR(rpp));
6934 			sfmmu_cache_flushcolor(PP_GET_VCOLOR(rpp),
6935 			    rpp->p_pagenum);
6936 		}
6937 #endif
6938 		/*
6939 		 * Copy the contents of the page.
6940 		 */
6941 		ppcopy_kernel(tpp, rpp);
6942 	}
6943 
6944 	tpp = targ;
6945 	rpp = repl;
6946 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6947 		/*
6948 		 * Copy attributes.  VAC consistency was handled above,
6949 		 * if required.
6950 		 */
6951 		ppattr = hat_page_getattr(tpp, (P_MOD | P_REF | P_RO));
6952 		page_clr_all_props(rpp, 0);
6953 		page_set_props(rpp, ppattr);
6954 		rpp->p_index = tpp->p_index;
6955 		tpp->p_index = 0;
6956 #ifdef VAC
6957 		rpp->p_vcolor = tpp->p_vcolor;
6958 #endif
6959 	}
6960 
6961 	/*
6962 	 * First, unsuspend the page, if we set the suspend bit, and transfer
6963 	 * the mapping list from the target page to the replacement page.
6964 	 * Next process postcallbacks; since pa_hment's are linked only to the
6965 	 * p_mapping list of root page, we don't iterate over the constituent
6966 	 * pages.
6967 	 */
6968 	hat_pagereload(targ, repl);
6969 
6970 suspend_fail:
6971 	hat_pageprocess_postcallbacks(repl, HAT_UNSUSPEND);
6972 
6973 	/*
6974 	 * Now lower our PIL and release any captured CPUs since we
6975 	 * are out of the "danger zone".  After this it will again be
6976 	 * safe to acquire adaptive mutex locks, or to drop them...
6977 	 */
6978 	if (old_pil != -1) {
6979 		splx(old_pil);
6980 	} else {
6981 		xc_dismissed(cpuset);
6982 	}
6983 
6984 	kpreempt_enable();
6985 
6986 	sfmmu_mlist_reloc_exit(low, high);
6987 
6988 	/*
6989 	 * Postsuspend callbacks should drop any locks held across
6990 	 * the suspend callbacks.  As before, we don't hold the mapping
6991 	 * list lock at this point.. our assumption is that the mapping
6992 	 * list still can't change due to our holding SE_EXCL lock and
6993 	 * there being no unlocked mappings left. Hence the restriction
6994 	 * on calling context to hat_delete_callback()
6995 	 */
6996 	hat_pageprocess_postcallbacks(repl, HAT_POSTUNSUSPEND);
6997 	if (ret != 0) {
6998 		/*
6999 		 * The second presuspend call failed: we got here through
7000 		 * the suspend_fail label above.
7001 		 */
7002 		ASSERT(ret != EIO);
7003 		PAGE_RELOCATE_LOG(target, replacement, ret, cap_cpus);
7004 		kreloc_thread = NULL;
7005 		mutex_exit(&kpr_mutex);
7006 		return (EAGAIN);
7007 	}
7008 
7009 	/*
7010 	 * Now that we're out of the performance critical section we can
7011 	 * take care of updating the hash table, since we still
7012 	 * hold all the pages locked SE_EXCL at this point we
7013 	 * needn't worry about things changing out from under us.
7014 	 */
7015 	tpp = targ;
7016 	rpp = repl;
7017 	for (i = 0; i < npages; i++, tpp++, rpp++) {
7018 
7019 		/*
7020 		 * replace targ with replacement in page_hash table
7021 		 */
7022 		targ = tpp;
7023 		page_relocate_hash(rpp, targ);
7024 
7025 		/*
7026 		 * concatenate target; caller of platform_page_relocate()
7027 		 * expects target to be concatenated after returning.
7028 		 */
7029 		ASSERT(targ->p_next == targ);
7030 		ASSERT(targ->p_prev == targ);
7031 		page_list_concat(&pl, &targ);
7032 	}
7033 
7034 	ASSERT(*target == pl);
7035 	*nrelocp = npages;
7036 	PAGE_RELOCATE_LOG(target, replacement, 0, cap_cpus);
7037 	kreloc_thread = NULL;
7038 	mutex_exit(&kpr_mutex);
7039 	return (0);
7040 }
7041 
7042 /*
7043  * Called when stray pa_hments are found attached to a page which is
7044  * being freed.  Notify the subsystem which attached the pa_hment of
7045  * the error if it registered a suitable handler, else panic.
7046  */
7047 static void
7048 sfmmu_pahment_leaked(struct pa_hment *pahmep)
7049 {
7050 	id_t cb_id = pahmep->cb_id;
7051 
7052 	ASSERT(cb_id >= (id_t)0 && cb_id < sfmmu_cb_nextid);
7053 	if (sfmmu_cb_table[cb_id].errhandler != NULL) {
7054 		if (sfmmu_cb_table[cb_id].errhandler(pahmep->addr, pahmep->len,
7055 		    HAT_CB_ERR_LEAKED, pahmep->pvt) == 0)
7056 			return;		/* non-fatal */
7057 	}
7058 	panic("pa_hment leaked: 0x%p", (void *)pahmep);
7059 }
7060 
7061 /*
7062  * Remove all mappings to page 'pp'.
7063  */
7064 int
7065 hat_pageunload(struct page *pp, uint_t forceflag)
7066 {
7067 	struct page *origpp = pp;
7068 	struct sf_hment *sfhme, *tmphme;
7069 	struct hme_blk *hmeblkp;
7070 	kmutex_t *pml;
7071 #ifdef VAC
7072 	kmutex_t *pmtx;
7073 #endif
7074 	cpuset_t cpuset, tset;
7075 	int index, cons;
7076 	int xhme_blks;
7077 	int pa_hments;
7078 
7079 	ASSERT(PAGE_EXCL(pp));
7080 
7081 retry_xhat:
7082 	tmphme = NULL;
7083 	xhme_blks = 0;
7084 	pa_hments = 0;
7085 	CPUSET_ZERO(cpuset);
7086 
7087 	pml = sfmmu_mlist_enter(pp);
7088 
7089 #ifdef VAC
7090 	if (pp->p_kpmref)
7091 		sfmmu_kpm_pageunload(pp);
7092 	ASSERT(!PP_ISMAPPED_KPM(pp));
7093 #endif
7094 
7095 	index = PP_MAPINDEX(pp);
7096 	cons = TTE8K;
7097 retry:
7098 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7099 		tmphme = sfhme->hme_next;
7100 
7101 		if (IS_PAHME(sfhme)) {
7102 			ASSERT(sfhme->hme_data != NULL);
7103 			pa_hments++;
7104 			continue;
7105 		}
7106 
7107 		hmeblkp = sfmmu_hmetohblk(sfhme);
7108 		if (hmeblkp->hblk_xhat_bit) {
7109 			struct xhat_hme_blk *xblk =
7110 			    (struct xhat_hme_blk *)hmeblkp;
7111 
7112 			(void) XHAT_PAGEUNLOAD(xblk->xhat_hme_blk_hat,
7113 			    pp, forceflag, XBLK2PROVBLK(xblk));
7114 
7115 			xhme_blks = 1;
7116 			continue;
7117 		}
7118 
7119 		/*
7120 		 * If there are kernel mappings don't unload them, they will
7121 		 * be suspended.
7122 		 */
7123 		if (forceflag == SFMMU_KERNEL_RELOC && hmeblkp->hblk_lckcnt &&
7124 		    hmeblkp->hblk_tag.htag_id == ksfmmup)
7125 			continue;
7126 
7127 		tset = sfmmu_pageunload(pp, sfhme, cons);
7128 		CPUSET_OR(cpuset, tset);
7129 	}
7130 
7131 	while (index != 0) {
7132 		index = index >> 1;
7133 		if (index != 0)
7134 			cons++;
7135 		if (index & 0x1) {
7136 			/* Go to leading page */
7137 			pp = PP_GROUPLEADER(pp, cons);
7138 			ASSERT(sfmmu_mlist_held(pp));
7139 			goto retry;
7140 		}
7141 	}
7142 
7143 	/*
7144 	 * cpuset may be empty if the page was only mapped by segkpm,
7145 	 * in which case we won't actually cross-trap.
7146 	 */
7147 	xt_sync(cpuset);
7148 
7149 	/*
7150 	 * The page should have no mappings at this point, unless
7151 	 * we were called from hat_page_relocate() in which case we
7152 	 * leave the locked mappings which will be suspended later.
7153 	 */
7154 	ASSERT(!PP_ISMAPPED(origpp) || xhme_blks || pa_hments ||
7155 	    (forceflag == SFMMU_KERNEL_RELOC));
7156 
7157 #ifdef VAC
7158 	if (PP_ISTNC(pp)) {
7159 		if (cons == TTE8K) {
7160 			pmtx = sfmmu_page_enter(pp);
7161 			PP_CLRTNC(pp);
7162 			sfmmu_page_exit(pmtx);
7163 		} else {
7164 			conv_tnc(pp, cons);
7165 		}
7166 	}
7167 #endif	/* VAC */
7168 
7169 	if (pa_hments && forceflag != SFMMU_KERNEL_RELOC) {
7170 		/*
7171 		 * Unlink any pa_hments and free them, calling back
7172 		 * the responsible subsystem to notify it of the error.
7173 		 * This can occur in situations such as drivers leaking
7174 		 * DMA handles: naughty, but common enough that we'd like
7175 		 * to keep the system running rather than bringing it
7176 		 * down with an obscure error like "pa_hment leaked"
7177 		 * which doesn't aid the user in debugging their driver.
7178 		 */
7179 		for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7180 			tmphme = sfhme->hme_next;
7181 			if (IS_PAHME(sfhme)) {
7182 				struct pa_hment *pahmep = sfhme->hme_data;
7183 				sfmmu_pahment_leaked(pahmep);
7184 				HME_SUB(sfhme, pp);
7185 				kmem_cache_free(pa_hment_cache, pahmep);
7186 			}
7187 		}
7188 
7189 		ASSERT(!PP_ISMAPPED(origpp) || xhme_blks);
7190 	}
7191 
7192 	sfmmu_mlist_exit(pml);
7193 
7194 	/*
7195 	 * XHAT may not have finished unloading pages
7196 	 * because some other thread was waiting for
7197 	 * mlist lock and XHAT_PAGEUNLOAD let it do
7198 	 * the job.
7199 	 */
7200 	if (xhme_blks) {
7201 		pp = origpp;
7202 		goto retry_xhat;
7203 	}
7204 
7205 	return (0);
7206 }
7207 
7208 cpuset_t
7209 sfmmu_pageunload(page_t *pp, struct sf_hment *sfhme, int cons)
7210 {
7211 	struct hme_blk *hmeblkp;
7212 	sfmmu_t *sfmmup;
7213 	tte_t tte, ttemod;
7214 #ifdef DEBUG
7215 	tte_t orig_old;
7216 #endif /* DEBUG */
7217 	caddr_t addr;
7218 	int ttesz;
7219 	int ret;
7220 	cpuset_t cpuset;
7221 
7222 	ASSERT(pp != NULL);
7223 	ASSERT(sfmmu_mlist_held(pp));
7224 	ASSERT(!PP_ISKAS(pp));
7225 
7226 	CPUSET_ZERO(cpuset);
7227 
7228 	hmeblkp = sfmmu_hmetohblk(sfhme);
7229 
7230 readtte:
7231 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7232 	if (TTE_IS_VALID(&tte)) {
7233 		sfmmup = hblktosfmmu(hmeblkp);
7234 		ttesz = get_hblk_ttesz(hmeblkp);
7235 		/*
7236 		 * Only unload mappings of 'cons' size.
7237 		 */
7238 		if (ttesz != cons)
7239 			return (cpuset);
7240 
7241 		/*
7242 		 * Note that we have p_mapping lock, but no hash lock here.
7243 		 * hblk_unload() has to have both hash lock AND p_mapping
7244 		 * lock before it tries to modify tte. So, the tte could
7245 		 * not become invalid in the sfmmu_modifytte_try() below.
7246 		 */
7247 		ttemod = tte;
7248 #ifdef DEBUG
7249 		orig_old = tte;
7250 #endif /* DEBUG */
7251 
7252 		TTE_SET_INVALID(&ttemod);
7253 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7254 		if (ret < 0) {
7255 #ifdef DEBUG
7256 			/* only R/M bits can change. */
7257 			chk_tte(&orig_old, &tte, &ttemod, hmeblkp);
7258 #endif /* DEBUG */
7259 			goto readtte;
7260 		}
7261 
7262 		if (ret == 0) {
7263 			panic("pageunload: cas failed?");
7264 		}
7265 
7266 		addr = tte_to_vaddr(hmeblkp, tte);
7267 
7268 		if (hmeblkp->hblk_shared) {
7269 			sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7270 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
7271 			sf_region_t *rgnp;
7272 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7273 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7274 			ASSERT(srdp != NULL);
7275 			rgnp = srdp->srd_hmergnp[rid];
7276 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
7277 			cpuset = sfmmu_rgntlb_demap(addr, rgnp, hmeblkp, 1);
7278 			sfmmu_ttesync(NULL, addr, &tte, pp);
7279 			ASSERT(rgnp->rgn_ttecnt[ttesz] > 0);
7280 			atomic_add_long(&rgnp->rgn_ttecnt[ttesz], -1);
7281 		} else {
7282 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
7283 			atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -1);
7284 
7285 			/*
7286 			 * We need to flush the page from the virtual cache
7287 			 * in order to prevent a virtual cache alias
7288 			 * inconsistency. The particular scenario we need
7289 			 * to worry about is:
7290 			 * Given:  va1 and va2 are two virtual address that
7291 			 * alias and will map the same physical address.
7292 			 * 1.   mapping exists from va1 to pa and data has
7293 			 *	been read into the cache.
7294 			 * 2.   unload va1.
7295 			 * 3.   load va2 and modify data using va2.
7296 			 * 4    unload va2.
7297 			 * 5.   load va1 and reference data.  Unless we flush
7298 			 *	the data cache when we unload we will get
7299 			 *	stale data.
7300 			 * This scenario is taken care of by using virtual
7301 			 * page coloring.
7302 			 */
7303 			if (sfmmup->sfmmu_ismhat) {
7304 				/*
7305 				 * Flush TSBs, TLBs and caches
7306 				 * of every process
7307 				 * sharing this ism segment.
7308 				 */
7309 				sfmmu_hat_lock_all();
7310 				mutex_enter(&ism_mlist_lock);
7311 				kpreempt_disable();
7312 				sfmmu_ismtlbcache_demap(addr, sfmmup, hmeblkp,
7313 				    pp->p_pagenum, CACHE_NO_FLUSH);
7314 				kpreempt_enable();
7315 				mutex_exit(&ism_mlist_lock);
7316 				sfmmu_hat_unlock_all();
7317 				cpuset = cpu_ready_set;
7318 			} else {
7319 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7320 				cpuset = sfmmup->sfmmu_cpusran;
7321 			}
7322 		}
7323 
7324 		/*
7325 		 * Hme_sub has to run after ttesync() and a_rss update.
7326 		 * See hblk_unload().
7327 		 */
7328 		HME_SUB(sfhme, pp);
7329 		membar_stst();
7330 
7331 		/*
7332 		 * We can not make ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
7333 		 * since pteload may have done a HME_ADD() right after
7334 		 * we did the HME_SUB() above. Hmecnt is now maintained
7335 		 * by cas only. no lock guranteed its value. The only
7336 		 * gurantee we have is the hmecnt should not be less than
7337 		 * what it should be so the hblk will not be taken away.
7338 		 * It's also important that we decremented the hmecnt after
7339 		 * we are done with hmeblkp so that this hmeblk won't be
7340 		 * stolen.
7341 		 */
7342 		ASSERT(hmeblkp->hblk_hmecnt > 0);
7343 		ASSERT(hmeblkp->hblk_vcnt > 0);
7344 		atomic_add_16(&hmeblkp->hblk_vcnt, -1);
7345 		atomic_add_16(&hmeblkp->hblk_hmecnt, -1);
7346 		/*
7347 		 * This is bug 4063182.
7348 		 * XXX: fixme
7349 		 * ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
7350 		 *	!hmeblkp->hblk_lckcnt);
7351 		 */
7352 	} else {
7353 		panic("invalid tte? pp %p &tte %p",
7354 		    (void *)pp, (void *)&tte);
7355 	}
7356 
7357 	return (cpuset);
7358 }
7359 
7360 /*
7361  * While relocating a kernel page, this function will move the mappings
7362  * from tpp to dpp and modify any associated data with these mappings.
7363  * It also unsuspends the suspended kernel mapping.
7364  */
7365 static void
7366 hat_pagereload(struct page *tpp, struct page *dpp)
7367 {
7368 	struct sf_hment *sfhme;
7369 	tte_t tte, ttemod;
7370 	int index, cons;
7371 
7372 	ASSERT(getpil() == PIL_MAX);
7373 	ASSERT(sfmmu_mlist_held(tpp));
7374 	ASSERT(sfmmu_mlist_held(dpp));
7375 
7376 	index = PP_MAPINDEX(tpp);
7377 	cons = TTE8K;
7378 
7379 	/* Update real mappings to the page */
7380 retry:
7381 	for (sfhme = tpp->p_mapping; sfhme != NULL; sfhme = sfhme->hme_next) {
7382 		if (IS_PAHME(sfhme))
7383 			continue;
7384 		sfmmu_copytte(&sfhme->hme_tte, &tte);
7385 		ttemod = tte;
7386 
7387 		/*
7388 		 * replace old pfn with new pfn in TTE
7389 		 */
7390 		PFN_TO_TTE(ttemod, dpp->p_pagenum);
7391 
7392 		/*
7393 		 * clear suspend bit
7394 		 */
7395 		ASSERT(TTE_IS_SUSPEND(&ttemod));
7396 		TTE_CLR_SUSPEND(&ttemod);
7397 
7398 		if (sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte) < 0)
7399 			panic("hat_pagereload(): sfmmu_modifytte_try() failed");
7400 
7401 		/*
7402 		 * set hme_page point to new page
7403 		 */
7404 		sfhme->hme_page = dpp;
7405 	}
7406 
7407 	/*
7408 	 * move p_mapping list from old page to new page
7409 	 */
7410 	dpp->p_mapping = tpp->p_mapping;
7411 	tpp->p_mapping = NULL;
7412 	dpp->p_share = tpp->p_share;
7413 	tpp->p_share = 0;
7414 
7415 	while (index != 0) {
7416 		index = index >> 1;
7417 		if (index != 0)
7418 			cons++;
7419 		if (index & 0x1) {
7420 			tpp = PP_GROUPLEADER(tpp, cons);
7421 			dpp = PP_GROUPLEADER(dpp, cons);
7422 			goto retry;
7423 		}
7424 	}
7425 
7426 	curthread->t_flag &= ~T_DONTDTRACE;
7427 	mutex_exit(&kpr_suspendlock);
7428 }
7429 
7430 uint_t
7431 hat_pagesync(struct page *pp, uint_t clearflag)
7432 {
7433 	struct sf_hment *sfhme, *tmphme = NULL;
7434 	struct hme_blk *hmeblkp;
7435 	kmutex_t *pml;
7436 	cpuset_t cpuset, tset;
7437 	int	index, cons;
7438 	extern	ulong_t po_share;
7439 	page_t	*save_pp = pp;
7440 	int	stop_on_sh = 0;
7441 	uint_t	shcnt;
7442 
7443 	CPUSET_ZERO(cpuset);
7444 
7445 	if (PP_ISRO(pp) && (clearflag & HAT_SYNC_STOPON_MOD)) {
7446 		return (PP_GENERIC_ATTR(pp));
7447 	}
7448 
7449 	if ((clearflag & HAT_SYNC_ZERORM) == 0) {
7450 		if ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(pp)) {
7451 			return (PP_GENERIC_ATTR(pp));
7452 		}
7453 		if ((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(pp)) {
7454 			return (PP_GENERIC_ATTR(pp));
7455 		}
7456 		if (clearflag & HAT_SYNC_STOPON_SHARED) {
7457 			if (pp->p_share > po_share) {
7458 				hat_page_setattr(pp, P_REF);
7459 				return (PP_GENERIC_ATTR(pp));
7460 			}
7461 			stop_on_sh = 1;
7462 			shcnt = 0;
7463 		}
7464 	}
7465 
7466 	clearflag &= ~HAT_SYNC_STOPON_SHARED;
7467 	pml = sfmmu_mlist_enter(pp);
7468 	index = PP_MAPINDEX(pp);
7469 	cons = TTE8K;
7470 retry:
7471 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7472 		/*
7473 		 * We need to save the next hment on the list since
7474 		 * it is possible for pagesync to remove an invalid hment
7475 		 * from the list.
7476 		 */
7477 		tmphme = sfhme->hme_next;
7478 		if (IS_PAHME(sfhme))
7479 			continue;
7480 		/*
7481 		 * If we are looking for large mappings and this hme doesn't
7482 		 * reach the range we are seeking, just ignore it.
7483 		 */
7484 		hmeblkp = sfmmu_hmetohblk(sfhme);
7485 		if (hmeblkp->hblk_xhat_bit)
7486 			continue;
7487 
7488 		if (hme_size(sfhme) < cons)
7489 			continue;
7490 
7491 		if (stop_on_sh) {
7492 			if (hmeblkp->hblk_shared) {
7493 				sf_srd_t *srdp = hblktosrd(hmeblkp);
7494 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7495 				sf_region_t *rgnp;
7496 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7497 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7498 				ASSERT(srdp != NULL);
7499 				rgnp = srdp->srd_hmergnp[rid];
7500 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
7501 				    rgnp, rid);
7502 				shcnt += rgnp->rgn_refcnt;
7503 			} else {
7504 				shcnt++;
7505 			}
7506 			if (shcnt > po_share) {
7507 				/*
7508 				 * tell the pager to spare the page this time
7509 				 * around.
7510 				 */
7511 				hat_page_setattr(save_pp, P_REF);
7512 				index = 0;
7513 				break;
7514 			}
7515 		}
7516 		tset = sfmmu_pagesync(pp, sfhme,
7517 		    clearflag & ~HAT_SYNC_STOPON_RM);
7518 		CPUSET_OR(cpuset, tset);
7519 
7520 		/*
7521 		 * If clearflag is HAT_SYNC_DONTZERO, break out as soon
7522 		 * as the "ref" or "mod" is set or share cnt exceeds po_share.
7523 		 */
7524 		if ((clearflag & ~HAT_SYNC_STOPON_RM) == HAT_SYNC_DONTZERO &&
7525 		    (((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(save_pp)) ||
7526 		    ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(save_pp)))) {
7527 			index = 0;
7528 			break;
7529 		}
7530 	}
7531 
7532 	while (index) {
7533 		index = index >> 1;
7534 		cons++;
7535 		if (index & 0x1) {
7536 			/* Go to leading page */
7537 			pp = PP_GROUPLEADER(pp, cons);
7538 			goto retry;
7539 		}
7540 	}
7541 
7542 	xt_sync(cpuset);
7543 	sfmmu_mlist_exit(pml);
7544 	return (PP_GENERIC_ATTR(save_pp));
7545 }
7546 
7547 /*
7548  * Get all the hardware dependent attributes for a page struct
7549  */
7550 static cpuset_t
7551 sfmmu_pagesync(struct page *pp, struct sf_hment *sfhme,
7552 	uint_t clearflag)
7553 {
7554 	caddr_t addr;
7555 	tte_t tte, ttemod;
7556 	struct hme_blk *hmeblkp;
7557 	int ret;
7558 	sfmmu_t *sfmmup;
7559 	cpuset_t cpuset;
7560 
7561 	ASSERT(pp != NULL);
7562 	ASSERT(sfmmu_mlist_held(pp));
7563 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
7564 	    (clearflag == HAT_SYNC_ZERORM));
7565 
7566 	SFMMU_STAT(sf_pagesync);
7567 
7568 	CPUSET_ZERO(cpuset);
7569 
7570 sfmmu_pagesync_retry:
7571 
7572 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7573 	if (TTE_IS_VALID(&tte)) {
7574 		hmeblkp = sfmmu_hmetohblk(sfhme);
7575 		sfmmup = hblktosfmmu(hmeblkp);
7576 		addr = tte_to_vaddr(hmeblkp, tte);
7577 		if (clearflag == HAT_SYNC_ZERORM) {
7578 			ttemod = tte;
7579 			TTE_CLR_RM(&ttemod);
7580 			ret = sfmmu_modifytte_try(&tte, &ttemod,
7581 			    &sfhme->hme_tte);
7582 			if (ret < 0) {
7583 				/*
7584 				 * cas failed and the new value is not what
7585 				 * we want.
7586 				 */
7587 				goto sfmmu_pagesync_retry;
7588 			}
7589 
7590 			if (ret > 0) {
7591 				/* we win the cas */
7592 				if (hmeblkp->hblk_shared) {
7593 					sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7594 					uint_t rid =
7595 					    hmeblkp->hblk_tag.htag_rid;
7596 					sf_region_t *rgnp;
7597 					ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7598 					ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7599 					ASSERT(srdp != NULL);
7600 					rgnp = srdp->srd_hmergnp[rid];
7601 					SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7602 					    srdp, rgnp, rid);
7603 					cpuset = sfmmu_rgntlb_demap(addr,
7604 					    rgnp, hmeblkp, 1);
7605 				} else {
7606 					sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
7607 					    0, 0);
7608 					cpuset = sfmmup->sfmmu_cpusran;
7609 				}
7610 			}
7611 		}
7612 		sfmmu_ttesync(hmeblkp->hblk_shared ? NULL : sfmmup, addr,
7613 		    &tte, pp);
7614 	}
7615 	return (cpuset);
7616 }
7617 
7618 /*
7619  * Remove write permission from a mappings to a page, so that
7620  * we can detect the next modification of it. This requires modifying
7621  * the TTE then invalidating (demap) any TLB entry using that TTE.
7622  * This code is similar to sfmmu_pagesync().
7623  */
7624 static cpuset_t
7625 sfmmu_pageclrwrt(struct page *pp, struct sf_hment *sfhme)
7626 {
7627 	caddr_t addr;
7628 	tte_t tte;
7629 	tte_t ttemod;
7630 	struct hme_blk *hmeblkp;
7631 	int ret;
7632 	sfmmu_t *sfmmup;
7633 	cpuset_t cpuset;
7634 
7635 	ASSERT(pp != NULL);
7636 	ASSERT(sfmmu_mlist_held(pp));
7637 
7638 	CPUSET_ZERO(cpuset);
7639 	SFMMU_STAT(sf_clrwrt);
7640 
7641 retry:
7642 
7643 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7644 	if (TTE_IS_VALID(&tte) && TTE_IS_WRITABLE(&tte)) {
7645 		hmeblkp = sfmmu_hmetohblk(sfhme);
7646 
7647 		/*
7648 		 * xhat mappings should never be to a VMODSORT page.
7649 		 */
7650 		ASSERT(hmeblkp->hblk_xhat_bit == 0);
7651 
7652 		sfmmup = hblktosfmmu(hmeblkp);
7653 		addr = tte_to_vaddr(hmeblkp, tte);
7654 
7655 		ttemod = tte;
7656 		TTE_CLR_WRT(&ttemod);
7657 		TTE_CLR_MOD(&ttemod);
7658 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7659 
7660 		/*
7661 		 * if cas failed and the new value is not what
7662 		 * we want retry
7663 		 */
7664 		if (ret < 0)
7665 			goto retry;
7666 
7667 		/* we win the cas */
7668 		if (ret > 0) {
7669 			if (hmeblkp->hblk_shared) {
7670 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7671 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7672 				sf_region_t *rgnp;
7673 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7674 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7675 				ASSERT(srdp != NULL);
7676 				rgnp = srdp->srd_hmergnp[rid];
7677 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7678 				    srdp, rgnp, rid);
7679 				cpuset = sfmmu_rgntlb_demap(addr,
7680 				    rgnp, hmeblkp, 1);
7681 			} else {
7682 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7683 				cpuset = sfmmup->sfmmu_cpusran;
7684 			}
7685 		}
7686 	}
7687 
7688 	return (cpuset);
7689 }
7690 
7691 /*
7692  * Walk all mappings of a page, removing write permission and clearing the
7693  * ref/mod bits. This code is similar to hat_pagesync()
7694  */
7695 static void
7696 hat_page_clrwrt(page_t *pp)
7697 {
7698 	struct sf_hment *sfhme;
7699 	struct sf_hment *tmphme = NULL;
7700 	kmutex_t *pml;
7701 	cpuset_t cpuset;
7702 	cpuset_t tset;
7703 	int	index;
7704 	int	 cons;
7705 
7706 	CPUSET_ZERO(cpuset);
7707 
7708 	pml = sfmmu_mlist_enter(pp);
7709 	index = PP_MAPINDEX(pp);
7710 	cons = TTE8K;
7711 retry:
7712 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7713 		tmphme = sfhme->hme_next;
7714 
7715 		/*
7716 		 * If we are looking for large mappings and this hme doesn't
7717 		 * reach the range we are seeking, just ignore its.
7718 		 */
7719 
7720 		if (hme_size(sfhme) < cons)
7721 			continue;
7722 
7723 		tset = sfmmu_pageclrwrt(pp, sfhme);
7724 		CPUSET_OR(cpuset, tset);
7725 	}
7726 
7727 	while (index) {
7728 		index = index >> 1;
7729 		cons++;
7730 		if (index & 0x1) {
7731 			/* Go to leading page */
7732 			pp = PP_GROUPLEADER(pp, cons);
7733 			goto retry;
7734 		}
7735 	}
7736 
7737 	xt_sync(cpuset);
7738 	sfmmu_mlist_exit(pml);
7739 }
7740 
7741 /*
7742  * Set the given REF/MOD/RO bits for the given page.
7743  * For a vnode with a sorted v_pages list, we need to change
7744  * the attributes and the v_pages list together under page_vnode_mutex.
7745  */
7746 void
7747 hat_page_setattr(page_t *pp, uint_t flag)
7748 {
7749 	vnode_t		*vp = pp->p_vnode;
7750 	page_t		**listp;
7751 	kmutex_t	*pmtx;
7752 	kmutex_t	*vphm = NULL;
7753 	int		noshuffle;
7754 
7755 	noshuffle = flag & P_NSH;
7756 	flag &= ~P_NSH;
7757 
7758 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO | P_EXEC)));
7759 
7760 	/*
7761 	 * nothing to do if attribute already set
7762 	 */
7763 	if ((pp->p_nrm & flag) == flag)
7764 		return;
7765 
7766 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp) &&
7767 	    !noshuffle) {
7768 		vphm = page_vnode_mutex(vp);
7769 		mutex_enter(vphm);
7770 	}
7771 
7772 	pmtx = sfmmu_page_enter(pp);
7773 	pp->p_nrm |= flag;
7774 	sfmmu_page_exit(pmtx);
7775 
7776 	if (vphm != NULL) {
7777 		/*
7778 		 * Some File Systems examine v_pages for NULL w/o
7779 		 * grabbing the vphm mutex. Must not let it become NULL when
7780 		 * pp is the only page on the list.
7781 		 */
7782 		if (pp->p_vpnext != pp) {
7783 			page_vpsub(&vp->v_pages, pp);
7784 			if (vp->v_pages != NULL)
7785 				listp = &vp->v_pages->p_vpprev->p_vpnext;
7786 			else
7787 				listp = &vp->v_pages;
7788 			page_vpadd(listp, pp);
7789 		}
7790 		mutex_exit(vphm);
7791 	}
7792 }
7793 
7794 void
7795 hat_page_clrattr(page_t *pp, uint_t flag)
7796 {
7797 	vnode_t		*vp = pp->p_vnode;
7798 	kmutex_t	*pmtx;
7799 
7800 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7801 
7802 	pmtx = sfmmu_page_enter(pp);
7803 
7804 	/*
7805 	 * Caller is expected to hold page's io lock for VMODSORT to work
7806 	 * correctly with pvn_vplist_dirty() and pvn_getdirty() when mod
7807 	 * bit is cleared.
7808 	 * We don't have assert to avoid tripping some existing third party
7809 	 * code. The dirty page is moved back to top of the v_page list
7810 	 * after IO is done in pvn_write_done().
7811 	 */
7812 	pp->p_nrm &= ~flag;
7813 	sfmmu_page_exit(pmtx);
7814 
7815 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) {
7816 
7817 		/*
7818 		 * VMODSORT works by removing write permissions and getting
7819 		 * a fault when a page is made dirty. At this point
7820 		 * we need to remove write permission from all mappings
7821 		 * to this page.
7822 		 */
7823 		hat_page_clrwrt(pp);
7824 	}
7825 }
7826 
7827 uint_t
7828 hat_page_getattr(page_t *pp, uint_t flag)
7829 {
7830 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7831 	return ((uint_t)(pp->p_nrm & flag));
7832 }
7833 
7834 /*
7835  * DEBUG kernels: verify that a kernel va<->pa translation
7836  * is safe by checking the underlying page_t is in a page
7837  * relocation-safe state.
7838  */
7839 #ifdef	DEBUG
7840 void
7841 sfmmu_check_kpfn(pfn_t pfn)
7842 {
7843 	page_t *pp;
7844 	int index, cons;
7845 
7846 	if (hat_check_vtop == 0)
7847 		return;
7848 
7849 	if (hat_kpr_enabled == 0 || kvseg.s_base == NULL || panicstr)
7850 		return;
7851 
7852 	pp = page_numtopp_nolock(pfn);
7853 	if (!pp)
7854 		return;
7855 
7856 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7857 		return;
7858 
7859 	/*
7860 	 * Handed a large kernel page, we dig up the root page since we
7861 	 * know the root page might have the lock also.
7862 	 */
7863 	if (pp->p_szc != 0) {
7864 		index = PP_MAPINDEX(pp);
7865 		cons = TTE8K;
7866 again:
7867 		while (index != 0) {
7868 			index >>= 1;
7869 			if (index != 0)
7870 				cons++;
7871 			if (index & 0x1) {
7872 				pp = PP_GROUPLEADER(pp, cons);
7873 				goto again;
7874 			}
7875 		}
7876 	}
7877 
7878 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7879 		return;
7880 
7881 	/*
7882 	 * Pages need to be locked or allocated "permanent" (either from
7883 	 * static_arena arena or explicitly setting PG_NORELOC when calling
7884 	 * page_create_va()) for VA->PA translations to be valid.
7885 	 */
7886 	if (!PP_ISNORELOC(pp))
7887 		panic("Illegal VA->PA translation, pp 0x%p not permanent",
7888 		    (void *)pp);
7889 	else
7890 		panic("Illegal VA->PA translation, pp 0x%p not locked",
7891 		    (void *)pp);
7892 }
7893 #endif	/* DEBUG */
7894 
7895 /*
7896  * Returns a page frame number for a given virtual address.
7897  * Returns PFN_INVALID to indicate an invalid mapping
7898  */
7899 pfn_t
7900 hat_getpfnum(struct hat *hat, caddr_t addr)
7901 {
7902 	pfn_t pfn;
7903 	tte_t tte;
7904 
7905 	/*
7906 	 * We would like to
7907 	 * ASSERT(AS_LOCK_HELD(as, &as->a_lock));
7908 	 * but we can't because the iommu driver will call this
7909 	 * routine at interrupt time and it can't grab the as lock
7910 	 * or it will deadlock: A thread could have the as lock
7911 	 * and be waiting for io.  The io can't complete
7912 	 * because the interrupt thread is blocked trying to grab
7913 	 * the as lock.
7914 	 */
7915 
7916 	ASSERT(hat->sfmmu_xhat_provider == NULL);
7917 
7918 	if (hat == ksfmmup) {
7919 		if (IS_KMEM_VA_LARGEPAGE(addr)) {
7920 			ASSERT(segkmem_lpszc > 0);
7921 			pfn = sfmmu_kvaszc2pfn(addr, segkmem_lpszc);
7922 			if (pfn != PFN_INVALID) {
7923 				sfmmu_check_kpfn(pfn);
7924 				return (pfn);
7925 			}
7926 		} else if (segkpm && IS_KPM_ADDR(addr)) {
7927 			return (sfmmu_kpm_vatopfn(addr));
7928 		}
7929 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
7930 		    == PFN_SUSPENDED) {
7931 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
7932 		}
7933 		sfmmu_check_kpfn(pfn);
7934 		return (pfn);
7935 	} else {
7936 		return (sfmmu_uvatopfn(addr, hat, NULL));
7937 	}
7938 }
7939 
7940 /*
7941  * hat_getkpfnum() is an obsolete DDI routine, and its use is discouraged.
7942  * Use hat_getpfnum(kas.a_hat, ...) instead.
7943  *
7944  * We'd like to return PFN_INVALID if the mappings have underlying page_t's
7945  * but can't right now due to the fact that some software has grown to use
7946  * this interface incorrectly. So for now when the interface is misused,
7947  * return a warning to the user that in the future it won't work in the
7948  * way they're abusing it, and carry on (after disabling page relocation).
7949  */
7950 pfn_t
7951 hat_getkpfnum(caddr_t addr)
7952 {
7953 	pfn_t pfn;
7954 	tte_t tte;
7955 	int badcaller = 0;
7956 	extern int segkmem_reloc;
7957 
7958 	if (segkpm && IS_KPM_ADDR(addr)) {
7959 		badcaller = 1;
7960 		pfn = sfmmu_kpm_vatopfn(addr);
7961 	} else {
7962 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
7963 		    == PFN_SUSPENDED) {
7964 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
7965 		}
7966 		badcaller = pf_is_memory(pfn);
7967 	}
7968 
7969 	if (badcaller) {
7970 		/*
7971 		 * We can't return PFN_INVALID or the caller may panic
7972 		 * or corrupt the system.  The only alternative is to
7973 		 * disable page relocation at this point for all kernel
7974 		 * memory.  This will impact any callers of page_relocate()
7975 		 * such as FMA or DR.
7976 		 *
7977 		 * RFE: Add junk here to spit out an ereport so the sysadmin
7978 		 * can be advised that he should upgrade his device driver
7979 		 * so that this doesn't happen.
7980 		 */
7981 		hat_getkpfnum_badcall(caller());
7982 		if (hat_kpr_enabled && segkmem_reloc) {
7983 			hat_kpr_enabled = 0;
7984 			segkmem_reloc = 0;
7985 			cmn_err(CE_WARN, "Kernel Page Relocation is DISABLED");
7986 		}
7987 	}
7988 	return (pfn);
7989 }
7990 
7991 /*
7992  * This routine will return both pfn and tte for the vaddr.
7993  */
7994 static pfn_t
7995 sfmmu_uvatopfn(caddr_t vaddr, struct hat *sfmmup, tte_t *ttep)
7996 {
7997 	struct hmehash_bucket *hmebp;
7998 	hmeblk_tag hblktag;
7999 	int hmeshift, hashno = 1;
8000 	struct hme_blk *hmeblkp = NULL;
8001 	tte_t tte;
8002 
8003 	struct sf_hment *sfhmep;
8004 	pfn_t pfn;
8005 
8006 	/* support for ISM */
8007 	ism_map_t	*ism_map;
8008 	ism_blk_t	*ism_blkp;
8009 	int		i;
8010 	sfmmu_t *ism_hatid = NULL;
8011 	sfmmu_t *locked_hatid = NULL;
8012 	sfmmu_t	*sv_sfmmup = sfmmup;
8013 	caddr_t	sv_vaddr = vaddr;
8014 	sf_srd_t *srdp;
8015 
8016 	if (ttep == NULL) {
8017 		ttep = &tte;
8018 	} else {
8019 		ttep->ll = 0;
8020 	}
8021 
8022 	ASSERT(sfmmup != ksfmmup);
8023 	SFMMU_STAT(sf_user_vtop);
8024 	/*
8025 	 * Set ism_hatid if vaddr falls in a ISM segment.
8026 	 */
8027 	ism_blkp = sfmmup->sfmmu_iblk;
8028 	if (ism_blkp != NULL) {
8029 		sfmmu_ismhat_enter(sfmmup, 0);
8030 		locked_hatid = sfmmup;
8031 	}
8032 	while (ism_blkp != NULL && ism_hatid == NULL) {
8033 		ism_map = ism_blkp->iblk_maps;
8034 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
8035 			if (vaddr >= ism_start(ism_map[i]) &&
8036 			    vaddr < ism_end(ism_map[i])) {
8037 				sfmmup = ism_hatid = ism_map[i].imap_ismhat;
8038 				vaddr = (caddr_t)(vaddr -
8039 				    ism_start(ism_map[i]));
8040 				break;
8041 			}
8042 		}
8043 		ism_blkp = ism_blkp->iblk_next;
8044 	}
8045 	if (locked_hatid) {
8046 		sfmmu_ismhat_exit(locked_hatid, 0);
8047 	}
8048 
8049 	hblktag.htag_id = sfmmup;
8050 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
8051 	do {
8052 		hmeshift = HME_HASH_SHIFT(hashno);
8053 		hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
8054 		hblktag.htag_rehash = hashno;
8055 		hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
8056 
8057 		SFMMU_HASH_LOCK(hmebp);
8058 
8059 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
8060 		if (hmeblkp != NULL) {
8061 			ASSERT(!hmeblkp->hblk_shared);
8062 			HBLKTOHME(sfhmep, hmeblkp, vaddr);
8063 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
8064 			SFMMU_HASH_UNLOCK(hmebp);
8065 			if (TTE_IS_VALID(ttep)) {
8066 				pfn = TTE_TO_PFN(vaddr, ttep);
8067 				return (pfn);
8068 			}
8069 			break;
8070 		}
8071 		SFMMU_HASH_UNLOCK(hmebp);
8072 		hashno++;
8073 	} while (HME_REHASH(sfmmup) && (hashno <= mmu_hashcnt));
8074 
8075 	if (SF_HMERGNMAP_ISNULL(sv_sfmmup)) {
8076 		return (PFN_INVALID);
8077 	}
8078 	srdp = sv_sfmmup->sfmmu_srdp;
8079 	ASSERT(srdp != NULL);
8080 	ASSERT(srdp->srd_refcnt != 0);
8081 	hblktag.htag_id = srdp;
8082 	hashno = 1;
8083 	do {
8084 		hmeshift = HME_HASH_SHIFT(hashno);
8085 		hblktag.htag_bspage = HME_HASH_BSPAGE(sv_vaddr, hmeshift);
8086 		hblktag.htag_rehash = hashno;
8087 		hmebp = HME_HASH_FUNCTION(srdp, sv_vaddr, hmeshift);
8088 
8089 		SFMMU_HASH_LOCK(hmebp);
8090 		for (hmeblkp = hmebp->hmeblkp; hmeblkp != NULL;
8091 		    hmeblkp = hmeblkp->hblk_next) {
8092 			uint_t rid;
8093 			sf_region_t *rgnp;
8094 			caddr_t rsaddr;
8095 			caddr_t readdr;
8096 
8097 			if (!HTAGS_EQ_SHME(hmeblkp->hblk_tag, hblktag,
8098 			    sv_sfmmup->sfmmu_hmeregion_map)) {
8099 				continue;
8100 			}
8101 			ASSERT(hmeblkp->hblk_shared);
8102 			rid = hmeblkp->hblk_tag.htag_rid;
8103 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8104 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8105 			rgnp = srdp->srd_hmergnp[rid];
8106 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
8107 			HBLKTOHME(sfhmep, hmeblkp, sv_vaddr);
8108 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
8109 			rsaddr = rgnp->rgn_saddr;
8110 			readdr = rsaddr + rgnp->rgn_size;
8111 #ifdef DEBUG
8112 			if (TTE_IS_VALID(ttep) ||
8113 			    get_hblk_ttesz(hmeblkp) > TTE8K) {
8114 				caddr_t eva = tte_to_evaddr(hmeblkp, ttep);
8115 				ASSERT(eva > sv_vaddr);
8116 				ASSERT(sv_vaddr >= rsaddr);
8117 				ASSERT(sv_vaddr < readdr);
8118 				ASSERT(eva <= readdr);
8119 			}
8120 #endif /* DEBUG */
8121 			/*
8122 			 * Continue the search if we
8123 			 * found an invalid 8K tte outside of the area
8124 			 * covered by this hmeblk's region.
8125 			 */
8126 			if (TTE_IS_VALID(ttep)) {
8127 				SFMMU_HASH_UNLOCK(hmebp);
8128 				pfn = TTE_TO_PFN(sv_vaddr, ttep);
8129 				return (pfn);
8130 			} else if (get_hblk_ttesz(hmeblkp) > TTE8K ||
8131 			    (sv_vaddr >= rsaddr && sv_vaddr < readdr)) {
8132 				SFMMU_HASH_UNLOCK(hmebp);
8133 				pfn = PFN_INVALID;
8134 				return (pfn);
8135 			}
8136 		}
8137 		SFMMU_HASH_UNLOCK(hmebp);
8138 		hashno++;
8139 	} while (hashno <= mmu_hashcnt);
8140 	return (PFN_INVALID);
8141 }
8142 
8143 
8144 /*
8145  * For compatability with AT&T and later optimizations
8146  */
8147 /* ARGSUSED */
8148 void
8149 hat_map(struct hat *hat, caddr_t addr, size_t len, uint_t flags)
8150 {
8151 	ASSERT(hat != NULL);
8152 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8153 }
8154 
8155 /*
8156  * Return the number of mappings to a particular page.  This number is an
8157  * approximation of the number of people sharing the page.
8158  *
8159  * shared hmeblks or ism hmeblks are counted as 1 mapping here.
8160  * hat_page_checkshare() can be used to compare threshold to share
8161  * count that reflects the number of region sharers albeit at higher cost.
8162  */
8163 ulong_t
8164 hat_page_getshare(page_t *pp)
8165 {
8166 	page_t *spp = pp;	/* start page */
8167 	kmutex_t *pml;
8168 	ulong_t	cnt;
8169 	int index, sz = TTE64K;
8170 
8171 	/*
8172 	 * We need to grab the mlist lock to make sure any outstanding
8173 	 * load/unloads complete.  Otherwise we could return zero
8174 	 * even though the unload(s) hasn't finished yet.
8175 	 */
8176 	pml = sfmmu_mlist_enter(spp);
8177 	cnt = spp->p_share;
8178 
8179 #ifdef VAC
8180 	if (kpm_enable)
8181 		cnt += spp->p_kpmref;
8182 #endif
8183 
8184 	/*
8185 	 * If we have any large mappings, we count the number of
8186 	 * mappings that this large page is part of.
8187 	 */
8188 	index = PP_MAPINDEX(spp);
8189 	index >>= 1;
8190 	while (index) {
8191 		pp = PP_GROUPLEADER(spp, sz);
8192 		if ((index & 0x1) && pp != spp) {
8193 			cnt += pp->p_share;
8194 			spp = pp;
8195 		}
8196 		index >>= 1;
8197 		sz++;
8198 	}
8199 	sfmmu_mlist_exit(pml);
8200 	return (cnt);
8201 }
8202 
8203 /*
8204  * Return 1 if the number of mappings exceeds sh_thresh. Return 0
8205  * otherwise. Count shared hmeblks by region's refcnt.
8206  */
8207 int
8208 hat_page_checkshare(page_t *pp, ulong_t sh_thresh)
8209 {
8210 	kmutex_t *pml;
8211 	ulong_t	cnt = 0;
8212 	int index, sz = TTE8K;
8213 	struct sf_hment *sfhme, *tmphme = NULL;
8214 	struct hme_blk *hmeblkp;
8215 
8216 	pml = sfmmu_mlist_enter(pp);
8217 
8218 	if (kpm_enable)
8219 		cnt = pp->p_kpmref;
8220 
8221 	if (pp->p_share + cnt > sh_thresh) {
8222 		sfmmu_mlist_exit(pml);
8223 		return (1);
8224 	}
8225 
8226 	index = PP_MAPINDEX(pp);
8227 
8228 again:
8229 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
8230 		tmphme = sfhme->hme_next;
8231 		if (IS_PAHME(sfhme)) {
8232 			continue;
8233 		}
8234 
8235 		hmeblkp = sfmmu_hmetohblk(sfhme);
8236 		if (hmeblkp->hblk_xhat_bit) {
8237 			cnt++;
8238 			if (cnt > sh_thresh) {
8239 				sfmmu_mlist_exit(pml);
8240 				return (1);
8241 			}
8242 			continue;
8243 		}
8244 		if (hme_size(sfhme) != sz) {
8245 			continue;
8246 		}
8247 
8248 		if (hmeblkp->hblk_shared) {
8249 			sf_srd_t *srdp = hblktosrd(hmeblkp);
8250 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
8251 			sf_region_t *rgnp;
8252 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8253 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8254 			ASSERT(srdp != NULL);
8255 			rgnp = srdp->srd_hmergnp[rid];
8256 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
8257 			    rgnp, rid);
8258 			cnt += rgnp->rgn_refcnt;
8259 		} else {
8260 			cnt++;
8261 		}
8262 		if (cnt > sh_thresh) {
8263 			sfmmu_mlist_exit(pml);
8264 			return (1);
8265 		}
8266 	}
8267 
8268 	index >>= 1;
8269 	sz++;
8270 	while (index) {
8271 		pp = PP_GROUPLEADER(pp, sz);
8272 		ASSERT(sfmmu_mlist_held(pp));
8273 		if (index & 0x1) {
8274 			goto again;
8275 		}
8276 		index >>= 1;
8277 		sz++;
8278 	}
8279 	sfmmu_mlist_exit(pml);
8280 	return (0);
8281 }
8282 
8283 /*
8284  * Unload all large mappings to the pp and reset the p_szc field of every
8285  * constituent page according to the remaining mappings.
8286  *
8287  * pp must be locked SE_EXCL. Even though no other constituent pages are
8288  * locked it's legal to unload the large mappings to the pp because all
8289  * constituent pages of large locked mappings have to be locked SE_SHARED.
8290  * This means if we have SE_EXCL lock on one of constituent pages none of the
8291  * large mappings to pp are locked.
8292  *
8293  * Decrease p_szc field starting from the last constituent page and ending
8294  * with the root page. This method is used because other threads rely on the
8295  * root's p_szc to find the lock to syncronize on. After a root page_t's p_szc
8296  * is demoted then other threads will succeed in sfmmu_mlspl_enter(). This
8297  * ensures that p_szc changes of the constituent pages appears atomic for all
8298  * threads that use sfmmu_mlspl_enter() to examine p_szc field.
8299  *
8300  * This mechanism is only used for file system pages where it's not always
8301  * possible to get SE_EXCL locks on all constituent pages to demote the size
8302  * code (as is done for anonymous or kernel large pages).
8303  *
8304  * See more comments in front of sfmmu_mlspl_enter().
8305  */
8306 void
8307 hat_page_demote(page_t *pp)
8308 {
8309 	int index;
8310 	int sz;
8311 	cpuset_t cpuset;
8312 	int sync = 0;
8313 	page_t *rootpp;
8314 	struct sf_hment *sfhme;
8315 	struct sf_hment *tmphme = NULL;
8316 	struct hme_blk *hmeblkp;
8317 	uint_t pszc;
8318 	page_t *lastpp;
8319 	cpuset_t tset;
8320 	pgcnt_t npgs;
8321 	kmutex_t *pml;
8322 	kmutex_t *pmtx = NULL;
8323 
8324 	ASSERT(PAGE_EXCL(pp));
8325 	ASSERT(!PP_ISFREE(pp));
8326 	ASSERT(!PP_ISKAS(pp));
8327 	ASSERT(page_szc_lock_assert(pp));
8328 	pml = sfmmu_mlist_enter(pp);
8329 
8330 	pszc = pp->p_szc;
8331 	if (pszc == 0) {
8332 		goto out;
8333 	}
8334 
8335 	index = PP_MAPINDEX(pp) >> 1;
8336 
8337 	if (index) {
8338 		CPUSET_ZERO(cpuset);
8339 		sz = TTE64K;
8340 		sync = 1;
8341 	}
8342 
8343 	while (index) {
8344 		if (!(index & 0x1)) {
8345 			index >>= 1;
8346 			sz++;
8347 			continue;
8348 		}
8349 		ASSERT(sz <= pszc);
8350 		rootpp = PP_GROUPLEADER(pp, sz);
8351 		for (sfhme = rootpp->p_mapping; sfhme; sfhme = tmphme) {
8352 			tmphme = sfhme->hme_next;
8353 			ASSERT(!IS_PAHME(sfhme));
8354 			hmeblkp = sfmmu_hmetohblk(sfhme);
8355 			if (hme_size(sfhme) != sz) {
8356 				continue;
8357 			}
8358 			if (hmeblkp->hblk_xhat_bit) {
8359 				cmn_err(CE_PANIC,
8360 				    "hat_page_demote: xhat hmeblk");
8361 			}
8362 			tset = sfmmu_pageunload(rootpp, sfhme, sz);
8363 			CPUSET_OR(cpuset, tset);
8364 		}
8365 		if (index >>= 1) {
8366 			sz++;
8367 		}
8368 	}
8369 
8370 	ASSERT(!PP_ISMAPPED_LARGE(pp));
8371 
8372 	if (sync) {
8373 		xt_sync(cpuset);
8374 #ifdef VAC
8375 		if (PP_ISTNC(pp)) {
8376 			conv_tnc(rootpp, sz);
8377 		}
8378 #endif	/* VAC */
8379 	}
8380 
8381 	pmtx = sfmmu_page_enter(pp);
8382 
8383 	ASSERT(pp->p_szc == pszc);
8384 	rootpp = PP_PAGEROOT(pp);
8385 	ASSERT(rootpp->p_szc == pszc);
8386 	lastpp = PP_PAGENEXT_N(rootpp, TTEPAGES(pszc) - 1);
8387 
8388 	while (lastpp != rootpp) {
8389 		sz = PP_MAPINDEX(lastpp) ? fnd_mapping_sz(lastpp) : 0;
8390 		ASSERT(sz < pszc);
8391 		npgs = (sz == 0) ? 1 : TTEPAGES(sz);
8392 		ASSERT(P2PHASE(lastpp->p_pagenum, npgs) == npgs - 1);
8393 		while (--npgs > 0) {
8394 			lastpp->p_szc = (uchar_t)sz;
8395 			lastpp = PP_PAGEPREV(lastpp);
8396 		}
8397 		if (sz) {
8398 			/*
8399 			 * make sure before current root's pszc
8400 			 * is updated all updates to constituent pages pszc
8401 			 * fields are globally visible.
8402 			 */
8403 			membar_producer();
8404 		}
8405 		lastpp->p_szc = sz;
8406 		ASSERT(IS_P2ALIGNED(lastpp->p_pagenum, TTEPAGES(sz)));
8407 		if (lastpp != rootpp) {
8408 			lastpp = PP_PAGEPREV(lastpp);
8409 		}
8410 	}
8411 	if (sz == 0) {
8412 		/* the loop above doesn't cover this case */
8413 		rootpp->p_szc = 0;
8414 	}
8415 out:
8416 	ASSERT(pp->p_szc == 0);
8417 	if (pmtx != NULL) {
8418 		sfmmu_page_exit(pmtx);
8419 	}
8420 	sfmmu_mlist_exit(pml);
8421 }
8422 
8423 /*
8424  * Refresh the HAT ismttecnt[] element for size szc.
8425  * Caller must have set ISM busy flag to prevent mapping
8426  * lists from changing while we're traversing them.
8427  */
8428 pgcnt_t
8429 ism_tsb_entries(sfmmu_t *sfmmup, int szc)
8430 {
8431 	ism_blk_t	*ism_blkp = sfmmup->sfmmu_iblk;
8432 	ism_map_t	*ism_map;
8433 	pgcnt_t		npgs = 0;
8434 	pgcnt_t		npgs_scd = 0;
8435 	int		j;
8436 	sf_scd_t	*scdp;
8437 	uchar_t		rid;
8438 
8439 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
8440 	scdp = sfmmup->sfmmu_scdp;
8441 
8442 	for (; ism_blkp != NULL; ism_blkp = ism_blkp->iblk_next) {
8443 		ism_map = ism_blkp->iblk_maps;
8444 		for (j = 0; ism_map[j].imap_ismhat && j < ISM_MAP_SLOTS; j++) {
8445 			rid = ism_map[j].imap_rid;
8446 			ASSERT(rid == SFMMU_INVALID_ISMRID ||
8447 			    rid < sfmmup->sfmmu_srdp->srd_next_ismrid);
8448 
8449 			if (scdp != NULL && rid != SFMMU_INVALID_ISMRID &&
8450 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
8451 				/* ISM is in sfmmup's SCD */
8452 				npgs_scd +=
8453 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8454 			} else {
8455 				/* ISMs is not in SCD */
8456 				npgs +=
8457 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8458 			}
8459 		}
8460 	}
8461 	sfmmup->sfmmu_ismttecnt[szc] = npgs;
8462 	sfmmup->sfmmu_scdismttecnt[szc] = npgs_scd;
8463 	return (npgs);
8464 }
8465 
8466 /*
8467  * Yield the memory claim requirement for an address space.
8468  *
8469  * This is currently implemented as the number of bytes that have active
8470  * hardware translations that have page structures.  Therefore, it can
8471  * underestimate the traditional resident set size, eg, if the
8472  * physical page is present and the hardware translation is missing;
8473  * and it can overestimate the rss, eg, if there are active
8474  * translations to a frame buffer with page structs.
8475  * Also, it does not take sharing into account.
8476  *
8477  * Note that we don't acquire locks here since this function is most often
8478  * called from the clock thread.
8479  */
8480 size_t
8481 hat_get_mapped_size(struct hat *hat)
8482 {
8483 	size_t		assize = 0;
8484 	int 		i;
8485 
8486 	if (hat == NULL)
8487 		return (0);
8488 
8489 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8490 
8491 	for (i = 0; i < mmu_page_sizes; i++)
8492 		assize += ((pgcnt_t)hat->sfmmu_ttecnt[i] +
8493 		    (pgcnt_t)hat->sfmmu_scdrttecnt[i]) * TTEBYTES(i);
8494 
8495 	if (hat->sfmmu_iblk == NULL)
8496 		return (assize);
8497 
8498 	for (i = 0; i < mmu_page_sizes; i++)
8499 		assize += ((pgcnt_t)hat->sfmmu_ismttecnt[i] +
8500 		    (pgcnt_t)hat->sfmmu_scdismttecnt[i]) * TTEBYTES(i);
8501 
8502 	return (assize);
8503 }
8504 
8505 int
8506 hat_stats_enable(struct hat *hat)
8507 {
8508 	hatlock_t	*hatlockp;
8509 
8510 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8511 
8512 	hatlockp = sfmmu_hat_enter(hat);
8513 	hat->sfmmu_rmstat++;
8514 	sfmmu_hat_exit(hatlockp);
8515 	return (1);
8516 }
8517 
8518 void
8519 hat_stats_disable(struct hat *hat)
8520 {
8521 	hatlock_t	*hatlockp;
8522 
8523 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8524 
8525 	hatlockp = sfmmu_hat_enter(hat);
8526 	hat->sfmmu_rmstat--;
8527 	sfmmu_hat_exit(hatlockp);
8528 }
8529 
8530 /*
8531  * Routines for entering or removing  ourselves from the
8532  * ism_hat's mapping list. This is used for both private and
8533  * SCD hats.
8534  */
8535 static void
8536 iment_add(struct ism_ment *iment,  struct hat *ism_hat)
8537 {
8538 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8539 
8540 	iment->iment_prev = NULL;
8541 	iment->iment_next = ism_hat->sfmmu_iment;
8542 	if (ism_hat->sfmmu_iment) {
8543 		ism_hat->sfmmu_iment->iment_prev = iment;
8544 	}
8545 	ism_hat->sfmmu_iment = iment;
8546 }
8547 
8548 static void
8549 iment_sub(struct ism_ment *iment, struct hat *ism_hat)
8550 {
8551 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8552 
8553 	if (ism_hat->sfmmu_iment == NULL) {
8554 		panic("ism map entry remove - no entries");
8555 	}
8556 
8557 	if (iment->iment_prev) {
8558 		ASSERT(ism_hat->sfmmu_iment != iment);
8559 		iment->iment_prev->iment_next = iment->iment_next;
8560 	} else {
8561 		ASSERT(ism_hat->sfmmu_iment == iment);
8562 		ism_hat->sfmmu_iment = iment->iment_next;
8563 	}
8564 
8565 	if (iment->iment_next) {
8566 		iment->iment_next->iment_prev = iment->iment_prev;
8567 	}
8568 
8569 	/*
8570 	 * zero out the entry
8571 	 */
8572 	iment->iment_next = NULL;
8573 	iment->iment_prev = NULL;
8574 	iment->iment_hat =  NULL;
8575 }
8576 
8577 /*
8578  * Hat_share()/unshare() return an (non-zero) error
8579  * when saddr and daddr are not properly aligned.
8580  *
8581  * The top level mapping element determines the alignment
8582  * requirement for saddr and daddr, depending on different
8583  * architectures.
8584  *
8585  * When hat_share()/unshare() are not supported,
8586  * HATOP_SHARE()/UNSHARE() return 0
8587  */
8588 int
8589 hat_share(struct hat *sfmmup, caddr_t addr,
8590 	struct hat *ism_hatid, caddr_t sptaddr, size_t len, uint_t ismszc)
8591 {
8592 	ism_blk_t	*ism_blkp;
8593 	ism_blk_t	*new_iblk;
8594 	ism_map_t 	*ism_map;
8595 	ism_ment_t	*ism_ment;
8596 	int		i, added;
8597 	hatlock_t	*hatlockp;
8598 	int		reload_mmu = 0;
8599 	uint_t		ismshift = page_get_shift(ismszc);
8600 	size_t		ismpgsz = page_get_pagesize(ismszc);
8601 	uint_t		ismmask = (uint_t)ismpgsz - 1;
8602 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8603 	ushort_t	ismhatflag;
8604 	hat_region_cookie_t rcookie;
8605 	sf_scd_t	*old_scdp;
8606 
8607 #ifdef DEBUG
8608 	caddr_t		eaddr = addr + len;
8609 #endif /* DEBUG */
8610 
8611 	ASSERT(ism_hatid != NULL && sfmmup != NULL);
8612 	ASSERT(sptaddr == ISMID_STARTADDR);
8613 	/*
8614 	 * Check the alignment.
8615 	 */
8616 	if (!ISM_ALIGNED(ismshift, addr) || !ISM_ALIGNED(ismshift, sptaddr))
8617 		return (EINVAL);
8618 
8619 	/*
8620 	 * Check size alignment.
8621 	 */
8622 	if (!ISM_ALIGNED(ismshift, len))
8623 		return (EINVAL);
8624 
8625 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
8626 
8627 	/*
8628 	 * Allocate ism_ment for the ism_hat's mapping list, and an
8629 	 * ism map blk in case we need one.  We must do our
8630 	 * allocations before acquiring locks to prevent a deadlock
8631 	 * in the kmem allocator on the mapping list lock.
8632 	 */
8633 	new_iblk = kmem_cache_alloc(ism_blk_cache, KM_SLEEP);
8634 	ism_ment = kmem_cache_alloc(ism_ment_cache, KM_SLEEP);
8635 
8636 	/*
8637 	 * Serialize ISM mappings with the ISM busy flag, and also the
8638 	 * trap handlers.
8639 	 */
8640 	sfmmu_ismhat_enter(sfmmup, 0);
8641 
8642 	/*
8643 	 * Allocate an ism map blk if necessary.
8644 	 */
8645 	if (sfmmup->sfmmu_iblk == NULL) {
8646 		sfmmup->sfmmu_iblk = new_iblk;
8647 		bzero(new_iblk, sizeof (*new_iblk));
8648 		new_iblk->iblk_nextpa = (uint64_t)-1;
8649 		membar_stst();	/* make sure next ptr visible to all CPUs */
8650 		sfmmup->sfmmu_ismblkpa = va_to_pa((caddr_t)new_iblk);
8651 		reload_mmu = 1;
8652 		new_iblk = NULL;
8653 	}
8654 
8655 #ifdef DEBUG
8656 	/*
8657 	 * Make sure mapping does not already exist.
8658 	 */
8659 	ism_blkp = sfmmup->sfmmu_iblk;
8660 	while (ism_blkp != NULL) {
8661 		ism_map = ism_blkp->iblk_maps;
8662 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
8663 			if ((addr >= ism_start(ism_map[i]) &&
8664 			    addr < ism_end(ism_map[i])) ||
8665 			    eaddr > ism_start(ism_map[i]) &&
8666 			    eaddr <= ism_end(ism_map[i])) {
8667 				panic("sfmmu_share: Already mapped!");
8668 			}
8669 		}
8670 		ism_blkp = ism_blkp->iblk_next;
8671 	}
8672 #endif /* DEBUG */
8673 
8674 	ASSERT(ismszc >= TTE4M);
8675 	if (ismszc == TTE4M) {
8676 		ismhatflag = HAT_4M_FLAG;
8677 	} else if (ismszc == TTE32M) {
8678 		ismhatflag = HAT_32M_FLAG;
8679 	} else if (ismszc == TTE256M) {
8680 		ismhatflag = HAT_256M_FLAG;
8681 	}
8682 	/*
8683 	 * Add mapping to first available mapping slot.
8684 	 */
8685 	ism_blkp = sfmmup->sfmmu_iblk;
8686 	added = 0;
8687 	while (!added) {
8688 		ism_map = ism_blkp->iblk_maps;
8689 		for (i = 0; i < ISM_MAP_SLOTS; i++)  {
8690 			if (ism_map[i].imap_ismhat == NULL) {
8691 
8692 				ism_map[i].imap_ismhat = ism_hatid;
8693 				ism_map[i].imap_vb_shift = (uchar_t)ismshift;
8694 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8695 				ism_map[i].imap_hatflags = ismhatflag;
8696 				ism_map[i].imap_sz_mask = ismmask;
8697 				/*
8698 				 * imap_seg is checked in ISM_CHECK to see if
8699 				 * non-NULL, then other info assumed valid.
8700 				 */
8701 				membar_stst();
8702 				ism_map[i].imap_seg = (uintptr_t)addr | sh_size;
8703 				ism_map[i].imap_ment = ism_ment;
8704 
8705 				/*
8706 				 * Now add ourselves to the ism_hat's
8707 				 * mapping list.
8708 				 */
8709 				ism_ment->iment_hat = sfmmup;
8710 				ism_ment->iment_base_va = addr;
8711 				ism_hatid->sfmmu_ismhat = 1;
8712 				mutex_enter(&ism_mlist_lock);
8713 				iment_add(ism_ment, ism_hatid);
8714 				mutex_exit(&ism_mlist_lock);
8715 				added = 1;
8716 				break;
8717 			}
8718 		}
8719 		if (!added && ism_blkp->iblk_next == NULL) {
8720 			ism_blkp->iblk_next = new_iblk;
8721 			new_iblk = NULL;
8722 			bzero(ism_blkp->iblk_next,
8723 			    sizeof (*ism_blkp->iblk_next));
8724 			ism_blkp->iblk_next->iblk_nextpa = (uint64_t)-1;
8725 			membar_stst();
8726 			ism_blkp->iblk_nextpa =
8727 			    va_to_pa((caddr_t)ism_blkp->iblk_next);
8728 		}
8729 		ism_blkp = ism_blkp->iblk_next;
8730 	}
8731 
8732 	/*
8733 	 * After calling hat_join_region, sfmmup may join a new SCD or
8734 	 * move from the old scd to a new scd, in which case, we want to
8735 	 * shrink the sfmmup's private tsb size, i.e., pass shrink to
8736 	 * sfmmu_check_page_sizes at the end of this routine.
8737 	 */
8738 	old_scdp = sfmmup->sfmmu_scdp;
8739 
8740 	rcookie = hat_join_region(sfmmup, addr, len, (void *)ism_hatid, 0,
8741 	    PROT_ALL, ismszc, NULL, HAT_REGION_ISM);
8742 	if (rcookie != HAT_INVALID_REGION_COOKIE) {
8743 		ism_map[i].imap_rid = (uchar_t)((uint64_t)rcookie);
8744 	}
8745 	/*
8746 	 * Update our counters for this sfmmup's ism mappings.
8747 	 */
8748 	for (i = 0; i <= ismszc; i++) {
8749 		if (!(disable_ism_large_pages & (1 << i)))
8750 			(void) ism_tsb_entries(sfmmup, i);
8751 	}
8752 
8753 	/*
8754 	 * For ISM and DISM we do not support 512K pages, so we only only
8755 	 * search the 4M and 8K/64K hashes for 4 pagesize cpus, and search the
8756 	 * 256M or 32M, and 4M and 8K/64K hashes for 6 pagesize cpus.
8757 	 *
8758 	 * Need to set 32M/256M ISM flags to make sure
8759 	 * sfmmu_check_page_sizes() enables them on Panther.
8760 	 */
8761 	ASSERT((disable_ism_large_pages & (1 << TTE512K)) != 0);
8762 
8763 	switch (ismszc) {
8764 	case TTE256M:
8765 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_256M_ISM)) {
8766 			hatlockp = sfmmu_hat_enter(sfmmup);
8767 			SFMMU_FLAGS_SET(sfmmup, HAT_256M_ISM);
8768 			sfmmu_hat_exit(hatlockp);
8769 		}
8770 		break;
8771 	case TTE32M:
8772 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_32M_ISM)) {
8773 			hatlockp = sfmmu_hat_enter(sfmmup);
8774 			SFMMU_FLAGS_SET(sfmmup, HAT_32M_ISM);
8775 			sfmmu_hat_exit(hatlockp);
8776 		}
8777 		break;
8778 	default:
8779 		break;
8780 	}
8781 
8782 	/*
8783 	 * If we updated the ismblkpa for this HAT we must make
8784 	 * sure all CPUs running this process reload their tsbmiss area.
8785 	 * Otherwise they will fail to load the mappings in the tsbmiss
8786 	 * handler and will loop calling pagefault().
8787 	 */
8788 	if (reload_mmu) {
8789 		hatlockp = sfmmu_hat_enter(sfmmup);
8790 		sfmmu_sync_mmustate(sfmmup);
8791 		sfmmu_hat_exit(hatlockp);
8792 	}
8793 
8794 	sfmmu_ismhat_exit(sfmmup, 0);
8795 
8796 	/*
8797 	 * Free up ismblk if we didn't use it.
8798 	 */
8799 	if (new_iblk != NULL)
8800 		kmem_cache_free(ism_blk_cache, new_iblk);
8801 
8802 	/*
8803 	 * Check TSB and TLB page sizes.
8804 	 */
8805 	if (sfmmup->sfmmu_scdp != NULL && old_scdp != sfmmup->sfmmu_scdp) {
8806 		sfmmu_check_page_sizes(sfmmup, 0);
8807 	} else {
8808 		sfmmu_check_page_sizes(sfmmup, 1);
8809 	}
8810 	return (0);
8811 }
8812 
8813 /*
8814  * hat_unshare removes exactly one ism_map from
8815  * this process's as.  It expects multiple calls
8816  * to hat_unshare for multiple shm segments.
8817  */
8818 void
8819 hat_unshare(struct hat *sfmmup, caddr_t addr, size_t len, uint_t ismszc)
8820 {
8821 	ism_map_t 	*ism_map;
8822 	ism_ment_t	*free_ment = NULL;
8823 	ism_blk_t	*ism_blkp;
8824 	struct hat	*ism_hatid;
8825 	int 		found, i;
8826 	hatlock_t	*hatlockp;
8827 	struct tsb_info	*tsbinfo;
8828 	uint_t		ismshift = page_get_shift(ismszc);
8829 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8830 	uchar_t		ism_rid;
8831 	sf_scd_t	*old_scdp;
8832 
8833 	ASSERT(ISM_ALIGNED(ismshift, addr));
8834 	ASSERT(ISM_ALIGNED(ismshift, len));
8835 	ASSERT(sfmmup != NULL);
8836 	ASSERT(sfmmup != ksfmmup);
8837 
8838 	if (sfmmup->sfmmu_xhat_provider) {
8839 		XHAT_UNSHARE(sfmmup, addr, len);
8840 		return;
8841 	} else {
8842 		/*
8843 		 * This must be a CPU HAT. If the address space has
8844 		 * XHATs attached, inform all XHATs that ISM segment
8845 		 * is going away
8846 		 */
8847 		ASSERT(sfmmup->sfmmu_as != NULL);
8848 		if (sfmmup->sfmmu_as->a_xhat != NULL)
8849 			xhat_unshare_all(sfmmup->sfmmu_as, addr, len);
8850 	}
8851 
8852 	/*
8853 	 * Make sure that during the entire time ISM mappings are removed,
8854 	 * the trap handlers serialize behind us, and that no one else
8855 	 * can be mucking with ISM mappings.  This also lets us get away
8856 	 * with not doing expensive cross calls to flush the TLB -- we
8857 	 * just discard the context, flush the entire TSB, and call it
8858 	 * a day.
8859 	 */
8860 	sfmmu_ismhat_enter(sfmmup, 0);
8861 
8862 	/*
8863 	 * Remove the mapping.
8864 	 *
8865 	 * We can't have any holes in the ism map.
8866 	 * The tsb miss code while searching the ism map will
8867 	 * stop on an empty map slot.  So we must move
8868 	 * everyone past the hole up 1 if any.
8869 	 *
8870 	 * Also empty ism map blks are not freed until the
8871 	 * process exits. This is to prevent a MT race condition
8872 	 * between sfmmu_unshare() and sfmmu_tsbmiss_exception().
8873 	 */
8874 	found = 0;
8875 	ism_blkp = sfmmup->sfmmu_iblk;
8876 	while (!found && ism_blkp != NULL) {
8877 		ism_map = ism_blkp->iblk_maps;
8878 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
8879 			if (addr == ism_start(ism_map[i]) &&
8880 			    sh_size == (size_t)(ism_size(ism_map[i]))) {
8881 				found = 1;
8882 				break;
8883 			}
8884 		}
8885 		if (!found)
8886 			ism_blkp = ism_blkp->iblk_next;
8887 	}
8888 
8889 	if (found) {
8890 		ism_hatid = ism_map[i].imap_ismhat;
8891 		ism_rid = ism_map[i].imap_rid;
8892 		ASSERT(ism_hatid != NULL);
8893 		ASSERT(ism_hatid->sfmmu_ismhat == 1);
8894 
8895 		/*
8896 		 * After hat_leave_region, the sfmmup may leave SCD,
8897 		 * in which case, we want to grow the private tsb size when
8898 		 * calling sfmmu_check_page_sizes at the end of the routine.
8899 		 */
8900 		old_scdp = sfmmup->sfmmu_scdp;
8901 		/*
8902 		 * Then remove ourselves from the region.
8903 		 */
8904 		if (ism_rid != SFMMU_INVALID_ISMRID) {
8905 			hat_leave_region(sfmmup, (void *)((uint64_t)ism_rid),
8906 			    HAT_REGION_ISM);
8907 		}
8908 
8909 		/*
8910 		 * And now guarantee that any other cpu
8911 		 * that tries to process an ISM miss
8912 		 * will go to tl=0.
8913 		 */
8914 		hatlockp = sfmmu_hat_enter(sfmmup);
8915 		sfmmu_invalidate_ctx(sfmmup);
8916 		sfmmu_hat_exit(hatlockp);
8917 
8918 		/*
8919 		 * Remove ourselves from the ism mapping list.
8920 		 */
8921 		mutex_enter(&ism_mlist_lock);
8922 		iment_sub(ism_map[i].imap_ment, ism_hatid);
8923 		mutex_exit(&ism_mlist_lock);
8924 		free_ment = ism_map[i].imap_ment;
8925 
8926 		/*
8927 		 * We delete the ism map by copying
8928 		 * the next map over the current one.
8929 		 * We will take the next one in the maps
8930 		 * array or from the next ism_blk.
8931 		 */
8932 		while (ism_blkp != NULL) {
8933 			ism_map = ism_blkp->iblk_maps;
8934 			while (i < (ISM_MAP_SLOTS - 1)) {
8935 				ism_map[i] = ism_map[i + 1];
8936 				i++;
8937 			}
8938 			/* i == (ISM_MAP_SLOTS - 1) */
8939 			ism_blkp = ism_blkp->iblk_next;
8940 			if (ism_blkp != NULL) {
8941 				ism_map[i] = ism_blkp->iblk_maps[0];
8942 				i = 0;
8943 			} else {
8944 				ism_map[i].imap_seg = 0;
8945 				ism_map[i].imap_vb_shift = 0;
8946 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8947 				ism_map[i].imap_hatflags = 0;
8948 				ism_map[i].imap_sz_mask = 0;
8949 				ism_map[i].imap_ismhat = NULL;
8950 				ism_map[i].imap_ment = NULL;
8951 			}
8952 		}
8953 
8954 		/*
8955 		 * Now flush entire TSB for the process, since
8956 		 * demapping page by page can be too expensive.
8957 		 * We don't have to flush the TLB here anymore
8958 		 * since we switch to a new TLB ctx instead.
8959 		 * Also, there is no need to flush if the process
8960 		 * is exiting since the TSB will be freed later.
8961 		 */
8962 		if (!sfmmup->sfmmu_free) {
8963 			hatlockp = sfmmu_hat_enter(sfmmup);
8964 			for (tsbinfo = sfmmup->sfmmu_tsb; tsbinfo != NULL;
8965 			    tsbinfo = tsbinfo->tsb_next) {
8966 				if (tsbinfo->tsb_flags & TSB_SWAPPED)
8967 					continue;
8968 				if (tsbinfo->tsb_flags & TSB_RELOC_FLAG) {
8969 					tsbinfo->tsb_flags |=
8970 					    TSB_FLUSH_NEEDED;
8971 					continue;
8972 				}
8973 
8974 				sfmmu_inv_tsb(tsbinfo->tsb_va,
8975 				    TSB_BYTES(tsbinfo->tsb_szc));
8976 			}
8977 			sfmmu_hat_exit(hatlockp);
8978 		}
8979 	}
8980 
8981 	/*
8982 	 * Update our counters for this sfmmup's ism mappings.
8983 	 */
8984 	for (i = 0; i <= ismszc; i++) {
8985 		if (!(disable_ism_large_pages & (1 << i)))
8986 			(void) ism_tsb_entries(sfmmup, i);
8987 	}
8988 
8989 	sfmmu_ismhat_exit(sfmmup, 0);
8990 
8991 	/*
8992 	 * We must do our freeing here after dropping locks
8993 	 * to prevent a deadlock in the kmem allocator on the
8994 	 * mapping list lock.
8995 	 */
8996 	if (free_ment != NULL)
8997 		kmem_cache_free(ism_ment_cache, free_ment);
8998 
8999 	/*
9000 	 * Check TSB and TLB page sizes if the process isn't exiting.
9001 	 */
9002 	if (!sfmmup->sfmmu_free) {
9003 		if (found && old_scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
9004 			sfmmu_check_page_sizes(sfmmup, 1);
9005 		} else {
9006 			sfmmu_check_page_sizes(sfmmup, 0);
9007 		}
9008 	}
9009 }
9010 
9011 /* ARGSUSED */
9012 static int
9013 sfmmu_idcache_constructor(void *buf, void *cdrarg, int kmflags)
9014 {
9015 	/* void *buf is sfmmu_t pointer */
9016 	bzero(buf, sizeof (sfmmu_t));
9017 
9018 	return (0);
9019 }
9020 
9021 /* ARGSUSED */
9022 static void
9023 sfmmu_idcache_destructor(void *buf, void *cdrarg)
9024 {
9025 	/* void *buf is sfmmu_t pointer */
9026 }
9027 
9028 /*
9029  * setup kmem hmeblks by bzeroing all members and initializing the nextpa
9030  * field to be the pa of this hmeblk
9031  */
9032 /* ARGSUSED */
9033 static int
9034 sfmmu_hblkcache_constructor(void *buf, void *cdrarg, int kmflags)
9035 {
9036 	struct hme_blk *hmeblkp;
9037 
9038 	bzero(buf, (size_t)cdrarg);
9039 	hmeblkp = (struct hme_blk *)buf;
9040 	hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
9041 
9042 #ifdef	HBLK_TRACE
9043 	mutex_init(&hmeblkp->hblk_audit_lock, NULL, MUTEX_DEFAULT, NULL);
9044 #endif	/* HBLK_TRACE */
9045 
9046 	return (0);
9047 }
9048 
9049 /* ARGSUSED */
9050 static void
9051 sfmmu_hblkcache_destructor(void *buf, void *cdrarg)
9052 {
9053 
9054 #ifdef	HBLK_TRACE
9055 
9056 	struct hme_blk *hmeblkp;
9057 
9058 	hmeblkp = (struct hme_blk *)buf;
9059 	mutex_destroy(&hmeblkp->hblk_audit_lock);
9060 
9061 #endif	/* HBLK_TRACE */
9062 }
9063 
9064 #define	SFMMU_CACHE_RECLAIM_SCAN_RATIO 8
9065 static int sfmmu_cache_reclaim_scan_ratio = SFMMU_CACHE_RECLAIM_SCAN_RATIO;
9066 /*
9067  * The kmem allocator will callback into our reclaim routine when the system
9068  * is running low in memory.  We traverse the hash and free up all unused but
9069  * still cached hme_blks.  We also traverse the free list and free them up
9070  * as well.
9071  */
9072 /*ARGSUSED*/
9073 static void
9074 sfmmu_hblkcache_reclaim(void *cdrarg)
9075 {
9076 	int i;
9077 	struct hmehash_bucket *hmebp;
9078 	struct hme_blk *hmeblkp, *nx_hblk, *pr_hblk = NULL;
9079 	static struct hmehash_bucket *uhmehash_reclaim_hand;
9080 	static struct hmehash_bucket *khmehash_reclaim_hand;
9081 	struct hme_blk *list = NULL, *last_hmeblkp;
9082 	cpuset_t cpuset = cpu_ready_set;
9083 	cpu_hme_pend_t *cpuhp;
9084 
9085 	/* Free up hmeblks on the cpu pending lists */
9086 	for (i = 0; i < NCPU; i++) {
9087 		cpuhp = &cpu_hme_pend[i];
9088 		if (cpuhp->chp_listp != NULL)  {
9089 			mutex_enter(&cpuhp->chp_mutex);
9090 			if (cpuhp->chp_listp == NULL) {
9091 				mutex_exit(&cpuhp->chp_mutex);
9092 				continue;
9093 			}
9094 			for (last_hmeblkp = cpuhp->chp_listp;
9095 			    last_hmeblkp->hblk_next != NULL;
9096 			    last_hmeblkp = last_hmeblkp->hblk_next)
9097 				;
9098 			last_hmeblkp->hblk_next = list;
9099 			list = cpuhp->chp_listp;
9100 			cpuhp->chp_listp = NULL;
9101 			cpuhp->chp_count = 0;
9102 			mutex_exit(&cpuhp->chp_mutex);
9103 		}
9104 
9105 	}
9106 
9107 	if (list != NULL) {
9108 		kpreempt_disable();
9109 		CPUSET_DEL(cpuset, CPU->cpu_id);
9110 		xt_sync(cpuset);
9111 		xt_sync(cpuset);
9112 		kpreempt_enable();
9113 		sfmmu_hblk_free(&list);
9114 		list = NULL;
9115 	}
9116 
9117 	hmebp = uhmehash_reclaim_hand;
9118 	if (hmebp == NULL || hmebp > &uhme_hash[UHMEHASH_SZ])
9119 		uhmehash_reclaim_hand = hmebp = uhme_hash;
9120 	uhmehash_reclaim_hand += UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
9121 
9122 	for (i = UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
9123 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
9124 			hmeblkp = hmebp->hmeblkp;
9125 			pr_hblk = NULL;
9126 			while (hmeblkp) {
9127 				nx_hblk = hmeblkp->hblk_next;
9128 				if (!hmeblkp->hblk_vcnt &&
9129 				    !hmeblkp->hblk_hmecnt) {
9130 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9131 					    pr_hblk, &list, 0);
9132 				} else {
9133 					pr_hblk = hmeblkp;
9134 				}
9135 				hmeblkp = nx_hblk;
9136 			}
9137 			SFMMU_HASH_UNLOCK(hmebp);
9138 		}
9139 		if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
9140 			hmebp = uhme_hash;
9141 	}
9142 
9143 	hmebp = khmehash_reclaim_hand;
9144 	if (hmebp == NULL || hmebp > &khme_hash[KHMEHASH_SZ])
9145 		khmehash_reclaim_hand = hmebp = khme_hash;
9146 	khmehash_reclaim_hand += KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
9147 
9148 	for (i = KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
9149 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
9150 			hmeblkp = hmebp->hmeblkp;
9151 			pr_hblk = NULL;
9152 			while (hmeblkp) {
9153 				nx_hblk = hmeblkp->hblk_next;
9154 				if (!hmeblkp->hblk_vcnt &&
9155 				    !hmeblkp->hblk_hmecnt) {
9156 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9157 					    pr_hblk, &list, 0);
9158 				} else {
9159 					pr_hblk = hmeblkp;
9160 				}
9161 				hmeblkp = nx_hblk;
9162 			}
9163 			SFMMU_HASH_UNLOCK(hmebp);
9164 		}
9165 		if (hmebp++ == &khme_hash[KHMEHASH_SZ])
9166 			hmebp = khme_hash;
9167 	}
9168 	sfmmu_hblks_list_purge(&list, 0);
9169 }
9170 
9171 /*
9172  * sfmmu_get_ppvcolor should become a vm_machdep or hatop interface.
9173  * same goes for sfmmu_get_addrvcolor().
9174  *
9175  * This function will return the virtual color for the specified page. The
9176  * virtual color corresponds to this page current mapping or its last mapping.
9177  * It is used by memory allocators to choose addresses with the correct
9178  * alignment so vac consistency is automatically maintained.  If the page
9179  * has no color it returns -1.
9180  */
9181 /*ARGSUSED*/
9182 int
9183 sfmmu_get_ppvcolor(struct page *pp)
9184 {
9185 #ifdef VAC
9186 	int color;
9187 
9188 	if (!(cache & CACHE_VAC) || PP_NEWPAGE(pp)) {
9189 		return (-1);
9190 	}
9191 	color = PP_GET_VCOLOR(pp);
9192 	ASSERT(color < mmu_btop(shm_alignment));
9193 	return (color);
9194 #else
9195 	return (-1);
9196 #endif	/* VAC */
9197 }
9198 
9199 /*
9200  * This function will return the desired alignment for vac consistency
9201  * (vac color) given a virtual address.  If no vac is present it returns -1.
9202  */
9203 /*ARGSUSED*/
9204 int
9205 sfmmu_get_addrvcolor(caddr_t vaddr)
9206 {
9207 #ifdef VAC
9208 	if (cache & CACHE_VAC) {
9209 		return (addr_to_vcolor(vaddr));
9210 	} else {
9211 		return (-1);
9212 	}
9213 #else
9214 	return (-1);
9215 #endif	/* VAC */
9216 }
9217 
9218 #ifdef VAC
9219 /*
9220  * Check for conflicts.
9221  * A conflict exists if the new and existent mappings do not match in
9222  * their "shm_alignment fields. If conflicts exist, the existant mappings
9223  * are flushed unless one of them is locked. If one of them is locked, then
9224  * the mappings are flushed and converted to non-cacheable mappings.
9225  */
9226 static void
9227 sfmmu_vac_conflict(struct hat *hat, caddr_t addr, page_t *pp)
9228 {
9229 	struct hat *tmphat;
9230 	struct sf_hment *sfhmep, *tmphme = NULL;
9231 	struct hme_blk *hmeblkp;
9232 	int vcolor;
9233 	tte_t tte;
9234 
9235 	ASSERT(sfmmu_mlist_held(pp));
9236 	ASSERT(!PP_ISNC(pp));		/* page better be cacheable */
9237 
9238 	vcolor = addr_to_vcolor(addr);
9239 	if (PP_NEWPAGE(pp)) {
9240 		PP_SET_VCOLOR(pp, vcolor);
9241 		return;
9242 	}
9243 
9244 	if (PP_GET_VCOLOR(pp) == vcolor) {
9245 		return;
9246 	}
9247 
9248 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
9249 		/*
9250 		 * Previous user of page had a different color
9251 		 * but since there are no current users
9252 		 * we just flush the cache and change the color.
9253 		 */
9254 		SFMMU_STAT(sf_pgcolor_conflict);
9255 		sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9256 		PP_SET_VCOLOR(pp, vcolor);
9257 		return;
9258 	}
9259 
9260 	/*
9261 	 * If we get here we have a vac conflict with a current
9262 	 * mapping.  VAC conflict policy is as follows.
9263 	 * - The default is to unload the other mappings unless:
9264 	 * - If we have a large mapping we uncache the page.
9265 	 * We need to uncache the rest of the large page too.
9266 	 * - If any of the mappings are locked we uncache the page.
9267 	 * - If the requested mapping is inconsistent
9268 	 * with another mapping and that mapping
9269 	 * is in the same address space we have to
9270 	 * make it non-cached.  The default thing
9271 	 * to do is unload the inconsistent mapping
9272 	 * but if they are in the same address space
9273 	 * we run the risk of unmapping the pc or the
9274 	 * stack which we will use as we return to the user,
9275 	 * in which case we can then fault on the thing
9276 	 * we just unloaded and get into an infinite loop.
9277 	 */
9278 	if (PP_ISMAPPED_LARGE(pp)) {
9279 		int sz;
9280 
9281 		/*
9282 		 * Existing mapping is for big pages. We don't unload
9283 		 * existing big mappings to satisfy new mappings.
9284 		 * Always convert all mappings to TNC.
9285 		 */
9286 		sz = fnd_mapping_sz(pp);
9287 		pp = PP_GROUPLEADER(pp, sz);
9288 		SFMMU_STAT_ADD(sf_uncache_conflict, TTEPAGES(sz));
9289 		sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH,
9290 		    TTEPAGES(sz));
9291 
9292 		return;
9293 	}
9294 
9295 	/*
9296 	 * check if any mapping is in same as or if it is locked
9297 	 * since in that case we need to uncache.
9298 	 */
9299 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9300 		tmphme = sfhmep->hme_next;
9301 		if (IS_PAHME(sfhmep))
9302 			continue;
9303 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9304 		if (hmeblkp->hblk_xhat_bit)
9305 			continue;
9306 		tmphat = hblktosfmmu(hmeblkp);
9307 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
9308 		ASSERT(TTE_IS_VALID(&tte));
9309 		if (hmeblkp->hblk_shared || tmphat == hat ||
9310 		    hmeblkp->hblk_lckcnt) {
9311 			/*
9312 			 * We have an uncache conflict
9313 			 */
9314 			SFMMU_STAT(sf_uncache_conflict);
9315 			sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH, 1);
9316 			return;
9317 		}
9318 	}
9319 
9320 	/*
9321 	 * We have an unload conflict
9322 	 * We have already checked for LARGE mappings, therefore
9323 	 * the remaining mapping(s) must be TTE8K.
9324 	 */
9325 	SFMMU_STAT(sf_unload_conflict);
9326 
9327 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9328 		tmphme = sfhmep->hme_next;
9329 		if (IS_PAHME(sfhmep))
9330 			continue;
9331 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9332 		if (hmeblkp->hblk_xhat_bit)
9333 			continue;
9334 		ASSERT(!hmeblkp->hblk_shared);
9335 		(void) sfmmu_pageunload(pp, sfhmep, TTE8K);
9336 	}
9337 
9338 	if (PP_ISMAPPED_KPM(pp))
9339 		sfmmu_kpm_vac_unload(pp, addr);
9340 
9341 	/*
9342 	 * Unloads only do TLB flushes so we need to flush the
9343 	 * cache here.
9344 	 */
9345 	sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9346 	PP_SET_VCOLOR(pp, vcolor);
9347 }
9348 
9349 /*
9350  * Whenever a mapping is unloaded and the page is in TNC state,
9351  * we see if the page can be made cacheable again. 'pp' is
9352  * the page that we just unloaded a mapping from, the size
9353  * of mapping that was unloaded is 'ottesz'.
9354  * Remark:
9355  * The recache policy for mpss pages can leave a performance problem
9356  * under the following circumstances:
9357  * . A large page in uncached mode has just been unmapped.
9358  * . All constituent pages are TNC due to a conflicting small mapping.
9359  * . There are many other, non conflicting, small mappings around for
9360  *   a lot of the constituent pages.
9361  * . We're called w/ the "old" groupleader page and the old ottesz,
9362  *   but this is irrelevant, since we're no more "PP_ISMAPPED_LARGE", so
9363  *   we end up w/ TTE8K or npages == 1.
9364  * . We call tst_tnc w/ the old groupleader only, and if there is no
9365  *   conflict, we re-cache only this page.
9366  * . All other small mappings are not checked and will be left in TNC mode.
9367  * The problem is not very serious because:
9368  * . mpss is actually only defined for heap and stack, so the probability
9369  *   is not very high that a large page mapping exists in parallel to a small
9370  *   one (this is possible, but seems to be bad programming style in the
9371  *   appl).
9372  * . The problem gets a little bit more serious, when those TNC pages
9373  *   have to be mapped into kernel space, e.g. for networking.
9374  * . When VAC alias conflicts occur in applications, this is regarded
9375  *   as an application bug. So if kstat's show them, the appl should
9376  *   be changed anyway.
9377  */
9378 void
9379 conv_tnc(page_t *pp, int ottesz)
9380 {
9381 	int cursz, dosz;
9382 	pgcnt_t curnpgs, dopgs;
9383 	pgcnt_t pg64k;
9384 	page_t *pp2;
9385 
9386 	/*
9387 	 * Determine how big a range we check for TNC and find
9388 	 * leader page. cursz is the size of the biggest
9389 	 * mapping that still exist on 'pp'.
9390 	 */
9391 	if (PP_ISMAPPED_LARGE(pp)) {
9392 		cursz = fnd_mapping_sz(pp);
9393 	} else {
9394 		cursz = TTE8K;
9395 	}
9396 
9397 	if (ottesz >= cursz) {
9398 		dosz = ottesz;
9399 		pp2 = pp;
9400 	} else {
9401 		dosz = cursz;
9402 		pp2 = PP_GROUPLEADER(pp, dosz);
9403 	}
9404 
9405 	pg64k = TTEPAGES(TTE64K);
9406 	dopgs = TTEPAGES(dosz);
9407 
9408 	ASSERT(dopgs == 1 || ((dopgs & (pg64k - 1)) == 0));
9409 
9410 	while (dopgs != 0) {
9411 		curnpgs = TTEPAGES(cursz);
9412 		if (tst_tnc(pp2, curnpgs)) {
9413 			SFMMU_STAT_ADD(sf_recache, curnpgs);
9414 			sfmmu_page_cache_array(pp2, HAT_CACHE, CACHE_NO_FLUSH,
9415 			    curnpgs);
9416 		}
9417 
9418 		ASSERT(dopgs >= curnpgs);
9419 		dopgs -= curnpgs;
9420 
9421 		if (dopgs == 0) {
9422 			break;
9423 		}
9424 
9425 		pp2 = PP_PAGENEXT_N(pp2, curnpgs);
9426 		if (((dopgs & (pg64k - 1)) == 0) && PP_ISMAPPED_LARGE(pp2)) {
9427 			cursz = fnd_mapping_sz(pp2);
9428 		} else {
9429 			cursz = TTE8K;
9430 		}
9431 	}
9432 }
9433 
9434 /*
9435  * Returns 1 if page(s) can be converted from TNC to cacheable setting,
9436  * returns 0 otherwise. Note that oaddr argument is valid for only
9437  * 8k pages.
9438  */
9439 int
9440 tst_tnc(page_t *pp, pgcnt_t npages)
9441 {
9442 	struct	sf_hment *sfhme;
9443 	struct	hme_blk *hmeblkp;
9444 	tte_t	tte;
9445 	caddr_t	vaddr;
9446 	int	clr_valid = 0;
9447 	int 	color, color1, bcolor;
9448 	int	i, ncolors;
9449 
9450 	ASSERT(pp != NULL);
9451 	ASSERT(!(cache & CACHE_WRITEBACK));
9452 
9453 	if (npages > 1) {
9454 		ncolors = CACHE_NUM_COLOR;
9455 	}
9456 
9457 	for (i = 0; i < npages; i++) {
9458 		ASSERT(sfmmu_mlist_held(pp));
9459 		ASSERT(PP_ISTNC(pp));
9460 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
9461 
9462 		if (PP_ISPNC(pp)) {
9463 			return (0);
9464 		}
9465 
9466 		clr_valid = 0;
9467 		if (PP_ISMAPPED_KPM(pp)) {
9468 			caddr_t kpmvaddr;
9469 
9470 			ASSERT(kpm_enable);
9471 			kpmvaddr = hat_kpm_page2va(pp, 1);
9472 			ASSERT(!(npages > 1 && IS_KPM_ALIAS_RANGE(kpmvaddr)));
9473 			color1 = addr_to_vcolor(kpmvaddr);
9474 			clr_valid = 1;
9475 		}
9476 
9477 		for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9478 			if (IS_PAHME(sfhme))
9479 				continue;
9480 			hmeblkp = sfmmu_hmetohblk(sfhme);
9481 			if (hmeblkp->hblk_xhat_bit)
9482 				continue;
9483 
9484 			sfmmu_copytte(&sfhme->hme_tte, &tte);
9485 			ASSERT(TTE_IS_VALID(&tte));
9486 
9487 			vaddr = tte_to_vaddr(hmeblkp, tte);
9488 			color = addr_to_vcolor(vaddr);
9489 
9490 			if (npages > 1) {
9491 				/*
9492 				 * If there is a big mapping, make sure
9493 				 * 8K mapping is consistent with the big
9494 				 * mapping.
9495 				 */
9496 				bcolor = i % ncolors;
9497 				if (color != bcolor) {
9498 					return (0);
9499 				}
9500 			}
9501 			if (!clr_valid) {
9502 				clr_valid = 1;
9503 				color1 = color;
9504 			}
9505 
9506 			if (color1 != color) {
9507 				return (0);
9508 			}
9509 		}
9510 
9511 		pp = PP_PAGENEXT(pp);
9512 	}
9513 
9514 	return (1);
9515 }
9516 
9517 void
9518 sfmmu_page_cache_array(page_t *pp, int flags, int cache_flush_flag,
9519 	pgcnt_t npages)
9520 {
9521 	kmutex_t *pmtx;
9522 	int i, ncolors, bcolor;
9523 	kpm_hlk_t *kpmp;
9524 	cpuset_t cpuset;
9525 
9526 	ASSERT(pp != NULL);
9527 	ASSERT(!(cache & CACHE_WRITEBACK));
9528 
9529 	kpmp = sfmmu_kpm_kpmp_enter(pp, npages);
9530 	pmtx = sfmmu_page_enter(pp);
9531 
9532 	/*
9533 	 * Fast path caching single unmapped page
9534 	 */
9535 	if (npages == 1 && !PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp) &&
9536 	    flags == HAT_CACHE) {
9537 		PP_CLRTNC(pp);
9538 		PP_CLRPNC(pp);
9539 		sfmmu_page_exit(pmtx);
9540 		sfmmu_kpm_kpmp_exit(kpmp);
9541 		return;
9542 	}
9543 
9544 	/*
9545 	 * We need to capture all cpus in order to change cacheability
9546 	 * because we can't allow one cpu to access the same physical
9547 	 * page using a cacheable and a non-cachebale mapping at the same
9548 	 * time. Since we may end up walking the ism mapping list
9549 	 * have to grab it's lock now since we can't after all the
9550 	 * cpus have been captured.
9551 	 */
9552 	sfmmu_hat_lock_all();
9553 	mutex_enter(&ism_mlist_lock);
9554 	kpreempt_disable();
9555 	cpuset = cpu_ready_set;
9556 	xc_attention(cpuset);
9557 
9558 	if (npages > 1) {
9559 		/*
9560 		 * Make sure all colors are flushed since the
9561 		 * sfmmu_page_cache() only flushes one color-
9562 		 * it does not know big pages.
9563 		 */
9564 		ncolors = CACHE_NUM_COLOR;
9565 		if (flags & HAT_TMPNC) {
9566 			for (i = 0; i < ncolors; i++) {
9567 				sfmmu_cache_flushcolor(i, pp->p_pagenum);
9568 			}
9569 			cache_flush_flag = CACHE_NO_FLUSH;
9570 		}
9571 	}
9572 
9573 	for (i = 0; i < npages; i++) {
9574 
9575 		ASSERT(sfmmu_mlist_held(pp));
9576 
9577 		if (!(flags == HAT_TMPNC && PP_ISTNC(pp))) {
9578 
9579 			if (npages > 1) {
9580 				bcolor = i % ncolors;
9581 			} else {
9582 				bcolor = NO_VCOLOR;
9583 			}
9584 
9585 			sfmmu_page_cache(pp, flags, cache_flush_flag,
9586 			    bcolor);
9587 		}
9588 
9589 		pp = PP_PAGENEXT(pp);
9590 	}
9591 
9592 	xt_sync(cpuset);
9593 	xc_dismissed(cpuset);
9594 	mutex_exit(&ism_mlist_lock);
9595 	sfmmu_hat_unlock_all();
9596 	sfmmu_page_exit(pmtx);
9597 	sfmmu_kpm_kpmp_exit(kpmp);
9598 	kpreempt_enable();
9599 }
9600 
9601 /*
9602  * This function changes the virtual cacheability of all mappings to a
9603  * particular page.  When changing from uncache to cacheable the mappings will
9604  * only be changed if all of them have the same virtual color.
9605  * We need to flush the cache in all cpus.  It is possible that
9606  * a process referenced a page as cacheable but has sinced exited
9607  * and cleared the mapping list.  We still to flush it but have no
9608  * state so all cpus is the only alternative.
9609  */
9610 static void
9611 sfmmu_page_cache(page_t *pp, int flags, int cache_flush_flag, int bcolor)
9612 {
9613 	struct	sf_hment *sfhme;
9614 	struct	hme_blk *hmeblkp;
9615 	sfmmu_t *sfmmup;
9616 	tte_t	tte, ttemod;
9617 	caddr_t	vaddr;
9618 	int	ret, color;
9619 	pfn_t	pfn;
9620 
9621 	color = bcolor;
9622 	pfn = pp->p_pagenum;
9623 
9624 	for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9625 
9626 		if (IS_PAHME(sfhme))
9627 			continue;
9628 		hmeblkp = sfmmu_hmetohblk(sfhme);
9629 
9630 		if (hmeblkp->hblk_xhat_bit)
9631 			continue;
9632 
9633 		sfmmu_copytte(&sfhme->hme_tte, &tte);
9634 		ASSERT(TTE_IS_VALID(&tte));
9635 		vaddr = tte_to_vaddr(hmeblkp, tte);
9636 		color = addr_to_vcolor(vaddr);
9637 
9638 #ifdef DEBUG
9639 		if ((flags & HAT_CACHE) && bcolor != NO_VCOLOR) {
9640 			ASSERT(color == bcolor);
9641 		}
9642 #endif
9643 
9644 		ASSERT(flags != HAT_TMPNC || color == PP_GET_VCOLOR(pp));
9645 
9646 		ttemod = tte;
9647 		if (flags & (HAT_UNCACHE | HAT_TMPNC)) {
9648 			TTE_CLR_VCACHEABLE(&ttemod);
9649 		} else {	/* flags & HAT_CACHE */
9650 			TTE_SET_VCACHEABLE(&ttemod);
9651 		}
9652 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
9653 		if (ret < 0) {
9654 			/*
9655 			 * Since all cpus are captured modifytte should not
9656 			 * fail.
9657 			 */
9658 			panic("sfmmu_page_cache: write to tte failed");
9659 		}
9660 
9661 		sfmmup = hblktosfmmu(hmeblkp);
9662 		if (cache_flush_flag == CACHE_FLUSH) {
9663 			/*
9664 			 * Flush TSBs, TLBs and caches
9665 			 */
9666 			if (hmeblkp->hblk_shared) {
9667 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9668 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9669 				sf_region_t *rgnp;
9670 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9671 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9672 				ASSERT(srdp != NULL);
9673 				rgnp = srdp->srd_hmergnp[rid];
9674 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9675 				    srdp, rgnp, rid);
9676 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9677 				    hmeblkp, 0);
9678 				sfmmu_cache_flush(pfn, addr_to_vcolor(vaddr));
9679 			} else if (sfmmup->sfmmu_ismhat) {
9680 				if (flags & HAT_CACHE) {
9681 					SFMMU_STAT(sf_ism_recache);
9682 				} else {
9683 					SFMMU_STAT(sf_ism_uncache);
9684 				}
9685 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9686 				    pfn, CACHE_FLUSH);
9687 			} else {
9688 				sfmmu_tlbcache_demap(vaddr, sfmmup, hmeblkp,
9689 				    pfn, 0, FLUSH_ALL_CPUS, CACHE_FLUSH, 1);
9690 			}
9691 
9692 			/*
9693 			 * all cache entries belonging to this pfn are
9694 			 * now flushed.
9695 			 */
9696 			cache_flush_flag = CACHE_NO_FLUSH;
9697 		} else {
9698 			/*
9699 			 * Flush only TSBs and TLBs.
9700 			 */
9701 			if (hmeblkp->hblk_shared) {
9702 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9703 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9704 				sf_region_t *rgnp;
9705 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9706 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9707 				ASSERT(srdp != NULL);
9708 				rgnp = srdp->srd_hmergnp[rid];
9709 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9710 				    srdp, rgnp, rid);
9711 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9712 				    hmeblkp, 0);
9713 			} else if (sfmmup->sfmmu_ismhat) {
9714 				if (flags & HAT_CACHE) {
9715 					SFMMU_STAT(sf_ism_recache);
9716 				} else {
9717 					SFMMU_STAT(sf_ism_uncache);
9718 				}
9719 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9720 				    pfn, CACHE_NO_FLUSH);
9721 			} else {
9722 				sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 1);
9723 			}
9724 		}
9725 	}
9726 
9727 	if (PP_ISMAPPED_KPM(pp))
9728 		sfmmu_kpm_page_cache(pp, flags, cache_flush_flag);
9729 
9730 	switch (flags) {
9731 
9732 		default:
9733 			panic("sfmmu_pagecache: unknown flags");
9734 			break;
9735 
9736 		case HAT_CACHE:
9737 			PP_CLRTNC(pp);
9738 			PP_CLRPNC(pp);
9739 			PP_SET_VCOLOR(pp, color);
9740 			break;
9741 
9742 		case HAT_TMPNC:
9743 			PP_SETTNC(pp);
9744 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9745 			break;
9746 
9747 		case HAT_UNCACHE:
9748 			PP_SETPNC(pp);
9749 			PP_CLRTNC(pp);
9750 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9751 			break;
9752 	}
9753 }
9754 #endif	/* VAC */
9755 
9756 
9757 /*
9758  * Wrapper routine used to return a context.
9759  *
9760  * It's the responsibility of the caller to guarantee that the
9761  * process serializes on calls here by taking the HAT lock for
9762  * the hat.
9763  *
9764  */
9765 static void
9766 sfmmu_get_ctx(sfmmu_t *sfmmup)
9767 {
9768 	mmu_ctx_t *mmu_ctxp;
9769 	uint_t pstate_save;
9770 	int ret;
9771 
9772 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9773 	ASSERT(sfmmup != ksfmmup);
9774 
9775 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID)) {
9776 		sfmmu_setup_tsbinfo(sfmmup);
9777 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ALLCTX_INVALID);
9778 	}
9779 
9780 	kpreempt_disable();
9781 
9782 	mmu_ctxp = CPU_MMU_CTXP(CPU);
9783 	ASSERT(mmu_ctxp);
9784 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
9785 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
9786 
9787 	/*
9788 	 * Do a wrap-around if cnum reaches the max # cnum supported by a MMU.
9789 	 */
9790 	if (mmu_ctxp->mmu_cnum == mmu_ctxp->mmu_nctxs)
9791 		sfmmu_ctx_wrap_around(mmu_ctxp);
9792 
9793 	/*
9794 	 * Let the MMU set up the page sizes to use for
9795 	 * this context in the TLB. Don't program 2nd dtlb for ism hat.
9796 	 */
9797 	if ((&mmu_set_ctx_page_sizes) && (sfmmup->sfmmu_ismhat == 0)) {
9798 		mmu_set_ctx_page_sizes(sfmmup);
9799 	}
9800 
9801 	/*
9802 	 * sfmmu_alloc_ctx and sfmmu_load_mmustate will be performed with
9803 	 * interrupts disabled to prevent race condition with wrap-around
9804 	 * ctx invalidatation. In sun4v, ctx invalidation also involves
9805 	 * a HV call to set the number of TSBs to 0. If interrupts are not
9806 	 * disabled until after sfmmu_load_mmustate is complete TSBs may
9807 	 * become assigned to INVALID_CONTEXT. This is not allowed.
9808 	 */
9809 	pstate_save = sfmmu_disable_intrs();
9810 
9811 	if (sfmmu_alloc_ctx(sfmmup, 1, CPU, SFMMU_PRIVATE) &&
9812 	    sfmmup->sfmmu_scdp != NULL) {
9813 		sf_scd_t *scdp = sfmmup->sfmmu_scdp;
9814 		sfmmu_t *scsfmmup = scdp->scd_sfmmup;
9815 		ret = sfmmu_alloc_ctx(scsfmmup, 1, CPU, SFMMU_SHARED);
9816 		/* debug purpose only */
9817 		ASSERT(!ret || scsfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
9818 		    != INVALID_CONTEXT);
9819 	}
9820 	sfmmu_load_mmustate(sfmmup);
9821 
9822 	sfmmu_enable_intrs(pstate_save);
9823 
9824 	kpreempt_enable();
9825 }
9826 
9827 /*
9828  * When all cnums are used up in a MMU, cnum will wrap around to the
9829  * next generation and start from 2.
9830  */
9831 static void
9832 sfmmu_ctx_wrap_around(mmu_ctx_t *mmu_ctxp)
9833 {
9834 
9835 	/* caller must have disabled the preemption */
9836 	ASSERT(curthread->t_preempt >= 1);
9837 	ASSERT(mmu_ctxp != NULL);
9838 
9839 	/* acquire Per-MMU (PM) spin lock */
9840 	mutex_enter(&mmu_ctxp->mmu_lock);
9841 
9842 	/* re-check to see if wrap-around is needed */
9843 	if (mmu_ctxp->mmu_cnum < mmu_ctxp->mmu_nctxs)
9844 		goto done;
9845 
9846 	SFMMU_MMU_STAT(mmu_wrap_around);
9847 
9848 	/* update gnum */
9849 	ASSERT(mmu_ctxp->mmu_gnum != 0);
9850 	mmu_ctxp->mmu_gnum++;
9851 	if (mmu_ctxp->mmu_gnum == 0 ||
9852 	    mmu_ctxp->mmu_gnum > MAX_SFMMU_GNUM_VAL) {
9853 		cmn_err(CE_PANIC, "mmu_gnum of mmu_ctx 0x%p is out of bound.",
9854 		    (void *)mmu_ctxp);
9855 	}
9856 
9857 	if (mmu_ctxp->mmu_ncpus > 1) {
9858 		cpuset_t cpuset;
9859 
9860 		membar_enter(); /* make sure updated gnum visible */
9861 
9862 		SFMMU_XCALL_STATS(NULL);
9863 
9864 		/* xcall to others on the same MMU to invalidate ctx */
9865 		cpuset = mmu_ctxp->mmu_cpuset;
9866 		ASSERT(CPU_IN_SET(cpuset, CPU->cpu_id));
9867 		CPUSET_DEL(cpuset, CPU->cpu_id);
9868 		CPUSET_AND(cpuset, cpu_ready_set);
9869 
9870 		/*
9871 		 * Pass in INVALID_CONTEXT as the first parameter to
9872 		 * sfmmu_raise_tsb_exception, which invalidates the context
9873 		 * of any process running on the CPUs in the MMU.
9874 		 */
9875 		xt_some(cpuset, sfmmu_raise_tsb_exception,
9876 		    INVALID_CONTEXT, INVALID_CONTEXT);
9877 		xt_sync(cpuset);
9878 
9879 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
9880 	}
9881 
9882 	if (sfmmu_getctx_sec() != INVALID_CONTEXT) {
9883 		sfmmu_setctx_sec(INVALID_CONTEXT);
9884 		sfmmu_clear_utsbinfo();
9885 	}
9886 
9887 	/*
9888 	 * No xcall is needed here. For sun4u systems all CPUs in context
9889 	 * domain share a single physical MMU therefore it's enough to flush
9890 	 * TLB on local CPU. On sun4v systems we use 1 global context
9891 	 * domain and flush all remote TLBs in sfmmu_raise_tsb_exception
9892 	 * handler. Note that vtag_flushall_uctxs() is called
9893 	 * for Ultra II machine, where the equivalent flushall functionality
9894 	 * is implemented in SW, and only user ctx TLB entries are flushed.
9895 	 */
9896 	if (&vtag_flushall_uctxs != NULL) {
9897 		vtag_flushall_uctxs();
9898 	} else {
9899 		vtag_flushall();
9900 	}
9901 
9902 	/* reset mmu cnum, skips cnum 0 and 1 */
9903 	mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
9904 
9905 done:
9906 	mutex_exit(&mmu_ctxp->mmu_lock);
9907 }
9908 
9909 
9910 /*
9911  * For multi-threaded process, set the process context to INVALID_CONTEXT
9912  * so that it faults and reloads the MMU state from TL=0. For single-threaded
9913  * process, we can just load the MMU state directly without having to
9914  * set context invalid. Caller must hold the hat lock since we don't
9915  * acquire it here.
9916  */
9917 static void
9918 sfmmu_sync_mmustate(sfmmu_t *sfmmup)
9919 {
9920 	uint_t cnum;
9921 	uint_t pstate_save;
9922 
9923 	ASSERT(sfmmup != ksfmmup);
9924 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9925 
9926 	kpreempt_disable();
9927 
9928 	/*
9929 	 * We check whether the pass'ed-in sfmmup is the same as the
9930 	 * current running proc. This is to makes sure the current proc
9931 	 * stays single-threaded if it already is.
9932 	 */
9933 	if ((sfmmup == curthread->t_procp->p_as->a_hat) &&
9934 	    (curthread->t_procp->p_lwpcnt == 1)) {
9935 		/* single-thread */
9936 		cnum = sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum;
9937 		if (cnum != INVALID_CONTEXT) {
9938 			uint_t curcnum;
9939 			/*
9940 			 * Disable interrupts to prevent race condition
9941 			 * with sfmmu_ctx_wrap_around ctx invalidation.
9942 			 * In sun4v, ctx invalidation involves setting
9943 			 * TSB to NULL, hence, interrupts should be disabled
9944 			 * untill after sfmmu_load_mmustate is completed.
9945 			 */
9946 			pstate_save = sfmmu_disable_intrs();
9947 			curcnum = sfmmu_getctx_sec();
9948 			if (curcnum == cnum)
9949 				sfmmu_load_mmustate(sfmmup);
9950 			sfmmu_enable_intrs(pstate_save);
9951 			ASSERT(curcnum == cnum || curcnum == INVALID_CONTEXT);
9952 		}
9953 	} else {
9954 		/*
9955 		 * multi-thread
9956 		 * or when sfmmup is not the same as the curproc.
9957 		 */
9958 		sfmmu_invalidate_ctx(sfmmup);
9959 	}
9960 
9961 	kpreempt_enable();
9962 }
9963 
9964 
9965 /*
9966  * Replace the specified TSB with a new TSB.  This function gets called when
9967  * we grow, shrink or swapin a TSB.  When swapping in a TSB (TSB_SWAPIN), the
9968  * TSB_FORCEALLOC flag may be used to force allocation of a minimum-sized TSB
9969  * (8K).
9970  *
9971  * Caller must hold the HAT lock, but should assume any tsb_info
9972  * pointers it has are no longer valid after calling this function.
9973  *
9974  * Return values:
9975  *	TSB_ALLOCFAIL	Failed to allocate a TSB, due to memory constraints
9976  *	TSB_LOSTRACE	HAT is busy, i.e. another thread is already doing
9977  *			something to this tsbinfo/TSB
9978  *	TSB_SUCCESS	Operation succeeded
9979  */
9980 static tsb_replace_rc_t
9981 sfmmu_replace_tsb(sfmmu_t *sfmmup, struct tsb_info *old_tsbinfo, uint_t szc,
9982     hatlock_t *hatlockp, uint_t flags)
9983 {
9984 	struct tsb_info *new_tsbinfo = NULL;
9985 	struct tsb_info *curtsb, *prevtsb;
9986 	uint_t tte_sz_mask;
9987 	int i;
9988 
9989 	ASSERT(sfmmup != ksfmmup);
9990 	ASSERT(sfmmup->sfmmu_ismhat == 0);
9991 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9992 	ASSERT(szc <= tsb_max_growsize);
9993 
9994 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_BUSY))
9995 		return (TSB_LOSTRACE);
9996 
9997 	/*
9998 	 * Find the tsb_info ahead of this one in the list, and
9999 	 * also make sure that the tsb_info passed in really
10000 	 * exists!
10001 	 */
10002 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
10003 	    curtsb != old_tsbinfo && curtsb != NULL;
10004 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
10005 		;
10006 	ASSERT(curtsb != NULL);
10007 
10008 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
10009 		/*
10010 		 * The process is swapped out, so just set the new size
10011 		 * code.  When it swaps back in, we'll allocate a new one
10012 		 * of the new chosen size.
10013 		 */
10014 		curtsb->tsb_szc = szc;
10015 		return (TSB_SUCCESS);
10016 	}
10017 	SFMMU_FLAGS_SET(sfmmup, HAT_BUSY);
10018 
10019 	tte_sz_mask = old_tsbinfo->tsb_ttesz_mask;
10020 
10021 	/*
10022 	 * All initialization is done inside of sfmmu_tsbinfo_alloc().
10023 	 * If we fail to allocate a TSB, exit.
10024 	 *
10025 	 * If tsb grows with new tsb size > 4M and old tsb size < 4M,
10026 	 * then try 4M slab after the initial alloc fails.
10027 	 *
10028 	 * If tsb swapin with tsb size > 4M, then try 4M after the
10029 	 * initial alloc fails.
10030 	 */
10031 	sfmmu_hat_exit(hatlockp);
10032 	if (sfmmu_tsbinfo_alloc(&new_tsbinfo, szc,
10033 	    tte_sz_mask, flags, sfmmup) &&
10034 	    (!(flags & (TSB_GROW | TSB_SWAPIN)) || (szc <= TSB_4M_SZCODE) ||
10035 	    (!(flags & TSB_SWAPIN) &&
10036 	    (old_tsbinfo->tsb_szc >= TSB_4M_SZCODE)) ||
10037 	    sfmmu_tsbinfo_alloc(&new_tsbinfo, TSB_4M_SZCODE,
10038 	    tte_sz_mask, flags, sfmmup))) {
10039 		(void) sfmmu_hat_enter(sfmmup);
10040 		if (!(flags & TSB_SWAPIN))
10041 			SFMMU_STAT(sf_tsb_resize_failures);
10042 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10043 		return (TSB_ALLOCFAIL);
10044 	}
10045 	(void) sfmmu_hat_enter(sfmmup);
10046 
10047 	/*
10048 	 * Re-check to make sure somebody else didn't muck with us while we
10049 	 * didn't hold the HAT lock.  If the process swapped out, fine, just
10050 	 * exit; this can happen if we try to shrink the TSB from the context
10051 	 * of another process (such as on an ISM unmap), though it is rare.
10052 	 */
10053 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
10054 		SFMMU_STAT(sf_tsb_resize_failures);
10055 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10056 		sfmmu_hat_exit(hatlockp);
10057 		sfmmu_tsbinfo_free(new_tsbinfo);
10058 		(void) sfmmu_hat_enter(sfmmup);
10059 		return (TSB_LOSTRACE);
10060 	}
10061 
10062 #ifdef	DEBUG
10063 	/* Reverify that the tsb_info still exists.. for debugging only */
10064 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
10065 	    curtsb != old_tsbinfo && curtsb != NULL;
10066 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
10067 		;
10068 	ASSERT(curtsb != NULL);
10069 #endif	/* DEBUG */
10070 
10071 	/*
10072 	 * Quiesce any CPUs running this process on their next TLB miss
10073 	 * so they atomically see the new tsb_info.  We temporarily set the
10074 	 * context to invalid context so new threads that come on processor
10075 	 * after we do the xcall to cpusran will also serialize behind the
10076 	 * HAT lock on TLB miss and will see the new TSB.  Since this short
10077 	 * race with a new thread coming on processor is relatively rare,
10078 	 * this synchronization mechanism should be cheaper than always
10079 	 * pausing all CPUs for the duration of the setup, which is what
10080 	 * the old implementation did.  This is particuarly true if we are
10081 	 * copying a huge chunk of memory around during that window.
10082 	 *
10083 	 * The memory barriers are to make sure things stay consistent
10084 	 * with resume() since it does not hold the HAT lock while
10085 	 * walking the list of tsb_info structures.
10086 	 */
10087 	if ((flags & TSB_SWAPIN) != TSB_SWAPIN) {
10088 		/* The TSB is either growing or shrinking. */
10089 		sfmmu_invalidate_ctx(sfmmup);
10090 	} else {
10091 		/*
10092 		 * It is illegal to swap in TSBs from a process other
10093 		 * than a process being swapped in.  This in turn
10094 		 * implies we do not have a valid MMU context here
10095 		 * since a process needs one to resolve translation
10096 		 * misses.
10097 		 */
10098 		ASSERT(curthread->t_procp->p_as->a_hat == sfmmup);
10099 	}
10100 
10101 #ifdef DEBUG
10102 	ASSERT(max_mmu_ctxdoms > 0);
10103 
10104 	/*
10105 	 * Process should have INVALID_CONTEXT on all MMUs
10106 	 */
10107 	for (i = 0; i < max_mmu_ctxdoms; i++) {
10108 
10109 		ASSERT(sfmmup->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
10110 	}
10111 #endif
10112 
10113 	new_tsbinfo->tsb_next = old_tsbinfo->tsb_next;
10114 	membar_stst();	/* strict ordering required */
10115 	if (prevtsb)
10116 		prevtsb->tsb_next = new_tsbinfo;
10117 	else
10118 		sfmmup->sfmmu_tsb = new_tsbinfo;
10119 	membar_enter();	/* make sure new TSB globally visible */
10120 
10121 	/*
10122 	 * We need to migrate TSB entries from the old TSB to the new TSB
10123 	 * if tsb_remap_ttes is set and the TSB is growing.
10124 	 */
10125 	if (tsb_remap_ttes && ((flags & TSB_GROW) == TSB_GROW))
10126 		sfmmu_copy_tsb(old_tsbinfo, new_tsbinfo);
10127 
10128 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10129 
10130 	/*
10131 	 * Drop the HAT lock to free our old tsb_info.
10132 	 */
10133 	sfmmu_hat_exit(hatlockp);
10134 
10135 	if ((flags & TSB_GROW) == TSB_GROW) {
10136 		SFMMU_STAT(sf_tsb_grow);
10137 	} else if ((flags & TSB_SHRINK) == TSB_SHRINK) {
10138 		SFMMU_STAT(sf_tsb_shrink);
10139 	}
10140 
10141 	sfmmu_tsbinfo_free(old_tsbinfo);
10142 
10143 	(void) sfmmu_hat_enter(sfmmup);
10144 	return (TSB_SUCCESS);
10145 }
10146 
10147 /*
10148  * This function will re-program hat pgsz array, and invalidate the
10149  * process' context, forcing the process to switch to another
10150  * context on the next TLB miss, and therefore start using the
10151  * TLB that is reprogrammed for the new page sizes.
10152  */
10153 void
10154 sfmmu_reprog_pgsz_arr(sfmmu_t *sfmmup, uint8_t *tmp_pgsz)
10155 {
10156 	int i;
10157 	hatlock_t *hatlockp = NULL;
10158 
10159 	hatlockp = sfmmu_hat_enter(sfmmup);
10160 	/* USIII+-IV+ optimization, requires hat lock */
10161 	if (tmp_pgsz) {
10162 		for (i = 0; i < mmu_page_sizes; i++)
10163 			sfmmup->sfmmu_pgsz[i] = tmp_pgsz[i];
10164 	}
10165 	SFMMU_STAT(sf_tlb_reprog_pgsz);
10166 
10167 	sfmmu_invalidate_ctx(sfmmup);
10168 
10169 	sfmmu_hat_exit(hatlockp);
10170 }
10171 
10172 /*
10173  * The scd_rttecnt field in the SCD must be updated to take account of the
10174  * regions which it contains.
10175  */
10176 static void
10177 sfmmu_set_scd_rttecnt(sf_srd_t *srdp, sf_scd_t *scdp)
10178 {
10179 	uint_t rid;
10180 	uint_t i, j;
10181 	ulong_t w;
10182 	sf_region_t *rgnp;
10183 
10184 	ASSERT(srdp != NULL);
10185 
10186 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
10187 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
10188 			continue;
10189 		}
10190 
10191 		j = 0;
10192 		while (w) {
10193 			if (!(w & 0x1)) {
10194 				j++;
10195 				w >>= 1;
10196 				continue;
10197 			}
10198 			rid = (i << BT_ULSHIFT) | j;
10199 			j++;
10200 			w >>= 1;
10201 
10202 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
10203 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
10204 			rgnp = srdp->srd_hmergnp[rid];
10205 			ASSERT(rgnp->rgn_refcnt > 0);
10206 			ASSERT(rgnp->rgn_id == rid);
10207 
10208 			scdp->scd_rttecnt[rgnp->rgn_pgszc] +=
10209 			    rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
10210 
10211 			/*
10212 			 * Maintain the tsb0 inflation cnt for the regions
10213 			 * in the SCD.
10214 			 */
10215 			if (rgnp->rgn_pgszc >= TTE4M) {
10216 				scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt +=
10217 				    rgnp->rgn_size >>
10218 				    (TTE_PAGE_SHIFT(TTE8K) + 2);
10219 			}
10220 		}
10221 	}
10222 }
10223 
10224 /*
10225  * This function assumes that there are either four or six supported page
10226  * sizes and at most two programmable TLBs, so we need to decide which
10227  * page sizes are most important and then tell the MMU layer so it
10228  * can adjust the TLB page sizes accordingly (if supported).
10229  *
10230  * If these assumptions change, this function will need to be
10231  * updated to support whatever the new limits are.
10232  *
10233  * The growing flag is nonzero if we are growing the address space,
10234  * and zero if it is shrinking.  This allows us to decide whether
10235  * to grow or shrink our TSB, depending upon available memory
10236  * conditions.
10237  */
10238 static void
10239 sfmmu_check_page_sizes(sfmmu_t *sfmmup, int growing)
10240 {
10241 	uint64_t ttecnt[MMU_PAGE_SIZES];
10242 	uint64_t tte8k_cnt, tte4m_cnt;
10243 	uint8_t i;
10244 	int sectsb_thresh;
10245 
10246 	/*
10247 	 * Kernel threads, processes with small address spaces not using
10248 	 * large pages, and dummy ISM HATs need not apply.
10249 	 */
10250 	if (sfmmup == ksfmmup || sfmmup->sfmmu_ismhat != NULL)
10251 		return;
10252 
10253 	if (!SFMMU_LGPGS_INUSE(sfmmup) &&
10254 	    sfmmup->sfmmu_ttecnt[TTE8K] <= tsb_rss_factor)
10255 		return;
10256 
10257 	for (i = 0; i < mmu_page_sizes; i++) {
10258 		ttecnt[i] = sfmmup->sfmmu_ttecnt[i] +
10259 		    sfmmup->sfmmu_ismttecnt[i];
10260 	}
10261 
10262 	/* Check pagesizes in use, and possibly reprogram DTLB. */
10263 	if (&mmu_check_page_sizes)
10264 		mmu_check_page_sizes(sfmmup, ttecnt);
10265 
10266 	/*
10267 	 * Calculate the number of 8k ttes to represent the span of these
10268 	 * pages.
10269 	 */
10270 	tte8k_cnt = ttecnt[TTE8K] +
10271 	    (ttecnt[TTE64K] << (MMU_PAGESHIFT64K - MMU_PAGESHIFT)) +
10272 	    (ttecnt[TTE512K] << (MMU_PAGESHIFT512K - MMU_PAGESHIFT));
10273 	if (mmu_page_sizes == max_mmu_page_sizes) {
10274 		tte4m_cnt = ttecnt[TTE4M] +
10275 		    (ttecnt[TTE32M] << (MMU_PAGESHIFT32M - MMU_PAGESHIFT4M)) +
10276 		    (ttecnt[TTE256M] << (MMU_PAGESHIFT256M - MMU_PAGESHIFT4M));
10277 	} else {
10278 		tte4m_cnt = ttecnt[TTE4M];
10279 	}
10280 
10281 	/*
10282 	 * Inflate tte8k_cnt to allow for region large page allocation failure.
10283 	 */
10284 	tte8k_cnt += sfmmup->sfmmu_tsb0_4minflcnt;
10285 
10286 	/*
10287 	 * Inflate TSB sizes by a factor of 2 if this process
10288 	 * uses 4M text pages to minimize extra conflict misses
10289 	 * in the first TSB since without counting text pages
10290 	 * 8K TSB may become too small.
10291 	 *
10292 	 * Also double the size of the second TSB to minimize
10293 	 * extra conflict misses due to competition between 4M text pages
10294 	 * and data pages.
10295 	 *
10296 	 * We need to adjust the second TSB allocation threshold by the
10297 	 * inflation factor, since there is no point in creating a second
10298 	 * TSB when we know all the mappings can fit in the I/D TLBs.
10299 	 */
10300 	sectsb_thresh = tsb_sectsb_threshold;
10301 	if (sfmmup->sfmmu_flags & HAT_4MTEXT_FLAG) {
10302 		tte8k_cnt <<= 1;
10303 		tte4m_cnt <<= 1;
10304 		sectsb_thresh <<= 1;
10305 	}
10306 
10307 	/*
10308 	 * Check to see if our TSB is the right size; we may need to
10309 	 * grow or shrink it.  If the process is small, our work is
10310 	 * finished at this point.
10311 	 */
10312 	if (tte8k_cnt <= tsb_rss_factor && tte4m_cnt <= sectsb_thresh) {
10313 		return;
10314 	}
10315 	sfmmu_size_tsb(sfmmup, growing, tte8k_cnt, tte4m_cnt, sectsb_thresh);
10316 }
10317 
10318 static void
10319 sfmmu_size_tsb(sfmmu_t *sfmmup, int growing, uint64_t tte8k_cnt,
10320 	uint64_t tte4m_cnt, int sectsb_thresh)
10321 {
10322 	int tsb_bits;
10323 	uint_t tsb_szc;
10324 	struct tsb_info *tsbinfop;
10325 	hatlock_t *hatlockp = NULL;
10326 
10327 	hatlockp = sfmmu_hat_enter(sfmmup);
10328 	ASSERT(hatlockp != NULL);
10329 	tsbinfop = sfmmup->sfmmu_tsb;
10330 	ASSERT(tsbinfop != NULL);
10331 
10332 	/*
10333 	 * If we're growing, select the size based on RSS.  If we're
10334 	 * shrinking, leave some room so we don't have to turn around and
10335 	 * grow again immediately.
10336 	 */
10337 	if (growing)
10338 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
10339 	else
10340 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt << 1);
10341 
10342 	if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10343 	    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10344 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10345 		    hatlockp, TSB_SHRINK);
10346 	} else if (growing && tsb_szc > tsbinfop->tsb_szc && TSB_OK_GROW()) {
10347 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10348 		    hatlockp, TSB_GROW);
10349 	}
10350 	tsbinfop = sfmmup->sfmmu_tsb;
10351 
10352 	/*
10353 	 * With the TLB and first TSB out of the way, we need to see if
10354 	 * we need a second TSB for 4M pages.  If we managed to reprogram
10355 	 * the TLB page sizes above, the process will start using this new
10356 	 * TSB right away; otherwise, it will start using it on the next
10357 	 * context switch.  Either way, it's no big deal so there's no
10358 	 * synchronization with the trap handlers here unless we grow the
10359 	 * TSB (in which case it's required to prevent using the old one
10360 	 * after it's freed). Note: second tsb is required for 32M/256M
10361 	 * page sizes.
10362 	 */
10363 	if (tte4m_cnt > sectsb_thresh) {
10364 		/*
10365 		 * If we're growing, select the size based on RSS.  If we're
10366 		 * shrinking, leave some room so we don't have to turn
10367 		 * around and grow again immediately.
10368 		 */
10369 		if (growing)
10370 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
10371 		else
10372 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt << 1);
10373 		if (tsbinfop->tsb_next == NULL) {
10374 			struct tsb_info *newtsb;
10375 			int allocflags = SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)?
10376 			    0 : TSB_ALLOC;
10377 
10378 			sfmmu_hat_exit(hatlockp);
10379 
10380 			/*
10381 			 * Try to allocate a TSB for 4[32|256]M pages.  If we
10382 			 * can't get the size we want, retry w/a minimum sized
10383 			 * TSB.  If that still didn't work, give up; we can
10384 			 * still run without one.
10385 			 */
10386 			tsb_bits = (mmu_page_sizes == max_mmu_page_sizes)?
10387 			    TSB4M|TSB32M|TSB256M:TSB4M;
10388 			if ((sfmmu_tsbinfo_alloc(&newtsb, tsb_szc, tsb_bits,
10389 			    allocflags, sfmmup)) &&
10390 			    (tsb_szc <= TSB_4M_SZCODE ||
10391 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
10392 			    tsb_bits, allocflags, sfmmup)) &&
10393 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_MIN_SZCODE,
10394 			    tsb_bits, allocflags, sfmmup)) {
10395 				return;
10396 			}
10397 
10398 			hatlockp = sfmmu_hat_enter(sfmmup);
10399 
10400 			sfmmu_invalidate_ctx(sfmmup);
10401 
10402 			if (sfmmup->sfmmu_tsb->tsb_next == NULL) {
10403 				sfmmup->sfmmu_tsb->tsb_next = newtsb;
10404 				SFMMU_STAT(sf_tsb_sectsb_create);
10405 				sfmmu_hat_exit(hatlockp);
10406 				return;
10407 			} else {
10408 				/*
10409 				 * It's annoying, but possible for us
10410 				 * to get here.. we dropped the HAT lock
10411 				 * because of locking order in the kmem
10412 				 * allocator, and while we were off getting
10413 				 * our memory, some other thread decided to
10414 				 * do us a favor and won the race to get a
10415 				 * second TSB for this process.  Sigh.
10416 				 */
10417 				sfmmu_hat_exit(hatlockp);
10418 				sfmmu_tsbinfo_free(newtsb);
10419 				return;
10420 			}
10421 		}
10422 
10423 		/*
10424 		 * We have a second TSB, see if it's big enough.
10425 		 */
10426 		tsbinfop = tsbinfop->tsb_next;
10427 
10428 		/*
10429 		 * Check to see if our second TSB is the right size;
10430 		 * we may need to grow or shrink it.
10431 		 * To prevent thrashing (e.g. growing the TSB on a
10432 		 * subsequent map operation), only try to shrink if
10433 		 * the TSB reach exceeds twice the virtual address
10434 		 * space size.
10435 		 */
10436 		if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10437 		    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10438 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10439 			    tsb_szc, hatlockp, TSB_SHRINK);
10440 		} else if (growing && tsb_szc > tsbinfop->tsb_szc &&
10441 		    TSB_OK_GROW()) {
10442 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10443 			    tsb_szc, hatlockp, TSB_GROW);
10444 		}
10445 	}
10446 
10447 	sfmmu_hat_exit(hatlockp);
10448 }
10449 
10450 /*
10451  * Free up a sfmmu
10452  * Since the sfmmu is currently embedded in the hat struct we simply zero
10453  * out our fields and free up the ism map blk list if any.
10454  */
10455 static void
10456 sfmmu_free_sfmmu(sfmmu_t *sfmmup)
10457 {
10458 	ism_blk_t	*blkp, *nx_blkp;
10459 #ifdef	DEBUG
10460 	ism_map_t	*map;
10461 	int 		i;
10462 #endif
10463 
10464 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
10465 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
10466 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
10467 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
10468 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
10469 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
10470 	ASSERT(SF_RGNMAP_ISNULL(sfmmup));
10471 
10472 	sfmmup->sfmmu_free = 0;
10473 	sfmmup->sfmmu_ismhat = 0;
10474 
10475 	blkp = sfmmup->sfmmu_iblk;
10476 	sfmmup->sfmmu_iblk = NULL;
10477 
10478 	while (blkp) {
10479 #ifdef	DEBUG
10480 		map = blkp->iblk_maps;
10481 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
10482 			ASSERT(map[i].imap_seg == 0);
10483 			ASSERT(map[i].imap_ismhat == NULL);
10484 			ASSERT(map[i].imap_ment == NULL);
10485 		}
10486 #endif
10487 		nx_blkp = blkp->iblk_next;
10488 		blkp->iblk_next = NULL;
10489 		blkp->iblk_nextpa = (uint64_t)-1;
10490 		kmem_cache_free(ism_blk_cache, blkp);
10491 		blkp = nx_blkp;
10492 	}
10493 }
10494 
10495 /*
10496  * Locking primitves accessed by HATLOCK macros
10497  */
10498 
10499 #define	SFMMU_SPL_MTX	(0x0)
10500 #define	SFMMU_ML_MTX	(0x1)
10501 
10502 #define	SFMMU_MLSPL_MTX(type, pg)	(((type) == SFMMU_SPL_MTX) ? \
10503 					    SPL_HASH(pg) : MLIST_HASH(pg))
10504 
10505 kmutex_t *
10506 sfmmu_page_enter(struct page *pp)
10507 {
10508 	return (sfmmu_mlspl_enter(pp, SFMMU_SPL_MTX));
10509 }
10510 
10511 void
10512 sfmmu_page_exit(kmutex_t *spl)
10513 {
10514 	mutex_exit(spl);
10515 }
10516 
10517 int
10518 sfmmu_page_spl_held(struct page *pp)
10519 {
10520 	return (sfmmu_mlspl_held(pp, SFMMU_SPL_MTX));
10521 }
10522 
10523 kmutex_t *
10524 sfmmu_mlist_enter(struct page *pp)
10525 {
10526 	return (sfmmu_mlspl_enter(pp, SFMMU_ML_MTX));
10527 }
10528 
10529 void
10530 sfmmu_mlist_exit(kmutex_t *mml)
10531 {
10532 	mutex_exit(mml);
10533 }
10534 
10535 int
10536 sfmmu_mlist_held(struct page *pp)
10537 {
10538 
10539 	return (sfmmu_mlspl_held(pp, SFMMU_ML_MTX));
10540 }
10541 
10542 /*
10543  * Common code for sfmmu_mlist_enter() and sfmmu_page_enter().  For
10544  * sfmmu_mlist_enter() case mml_table lock array is used and for
10545  * sfmmu_page_enter() sfmmu_page_lock lock array is used.
10546  *
10547  * The lock is taken on a root page so that it protects an operation on all
10548  * constituent pages of a large page pp belongs to.
10549  *
10550  * The routine takes a lock from the appropriate array. The lock is determined
10551  * by hashing the root page. After taking the lock this routine checks if the
10552  * root page has the same size code that was used to determine the root (i.e
10553  * that root hasn't changed).  If root page has the expected p_szc field we
10554  * have the right lock and it's returned to the caller. If root's p_szc
10555  * decreased we release the lock and retry from the beginning.  This case can
10556  * happen due to hat_page_demote() decreasing p_szc between our load of p_szc
10557  * value and taking the lock. The number of retries due to p_szc decrease is
10558  * limited by the maximum p_szc value. If p_szc is 0 we return the lock
10559  * determined by hashing pp itself.
10560  *
10561  * If our caller doesn't hold a SE_SHARED or SE_EXCL lock on pp it's also
10562  * possible that p_szc can increase. To increase p_szc a thread has to lock
10563  * all constituent pages EXCL and do hat_pageunload() on all of them. All the
10564  * callers that don't hold a page locked recheck if hmeblk through which pp
10565  * was found still maps this pp.  If it doesn't map it anymore returned lock
10566  * is immediately dropped. Therefore if sfmmu_mlspl_enter() hits the case of
10567  * p_szc increase after taking the lock it returns this lock without further
10568  * retries because in this case the caller doesn't care about which lock was
10569  * taken. The caller will drop it right away.
10570  *
10571  * After the routine returns it's guaranteed that hat_page_demote() can't
10572  * change p_szc field of any of constituent pages of a large page pp belongs
10573  * to as long as pp was either locked at least SHARED prior to this call or
10574  * the caller finds that hment that pointed to this pp still references this
10575  * pp (this also assumes that the caller holds hme hash bucket lock so that
10576  * the same pp can't be remapped into the same hmeblk after it was unmapped by
10577  * hat_pageunload()).
10578  */
10579 static kmutex_t *
10580 sfmmu_mlspl_enter(struct page *pp, int type)
10581 {
10582 	kmutex_t	*mtx;
10583 	uint_t		prev_rszc = UINT_MAX;
10584 	page_t		*rootpp;
10585 	uint_t		szc;
10586 	uint_t		rszc;
10587 	uint_t		pszc = pp->p_szc;
10588 
10589 	ASSERT(pp != NULL);
10590 
10591 again:
10592 	if (pszc == 0) {
10593 		mtx = SFMMU_MLSPL_MTX(type, pp);
10594 		mutex_enter(mtx);
10595 		return (mtx);
10596 	}
10597 
10598 	/* The lock lives in the root page */
10599 	rootpp = PP_GROUPLEADER(pp, pszc);
10600 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10601 	mutex_enter(mtx);
10602 
10603 	/*
10604 	 * Return mml in the following 3 cases:
10605 	 *
10606 	 * 1) If pp itself is root since if its p_szc decreased before we took
10607 	 * the lock pp is still the root of smaller szc page. And if its p_szc
10608 	 * increased it doesn't matter what lock we return (see comment in
10609 	 * front of this routine).
10610 	 *
10611 	 * 2) If pp's not root but rootpp is the root of a rootpp->p_szc size
10612 	 * large page we have the right lock since any previous potential
10613 	 * hat_page_demote() is done demoting from greater than current root's
10614 	 * p_szc because hat_page_demote() changes root's p_szc last. No
10615 	 * further hat_page_demote() can start or be in progress since it
10616 	 * would need the same lock we currently hold.
10617 	 *
10618 	 * 3) If rootpp's p_szc increased since previous iteration it doesn't
10619 	 * matter what lock we return (see comment in front of this routine).
10620 	 */
10621 	if (pp == rootpp || (rszc = rootpp->p_szc) == pszc ||
10622 	    rszc >= prev_rszc) {
10623 		return (mtx);
10624 	}
10625 
10626 	/*
10627 	 * hat_page_demote() could have decreased root's p_szc.
10628 	 * In this case pp's p_szc must also be smaller than pszc.
10629 	 * Retry.
10630 	 */
10631 	if (rszc < pszc) {
10632 		szc = pp->p_szc;
10633 		if (szc < pszc) {
10634 			mutex_exit(mtx);
10635 			pszc = szc;
10636 			goto again;
10637 		}
10638 		/*
10639 		 * pp's p_szc increased after it was decreased.
10640 		 * page cannot be mapped. Return current lock. The caller
10641 		 * will drop it right away.
10642 		 */
10643 		return (mtx);
10644 	}
10645 
10646 	/*
10647 	 * root's p_szc is greater than pp's p_szc.
10648 	 * hat_page_demote() is not done with all pages
10649 	 * yet. Wait for it to complete.
10650 	 */
10651 	mutex_exit(mtx);
10652 	rootpp = PP_GROUPLEADER(rootpp, rszc);
10653 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10654 	mutex_enter(mtx);
10655 	mutex_exit(mtx);
10656 	prev_rszc = rszc;
10657 	goto again;
10658 }
10659 
10660 static int
10661 sfmmu_mlspl_held(struct page *pp, int type)
10662 {
10663 	kmutex_t	*mtx;
10664 
10665 	ASSERT(pp != NULL);
10666 	/* The lock lives in the root page */
10667 	pp = PP_PAGEROOT(pp);
10668 	ASSERT(pp != NULL);
10669 
10670 	mtx = SFMMU_MLSPL_MTX(type, pp);
10671 	return (MUTEX_HELD(mtx));
10672 }
10673 
10674 static uint_t
10675 sfmmu_get_free_hblk(struct hme_blk **hmeblkpp, uint_t critical)
10676 {
10677 	struct  hme_blk *hblkp;
10678 
10679 
10680 	if (freehblkp != NULL) {
10681 		mutex_enter(&freehblkp_lock);
10682 		if (freehblkp != NULL) {
10683 			/*
10684 			 * If the current thread is owning hblk_reserve OR
10685 			 * critical request from sfmmu_hblk_steal()
10686 			 * let it succeed even if freehblkcnt is really low.
10687 			 */
10688 			if (freehblkcnt <= HBLK_RESERVE_MIN && !critical) {
10689 				SFMMU_STAT(sf_get_free_throttle);
10690 				mutex_exit(&freehblkp_lock);
10691 				return (0);
10692 			}
10693 			freehblkcnt--;
10694 			*hmeblkpp = freehblkp;
10695 			hblkp = *hmeblkpp;
10696 			freehblkp = hblkp->hblk_next;
10697 			mutex_exit(&freehblkp_lock);
10698 			hblkp->hblk_next = NULL;
10699 			SFMMU_STAT(sf_get_free_success);
10700 
10701 			ASSERT(hblkp->hblk_hmecnt == 0);
10702 			ASSERT(hblkp->hblk_vcnt == 0);
10703 			ASSERT(hblkp->hblk_nextpa == va_to_pa((caddr_t)hblkp));
10704 
10705 			return (1);
10706 		}
10707 		mutex_exit(&freehblkp_lock);
10708 	}
10709 
10710 	/* Check cpu hblk pending queues */
10711 	if ((*hmeblkpp = sfmmu_check_pending_hblks(TTE8K)) != NULL) {
10712 		hblkp = *hmeblkpp;
10713 		hblkp->hblk_next = NULL;
10714 		hblkp->hblk_nextpa = va_to_pa((caddr_t)hblkp);
10715 
10716 		ASSERT(hblkp->hblk_hmecnt == 0);
10717 		ASSERT(hblkp->hblk_vcnt == 0);
10718 
10719 		return (1);
10720 	}
10721 
10722 	SFMMU_STAT(sf_get_free_fail);
10723 	return (0);
10724 }
10725 
10726 static uint_t
10727 sfmmu_put_free_hblk(struct hme_blk *hmeblkp, uint_t critical)
10728 {
10729 	struct  hme_blk *hblkp;
10730 
10731 	ASSERT(hmeblkp->hblk_hmecnt == 0);
10732 	ASSERT(hmeblkp->hblk_vcnt == 0);
10733 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
10734 
10735 	/*
10736 	 * If the current thread is mapping into kernel space,
10737 	 * let it succede even if freehblkcnt is max
10738 	 * so that it will avoid freeing it to kmem.
10739 	 * This will prevent stack overflow due to
10740 	 * possible recursion since kmem_cache_free()
10741 	 * might require creation of a slab which
10742 	 * in turn needs an hmeblk to map that slab;
10743 	 * let's break this vicious chain at the first
10744 	 * opportunity.
10745 	 */
10746 	if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10747 		mutex_enter(&freehblkp_lock);
10748 		if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10749 			SFMMU_STAT(sf_put_free_success);
10750 			freehblkcnt++;
10751 			hmeblkp->hblk_next = freehblkp;
10752 			freehblkp = hmeblkp;
10753 			mutex_exit(&freehblkp_lock);
10754 			return (1);
10755 		}
10756 		mutex_exit(&freehblkp_lock);
10757 	}
10758 
10759 	/*
10760 	 * Bring down freehblkcnt to HBLK_RESERVE_CNT. We are here
10761 	 * only if freehblkcnt is at least HBLK_RESERVE_CNT *and*
10762 	 * we are not in the process of mapping into kernel space.
10763 	 */
10764 	ASSERT(!critical);
10765 	while (freehblkcnt > HBLK_RESERVE_CNT) {
10766 		mutex_enter(&freehblkp_lock);
10767 		if (freehblkcnt > HBLK_RESERVE_CNT) {
10768 			freehblkcnt--;
10769 			hblkp = freehblkp;
10770 			freehblkp = hblkp->hblk_next;
10771 			mutex_exit(&freehblkp_lock);
10772 			ASSERT(get_hblk_cache(hblkp) == sfmmu8_cache);
10773 			kmem_cache_free(sfmmu8_cache, hblkp);
10774 			continue;
10775 		}
10776 		mutex_exit(&freehblkp_lock);
10777 	}
10778 	SFMMU_STAT(sf_put_free_fail);
10779 	return (0);
10780 }
10781 
10782 static void
10783 sfmmu_hblk_swap(struct hme_blk *new)
10784 {
10785 	struct hme_blk *old, *hblkp, *prev;
10786 	uint64_t newpa;
10787 	caddr_t	base, vaddr, endaddr;
10788 	struct hmehash_bucket *hmebp;
10789 	struct sf_hment *osfhme, *nsfhme;
10790 	page_t *pp;
10791 	kmutex_t *pml;
10792 	tte_t tte;
10793 	struct hme_blk *list = NULL;
10794 
10795 #ifdef	DEBUG
10796 	hmeblk_tag		hblktag;
10797 	struct hme_blk		*found;
10798 #endif
10799 	old = HBLK_RESERVE;
10800 	ASSERT(!old->hblk_shared);
10801 
10802 	/*
10803 	 * save pa before bcopy clobbers it
10804 	 */
10805 	newpa = new->hblk_nextpa;
10806 
10807 	base = (caddr_t)get_hblk_base(old);
10808 	endaddr = base + get_hblk_span(old);
10809 
10810 	/*
10811 	 * acquire hash bucket lock.
10812 	 */
10813 	hmebp = sfmmu_tteload_acquire_hashbucket(ksfmmup, base, TTE8K,
10814 	    SFMMU_INVALID_SHMERID);
10815 
10816 	/*
10817 	 * copy contents from old to new
10818 	 */
10819 	bcopy((void *)old, (void *)new, HME8BLK_SZ);
10820 
10821 	/*
10822 	 * add new to hash chain
10823 	 */
10824 	sfmmu_hblk_hash_add(hmebp, new, newpa);
10825 
10826 	/*
10827 	 * search hash chain for hblk_reserve; this needs to be performed
10828 	 * after adding new, otherwise prev won't correspond to the hblk which
10829 	 * is prior to old in hash chain when we call sfmmu_hblk_hash_rm to
10830 	 * remove old later.
10831 	 */
10832 	for (prev = NULL,
10833 	    hblkp = hmebp->hmeblkp; hblkp != NULL && hblkp != old;
10834 	    prev = hblkp, hblkp = hblkp->hblk_next)
10835 		;
10836 
10837 	if (hblkp != old)
10838 		panic("sfmmu_hblk_swap: hblk_reserve not found");
10839 
10840 	/*
10841 	 * p_mapping list is still pointing to hments in hblk_reserve;
10842 	 * fix up p_mapping list so that they point to hments in new.
10843 	 *
10844 	 * Since all these mappings are created by hblk_reserve_thread
10845 	 * on the way and it's using at least one of the buffers from each of
10846 	 * the newly minted slabs, there is no danger of any of these
10847 	 * mappings getting unloaded by another thread.
10848 	 *
10849 	 * tsbmiss could only modify ref/mod bits of hments in old/new.
10850 	 * Since all of these hments hold mappings established by segkmem
10851 	 * and mappings in segkmem are setup with HAT_NOSYNC, ref/mod bits
10852 	 * have no meaning for the mappings in hblk_reserve.  hments in
10853 	 * old and new are identical except for ref/mod bits.
10854 	 */
10855 	for (vaddr = base; vaddr < endaddr; vaddr += TTEBYTES(TTE8K)) {
10856 
10857 		HBLKTOHME(osfhme, old, vaddr);
10858 		sfmmu_copytte(&osfhme->hme_tte, &tte);
10859 
10860 		if (TTE_IS_VALID(&tte)) {
10861 			if ((pp = osfhme->hme_page) == NULL)
10862 				panic("sfmmu_hblk_swap: page not mapped");
10863 
10864 			pml = sfmmu_mlist_enter(pp);
10865 
10866 			if (pp != osfhme->hme_page)
10867 				panic("sfmmu_hblk_swap: mapping changed");
10868 
10869 			HBLKTOHME(nsfhme, new, vaddr);
10870 
10871 			HME_ADD(nsfhme, pp);
10872 			HME_SUB(osfhme, pp);
10873 
10874 			sfmmu_mlist_exit(pml);
10875 		}
10876 	}
10877 
10878 	/*
10879 	 * remove old from hash chain
10880 	 */
10881 	sfmmu_hblk_hash_rm(hmebp, old, prev, &list, 1);
10882 
10883 #ifdef	DEBUG
10884 
10885 	hblktag.htag_id = ksfmmup;
10886 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
10887 	hblktag.htag_bspage = HME_HASH_BSPAGE(base, HME_HASH_SHIFT(TTE8K));
10888 	hblktag.htag_rehash = HME_HASH_REHASH(TTE8K);
10889 	HME_HASH_FAST_SEARCH(hmebp, hblktag, found);
10890 
10891 	if (found != new)
10892 		panic("sfmmu_hblk_swap: new hblk not found");
10893 #endif
10894 
10895 	SFMMU_HASH_UNLOCK(hmebp);
10896 
10897 	/*
10898 	 * Reset hblk_reserve
10899 	 */
10900 	bzero((void *)old, HME8BLK_SZ);
10901 	old->hblk_nextpa = va_to_pa((caddr_t)old);
10902 }
10903 
10904 /*
10905  * Grab the mlist mutex for both pages passed in.
10906  *
10907  * low and high will be returned as pointers to the mutexes for these pages.
10908  * low refers to the mutex residing in the lower bin of the mlist hash, while
10909  * high refers to the mutex residing in the higher bin of the mlist hash.  This
10910  * is due to the locking order restrictions on the same thread grabbing
10911  * multiple mlist mutexes.  The low lock must be acquired before the high lock.
10912  *
10913  * If both pages hash to the same mutex, only grab that single mutex, and
10914  * high will be returned as NULL
10915  * If the pages hash to different bins in the hash, grab the lower addressed
10916  * lock first and then the higher addressed lock in order to follow the locking
10917  * rules involved with the same thread grabbing multiple mlist mutexes.
10918  * low and high will both have non-NULL values.
10919  */
10920 static void
10921 sfmmu_mlist_reloc_enter(struct page *targ, struct page *repl,
10922     kmutex_t **low, kmutex_t **high)
10923 {
10924 	kmutex_t	*mml_targ, *mml_repl;
10925 
10926 	/*
10927 	 * no need to do the dance around szc as in sfmmu_mlist_enter()
10928 	 * because this routine is only called by hat_page_relocate() and all
10929 	 * targ and repl pages are already locked EXCL so szc can't change.
10930 	 */
10931 
10932 	mml_targ = MLIST_HASH(PP_PAGEROOT(targ));
10933 	mml_repl = MLIST_HASH(PP_PAGEROOT(repl));
10934 
10935 	if (mml_targ == mml_repl) {
10936 		*low = mml_targ;
10937 		*high = NULL;
10938 	} else {
10939 		if (mml_targ < mml_repl) {
10940 			*low = mml_targ;
10941 			*high = mml_repl;
10942 		} else {
10943 			*low = mml_repl;
10944 			*high = mml_targ;
10945 		}
10946 	}
10947 
10948 	mutex_enter(*low);
10949 	if (*high)
10950 		mutex_enter(*high);
10951 }
10952 
10953 static void
10954 sfmmu_mlist_reloc_exit(kmutex_t *low, kmutex_t *high)
10955 {
10956 	if (high)
10957 		mutex_exit(high);
10958 	mutex_exit(low);
10959 }
10960 
10961 static hatlock_t *
10962 sfmmu_hat_enter(sfmmu_t *sfmmup)
10963 {
10964 	hatlock_t	*hatlockp;
10965 
10966 	if (sfmmup != ksfmmup) {
10967 		hatlockp = TSB_HASH(sfmmup);
10968 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
10969 		return (hatlockp);
10970 	}
10971 	return (NULL);
10972 }
10973 
10974 static hatlock_t *
10975 sfmmu_hat_tryenter(sfmmu_t *sfmmup)
10976 {
10977 	hatlock_t	*hatlockp;
10978 
10979 	if (sfmmup != ksfmmup) {
10980 		hatlockp = TSB_HASH(sfmmup);
10981 		if (mutex_tryenter(HATLOCK_MUTEXP(hatlockp)) == 0)
10982 			return (NULL);
10983 		return (hatlockp);
10984 	}
10985 	return (NULL);
10986 }
10987 
10988 static void
10989 sfmmu_hat_exit(hatlock_t *hatlockp)
10990 {
10991 	if (hatlockp != NULL)
10992 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
10993 }
10994 
10995 static void
10996 sfmmu_hat_lock_all(void)
10997 {
10998 	int i;
10999 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
11000 		mutex_enter(HATLOCK_MUTEXP(&hat_lock[i]));
11001 }
11002 
11003 static void
11004 sfmmu_hat_unlock_all(void)
11005 {
11006 	int i;
11007 	for (i = SFMMU_NUM_LOCK - 1; i >= 0; i--)
11008 		mutex_exit(HATLOCK_MUTEXP(&hat_lock[i]));
11009 }
11010 
11011 int
11012 sfmmu_hat_lock_held(sfmmu_t *sfmmup)
11013 {
11014 	ASSERT(sfmmup != ksfmmup);
11015 	return (MUTEX_HELD(HATLOCK_MUTEXP(TSB_HASH(sfmmup))));
11016 }
11017 
11018 /*
11019  * Locking primitives to provide consistency between ISM unmap
11020  * and other operations.  Since ISM unmap can take a long time, we
11021  * use HAT_ISMBUSY flag (protected by the hatlock) to avoid creating
11022  * contention on the hatlock buckets while ISM segments are being
11023  * unmapped.  The tradeoff is that the flags don't prevent priority
11024  * inversion from occurring, so we must request kernel priority in
11025  * case we have to sleep to keep from getting buried while holding
11026  * the HAT_ISMBUSY flag set, which in turn could block other kernel
11027  * threads from running (for example, in sfmmu_uvatopfn()).
11028  */
11029 static void
11030 sfmmu_ismhat_enter(sfmmu_t *sfmmup, int hatlock_held)
11031 {
11032 	hatlock_t *hatlockp;
11033 
11034 	THREAD_KPRI_REQUEST();
11035 	if (!hatlock_held)
11036 		hatlockp = sfmmu_hat_enter(sfmmup);
11037 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY))
11038 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
11039 	SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
11040 	if (!hatlock_held)
11041 		sfmmu_hat_exit(hatlockp);
11042 }
11043 
11044 static void
11045 sfmmu_ismhat_exit(sfmmu_t *sfmmup, int hatlock_held)
11046 {
11047 	hatlock_t *hatlockp;
11048 
11049 	if (!hatlock_held)
11050 		hatlockp = sfmmu_hat_enter(sfmmup);
11051 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
11052 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
11053 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11054 	if (!hatlock_held)
11055 		sfmmu_hat_exit(hatlockp);
11056 	THREAD_KPRI_RELEASE();
11057 }
11058 
11059 /*
11060  *
11061  * Algorithm:
11062  *
11063  * (1) if segkmem is not ready, allocate hblk from an array of pre-alloc'ed
11064  *	hblks.
11065  *
11066  * (2) if we are allocating an hblk for mapping a slab in sfmmu_cache,
11067  *
11068  * 		(a) try to return an hblk from reserve pool of free hblks;
11069  *		(b) if the reserve pool is empty, acquire hblk_reserve_lock
11070  *		    and return hblk_reserve.
11071  *
11072  * (3) call kmem_cache_alloc() to allocate hblk;
11073  *
11074  *		(a) if hblk_reserve_lock is held by the current thread,
11075  *		    atomically replace hblk_reserve by the hblk that is
11076  *		    returned by kmem_cache_alloc; release hblk_reserve_lock
11077  *		    and call kmem_cache_alloc() again.
11078  *		(b) if reserve pool is not full, add the hblk that is
11079  *		    returned by kmem_cache_alloc to reserve pool and
11080  *		    call kmem_cache_alloc again.
11081  *
11082  */
11083 static struct hme_blk *
11084 sfmmu_hblk_alloc(sfmmu_t *sfmmup, caddr_t vaddr,
11085 	struct hmehash_bucket *hmebp, uint_t size, hmeblk_tag hblktag,
11086 	uint_t flags, uint_t rid)
11087 {
11088 	struct hme_blk *hmeblkp = NULL;
11089 	struct hme_blk *newhblkp;
11090 	struct hme_blk *shw_hblkp = NULL;
11091 	struct kmem_cache *sfmmu_cache = NULL;
11092 	uint64_t hblkpa;
11093 	ulong_t index;
11094 	uint_t owner;		/* set to 1 if using hblk_reserve */
11095 	uint_t forcefree;
11096 	int sleep;
11097 	sf_srd_t *srdp;
11098 	sf_region_t *rgnp;
11099 
11100 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11101 	ASSERT(hblktag.htag_rid == rid);
11102 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
11103 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
11104 	    IS_P2ALIGNED(vaddr, TTEBYTES(size)));
11105 
11106 	/*
11107 	 * If segkmem is not created yet, allocate from static hmeblks
11108 	 * created at the end of startup_modules().  See the block comment
11109 	 * in startup_modules() describing how we estimate the number of
11110 	 * static hmeblks that will be needed during re-map.
11111 	 */
11112 	if (!hblk_alloc_dynamic) {
11113 
11114 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11115 
11116 		if (size == TTE8K) {
11117 			index = nucleus_hblk8.index;
11118 			if (index >= nucleus_hblk8.len) {
11119 				/*
11120 				 * If we panic here, see startup_modules() to
11121 				 * make sure that we are calculating the
11122 				 * number of hblk8's that we need correctly.
11123 				 */
11124 				prom_panic("no nucleus hblk8 to allocate");
11125 			}
11126 			hmeblkp =
11127 			    (struct hme_blk *)&nucleus_hblk8.list[index];
11128 			nucleus_hblk8.index++;
11129 			SFMMU_STAT(sf_hblk8_nalloc);
11130 		} else {
11131 			index = nucleus_hblk1.index;
11132 			if (nucleus_hblk1.index >= nucleus_hblk1.len) {
11133 				/*
11134 				 * If we panic here, see startup_modules().
11135 				 * Most likely you need to update the
11136 				 * calculation of the number of hblk1 elements
11137 				 * that the kernel needs to boot.
11138 				 */
11139 				prom_panic("no nucleus hblk1 to allocate");
11140 			}
11141 			hmeblkp =
11142 			    (struct hme_blk *)&nucleus_hblk1.list[index];
11143 			nucleus_hblk1.index++;
11144 			SFMMU_STAT(sf_hblk1_nalloc);
11145 		}
11146 
11147 		goto hblk_init;
11148 	}
11149 
11150 	SFMMU_HASH_UNLOCK(hmebp);
11151 
11152 	if (sfmmup != KHATID && !SFMMU_IS_SHMERID_VALID(rid)) {
11153 		if (mmu_page_sizes == max_mmu_page_sizes) {
11154 			if (size < TTE256M)
11155 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
11156 				    size, flags);
11157 		} else {
11158 			if (size < TTE4M)
11159 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
11160 				    size, flags);
11161 		}
11162 	} else if (SFMMU_IS_SHMERID_VALID(rid)) {
11163 		/*
11164 		 * Shared hmes use per region bitmaps in rgn_hmeflag
11165 		 * rather than shadow hmeblks to keep track of the
11166 		 * mapping sizes which have been allocated for the region.
11167 		 * Here we cleanup old invalid hmeblks with this rid,
11168 		 * which may be left around by pageunload().
11169 		 */
11170 		int ttesz;
11171 		caddr_t va;
11172 		caddr_t	eva = vaddr + TTEBYTES(size);
11173 
11174 		ASSERT(sfmmup != KHATID);
11175 
11176 		srdp = sfmmup->sfmmu_srdp;
11177 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11178 		rgnp = srdp->srd_hmergnp[rid];
11179 		ASSERT(rgnp != NULL && rgnp->rgn_id == rid);
11180 		ASSERT(rgnp->rgn_refcnt != 0);
11181 		ASSERT(size <= rgnp->rgn_pgszc);
11182 
11183 		ttesz = HBLK_MIN_TTESZ;
11184 		do {
11185 			if (!(rgnp->rgn_hmeflags & (0x1 << ttesz))) {
11186 				continue;
11187 			}
11188 
11189 			if (ttesz > size && ttesz != HBLK_MIN_TTESZ) {
11190 				sfmmu_cleanup_rhblk(srdp, vaddr, rid, ttesz);
11191 			} else if (ttesz < size) {
11192 				for (va = vaddr; va < eva;
11193 				    va += TTEBYTES(ttesz)) {
11194 					sfmmu_cleanup_rhblk(srdp, va, rid,
11195 					    ttesz);
11196 				}
11197 			}
11198 		} while (++ttesz <= rgnp->rgn_pgszc);
11199 	}
11200 
11201 fill_hblk:
11202 	owner = (hblk_reserve_thread == curthread) ? 1 : 0;
11203 
11204 	if (owner && size == TTE8K) {
11205 
11206 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11207 		/*
11208 		 * We are really in a tight spot. We already own
11209 		 * hblk_reserve and we need another hblk.  In anticipation
11210 		 * of this kind of scenario, we specifically set aside
11211 		 * HBLK_RESERVE_MIN number of hblks to be used exclusively
11212 		 * by owner of hblk_reserve.
11213 		 */
11214 		SFMMU_STAT(sf_hblk_recurse_cnt);
11215 
11216 		if (!sfmmu_get_free_hblk(&hmeblkp, 1))
11217 			panic("sfmmu_hblk_alloc: reserve list is empty");
11218 
11219 		goto hblk_verify;
11220 	}
11221 
11222 	ASSERT(!owner);
11223 
11224 	if ((flags & HAT_NO_KALLOC) == 0) {
11225 
11226 		sfmmu_cache = ((size == TTE8K) ? sfmmu8_cache : sfmmu1_cache);
11227 		sleep = ((sfmmup == KHATID) ? KM_NOSLEEP : KM_SLEEP);
11228 
11229 		if ((hmeblkp = kmem_cache_alloc(sfmmu_cache, sleep)) == NULL) {
11230 			hmeblkp = sfmmu_hblk_steal(size);
11231 		} else {
11232 			/*
11233 			 * if we are the owner of hblk_reserve,
11234 			 * swap hblk_reserve with hmeblkp and
11235 			 * start a fresh life.  Hope things go
11236 			 * better this time.
11237 			 */
11238 			if (hblk_reserve_thread == curthread) {
11239 				ASSERT(sfmmu_cache == sfmmu8_cache);
11240 				sfmmu_hblk_swap(hmeblkp);
11241 				hblk_reserve_thread = NULL;
11242 				mutex_exit(&hblk_reserve_lock);
11243 				goto fill_hblk;
11244 			}
11245 			/*
11246 			 * let's donate this hblk to our reserve list if
11247 			 * we are not mapping kernel range
11248 			 */
11249 			if (size == TTE8K && sfmmup != KHATID) {
11250 				if (sfmmu_put_free_hblk(hmeblkp, 0))
11251 					goto fill_hblk;
11252 			}
11253 		}
11254 	} else {
11255 		/*
11256 		 * We are here to map the slab in sfmmu8_cache; let's
11257 		 * check if we could tap our reserve list; if successful,
11258 		 * this will avoid the pain of going thru sfmmu_hblk_swap
11259 		 */
11260 		SFMMU_STAT(sf_hblk_slab_cnt);
11261 		if (!sfmmu_get_free_hblk(&hmeblkp, 0)) {
11262 			/*
11263 			 * let's start hblk_reserve dance
11264 			 */
11265 			SFMMU_STAT(sf_hblk_reserve_cnt);
11266 			owner = 1;
11267 			mutex_enter(&hblk_reserve_lock);
11268 			hmeblkp = HBLK_RESERVE;
11269 			hblk_reserve_thread = curthread;
11270 		}
11271 	}
11272 
11273 hblk_verify:
11274 	ASSERT(hmeblkp != NULL);
11275 	set_hblk_sz(hmeblkp, size);
11276 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
11277 	SFMMU_HASH_LOCK(hmebp);
11278 	HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11279 	if (newhblkp != NULL) {
11280 		SFMMU_HASH_UNLOCK(hmebp);
11281 		if (hmeblkp != HBLK_RESERVE) {
11282 			/*
11283 			 * This is really tricky!
11284 			 *
11285 			 * vmem_alloc(vmem_seg_arena)
11286 			 *  vmem_alloc(vmem_internal_arena)
11287 			 *   segkmem_alloc(heap_arena)
11288 			 *    vmem_alloc(heap_arena)
11289 			 *    page_create()
11290 			 *    hat_memload()
11291 			 *	kmem_cache_free()
11292 			 *	 kmem_cache_alloc()
11293 			 *	  kmem_slab_create()
11294 			 *	   vmem_alloc(kmem_internal_arena)
11295 			 *	    segkmem_alloc(heap_arena)
11296 			 *		vmem_alloc(heap_arena)
11297 			 *		page_create()
11298 			 *		hat_memload()
11299 			 *		  kmem_cache_free()
11300 			 *		...
11301 			 *
11302 			 * Thus, hat_memload() could call kmem_cache_free
11303 			 * for enough number of times that we could easily
11304 			 * hit the bottom of the stack or run out of reserve
11305 			 * list of vmem_seg structs.  So, we must donate
11306 			 * this hblk to reserve list if it's allocated
11307 			 * from sfmmu8_cache *and* mapping kernel range.
11308 			 * We don't need to worry about freeing hmeblk1's
11309 			 * to kmem since they don't map any kmem slabs.
11310 			 *
11311 			 * Note: When segkmem supports largepages, we must
11312 			 * free hmeblk1's to reserve list as well.
11313 			 */
11314 			forcefree = (sfmmup == KHATID) ? 1 : 0;
11315 			if (size == TTE8K &&
11316 			    sfmmu_put_free_hblk(hmeblkp, forcefree)) {
11317 				goto re_verify;
11318 			}
11319 			ASSERT(sfmmup != KHATID);
11320 			kmem_cache_free(get_hblk_cache(hmeblkp), hmeblkp);
11321 		} else {
11322 			/*
11323 			 * Hey! we don't need hblk_reserve any more.
11324 			 */
11325 			ASSERT(owner);
11326 			hblk_reserve_thread = NULL;
11327 			mutex_exit(&hblk_reserve_lock);
11328 			owner = 0;
11329 		}
11330 re_verify:
11331 		/*
11332 		 * let's check if the goodies are still present
11333 		 */
11334 		SFMMU_HASH_LOCK(hmebp);
11335 		HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11336 		if (newhblkp != NULL) {
11337 			/*
11338 			 * return newhblkp if it's not hblk_reserve;
11339 			 * if newhblkp is hblk_reserve, return it
11340 			 * _only if_ we are the owner of hblk_reserve.
11341 			 */
11342 			if (newhblkp != HBLK_RESERVE || owner) {
11343 				ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
11344 				    newhblkp->hblk_shared);
11345 				ASSERT(SFMMU_IS_SHMERID_VALID(rid) ||
11346 				    !newhblkp->hblk_shared);
11347 				return (newhblkp);
11348 			} else {
11349 				/*
11350 				 * we just hit hblk_reserve in the hash and
11351 				 * we are not the owner of that;
11352 				 *
11353 				 * block until hblk_reserve_thread completes
11354 				 * swapping hblk_reserve and try the dance
11355 				 * once again.
11356 				 */
11357 				SFMMU_HASH_UNLOCK(hmebp);
11358 				mutex_enter(&hblk_reserve_lock);
11359 				mutex_exit(&hblk_reserve_lock);
11360 				SFMMU_STAT(sf_hblk_reserve_hit);
11361 				goto fill_hblk;
11362 			}
11363 		} else {
11364 			/*
11365 			 * it's no more! try the dance once again.
11366 			 */
11367 			SFMMU_HASH_UNLOCK(hmebp);
11368 			goto fill_hblk;
11369 		}
11370 	}
11371 
11372 hblk_init:
11373 	if (SFMMU_IS_SHMERID_VALID(rid)) {
11374 		uint16_t tteflag = 0x1 <<
11375 		    ((size < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ : size);
11376 
11377 		if (!(rgnp->rgn_hmeflags & tteflag)) {
11378 			atomic_or_16(&rgnp->rgn_hmeflags, tteflag);
11379 		}
11380 		hmeblkp->hblk_shared = 1;
11381 	} else {
11382 		hmeblkp->hblk_shared = 0;
11383 	}
11384 	set_hblk_sz(hmeblkp, size);
11385 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11386 	hmeblkp->hblk_next = (struct hme_blk *)NULL;
11387 	hmeblkp->hblk_tag = hblktag;
11388 	hmeblkp->hblk_shadow = shw_hblkp;
11389 	hblkpa = hmeblkp->hblk_nextpa;
11390 	hmeblkp->hblk_nextpa = HMEBLK_ENDPA;
11391 
11392 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
11393 	ASSERT(get_hblk_span(hmeblkp) == HMEBLK_SPAN(size));
11394 	ASSERT(hmeblkp->hblk_hmecnt == 0);
11395 	ASSERT(hmeblkp->hblk_vcnt == 0);
11396 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11397 	ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
11398 	sfmmu_hblk_hash_add(hmebp, hmeblkp, hblkpa);
11399 	return (hmeblkp);
11400 }
11401 
11402 /*
11403  * This function cleans up the hme_blk and returns it to the free list.
11404  */
11405 /* ARGSUSED */
11406 static void
11407 sfmmu_hblk_free(struct hme_blk **listp)
11408 {
11409 	struct hme_blk *hmeblkp, *next_hmeblkp;
11410 	int		size;
11411 	uint_t		critical;
11412 	uint64_t	hblkpa;
11413 
11414 	ASSERT(*listp != NULL);
11415 
11416 	hmeblkp = *listp;
11417 	while (hmeblkp != NULL) {
11418 		next_hmeblkp = hmeblkp->hblk_next;
11419 		ASSERT(!hmeblkp->hblk_hmecnt);
11420 		ASSERT(!hmeblkp->hblk_vcnt);
11421 		ASSERT(!hmeblkp->hblk_lckcnt);
11422 		ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
11423 		ASSERT(hmeblkp->hblk_shared == 0);
11424 		ASSERT(hmeblkp->hblk_shw_bit == 0);
11425 		ASSERT(hmeblkp->hblk_shadow == NULL);
11426 
11427 		hblkpa = va_to_pa((caddr_t)hmeblkp);
11428 		ASSERT(hblkpa != (uint64_t)-1);
11429 		critical = (hblktosfmmu(hmeblkp) == KHATID) ? 1 : 0;
11430 
11431 		size = get_hblk_ttesz(hmeblkp);
11432 		hmeblkp->hblk_next = NULL;
11433 		hmeblkp->hblk_nextpa = hblkpa;
11434 
11435 		if (hmeblkp->hblk_nuc_bit == 0) {
11436 
11437 			if (size != TTE8K ||
11438 			    !sfmmu_put_free_hblk(hmeblkp, critical))
11439 				kmem_cache_free(get_hblk_cache(hmeblkp),
11440 				    hmeblkp);
11441 		}
11442 		hmeblkp = next_hmeblkp;
11443 	}
11444 }
11445 
11446 #define	BUCKETS_TO_SEARCH_BEFORE_UNLOAD	30
11447 #define	SFMMU_HBLK_STEAL_THRESHOLD 5
11448 
11449 static uint_t sfmmu_hblk_steal_twice;
11450 static uint_t sfmmu_hblk_steal_count, sfmmu_hblk_steal_unload_count;
11451 
11452 /*
11453  * Steal a hmeblk from user or kernel hme hash lists.
11454  * For 8K tte grab one from reserve pool (freehblkp) before proceeding to
11455  * steal and if we fail to steal after SFMMU_HBLK_STEAL_THRESHOLD attempts
11456  * tap into critical reserve of freehblkp.
11457  * Note: We remain looping in this routine until we find one.
11458  */
11459 static struct hme_blk *
11460 sfmmu_hblk_steal(int size)
11461 {
11462 	static struct hmehash_bucket *uhmehash_steal_hand = NULL;
11463 	struct hmehash_bucket *hmebp;
11464 	struct hme_blk *hmeblkp = NULL, *pr_hblk;
11465 	uint64_t hblkpa;
11466 	int i;
11467 	uint_t loop_cnt = 0, critical;
11468 
11469 	for (;;) {
11470 		/* Check cpu hblk pending queues */
11471 		if ((hmeblkp = sfmmu_check_pending_hblks(size)) != NULL) {
11472 			hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
11473 			ASSERT(hmeblkp->hblk_hmecnt == 0);
11474 			ASSERT(hmeblkp->hblk_vcnt == 0);
11475 			return (hmeblkp);
11476 		}
11477 
11478 		if (size == TTE8K) {
11479 			critical =
11480 			    (++loop_cnt > SFMMU_HBLK_STEAL_THRESHOLD) ? 1 : 0;
11481 			if (sfmmu_get_free_hblk(&hmeblkp, critical))
11482 				return (hmeblkp);
11483 		}
11484 
11485 		hmebp = (uhmehash_steal_hand == NULL) ? uhme_hash :
11486 		    uhmehash_steal_hand;
11487 		ASSERT(hmebp >= uhme_hash && hmebp <= &uhme_hash[UHMEHASH_SZ]);
11488 
11489 		for (i = 0; hmeblkp == NULL && i <= UHMEHASH_SZ +
11490 		    BUCKETS_TO_SEARCH_BEFORE_UNLOAD; i++) {
11491 			SFMMU_HASH_LOCK(hmebp);
11492 			hmeblkp = hmebp->hmeblkp;
11493 			hblkpa = hmebp->hmeh_nextpa;
11494 			pr_hblk = NULL;
11495 			while (hmeblkp) {
11496 				/*
11497 				 * check if it is a hmeblk that is not locked
11498 				 * and not shared. skip shadow hmeblks with
11499 				 * shadow_mask set i.e valid count non zero.
11500 				 */
11501 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11502 				    (hmeblkp->hblk_shw_bit == 0 ||
11503 				    hmeblkp->hblk_vcnt == 0) &&
11504 				    (hmeblkp->hblk_lckcnt == 0)) {
11505 					/*
11506 					 * there is a high probability that we
11507 					 * will find a free one. search some
11508 					 * buckets for a free hmeblk initially
11509 					 * before unloading a valid hmeblk.
11510 					 */
11511 					if ((hmeblkp->hblk_vcnt == 0 &&
11512 					    hmeblkp->hblk_hmecnt == 0) || (i >=
11513 					    BUCKETS_TO_SEARCH_BEFORE_UNLOAD)) {
11514 						if (sfmmu_steal_this_hblk(hmebp,
11515 						    hmeblkp, hblkpa, pr_hblk)) {
11516 							/*
11517 							 * Hblk is unloaded
11518 							 * successfully
11519 							 */
11520 							break;
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++ == &uhme_hash[UHMEHASH_SZ])
11531 				hmebp = uhme_hash;
11532 		}
11533 		uhmehash_steal_hand = hmebp;
11534 
11535 		if (hmeblkp != NULL)
11536 			break;
11537 
11538 		/*
11539 		 * in the worst case, look for a free one in the kernel
11540 		 * hash table.
11541 		 */
11542 		for (i = 0, hmebp = khme_hash; i <= KHMEHASH_SZ; i++) {
11543 			SFMMU_HASH_LOCK(hmebp);
11544 			hmeblkp = hmebp->hmeblkp;
11545 			hblkpa = hmebp->hmeh_nextpa;
11546 			pr_hblk = NULL;
11547 			while (hmeblkp) {
11548 				/*
11549 				 * check if it is free hmeblk
11550 				 */
11551 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11552 				    (hmeblkp->hblk_lckcnt == 0) &&
11553 				    (hmeblkp->hblk_vcnt == 0) &&
11554 				    (hmeblkp->hblk_hmecnt == 0)) {
11555 					if (sfmmu_steal_this_hblk(hmebp,
11556 					    hmeblkp, hblkpa, pr_hblk)) {
11557 						break;
11558 					} else {
11559 						/*
11560 						 * Cannot fail since we have
11561 						 * hash lock.
11562 						 */
11563 						panic("fail to steal?");
11564 					}
11565 				}
11566 
11567 				pr_hblk = hmeblkp;
11568 				hblkpa = hmeblkp->hblk_nextpa;
11569 				hmeblkp = hmeblkp->hblk_next;
11570 			}
11571 
11572 			SFMMU_HASH_UNLOCK(hmebp);
11573 			if (hmebp++ == &khme_hash[KHMEHASH_SZ])
11574 				hmebp = khme_hash;
11575 		}
11576 
11577 		if (hmeblkp != NULL)
11578 			break;
11579 		sfmmu_hblk_steal_twice++;
11580 	}
11581 	return (hmeblkp);
11582 }
11583 
11584 /*
11585  * This routine does real work to prepare a hblk to be "stolen" by
11586  * unloading the mappings, updating shadow counts ....
11587  * It returns 1 if the block is ready to be reused (stolen), or 0
11588  * means the block cannot be stolen yet- pageunload is still working
11589  * on this hblk.
11590  */
11591 static int
11592 sfmmu_steal_this_hblk(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
11593 	uint64_t hblkpa, struct hme_blk *pr_hblk)
11594 {
11595 	int shw_size, vshift;
11596 	struct hme_blk *shw_hblkp;
11597 	caddr_t vaddr;
11598 	uint_t shw_mask, newshw_mask;
11599 	struct hme_blk *list = NULL;
11600 
11601 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11602 
11603 	/*
11604 	 * check if the hmeblk is free, unload if necessary
11605 	 */
11606 	if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11607 		sfmmu_t *sfmmup;
11608 		demap_range_t dmr;
11609 
11610 		sfmmup = hblktosfmmu(hmeblkp);
11611 		if (hmeblkp->hblk_shared || sfmmup->sfmmu_ismhat) {
11612 			return (0);
11613 		}
11614 		DEMAP_RANGE_INIT(sfmmup, &dmr);
11615 		(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
11616 		    (caddr_t)get_hblk_base(hmeblkp),
11617 		    get_hblk_endaddr(hmeblkp), &dmr, HAT_UNLOAD);
11618 		DEMAP_RANGE_FLUSH(&dmr);
11619 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11620 			/*
11621 			 * Pageunload is working on the same hblk.
11622 			 */
11623 			return (0);
11624 		}
11625 
11626 		sfmmu_hblk_steal_unload_count++;
11627 	}
11628 
11629 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11630 	ASSERT(hmeblkp->hblk_vcnt == 0 && hmeblkp->hblk_hmecnt == 0);
11631 
11632 	sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 1);
11633 	hmeblkp->hblk_nextpa = hblkpa;
11634 
11635 	shw_hblkp = hmeblkp->hblk_shadow;
11636 	if (shw_hblkp) {
11637 		ASSERT(!hmeblkp->hblk_shared);
11638 		shw_size = get_hblk_ttesz(shw_hblkp);
11639 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
11640 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
11641 		ASSERT(vshift < 8);
11642 		/*
11643 		 * Atomically clear shadow mask bit
11644 		 */
11645 		do {
11646 			shw_mask = shw_hblkp->hblk_shw_mask;
11647 			ASSERT(shw_mask & (1 << vshift));
11648 			newshw_mask = shw_mask & ~(1 << vshift);
11649 			newshw_mask = cas32(&shw_hblkp->hblk_shw_mask,
11650 			    shw_mask, newshw_mask);
11651 		} while (newshw_mask != shw_mask);
11652 		hmeblkp->hblk_shadow = NULL;
11653 	}
11654 
11655 	/*
11656 	 * remove shadow bit if we are stealing an unused shadow hmeblk.
11657 	 * sfmmu_hblk_alloc needs it that way, will set shadow bit later if
11658 	 * we are indeed allocating a shadow hmeblk.
11659 	 */
11660 	hmeblkp->hblk_shw_bit = 0;
11661 
11662 	if (hmeblkp->hblk_shared) {
11663 		sf_srd_t	*srdp;
11664 		sf_region_t	*rgnp;
11665 		uint_t		rid;
11666 
11667 		srdp = hblktosrd(hmeblkp);
11668 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11669 		rid = hmeblkp->hblk_tag.htag_rid;
11670 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
11671 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
11672 		rgnp = srdp->srd_hmergnp[rid];
11673 		ASSERT(rgnp != NULL);
11674 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
11675 		hmeblkp->hblk_shared = 0;
11676 	}
11677 
11678 	sfmmu_hblk_steal_count++;
11679 	SFMMU_STAT(sf_steal_count);
11680 
11681 	return (1);
11682 }
11683 
11684 struct hme_blk *
11685 sfmmu_hmetohblk(struct sf_hment *sfhme)
11686 {
11687 	struct hme_blk *hmeblkp;
11688 	struct sf_hment *sfhme0;
11689 	struct hme_blk *hblk_dummy = 0;
11690 
11691 	/*
11692 	 * No dummy sf_hments, please.
11693 	 */
11694 	ASSERT(sfhme->hme_tte.ll != 0);
11695 
11696 	sfhme0 = sfhme - sfhme->hme_tte.tte_hmenum;
11697 	hmeblkp = (struct hme_blk *)((uintptr_t)sfhme0 -
11698 	    (uintptr_t)&hblk_dummy->hblk_hme[0]);
11699 
11700 	return (hmeblkp);
11701 }
11702 
11703 /*
11704  * On swapin, get appropriately sized TSB(s) and clear the HAT_SWAPPED flag.
11705  * If we can't get appropriately sized TSB(s), try for 8K TSB(s) using
11706  * KM_SLEEP allocation.
11707  *
11708  * Return 0 on success, -1 otherwise.
11709  */
11710 static void
11711 sfmmu_tsb_swapin(sfmmu_t *sfmmup, hatlock_t *hatlockp)
11712 {
11713 	struct tsb_info *tsbinfop, *next;
11714 	tsb_replace_rc_t rc;
11715 	boolean_t gotfirst = B_FALSE;
11716 
11717 	ASSERT(sfmmup != ksfmmup);
11718 	ASSERT(sfmmu_hat_lock_held(sfmmup));
11719 
11720 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPIN)) {
11721 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
11722 	}
11723 
11724 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
11725 		SFMMU_FLAGS_SET(sfmmup, HAT_SWAPIN);
11726 	} else {
11727 		return;
11728 	}
11729 
11730 	ASSERT(sfmmup->sfmmu_tsb != NULL);
11731 
11732 	/*
11733 	 * Loop over all tsbinfo's replacing them with ones that actually have
11734 	 * a TSB.  If any of the replacements ever fail, bail out of the loop.
11735 	 */
11736 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; tsbinfop = next) {
11737 		ASSERT(tsbinfop->tsb_flags & TSB_SWAPPED);
11738 		next = tsbinfop->tsb_next;
11739 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, tsbinfop->tsb_szc,
11740 		    hatlockp, TSB_SWAPIN);
11741 		if (rc != TSB_SUCCESS) {
11742 			break;
11743 		}
11744 		gotfirst = B_TRUE;
11745 	}
11746 
11747 	switch (rc) {
11748 	case TSB_SUCCESS:
11749 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11750 		cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11751 		return;
11752 	case TSB_LOSTRACE:
11753 		break;
11754 	case TSB_ALLOCFAIL:
11755 		break;
11756 	default:
11757 		panic("sfmmu_replace_tsb returned unrecognized failure code "
11758 		    "%d", rc);
11759 	}
11760 
11761 	/*
11762 	 * In this case, we failed to get one of our TSBs.  If we failed to
11763 	 * get the first TSB, get one of minimum size (8KB).  Walk the list
11764 	 * and throw away the tsbinfos, starting where the allocation failed;
11765 	 * we can get by with just one TSB as long as we don't leave the
11766 	 * SWAPPED tsbinfo structures lying around.
11767 	 */
11768 	tsbinfop = sfmmup->sfmmu_tsb;
11769 	next = tsbinfop->tsb_next;
11770 	tsbinfop->tsb_next = NULL;
11771 
11772 	sfmmu_hat_exit(hatlockp);
11773 	for (tsbinfop = next; tsbinfop != NULL; tsbinfop = next) {
11774 		next = tsbinfop->tsb_next;
11775 		sfmmu_tsbinfo_free(tsbinfop);
11776 	}
11777 	hatlockp = sfmmu_hat_enter(sfmmup);
11778 
11779 	/*
11780 	 * If we don't have any TSBs, get a single 8K TSB for 8K, 64K and 512K
11781 	 * pages.
11782 	 */
11783 	if (!gotfirst) {
11784 		tsbinfop = sfmmup->sfmmu_tsb;
11785 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, TSB_MIN_SZCODE,
11786 		    hatlockp, TSB_SWAPIN | TSB_FORCEALLOC);
11787 		ASSERT(rc == TSB_SUCCESS);
11788 	}
11789 
11790 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11791 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11792 }
11793 
11794 static int
11795 sfmmu_is_rgnva(sf_srd_t *srdp, caddr_t addr, ulong_t w, ulong_t bmw)
11796 {
11797 	ulong_t bix = 0;
11798 	uint_t rid;
11799 	sf_region_t *rgnp;
11800 
11801 	ASSERT(srdp != NULL);
11802 	ASSERT(srdp->srd_refcnt != 0);
11803 
11804 	w <<= BT_ULSHIFT;
11805 	while (bmw) {
11806 		if (!(bmw & 0x1)) {
11807 			bix++;
11808 			bmw >>= 1;
11809 			continue;
11810 		}
11811 		rid = w | bix;
11812 		rgnp = srdp->srd_hmergnp[rid];
11813 		ASSERT(rgnp->rgn_refcnt > 0);
11814 		ASSERT(rgnp->rgn_id == rid);
11815 		if (addr < rgnp->rgn_saddr ||
11816 		    addr >= (rgnp->rgn_saddr + rgnp->rgn_size)) {
11817 			bix++;
11818 			bmw >>= 1;
11819 		} else {
11820 			return (1);
11821 		}
11822 	}
11823 	return (0);
11824 }
11825 
11826 /*
11827  * Handle exceptions for low level tsb_handler.
11828  *
11829  * There are many scenarios that could land us here:
11830  *
11831  * If the context is invalid we land here. The context can be invalid
11832  * for 3 reasons: 1) we couldn't allocate a new context and now need to
11833  * perform a wrap around operation in order to allocate a new context.
11834  * 2) Context was invalidated to change pagesize programming 3) ISMs or
11835  * TSBs configuration is changeing for this process and we are forced into
11836  * here to do a syncronization operation. If the context is valid we can
11837  * be here from window trap hanlder. In this case just call trap to handle
11838  * the fault.
11839  *
11840  * Note that the process will run in INVALID_CONTEXT before
11841  * faulting into here and subsequently loading the MMU registers
11842  * (including the TSB base register) associated with this process.
11843  * For this reason, the trap handlers must all test for
11844  * INVALID_CONTEXT before attempting to access any registers other
11845  * than the context registers.
11846  */
11847 void
11848 sfmmu_tsbmiss_exception(struct regs *rp, uintptr_t tagaccess, uint_t traptype)
11849 {
11850 	sfmmu_t *sfmmup, *shsfmmup;
11851 	uint_t ctxtype;
11852 	klwp_id_t lwp;
11853 	char lwp_save_state;
11854 	hatlock_t *hatlockp, *shatlockp;
11855 	struct tsb_info *tsbinfop;
11856 	struct tsbmiss *tsbmp;
11857 	sf_scd_t *scdp;
11858 
11859 	SFMMU_STAT(sf_tsb_exceptions);
11860 	SFMMU_MMU_STAT(mmu_tsb_exceptions);
11861 	sfmmup = astosfmmu(curthread->t_procp->p_as);
11862 	/*
11863 	 * note that in sun4u, tagacces register contains ctxnum
11864 	 * while sun4v passes ctxtype in the tagaccess register.
11865 	 */
11866 	ctxtype = tagaccess & TAGACC_CTX_MASK;
11867 
11868 	ASSERT(sfmmup != ksfmmup && ctxtype != KCONTEXT);
11869 	ASSERT(sfmmup->sfmmu_ismhat == 0);
11870 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED) ||
11871 	    ctxtype == INVALID_CONTEXT);
11872 
11873 	if (ctxtype != INVALID_CONTEXT && traptype != T_DATA_PROT) {
11874 		/*
11875 		 * We may land here because shme bitmap and pagesize
11876 		 * flags are updated lazily in tsbmiss area on other cpus.
11877 		 * If we detect here that tsbmiss area is out of sync with
11878 		 * sfmmu update it and retry the trapped instruction.
11879 		 * Otherwise call trap().
11880 		 */
11881 		int ret = 0;
11882 		uchar_t tteflag_mask = (1 << TTE64K) | (1 << TTE8K);
11883 		caddr_t addr = (caddr_t)(tagaccess & TAGACC_VADDR_MASK);
11884 
11885 		/*
11886 		 * Must set lwp state to LWP_SYS before
11887 		 * trying to acquire any adaptive lock
11888 		 */
11889 		lwp = ttolwp(curthread);
11890 		ASSERT(lwp);
11891 		lwp_save_state = lwp->lwp_state;
11892 		lwp->lwp_state = LWP_SYS;
11893 
11894 		hatlockp = sfmmu_hat_enter(sfmmup);
11895 		kpreempt_disable();
11896 		tsbmp = &tsbmiss_area[CPU->cpu_id];
11897 		ASSERT(sfmmup == tsbmp->usfmmup);
11898 		if (((tsbmp->uhat_tteflags ^ sfmmup->sfmmu_tteflags) &
11899 		    ~tteflag_mask) ||
11900 		    ((tsbmp->uhat_rtteflags ^  sfmmup->sfmmu_rtteflags) &
11901 		    ~tteflag_mask)) {
11902 			tsbmp->uhat_tteflags = sfmmup->sfmmu_tteflags;
11903 			tsbmp->uhat_rtteflags = sfmmup->sfmmu_rtteflags;
11904 			ret = 1;
11905 		}
11906 		if (sfmmup->sfmmu_srdp != NULL) {
11907 			ulong_t *sm = sfmmup->sfmmu_hmeregion_map.bitmap;
11908 			ulong_t *tm = tsbmp->shmermap;
11909 			ulong_t i;
11910 			for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
11911 				ulong_t d = tm[i] ^ sm[i];
11912 				if (d) {
11913 					if (d & sm[i]) {
11914 						if (!ret && sfmmu_is_rgnva(
11915 						    sfmmup->sfmmu_srdp,
11916 						    addr, i, d & sm[i])) {
11917 							ret = 1;
11918 						}
11919 					}
11920 					tm[i] = sm[i];
11921 				}
11922 			}
11923 		}
11924 		kpreempt_enable();
11925 		sfmmu_hat_exit(hatlockp);
11926 		lwp->lwp_state = lwp_save_state;
11927 		if (ret) {
11928 			return;
11929 		}
11930 	} else if (ctxtype == INVALID_CONTEXT) {
11931 		/*
11932 		 * First, make sure we come out of here with a valid ctx,
11933 		 * since if we don't get one we'll simply loop on the
11934 		 * faulting instruction.
11935 		 *
11936 		 * If the ISM mappings are changing, the TSB is relocated,
11937 		 * the process is swapped, the process is joining SCD or
11938 		 * leaving SCD or shared regions we serialize behind the
11939 		 * controlling thread with hat lock, sfmmu_flags and
11940 		 * sfmmu_tsb_cv condition variable.
11941 		 */
11942 
11943 		/*
11944 		 * Must set lwp state to LWP_SYS before
11945 		 * trying to acquire any adaptive lock
11946 		 */
11947 		lwp = ttolwp(curthread);
11948 		ASSERT(lwp);
11949 		lwp_save_state = lwp->lwp_state;
11950 		lwp->lwp_state = LWP_SYS;
11951 
11952 		hatlockp = sfmmu_hat_enter(sfmmup);
11953 retry:
11954 		if ((scdp = sfmmup->sfmmu_scdp) != NULL) {
11955 			shsfmmup = scdp->scd_sfmmup;
11956 			ASSERT(shsfmmup != NULL);
11957 
11958 			for (tsbinfop = shsfmmup->sfmmu_tsb; tsbinfop != NULL;
11959 			    tsbinfop = tsbinfop->tsb_next) {
11960 				if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
11961 					/* drop the private hat lock */
11962 					sfmmu_hat_exit(hatlockp);
11963 					/* acquire the shared hat lock */
11964 					shatlockp = sfmmu_hat_enter(shsfmmup);
11965 					/*
11966 					 * recheck to see if anything changed
11967 					 * after we drop the private hat lock.
11968 					 */
11969 					if (sfmmup->sfmmu_scdp == scdp &&
11970 					    shsfmmup == scdp->scd_sfmmup) {
11971 						sfmmu_tsb_chk_reloc(shsfmmup,
11972 						    shatlockp);
11973 					}
11974 					sfmmu_hat_exit(shatlockp);
11975 					hatlockp = sfmmu_hat_enter(sfmmup);
11976 					goto retry;
11977 				}
11978 			}
11979 		}
11980 
11981 		for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
11982 		    tsbinfop = tsbinfop->tsb_next) {
11983 			if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
11984 				cv_wait(&sfmmup->sfmmu_tsb_cv,
11985 				    HATLOCK_MUTEXP(hatlockp));
11986 				goto retry;
11987 			}
11988 		}
11989 
11990 		/*
11991 		 * Wait for ISM maps to be updated.
11992 		 */
11993 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
11994 			cv_wait(&sfmmup->sfmmu_tsb_cv,
11995 			    HATLOCK_MUTEXP(hatlockp));
11996 			goto retry;
11997 		}
11998 
11999 		/* Is this process joining an SCD? */
12000 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
12001 			/*
12002 			 * Flush private TSB and setup shared TSB.
12003 			 * sfmmu_finish_join_scd() does not drop the
12004 			 * hat lock.
12005 			 */
12006 			sfmmu_finish_join_scd(sfmmup);
12007 			SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
12008 		}
12009 
12010 		/*
12011 		 * If we're swapping in, get TSB(s).  Note that we must do
12012 		 * this before we get a ctx or load the MMU state.  Once
12013 		 * we swap in we have to recheck to make sure the TSB(s) and
12014 		 * ISM mappings didn't change while we slept.
12015 		 */
12016 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
12017 			sfmmu_tsb_swapin(sfmmup, hatlockp);
12018 			goto retry;
12019 		}
12020 
12021 		sfmmu_get_ctx(sfmmup);
12022 
12023 		sfmmu_hat_exit(hatlockp);
12024 		/*
12025 		 * Must restore lwp_state if not calling
12026 		 * trap() for further processing. Restore
12027 		 * it anyway.
12028 		 */
12029 		lwp->lwp_state = lwp_save_state;
12030 		return;
12031 	}
12032 	trap(rp, (caddr_t)tagaccess, traptype, 0);
12033 }
12034 
12035 static void
12036 sfmmu_tsb_chk_reloc(sfmmu_t *sfmmup, hatlock_t *hatlockp)
12037 {
12038 	struct tsb_info *tp;
12039 
12040 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12041 
12042 	for (tp = sfmmup->sfmmu_tsb; tp != NULL; tp = tp->tsb_next) {
12043 		if (tp->tsb_flags & TSB_RELOC_FLAG) {
12044 			cv_wait(&sfmmup->sfmmu_tsb_cv,
12045 			    HATLOCK_MUTEXP(hatlockp));
12046 			break;
12047 		}
12048 	}
12049 }
12050 
12051 /*
12052  * sfmmu_vatopfn_suspended is called from GET_TTE when TL=0 and
12053  * TTE_SUSPENDED bit set in tte we block on aquiring a page lock
12054  * rather than spinning to avoid send mondo timeouts with
12055  * interrupts enabled. When the lock is acquired it is immediately
12056  * released and we return back to sfmmu_vatopfn just after
12057  * the GET_TTE call.
12058  */
12059 void
12060 sfmmu_vatopfn_suspended(caddr_t vaddr, sfmmu_t *sfmmu, tte_t *ttep)
12061 {
12062 	struct page	**pp;
12063 
12064 	(void) as_pagelock(sfmmu->sfmmu_as, &pp, vaddr, TTE_CSZ(ttep), S_WRITE);
12065 	as_pageunlock(sfmmu->sfmmu_as, pp, vaddr, TTE_CSZ(ttep), S_WRITE);
12066 }
12067 
12068 /*
12069  * sfmmu_tsbmiss_suspended is called from GET_TTE when TL>0 and
12070  * TTE_SUSPENDED bit set in tte. We do this so that we can handle
12071  * cross traps which cannot be handled while spinning in the
12072  * trap handlers. Simply enter and exit the kpr_suspendlock spin
12073  * mutex, which is held by the holder of the suspend bit, and then
12074  * retry the trapped instruction after unwinding.
12075  */
12076 /*ARGSUSED*/
12077 void
12078 sfmmu_tsbmiss_suspended(struct regs *rp, uintptr_t tagacc, uint_t traptype)
12079 {
12080 	ASSERT(curthread != kreloc_thread);
12081 	mutex_enter(&kpr_suspendlock);
12082 	mutex_exit(&kpr_suspendlock);
12083 }
12084 
12085 /*
12086  * This routine could be optimized to reduce the number of xcalls by flushing
12087  * the entire TLBs if region reference count is above some threshold but the
12088  * tradeoff will depend on the size of the TLB. So for now flush the specific
12089  * page a context at a time.
12090  *
12091  * If uselocks is 0 then it's called after all cpus were captured and all the
12092  * hat locks were taken. In this case don't take the region lock by relying on
12093  * the order of list region update operations in hat_join_region(),
12094  * hat_leave_region() and hat_dup_region(). The ordering in those routines
12095  * guarantees that list is always forward walkable and reaches active sfmmus
12096  * regardless of where xc_attention() captures a cpu.
12097  */
12098 cpuset_t
12099 sfmmu_rgntlb_demap(caddr_t addr, sf_region_t *rgnp,
12100     struct hme_blk *hmeblkp, int uselocks)
12101 {
12102 	sfmmu_t	*sfmmup;
12103 	cpuset_t cpuset;
12104 	cpuset_t rcpuset;
12105 	hatlock_t *hatlockp;
12106 	uint_t rid = rgnp->rgn_id;
12107 	sf_rgn_link_t *rlink;
12108 	sf_scd_t *scdp;
12109 
12110 	ASSERT(hmeblkp->hblk_shared);
12111 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
12112 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
12113 
12114 	CPUSET_ZERO(rcpuset);
12115 	if (uselocks) {
12116 		mutex_enter(&rgnp->rgn_mutex);
12117 	}
12118 	sfmmup = rgnp->rgn_sfmmu_head;
12119 	while (sfmmup != NULL) {
12120 		if (uselocks) {
12121 			hatlockp = sfmmu_hat_enter(sfmmup);
12122 		}
12123 
12124 		/*
12125 		 * When an SCD is created the SCD hat is linked on the sfmmu
12126 		 * region lists for each hme region which is part of the
12127 		 * SCD. If we find an SCD hat, when walking these lists,
12128 		 * then we flush the shared TSBs, if we find a private hat,
12129 		 * which is part of an SCD, but where the region
12130 		 * is not part of the SCD then we flush the private TSBs.
12131 		 */
12132 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12133 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
12134 			scdp = sfmmup->sfmmu_scdp;
12135 			if (SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
12136 				if (uselocks) {
12137 					sfmmu_hat_exit(hatlockp);
12138 				}
12139 				goto next;
12140 			}
12141 		}
12142 
12143 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12144 
12145 		kpreempt_disable();
12146 		cpuset = sfmmup->sfmmu_cpusran;
12147 		CPUSET_AND(cpuset, cpu_ready_set);
12148 		CPUSET_DEL(cpuset, CPU->cpu_id);
12149 		SFMMU_XCALL_STATS(sfmmup);
12150 		xt_some(cpuset, vtag_flushpage_tl1,
12151 		    (uint64_t)addr, (uint64_t)sfmmup);
12152 		vtag_flushpage(addr, (uint64_t)sfmmup);
12153 		if (uselocks) {
12154 			sfmmu_hat_exit(hatlockp);
12155 		}
12156 		kpreempt_enable();
12157 		CPUSET_OR(rcpuset, cpuset);
12158 
12159 next:
12160 		/* LINTED: constant in conditional context */
12161 		SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
12162 		ASSERT(rlink != NULL);
12163 		sfmmup = rlink->next;
12164 	}
12165 	if (uselocks) {
12166 		mutex_exit(&rgnp->rgn_mutex);
12167 	}
12168 	return (rcpuset);
12169 }
12170 
12171 /*
12172  * This routine takes an sfmmu pointer and the va for an adddress in an
12173  * ISM region as input and returns the corresponding region id in ism_rid.
12174  * The return value of 1 indicates that a region has been found and ism_rid
12175  * is valid, otherwise 0 is returned.
12176  */
12177 static int
12178 find_ism_rid(sfmmu_t *sfmmup, sfmmu_t *ism_sfmmup, caddr_t va, uint_t *ism_rid)
12179 {
12180 	ism_blk_t	*ism_blkp;
12181 	int		i;
12182 	ism_map_t	*ism_map;
12183 #ifdef DEBUG
12184 	struct hat	*ism_hatid;
12185 #endif
12186 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12187 
12188 	ism_blkp = sfmmup->sfmmu_iblk;
12189 	while (ism_blkp != NULL) {
12190 		ism_map = ism_blkp->iblk_maps;
12191 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
12192 			if ((va >= ism_start(ism_map[i])) &&
12193 			    (va < ism_end(ism_map[i]))) {
12194 
12195 				*ism_rid = ism_map[i].imap_rid;
12196 #ifdef DEBUG
12197 				ism_hatid = ism_map[i].imap_ismhat;
12198 				ASSERT(ism_hatid == ism_sfmmup);
12199 				ASSERT(ism_hatid->sfmmu_ismhat);
12200 #endif
12201 				return (1);
12202 			}
12203 		}
12204 		ism_blkp = ism_blkp->iblk_next;
12205 	}
12206 	return (0);
12207 }
12208 
12209 /*
12210  * Special routine to flush out ism mappings- TSBs, TLBs and D-caches.
12211  * This routine may be called with all cpu's captured. Therefore, the
12212  * caller is responsible for holding all locks and disabling kernel
12213  * preemption.
12214  */
12215 /* ARGSUSED */
12216 static void
12217 sfmmu_ismtlbcache_demap(caddr_t addr, sfmmu_t *ism_sfmmup,
12218 	struct hme_blk *hmeblkp, pfn_t pfnum, int cache_flush_flag)
12219 {
12220 	cpuset_t 	cpuset;
12221 	caddr_t 	va;
12222 	ism_ment_t	*ment;
12223 	sfmmu_t		*sfmmup;
12224 #ifdef VAC
12225 	int 		vcolor;
12226 #endif
12227 
12228 	sf_scd_t	*scdp;
12229 	uint_t		ism_rid;
12230 
12231 	ASSERT(!hmeblkp->hblk_shared);
12232 	/*
12233 	 * Walk the ism_hat's mapping list and flush the page
12234 	 * from every hat sharing this ism_hat. This routine
12235 	 * may be called while all cpu's have been captured.
12236 	 * Therefore we can't attempt to grab any locks. For now
12237 	 * this means we will protect the ism mapping list under
12238 	 * a single lock which will be grabbed by the caller.
12239 	 * If hat_share/unshare scalibility becomes a performance
12240 	 * problem then we may need to re-think ism mapping list locking.
12241 	 */
12242 	ASSERT(ism_sfmmup->sfmmu_ismhat);
12243 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
12244 	addr = addr - ISMID_STARTADDR;
12245 
12246 	for (ment = ism_sfmmup->sfmmu_iment; ment; ment = ment->iment_next) {
12247 
12248 		sfmmup = ment->iment_hat;
12249 
12250 		va = ment->iment_base_va;
12251 		va = (caddr_t)((uintptr_t)va  + (uintptr_t)addr);
12252 
12253 		/*
12254 		 * When an SCD is created the SCD hat is linked on the ism
12255 		 * mapping lists for each ISM segment which is part of the
12256 		 * SCD. If we find an SCD hat, when walking these lists,
12257 		 * then we flush the shared TSBs, if we find a private hat,
12258 		 * which is part of an SCD, but where the region
12259 		 * corresponding to this va is not part of the SCD then we
12260 		 * flush the private TSBs.
12261 		 */
12262 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12263 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD) &&
12264 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
12265 			if (!find_ism_rid(sfmmup, ism_sfmmup, va,
12266 			    &ism_rid)) {
12267 				cmn_err(CE_PANIC,
12268 				    "can't find matching ISM rid!");
12269 			}
12270 
12271 			scdp = sfmmup->sfmmu_scdp;
12272 			if (SFMMU_IS_ISMRID_VALID(ism_rid) &&
12273 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map,
12274 			    ism_rid)) {
12275 				continue;
12276 			}
12277 		}
12278 		SFMMU_UNLOAD_TSB(va, sfmmup, hmeblkp, 1);
12279 
12280 		cpuset = sfmmup->sfmmu_cpusran;
12281 		CPUSET_AND(cpuset, cpu_ready_set);
12282 		CPUSET_DEL(cpuset, CPU->cpu_id);
12283 		SFMMU_XCALL_STATS(sfmmup);
12284 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)va,
12285 		    (uint64_t)sfmmup);
12286 		vtag_flushpage(va, (uint64_t)sfmmup);
12287 
12288 #ifdef VAC
12289 		/*
12290 		 * Flush D$
12291 		 * When flushing D$ we must flush all
12292 		 * cpu's. See sfmmu_cache_flush().
12293 		 */
12294 		if (cache_flush_flag == CACHE_FLUSH) {
12295 			cpuset = cpu_ready_set;
12296 			CPUSET_DEL(cpuset, CPU->cpu_id);
12297 
12298 			SFMMU_XCALL_STATS(sfmmup);
12299 			vcolor = addr_to_vcolor(va);
12300 			xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12301 			vac_flushpage(pfnum, vcolor);
12302 		}
12303 #endif	/* VAC */
12304 	}
12305 }
12306 
12307 /*
12308  * Demaps the TSB, CPU caches, and flushes all TLBs on all CPUs of
12309  * a particular virtual address and ctx.  If noflush is set we do not
12310  * flush the TLB/TSB.  This function may or may not be called with the
12311  * HAT lock held.
12312  */
12313 static void
12314 sfmmu_tlbcache_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12315 	pfn_t pfnum, int tlb_noflush, int cpu_flag, int cache_flush_flag,
12316 	int hat_lock_held)
12317 {
12318 #ifdef VAC
12319 	int vcolor;
12320 #endif
12321 	cpuset_t cpuset;
12322 	hatlock_t *hatlockp;
12323 
12324 	ASSERT(!hmeblkp->hblk_shared);
12325 
12326 #if defined(lint) && !defined(VAC)
12327 	pfnum = pfnum;
12328 	cpu_flag = cpu_flag;
12329 	cache_flush_flag = cache_flush_flag;
12330 #endif
12331 
12332 	/*
12333 	 * There is no longer a need to protect against ctx being
12334 	 * stolen here since we don't store the ctx in the TSB anymore.
12335 	 */
12336 #ifdef VAC
12337 	vcolor = addr_to_vcolor(addr);
12338 #endif
12339 
12340 	/*
12341 	 * We must hold the hat lock during the flush of TLB,
12342 	 * to avoid a race with sfmmu_invalidate_ctx(), where
12343 	 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12344 	 * causing TLB demap routine to skip flush on that MMU.
12345 	 * If the context on a MMU has already been set to
12346 	 * INVALID_CONTEXT, we just get an extra flush on
12347 	 * that MMU.
12348 	 */
12349 	if (!hat_lock_held && !tlb_noflush)
12350 		hatlockp = sfmmu_hat_enter(sfmmup);
12351 
12352 	kpreempt_disable();
12353 	if (!tlb_noflush) {
12354 		/*
12355 		 * Flush the TSB and TLB.
12356 		 */
12357 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12358 
12359 		cpuset = sfmmup->sfmmu_cpusran;
12360 		CPUSET_AND(cpuset, cpu_ready_set);
12361 		CPUSET_DEL(cpuset, CPU->cpu_id);
12362 
12363 		SFMMU_XCALL_STATS(sfmmup);
12364 
12365 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
12366 		    (uint64_t)sfmmup);
12367 
12368 		vtag_flushpage(addr, (uint64_t)sfmmup);
12369 	}
12370 
12371 	if (!hat_lock_held && !tlb_noflush)
12372 		sfmmu_hat_exit(hatlockp);
12373 
12374 #ifdef VAC
12375 	/*
12376 	 * Flush the D$
12377 	 *
12378 	 * Even if the ctx is stolen, we need to flush the
12379 	 * cache. Our ctx stealer only flushes the TLBs.
12380 	 */
12381 	if (cache_flush_flag == CACHE_FLUSH) {
12382 		if (cpu_flag & FLUSH_ALL_CPUS) {
12383 			cpuset = cpu_ready_set;
12384 		} else {
12385 			cpuset = sfmmup->sfmmu_cpusran;
12386 			CPUSET_AND(cpuset, cpu_ready_set);
12387 		}
12388 		CPUSET_DEL(cpuset, CPU->cpu_id);
12389 		SFMMU_XCALL_STATS(sfmmup);
12390 		xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12391 		vac_flushpage(pfnum, vcolor);
12392 	}
12393 #endif	/* VAC */
12394 	kpreempt_enable();
12395 }
12396 
12397 /*
12398  * Demaps the TSB and flushes all TLBs on all cpus for a particular virtual
12399  * address and ctx.  If noflush is set we do not currently do anything.
12400  * This function may or may not be called with the HAT lock held.
12401  */
12402 static void
12403 sfmmu_tlb_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12404 	int tlb_noflush, int hat_lock_held)
12405 {
12406 	cpuset_t cpuset;
12407 	hatlock_t *hatlockp;
12408 
12409 	ASSERT(!hmeblkp->hblk_shared);
12410 
12411 	/*
12412 	 * If the process is exiting we have nothing to do.
12413 	 */
12414 	if (tlb_noflush)
12415 		return;
12416 
12417 	/*
12418 	 * Flush TSB.
12419 	 */
12420 	if (!hat_lock_held)
12421 		hatlockp = sfmmu_hat_enter(sfmmup);
12422 	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12423 
12424 	kpreempt_disable();
12425 
12426 	cpuset = sfmmup->sfmmu_cpusran;
12427 	CPUSET_AND(cpuset, cpu_ready_set);
12428 	CPUSET_DEL(cpuset, CPU->cpu_id);
12429 
12430 	SFMMU_XCALL_STATS(sfmmup);
12431 	xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, (uint64_t)sfmmup);
12432 
12433 	vtag_flushpage(addr, (uint64_t)sfmmup);
12434 
12435 	if (!hat_lock_held)
12436 		sfmmu_hat_exit(hatlockp);
12437 
12438 	kpreempt_enable();
12439 
12440 }
12441 
12442 /*
12443  * Special case of sfmmu_tlb_demap for MMU_PAGESIZE hblks. Use the xcall
12444  * call handler that can flush a range of pages to save on xcalls.
12445  */
12446 static int sfmmu_xcall_save;
12447 
12448 /*
12449  * this routine is never used for demaping addresses backed by SRD hmeblks.
12450  */
12451 static void
12452 sfmmu_tlb_range_demap(demap_range_t *dmrp)
12453 {
12454 	sfmmu_t *sfmmup = dmrp->dmr_sfmmup;
12455 	hatlock_t *hatlockp;
12456 	cpuset_t cpuset;
12457 	uint64_t sfmmu_pgcnt;
12458 	pgcnt_t pgcnt = 0;
12459 	int pgunload = 0;
12460 	int dirtypg = 0;
12461 	caddr_t addr = dmrp->dmr_addr;
12462 	caddr_t eaddr;
12463 	uint64_t bitvec = dmrp->dmr_bitvec;
12464 
12465 	ASSERT(bitvec & 1);
12466 
12467 	/*
12468 	 * Flush TSB and calculate number of pages to flush.
12469 	 */
12470 	while (bitvec != 0) {
12471 		dirtypg = 0;
12472 		/*
12473 		 * Find the first page to flush and then count how many
12474 		 * pages there are after it that also need to be flushed.
12475 		 * This way the number of TSB flushes is minimized.
12476 		 */
12477 		while ((bitvec & 1) == 0) {
12478 			pgcnt++;
12479 			addr += MMU_PAGESIZE;
12480 			bitvec >>= 1;
12481 		}
12482 		while (bitvec & 1) {
12483 			dirtypg++;
12484 			bitvec >>= 1;
12485 		}
12486 		eaddr = addr + ptob(dirtypg);
12487 		hatlockp = sfmmu_hat_enter(sfmmup);
12488 		sfmmu_unload_tsb_range(sfmmup, addr, eaddr, TTE8K);
12489 		sfmmu_hat_exit(hatlockp);
12490 		pgunload += dirtypg;
12491 		addr = eaddr;
12492 		pgcnt += dirtypg;
12493 	}
12494 
12495 	ASSERT((pgcnt<<MMU_PAGESHIFT) <= dmrp->dmr_endaddr - dmrp->dmr_addr);
12496 	if (sfmmup->sfmmu_free == 0) {
12497 		addr = dmrp->dmr_addr;
12498 		bitvec = dmrp->dmr_bitvec;
12499 
12500 		/*
12501 		 * make sure it has SFMMU_PGCNT_SHIFT bits only,
12502 		 * as it will be used to pack argument for xt_some
12503 		 */
12504 		ASSERT((pgcnt > 0) &&
12505 		    (pgcnt <= (1 << SFMMU_PGCNT_SHIFT)));
12506 
12507 		/*
12508 		 * Encode pgcnt as (pgcnt -1 ), and pass (pgcnt - 1) in
12509 		 * the low 6 bits of sfmmup. This is doable since pgcnt
12510 		 * always >= 1.
12511 		 */
12512 		ASSERT(!((uint64_t)sfmmup & SFMMU_PGCNT_MASK));
12513 		sfmmu_pgcnt = (uint64_t)sfmmup |
12514 		    ((pgcnt - 1) & SFMMU_PGCNT_MASK);
12515 
12516 		/*
12517 		 * We must hold the hat lock during the flush of TLB,
12518 		 * to avoid a race with sfmmu_invalidate_ctx(), where
12519 		 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12520 		 * causing TLB demap routine to skip flush on that MMU.
12521 		 * If the context on a MMU has already been set to
12522 		 * INVALID_CONTEXT, we just get an extra flush on
12523 		 * that MMU.
12524 		 */
12525 		hatlockp = sfmmu_hat_enter(sfmmup);
12526 		kpreempt_disable();
12527 
12528 		cpuset = sfmmup->sfmmu_cpusran;
12529 		CPUSET_AND(cpuset, cpu_ready_set);
12530 		CPUSET_DEL(cpuset, CPU->cpu_id);
12531 
12532 		SFMMU_XCALL_STATS(sfmmup);
12533 		xt_some(cpuset, vtag_flush_pgcnt_tl1, (uint64_t)addr,
12534 		    sfmmu_pgcnt);
12535 
12536 		for (; bitvec != 0; bitvec >>= 1) {
12537 			if (bitvec & 1)
12538 				vtag_flushpage(addr, (uint64_t)sfmmup);
12539 			addr += MMU_PAGESIZE;
12540 		}
12541 		kpreempt_enable();
12542 		sfmmu_hat_exit(hatlockp);
12543 
12544 		sfmmu_xcall_save += (pgunload-1);
12545 	}
12546 	dmrp->dmr_bitvec = 0;
12547 }
12548 
12549 /*
12550  * In cases where we need to synchronize with TLB/TSB miss trap
12551  * handlers, _and_ need to flush the TLB, it's a lot easier to
12552  * throw away the context from the process than to do a
12553  * special song and dance to keep things consistent for the
12554  * handlers.
12555  *
12556  * Since the process suddenly ends up without a context and our caller
12557  * holds the hat lock, threads that fault after this function is called
12558  * will pile up on the lock.  We can then do whatever we need to
12559  * atomically from the context of the caller.  The first blocked thread
12560  * to resume executing will get the process a new context, and the
12561  * process will resume executing.
12562  *
12563  * One added advantage of this approach is that on MMUs that
12564  * support a "flush all" operation, we will delay the flush until
12565  * cnum wrap-around, and then flush the TLB one time.  This
12566  * is rather rare, so it's a lot less expensive than making 8000
12567  * x-calls to flush the TLB 8000 times.
12568  *
12569  * A per-process (PP) lock is used to synchronize ctx allocations in
12570  * resume() and ctx invalidations here.
12571  */
12572 static void
12573 sfmmu_invalidate_ctx(sfmmu_t *sfmmup)
12574 {
12575 	cpuset_t cpuset;
12576 	int cnum, currcnum;
12577 	mmu_ctx_t *mmu_ctxp;
12578 	int i;
12579 	uint_t pstate_save;
12580 
12581 	SFMMU_STAT(sf_ctx_inv);
12582 
12583 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12584 	ASSERT(sfmmup != ksfmmup);
12585 
12586 	kpreempt_disable();
12587 
12588 	mmu_ctxp = CPU_MMU_CTXP(CPU);
12589 	ASSERT(mmu_ctxp);
12590 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
12591 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
12592 
12593 	currcnum = sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum;
12594 
12595 	pstate_save = sfmmu_disable_intrs();
12596 
12597 	lock_set(&sfmmup->sfmmu_ctx_lock);	/* acquire PP lock */
12598 	/* set HAT cnum invalid across all context domains. */
12599 	for (i = 0; i < max_mmu_ctxdoms; i++) {
12600 
12601 		cnum = 	sfmmup->sfmmu_ctxs[i].cnum;
12602 		if (cnum == INVALID_CONTEXT) {
12603 			continue;
12604 		}
12605 
12606 		sfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
12607 	}
12608 	membar_enter();	/* make sure globally visible to all CPUs */
12609 	lock_clear(&sfmmup->sfmmu_ctx_lock);	/* release PP lock */
12610 
12611 	sfmmu_enable_intrs(pstate_save);
12612 
12613 	cpuset = sfmmup->sfmmu_cpusran;
12614 	CPUSET_DEL(cpuset, CPU->cpu_id);
12615 	CPUSET_AND(cpuset, cpu_ready_set);
12616 	if (!CPUSET_ISNULL(cpuset)) {
12617 		SFMMU_XCALL_STATS(sfmmup);
12618 		xt_some(cpuset, sfmmu_raise_tsb_exception,
12619 		    (uint64_t)sfmmup, INVALID_CONTEXT);
12620 		xt_sync(cpuset);
12621 		SFMMU_STAT(sf_tsb_raise_exception);
12622 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
12623 	}
12624 
12625 	/*
12626 	 * If the hat to-be-invalidated is the same as the current
12627 	 * process on local CPU we need to invalidate
12628 	 * this CPU context as well.
12629 	 */
12630 	if ((sfmmu_getctx_sec() == currcnum) &&
12631 	    (currcnum != INVALID_CONTEXT)) {
12632 		/* sets shared context to INVALID too */
12633 		sfmmu_setctx_sec(INVALID_CONTEXT);
12634 		sfmmu_clear_utsbinfo();
12635 	}
12636 
12637 	SFMMU_FLAGS_SET(sfmmup, HAT_ALLCTX_INVALID);
12638 
12639 	kpreempt_enable();
12640 
12641 	/*
12642 	 * we hold the hat lock, so nobody should allocate a context
12643 	 * for us yet
12644 	 */
12645 	ASSERT(sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum == INVALID_CONTEXT);
12646 }
12647 
12648 #ifdef VAC
12649 /*
12650  * We need to flush the cache in all cpus.  It is possible that
12651  * a process referenced a page as cacheable but has sinced exited
12652  * and cleared the mapping list.  We still to flush it but have no
12653  * state so all cpus is the only alternative.
12654  */
12655 void
12656 sfmmu_cache_flush(pfn_t pfnum, int vcolor)
12657 {
12658 	cpuset_t cpuset;
12659 
12660 	kpreempt_disable();
12661 	cpuset = cpu_ready_set;
12662 	CPUSET_DEL(cpuset, CPU->cpu_id);
12663 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12664 	xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12665 	xt_sync(cpuset);
12666 	vac_flushpage(pfnum, vcolor);
12667 	kpreempt_enable();
12668 }
12669 
12670 void
12671 sfmmu_cache_flushcolor(int vcolor, pfn_t pfnum)
12672 {
12673 	cpuset_t cpuset;
12674 
12675 	ASSERT(vcolor >= 0);
12676 
12677 	kpreempt_disable();
12678 	cpuset = cpu_ready_set;
12679 	CPUSET_DEL(cpuset, CPU->cpu_id);
12680 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12681 	xt_some(cpuset, vac_flushcolor_tl1, vcolor, pfnum);
12682 	xt_sync(cpuset);
12683 	vac_flushcolor(vcolor, pfnum);
12684 	kpreempt_enable();
12685 }
12686 #endif	/* VAC */
12687 
12688 /*
12689  * We need to prevent processes from accessing the TSB using a cached physical
12690  * address.  It's alright if they try to access the TSB via virtual address
12691  * since they will just fault on that virtual address once the mapping has
12692  * been suspended.
12693  */
12694 #pragma weak sendmondo_in_recover
12695 
12696 /* ARGSUSED */
12697 static int
12698 sfmmu_tsb_pre_relocator(caddr_t va, uint_t tsbsz, uint_t flags, void *tsbinfo)
12699 {
12700 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12701 	sfmmu_t *sfmmup = tsbinfop->tsb_sfmmu;
12702 	hatlock_t *hatlockp;
12703 	sf_scd_t *scdp;
12704 
12705 	if (flags != HAT_PRESUSPEND)
12706 		return (0);
12707 
12708 	/*
12709 	 * If tsb is a shared TSB with TSB_SHAREDCTX set, sfmmup must
12710 	 * be a shared hat, then set SCD's tsbinfo's flag.
12711 	 * If tsb is not shared, sfmmup is a private hat, then set
12712 	 * its private tsbinfo's flag.
12713 	 */
12714 	hatlockp = sfmmu_hat_enter(sfmmup);
12715 	tsbinfop->tsb_flags |= TSB_RELOC_FLAG;
12716 
12717 	if (!(tsbinfop->tsb_flags & TSB_SHAREDCTX)) {
12718 		sfmmu_tsb_inv_ctx(sfmmup);
12719 		sfmmu_hat_exit(hatlockp);
12720 	} else {
12721 		/* release lock on the shared hat */
12722 		sfmmu_hat_exit(hatlockp);
12723 		/* sfmmup is a shared hat */
12724 		ASSERT(sfmmup->sfmmu_scdhat);
12725 		scdp = sfmmup->sfmmu_scdp;
12726 		ASSERT(scdp != NULL);
12727 		/* get private hat from the scd list */
12728 		mutex_enter(&scdp->scd_mutex);
12729 		sfmmup = scdp->scd_sf_list;
12730 		while (sfmmup != NULL) {
12731 			hatlockp = sfmmu_hat_enter(sfmmup);
12732 			/*
12733 			 * We do not call sfmmu_tsb_inv_ctx here because
12734 			 * sendmondo_in_recover check is only needed for
12735 			 * sun4u.
12736 			 */
12737 			sfmmu_invalidate_ctx(sfmmup);
12738 			sfmmu_hat_exit(hatlockp);
12739 			sfmmup = sfmmup->sfmmu_scd_link.next;
12740 
12741 		}
12742 		mutex_exit(&scdp->scd_mutex);
12743 	}
12744 	return (0);
12745 }
12746 
12747 static void
12748 sfmmu_tsb_inv_ctx(sfmmu_t *sfmmup)
12749 {
12750 	extern uint32_t sendmondo_in_recover;
12751 
12752 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12753 
12754 	/*
12755 	 * For Cheetah+ Erratum 25:
12756 	 * Wait for any active recovery to finish.  We can't risk
12757 	 * relocating the TSB of the thread running mondo_recover_proc()
12758 	 * since, if we did that, we would deadlock.  The scenario we are
12759 	 * trying to avoid is as follows:
12760 	 *
12761 	 * THIS CPU			RECOVER CPU
12762 	 * --------			-----------
12763 	 *				Begins recovery, walking through TSB
12764 	 * hat_pagesuspend() TSB TTE
12765 	 *				TLB miss on TSB TTE, spins at TL1
12766 	 * xt_sync()
12767 	 *	send_mondo_timeout()
12768 	 *	mondo_recover_proc()
12769 	 *	((deadlocked))
12770 	 *
12771 	 * The second half of the workaround is that mondo_recover_proc()
12772 	 * checks to see if the tsb_info has the RELOC flag set, and if it
12773 	 * does, it skips over that TSB without ever touching tsbinfop->tsb_va
12774 	 * and hence avoiding the TLB miss that could result in a deadlock.
12775 	 */
12776 	if (&sendmondo_in_recover) {
12777 		membar_enter();	/* make sure RELOC flag visible */
12778 		while (sendmondo_in_recover) {
12779 			drv_usecwait(1);
12780 			membar_consumer();
12781 		}
12782 	}
12783 
12784 	sfmmu_invalidate_ctx(sfmmup);
12785 }
12786 
12787 /* ARGSUSED */
12788 static int
12789 sfmmu_tsb_post_relocator(caddr_t va, uint_t tsbsz, uint_t flags,
12790 	void *tsbinfo, pfn_t newpfn)
12791 {
12792 	hatlock_t *hatlockp;
12793 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12794 	sfmmu_t	*sfmmup = tsbinfop->tsb_sfmmu;
12795 
12796 	if (flags != HAT_POSTUNSUSPEND)
12797 		return (0);
12798 
12799 	hatlockp = sfmmu_hat_enter(sfmmup);
12800 
12801 	SFMMU_STAT(sf_tsb_reloc);
12802 
12803 	/*
12804 	 * The process may have swapped out while we were relocating one
12805 	 * of its TSBs.  If so, don't bother doing the setup since the
12806 	 * process can't be using the memory anymore.
12807 	 */
12808 	if ((tsbinfop->tsb_flags & TSB_SWAPPED) == 0) {
12809 		ASSERT(va == tsbinfop->tsb_va);
12810 		sfmmu_tsbinfo_setup_phys(tsbinfop, newpfn);
12811 
12812 		if (tsbinfop->tsb_flags & TSB_FLUSH_NEEDED) {
12813 			sfmmu_inv_tsb(tsbinfop->tsb_va,
12814 			    TSB_BYTES(tsbinfop->tsb_szc));
12815 			tsbinfop->tsb_flags &= ~TSB_FLUSH_NEEDED;
12816 		}
12817 	}
12818 
12819 	membar_exit();
12820 	tsbinfop->tsb_flags &= ~TSB_RELOC_FLAG;
12821 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
12822 
12823 	sfmmu_hat_exit(hatlockp);
12824 
12825 	return (0);
12826 }
12827 
12828 /*
12829  * Allocate and initialize a tsb_info structure.  Note that we may or may not
12830  * allocate a TSB here, depending on the flags passed in.
12831  */
12832 static int
12833 sfmmu_tsbinfo_alloc(struct tsb_info **tsbinfopp, int tsb_szc, int tte_sz_mask,
12834 	uint_t flags, sfmmu_t *sfmmup)
12835 {
12836 	int err;
12837 
12838 	*tsbinfopp = (struct tsb_info *)kmem_cache_alloc(
12839 	    sfmmu_tsbinfo_cache, KM_SLEEP);
12840 
12841 	if ((err = sfmmu_init_tsbinfo(*tsbinfopp, tte_sz_mask,
12842 	    tsb_szc, flags, sfmmup)) != 0) {
12843 		kmem_cache_free(sfmmu_tsbinfo_cache, *tsbinfopp);
12844 		SFMMU_STAT(sf_tsb_allocfail);
12845 		*tsbinfopp = NULL;
12846 		return (err);
12847 	}
12848 	SFMMU_STAT(sf_tsb_alloc);
12849 
12850 	/*
12851 	 * Bump the TSB size counters for this TSB size.
12852 	 */
12853 	(*(((int *)&sfmmu_tsbsize_stat) + tsb_szc))++;
12854 	return (0);
12855 }
12856 
12857 static void
12858 sfmmu_tsb_free(struct tsb_info *tsbinfo)
12859 {
12860 	caddr_t tsbva = tsbinfo->tsb_va;
12861 	uint_t tsb_size = TSB_BYTES(tsbinfo->tsb_szc);
12862 	struct kmem_cache *kmem_cachep = tsbinfo->tsb_cache;
12863 	vmem_t	*vmp = tsbinfo->tsb_vmp;
12864 
12865 	/*
12866 	 * If we allocated this TSB from relocatable kernel memory, then we
12867 	 * need to uninstall the callback handler.
12868 	 */
12869 	if (tsbinfo->tsb_cache != sfmmu_tsb8k_cache) {
12870 		uintptr_t slab_mask;
12871 		caddr_t slab_vaddr;
12872 		page_t **ppl;
12873 		int ret;
12874 
12875 		ASSERT(tsb_size <= MMU_PAGESIZE4M || use_bigtsb_arena);
12876 		if (tsb_size > MMU_PAGESIZE4M)
12877 			slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
12878 		else
12879 			slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
12880 		slab_vaddr = (caddr_t)((uintptr_t)tsbva & slab_mask);
12881 
12882 		ret = as_pagelock(&kas, &ppl, slab_vaddr, PAGESIZE, S_WRITE);
12883 		ASSERT(ret == 0);
12884 		hat_delete_callback(tsbva, (uint_t)tsb_size, (void *)tsbinfo,
12885 		    0, NULL);
12886 		as_pageunlock(&kas, ppl, slab_vaddr, PAGESIZE, S_WRITE);
12887 	}
12888 
12889 	if (kmem_cachep != NULL) {
12890 		kmem_cache_free(kmem_cachep, tsbva);
12891 	} else {
12892 		vmem_xfree(vmp, (void *)tsbva, tsb_size);
12893 	}
12894 	tsbinfo->tsb_va = (caddr_t)0xbad00bad;
12895 	atomic_add_64(&tsb_alloc_bytes, -(int64_t)tsb_size);
12896 }
12897 
12898 static void
12899 sfmmu_tsbinfo_free(struct tsb_info *tsbinfo)
12900 {
12901 	if ((tsbinfo->tsb_flags & TSB_SWAPPED) == 0) {
12902 		sfmmu_tsb_free(tsbinfo);
12903 	}
12904 	kmem_cache_free(sfmmu_tsbinfo_cache, tsbinfo);
12905 
12906 }
12907 
12908 /*
12909  * Setup all the references to physical memory for this tsbinfo.
12910  * The underlying page(s) must be locked.
12911  */
12912 static void
12913 sfmmu_tsbinfo_setup_phys(struct tsb_info *tsbinfo, pfn_t pfn)
12914 {
12915 	ASSERT(pfn != PFN_INVALID);
12916 	ASSERT(pfn == va_to_pfn(tsbinfo->tsb_va));
12917 
12918 #ifndef sun4v
12919 	if (tsbinfo->tsb_szc == 0) {
12920 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn,
12921 		    PROT_WRITE|PROT_READ, TTE8K);
12922 	} else {
12923 		/*
12924 		 * Round down PA and use a large mapping; the handlers will
12925 		 * compute the TSB pointer at the correct offset into the
12926 		 * big virtual page.  NOTE: this assumes all TSBs larger
12927 		 * than 8K must come from physically contiguous slabs of
12928 		 * size tsb_slab_size.
12929 		 */
12930 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn & ~tsb_slab_mask,
12931 		    PROT_WRITE|PROT_READ, tsb_slab_ttesz);
12932 	}
12933 	tsbinfo->tsb_pa = ptob(pfn);
12934 
12935 	TTE_SET_LOCKED(&tsbinfo->tsb_tte); /* lock the tte into dtlb */
12936 	TTE_SET_MOD(&tsbinfo->tsb_tte);    /* enable writes */
12937 
12938 	ASSERT(TTE_IS_PRIVILEGED(&tsbinfo->tsb_tte));
12939 	ASSERT(TTE_IS_LOCKED(&tsbinfo->tsb_tte));
12940 #else /* sun4v */
12941 	tsbinfo->tsb_pa = ptob(pfn);
12942 #endif /* sun4v */
12943 }
12944 
12945 
12946 /*
12947  * Returns zero on success, ENOMEM if over the high water mark,
12948  * or EAGAIN if the caller needs to retry with a smaller TSB
12949  * size (or specify TSB_FORCEALLOC if the allocation can't fail).
12950  *
12951  * This call cannot fail to allocate a TSB if TSB_FORCEALLOC
12952  * is specified and the TSB requested is PAGESIZE, though it
12953  * may sleep waiting for memory if sufficient memory is not
12954  * available.
12955  */
12956 static int
12957 sfmmu_init_tsbinfo(struct tsb_info *tsbinfo, int tteszmask,
12958     int tsbcode, uint_t flags, sfmmu_t *sfmmup)
12959 {
12960 	caddr_t vaddr = NULL;
12961 	caddr_t slab_vaddr;
12962 	uintptr_t slab_mask;
12963 	int tsbbytes = TSB_BYTES(tsbcode);
12964 	int lowmem = 0;
12965 	struct kmem_cache *kmem_cachep = NULL;
12966 	vmem_t *vmp = NULL;
12967 	lgrp_id_t lgrpid = LGRP_NONE;
12968 	pfn_t pfn;
12969 	uint_t cbflags = HAC_SLEEP;
12970 	page_t **pplist;
12971 	int ret;
12972 
12973 	ASSERT(tsbbytes <= MMU_PAGESIZE4M || use_bigtsb_arena);
12974 	if (tsbbytes > MMU_PAGESIZE4M)
12975 		slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
12976 	else
12977 		slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
12978 
12979 	if (flags & (TSB_FORCEALLOC | TSB_SWAPIN | TSB_GROW | TSB_SHRINK))
12980 		flags |= TSB_ALLOC;
12981 
12982 	ASSERT((flags & TSB_FORCEALLOC) == 0 || tsbcode == TSB_MIN_SZCODE);
12983 
12984 	tsbinfo->tsb_sfmmu = sfmmup;
12985 
12986 	/*
12987 	 * If not allocating a TSB, set up the tsbinfo, set TSB_SWAPPED, and
12988 	 * return.
12989 	 */
12990 	if ((flags & TSB_ALLOC) == 0) {
12991 		tsbinfo->tsb_szc = tsbcode;
12992 		tsbinfo->tsb_ttesz_mask = tteszmask;
12993 		tsbinfo->tsb_va = (caddr_t)0xbadbadbeef;
12994 		tsbinfo->tsb_pa = -1;
12995 		tsbinfo->tsb_tte.ll = 0;
12996 		tsbinfo->tsb_next = NULL;
12997 		tsbinfo->tsb_flags = TSB_SWAPPED;
12998 		tsbinfo->tsb_cache = NULL;
12999 		tsbinfo->tsb_vmp = NULL;
13000 		return (0);
13001 	}
13002 
13003 #ifdef DEBUG
13004 	/*
13005 	 * For debugging:
13006 	 * Randomly force allocation failures every tsb_alloc_mtbf
13007 	 * tries if TSB_FORCEALLOC is not specified.  This will
13008 	 * return ENOMEM if tsb_alloc_mtbf is odd, or EAGAIN if
13009 	 * it is even, to allow testing of both failure paths...
13010 	 */
13011 	if (tsb_alloc_mtbf && ((flags & TSB_FORCEALLOC) == 0) &&
13012 	    (tsb_alloc_count++ == tsb_alloc_mtbf)) {
13013 		tsb_alloc_count = 0;
13014 		tsb_alloc_fail_mtbf++;
13015 		return ((tsb_alloc_mtbf & 1)? ENOMEM : EAGAIN);
13016 	}
13017 #endif	/* DEBUG */
13018 
13019 	/*
13020 	 * Enforce high water mark if we are not doing a forced allocation
13021 	 * and are not shrinking a process' TSB.
13022 	 */
13023 	if ((flags & TSB_SHRINK) == 0 &&
13024 	    (tsbbytes + tsb_alloc_bytes) > tsb_alloc_hiwater) {
13025 		if ((flags & TSB_FORCEALLOC) == 0)
13026 			return (ENOMEM);
13027 		lowmem = 1;
13028 	}
13029 
13030 	/*
13031 	 * Allocate from the correct location based upon the size of the TSB
13032 	 * compared to the base page size, and what memory conditions dictate.
13033 	 * Note we always do nonblocking allocations from the TSB arena since
13034 	 * we don't want memory fragmentation to cause processes to block
13035 	 * indefinitely waiting for memory; until the kernel algorithms that
13036 	 * coalesce large pages are improved this is our best option.
13037 	 *
13038 	 * Algorithm:
13039 	 *	If allocating a "large" TSB (>8K), allocate from the
13040 	 *		appropriate kmem_tsb_default_arena vmem arena
13041 	 *	else if low on memory or the TSB_FORCEALLOC flag is set or
13042 	 *	tsb_forceheap is set
13043 	 *		Allocate from kernel heap via sfmmu_tsb8k_cache with
13044 	 *		KM_SLEEP (never fails)
13045 	 *	else
13046 	 *		Allocate from appropriate sfmmu_tsb_cache with
13047 	 *		KM_NOSLEEP
13048 	 *	endif
13049 	 */
13050 	if (tsb_lgrp_affinity)
13051 		lgrpid = lgrp_home_id(curthread);
13052 	if (lgrpid == LGRP_NONE)
13053 		lgrpid = 0;	/* use lgrp of boot CPU */
13054 
13055 	if (tsbbytes > MMU_PAGESIZE) {
13056 		if (tsbbytes > MMU_PAGESIZE4M) {
13057 			vmp = kmem_bigtsb_default_arena[lgrpid];
13058 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
13059 			    0, 0, NULL, NULL, VM_NOSLEEP);
13060 		} else {
13061 			vmp = kmem_tsb_default_arena[lgrpid];
13062 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
13063 			    0, 0, NULL, NULL, VM_NOSLEEP);
13064 		}
13065 #ifdef	DEBUG
13066 	} else if (lowmem || (flags & TSB_FORCEALLOC) || tsb_forceheap) {
13067 #else	/* !DEBUG */
13068 	} else if (lowmem || (flags & TSB_FORCEALLOC)) {
13069 #endif	/* DEBUG */
13070 		kmem_cachep = sfmmu_tsb8k_cache;
13071 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_SLEEP);
13072 		ASSERT(vaddr != NULL);
13073 	} else {
13074 		kmem_cachep = sfmmu_tsb_cache[lgrpid];
13075 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_NOSLEEP);
13076 	}
13077 
13078 	tsbinfo->tsb_cache = kmem_cachep;
13079 	tsbinfo->tsb_vmp = vmp;
13080 
13081 	if (vaddr == NULL) {
13082 		return (EAGAIN);
13083 	}
13084 
13085 	atomic_add_64(&tsb_alloc_bytes, (int64_t)tsbbytes);
13086 	kmem_cachep = tsbinfo->tsb_cache;
13087 
13088 	/*
13089 	 * If we are allocating from outside the cage, then we need to
13090 	 * register a relocation callback handler.  Note that for now
13091 	 * since pseudo mappings always hang off of the slab's root page,
13092 	 * we need only lock the first 8K of the TSB slab.  This is a bit
13093 	 * hacky but it is good for performance.
13094 	 */
13095 	if (kmem_cachep != sfmmu_tsb8k_cache) {
13096 		slab_vaddr = (caddr_t)((uintptr_t)vaddr & slab_mask);
13097 		ret = as_pagelock(&kas, &pplist, slab_vaddr, PAGESIZE, S_WRITE);
13098 		ASSERT(ret == 0);
13099 		ret = hat_add_callback(sfmmu_tsb_cb_id, vaddr, (uint_t)tsbbytes,
13100 		    cbflags, (void *)tsbinfo, &pfn, NULL);
13101 
13102 		/*
13103 		 * Need to free up resources if we could not successfully
13104 		 * add the callback function and return an error condition.
13105 		 */
13106 		if (ret != 0) {
13107 			if (kmem_cachep) {
13108 				kmem_cache_free(kmem_cachep, vaddr);
13109 			} else {
13110 				vmem_xfree(vmp, (void *)vaddr, tsbbytes);
13111 			}
13112 			as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE,
13113 			    S_WRITE);
13114 			return (EAGAIN);
13115 		}
13116 	} else {
13117 		/*
13118 		 * Since allocation of 8K TSBs from heap is rare and occurs
13119 		 * during memory pressure we allocate them from permanent
13120 		 * memory rather than using callbacks to get the PFN.
13121 		 */
13122 		pfn = hat_getpfnum(kas.a_hat, vaddr);
13123 	}
13124 
13125 	tsbinfo->tsb_va = vaddr;
13126 	tsbinfo->tsb_szc = tsbcode;
13127 	tsbinfo->tsb_ttesz_mask = tteszmask;
13128 	tsbinfo->tsb_next = NULL;
13129 	tsbinfo->tsb_flags = 0;
13130 
13131 	sfmmu_tsbinfo_setup_phys(tsbinfo, pfn);
13132 
13133 	sfmmu_inv_tsb(vaddr, tsbbytes);
13134 
13135 	if (kmem_cachep != sfmmu_tsb8k_cache) {
13136 		as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE, S_WRITE);
13137 	}
13138 
13139 	return (0);
13140 }
13141 
13142 /*
13143  * Initialize per cpu tsb and per cpu tsbmiss_area
13144  */
13145 void
13146 sfmmu_init_tsbs(void)
13147 {
13148 	int i;
13149 	struct tsbmiss	*tsbmissp;
13150 	struct kpmtsbm	*kpmtsbmp;
13151 #ifndef sun4v
13152 	extern int	dcache_line_mask;
13153 #endif /* sun4v */
13154 	extern uint_t	vac_colors;
13155 
13156 	/*
13157 	 * Init. tsb miss area.
13158 	 */
13159 	tsbmissp = tsbmiss_area;
13160 
13161 	for (i = 0; i < NCPU; tsbmissp++, i++) {
13162 		/*
13163 		 * initialize the tsbmiss area.
13164 		 * Do this for all possible CPUs as some may be added
13165 		 * while the system is running. There is no cost to this.
13166 		 */
13167 		tsbmissp->ksfmmup = ksfmmup;
13168 #ifndef sun4v
13169 		tsbmissp->dcache_line_mask = (uint16_t)dcache_line_mask;
13170 #endif /* sun4v */
13171 		tsbmissp->khashstart =
13172 		    (struct hmehash_bucket *)va_to_pa((caddr_t)khme_hash);
13173 		tsbmissp->uhashstart =
13174 		    (struct hmehash_bucket *)va_to_pa((caddr_t)uhme_hash);
13175 		tsbmissp->khashsz = khmehash_num;
13176 		tsbmissp->uhashsz = uhmehash_num;
13177 	}
13178 
13179 	sfmmu_tsb_cb_id = hat_register_callback('T'<<16 | 'S' << 8 | 'B',
13180 	    sfmmu_tsb_pre_relocator, sfmmu_tsb_post_relocator, NULL, 0);
13181 
13182 	if (kpm_enable == 0)
13183 		return;
13184 
13185 	/* -- Begin KPM specific init -- */
13186 
13187 	if (kpm_smallpages) {
13188 		/*
13189 		 * If we're using base pagesize pages for seg_kpm
13190 		 * mappings, we use the kernel TSB since we can't afford
13191 		 * to allocate a second huge TSB for these mappings.
13192 		 */
13193 		kpm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13194 		kpm_tsbsz = ktsb_szcode;
13195 		kpmsm_tsbbase = kpm_tsbbase;
13196 		kpmsm_tsbsz = kpm_tsbsz;
13197 	} else {
13198 		/*
13199 		 * In VAC conflict case, just put the entries in the
13200 		 * kernel 8K indexed TSB for now so we can find them.
13201 		 * This could really be changed in the future if we feel
13202 		 * the need...
13203 		 */
13204 		kpmsm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13205 		kpmsm_tsbsz = ktsb_szcode;
13206 		kpm_tsbbase = ktsb_phys? ktsb4m_pbase : (uint64_t)ktsb4m_base;
13207 		kpm_tsbsz = ktsb4m_szcode;
13208 	}
13209 
13210 	kpmtsbmp = kpmtsbm_area;
13211 	for (i = 0; i < NCPU; kpmtsbmp++, i++) {
13212 		/*
13213 		 * Initialize the kpmtsbm area.
13214 		 * Do this for all possible CPUs as some may be added
13215 		 * while the system is running. There is no cost to this.
13216 		 */
13217 		kpmtsbmp->vbase = kpm_vbase;
13218 		kpmtsbmp->vend = kpm_vbase + kpm_size * vac_colors;
13219 		kpmtsbmp->sz_shift = kpm_size_shift;
13220 		kpmtsbmp->kpmp_shift = kpmp_shift;
13221 		kpmtsbmp->kpmp2pshft = (uchar_t)kpmp2pshft;
13222 		if (kpm_smallpages == 0) {
13223 			kpmtsbmp->kpmp_table_sz = kpmp_table_sz;
13224 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_table);
13225 		} else {
13226 			kpmtsbmp->kpmp_table_sz = kpmp_stable_sz;
13227 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_stable);
13228 		}
13229 		kpmtsbmp->msegphashpa = va_to_pa(memseg_phash);
13230 		kpmtsbmp->flags = KPMTSBM_ENABLE_FLAG;
13231 #ifdef	DEBUG
13232 		kpmtsbmp->flags |= (kpm_tsbmtl) ?  KPMTSBM_TLTSBM_FLAG : 0;
13233 #endif	/* DEBUG */
13234 		if (ktsb_phys)
13235 			kpmtsbmp->flags |= KPMTSBM_TSBPHYS_FLAG;
13236 	}
13237 
13238 	/* -- End KPM specific init -- */
13239 }
13240 
13241 /* Avoid using sfmmu_tsbinfo_alloc() to avoid kmem_alloc - no real reason */
13242 struct tsb_info ktsb_info[2];
13243 
13244 /*
13245  * Called from hat_kern_setup() to setup the tsb_info for ksfmmup.
13246  */
13247 void
13248 sfmmu_init_ktsbinfo()
13249 {
13250 	ASSERT(ksfmmup != NULL);
13251 	ASSERT(ksfmmup->sfmmu_tsb == NULL);
13252 	/*
13253 	 * Allocate tsbinfos for kernel and copy in data
13254 	 * to make debug easier and sun4v setup easier.
13255 	 */
13256 	ktsb_info[0].tsb_sfmmu = ksfmmup;
13257 	ktsb_info[0].tsb_szc = ktsb_szcode;
13258 	ktsb_info[0].tsb_ttesz_mask = TSB8K|TSB64K|TSB512K;
13259 	ktsb_info[0].tsb_va = ktsb_base;
13260 	ktsb_info[0].tsb_pa = ktsb_pbase;
13261 	ktsb_info[0].tsb_flags = 0;
13262 	ktsb_info[0].tsb_tte.ll = 0;
13263 	ktsb_info[0].tsb_cache = NULL;
13264 
13265 	ktsb_info[1].tsb_sfmmu = ksfmmup;
13266 	ktsb_info[1].tsb_szc = ktsb4m_szcode;
13267 	ktsb_info[1].tsb_ttesz_mask = TSB4M;
13268 	ktsb_info[1].tsb_va = ktsb4m_base;
13269 	ktsb_info[1].tsb_pa = ktsb4m_pbase;
13270 	ktsb_info[1].tsb_flags = 0;
13271 	ktsb_info[1].tsb_tte.ll = 0;
13272 	ktsb_info[1].tsb_cache = NULL;
13273 
13274 	/* Link them into ksfmmup. */
13275 	ktsb_info[0].tsb_next = &ktsb_info[1];
13276 	ktsb_info[1].tsb_next = NULL;
13277 	ksfmmup->sfmmu_tsb = &ktsb_info[0];
13278 
13279 	sfmmu_setup_tsbinfo(ksfmmup);
13280 }
13281 
13282 /*
13283  * Cache the last value returned from va_to_pa().  If the VA specified
13284  * in the current call to cached_va_to_pa() maps to the same Page (as the
13285  * previous call to cached_va_to_pa()), then compute the PA using
13286  * cached info, else call va_to_pa().
13287  *
13288  * Note: this function is neither MT-safe nor consistent in the presence
13289  * of multiple, interleaved threads.  This function was created to enable
13290  * an optimization used during boot (at a point when there's only one thread
13291  * executing on the "boot CPU", and before startup_vm() has been called).
13292  */
13293 static uint64_t
13294 cached_va_to_pa(void *vaddr)
13295 {
13296 	static uint64_t prev_vaddr_base = 0;
13297 	static uint64_t prev_pfn = 0;
13298 
13299 	if ((((uint64_t)vaddr) & MMU_PAGEMASK) == prev_vaddr_base) {
13300 		return (prev_pfn | ((uint64_t)vaddr & MMU_PAGEOFFSET));
13301 	} else {
13302 		uint64_t pa = va_to_pa(vaddr);
13303 
13304 		if (pa != ((uint64_t)-1)) {
13305 			/*
13306 			 * Computed physical address is valid.  Cache its
13307 			 * related info for the next cached_va_to_pa() call.
13308 			 */
13309 			prev_pfn = pa & MMU_PAGEMASK;
13310 			prev_vaddr_base = ((uint64_t)vaddr) & MMU_PAGEMASK;
13311 		}
13312 
13313 		return (pa);
13314 	}
13315 }
13316 
13317 /*
13318  * Carve up our nucleus hblk region.  We may allocate more hblks than
13319  * asked due to rounding errors but we are guaranteed to have at least
13320  * enough space to allocate the requested number of hblk8's and hblk1's.
13321  */
13322 void
13323 sfmmu_init_nucleus_hblks(caddr_t addr, size_t size, int nhblk8, int nhblk1)
13324 {
13325 	struct hme_blk *hmeblkp;
13326 	size_t hme8blk_sz, hme1blk_sz;
13327 	size_t i;
13328 	size_t hblk8_bound;
13329 	ulong_t j = 0, k = 0;
13330 
13331 	ASSERT(addr != NULL && size != 0);
13332 
13333 	/* Need to use proper structure alignment */
13334 	hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t));
13335 	hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t));
13336 
13337 	nucleus_hblk8.list = (void *)addr;
13338 	nucleus_hblk8.index = 0;
13339 
13340 	/*
13341 	 * Use as much memory as possible for hblk8's since we
13342 	 * expect all bop_alloc'ed memory to be allocated in 8k chunks.
13343 	 * We need to hold back enough space for the hblk1's which
13344 	 * we'll allocate next.
13345 	 */
13346 	hblk8_bound = size - (nhblk1 * hme1blk_sz) - hme8blk_sz;
13347 	for (i = 0; i <= hblk8_bound; i += hme8blk_sz, j++) {
13348 		hmeblkp = (struct hme_blk *)addr;
13349 		addr += hme8blk_sz;
13350 		hmeblkp->hblk_nuc_bit = 1;
13351 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13352 	}
13353 	nucleus_hblk8.len = j;
13354 	ASSERT(j >= nhblk8);
13355 	SFMMU_STAT_ADD(sf_hblk8_ncreate, j);
13356 
13357 	nucleus_hblk1.list = (void *)addr;
13358 	nucleus_hblk1.index = 0;
13359 	for (; i <= (size - hme1blk_sz); i += hme1blk_sz, k++) {
13360 		hmeblkp = (struct hme_blk *)addr;
13361 		addr += hme1blk_sz;
13362 		hmeblkp->hblk_nuc_bit = 1;
13363 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13364 	}
13365 	ASSERT(k >= nhblk1);
13366 	nucleus_hblk1.len = k;
13367 	SFMMU_STAT_ADD(sf_hblk1_ncreate, k);
13368 }
13369 
13370 /*
13371  * This function is currently not supported on this platform. For what
13372  * it's supposed to do, see hat.c and hat_srmmu.c
13373  */
13374 /* ARGSUSED */
13375 faultcode_t
13376 hat_softlock(struct hat *hat, caddr_t addr, size_t *lenp, page_t **ppp,
13377     uint_t flags)
13378 {
13379 	ASSERT(hat->sfmmu_xhat_provider == NULL);
13380 	return (FC_NOSUPPORT);
13381 }
13382 
13383 /*
13384  * Searchs the mapping list of the page for a mapping of the same size. If not
13385  * found the corresponding bit is cleared in the p_index field. When large
13386  * pages are more prevalent in the system, we can maintain the mapping list
13387  * in order and we don't have to traverse the list each time. Just check the
13388  * next and prev entries, and if both are of different size, we clear the bit.
13389  */
13390 static void
13391 sfmmu_rm_large_mappings(page_t *pp, int ttesz)
13392 {
13393 	struct sf_hment *sfhmep;
13394 	struct hme_blk *hmeblkp;
13395 	int	index;
13396 	pgcnt_t	npgs;
13397 
13398 	ASSERT(ttesz > TTE8K);
13399 
13400 	ASSERT(sfmmu_mlist_held(pp));
13401 
13402 	ASSERT(PP_ISMAPPED_LARGE(pp));
13403 
13404 	/*
13405 	 * Traverse mapping list looking for another mapping of same size.
13406 	 * since we only want to clear index field if all mappings of
13407 	 * that size are gone.
13408 	 */
13409 
13410 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
13411 		if (IS_PAHME(sfhmep))
13412 			continue;
13413 		hmeblkp = sfmmu_hmetohblk(sfhmep);
13414 		if (hmeblkp->hblk_xhat_bit)
13415 			continue;
13416 		if (hme_size(sfhmep) == ttesz) {
13417 			/*
13418 			 * another mapping of the same size. don't clear index.
13419 			 */
13420 			return;
13421 		}
13422 	}
13423 
13424 	/*
13425 	 * Clear the p_index bit for large page.
13426 	 */
13427 	index = PAGESZ_TO_INDEX(ttesz);
13428 	npgs = TTEPAGES(ttesz);
13429 	while (npgs-- > 0) {
13430 		ASSERT(pp->p_index & index);
13431 		pp->p_index &= ~index;
13432 		pp = PP_PAGENEXT(pp);
13433 	}
13434 }
13435 
13436 /*
13437  * return supported features
13438  */
13439 /* ARGSUSED */
13440 int
13441 hat_supported(enum hat_features feature, void *arg)
13442 {
13443 	switch (feature) {
13444 	case    HAT_SHARED_PT:
13445 	case	HAT_DYNAMIC_ISM_UNMAP:
13446 	case	HAT_VMODSORT:
13447 		return (1);
13448 	case	HAT_SHARED_REGIONS:
13449 		if (shctx_on)
13450 			return (1);
13451 		else
13452 			return (0);
13453 	default:
13454 		return (0);
13455 	}
13456 }
13457 
13458 void
13459 hat_enter(struct hat *hat)
13460 {
13461 	hatlock_t	*hatlockp;
13462 
13463 	if (hat != ksfmmup) {
13464 		hatlockp = TSB_HASH(hat);
13465 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
13466 	}
13467 }
13468 
13469 void
13470 hat_exit(struct hat *hat)
13471 {
13472 	hatlock_t	*hatlockp;
13473 
13474 	if (hat != ksfmmup) {
13475 		hatlockp = TSB_HASH(hat);
13476 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
13477 	}
13478 }
13479 
13480 /*ARGSUSED*/
13481 void
13482 hat_reserve(struct as *as, caddr_t addr, size_t len)
13483 {
13484 }
13485 
13486 static void
13487 hat_kstat_init(void)
13488 {
13489 	kstat_t *ksp;
13490 
13491 	ksp = kstat_create("unix", 0, "sfmmu_global_stat", "hat",
13492 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_global_stat),
13493 	    KSTAT_FLAG_VIRTUAL);
13494 	if (ksp) {
13495 		ksp->ks_data = (void *) &sfmmu_global_stat;
13496 		kstat_install(ksp);
13497 	}
13498 	ksp = kstat_create("unix", 0, "sfmmu_tsbsize_stat", "hat",
13499 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_tsbsize_stat),
13500 	    KSTAT_FLAG_VIRTUAL);
13501 	if (ksp) {
13502 		ksp->ks_data = (void *) &sfmmu_tsbsize_stat;
13503 		kstat_install(ksp);
13504 	}
13505 	ksp = kstat_create("unix", 0, "sfmmu_percpu_stat", "hat",
13506 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_percpu_stat) * NCPU,
13507 	    KSTAT_FLAG_WRITABLE);
13508 	if (ksp) {
13509 		ksp->ks_update = sfmmu_kstat_percpu_update;
13510 		kstat_install(ksp);
13511 	}
13512 }
13513 
13514 /* ARGSUSED */
13515 static int
13516 sfmmu_kstat_percpu_update(kstat_t *ksp, int rw)
13517 {
13518 	struct sfmmu_percpu_stat *cpu_kstat = ksp->ks_data;
13519 	struct tsbmiss *tsbm = tsbmiss_area;
13520 	struct kpmtsbm *kpmtsbm = kpmtsbm_area;
13521 	int i;
13522 
13523 	ASSERT(cpu_kstat);
13524 	if (rw == KSTAT_READ) {
13525 		for (i = 0; i < NCPU; cpu_kstat++, tsbm++, kpmtsbm++, i++) {
13526 			cpu_kstat->sf_itlb_misses = 0;
13527 			cpu_kstat->sf_dtlb_misses = 0;
13528 			cpu_kstat->sf_utsb_misses = tsbm->utsb_misses -
13529 			    tsbm->uprot_traps;
13530 			cpu_kstat->sf_ktsb_misses = tsbm->ktsb_misses +
13531 			    kpmtsbm->kpm_tsb_misses - tsbm->kprot_traps;
13532 			cpu_kstat->sf_tsb_hits = 0;
13533 			cpu_kstat->sf_umod_faults = tsbm->uprot_traps;
13534 			cpu_kstat->sf_kmod_faults = tsbm->kprot_traps;
13535 		}
13536 	} else {
13537 		/* KSTAT_WRITE is used to clear stats */
13538 		for (i = 0; i < NCPU; tsbm++, kpmtsbm++, i++) {
13539 			tsbm->utsb_misses = 0;
13540 			tsbm->ktsb_misses = 0;
13541 			tsbm->uprot_traps = 0;
13542 			tsbm->kprot_traps = 0;
13543 			kpmtsbm->kpm_dtlb_misses = 0;
13544 			kpmtsbm->kpm_tsb_misses = 0;
13545 		}
13546 	}
13547 	return (0);
13548 }
13549 
13550 #ifdef	DEBUG
13551 
13552 tte_t  *gorig[NCPU], *gcur[NCPU], *gnew[NCPU];
13553 
13554 /*
13555  * A tte checker. *orig_old is the value we read before cas.
13556  *	*cur is the value returned by cas.
13557  *	*new is the desired value when we do the cas.
13558  *
13559  *	*hmeblkp is currently unused.
13560  */
13561 
13562 /* ARGSUSED */
13563 void
13564 chk_tte(tte_t *orig_old, tte_t *cur, tte_t *new, struct hme_blk *hmeblkp)
13565 {
13566 	pfn_t i, j, k;
13567 	int cpuid = CPU->cpu_id;
13568 
13569 	gorig[cpuid] = orig_old;
13570 	gcur[cpuid] = cur;
13571 	gnew[cpuid] = new;
13572 
13573 #ifdef lint
13574 	hmeblkp = hmeblkp;
13575 #endif
13576 
13577 	if (TTE_IS_VALID(orig_old)) {
13578 		if (TTE_IS_VALID(cur)) {
13579 			i = TTE_TO_TTEPFN(orig_old);
13580 			j = TTE_TO_TTEPFN(cur);
13581 			k = TTE_TO_TTEPFN(new);
13582 			if (i != j) {
13583 				/* remap error? */
13584 				panic("chk_tte: bad pfn, 0x%lx, 0x%lx", i, j);
13585 			}
13586 
13587 			if (i != k) {
13588 				/* remap error? */
13589 				panic("chk_tte: bad pfn2, 0x%lx, 0x%lx", i, k);
13590 			}
13591 		} else {
13592 			if (TTE_IS_VALID(new)) {
13593 				panic("chk_tte: invalid cur? ");
13594 			}
13595 
13596 			i = TTE_TO_TTEPFN(orig_old);
13597 			k = TTE_TO_TTEPFN(new);
13598 			if (i != k) {
13599 				panic("chk_tte: bad pfn3, 0x%lx, 0x%lx", i, k);
13600 			}
13601 		}
13602 	} else {
13603 		if (TTE_IS_VALID(cur)) {
13604 			j = TTE_TO_TTEPFN(cur);
13605 			if (TTE_IS_VALID(new)) {
13606 				k = TTE_TO_TTEPFN(new);
13607 				if (j != k) {
13608 					panic("chk_tte: bad pfn4, 0x%lx, 0x%lx",
13609 					    j, k);
13610 				}
13611 			} else {
13612 				panic("chk_tte: why here?");
13613 			}
13614 		} else {
13615 			if (!TTE_IS_VALID(new)) {
13616 				panic("chk_tte: why here2 ?");
13617 			}
13618 		}
13619 	}
13620 }
13621 
13622 #endif /* DEBUG */
13623 
13624 extern void prefetch_tsbe_read(struct tsbe *);
13625 extern void prefetch_tsbe_write(struct tsbe *);
13626 
13627 
13628 /*
13629  * We want to prefetch 7 cache lines ahead for our read prefetch.  This gives
13630  * us optimal performance on Cheetah+.  You can only have 8 outstanding
13631  * prefetches at any one time, so we opted for 7 read prefetches and 1 write
13632  * prefetch to make the most utilization of the prefetch capability.
13633  */
13634 #define	TSBE_PREFETCH_STRIDE (7)
13635 
13636 void
13637 sfmmu_copy_tsb(struct tsb_info *old_tsbinfo, struct tsb_info *new_tsbinfo)
13638 {
13639 	int old_bytes = TSB_BYTES(old_tsbinfo->tsb_szc);
13640 	int new_bytes = TSB_BYTES(new_tsbinfo->tsb_szc);
13641 	int old_entries = TSB_ENTRIES(old_tsbinfo->tsb_szc);
13642 	int new_entries = TSB_ENTRIES(new_tsbinfo->tsb_szc);
13643 	struct tsbe *old;
13644 	struct tsbe *new;
13645 	struct tsbe *new_base = (struct tsbe *)new_tsbinfo->tsb_va;
13646 	uint64_t va;
13647 	int new_offset;
13648 	int i;
13649 	int vpshift;
13650 	int last_prefetch;
13651 
13652 	if (old_bytes == new_bytes) {
13653 		bcopy(old_tsbinfo->tsb_va, new_tsbinfo->tsb_va, new_bytes);
13654 	} else {
13655 
13656 		/*
13657 		 * A TSBE is 16 bytes which means there are four TSBE's per
13658 		 * P$ line (64 bytes), thus every 4 TSBE's we prefetch.
13659 		 */
13660 		old = (struct tsbe *)old_tsbinfo->tsb_va;
13661 		last_prefetch = old_entries - (4*(TSBE_PREFETCH_STRIDE+1));
13662 		for (i = 0; i < old_entries; i++, old++) {
13663 			if (((i & (4-1)) == 0) && (i < last_prefetch))
13664 				prefetch_tsbe_read(old);
13665 			if (!old->tte_tag.tag_invalid) {
13666 				/*
13667 				 * We have a valid TTE to remap.  Check the
13668 				 * size.  We won't remap 64K or 512K TTEs
13669 				 * because they span more than one TSB entry
13670 				 * and are indexed using an 8K virt. page.
13671 				 * Ditto for 32M and 256M TTEs.
13672 				 */
13673 				if (TTE_CSZ(&old->tte_data) == TTE64K ||
13674 				    TTE_CSZ(&old->tte_data) == TTE512K)
13675 					continue;
13676 				if (mmu_page_sizes == max_mmu_page_sizes) {
13677 					if (TTE_CSZ(&old->tte_data) == TTE32M ||
13678 					    TTE_CSZ(&old->tte_data) == TTE256M)
13679 						continue;
13680 				}
13681 
13682 				/* clear the lower 22 bits of the va */
13683 				va = *(uint64_t *)old << 22;
13684 				/* turn va into a virtual pfn */
13685 				va >>= 22 - TSB_START_SIZE;
13686 				/*
13687 				 * or in bits from the offset in the tsb
13688 				 * to get the real virtual pfn. These
13689 				 * correspond to bits [21:13] in the va
13690 				 */
13691 				vpshift =
13692 				    TTE_BSZS_SHIFT(TTE_CSZ(&old->tte_data)) &
13693 				    0x1ff;
13694 				va |= (i << vpshift);
13695 				va >>= vpshift;
13696 				new_offset = va & (new_entries - 1);
13697 				new = new_base + new_offset;
13698 				prefetch_tsbe_write(new);
13699 				*new = *old;
13700 			}
13701 		}
13702 	}
13703 }
13704 
13705 /*
13706  * unused in sfmmu
13707  */
13708 void
13709 hat_dump(void)
13710 {
13711 }
13712 
13713 /*
13714  * Called when a thread is exiting and we have switched to the kernel address
13715  * space.  Perform the same VM initialization resume() uses when switching
13716  * processes.
13717  *
13718  * Note that sfmmu_load_mmustate() is currently a no-op for kernel threads, but
13719  * we call it anyway in case the semantics change in the future.
13720  */
13721 /*ARGSUSED*/
13722 void
13723 hat_thread_exit(kthread_t *thd)
13724 {
13725 	uint_t pgsz_cnum;
13726 	uint_t pstate_save;
13727 
13728 	ASSERT(thd->t_procp->p_as == &kas);
13729 
13730 	pgsz_cnum = KCONTEXT;
13731 #ifdef sun4u
13732 	pgsz_cnum |= (ksfmmup->sfmmu_cext << CTXREG_EXT_SHIFT);
13733 #endif
13734 
13735 	/*
13736 	 * Note that sfmmu_load_mmustate() is currently a no-op for
13737 	 * kernel threads. We need to disable interrupts here,
13738 	 * simply because otherwise sfmmu_load_mmustate() would panic
13739 	 * if the caller does not disable interrupts.
13740 	 */
13741 	pstate_save = sfmmu_disable_intrs();
13742 
13743 	/* Compatibility Note: hw takes care of MMU_SCONTEXT1 */
13744 	sfmmu_setctx_sec(pgsz_cnum);
13745 	sfmmu_load_mmustate(ksfmmup);
13746 	sfmmu_enable_intrs(pstate_save);
13747 }
13748 
13749 
13750 /*
13751  * SRD support
13752  */
13753 #define	SRD_HASH_FUNCTION(vp)	(((((uintptr_t)(vp)) >> 4) ^ \
13754 				    (((uintptr_t)(vp)) >> 11)) & \
13755 				    srd_hashmask)
13756 
13757 /*
13758  * Attach the process to the srd struct associated with the exec vnode
13759  * from which the process is started.
13760  */
13761 void
13762 hat_join_srd(struct hat *sfmmup, vnode_t *evp)
13763 {
13764 	uint_t hash = SRD_HASH_FUNCTION(evp);
13765 	sf_srd_t *srdp;
13766 	sf_srd_t *newsrdp;
13767 
13768 	ASSERT(sfmmup != ksfmmup);
13769 	ASSERT(sfmmup->sfmmu_srdp == NULL);
13770 
13771 	if (!shctx_on) {
13772 		return;
13773 	}
13774 
13775 	VN_HOLD(evp);
13776 
13777 	if (srd_buckets[hash].srdb_srdp != NULL) {
13778 		mutex_enter(&srd_buckets[hash].srdb_lock);
13779 		for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13780 		    srdp = srdp->srd_hash) {
13781 			if (srdp->srd_evp == evp) {
13782 				ASSERT(srdp->srd_refcnt >= 0);
13783 				sfmmup->sfmmu_srdp = srdp;
13784 				atomic_add_32(
13785 				    (volatile uint_t *)&srdp->srd_refcnt, 1);
13786 				mutex_exit(&srd_buckets[hash].srdb_lock);
13787 				return;
13788 			}
13789 		}
13790 		mutex_exit(&srd_buckets[hash].srdb_lock);
13791 	}
13792 	newsrdp = kmem_cache_alloc(srd_cache, KM_SLEEP);
13793 	ASSERT(newsrdp->srd_next_ismrid == 0 && newsrdp->srd_next_hmerid == 0);
13794 
13795 	newsrdp->srd_evp = evp;
13796 	newsrdp->srd_refcnt = 1;
13797 	newsrdp->srd_hmergnfree = NULL;
13798 	newsrdp->srd_ismrgnfree = NULL;
13799 
13800 	mutex_enter(&srd_buckets[hash].srdb_lock);
13801 	for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13802 	    srdp = srdp->srd_hash) {
13803 		if (srdp->srd_evp == evp) {
13804 			ASSERT(srdp->srd_refcnt >= 0);
13805 			sfmmup->sfmmu_srdp = srdp;
13806 			atomic_add_32((volatile uint_t *)&srdp->srd_refcnt, 1);
13807 			mutex_exit(&srd_buckets[hash].srdb_lock);
13808 			kmem_cache_free(srd_cache, newsrdp);
13809 			return;
13810 		}
13811 	}
13812 	newsrdp->srd_hash = srd_buckets[hash].srdb_srdp;
13813 	srd_buckets[hash].srdb_srdp = newsrdp;
13814 	sfmmup->sfmmu_srdp = newsrdp;
13815 
13816 	mutex_exit(&srd_buckets[hash].srdb_lock);
13817 
13818 }
13819 
13820 static void
13821 sfmmu_leave_srd(sfmmu_t *sfmmup)
13822 {
13823 	vnode_t *evp;
13824 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
13825 	uint_t hash;
13826 	sf_srd_t **prev_srdpp;
13827 	sf_region_t *rgnp;
13828 	sf_region_t *nrgnp;
13829 #ifdef DEBUG
13830 	int rgns = 0;
13831 #endif
13832 	int i;
13833 
13834 	ASSERT(sfmmup != ksfmmup);
13835 	ASSERT(srdp != NULL);
13836 	ASSERT(srdp->srd_refcnt > 0);
13837 	ASSERT(sfmmup->sfmmu_scdp == NULL);
13838 	ASSERT(sfmmup->sfmmu_free == 1);
13839 
13840 	sfmmup->sfmmu_srdp = NULL;
13841 	evp = srdp->srd_evp;
13842 	ASSERT(evp != NULL);
13843 	if (atomic_add_32_nv(
13844 	    (volatile uint_t *)&srdp->srd_refcnt, -1)) {
13845 		VN_RELE(evp);
13846 		return;
13847 	}
13848 
13849 	hash = SRD_HASH_FUNCTION(evp);
13850 	mutex_enter(&srd_buckets[hash].srdb_lock);
13851 	for (prev_srdpp = &srd_buckets[hash].srdb_srdp;
13852 	    (srdp = *prev_srdpp) != NULL; prev_srdpp = &srdp->srd_hash) {
13853 		if (srdp->srd_evp == evp) {
13854 			break;
13855 		}
13856 	}
13857 	if (srdp == NULL || srdp->srd_refcnt) {
13858 		mutex_exit(&srd_buckets[hash].srdb_lock);
13859 		VN_RELE(evp);
13860 		return;
13861 	}
13862 	*prev_srdpp = srdp->srd_hash;
13863 	mutex_exit(&srd_buckets[hash].srdb_lock);
13864 
13865 	ASSERT(srdp->srd_refcnt == 0);
13866 	VN_RELE(evp);
13867 
13868 #ifdef DEBUG
13869 	for (i = 0; i < SFMMU_MAX_REGION_BUCKETS; i++) {
13870 		ASSERT(srdp->srd_rgnhash[i] == NULL);
13871 	}
13872 #endif /* DEBUG */
13873 
13874 	/* free each hme regions in the srd */
13875 	for (rgnp = srdp->srd_hmergnfree; rgnp != NULL; rgnp = nrgnp) {
13876 		nrgnp = rgnp->rgn_next;
13877 		ASSERT(rgnp->rgn_id < srdp->srd_next_hmerid);
13878 		ASSERT(rgnp->rgn_refcnt == 0);
13879 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13880 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13881 		ASSERT(rgnp->rgn_hmeflags == 0);
13882 		ASSERT(srdp->srd_hmergnp[rgnp->rgn_id] == rgnp);
13883 #ifdef DEBUG
13884 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13885 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13886 		}
13887 		rgns++;
13888 #endif /* DEBUG */
13889 		kmem_cache_free(region_cache, rgnp);
13890 	}
13891 	ASSERT(rgns == srdp->srd_next_hmerid);
13892 
13893 #ifdef DEBUG
13894 	rgns = 0;
13895 #endif
13896 	/* free each ism rgns in the srd */
13897 	for (rgnp = srdp->srd_ismrgnfree; rgnp != NULL; rgnp = nrgnp) {
13898 		nrgnp = rgnp->rgn_next;
13899 		ASSERT(rgnp->rgn_id < srdp->srd_next_ismrid);
13900 		ASSERT(rgnp->rgn_refcnt == 0);
13901 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13902 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13903 		ASSERT(srdp->srd_ismrgnp[rgnp->rgn_id] == rgnp);
13904 #ifdef DEBUG
13905 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13906 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13907 		}
13908 		rgns++;
13909 #endif /* DEBUG */
13910 		kmem_cache_free(region_cache, rgnp);
13911 	}
13912 	ASSERT(rgns == srdp->srd_next_ismrid);
13913 	ASSERT(srdp->srd_ismbusyrgns == 0);
13914 	ASSERT(srdp->srd_hmebusyrgns == 0);
13915 
13916 	srdp->srd_next_ismrid = 0;
13917 	srdp->srd_next_hmerid = 0;
13918 
13919 	bzero((void *)srdp->srd_ismrgnp,
13920 	    sizeof (sf_region_t *) * SFMMU_MAX_ISM_REGIONS);
13921 	bzero((void *)srdp->srd_hmergnp,
13922 	    sizeof (sf_region_t *) * SFMMU_MAX_HME_REGIONS);
13923 
13924 	ASSERT(srdp->srd_scdp == NULL);
13925 	kmem_cache_free(srd_cache, srdp);
13926 }
13927 
13928 /* ARGSUSED */
13929 static int
13930 sfmmu_srdcache_constructor(void *buf, void *cdrarg, int kmflags)
13931 {
13932 	sf_srd_t *srdp = (sf_srd_t *)buf;
13933 	bzero(buf, sizeof (*srdp));
13934 
13935 	mutex_init(&srdp->srd_mutex, NULL, MUTEX_DEFAULT, NULL);
13936 	mutex_init(&srdp->srd_scd_mutex, NULL, MUTEX_DEFAULT, NULL);
13937 	return (0);
13938 }
13939 
13940 /* ARGSUSED */
13941 static void
13942 sfmmu_srdcache_destructor(void *buf, void *cdrarg)
13943 {
13944 	sf_srd_t *srdp = (sf_srd_t *)buf;
13945 
13946 	mutex_destroy(&srdp->srd_mutex);
13947 	mutex_destroy(&srdp->srd_scd_mutex);
13948 }
13949 
13950 /*
13951  * The caller makes sure hat_join_region()/hat_leave_region() can't be called
13952  * at the same time for the same process and address range. This is ensured by
13953  * the fact that address space is locked as writer when a process joins the
13954  * regions. Therefore there's no need to hold an srd lock during the entire
13955  * execution of hat_join_region()/hat_leave_region().
13956  */
13957 
13958 #define	RGN_HASH_FUNCTION(obj)	(((((uintptr_t)(obj)) >> 4) ^ \
13959 				    (((uintptr_t)(obj)) >> 11)) & \
13960 					srd_rgn_hashmask)
13961 /*
13962  * This routine implements the shared context functionality required when
13963  * attaching a segment to an address space. It must be called from
13964  * hat_share() for D(ISM) segments and from segvn_create() for segments
13965  * with the MAP_PRIVATE and MAP_TEXT flags set. It returns a region_cookie
13966  * which is saved in the private segment data for hme segments and
13967  * the ism_map structure for ism segments.
13968  */
13969 hat_region_cookie_t
13970 hat_join_region(struct hat *sfmmup,
13971 	caddr_t r_saddr,
13972 	size_t r_size,
13973 	void *r_obj,
13974 	u_offset_t r_objoff,
13975 	uchar_t r_perm,
13976 	uchar_t r_pgszc,
13977 	hat_rgn_cb_func_t r_cb_function,
13978 	uint_t flags)
13979 {
13980 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
13981 	uint_t rhash;
13982 	uint_t rid;
13983 	hatlock_t *hatlockp;
13984 	sf_region_t *rgnp;
13985 	sf_region_t *new_rgnp = NULL;
13986 	int i;
13987 	uint16_t *nextidp;
13988 	sf_region_t **freelistp;
13989 	int maxids;
13990 	sf_region_t **rarrp;
13991 	uint16_t *busyrgnsp;
13992 	ulong_t rttecnt;
13993 	uchar_t tteflag;
13994 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
13995 	int text = (r_type == HAT_REGION_TEXT);
13996 
13997 	if (srdp == NULL || r_size == 0) {
13998 		return (HAT_INVALID_REGION_COOKIE);
13999 	}
14000 
14001 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
14002 	ASSERT(sfmmup != ksfmmup);
14003 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
14004 	ASSERT(srdp->srd_refcnt > 0);
14005 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
14006 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
14007 	ASSERT(r_pgszc < mmu_page_sizes);
14008 	if (!IS_P2ALIGNED(r_saddr, TTEBYTES(r_pgszc)) ||
14009 	    !IS_P2ALIGNED(r_size, TTEBYTES(r_pgszc))) {
14010 		panic("hat_join_region: region addr or size is not aligned\n");
14011 	}
14012 
14013 
14014 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
14015 	    SFMMU_REGION_HME;
14016 	/*
14017 	 * Currently only support shared hmes for the read only main text
14018 	 * region.
14019 	 */
14020 	if (r_type == SFMMU_REGION_HME && ((r_obj != srdp->srd_evp) ||
14021 	    (r_perm & PROT_WRITE))) {
14022 		return (HAT_INVALID_REGION_COOKIE);
14023 	}
14024 
14025 	rhash = RGN_HASH_FUNCTION(r_obj);
14026 
14027 	if (r_type == SFMMU_REGION_ISM) {
14028 		nextidp = &srdp->srd_next_ismrid;
14029 		freelistp = &srdp->srd_ismrgnfree;
14030 		maxids = SFMMU_MAX_ISM_REGIONS;
14031 		rarrp = srdp->srd_ismrgnp;
14032 		busyrgnsp = &srdp->srd_ismbusyrgns;
14033 	} else {
14034 		nextidp = &srdp->srd_next_hmerid;
14035 		freelistp = &srdp->srd_hmergnfree;
14036 		maxids = SFMMU_MAX_HME_REGIONS;
14037 		rarrp = srdp->srd_hmergnp;
14038 		busyrgnsp = &srdp->srd_hmebusyrgns;
14039 	}
14040 
14041 	mutex_enter(&srdp->srd_mutex);
14042 
14043 	for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
14044 	    rgnp = rgnp->rgn_hash) {
14045 		if (rgnp->rgn_saddr == r_saddr && rgnp->rgn_size == r_size &&
14046 		    rgnp->rgn_obj == r_obj && rgnp->rgn_objoff == r_objoff &&
14047 		    rgnp->rgn_perm == r_perm && rgnp->rgn_pgszc == r_pgszc) {
14048 			break;
14049 		}
14050 	}
14051 
14052 rfound:
14053 	if (rgnp != NULL) {
14054 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14055 		ASSERT(rgnp->rgn_cb_function == r_cb_function);
14056 		ASSERT(rgnp->rgn_refcnt >= 0);
14057 		rid = rgnp->rgn_id;
14058 		ASSERT(rid < maxids);
14059 		ASSERT(rarrp[rid] == rgnp);
14060 		ASSERT(rid < *nextidp);
14061 		atomic_add_32((volatile uint_t *)&rgnp->rgn_refcnt, 1);
14062 		mutex_exit(&srdp->srd_mutex);
14063 		if (new_rgnp != NULL) {
14064 			kmem_cache_free(region_cache, new_rgnp);
14065 		}
14066 		if (r_type == SFMMU_REGION_HME) {
14067 			int myjoin =
14068 			    (sfmmup == astosfmmu(curthread->t_procp->p_as));
14069 
14070 			sfmmu_link_to_hmeregion(sfmmup, rgnp);
14071 			/*
14072 			 * bitmap should be updated after linking sfmmu on
14073 			 * region list so that pageunload() doesn't skip
14074 			 * TSB/TLB flush. As soon as bitmap is updated another
14075 			 * thread in this process can already start accessing
14076 			 * this region.
14077 			 */
14078 			/*
14079 			 * Normally ttecnt accounting is done as part of
14080 			 * pagefault handling. But a process may not take any
14081 			 * pagefaults on shared hmeblks created by some other
14082 			 * process. To compensate for this assume that the
14083 			 * entire region will end up faulted in using
14084 			 * the region's pagesize.
14085 			 *
14086 			 */
14087 			if (r_pgszc > TTE8K) {
14088 				tteflag = 1 << r_pgszc;
14089 				if (disable_large_pages & tteflag) {
14090 					tteflag = 0;
14091 				}
14092 			} else {
14093 				tteflag = 0;
14094 			}
14095 			if (tteflag && !(sfmmup->sfmmu_rtteflags & tteflag)) {
14096 				hatlockp = sfmmu_hat_enter(sfmmup);
14097 				sfmmup->sfmmu_rtteflags |= tteflag;
14098 				sfmmu_hat_exit(hatlockp);
14099 			}
14100 			hatlockp = sfmmu_hat_enter(sfmmup);
14101 
14102 			/*
14103 			 * Preallocate 1/4 of ttecnt's in 8K TSB for >= 4M
14104 			 * region to allow for large page allocation failure.
14105 			 */
14106 			if (r_pgszc >= TTE4M) {
14107 				sfmmup->sfmmu_tsb0_4minflcnt +=
14108 				    r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14109 			}
14110 
14111 			/* update sfmmu_ttecnt with the shme rgn ttecnt */
14112 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14113 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14114 			    rttecnt);
14115 
14116 			if (text && r_pgszc >= TTE4M &&
14117 			    (tteflag || ((disable_large_pages >> TTE4M) &
14118 			    ((1 << (r_pgszc - TTE4M + 1)) - 1))) &&
14119 			    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
14120 				SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
14121 			}
14122 
14123 			sfmmu_hat_exit(hatlockp);
14124 			/*
14125 			 * On Panther we need to make sure TLB is programmed
14126 			 * to accept 32M/256M pages.  Call
14127 			 * sfmmu_check_page_sizes() now to make sure TLB is
14128 			 * setup before making hmeregions visible to other
14129 			 * threads.
14130 			 */
14131 			sfmmu_check_page_sizes(sfmmup, 1);
14132 			hatlockp = sfmmu_hat_enter(sfmmup);
14133 			SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14134 
14135 			/*
14136 			 * if context is invalid tsb miss exception code will
14137 			 * call sfmmu_check_page_sizes() and update tsbmiss
14138 			 * area later.
14139 			 */
14140 			kpreempt_disable();
14141 			if (myjoin &&
14142 			    (sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
14143 			    != INVALID_CONTEXT)) {
14144 				struct tsbmiss *tsbmp;
14145 
14146 				tsbmp = &tsbmiss_area[CPU->cpu_id];
14147 				ASSERT(sfmmup == tsbmp->usfmmup);
14148 				BT_SET(tsbmp->shmermap, rid);
14149 				if (r_pgszc > TTE64K) {
14150 					tsbmp->uhat_rtteflags |= tteflag;
14151 				}
14152 
14153 			}
14154 			kpreempt_enable();
14155 
14156 			sfmmu_hat_exit(hatlockp);
14157 			ASSERT((hat_region_cookie_t)((uint64_t)rid) !=
14158 			    HAT_INVALID_REGION_COOKIE);
14159 		} else {
14160 			hatlockp = sfmmu_hat_enter(sfmmup);
14161 			SF_RGNMAP_ADD(sfmmup->sfmmu_ismregion_map, rid);
14162 			sfmmu_hat_exit(hatlockp);
14163 		}
14164 		ASSERT(rid < maxids);
14165 
14166 		if (r_type == SFMMU_REGION_ISM) {
14167 			sfmmu_find_scd(sfmmup);
14168 		}
14169 		return ((hat_region_cookie_t)((uint64_t)rid));
14170 	}
14171 
14172 	ASSERT(new_rgnp == NULL);
14173 
14174 	if (*busyrgnsp >= maxids) {
14175 		mutex_exit(&srdp->srd_mutex);
14176 		return (HAT_INVALID_REGION_COOKIE);
14177 	}
14178 
14179 	ASSERT(MUTEX_HELD(&srdp->srd_mutex));
14180 	if (*freelistp != NULL) {
14181 		rgnp = *freelistp;
14182 		*freelistp = rgnp->rgn_next;
14183 		ASSERT(rgnp->rgn_id < *nextidp);
14184 		ASSERT(rgnp->rgn_id < maxids);
14185 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
14186 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK)
14187 		    == r_type);
14188 		ASSERT(rarrp[rgnp->rgn_id] == rgnp);
14189 		ASSERT(rgnp->rgn_hmeflags == 0);
14190 	} else {
14191 		/*
14192 		 * release local locks before memory allocation.
14193 		 */
14194 		mutex_exit(&srdp->srd_mutex);
14195 
14196 		new_rgnp = kmem_cache_alloc(region_cache, KM_SLEEP);
14197 
14198 		mutex_enter(&srdp->srd_mutex);
14199 		for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
14200 		    rgnp = rgnp->rgn_hash) {
14201 			if (rgnp->rgn_saddr == r_saddr &&
14202 			    rgnp->rgn_size == r_size &&
14203 			    rgnp->rgn_obj == r_obj &&
14204 			    rgnp->rgn_objoff == r_objoff &&
14205 			    rgnp->rgn_perm == r_perm &&
14206 			    rgnp->rgn_pgszc == r_pgszc) {
14207 				break;
14208 			}
14209 		}
14210 		if (rgnp != NULL) {
14211 			goto rfound;
14212 		}
14213 
14214 		if (*nextidp >= maxids) {
14215 			mutex_exit(&srdp->srd_mutex);
14216 			goto fail;
14217 		}
14218 		rgnp = new_rgnp;
14219 		new_rgnp = NULL;
14220 		rgnp->rgn_id = (*nextidp)++;
14221 		ASSERT(rgnp->rgn_id < maxids);
14222 		ASSERT(rarrp[rgnp->rgn_id] == NULL);
14223 		rarrp[rgnp->rgn_id] = rgnp;
14224 	}
14225 
14226 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14227 	ASSERT(rgnp->rgn_hmeflags == 0);
14228 #ifdef DEBUG
14229 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14230 		ASSERT(rgnp->rgn_ttecnt[i] == 0);
14231 	}
14232 #endif
14233 	rgnp->rgn_saddr = r_saddr;
14234 	rgnp->rgn_size = r_size;
14235 	rgnp->rgn_obj = r_obj;
14236 	rgnp->rgn_objoff = r_objoff;
14237 	rgnp->rgn_perm = r_perm;
14238 	rgnp->rgn_pgszc = r_pgszc;
14239 	rgnp->rgn_flags = r_type;
14240 	rgnp->rgn_refcnt = 0;
14241 	rgnp->rgn_cb_function = r_cb_function;
14242 	rgnp->rgn_hash = srdp->srd_rgnhash[rhash];
14243 	srdp->srd_rgnhash[rhash] = rgnp;
14244 	(*busyrgnsp)++;
14245 	ASSERT(*busyrgnsp <= maxids);
14246 	goto rfound;
14247 
14248 fail:
14249 	ASSERT(new_rgnp != NULL);
14250 	kmem_cache_free(region_cache, new_rgnp);
14251 	return (HAT_INVALID_REGION_COOKIE);
14252 }
14253 
14254 /*
14255  * This function implements the shared context functionality required
14256  * when detaching a segment from an address space. It must be called
14257  * from hat_unshare() for all D(ISM) segments and from segvn_unmap(),
14258  * for segments with a valid region_cookie.
14259  * It will also be called from all seg_vn routines which change a
14260  * segment's attributes such as segvn_setprot(), segvn_setpagesize(),
14261  * segvn_clrszc() & segvn_advise(), as well as in the case of COW fault
14262  * from segvn_fault().
14263  */
14264 void
14265 hat_leave_region(struct hat *sfmmup, hat_region_cookie_t rcookie, uint_t flags)
14266 {
14267 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14268 	sf_scd_t *scdp;
14269 	uint_t rhash;
14270 	uint_t rid = (uint_t)((uint64_t)rcookie);
14271 	hatlock_t *hatlockp = NULL;
14272 	sf_region_t *rgnp;
14273 	sf_region_t **prev_rgnpp;
14274 	sf_region_t *cur_rgnp;
14275 	void *r_obj;
14276 	int i;
14277 	caddr_t	r_saddr;
14278 	caddr_t r_eaddr;
14279 	size_t	r_size;
14280 	uchar_t	r_pgszc;
14281 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
14282 
14283 	ASSERT(sfmmup != ksfmmup);
14284 	ASSERT(srdp != NULL);
14285 	ASSERT(srdp->srd_refcnt > 0);
14286 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
14287 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
14288 	ASSERT(!sfmmup->sfmmu_free || sfmmup->sfmmu_scdp == NULL);
14289 
14290 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
14291 	    SFMMU_REGION_HME;
14292 
14293 	if (r_type == SFMMU_REGION_ISM) {
14294 		ASSERT(SFMMU_IS_ISMRID_VALID(rid));
14295 		ASSERT(rid < SFMMU_MAX_ISM_REGIONS);
14296 		rgnp = srdp->srd_ismrgnp[rid];
14297 	} else {
14298 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14299 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14300 		rgnp = srdp->srd_hmergnp[rid];
14301 	}
14302 	ASSERT(rgnp != NULL);
14303 	ASSERT(rgnp->rgn_id == rid);
14304 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14305 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14306 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
14307 
14308 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
14309 	if (r_type == SFMMU_REGION_HME && sfmmup->sfmmu_as->a_xhat != NULL) {
14310 		xhat_unload_callback_all(sfmmup->sfmmu_as, rgnp->rgn_saddr,
14311 		    rgnp->rgn_size, 0, NULL);
14312 	}
14313 
14314 	if (sfmmup->sfmmu_free) {
14315 		ulong_t rttecnt;
14316 		r_pgszc = rgnp->rgn_pgszc;
14317 		r_size = rgnp->rgn_size;
14318 
14319 		ASSERT(sfmmup->sfmmu_scdp == NULL);
14320 		if (r_type == SFMMU_REGION_ISM) {
14321 			SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14322 		} else {
14323 			/* update shme rgns ttecnt in sfmmu_ttecnt */
14324 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14325 			ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14326 
14327 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14328 			    -rttecnt);
14329 
14330 			SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14331 		}
14332 	} else if (r_type == SFMMU_REGION_ISM) {
14333 		hatlockp = sfmmu_hat_enter(sfmmup);
14334 		ASSERT(rid < srdp->srd_next_ismrid);
14335 		SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14336 		scdp = sfmmup->sfmmu_scdp;
14337 		if (scdp != NULL &&
14338 		    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
14339 			sfmmu_leave_scd(sfmmup, r_type);
14340 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14341 		}
14342 		sfmmu_hat_exit(hatlockp);
14343 	} else {
14344 		ulong_t rttecnt;
14345 		r_pgszc = rgnp->rgn_pgszc;
14346 		r_saddr = rgnp->rgn_saddr;
14347 		r_size = rgnp->rgn_size;
14348 		r_eaddr = r_saddr + r_size;
14349 
14350 		ASSERT(r_type == SFMMU_REGION_HME);
14351 		hatlockp = sfmmu_hat_enter(sfmmup);
14352 		ASSERT(rid < srdp->srd_next_hmerid);
14353 		SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14354 
14355 		/*
14356 		 * If region is part of an SCD call sfmmu_leave_scd().
14357 		 * Otherwise if process is not exiting and has valid context
14358 		 * just drop the context on the floor to lose stale TLB
14359 		 * entries and force the update of tsb miss area to reflect
14360 		 * the new region map. After that clean our TSB entries.
14361 		 */
14362 		scdp = sfmmup->sfmmu_scdp;
14363 		if (scdp != NULL &&
14364 		    SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
14365 			sfmmu_leave_scd(sfmmup, r_type);
14366 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14367 		}
14368 		sfmmu_invalidate_ctx(sfmmup);
14369 
14370 		i = TTE8K;
14371 		while (i < mmu_page_sizes) {
14372 			if (rgnp->rgn_ttecnt[i] != 0) {
14373 				sfmmu_unload_tsb_range(sfmmup, r_saddr,
14374 				    r_eaddr, i);
14375 				if (i < TTE4M) {
14376 					i = TTE4M;
14377 					continue;
14378 				} else {
14379 					break;
14380 				}
14381 			}
14382 			i++;
14383 		}
14384 		/* Remove the preallocated 1/4 8k ttecnt for 4M regions. */
14385 		if (r_pgszc >= TTE4M) {
14386 			rttecnt = r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14387 			ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
14388 			    rttecnt);
14389 			sfmmup->sfmmu_tsb0_4minflcnt -= rttecnt;
14390 		}
14391 
14392 		/* update shme rgns ttecnt in sfmmu_ttecnt */
14393 		rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14394 		ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14395 		atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc], -rttecnt);
14396 
14397 		sfmmu_hat_exit(hatlockp);
14398 		if (scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
14399 			/* sfmmup left the scd, grow private tsb */
14400 			sfmmu_check_page_sizes(sfmmup, 1);
14401 		} else {
14402 			sfmmu_check_page_sizes(sfmmup, 0);
14403 		}
14404 	}
14405 
14406 	if (r_type == SFMMU_REGION_HME) {
14407 		sfmmu_unlink_from_hmeregion(sfmmup, rgnp);
14408 	}
14409 
14410 	r_obj = rgnp->rgn_obj;
14411 	if (atomic_add_32_nv((volatile uint_t *)&rgnp->rgn_refcnt, -1)) {
14412 		return;
14413 	}
14414 
14415 	/*
14416 	 * looks like nobody uses this region anymore. Free it.
14417 	 */
14418 	rhash = RGN_HASH_FUNCTION(r_obj);
14419 	mutex_enter(&srdp->srd_mutex);
14420 	for (prev_rgnpp = &srdp->srd_rgnhash[rhash];
14421 	    (cur_rgnp = *prev_rgnpp) != NULL;
14422 	    prev_rgnpp = &cur_rgnp->rgn_hash) {
14423 		if (cur_rgnp == rgnp && cur_rgnp->rgn_refcnt == 0) {
14424 			break;
14425 		}
14426 	}
14427 
14428 	if (cur_rgnp == NULL) {
14429 		mutex_exit(&srdp->srd_mutex);
14430 		return;
14431 	}
14432 
14433 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14434 	*prev_rgnpp = rgnp->rgn_hash;
14435 	if (r_type == SFMMU_REGION_ISM) {
14436 		rgnp->rgn_flags |= SFMMU_REGION_FREE;
14437 		ASSERT(rid < srdp->srd_next_ismrid);
14438 		rgnp->rgn_next = srdp->srd_ismrgnfree;
14439 		srdp->srd_ismrgnfree = rgnp;
14440 		ASSERT(srdp->srd_ismbusyrgns > 0);
14441 		srdp->srd_ismbusyrgns--;
14442 		mutex_exit(&srdp->srd_mutex);
14443 		return;
14444 	}
14445 	mutex_exit(&srdp->srd_mutex);
14446 
14447 	/*
14448 	 * Destroy region's hmeblks.
14449 	 */
14450 	sfmmu_unload_hmeregion(srdp, rgnp);
14451 
14452 	rgnp->rgn_hmeflags = 0;
14453 
14454 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14455 	ASSERT(rgnp->rgn_id == rid);
14456 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14457 		rgnp->rgn_ttecnt[i] = 0;
14458 	}
14459 	rgnp->rgn_flags |= SFMMU_REGION_FREE;
14460 	mutex_enter(&srdp->srd_mutex);
14461 	ASSERT(rid < srdp->srd_next_hmerid);
14462 	rgnp->rgn_next = srdp->srd_hmergnfree;
14463 	srdp->srd_hmergnfree = rgnp;
14464 	ASSERT(srdp->srd_hmebusyrgns > 0);
14465 	srdp->srd_hmebusyrgns--;
14466 	mutex_exit(&srdp->srd_mutex);
14467 }
14468 
14469 /*
14470  * For now only called for hmeblk regions and not for ISM regions.
14471  */
14472 void
14473 hat_dup_region(struct hat *sfmmup, hat_region_cookie_t rcookie)
14474 {
14475 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14476 	uint_t rid = (uint_t)((uint64_t)rcookie);
14477 	sf_region_t *rgnp;
14478 	sf_rgn_link_t *rlink;
14479 	sf_rgn_link_t *hrlink;
14480 	ulong_t	rttecnt;
14481 
14482 	ASSERT(sfmmup != ksfmmup);
14483 	ASSERT(srdp != NULL);
14484 	ASSERT(srdp->srd_refcnt > 0);
14485 
14486 	ASSERT(rid < srdp->srd_next_hmerid);
14487 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14488 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14489 
14490 	rgnp = srdp->srd_hmergnp[rid];
14491 	ASSERT(rgnp->rgn_refcnt > 0);
14492 	ASSERT(rgnp->rgn_id == rid);
14493 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == SFMMU_REGION_HME);
14494 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14495 
14496 	atomic_add_32((volatile uint_t *)&rgnp->rgn_refcnt, 1);
14497 
14498 	/* LINTED: constant in conditional context */
14499 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 0);
14500 	ASSERT(rlink != NULL);
14501 	mutex_enter(&rgnp->rgn_mutex);
14502 	ASSERT(rgnp->rgn_sfmmu_head != NULL);
14503 	/* LINTED: constant in conditional context */
14504 	SFMMU_HMERID2RLINKP(rgnp->rgn_sfmmu_head, rid, hrlink, 0, 0);
14505 	ASSERT(hrlink != NULL);
14506 	ASSERT(hrlink->prev == NULL);
14507 	rlink->next = rgnp->rgn_sfmmu_head;
14508 	rlink->prev = NULL;
14509 	hrlink->prev = sfmmup;
14510 	/*
14511 	 * make sure rlink's next field is correct
14512 	 * before making this link visible.
14513 	 */
14514 	membar_stst();
14515 	rgnp->rgn_sfmmu_head = sfmmup;
14516 	mutex_exit(&rgnp->rgn_mutex);
14517 
14518 	/* update sfmmu_ttecnt with the shme rgn ttecnt */
14519 	rttecnt = rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
14520 	atomic_add_long(&sfmmup->sfmmu_ttecnt[rgnp->rgn_pgszc], rttecnt);
14521 	/* update tsb0 inflation count */
14522 	if (rgnp->rgn_pgszc >= TTE4M) {
14523 		sfmmup->sfmmu_tsb0_4minflcnt +=
14524 		    rgnp->rgn_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14525 	}
14526 	/*
14527 	 * Update regionid bitmask without hat lock since no other thread
14528 	 * can update this region bitmask right now.
14529 	 */
14530 	SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14531 }
14532 
14533 /* ARGSUSED */
14534 static int
14535 sfmmu_rgncache_constructor(void *buf, void *cdrarg, int kmflags)
14536 {
14537 	sf_region_t *rgnp = (sf_region_t *)buf;
14538 	bzero(buf, sizeof (*rgnp));
14539 
14540 	mutex_init(&rgnp->rgn_mutex, NULL, MUTEX_DEFAULT, NULL);
14541 
14542 	return (0);
14543 }
14544 
14545 /* ARGSUSED */
14546 static void
14547 sfmmu_rgncache_destructor(void *buf, void *cdrarg)
14548 {
14549 	sf_region_t *rgnp = (sf_region_t *)buf;
14550 	mutex_destroy(&rgnp->rgn_mutex);
14551 }
14552 
14553 static int
14554 sfrgnmap_isnull(sf_region_map_t *map)
14555 {
14556 	int i;
14557 
14558 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14559 		if (map->bitmap[i] != 0) {
14560 			return (0);
14561 		}
14562 	}
14563 	return (1);
14564 }
14565 
14566 static int
14567 sfhmergnmap_isnull(sf_hmeregion_map_t *map)
14568 {
14569 	int i;
14570 
14571 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
14572 		if (map->bitmap[i] != 0) {
14573 			return (0);
14574 		}
14575 	}
14576 	return (1);
14577 }
14578 
14579 #ifdef DEBUG
14580 static void
14581 check_scd_sfmmu_list(sfmmu_t **headp, sfmmu_t *sfmmup, int onlist)
14582 {
14583 	sfmmu_t *sp;
14584 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14585 
14586 	for (sp = *headp; sp != NULL; sp = sp->sfmmu_scd_link.next) {
14587 		ASSERT(srdp == sp->sfmmu_srdp);
14588 		if (sp == sfmmup) {
14589 			if (onlist) {
14590 				return;
14591 			} else {
14592 				panic("shctx: sfmmu 0x%p found on scd"
14593 				    "list 0x%p", (void *)sfmmup,
14594 				    (void *)*headp);
14595 			}
14596 		}
14597 	}
14598 	if (onlist) {
14599 		panic("shctx: sfmmu 0x%p not found on scd list 0x%p",
14600 		    (void *)sfmmup, (void *)*headp);
14601 	} else {
14602 		return;
14603 	}
14604 }
14605 #else /* DEBUG */
14606 #define	check_scd_sfmmu_list(headp, sfmmup, onlist)
14607 #endif /* DEBUG */
14608 
14609 /*
14610  * Removes an sfmmu from the SCD sfmmu list.
14611  */
14612 static void
14613 sfmmu_from_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14614 {
14615 	ASSERT(sfmmup->sfmmu_srdp != NULL);
14616 	check_scd_sfmmu_list(headp, sfmmup, 1);
14617 	if (sfmmup->sfmmu_scd_link.prev != NULL) {
14618 		ASSERT(*headp != sfmmup);
14619 		sfmmup->sfmmu_scd_link.prev->sfmmu_scd_link.next =
14620 		    sfmmup->sfmmu_scd_link.next;
14621 	} else {
14622 		ASSERT(*headp == sfmmup);
14623 		*headp = sfmmup->sfmmu_scd_link.next;
14624 	}
14625 	if (sfmmup->sfmmu_scd_link.next != NULL) {
14626 		sfmmup->sfmmu_scd_link.next->sfmmu_scd_link.prev =
14627 		    sfmmup->sfmmu_scd_link.prev;
14628 	}
14629 }
14630 
14631 
14632 /*
14633  * Adds an sfmmu to the start of the queue.
14634  */
14635 static void
14636 sfmmu_to_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14637 {
14638 	check_scd_sfmmu_list(headp, sfmmup, 0);
14639 	sfmmup->sfmmu_scd_link.prev = NULL;
14640 	sfmmup->sfmmu_scd_link.next = *headp;
14641 	if (*headp != NULL)
14642 		(*headp)->sfmmu_scd_link.prev = sfmmup;
14643 	*headp = sfmmup;
14644 }
14645 
14646 /*
14647  * Remove an scd from the start of the queue.
14648  */
14649 static void
14650 sfmmu_remove_scd(sf_scd_t **headp, sf_scd_t *scdp)
14651 {
14652 	if (scdp->scd_prev != NULL) {
14653 		ASSERT(*headp != scdp);
14654 		scdp->scd_prev->scd_next = scdp->scd_next;
14655 	} else {
14656 		ASSERT(*headp == scdp);
14657 		*headp = scdp->scd_next;
14658 	}
14659 
14660 	if (scdp->scd_next != NULL) {
14661 		scdp->scd_next->scd_prev = scdp->scd_prev;
14662 	}
14663 }
14664 
14665 /*
14666  * Add an scd to the start of the queue.
14667  */
14668 static void
14669 sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *scdp)
14670 {
14671 	scdp->scd_prev = NULL;
14672 	scdp->scd_next = *headp;
14673 	if (*headp != NULL) {
14674 		(*headp)->scd_prev = scdp;
14675 	}
14676 	*headp = scdp;
14677 }
14678 
14679 static int
14680 sfmmu_alloc_scd_tsbs(sf_srd_t *srdp, sf_scd_t *scdp)
14681 {
14682 	uint_t rid;
14683 	uint_t i;
14684 	uint_t j;
14685 	ulong_t w;
14686 	sf_region_t *rgnp;
14687 	ulong_t tte8k_cnt = 0;
14688 	ulong_t tte4m_cnt = 0;
14689 	uint_t tsb_szc;
14690 	sfmmu_t *scsfmmup = scdp->scd_sfmmup;
14691 	sfmmu_t	*ism_hatid;
14692 	struct tsb_info *newtsb;
14693 	int szc;
14694 
14695 	ASSERT(srdp != NULL);
14696 
14697 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14698 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14699 			continue;
14700 		}
14701 		j = 0;
14702 		while (w) {
14703 			if (!(w & 0x1)) {
14704 				j++;
14705 				w >>= 1;
14706 				continue;
14707 			}
14708 			rid = (i << BT_ULSHIFT) | j;
14709 			j++;
14710 			w >>= 1;
14711 
14712 			if (rid < SFMMU_MAX_HME_REGIONS) {
14713 				rgnp = srdp->srd_hmergnp[rid];
14714 				ASSERT(rgnp->rgn_id == rid);
14715 				ASSERT(rgnp->rgn_refcnt > 0);
14716 
14717 				if (rgnp->rgn_pgszc < TTE4M) {
14718 					tte8k_cnt += rgnp->rgn_size >>
14719 					    TTE_PAGE_SHIFT(TTE8K);
14720 				} else {
14721 					ASSERT(rgnp->rgn_pgszc >= TTE4M);
14722 					tte4m_cnt += rgnp->rgn_size >>
14723 					    TTE_PAGE_SHIFT(TTE4M);
14724 					/*
14725 					 * Inflate SCD tsb0 by preallocating
14726 					 * 1/4 8k ttecnt for 4M regions to
14727 					 * allow for lgpg alloc failure.
14728 					 */
14729 					tte8k_cnt += rgnp->rgn_size >>
14730 					    (TTE_PAGE_SHIFT(TTE8K) + 2);
14731 				}
14732 			} else {
14733 				rid -= SFMMU_MAX_HME_REGIONS;
14734 				rgnp = srdp->srd_ismrgnp[rid];
14735 				ASSERT(rgnp->rgn_id == rid);
14736 				ASSERT(rgnp->rgn_refcnt > 0);
14737 
14738 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14739 				ASSERT(ism_hatid->sfmmu_ismhat);
14740 
14741 				for (szc = 0; szc < TTE4M; szc++) {
14742 					tte8k_cnt +=
14743 					    ism_hatid->sfmmu_ttecnt[szc] <<
14744 					    TTE_BSZS_SHIFT(szc);
14745 				}
14746 
14747 				ASSERT(rgnp->rgn_pgszc >= TTE4M);
14748 				if (rgnp->rgn_pgszc >= TTE4M) {
14749 					tte4m_cnt += rgnp->rgn_size >>
14750 					    TTE_PAGE_SHIFT(TTE4M);
14751 				}
14752 			}
14753 		}
14754 	}
14755 
14756 	tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
14757 
14758 	/* Allocate both the SCD TSBs here. */
14759 	if (sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14760 	    tsb_szc, TSB8K|TSB64K|TSB512K, TSB_ALLOC, scsfmmup) &&
14761 	    (tsb_szc <= TSB_4M_SZCODE ||
14762 	    sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14763 	    TSB_4M_SZCODE, TSB8K|TSB64K|TSB512K,
14764 	    TSB_ALLOC, scsfmmup))) {
14765 
14766 		SFMMU_STAT(sf_scd_1sttsb_allocfail);
14767 		return (TSB_ALLOCFAIL);
14768 	} else {
14769 		scsfmmup->sfmmu_tsb->tsb_flags |= TSB_SHAREDCTX;
14770 
14771 		if (tte4m_cnt) {
14772 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
14773 			if (sfmmu_tsbinfo_alloc(&newtsb, tsb_szc,
14774 			    TSB4M|TSB32M|TSB256M, TSB_ALLOC, scsfmmup) &&
14775 			    (tsb_szc <= TSB_4M_SZCODE ||
14776 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
14777 			    TSB4M|TSB32M|TSB256M,
14778 			    TSB_ALLOC, scsfmmup))) {
14779 				/*
14780 				 * If we fail to allocate the 2nd shared tsb,
14781 				 * just free the 1st tsb, return failure.
14782 				 */
14783 				sfmmu_tsbinfo_free(scsfmmup->sfmmu_tsb);
14784 				SFMMU_STAT(sf_scd_2ndtsb_allocfail);
14785 				return (TSB_ALLOCFAIL);
14786 			} else {
14787 				ASSERT(scsfmmup->sfmmu_tsb->tsb_next == NULL);
14788 				newtsb->tsb_flags |= TSB_SHAREDCTX;
14789 				scsfmmup->sfmmu_tsb->tsb_next = newtsb;
14790 				SFMMU_STAT(sf_scd_2ndtsb_alloc);
14791 			}
14792 		}
14793 		SFMMU_STAT(sf_scd_1sttsb_alloc);
14794 	}
14795 	return (TSB_SUCCESS);
14796 }
14797 
14798 static void
14799 sfmmu_free_scd_tsbs(sfmmu_t *scd_sfmmu)
14800 {
14801 	while (scd_sfmmu->sfmmu_tsb != NULL) {
14802 		struct tsb_info *next = scd_sfmmu->sfmmu_tsb->tsb_next;
14803 		sfmmu_tsbinfo_free(scd_sfmmu->sfmmu_tsb);
14804 		scd_sfmmu->sfmmu_tsb = next;
14805 	}
14806 }
14807 
14808 /*
14809  * Link the sfmmu onto the hme region list.
14810  */
14811 void
14812 sfmmu_link_to_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14813 {
14814 	uint_t rid;
14815 	sf_rgn_link_t *rlink;
14816 	sfmmu_t *head;
14817 	sf_rgn_link_t *hrlink;
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, 1, 1);
14824 	ASSERT(rlink != NULL);
14825 	mutex_enter(&rgnp->rgn_mutex);
14826 	if ((head = rgnp->rgn_sfmmu_head) == NULL) {
14827 		rlink->next = NULL;
14828 		rlink->prev = NULL;
14829 		/*
14830 		 * make sure rlink's next field is NULL
14831 		 * before making this link visible.
14832 		 */
14833 		membar_stst();
14834 		rgnp->rgn_sfmmu_head = sfmmup;
14835 	} else {
14836 		/* LINTED: constant in conditional context */
14837 		SFMMU_HMERID2RLINKP(head, rid, hrlink, 0, 0);
14838 		ASSERT(hrlink != NULL);
14839 		ASSERT(hrlink->prev == NULL);
14840 		rlink->next = head;
14841 		rlink->prev = NULL;
14842 		hrlink->prev = sfmmup;
14843 		/*
14844 		 * make sure rlink's next field is correct
14845 		 * before making this link visible.
14846 		 */
14847 		membar_stst();
14848 		rgnp->rgn_sfmmu_head = sfmmup;
14849 	}
14850 	mutex_exit(&rgnp->rgn_mutex);
14851 }
14852 
14853 /*
14854  * Unlink the sfmmu from the hme region list.
14855  */
14856 void
14857 sfmmu_unlink_from_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14858 {
14859 	uint_t rid;
14860 	sf_rgn_link_t *rlink;
14861 
14862 	rid = rgnp->rgn_id;
14863 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14864 
14865 	/* LINTED: constant in conditional context */
14866 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
14867 	ASSERT(rlink != NULL);
14868 	mutex_enter(&rgnp->rgn_mutex);
14869 	if (rgnp->rgn_sfmmu_head == sfmmup) {
14870 		sfmmu_t *next = rlink->next;
14871 		rgnp->rgn_sfmmu_head = next;
14872 		/*
14873 		 * if we are stopped by xc_attention() after this
14874 		 * point the forward link walking in
14875 		 * sfmmu_rgntlb_demap() will work correctly since the
14876 		 * head correctly points to the next element.
14877 		 */
14878 		membar_stst();
14879 		rlink->next = NULL;
14880 		ASSERT(rlink->prev == NULL);
14881 		if (next != NULL) {
14882 			sf_rgn_link_t *nrlink;
14883 			/* LINTED: constant in conditional context */
14884 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14885 			ASSERT(nrlink != NULL);
14886 			ASSERT(nrlink->prev == sfmmup);
14887 			nrlink->prev = NULL;
14888 		}
14889 	} else {
14890 		sfmmu_t *next = rlink->next;
14891 		sfmmu_t *prev = rlink->prev;
14892 		sf_rgn_link_t *prlink;
14893 
14894 		ASSERT(prev != NULL);
14895 		/* LINTED: constant in conditional context */
14896 		SFMMU_HMERID2RLINKP(prev, rid, prlink, 0, 0);
14897 		ASSERT(prlink != NULL);
14898 		ASSERT(prlink->next == sfmmup);
14899 		prlink->next = next;
14900 		/*
14901 		 * if we are stopped by xc_attention()
14902 		 * after this point the forward link walking
14903 		 * will work correctly since the prev element
14904 		 * correctly points to the next element.
14905 		 */
14906 		membar_stst();
14907 		rlink->next = NULL;
14908 		rlink->prev = NULL;
14909 		if (next != NULL) {
14910 			sf_rgn_link_t *nrlink;
14911 			/* LINTED: constant in conditional context */
14912 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14913 			ASSERT(nrlink != NULL);
14914 			ASSERT(nrlink->prev == sfmmup);
14915 			nrlink->prev = prev;
14916 		}
14917 	}
14918 	mutex_exit(&rgnp->rgn_mutex);
14919 }
14920 
14921 /*
14922  * Link scd sfmmu onto ism or hme region list for each region in the
14923  * scd region map.
14924  */
14925 void
14926 sfmmu_link_scd_to_regions(sf_srd_t *srdp, sf_scd_t *scdp)
14927 {
14928 	uint_t rid;
14929 	uint_t i;
14930 	uint_t j;
14931 	ulong_t w;
14932 	sf_region_t *rgnp;
14933 	sfmmu_t *scsfmmup;
14934 
14935 	scsfmmup = scdp->scd_sfmmup;
14936 	ASSERT(scsfmmup->sfmmu_scdhat);
14937 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14938 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14939 			continue;
14940 		}
14941 		j = 0;
14942 		while (w) {
14943 			if (!(w & 0x1)) {
14944 				j++;
14945 				w >>= 1;
14946 				continue;
14947 			}
14948 			rid = (i << BT_ULSHIFT) | j;
14949 			j++;
14950 			w >>= 1;
14951 
14952 			if (rid < SFMMU_MAX_HME_REGIONS) {
14953 				rgnp = srdp->srd_hmergnp[rid];
14954 				ASSERT(rgnp->rgn_id == rid);
14955 				ASSERT(rgnp->rgn_refcnt > 0);
14956 				sfmmu_link_to_hmeregion(scsfmmup, rgnp);
14957 			} else {
14958 				sfmmu_t *ism_hatid = NULL;
14959 				ism_ment_t *ism_ment;
14960 				rid -= SFMMU_MAX_HME_REGIONS;
14961 				rgnp = srdp->srd_ismrgnp[rid];
14962 				ASSERT(rgnp->rgn_id == rid);
14963 				ASSERT(rgnp->rgn_refcnt > 0);
14964 
14965 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14966 				ASSERT(ism_hatid->sfmmu_ismhat);
14967 				ism_ment = &scdp->scd_ism_links[rid];
14968 				ism_ment->iment_hat = scsfmmup;
14969 				ism_ment->iment_base_va = rgnp->rgn_saddr;
14970 				mutex_enter(&ism_mlist_lock);
14971 				iment_add(ism_ment, ism_hatid);
14972 				mutex_exit(&ism_mlist_lock);
14973 
14974 			}
14975 		}
14976 	}
14977 }
14978 /*
14979  * Unlink scd sfmmu from ism or hme region list for each region in the
14980  * scd region map.
14981  */
14982 void
14983 sfmmu_unlink_scd_from_regions(sf_srd_t *srdp, sf_scd_t *scdp)
14984 {
14985 	uint_t rid;
14986 	uint_t i;
14987 	uint_t j;
14988 	ulong_t w;
14989 	sf_region_t *rgnp;
14990 	sfmmu_t *scsfmmup;
14991 
14992 	scsfmmup = scdp->scd_sfmmup;
14993 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14994 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14995 			continue;
14996 		}
14997 		j = 0;
14998 		while (w) {
14999 			if (!(w & 0x1)) {
15000 				j++;
15001 				w >>= 1;
15002 				continue;
15003 			}
15004 			rid = (i << BT_ULSHIFT) | j;
15005 			j++;
15006 			w >>= 1;
15007 
15008 			if (rid < SFMMU_MAX_HME_REGIONS) {
15009 				rgnp = srdp->srd_hmergnp[rid];
15010 				ASSERT(rgnp->rgn_id == rid);
15011 				ASSERT(rgnp->rgn_refcnt > 0);
15012 				sfmmu_unlink_from_hmeregion(scsfmmup,
15013 				    rgnp);
15014 
15015 			} else {
15016 				sfmmu_t *ism_hatid = NULL;
15017 				ism_ment_t *ism_ment;
15018 				rid -= SFMMU_MAX_HME_REGIONS;
15019 				rgnp = srdp->srd_ismrgnp[rid];
15020 				ASSERT(rgnp->rgn_id == rid);
15021 				ASSERT(rgnp->rgn_refcnt > 0);
15022 
15023 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
15024 				ASSERT(ism_hatid->sfmmu_ismhat);
15025 				ism_ment = &scdp->scd_ism_links[rid];
15026 				ASSERT(ism_ment->iment_hat == scdp->scd_sfmmup);
15027 				ASSERT(ism_ment->iment_base_va ==
15028 				    rgnp->rgn_saddr);
15029 				ism_ment->iment_hat = NULL;
15030 				ism_ment->iment_base_va = 0;
15031 				mutex_enter(&ism_mlist_lock);
15032 				iment_sub(ism_ment, ism_hatid);
15033 				mutex_exit(&ism_mlist_lock);
15034 
15035 			}
15036 		}
15037 	}
15038 }
15039 /*
15040  * Allocates and initialises a new SCD structure, this is called with
15041  * the srd_scd_mutex held and returns with the reference count
15042  * initialised to 1.
15043  */
15044 static sf_scd_t *
15045 sfmmu_alloc_scd(sf_srd_t *srdp, sf_region_map_t *new_map)
15046 {
15047 	sf_scd_t *new_scdp;
15048 	sfmmu_t *scsfmmup;
15049 	int i;
15050 
15051 	ASSERT(MUTEX_HELD(&srdp->srd_scd_mutex));
15052 	new_scdp = kmem_cache_alloc(scd_cache, KM_SLEEP);
15053 
15054 	scsfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
15055 	new_scdp->scd_sfmmup = scsfmmup;
15056 	scsfmmup->sfmmu_srdp = srdp;
15057 	scsfmmup->sfmmu_scdp = new_scdp;
15058 	scsfmmup->sfmmu_tsb0_4minflcnt = 0;
15059 	scsfmmup->sfmmu_scdhat = 1;
15060 	CPUSET_ALL(scsfmmup->sfmmu_cpusran);
15061 	bzero(scsfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
15062 
15063 	ASSERT(max_mmu_ctxdoms > 0);
15064 	for (i = 0; i < max_mmu_ctxdoms; i++) {
15065 		scsfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
15066 		scsfmmup->sfmmu_ctxs[i].gnum = 0;
15067 	}
15068 
15069 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
15070 		new_scdp->scd_rttecnt[i] = 0;
15071 	}
15072 
15073 	new_scdp->scd_region_map = *new_map;
15074 	new_scdp->scd_refcnt = 1;
15075 	if (sfmmu_alloc_scd_tsbs(srdp, new_scdp) != TSB_SUCCESS) {
15076 		kmem_cache_free(scd_cache, new_scdp);
15077 		kmem_cache_free(sfmmuid_cache, scsfmmup);
15078 		return (NULL);
15079 	}
15080 	if (&mmu_init_scd) {
15081 		mmu_init_scd(new_scdp);
15082 	}
15083 	return (new_scdp);
15084 }
15085 
15086 /*
15087  * The first phase of a process joining an SCD. The hat structure is
15088  * linked to the SCD queue and then the HAT_JOIN_SCD sfmmu flag is set
15089  * and a cross-call with context invalidation is used to cause the
15090  * remaining work to be carried out in the sfmmu_tsbmiss_exception()
15091  * routine.
15092  */
15093 static void
15094 sfmmu_join_scd(sf_scd_t *scdp, sfmmu_t *sfmmup)
15095 {
15096 	hatlock_t *hatlockp;
15097 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15098 	int i;
15099 	sf_scd_t *old_scdp;
15100 
15101 	ASSERT(srdp != NULL);
15102 	ASSERT(scdp != NULL);
15103 	ASSERT(scdp->scd_refcnt > 0);
15104 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15105 
15106 	if ((old_scdp = sfmmup->sfmmu_scdp) != NULL) {
15107 		ASSERT(old_scdp != scdp);
15108 
15109 		mutex_enter(&old_scdp->scd_mutex);
15110 		sfmmu_from_scd_list(&old_scdp->scd_sf_list, sfmmup);
15111 		mutex_exit(&old_scdp->scd_mutex);
15112 		/*
15113 		 * sfmmup leaves the old scd. Update sfmmu_ttecnt to
15114 		 * include the shme rgn ttecnt for rgns that
15115 		 * were in the old SCD
15116 		 */
15117 		for (i = 0; i < mmu_page_sizes; i++) {
15118 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15119 			    old_scdp->scd_rttecnt[i]);
15120 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15121 			    sfmmup->sfmmu_scdrttecnt[i]);
15122 		}
15123 	}
15124 
15125 	/*
15126 	 * Move sfmmu to the scd lists.
15127 	 */
15128 	mutex_enter(&scdp->scd_mutex);
15129 	sfmmu_to_scd_list(&scdp->scd_sf_list, sfmmup);
15130 	mutex_exit(&scdp->scd_mutex);
15131 	SF_SCD_INCR_REF(scdp);
15132 
15133 	hatlockp = sfmmu_hat_enter(sfmmup);
15134 	/*
15135 	 * For a multi-thread process, we must stop
15136 	 * all the other threads before joining the scd.
15137 	 */
15138 
15139 	SFMMU_FLAGS_SET(sfmmup, HAT_JOIN_SCD);
15140 
15141 	sfmmu_invalidate_ctx(sfmmup);
15142 	sfmmup->sfmmu_scdp = scdp;
15143 
15144 	/*
15145 	 * Copy scd_rttecnt into sfmmup's sfmmu_scdrttecnt, and update
15146 	 * sfmmu_ttecnt to not include the rgn ttecnt just joined in SCD.
15147 	 */
15148 	for (i = 0; i < mmu_page_sizes; i++) {
15149 		sfmmup->sfmmu_scdrttecnt[i] = scdp->scd_rttecnt[i];
15150 		ASSERT(sfmmup->sfmmu_ttecnt[i] >= scdp->scd_rttecnt[i]);
15151 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15152 		    -sfmmup->sfmmu_scdrttecnt[i]);
15153 	}
15154 	/* update tsb0 inflation count */
15155 	if (old_scdp != NULL) {
15156 		sfmmup->sfmmu_tsb0_4minflcnt +=
15157 		    old_scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15158 	}
15159 	ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
15160 	    scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt);
15161 	sfmmup->sfmmu_tsb0_4minflcnt -= scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15162 
15163 	sfmmu_hat_exit(hatlockp);
15164 
15165 	if (old_scdp != NULL) {
15166 		SF_SCD_DECR_REF(srdp, old_scdp);
15167 	}
15168 
15169 }
15170 
15171 /*
15172  * This routine is called by a process to become part of an SCD. It is called
15173  * from sfmmu_tsbmiss_exception() once most of the initial work has been
15174  * done by sfmmu_join_scd(). This routine must not drop the hat lock.
15175  */
15176 static void
15177 sfmmu_finish_join_scd(sfmmu_t *sfmmup)
15178 {
15179 	struct tsb_info	*tsbinfop;
15180 
15181 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15182 	ASSERT(sfmmup->sfmmu_scdp != NULL);
15183 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD));
15184 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15185 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID));
15186 
15187 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
15188 	    tsbinfop = tsbinfop->tsb_next) {
15189 		if (tsbinfop->tsb_flags & TSB_SWAPPED) {
15190 			continue;
15191 		}
15192 		ASSERT(!(tsbinfop->tsb_flags & TSB_RELOC_FLAG));
15193 
15194 		sfmmu_inv_tsb(tsbinfop->tsb_va,
15195 		    TSB_BYTES(tsbinfop->tsb_szc));
15196 	}
15197 
15198 	/* Set HAT_CTX1_FLAG for all SCD ISMs */
15199 	sfmmu_ism_hatflags(sfmmup, 1);
15200 
15201 	SFMMU_STAT(sf_join_scd);
15202 }
15203 
15204 /*
15205  * This routine is called in order to check if there is an SCD which matches
15206  * the process's region map if not then a new SCD may be created.
15207  */
15208 static void
15209 sfmmu_find_scd(sfmmu_t *sfmmup)
15210 {
15211 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15212 	sf_scd_t *scdp, *new_scdp;
15213 	int ret;
15214 
15215 	ASSERT(srdp != NULL);
15216 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15217 
15218 	mutex_enter(&srdp->srd_scd_mutex);
15219 	for (scdp = srdp->srd_scdp; scdp != NULL;
15220 	    scdp = scdp->scd_next) {
15221 		SF_RGNMAP_EQUAL(&scdp->scd_region_map,
15222 		    &sfmmup->sfmmu_region_map, ret);
15223 		if (ret == 1) {
15224 			SF_SCD_INCR_REF(scdp);
15225 			mutex_exit(&srdp->srd_scd_mutex);
15226 			sfmmu_join_scd(scdp, sfmmup);
15227 			ASSERT(scdp->scd_refcnt >= 2);
15228 			atomic_add_32((volatile uint32_t *)
15229 			    &scdp->scd_refcnt, -1);
15230 			return;
15231 		} else {
15232 			/*
15233 			 * If the sfmmu region map is a subset of the scd
15234 			 * region map, then the assumption is that this process
15235 			 * will continue attaching to ISM segments until the
15236 			 * region maps are equal.
15237 			 */
15238 			SF_RGNMAP_IS_SUBSET(&scdp->scd_region_map,
15239 			    &sfmmup->sfmmu_region_map, ret);
15240 			if (ret == 1) {
15241 				mutex_exit(&srdp->srd_scd_mutex);
15242 				return;
15243 			}
15244 		}
15245 	}
15246 
15247 	ASSERT(scdp == NULL);
15248 	/*
15249 	 * No matching SCD has been found, create a new one.
15250 	 */
15251 	if ((new_scdp = sfmmu_alloc_scd(srdp, &sfmmup->sfmmu_region_map)) ==
15252 	    NULL) {
15253 		mutex_exit(&srdp->srd_scd_mutex);
15254 		return;
15255 	}
15256 
15257 	/*
15258 	 * sfmmu_alloc_scd() returns with a ref count of 1 on the scd.
15259 	 */
15260 
15261 	/* Set scd_rttecnt for shme rgns in SCD */
15262 	sfmmu_set_scd_rttecnt(srdp, new_scdp);
15263 
15264 	/*
15265 	 * Link scd onto srd_scdp list and scd sfmmu onto region/iment lists.
15266 	 */
15267 	sfmmu_link_scd_to_regions(srdp, new_scdp);
15268 	sfmmu_add_scd(&srdp->srd_scdp, new_scdp);
15269 	SFMMU_STAT_ADD(sf_create_scd, 1);
15270 
15271 	mutex_exit(&srdp->srd_scd_mutex);
15272 	sfmmu_join_scd(new_scdp, sfmmup);
15273 	ASSERT(new_scdp->scd_refcnt >= 2);
15274 	atomic_add_32((volatile uint32_t *)&new_scdp->scd_refcnt, -1);
15275 }
15276 
15277 /*
15278  * This routine is called by a process to remove itself from an SCD. It is
15279  * either called when the processes has detached from a segment or from
15280  * hat_free_start() as a result of calling exit.
15281  */
15282 static void
15283 sfmmu_leave_scd(sfmmu_t *sfmmup, uchar_t r_type)
15284 {
15285 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15286 	sf_srd_t *srdp =  sfmmup->sfmmu_srdp;
15287 	hatlock_t *hatlockp = TSB_HASH(sfmmup);
15288 	int i;
15289 
15290 	ASSERT(scdp != NULL);
15291 	ASSERT(srdp != NULL);
15292 
15293 	if (sfmmup->sfmmu_free) {
15294 		/*
15295 		 * If the process is part of an SCD the sfmmu is unlinked
15296 		 * from scd_sf_list.
15297 		 */
15298 		mutex_enter(&scdp->scd_mutex);
15299 		sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15300 		mutex_exit(&scdp->scd_mutex);
15301 		/*
15302 		 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15303 		 * are about to leave the SCD
15304 		 */
15305 		for (i = 0; i < mmu_page_sizes; i++) {
15306 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15307 			    scdp->scd_rttecnt[i]);
15308 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15309 			    sfmmup->sfmmu_scdrttecnt[i]);
15310 			sfmmup->sfmmu_scdrttecnt[i] = 0;
15311 		}
15312 		sfmmup->sfmmu_scdp = NULL;
15313 
15314 		SF_SCD_DECR_REF(srdp, scdp);
15315 		return;
15316 	}
15317 
15318 	ASSERT(r_type != SFMMU_REGION_ISM ||
15319 	    SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15320 	ASSERT(scdp->scd_refcnt);
15321 	ASSERT(!sfmmup->sfmmu_free);
15322 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15323 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15324 
15325 	/*
15326 	 * Wait for ISM maps to be updated.
15327 	 */
15328 	if (r_type != SFMMU_REGION_ISM) {
15329 		while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY) &&
15330 		    sfmmup->sfmmu_scdp != NULL) {
15331 			cv_wait(&sfmmup->sfmmu_tsb_cv,
15332 			    HATLOCK_MUTEXP(hatlockp));
15333 		}
15334 
15335 		if (sfmmup->sfmmu_scdp == NULL) {
15336 			sfmmu_hat_exit(hatlockp);
15337 			return;
15338 		}
15339 		SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
15340 	}
15341 
15342 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
15343 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
15344 		/*
15345 		 * Since HAT_JOIN_SCD was set our context
15346 		 * is still invalid.
15347 		 */
15348 	} else {
15349 		/*
15350 		 * For a multi-thread process, we must stop
15351 		 * all the other threads before leaving the scd.
15352 		 */
15353 
15354 		sfmmu_invalidate_ctx(sfmmup);
15355 	}
15356 
15357 	/* Clear all the rid's for ISM, delete flags, etc */
15358 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15359 	sfmmu_ism_hatflags(sfmmup, 0);
15360 
15361 	/*
15362 	 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15363 	 * are in SCD before this sfmmup leaves the SCD.
15364 	 */
15365 	for (i = 0; i < mmu_page_sizes; i++) {
15366 		ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15367 		    scdp->scd_rttecnt[i]);
15368 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15369 		    sfmmup->sfmmu_scdrttecnt[i]);
15370 		sfmmup->sfmmu_scdrttecnt[i] = 0;
15371 		/* update ismttecnt to include SCD ism before hat leaves SCD */
15372 		sfmmup->sfmmu_ismttecnt[i] += sfmmup->sfmmu_scdismttecnt[i];
15373 		sfmmup->sfmmu_scdismttecnt[i] = 0;
15374 	}
15375 	/* update tsb0 inflation count */
15376 	sfmmup->sfmmu_tsb0_4minflcnt += scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15377 
15378 	if (r_type != SFMMU_REGION_ISM) {
15379 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
15380 	}
15381 	sfmmup->sfmmu_scdp = NULL;
15382 
15383 	sfmmu_hat_exit(hatlockp);
15384 
15385 	/*
15386 	 * Unlink sfmmu from scd_sf_list this can be done without holding
15387 	 * the hat lock as we hold the sfmmu_as lock which prevents
15388 	 * hat_join_region from adding this thread to the scd again. Other
15389 	 * threads check if sfmmu_scdp is NULL under hat lock and if it's NULL
15390 	 * they won't get here, since sfmmu_leave_scd() clears sfmmu_scdp
15391 	 * while holding the hat lock.
15392 	 */
15393 	mutex_enter(&scdp->scd_mutex);
15394 	sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15395 	mutex_exit(&scdp->scd_mutex);
15396 	SFMMU_STAT(sf_leave_scd);
15397 
15398 	SF_SCD_DECR_REF(srdp, scdp);
15399 	hatlockp = sfmmu_hat_enter(sfmmup);
15400 
15401 }
15402 
15403 /*
15404  * Unlink and free up an SCD structure with a reference count of 0.
15405  */
15406 static void
15407 sfmmu_destroy_scd(sf_srd_t *srdp, sf_scd_t *scdp, sf_region_map_t *scd_rmap)
15408 {
15409 	sfmmu_t *scsfmmup;
15410 	sf_scd_t *sp;
15411 	hatlock_t *shatlockp;
15412 	int i, ret;
15413 
15414 	mutex_enter(&srdp->srd_scd_mutex);
15415 	for (sp = srdp->srd_scdp; sp != NULL; sp = sp->scd_next) {
15416 		if (sp == scdp)
15417 			break;
15418 	}
15419 	if (sp == NULL || sp->scd_refcnt) {
15420 		mutex_exit(&srdp->srd_scd_mutex);
15421 		return;
15422 	}
15423 
15424 	/*
15425 	 * It is possible that the scd has been freed and reallocated with a
15426 	 * different region map while we've been waiting for the srd_scd_mutex.
15427 	 */
15428 	SF_RGNMAP_EQUAL(scd_rmap, &sp->scd_region_map, ret);
15429 	if (ret != 1) {
15430 		mutex_exit(&srdp->srd_scd_mutex);
15431 		return;
15432 	}
15433 
15434 	ASSERT(scdp->scd_sf_list == NULL);
15435 	/*
15436 	 * Unlink scd from srd_scdp list.
15437 	 */
15438 	sfmmu_remove_scd(&srdp->srd_scdp, scdp);
15439 	mutex_exit(&srdp->srd_scd_mutex);
15440 
15441 	sfmmu_unlink_scd_from_regions(srdp, scdp);
15442 
15443 	/* Clear shared context tsb and release ctx */
15444 	scsfmmup = scdp->scd_sfmmup;
15445 
15446 	/*
15447 	 * create a barrier so that scd will not be destroyed
15448 	 * if other thread still holds the same shared hat lock.
15449 	 * E.g., sfmmu_tsbmiss_exception() needs to acquire the
15450 	 * shared hat lock before checking the shared tsb reloc flag.
15451 	 */
15452 	shatlockp = sfmmu_hat_enter(scsfmmup);
15453 	sfmmu_hat_exit(shatlockp);
15454 
15455 	sfmmu_free_scd_tsbs(scsfmmup);
15456 
15457 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
15458 		if (scsfmmup->sfmmu_hmeregion_links[i] != NULL) {
15459 			kmem_free(scsfmmup->sfmmu_hmeregion_links[i],
15460 			    SFMMU_L2_HMERLINKS_SIZE);
15461 			scsfmmup->sfmmu_hmeregion_links[i] = NULL;
15462 		}
15463 	}
15464 	kmem_cache_free(sfmmuid_cache, scsfmmup);
15465 	kmem_cache_free(scd_cache, scdp);
15466 	SFMMU_STAT(sf_destroy_scd);
15467 }
15468 
15469 /*
15470  * Modifies the HAT_CTX1_FLAG for each of the ISM segments which correspond to
15471  * bits which are set in the ism_region_map parameter. This flag indicates to
15472  * the tsbmiss handler that mapping for these segments should be loaded using
15473  * the shared context.
15474  */
15475 static void
15476 sfmmu_ism_hatflags(sfmmu_t *sfmmup, int addflag)
15477 {
15478 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15479 	ism_blk_t *ism_blkp;
15480 	ism_map_t *ism_map;
15481 	int i, rid;
15482 
15483 	ASSERT(sfmmup->sfmmu_iblk != NULL);
15484 	ASSERT(scdp != NULL);
15485 	/*
15486 	 * Note that the caller either set HAT_ISMBUSY flag or checked
15487 	 * under hat lock that HAT_ISMBUSY was not set by another thread.
15488 	 */
15489 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15490 
15491 	ism_blkp = sfmmup->sfmmu_iblk;
15492 	while (ism_blkp != NULL) {
15493 		ism_map = ism_blkp->iblk_maps;
15494 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
15495 			rid = ism_map[i].imap_rid;
15496 			if (rid == SFMMU_INVALID_ISMRID) {
15497 				continue;
15498 			}
15499 			ASSERT(rid >= 0 && rid < SFMMU_MAX_ISM_REGIONS);
15500 			if (SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid) &&
15501 			    addflag) {
15502 				ism_map[i].imap_hatflags |=
15503 				    HAT_CTX1_FLAG;
15504 			} else {
15505 				ism_map[i].imap_hatflags &=
15506 				    ~HAT_CTX1_FLAG;
15507 			}
15508 		}
15509 		ism_blkp = ism_blkp->iblk_next;
15510 	}
15511 }
15512 
15513 static int
15514 sfmmu_srd_lock_held(sf_srd_t *srdp)
15515 {
15516 	return (MUTEX_HELD(&srdp->srd_mutex));
15517 }
15518 
15519 /* ARGSUSED */
15520 static int
15521 sfmmu_scdcache_constructor(void *buf, void *cdrarg, int kmflags)
15522 {
15523 	sf_scd_t *scdp = (sf_scd_t *)buf;
15524 
15525 	bzero(buf, sizeof (sf_scd_t));
15526 	mutex_init(&scdp->scd_mutex, NULL, MUTEX_DEFAULT, NULL);
15527 	return (0);
15528 }
15529 
15530 /* ARGSUSED */
15531 static void
15532 sfmmu_scdcache_destructor(void *buf, void *cdrarg)
15533 {
15534 	sf_scd_t *scdp = (sf_scd_t *)buf;
15535 
15536 	mutex_destroy(&scdp->scd_mutex);
15537 }
15538 
15539 /*
15540  * The listp parameter is a pointer to a list of hmeblks which are partially
15541  * freed as result of calling sfmmu_hblk_hash_rm(), the last phase of the
15542  * freeing process is to cross-call all cpus to ensure that there are no
15543  * remaining cached references.
15544  *
15545  * If the local generation number is less than the global then we can free
15546  * hmeblks which are already on the pending queue as another cpu has completed
15547  * the cross-call.
15548  *
15549  * We cross-call to make sure that there are no threads on other cpus accessing
15550  * these hmblks and then complete the process of freeing them under the
15551  * following conditions:
15552  * 	The total number of pending hmeblks is greater than the threshold
15553  *	The reserve list has fewer than HBLK_RESERVE_CNT hmeblks
15554  *	It is at least 1 second since the last time we cross-called
15555  *
15556  * Otherwise, we add the hmeblks to the per-cpu pending queue.
15557  */
15558 static void
15559 sfmmu_hblks_list_purge(struct hme_blk **listp, int dontfree)
15560 {
15561 	struct hme_blk *hblkp, *pr_hblkp = NULL;
15562 	int		count = 0;
15563 	cpuset_t	cpuset = cpu_ready_set;
15564 	cpu_hme_pend_t	*cpuhp;
15565 	timestruc_t	now;
15566 	int		one_second_expired = 0;
15567 
15568 	gethrestime_lasttick(&now);
15569 
15570 	for (hblkp = *listp; hblkp != NULL; hblkp = hblkp->hblk_next) {
15571 		ASSERT(hblkp->hblk_shw_bit == 0);
15572 		ASSERT(hblkp->hblk_shared == 0);
15573 		count++;
15574 		pr_hblkp = hblkp;
15575 	}
15576 
15577 	cpuhp = &cpu_hme_pend[CPU->cpu_seqid];
15578 	mutex_enter(&cpuhp->chp_mutex);
15579 
15580 	if ((cpuhp->chp_count + count) == 0) {
15581 		mutex_exit(&cpuhp->chp_mutex);
15582 		return;
15583 	}
15584 
15585 	if ((now.tv_sec - cpuhp->chp_timestamp) > 1) {
15586 		one_second_expired  = 1;
15587 	}
15588 
15589 	if (!dontfree && (freehblkcnt < HBLK_RESERVE_CNT ||
15590 	    (cpuhp->chp_count + count) > cpu_hme_pend_thresh ||
15591 	    one_second_expired)) {
15592 		/* Append global list to local */
15593 		if (pr_hblkp == NULL) {
15594 			*listp = cpuhp->chp_listp;
15595 		} else {
15596 			pr_hblkp->hblk_next = cpuhp->chp_listp;
15597 		}
15598 		cpuhp->chp_listp = NULL;
15599 		cpuhp->chp_count = 0;
15600 		cpuhp->chp_timestamp = now.tv_sec;
15601 		mutex_exit(&cpuhp->chp_mutex);
15602 
15603 		kpreempt_disable();
15604 		CPUSET_DEL(cpuset, CPU->cpu_id);
15605 		xt_sync(cpuset);
15606 		xt_sync(cpuset);
15607 		kpreempt_enable();
15608 
15609 		/*
15610 		 * At this stage we know that no trap handlers on other
15611 		 * cpus can have references to hmeblks on the list.
15612 		 */
15613 		sfmmu_hblk_free(listp);
15614 	} else if (*listp != NULL) {
15615 		pr_hblkp->hblk_next = cpuhp->chp_listp;
15616 		cpuhp->chp_listp = *listp;
15617 		cpuhp->chp_count += count;
15618 		*listp = NULL;
15619 		mutex_exit(&cpuhp->chp_mutex);
15620 	} else {
15621 		mutex_exit(&cpuhp->chp_mutex);
15622 	}
15623 }
15624 
15625 /*
15626  * Add an hmeblk to the the hash list.
15627  */
15628 void
15629 sfmmu_hblk_hash_add(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
15630 	uint64_t hblkpa)
15631 {
15632 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
15633 #ifdef	DEBUG
15634 	if (hmebp->hmeblkp == NULL) {
15635 		ASSERT(hmebp->hmeh_nextpa == HMEBLK_ENDPA);
15636 	}
15637 #endif /* DEBUG */
15638 
15639 	hmeblkp->hblk_nextpa = hmebp->hmeh_nextpa;
15640 	/*
15641 	 * Since the TSB miss handler now does not lock the hash chain before
15642 	 * walking it, make sure that the hmeblks nextpa is globally visible
15643 	 * before we make the hmeblk globally visible by updating the chain root
15644 	 * pointer in the hash bucket.
15645 	 */
15646 	membar_producer();
15647 	hmebp->hmeh_nextpa = hblkpa;
15648 	hmeblkp->hblk_next = hmebp->hmeblkp;
15649 	hmebp->hmeblkp = hmeblkp;
15650 
15651 }
15652 
15653 /*
15654  * This function is the first part of a 2 part process to remove an hmeblk
15655  * from the hash chain. In this phase we unlink the hmeblk from the hash chain
15656  * but leave the next physical pointer unchanged. The hmeblk is then linked onto
15657  * a per-cpu pending list using the virtual address pointer.
15658  *
15659  * TSB miss trap handlers that start after this phase will no longer see
15660  * this hmeblk. TSB miss handlers that still cache this hmeblk in a register
15661  * can still use it for further chain traversal because we haven't yet modifed
15662  * the next physical pointer or freed it.
15663  *
15664  * In the second phase of hmeblk removal we'll issue a barrier xcall before
15665  * we reuse or free this hmeblk. This will make sure all lingering references to
15666  * the hmeblk after first phase disappear before we finally reclaim it.
15667  * This scheme eliminates the need for TSB miss handlers to lock hmeblk chains
15668  * during their traversal.
15669  *
15670  * The hmehash_mutex must be held when calling this function.
15671  *
15672  * Input:
15673  *	 hmebp - hme hash bucket pointer
15674  *	 hmeblkp - address of hmeblk to be removed
15675  *	 pr_hblk - virtual address of previous hmeblkp
15676  *	 listp - pointer to list of hmeblks linked by virtual address
15677  *	 free_now flag - indicates that a complete removal from the hash chains
15678  *			 is necessary.
15679  *
15680  * It is inefficient to use the free_now flag as a cross-call is required to
15681  * remove a single hmeblk from the hash chain but is necessary when hmeblks are
15682  * in short supply.
15683  */
15684 void
15685 sfmmu_hblk_hash_rm(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
15686     struct hme_blk *pr_hblk, struct hme_blk **listp,
15687     int free_now)
15688 {
15689 	int shw_size, vshift;
15690 	struct hme_blk *shw_hblkp;
15691 	uint_t		shw_mask, newshw_mask;
15692 	caddr_t		vaddr;
15693 	int		size;
15694 	cpuset_t cpuset = cpu_ready_set;
15695 
15696 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
15697 
15698 	if (hmebp->hmeblkp == hmeblkp) {
15699 		hmebp->hmeh_nextpa = hmeblkp->hblk_nextpa;
15700 		hmebp->hmeblkp = hmeblkp->hblk_next;
15701 	} else {
15702 		pr_hblk->hblk_nextpa = hmeblkp->hblk_nextpa;
15703 		pr_hblk->hblk_next = hmeblkp->hblk_next;
15704 	}
15705 
15706 	size = get_hblk_ttesz(hmeblkp);
15707 	shw_hblkp = hmeblkp->hblk_shadow;
15708 	if (shw_hblkp) {
15709 		ASSERT(hblktosfmmu(hmeblkp) != KHATID);
15710 		ASSERT(!hmeblkp->hblk_shared);
15711 #ifdef	DEBUG
15712 		if (mmu_page_sizes == max_mmu_page_sizes) {
15713 			ASSERT(size < TTE256M);
15714 		} else {
15715 			ASSERT(size < TTE4M);
15716 		}
15717 #endif /* DEBUG */
15718 
15719 		shw_size = get_hblk_ttesz(shw_hblkp);
15720 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
15721 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
15722 		ASSERT(vshift < 8);
15723 		/*
15724 		 * Atomically clear shadow mask bit
15725 		 */
15726 		do {
15727 			shw_mask = shw_hblkp->hblk_shw_mask;
15728 			ASSERT(shw_mask & (1 << vshift));
15729 			newshw_mask = shw_mask & ~(1 << vshift);
15730 			newshw_mask = cas32(&shw_hblkp->hblk_shw_mask,
15731 			    shw_mask, newshw_mask);
15732 		} while (newshw_mask != shw_mask);
15733 		hmeblkp->hblk_shadow = NULL;
15734 	}
15735 	hmeblkp->hblk_shw_bit = 0;
15736 
15737 	if (hmeblkp->hblk_shared) {
15738 #ifdef	DEBUG
15739 		sf_srd_t	*srdp;
15740 		sf_region_t	*rgnp;
15741 		uint_t		rid;
15742 
15743 		srdp = hblktosrd(hmeblkp);
15744 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
15745 		rid = hmeblkp->hblk_tag.htag_rid;
15746 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
15747 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
15748 		rgnp = srdp->srd_hmergnp[rid];
15749 		ASSERT(rgnp != NULL);
15750 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
15751 #endif /* DEBUG */
15752 		hmeblkp->hblk_shared = 0;
15753 	}
15754 	if (free_now) {
15755 		kpreempt_disable();
15756 		CPUSET_DEL(cpuset, CPU->cpu_id);
15757 		xt_sync(cpuset);
15758 		xt_sync(cpuset);
15759 		kpreempt_enable();
15760 
15761 		hmeblkp->hblk_nextpa = HMEBLK_ENDPA;
15762 		hmeblkp->hblk_next = NULL;
15763 	} else {
15764 		/* Append hmeblkp to listp for processing later. */
15765 		hmeblkp->hblk_next = *listp;
15766 		*listp = hmeblkp;
15767 	}
15768 }
15769 
15770 /*
15771  * This routine is called when memory is in short supply and returns a free
15772  * hmeblk of the requested size from the cpu pending lists.
15773  */
15774 static struct hme_blk *
15775 sfmmu_check_pending_hblks(int size)
15776 {
15777 	int i;
15778 	struct hme_blk *hmeblkp = NULL, *last_hmeblkp;
15779 	int found_hmeblk;
15780 	cpuset_t cpuset = cpu_ready_set;
15781 	cpu_hme_pend_t *cpuhp;
15782 
15783 	/* Flush cpu hblk pending queues */
15784 	for (i = 0; i < NCPU; i++) {
15785 		cpuhp = &cpu_hme_pend[i];
15786 		if (cpuhp->chp_listp != NULL)  {
15787 			mutex_enter(&cpuhp->chp_mutex);
15788 			if (cpuhp->chp_listp == NULL)  {
15789 				mutex_exit(&cpuhp->chp_mutex);
15790 				continue;
15791 			}
15792 			found_hmeblk = 0;
15793 			last_hmeblkp = NULL;
15794 			for (hmeblkp = cpuhp->chp_listp; hmeblkp != NULL;
15795 			    hmeblkp = hmeblkp->hblk_next) {
15796 				if (get_hblk_ttesz(hmeblkp) == size) {
15797 					if (last_hmeblkp == NULL) {
15798 						cpuhp->chp_listp =
15799 						    hmeblkp->hblk_next;
15800 					} else {
15801 						last_hmeblkp->hblk_next =
15802 						    hmeblkp->hblk_next;
15803 					}
15804 					ASSERT(cpuhp->chp_count > 0);
15805 					cpuhp->chp_count--;
15806 					found_hmeblk = 1;
15807 					break;
15808 				} else {
15809 					last_hmeblkp = hmeblkp;
15810 				}
15811 			}
15812 			mutex_exit(&cpuhp->chp_mutex);
15813 
15814 			if (found_hmeblk) {
15815 				kpreempt_disable();
15816 				CPUSET_DEL(cpuset, CPU->cpu_id);
15817 				xt_sync(cpuset);
15818 				xt_sync(cpuset);
15819 				kpreempt_enable();
15820 				return (hmeblkp);
15821 			}
15822 		}
15823 	}
15824 	return (NULL);
15825 }
15826