xref: /titanic_41/usr/src/uts/sfmmu/vm/hat_sfmmu.c (revision fdc35dd8859c711d510ef5e9397204a1d98e23c5)
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 (c) 1993, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 /*
25  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
26  */
27 
28 /*
29  * VM - Hardware Address Translation management for Spitfire MMU.
30  *
31  * This file implements the machine specific hardware translation
32  * needed by the VM system.  The machine independent interface is
33  * described in <vm/hat.h> while the machine dependent interface
34  * and data structures are described in <vm/hat_sfmmu.h>.
35  *
36  * The hat layer manages the address translation hardware as a cache
37  * driven by calls from the higher levels in the VM system.
38  */
39 
40 #include <sys/types.h>
41 #include <sys/kstat.h>
42 #include <vm/hat.h>
43 #include <vm/hat_sfmmu.h>
44 #include <vm/page.h>
45 #include <sys/pte.h>
46 #include <sys/systm.h>
47 #include <sys/mman.h>
48 #include <sys/sysmacros.h>
49 #include <sys/machparam.h>
50 #include <sys/vtrace.h>
51 #include <sys/kmem.h>
52 #include <sys/mmu.h>
53 #include <sys/cmn_err.h>
54 #include <sys/cpu.h>
55 #include <sys/cpuvar.h>
56 #include <sys/debug.h>
57 #include <sys/lgrp.h>
58 #include <sys/archsystm.h>
59 #include <sys/machsystm.h>
60 #include <sys/vmsystm.h>
61 #include <vm/as.h>
62 #include <vm/seg.h>
63 #include <vm/seg_kp.h>
64 #include <vm/seg_kmem.h>
65 #include <vm/seg_kpm.h>
66 #include <vm/rm.h>
67 #include <sys/t_lock.h>
68 #include <sys/obpdefs.h>
69 #include <sys/vm_machparam.h>
70 #include <sys/var.h>
71 #include <sys/trap.h>
72 #include <sys/machtrap.h>
73 #include <sys/scb.h>
74 #include <sys/bitmap.h>
75 #include <sys/machlock.h>
76 #include <sys/membar.h>
77 #include <sys/atomic.h>
78 #include <sys/cpu_module.h>
79 #include <sys/prom_debug.h>
80 #include <sys/ksynch.h>
81 #include <sys/mem_config.h>
82 #include <sys/mem_cage.h>
83 #include <vm/vm_dep.h>
84 #include <vm/xhat_sfmmu.h>
85 #include <sys/fpu/fpusystm.h>
86 #include <vm/mach_kpm.h>
87 #include <sys/callb.h>
88 
89 #ifdef	DEBUG
90 #define	SFMMU_VALIDATE_HMERID(hat, rid, saddr, len)			\
91 	if (SFMMU_IS_SHMERID_VALID(rid)) {				\
92 		caddr_t _eaddr = (saddr) + (len);			\
93 		sf_srd_t *_srdp;					\
94 		sf_region_t *_rgnp;					\
95 		ASSERT((rid) < SFMMU_MAX_HME_REGIONS);			\
96 		ASSERT(SF_RGNMAP_TEST(hat->sfmmu_hmeregion_map, rid));	\
97 		ASSERT((hat) != ksfmmup);				\
98 		_srdp = (hat)->sfmmu_srdp;				\
99 		ASSERT(_srdp != NULL);					\
100 		ASSERT(_srdp->srd_refcnt != 0);				\
101 		_rgnp = _srdp->srd_hmergnp[(rid)];			\
102 		ASSERT(_rgnp != NULL && _rgnp->rgn_id == rid);		\
103 		ASSERT(_rgnp->rgn_refcnt != 0);				\
104 		ASSERT(!(_rgnp->rgn_flags & SFMMU_REGION_FREE));	\
105 		ASSERT((_rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) ==	\
106 		    SFMMU_REGION_HME);					\
107 		ASSERT((saddr) >= _rgnp->rgn_saddr);			\
108 		ASSERT((saddr) < _rgnp->rgn_saddr + _rgnp->rgn_size);	\
109 		ASSERT(_eaddr > _rgnp->rgn_saddr);			\
110 		ASSERT(_eaddr <= _rgnp->rgn_saddr + _rgnp->rgn_size);	\
111 	}
112 
113 #define	SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid) 	 	 \
114 {						 			 \
115 		caddr_t _hsva;						 \
116 		caddr_t _heva;						 \
117 		caddr_t _rsva;					 	 \
118 		caddr_t _reva;					 	 \
119 		int	_ttesz = get_hblk_ttesz(hmeblkp);		 \
120 		int	_flagtte;					 \
121 		ASSERT((srdp)->srd_refcnt != 0);			 \
122 		ASSERT((rid) < SFMMU_MAX_HME_REGIONS);			 \
123 		ASSERT((rgnp)->rgn_id == rid);				 \
124 		ASSERT(!((rgnp)->rgn_flags & SFMMU_REGION_FREE));	 \
125 		ASSERT(((rgnp)->rgn_flags & SFMMU_REGION_TYPE_MASK) ==	 \
126 		    SFMMU_REGION_HME);					 \
127 		ASSERT(_ttesz <= (rgnp)->rgn_pgszc);			 \
128 		_hsva = (caddr_t)get_hblk_base(hmeblkp);		 \
129 		_heva = get_hblk_endaddr(hmeblkp);			 \
130 		_rsva = (caddr_t)P2ALIGN(				 \
131 		    (uintptr_t)(rgnp)->rgn_saddr, HBLK_MIN_BYTES);	 \
132 		_reva = (caddr_t)P2ROUNDUP(				 \
133 		    (uintptr_t)((rgnp)->rgn_saddr + (rgnp)->rgn_size),	 \
134 		    HBLK_MIN_BYTES);					 \
135 		ASSERT(_hsva >= _rsva);				 	 \
136 		ASSERT(_hsva < _reva);				 	 \
137 		ASSERT(_heva > _rsva);				 	 \
138 		ASSERT(_heva <= _reva);				 	 \
139 		_flagtte = (_ttesz < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ :  \
140 			_ttesz;						 \
141 		ASSERT(rgnp->rgn_hmeflags & (0x1 << _flagtte));		 \
142 }
143 
144 #else /* DEBUG */
145 #define	SFMMU_VALIDATE_HMERID(hat, rid, addr, len)
146 #define	SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid)
147 #endif /* DEBUG */
148 
149 #if defined(SF_ERRATA_57)
150 extern caddr_t errata57_limit;
151 #endif
152 
153 #define	HME8BLK_SZ_RND		((roundup(HME8BLK_SZ, sizeof (int64_t))) /  \
154 				(sizeof (int64_t)))
155 #define	HBLK_RESERVE		((struct hme_blk *)hblk_reserve)
156 
157 #define	HBLK_RESERVE_CNT	128
158 #define	HBLK_RESERVE_MIN	20
159 
160 static struct hme_blk		*freehblkp;
161 static kmutex_t			freehblkp_lock;
162 static int			freehblkcnt;
163 
164 static int64_t			hblk_reserve[HME8BLK_SZ_RND];
165 static kmutex_t			hblk_reserve_lock;
166 static kthread_t		*hblk_reserve_thread;
167 
168 static nucleus_hblk8_info_t	nucleus_hblk8;
169 static nucleus_hblk1_info_t	nucleus_hblk1;
170 
171 /*
172  * Data to manage per-cpu hmeblk pending queues, hmeblks are queued here
173  * after the initial phase of removing an hmeblk from the hash chain, see
174  * the detailed comment in sfmmu_hblk_hash_rm() for further details.
175  */
176 static cpu_hme_pend_t		*cpu_hme_pend;
177 static uint_t			cpu_hme_pend_thresh;
178 /*
179  * SFMMU specific hat functions
180  */
181 void	hat_pagecachectl(struct page *, int);
182 
183 /* flags for hat_pagecachectl */
184 #define	HAT_CACHE	0x1
185 #define	HAT_UNCACHE	0x2
186 #define	HAT_TMPNC	0x4
187 
188 /*
189  * Flag to allow the creation of non-cacheable translations
190  * to system memory. It is off by default. At the moment this
191  * flag is used by the ecache error injector. The error injector
192  * will turn it on when creating such a translation then shut it
193  * off when it's finished.
194  */
195 
196 int	sfmmu_allow_nc_trans = 0;
197 
198 /*
199  * Flag to disable large page support.
200  * 	value of 1 => disable all large pages.
201  *	bits 1, 2, and 3 are to disable 64K, 512K and 4M pages respectively.
202  *
203  * For example, use the value 0x4 to disable 512K pages.
204  *
205  */
206 #define	LARGE_PAGES_OFF		0x1
207 
208 /*
209  * The disable_large_pages and disable_ism_large_pages variables control
210  * hat_memload_array and the page sizes to be used by ISM and the kernel.
211  *
212  * The disable_auto_data_large_pages and disable_auto_text_large_pages variables
213  * are only used to control which OOB pages to use at upper VM segment creation
214  * time, and are set in hat_init_pagesizes and used in the map_pgsz* routines.
215  * Their values may come from platform or CPU specific code to disable page
216  * sizes that should not be used.
217  *
218  * WARNING: 512K pages are currently not supported for ISM/DISM.
219  */
220 uint_t	disable_large_pages = 0;
221 uint_t	disable_ism_large_pages = (1 << TTE512K);
222 uint_t	disable_auto_data_large_pages = 0;
223 uint_t	disable_auto_text_large_pages = 0;
224 
225 /*
226  * Private sfmmu data structures for hat management
227  */
228 static struct kmem_cache *sfmmuid_cache;
229 static struct kmem_cache *mmuctxdom_cache;
230 
231 /*
232  * Private sfmmu data structures for tsb management
233  */
234 static struct kmem_cache *sfmmu_tsbinfo_cache;
235 static struct kmem_cache *sfmmu_tsb8k_cache;
236 static struct kmem_cache *sfmmu_tsb_cache[NLGRPS_MAX];
237 static vmem_t *kmem_bigtsb_arena;
238 static vmem_t *kmem_tsb_arena;
239 
240 /*
241  * sfmmu static variables for hmeblk resource management.
242  */
243 static vmem_t *hat_memload1_arena; /* HAT translation arena for sfmmu1_cache */
244 static struct kmem_cache *sfmmu8_cache;
245 static struct kmem_cache *sfmmu1_cache;
246 static struct kmem_cache *pa_hment_cache;
247 
248 static kmutex_t 	ism_mlist_lock;	/* mutex for ism mapping list */
249 /*
250  * private data for ism
251  */
252 static struct kmem_cache *ism_blk_cache;
253 static struct kmem_cache *ism_ment_cache;
254 #define	ISMID_STARTADDR	NULL
255 
256 /*
257  * Region management data structures and function declarations.
258  */
259 
260 static void	sfmmu_leave_srd(sfmmu_t *);
261 static int	sfmmu_srdcache_constructor(void *, void *, int);
262 static void	sfmmu_srdcache_destructor(void *, void *);
263 static int	sfmmu_rgncache_constructor(void *, void *, int);
264 static void	sfmmu_rgncache_destructor(void *, void *);
265 static int	sfrgnmap_isnull(sf_region_map_t *);
266 static int	sfhmergnmap_isnull(sf_hmeregion_map_t *);
267 static int	sfmmu_scdcache_constructor(void *, void *, int);
268 static void	sfmmu_scdcache_destructor(void *, void *);
269 static void	sfmmu_rgn_cb_noop(caddr_t, caddr_t, caddr_t,
270     size_t, void *, u_offset_t);
271 
272 static uint_t srd_hashmask = SFMMU_MAX_SRD_BUCKETS - 1;
273 static sf_srd_bucket_t *srd_buckets;
274 static struct kmem_cache *srd_cache;
275 static uint_t srd_rgn_hashmask = SFMMU_MAX_REGION_BUCKETS - 1;
276 static struct kmem_cache *region_cache;
277 static struct kmem_cache *scd_cache;
278 
279 #ifdef sun4v
280 int use_bigtsb_arena = 1;
281 #else
282 int use_bigtsb_arena = 0;
283 #endif
284 
285 /* External /etc/system tunable, for turning on&off the shctx support */
286 int disable_shctx = 0;
287 /* Internal variable, set by MD if the HW supports shctx feature */
288 int shctx_on = 0;
289 
290 #ifdef DEBUG
291 static void check_scd_sfmmu_list(sfmmu_t **, sfmmu_t *, int);
292 #endif
293 static void sfmmu_to_scd_list(sfmmu_t **, sfmmu_t *);
294 static void sfmmu_from_scd_list(sfmmu_t **, sfmmu_t *);
295 
296 static sf_scd_t *sfmmu_alloc_scd(sf_srd_t *, sf_region_map_t *);
297 static void sfmmu_find_scd(sfmmu_t *);
298 static void sfmmu_join_scd(sf_scd_t *, sfmmu_t *);
299 static void sfmmu_finish_join_scd(sfmmu_t *);
300 static void sfmmu_leave_scd(sfmmu_t *, uchar_t);
301 static void sfmmu_destroy_scd(sf_srd_t *, sf_scd_t *, sf_region_map_t *);
302 static int sfmmu_alloc_scd_tsbs(sf_srd_t *, sf_scd_t *);
303 static void sfmmu_free_scd_tsbs(sfmmu_t *);
304 static void sfmmu_tsb_inv_ctx(sfmmu_t *);
305 static int find_ism_rid(sfmmu_t *, sfmmu_t *, caddr_t, uint_t *);
306 static void sfmmu_ism_hatflags(sfmmu_t *, int);
307 static int sfmmu_srd_lock_held(sf_srd_t *);
308 static void sfmmu_remove_scd(sf_scd_t **, sf_scd_t *);
309 static void sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *);
310 static void sfmmu_link_scd_to_regions(sf_srd_t *, sf_scd_t *);
311 static void sfmmu_unlink_scd_from_regions(sf_srd_t *, sf_scd_t *);
312 static void sfmmu_link_to_hmeregion(sfmmu_t *, sf_region_t *);
313 static void sfmmu_unlink_from_hmeregion(sfmmu_t *, sf_region_t *);
314 
315 /*
316  * ``hat_lock'' is a hashed mutex lock for protecting sfmmu TSB lists,
317  * HAT flags, synchronizing TLB/TSB coherency, and context management.
318  * The lock is hashed on the sfmmup since the case where we need to lock
319  * all processes is rare but does occur (e.g. we need to unload a shared
320  * mapping from all processes using the mapping).  We have a lot of buckets,
321  * and each slab of sfmmu_t's can use about a quarter of them, giving us
322  * a fairly good distribution without wasting too much space and overhead
323  * when we have to grab them all.
324  */
325 #define	SFMMU_NUM_LOCK	128		/* must be power of two */
326 hatlock_t	hat_lock[SFMMU_NUM_LOCK];
327 
328 /*
329  * Hash algorithm optimized for a small number of slabs.
330  *  7 is (highbit((sizeof sfmmu_t)) - 1)
331  * This hash algorithm is based upon the knowledge that sfmmu_t's come from a
332  * kmem_cache, and thus they will be sequential within that cache.  In
333  * addition, each new slab will have a different "color" up to cache_maxcolor
334  * which will skew the hashing for each successive slab which is allocated.
335  * If the size of sfmmu_t changed to a larger size, this algorithm may need
336  * to be revisited.
337  */
338 #define	TSB_HASH_SHIFT_BITS (7)
339 #define	PTR_HASH(x) ((uintptr_t)x >> TSB_HASH_SHIFT_BITS)
340 
341 #ifdef DEBUG
342 int tsb_hash_debug = 0;
343 #define	TSB_HASH(sfmmup)	\
344 	(tsb_hash_debug ? &hat_lock[0] : \
345 	&hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)])
346 #else	/* DEBUG */
347 #define	TSB_HASH(sfmmup)	&hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)]
348 #endif	/* DEBUG */
349 
350 
351 /* sfmmu_replace_tsb() return codes. */
352 typedef enum tsb_replace_rc {
353 	TSB_SUCCESS,
354 	TSB_ALLOCFAIL,
355 	TSB_LOSTRACE,
356 	TSB_ALREADY_SWAPPED,
357 	TSB_CANTGROW
358 } tsb_replace_rc_t;
359 
360 /*
361  * Flags for TSB allocation routines.
362  */
363 #define	TSB_ALLOC	0x01
364 #define	TSB_FORCEALLOC	0x02
365 #define	TSB_GROW	0x04
366 #define	TSB_SHRINK	0x08
367 #define	TSB_SWAPIN	0x10
368 
369 /*
370  * Support for HAT callbacks.
371  */
372 #define	SFMMU_MAX_RELOC_CALLBACKS	10
373 int sfmmu_max_cb_id = SFMMU_MAX_RELOC_CALLBACKS;
374 static id_t sfmmu_cb_nextid = 0;
375 static id_t sfmmu_tsb_cb_id;
376 struct sfmmu_callback *sfmmu_cb_table;
377 
378 kmutex_t	kpr_mutex;
379 kmutex_t	kpr_suspendlock;
380 kthread_t	*kreloc_thread;
381 
382 /*
383  * Enable VA->PA translation sanity checking on DEBUG kernels.
384  * Disabled by default.  This is incompatible with some
385  * drivers (error injector, RSM) so if it breaks you get
386  * to keep both pieces.
387  */
388 int hat_check_vtop = 0;
389 
390 /*
391  * Private sfmmu routines (prototypes)
392  */
393 static struct hme_blk *sfmmu_shadow_hcreate(sfmmu_t *, caddr_t, int, uint_t);
394 static struct 	hme_blk *sfmmu_hblk_alloc(sfmmu_t *, caddr_t,
395 			struct hmehash_bucket *, uint_t, hmeblk_tag, uint_t,
396 			uint_t);
397 static caddr_t	sfmmu_hblk_unload(struct hat *, struct hme_blk *, caddr_t,
398 			caddr_t, demap_range_t *, uint_t);
399 static caddr_t	sfmmu_hblk_sync(struct hat *, struct hme_blk *, caddr_t,
400 			caddr_t, int);
401 static void	sfmmu_hblk_free(struct hme_blk **);
402 static void	sfmmu_hblks_list_purge(struct hme_blk **, int);
403 static uint_t	sfmmu_get_free_hblk(struct hme_blk **, uint_t);
404 static uint_t	sfmmu_put_free_hblk(struct hme_blk *, uint_t);
405 static struct hme_blk *sfmmu_hblk_steal(int);
406 static int	sfmmu_steal_this_hblk(struct hmehash_bucket *,
407 			struct hme_blk *, uint64_t, struct hme_blk *);
408 static caddr_t	sfmmu_hblk_unlock(struct hme_blk *, caddr_t, caddr_t);
409 
410 static void	hat_do_memload_array(struct hat *, caddr_t, size_t,
411 		    struct page **, uint_t, uint_t, uint_t);
412 static void	hat_do_memload(struct hat *, caddr_t, struct page *,
413 		    uint_t, uint_t, uint_t);
414 static void	sfmmu_memload_batchsmall(struct hat *, caddr_t, page_t **,
415 		    uint_t, uint_t, pgcnt_t, uint_t);
416 void		sfmmu_tteload(struct hat *, tte_t *, caddr_t, page_t *,
417 			uint_t);
418 static int	sfmmu_tteload_array(sfmmu_t *, tte_t *, caddr_t, page_t **,
419 			uint_t, uint_t);
420 static struct hmehash_bucket *sfmmu_tteload_acquire_hashbucket(sfmmu_t *,
421 					caddr_t, int, uint_t);
422 static struct hme_blk *sfmmu_tteload_find_hmeblk(sfmmu_t *,
423 			struct hmehash_bucket *, caddr_t, uint_t, uint_t,
424 			uint_t);
425 static int	sfmmu_tteload_addentry(sfmmu_t *, struct hme_blk *, tte_t *,
426 			caddr_t, page_t **, uint_t, uint_t);
427 static void	sfmmu_tteload_release_hashbucket(struct hmehash_bucket *);
428 
429 static int	sfmmu_pagearray_setup(caddr_t, page_t **, tte_t *, int);
430 static pfn_t	sfmmu_uvatopfn(caddr_t, sfmmu_t *, tte_t *);
431 void		sfmmu_memtte(tte_t *, pfn_t, uint_t, int);
432 #ifdef VAC
433 static void	sfmmu_vac_conflict(struct hat *, caddr_t, page_t *);
434 static int	sfmmu_vacconflict_array(caddr_t, page_t *, int *);
435 int	tst_tnc(page_t *pp, pgcnt_t);
436 void	conv_tnc(page_t *pp, int);
437 #endif
438 
439 static void	sfmmu_get_ctx(sfmmu_t *);
440 static void	sfmmu_free_sfmmu(sfmmu_t *);
441 
442 static void	sfmmu_ttesync(struct hat *, caddr_t, tte_t *, page_t *);
443 static void	sfmmu_chgattr(struct hat *, caddr_t, size_t, uint_t, int);
444 
445 cpuset_t	sfmmu_pageunload(page_t *, struct sf_hment *, int);
446 static void	hat_pagereload(struct page *, struct page *);
447 static cpuset_t	sfmmu_pagesync(page_t *, struct sf_hment *, uint_t);
448 #ifdef VAC
449 void	sfmmu_page_cache_array(page_t *, int, int, pgcnt_t);
450 static void	sfmmu_page_cache(page_t *, int, int, int);
451 #endif
452 
453 cpuset_t	sfmmu_rgntlb_demap(caddr_t, sf_region_t *,
454     struct hme_blk *, int);
455 static void	sfmmu_tlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *,
456 			pfn_t, int, int, int, int);
457 static void	sfmmu_ismtlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *,
458 			pfn_t, int);
459 static void	sfmmu_tlb_demap(caddr_t, sfmmu_t *, struct hme_blk *, int, int);
460 static void	sfmmu_tlb_range_demap(demap_range_t *);
461 static void	sfmmu_invalidate_ctx(sfmmu_t *);
462 static void	sfmmu_sync_mmustate(sfmmu_t *);
463 
464 static void 	sfmmu_tsbinfo_setup_phys(struct tsb_info *, pfn_t);
465 static int	sfmmu_tsbinfo_alloc(struct tsb_info **, int, int, uint_t,
466 			sfmmu_t *);
467 static void	sfmmu_tsb_free(struct tsb_info *);
468 static void	sfmmu_tsbinfo_free(struct tsb_info *);
469 static int	sfmmu_init_tsbinfo(struct tsb_info *, int, int, uint_t,
470 			sfmmu_t *);
471 static void	sfmmu_tsb_chk_reloc(sfmmu_t *, hatlock_t *);
472 static void	sfmmu_tsb_swapin(sfmmu_t *, hatlock_t *);
473 static int	sfmmu_select_tsb_szc(pgcnt_t);
474 static void	sfmmu_mod_tsb(sfmmu_t *, caddr_t, tte_t *, int);
475 #define		sfmmu_load_tsb(sfmmup, vaddr, tte, szc) \
476 	sfmmu_mod_tsb(sfmmup, vaddr, tte, szc)
477 #define		sfmmu_unload_tsb(sfmmup, vaddr, szc)    \
478 	sfmmu_mod_tsb(sfmmup, vaddr, NULL, szc)
479 static void	sfmmu_copy_tsb(struct tsb_info *, struct tsb_info *);
480 static tsb_replace_rc_t sfmmu_replace_tsb(sfmmu_t *, struct tsb_info *, uint_t,
481     hatlock_t *, uint_t);
482 static void	sfmmu_size_tsb(sfmmu_t *, int, uint64_t, uint64_t, int);
483 
484 #ifdef VAC
485 void	sfmmu_cache_flush(pfn_t, int);
486 void	sfmmu_cache_flushcolor(int, pfn_t);
487 #endif
488 static caddr_t	sfmmu_hblk_chgattr(sfmmu_t *, struct hme_blk *, caddr_t,
489 			caddr_t, demap_range_t *, uint_t, int);
490 
491 static uint64_t	sfmmu_vtop_attr(uint_t, int mode, tte_t *);
492 static uint_t	sfmmu_ptov_attr(tte_t *);
493 static caddr_t	sfmmu_hblk_chgprot(sfmmu_t *, struct hme_blk *, caddr_t,
494 			caddr_t, demap_range_t *, uint_t);
495 static uint_t	sfmmu_vtop_prot(uint_t, uint_t *);
496 static int	sfmmu_idcache_constructor(void *, void *, int);
497 static void	sfmmu_idcache_destructor(void *, void *);
498 static int	sfmmu_hblkcache_constructor(void *, void *, int);
499 static void	sfmmu_hblkcache_destructor(void *, void *);
500 static void	sfmmu_hblkcache_reclaim(void *);
501 static void	sfmmu_shadow_hcleanup(sfmmu_t *, struct hme_blk *,
502 			struct hmehash_bucket *);
503 static void	sfmmu_hblk_hash_rm(struct hmehash_bucket *, struct hme_blk *,
504 			struct hme_blk *, struct hme_blk **, int);
505 static void	sfmmu_hblk_hash_add(struct hmehash_bucket *, struct hme_blk *,
506 			uint64_t);
507 static struct hme_blk *sfmmu_check_pending_hblks(int);
508 static void	sfmmu_free_hblks(sfmmu_t *, caddr_t, caddr_t, int);
509 static void	sfmmu_cleanup_rhblk(sf_srd_t *, caddr_t, uint_t, int);
510 static void	sfmmu_unload_hmeregion_va(sf_srd_t *, uint_t, caddr_t, caddr_t,
511 			int, caddr_t *);
512 static void	sfmmu_unload_hmeregion(sf_srd_t *, sf_region_t *);
513 
514 static void	sfmmu_rm_large_mappings(page_t *, int);
515 
516 static void	hat_lock_init(void);
517 static void	hat_kstat_init(void);
518 static int	sfmmu_kstat_percpu_update(kstat_t *ksp, int rw);
519 static void	sfmmu_set_scd_rttecnt(sf_srd_t *, sf_scd_t *);
520 static	int	sfmmu_is_rgnva(sf_srd_t *, caddr_t, ulong_t, ulong_t);
521 static void	sfmmu_check_page_sizes(sfmmu_t *, int);
522 int	fnd_mapping_sz(page_t *);
523 static void	iment_add(struct ism_ment *,  struct hat *);
524 static void	iment_sub(struct ism_ment *, struct hat *);
525 static pgcnt_t	ism_tsb_entries(sfmmu_t *, int szc);
526 extern void	sfmmu_setup_tsbinfo(sfmmu_t *);
527 extern void	sfmmu_clear_utsbinfo(void);
528 
529 static void		sfmmu_ctx_wrap_around(mmu_ctx_t *, boolean_t);
530 
531 extern int vpm_enable;
532 
533 /* kpm globals */
534 #ifdef	DEBUG
535 /*
536  * Enable trap level tsbmiss handling
537  */
538 int	kpm_tsbmtl = 1;
539 
540 /*
541  * Flush the TLB on kpm mapout. Note: Xcalls are used (again) for the
542  * required TLB shootdowns in this case, so handle w/ care. Off by default.
543  */
544 int	kpm_tlb_flush;
545 #endif	/* DEBUG */
546 
547 static void	*sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *, size_t, int);
548 
549 #ifdef DEBUG
550 static void	sfmmu_check_hblk_flist();
551 #endif
552 
553 /*
554  * Semi-private sfmmu data structures.  Some of them are initialize in
555  * startup or in hat_init. Some of them are private but accessed by
556  * assembly code or mach_sfmmu.c
557  */
558 struct hmehash_bucket *uhme_hash;	/* user hmeblk hash table */
559 struct hmehash_bucket *khme_hash;	/* kernel hmeblk hash table */
560 uint64_t	uhme_hash_pa;		/* PA of uhme_hash */
561 uint64_t	khme_hash_pa;		/* PA of khme_hash */
562 int 		uhmehash_num;		/* # of buckets in user hash table */
563 int 		khmehash_num;		/* # of buckets in kernel hash table */
564 
565 uint_t		max_mmu_ctxdoms = 0;	/* max context domains in the system */
566 mmu_ctx_t	**mmu_ctxs_tbl;		/* global array of context domains */
567 uint64_t	mmu_saved_gnum = 0;	/* to init incoming MMUs' gnums */
568 
569 #define	DEFAULT_NUM_CTXS_PER_MMU 8192
570 static uint_t	nctxs = DEFAULT_NUM_CTXS_PER_MMU;
571 
572 int		cache;			/* describes system cache */
573 
574 caddr_t		ktsb_base;		/* kernel 8k-indexed tsb base address */
575 uint64_t	ktsb_pbase;		/* kernel 8k-indexed tsb phys address */
576 int		ktsb_szcode;		/* kernel 8k-indexed tsb size code */
577 int		ktsb_sz;		/* kernel 8k-indexed tsb size */
578 
579 caddr_t		ktsb4m_base;		/* kernel 4m-indexed tsb base address */
580 uint64_t	ktsb4m_pbase;		/* kernel 4m-indexed tsb phys address */
581 int		ktsb4m_szcode;		/* kernel 4m-indexed tsb size code */
582 int		ktsb4m_sz;		/* kernel 4m-indexed tsb size */
583 
584 uint64_t	kpm_tsbbase;		/* kernel seg_kpm 4M TSB base address */
585 int		kpm_tsbsz;		/* kernel seg_kpm 4M TSB size code */
586 uint64_t	kpmsm_tsbbase;		/* kernel seg_kpm 8K TSB base address */
587 int		kpmsm_tsbsz;		/* kernel seg_kpm 8K TSB size code */
588 
589 #ifndef sun4v
590 int		utsb_dtlb_ttenum = -1;	/* index in TLB for utsb locked TTE */
591 int		utsb4m_dtlb_ttenum = -1; /* index in TLB for 4M TSB TTE */
592 int		dtlb_resv_ttenum;	/* index in TLB of first reserved TTE */
593 caddr_t		utsb_vabase;		/* reserved kernel virtual memory */
594 caddr_t		utsb4m_vabase;		/* for trap handler TSB accesses */
595 #endif /* sun4v */
596 uint64_t	tsb_alloc_bytes = 0;	/* bytes allocated to TSBs */
597 vmem_t		*kmem_tsb_default_arena[NLGRPS_MAX];	/* For dynamic TSBs */
598 vmem_t		*kmem_bigtsb_default_arena[NLGRPS_MAX]; /* dynamic 256M TSBs */
599 
600 /*
601  * Size to use for TSB slabs.  Future platforms that support page sizes
602  * larger than 4M may wish to change these values, and provide their own
603  * assembly macros for building and decoding the TSB base register contents.
604  * Note disable_large_pages will override the value set here.
605  */
606 static	uint_t tsb_slab_ttesz = TTE4M;
607 size_t	tsb_slab_size = MMU_PAGESIZE4M;
608 uint_t	tsb_slab_shift = MMU_PAGESHIFT4M;
609 /* PFN mask for TTE */
610 size_t	tsb_slab_mask = MMU_PAGEOFFSET4M >> MMU_PAGESHIFT;
611 
612 /*
613  * Size to use for TSB slabs.  These are used only when 256M tsb arenas
614  * exist.
615  */
616 static uint_t	bigtsb_slab_ttesz = TTE256M;
617 static size_t	bigtsb_slab_size = MMU_PAGESIZE256M;
618 static uint_t	bigtsb_slab_shift = MMU_PAGESHIFT256M;
619 /* 256M page alignment for 8K pfn */
620 static size_t	bigtsb_slab_mask = MMU_PAGEOFFSET256M >> MMU_PAGESHIFT;
621 
622 /* largest TSB size to grow to, will be smaller on smaller memory systems */
623 static int	tsb_max_growsize = 0;
624 
625 /*
626  * Tunable parameters dealing with TSB policies.
627  */
628 
629 /*
630  * This undocumented tunable forces all 8K TSBs to be allocated from
631  * the kernel heap rather than from the kmem_tsb_default_arena arenas.
632  */
633 #ifdef	DEBUG
634 int	tsb_forceheap = 0;
635 #endif	/* DEBUG */
636 
637 /*
638  * Decide whether to use per-lgroup arenas, or one global set of
639  * TSB arenas.  The default is not to break up per-lgroup, since
640  * most platforms don't recognize any tangible benefit from it.
641  */
642 int	tsb_lgrp_affinity = 0;
643 
644 /*
645  * Used for growing the TSB based on the process RSS.
646  * tsb_rss_factor is based on the smallest TSB, and is
647  * shifted by the TSB size to determine if we need to grow.
648  * The default will grow the TSB if the number of TTEs for
649  * this page size exceeds 75% of the number of TSB entries,
650  * which should _almost_ eliminate all conflict misses
651  * (at the expense of using up lots and lots of memory).
652  */
653 #define	TSB_RSS_FACTOR		(TSB_ENTRIES(TSB_MIN_SZCODE) * 0.75)
654 #define	SFMMU_RSS_TSBSIZE(tsbszc)	(tsb_rss_factor << tsbszc)
655 #define	SELECT_TSB_SIZECODE(pgcnt) ( \
656 	(enable_tsb_rss_sizing)? sfmmu_select_tsb_szc(pgcnt) : \
657 	default_tsb_size)
658 #define	TSB_OK_SHRINK()	\
659 	(tsb_alloc_bytes > tsb_alloc_hiwater || freemem < desfree)
660 #define	TSB_OK_GROW()	\
661 	(tsb_alloc_bytes < tsb_alloc_hiwater && freemem > desfree)
662 
663 int	enable_tsb_rss_sizing = 1;
664 int	tsb_rss_factor	= (int)TSB_RSS_FACTOR;
665 
666 /* which TSB size code to use for new address spaces or if rss sizing off */
667 int default_tsb_size = TSB_8K_SZCODE;
668 
669 static uint64_t tsb_alloc_hiwater; /* limit TSB reserved memory */
670 uint64_t tsb_alloc_hiwater_factor; /* tsb_alloc_hiwater = physmem / this */
671 #define	TSB_ALLOC_HIWATER_FACTOR_DEFAULT	32
672 
673 #ifdef DEBUG
674 static int tsb_random_size = 0;	/* set to 1 to test random tsb sizes on alloc */
675 static int tsb_grow_stress = 0;	/* if set to 1, keep replacing TSB w/ random */
676 static int tsb_alloc_mtbf = 0;	/* fail allocation every n attempts */
677 static int tsb_alloc_fail_mtbf = 0;
678 static int tsb_alloc_count = 0;
679 #endif /* DEBUG */
680 
681 /* if set to 1, will remap valid TTEs when growing TSB. */
682 int tsb_remap_ttes = 1;
683 
684 /*
685  * If we have more than this many mappings, allocate a second TSB.
686  * This default is chosen because the I/D fully associative TLBs are
687  * assumed to have at least 8 available entries. Platforms with a
688  * larger fully-associative TLB could probably override the default.
689  */
690 
691 #ifdef sun4v
692 int tsb_sectsb_threshold = 0;
693 #else
694 int tsb_sectsb_threshold = 8;
695 #endif
696 
697 /*
698  * kstat data
699  */
700 struct sfmmu_global_stat sfmmu_global_stat;
701 struct sfmmu_tsbsize_stat sfmmu_tsbsize_stat;
702 
703 /*
704  * Global data
705  */
706 sfmmu_t 	*ksfmmup;		/* kernel's hat id */
707 
708 #ifdef DEBUG
709 static void	chk_tte(tte_t *, tte_t *, tte_t *, struct hme_blk *);
710 #endif
711 
712 /* sfmmu locking operations */
713 static kmutex_t *sfmmu_mlspl_enter(struct page *, int);
714 static int	sfmmu_mlspl_held(struct page *, int);
715 
716 kmutex_t *sfmmu_page_enter(page_t *);
717 void	sfmmu_page_exit(kmutex_t *);
718 int	sfmmu_page_spl_held(struct page *);
719 
720 /* sfmmu internal locking operations - accessed directly */
721 static void	sfmmu_mlist_reloc_enter(page_t *, page_t *,
722 				kmutex_t **, kmutex_t **);
723 static void	sfmmu_mlist_reloc_exit(kmutex_t *, kmutex_t *);
724 static hatlock_t *
725 		sfmmu_hat_enter(sfmmu_t *);
726 static hatlock_t *
727 		sfmmu_hat_tryenter(sfmmu_t *);
728 static void	sfmmu_hat_exit(hatlock_t *);
729 static void	sfmmu_hat_lock_all(void);
730 static void	sfmmu_hat_unlock_all(void);
731 static void	sfmmu_ismhat_enter(sfmmu_t *, int);
732 static void	sfmmu_ismhat_exit(sfmmu_t *, int);
733 
734 kpm_hlk_t	*kpmp_table;
735 uint_t		kpmp_table_sz;	/* must be a power of 2 */
736 uchar_t		kpmp_shift;
737 
738 kpm_shlk_t	*kpmp_stable;
739 uint_t		kpmp_stable_sz;	/* must be a power of 2 */
740 
741 /*
742  * SPL_TABLE_SIZE is 2 * NCPU, but no smaller than 128.
743  * SPL_SHIFT is log2(SPL_TABLE_SIZE).
744  */
745 #if ((2*NCPU_P2) > 128)
746 #define	SPL_SHIFT	((unsigned)(NCPU_LOG2 + 1))
747 #else
748 #define	SPL_SHIFT	7U
749 #endif
750 #define	SPL_TABLE_SIZE	(1U << SPL_SHIFT)
751 #define	SPL_MASK	(SPL_TABLE_SIZE - 1)
752 
753 /*
754  * We shift by PP_SHIFT to take care of the low-order 0 bits of a page_t
755  * and by multiples of SPL_SHIFT to get as many varied bits as we can.
756  */
757 #define	SPL_INDEX(pp) \
758 	((((uintptr_t)(pp) >> PP_SHIFT) ^ \
759 	((uintptr_t)(pp) >> (PP_SHIFT + SPL_SHIFT)) ^ \
760 	((uintptr_t)(pp) >> (PP_SHIFT + SPL_SHIFT * 2)) ^ \
761 	((uintptr_t)(pp) >> (PP_SHIFT + SPL_SHIFT * 3))) & \
762 	SPL_MASK)
763 
764 #define	SPL_HASH(pp)    \
765 	(&sfmmu_page_lock[SPL_INDEX(pp)].pad_mutex)
766 
767 static	pad_mutex_t	sfmmu_page_lock[SPL_TABLE_SIZE];
768 
769 /* Array of mutexes protecting a page's mapping list and p_nrm field. */
770 
771 #define	MML_TABLE_SIZE	SPL_TABLE_SIZE
772 #define	MLIST_HASH(pp)	(&mml_table[SPL_INDEX(pp)].pad_mutex)
773 
774 static pad_mutex_t	mml_table[MML_TABLE_SIZE];
775 
776 /*
777  * hat_unload_callback() will group together callbacks in order
778  * to avoid xt_sync() calls.  This is the maximum size of the group.
779  */
780 #define	MAX_CB_ADDR	32
781 
782 tte_t	hw_tte;
783 static ulong_t sfmmu_dmr_maxbit = DMR_MAXBIT;
784 
785 static char	*mmu_ctx_kstat_names[] = {
786 	"mmu_ctx_tsb_exceptions",
787 	"mmu_ctx_tsb_raise_exception",
788 	"mmu_ctx_wrap_around",
789 };
790 
791 /*
792  * Wrapper for vmem_xalloc since vmem_create only allows limited
793  * parameters for vm_source_alloc functions.  This function allows us
794  * to specify alignment consistent with the size of the object being
795  * allocated.
796  */
797 static void *
798 sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag)
799 {
800 	return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag));
801 }
802 
803 /* Common code for setting tsb_alloc_hiwater. */
804 #define	SFMMU_SET_TSB_ALLOC_HIWATER(pages)	tsb_alloc_hiwater = \
805 		ptob(pages) / tsb_alloc_hiwater_factor
806 
807 /*
808  * Set tsb_max_growsize to allow at most all of physical memory to be mapped by
809  * a single TSB.  physmem is the number of physical pages so we need physmem 8K
810  * TTEs to represent all those physical pages.  We round this up by using
811  * 1<<highbit().  To figure out which size code to use, remember that the size
812  * code is just an amount to shift the smallest TSB size to get the size of
813  * this TSB.  So we subtract that size, TSB_START_SIZE, from highbit() (or
814  * highbit() - 1) to get the size code for the smallest TSB that can represent
815  * all of physical memory, while erring on the side of too much.
816  *
817  * Restrict tsb_max_growsize to make sure that:
818  *	1) TSBs can't grow larger than the TSB slab size
819  *	2) TSBs can't grow larger than UTSB_MAX_SZCODE.
820  */
821 #define	SFMMU_SET_TSB_MAX_GROWSIZE(pages) {				\
822 	int	_i, _szc, _slabszc, _tsbszc;				\
823 									\
824 	_i = highbit(pages);						\
825 	if ((1 << (_i - 1)) == (pages))					\
826 		_i--;		/* 2^n case, round down */              \
827 	_szc = _i - TSB_START_SIZE;					\
828 	_slabszc = bigtsb_slab_shift - (TSB_START_SIZE + TSB_ENTRY_SHIFT); \
829 	_tsbszc = MIN(_szc, _slabszc);                                  \
830 	tsb_max_growsize = MIN(_tsbszc, UTSB_MAX_SZCODE);               \
831 }
832 
833 /*
834  * Given a pointer to an sfmmu and a TTE size code, return a pointer to the
835  * tsb_info which handles that TTE size.
836  */
837 #define	SFMMU_GET_TSBINFO(tsbinfop, sfmmup, tte_szc) {			\
838 	(tsbinfop) = (sfmmup)->sfmmu_tsb;				\
839 	ASSERT(((tsbinfop)->tsb_flags & TSB_SHAREDCTX) ||		\
840 	    sfmmu_hat_lock_held(sfmmup));				\
841 	if ((tte_szc) >= TTE4M)	{					\
842 		ASSERT((tsbinfop) != NULL);				\
843 		(tsbinfop) = (tsbinfop)->tsb_next;			\
844 	}								\
845 }
846 
847 /*
848  * Macro to use to unload entries from the TSB.
849  * It has knowledge of which page sizes get replicated in the TSB
850  * and will call the appropriate unload routine for the appropriate size.
851  */
852 #define	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, ismhat)		\
853 {									\
854 	int ttesz = get_hblk_ttesz(hmeblkp);				\
855 	if (ttesz == TTE8K || ttesz == TTE4M) {				\
856 		sfmmu_unload_tsb(sfmmup, addr, ttesz);			\
857 	} else {							\
858 		caddr_t sva = ismhat ? addr : 				\
859 		    (caddr_t)get_hblk_base(hmeblkp);			\
860 		caddr_t eva = sva + get_hblk_span(hmeblkp);		\
861 		ASSERT(addr >= sva && addr < eva);			\
862 		sfmmu_unload_tsb_range(sfmmup, sva, eva, ttesz);	\
863 	}								\
864 }
865 
866 
867 /* Update tsb_alloc_hiwater after memory is configured. */
868 /*ARGSUSED*/
869 static void
870 sfmmu_update_post_add(void *arg, pgcnt_t delta_pages)
871 {
872 	/* Assumes physmem has already been updated. */
873 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
874 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
875 }
876 
877 /*
878  * Update tsb_alloc_hiwater before memory is deleted.  We'll do nothing here
879  * and update tsb_alloc_hiwater and tsb_max_growsize after the memory is
880  * deleted.
881  */
882 /*ARGSUSED*/
883 static int
884 sfmmu_update_pre_del(void *arg, pgcnt_t delta_pages)
885 {
886 	return (0);
887 }
888 
889 /* Update tsb_alloc_hiwater after memory fails to be unconfigured. */
890 /*ARGSUSED*/
891 static void
892 sfmmu_update_post_del(void *arg, pgcnt_t delta_pages, int cancelled)
893 {
894 	/*
895 	 * Whether the delete was cancelled or not, just go ahead and update
896 	 * tsb_alloc_hiwater and tsb_max_growsize.
897 	 */
898 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
899 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
900 }
901 
902 static kphysm_setup_vector_t sfmmu_update_vec = {
903 	KPHYSM_SETUP_VECTOR_VERSION,	/* version */
904 	sfmmu_update_post_add,		/* post_add */
905 	sfmmu_update_pre_del,		/* pre_del */
906 	sfmmu_update_post_del		/* post_del */
907 };
908 
909 
910 /*
911  * HME_BLK HASH PRIMITIVES
912  */
913 
914 /*
915  * Enter a hme on the mapping list for page pp.
916  * When large pages are more prevalent in the system we might want to
917  * keep the mapping list in ascending order by the hment size. For now,
918  * small pages are more frequent, so don't slow it down.
919  */
920 #define	HME_ADD(hme, pp)					\
921 {								\
922 	ASSERT(sfmmu_mlist_held(pp));				\
923 								\
924 	hme->hme_prev = NULL;					\
925 	hme->hme_next = pp->p_mapping;				\
926 	hme->hme_page = pp;					\
927 	if (pp->p_mapping) {					\
928 		((struct sf_hment *)(pp->p_mapping))->hme_prev = hme;\
929 		ASSERT(pp->p_share > 0);			\
930 	} else  {						\
931 		/* EMPTY */					\
932 		ASSERT(pp->p_share == 0);			\
933 	}							\
934 	pp->p_mapping = hme;					\
935 	pp->p_share++;						\
936 }
937 
938 /*
939  * Enter a hme on the mapping list for page pp.
940  * If we are unmapping a large translation, we need to make sure that the
941  * change is reflect in the corresponding bit of the p_index field.
942  */
943 #define	HME_SUB(hme, pp)					\
944 {								\
945 	ASSERT(sfmmu_mlist_held(pp));				\
946 	ASSERT(hme->hme_page == pp || IS_PAHME(hme));		\
947 								\
948 	if (pp->p_mapping == NULL) {				\
949 		panic("hme_remove - no mappings");		\
950 	}							\
951 								\
952 	membar_stst();	/* ensure previous stores finish */	\
953 								\
954 	ASSERT(pp->p_share > 0);				\
955 	pp->p_share--;						\
956 								\
957 	if (hme->hme_prev) {					\
958 		ASSERT(pp->p_mapping != hme);			\
959 		ASSERT(hme->hme_prev->hme_page == pp ||		\
960 			IS_PAHME(hme->hme_prev));		\
961 		hme->hme_prev->hme_next = hme->hme_next;	\
962 	} else {						\
963 		ASSERT(pp->p_mapping == hme);			\
964 		pp->p_mapping = hme->hme_next;			\
965 		ASSERT((pp->p_mapping == NULL) ?		\
966 			(pp->p_share == 0) : 1);		\
967 	}							\
968 								\
969 	if (hme->hme_next) {					\
970 		ASSERT(hme->hme_next->hme_page == pp ||		\
971 			IS_PAHME(hme->hme_next));		\
972 		hme->hme_next->hme_prev = hme->hme_prev;	\
973 	}							\
974 								\
975 	/* zero out the entry */				\
976 	hme->hme_next = NULL;					\
977 	hme->hme_prev = NULL;					\
978 	hme->hme_page = NULL;					\
979 								\
980 	if (hme_size(hme) > TTE8K) {				\
981 		/* remove mappings for remainder of large pg */	\
982 		sfmmu_rm_large_mappings(pp, hme_size(hme));	\
983 	}							\
984 }
985 
986 /*
987  * This function returns the hment given the hme_blk and a vaddr.
988  * It assumes addr has already been checked to belong to hme_blk's
989  * range.
990  */
991 #define	HBLKTOHME(hment, hmeblkp, addr)					\
992 {									\
993 	int index;							\
994 	HBLKTOHME_IDX(hment, hmeblkp, addr, index)			\
995 }
996 
997 /*
998  * Version of HBLKTOHME that also returns the index in hmeblkp
999  * of the hment.
1000  */
1001 #define	HBLKTOHME_IDX(hment, hmeblkp, addr, idx)			\
1002 {									\
1003 	ASSERT(in_hblk_range((hmeblkp), (addr)));			\
1004 									\
1005 	if (get_hblk_ttesz(hmeblkp) == TTE8K) {				\
1006 		idx = (((uintptr_t)(addr) >> MMU_PAGESHIFT) & (NHMENTS-1)); \
1007 	} else								\
1008 		idx = 0;						\
1009 									\
1010 	(hment) = &(hmeblkp)->hblk_hme[idx];				\
1011 }
1012 
1013 /*
1014  * Disable any page sizes not supported by the CPU
1015  */
1016 void
1017 hat_init_pagesizes()
1018 {
1019 	int 		i;
1020 
1021 	mmu_exported_page_sizes = 0;
1022 	for (i = TTE8K; i < max_mmu_page_sizes; i++) {
1023 
1024 		szc_2_userszc[i] = (uint_t)-1;
1025 		userszc_2_szc[i] = (uint_t)-1;
1026 
1027 		if ((mmu_exported_pagesize_mask & (1 << i)) == 0) {
1028 			disable_large_pages |= (1 << i);
1029 		} else {
1030 			szc_2_userszc[i] = mmu_exported_page_sizes;
1031 			userszc_2_szc[mmu_exported_page_sizes] = i;
1032 			mmu_exported_page_sizes++;
1033 		}
1034 	}
1035 
1036 	disable_ism_large_pages |= disable_large_pages;
1037 	disable_auto_data_large_pages = disable_large_pages;
1038 	disable_auto_text_large_pages = disable_large_pages;
1039 
1040 	/*
1041 	 * Initialize mmu-specific large page sizes.
1042 	 */
1043 	if (&mmu_large_pages_disabled) {
1044 		disable_large_pages |= mmu_large_pages_disabled(HAT_LOAD);
1045 		disable_ism_large_pages |=
1046 		    mmu_large_pages_disabled(HAT_LOAD_SHARE);
1047 		disable_auto_data_large_pages |=
1048 		    mmu_large_pages_disabled(HAT_AUTO_DATA);
1049 		disable_auto_text_large_pages |=
1050 		    mmu_large_pages_disabled(HAT_AUTO_TEXT);
1051 	}
1052 }
1053 
1054 /*
1055  * Initialize the hardware address translation structures.
1056  */
1057 void
1058 hat_init(void)
1059 {
1060 	int 		i;
1061 	uint_t		sz;
1062 	size_t		size;
1063 
1064 	hat_lock_init();
1065 	hat_kstat_init();
1066 
1067 	/*
1068 	 * Hardware-only bits in a TTE
1069 	 */
1070 	MAKE_TTE_MASK(&hw_tte);
1071 
1072 	hat_init_pagesizes();
1073 
1074 	/* Initialize the hash locks */
1075 	for (i = 0; i < khmehash_num; i++) {
1076 		mutex_init(&khme_hash[i].hmehash_mutex, NULL,
1077 		    MUTEX_DEFAULT, NULL);
1078 		khme_hash[i].hmeh_nextpa = HMEBLK_ENDPA;
1079 	}
1080 	for (i = 0; i < uhmehash_num; i++) {
1081 		mutex_init(&uhme_hash[i].hmehash_mutex, NULL,
1082 		    MUTEX_DEFAULT, NULL);
1083 		uhme_hash[i].hmeh_nextpa = HMEBLK_ENDPA;
1084 	}
1085 	khmehash_num--;		/* make sure counter starts from 0 */
1086 	uhmehash_num--;		/* make sure counter starts from 0 */
1087 
1088 	/*
1089 	 * Allocate context domain structures.
1090 	 *
1091 	 * A platform may choose to modify max_mmu_ctxdoms in
1092 	 * set_platform_defaults(). If a platform does not define
1093 	 * a set_platform_defaults() or does not choose to modify
1094 	 * max_mmu_ctxdoms, it gets one MMU context domain for every CPU.
1095 	 *
1096 	 * For all platforms that have CPUs sharing MMUs, this
1097 	 * value must be defined.
1098 	 */
1099 	if (max_mmu_ctxdoms == 0)
1100 		max_mmu_ctxdoms = max_ncpus;
1101 
1102 	size = max_mmu_ctxdoms * sizeof (mmu_ctx_t *);
1103 	mmu_ctxs_tbl = kmem_zalloc(size, KM_SLEEP);
1104 
1105 	/* mmu_ctx_t is 64 bytes aligned */
1106 	mmuctxdom_cache = kmem_cache_create("mmuctxdom_cache",
1107 	    sizeof (mmu_ctx_t), 64, NULL, NULL, NULL, NULL, NULL, 0);
1108 	/*
1109 	 * MMU context domain initialization for the Boot CPU.
1110 	 * This needs the context domains array allocated above.
1111 	 */
1112 	mutex_enter(&cpu_lock);
1113 	sfmmu_cpu_init(CPU);
1114 	mutex_exit(&cpu_lock);
1115 
1116 	/*
1117 	 * Intialize ism mapping list lock.
1118 	 */
1119 
1120 	mutex_init(&ism_mlist_lock, NULL, MUTEX_DEFAULT, NULL);
1121 
1122 	/*
1123 	 * Each sfmmu structure carries an array of MMU context info
1124 	 * structures, one per context domain. The size of this array depends
1125 	 * on the maximum number of context domains. So, the size of the
1126 	 * sfmmu structure varies per platform.
1127 	 *
1128 	 * sfmmu is allocated from static arena, because trap
1129 	 * handler at TL > 0 is not allowed to touch kernel relocatable
1130 	 * memory. sfmmu's alignment is changed to 64 bytes from
1131 	 * default 8 bytes, as the lower 6 bits will be used to pass
1132 	 * pgcnt to vtag_flush_pgcnt_tl1.
1133 	 */
1134 	size = sizeof (sfmmu_t) + sizeof (sfmmu_ctx_t) * (max_mmu_ctxdoms - 1);
1135 
1136 	sfmmuid_cache = kmem_cache_create("sfmmuid_cache", size,
1137 	    64, sfmmu_idcache_constructor, sfmmu_idcache_destructor,
1138 	    NULL, NULL, static_arena, 0);
1139 
1140 	sfmmu_tsbinfo_cache = kmem_cache_create("sfmmu_tsbinfo_cache",
1141 	    sizeof (struct tsb_info), 0, NULL, NULL, NULL, NULL, NULL, 0);
1142 
1143 	/*
1144 	 * Since we only use the tsb8k cache to "borrow" pages for TSBs
1145 	 * from the heap when low on memory or when TSB_FORCEALLOC is
1146 	 * specified, don't use magazines to cache them--we want to return
1147 	 * them to the system as quickly as possible.
1148 	 */
1149 	sfmmu_tsb8k_cache = kmem_cache_create("sfmmu_tsb8k_cache",
1150 	    MMU_PAGESIZE, MMU_PAGESIZE, NULL, NULL, NULL, NULL,
1151 	    static_arena, KMC_NOMAGAZINE);
1152 
1153 	/*
1154 	 * Set tsb_alloc_hiwater to 1/tsb_alloc_hiwater_factor of physical
1155 	 * memory, which corresponds to the old static reserve for TSBs.
1156 	 * tsb_alloc_hiwater_factor defaults to 32.  This caps the amount of
1157 	 * memory we'll allocate for TSB slabs; beyond this point TSB
1158 	 * allocations will be taken from the kernel heap (via
1159 	 * sfmmu_tsb8k_cache) and will be throttled as would any other kmem
1160 	 * consumer.
1161 	 */
1162 	if (tsb_alloc_hiwater_factor == 0) {
1163 		tsb_alloc_hiwater_factor = TSB_ALLOC_HIWATER_FACTOR_DEFAULT;
1164 	}
1165 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
1166 
1167 	for (sz = tsb_slab_ttesz; sz > 0; sz--) {
1168 		if (!(disable_large_pages & (1 << sz)))
1169 			break;
1170 	}
1171 
1172 	if (sz < tsb_slab_ttesz) {
1173 		tsb_slab_ttesz = sz;
1174 		tsb_slab_shift = MMU_PAGESHIFT + (sz << 1) + sz;
1175 		tsb_slab_size = 1 << tsb_slab_shift;
1176 		tsb_slab_mask = (1 << (tsb_slab_shift - MMU_PAGESHIFT)) - 1;
1177 		use_bigtsb_arena = 0;
1178 	} else if (use_bigtsb_arena &&
1179 	    (disable_large_pages & (1 << bigtsb_slab_ttesz))) {
1180 		use_bigtsb_arena = 0;
1181 	}
1182 
1183 	if (!use_bigtsb_arena) {
1184 		bigtsb_slab_shift = tsb_slab_shift;
1185 	}
1186 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
1187 
1188 	/*
1189 	 * On smaller memory systems, allocate TSB memory in smaller chunks
1190 	 * than the default 4M slab size. We also honor disable_large_pages
1191 	 * here.
1192 	 *
1193 	 * The trap handlers need to be patched with the final slab shift,
1194 	 * since they need to be able to construct the TSB pointer at runtime.
1195 	 */
1196 	if ((tsb_max_growsize <= TSB_512K_SZCODE) &&
1197 	    !(disable_large_pages & (1 << TTE512K))) {
1198 		tsb_slab_ttesz = TTE512K;
1199 		tsb_slab_shift = MMU_PAGESHIFT512K;
1200 		tsb_slab_size = MMU_PAGESIZE512K;
1201 		tsb_slab_mask = MMU_PAGEOFFSET512K >> MMU_PAGESHIFT;
1202 		use_bigtsb_arena = 0;
1203 	}
1204 
1205 	if (!use_bigtsb_arena) {
1206 		bigtsb_slab_ttesz = tsb_slab_ttesz;
1207 		bigtsb_slab_shift = tsb_slab_shift;
1208 		bigtsb_slab_size = tsb_slab_size;
1209 		bigtsb_slab_mask = tsb_slab_mask;
1210 	}
1211 
1212 
1213 	/*
1214 	 * Set up memory callback to update tsb_alloc_hiwater and
1215 	 * tsb_max_growsize.
1216 	 */
1217 	i = kphysm_setup_func_register(&sfmmu_update_vec, (void *) 0);
1218 	ASSERT(i == 0);
1219 
1220 	/*
1221 	 * kmem_tsb_arena is the source from which large TSB slabs are
1222 	 * drawn.  The quantum of this arena corresponds to the largest
1223 	 * TSB size we can dynamically allocate for user processes.
1224 	 * Currently it must also be a supported page size since we
1225 	 * use exactly one translation entry to map each slab page.
1226 	 *
1227 	 * The per-lgroup kmem_tsb_default_arena arenas are the arenas from
1228 	 * which most TSBs are allocated.  Since most TSB allocations are
1229 	 * typically 8K we have a kmem cache we stack on top of each
1230 	 * kmem_tsb_default_arena to speed up those allocations.
1231 	 *
1232 	 * Note the two-level scheme of arenas is required only
1233 	 * because vmem_create doesn't allow us to specify alignment
1234 	 * requirements.  If this ever changes the code could be
1235 	 * simplified to use only one level of arenas.
1236 	 *
1237 	 * If 256M page support exists on sun4v, 256MB kmem_bigtsb_arena
1238 	 * will be provided in addition to the 4M kmem_tsb_arena.
1239 	 */
1240 	if (use_bigtsb_arena) {
1241 		kmem_bigtsb_arena = vmem_create("kmem_bigtsb", NULL, 0,
1242 		    bigtsb_slab_size, sfmmu_vmem_xalloc_aligned_wrapper,
1243 		    vmem_xfree, heap_arena, 0, VM_SLEEP);
1244 	}
1245 
1246 	kmem_tsb_arena = vmem_create("kmem_tsb", NULL, 0, tsb_slab_size,
1247 	    sfmmu_vmem_xalloc_aligned_wrapper,
1248 	    vmem_xfree, heap_arena, 0, VM_SLEEP);
1249 
1250 	if (tsb_lgrp_affinity) {
1251 		char s[50];
1252 		for (i = 0; i < NLGRPS_MAX; i++) {
1253 			if (use_bigtsb_arena) {
1254 				(void) sprintf(s, "kmem_bigtsb_lgrp%d", i);
1255 				kmem_bigtsb_default_arena[i] = vmem_create(s,
1256 				    NULL, 0, 2 * tsb_slab_size,
1257 				    sfmmu_tsb_segkmem_alloc,
1258 				    sfmmu_tsb_segkmem_free, kmem_bigtsb_arena,
1259 				    0, VM_SLEEP | VM_BESTFIT);
1260 			}
1261 
1262 			(void) sprintf(s, "kmem_tsb_lgrp%d", i);
1263 			kmem_tsb_default_arena[i] = vmem_create(s,
1264 			    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1265 			    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1266 			    VM_SLEEP | VM_BESTFIT);
1267 
1268 			(void) sprintf(s, "sfmmu_tsb_lgrp%d_cache", i);
1269 			sfmmu_tsb_cache[i] = kmem_cache_create(s,
1270 			    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1271 			    kmem_tsb_default_arena[i], 0);
1272 		}
1273 	} else {
1274 		if (use_bigtsb_arena) {
1275 			kmem_bigtsb_default_arena[0] =
1276 			    vmem_create("kmem_bigtsb_default", NULL, 0,
1277 			    2 * tsb_slab_size, sfmmu_tsb_segkmem_alloc,
1278 			    sfmmu_tsb_segkmem_free, kmem_bigtsb_arena, 0,
1279 			    VM_SLEEP | VM_BESTFIT);
1280 		}
1281 
1282 		kmem_tsb_default_arena[0] = vmem_create("kmem_tsb_default",
1283 		    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1284 		    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1285 		    VM_SLEEP | VM_BESTFIT);
1286 		sfmmu_tsb_cache[0] = kmem_cache_create("sfmmu_tsb_cache",
1287 		    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1288 		    kmem_tsb_default_arena[0], 0);
1289 	}
1290 
1291 	sfmmu8_cache = kmem_cache_create("sfmmu8_cache", HME8BLK_SZ,
1292 	    HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1293 	    sfmmu_hblkcache_destructor,
1294 	    sfmmu_hblkcache_reclaim, (void *)HME8BLK_SZ,
1295 	    hat_memload_arena, KMC_NOHASH);
1296 
1297 	hat_memload1_arena = vmem_create("hat_memload1", NULL, 0, PAGESIZE,
1298 	    segkmem_alloc_permanent, segkmem_free, heap_arena, 0,
1299 	    VMC_DUMPSAFE | VM_SLEEP);
1300 
1301 	sfmmu1_cache = kmem_cache_create("sfmmu1_cache", HME1BLK_SZ,
1302 	    HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1303 	    sfmmu_hblkcache_destructor,
1304 	    NULL, (void *)HME1BLK_SZ,
1305 	    hat_memload1_arena, KMC_NOHASH);
1306 
1307 	pa_hment_cache = kmem_cache_create("pa_hment_cache", PAHME_SZ,
1308 	    0, NULL, NULL, NULL, NULL, static_arena, KMC_NOHASH);
1309 
1310 	ism_blk_cache = kmem_cache_create("ism_blk_cache",
1311 	    sizeof (ism_blk_t), ecache_alignsize, NULL, NULL,
1312 	    NULL, NULL, static_arena, KMC_NOHASH);
1313 
1314 	ism_ment_cache = kmem_cache_create("ism_ment_cache",
1315 	    sizeof (ism_ment_t), 0, NULL, NULL,
1316 	    NULL, NULL, NULL, 0);
1317 
1318 	/*
1319 	 * We grab the first hat for the kernel,
1320 	 */
1321 	AS_LOCK_ENTER(&kas, RW_WRITER);
1322 	kas.a_hat = hat_alloc(&kas);
1323 	AS_LOCK_EXIT(&kas);
1324 
1325 	/*
1326 	 * Initialize hblk_reserve.
1327 	 */
1328 	((struct hme_blk *)hblk_reserve)->hblk_nextpa =
1329 	    va_to_pa((caddr_t)hblk_reserve);
1330 
1331 #ifndef UTSB_PHYS
1332 	/*
1333 	 * Reserve some kernel virtual address space for the locked TTEs
1334 	 * that allow us to probe the TSB from TL>0.
1335 	 */
1336 	utsb_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1337 	    0, 0, NULL, NULL, VM_SLEEP);
1338 	utsb4m_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1339 	    0, 0, NULL, NULL, VM_SLEEP);
1340 #endif
1341 
1342 #ifdef VAC
1343 	/*
1344 	 * The big page VAC handling code assumes VAC
1345 	 * will not be bigger than the smallest big
1346 	 * page- which is 64K.
1347 	 */
1348 	if (TTEPAGES(TTE64K) < CACHE_NUM_COLOR) {
1349 		cmn_err(CE_PANIC, "VAC too big!");
1350 	}
1351 #endif
1352 
1353 	(void) xhat_init();
1354 
1355 	uhme_hash_pa = va_to_pa(uhme_hash);
1356 	khme_hash_pa = va_to_pa(khme_hash);
1357 
1358 	/*
1359 	 * Initialize relocation locks. kpr_suspendlock is held
1360 	 * at PIL_MAX to prevent interrupts from pinning the holder
1361 	 * of a suspended TTE which may access it leading to a
1362 	 * deadlock condition.
1363 	 */
1364 	mutex_init(&kpr_mutex, NULL, MUTEX_DEFAULT, NULL);
1365 	mutex_init(&kpr_suspendlock, NULL, MUTEX_SPIN, (void *)PIL_MAX);
1366 
1367 	/*
1368 	 * If Shared context support is disabled via /etc/system
1369 	 * set shctx_on to 0 here if it was set to 1 earlier in boot
1370 	 * sequence by cpu module initialization code.
1371 	 */
1372 	if (shctx_on && disable_shctx) {
1373 		shctx_on = 0;
1374 	}
1375 
1376 	if (shctx_on) {
1377 		srd_buckets = kmem_zalloc(SFMMU_MAX_SRD_BUCKETS *
1378 		    sizeof (srd_buckets[0]), KM_SLEEP);
1379 		for (i = 0; i < SFMMU_MAX_SRD_BUCKETS; i++) {
1380 			mutex_init(&srd_buckets[i].srdb_lock, NULL,
1381 			    MUTEX_DEFAULT, NULL);
1382 		}
1383 
1384 		srd_cache = kmem_cache_create("srd_cache", sizeof (sf_srd_t),
1385 		    0, sfmmu_srdcache_constructor, sfmmu_srdcache_destructor,
1386 		    NULL, NULL, NULL, 0);
1387 		region_cache = kmem_cache_create("region_cache",
1388 		    sizeof (sf_region_t), 0, sfmmu_rgncache_constructor,
1389 		    sfmmu_rgncache_destructor, NULL, NULL, NULL, 0);
1390 		scd_cache = kmem_cache_create("scd_cache", sizeof (sf_scd_t),
1391 		    0, sfmmu_scdcache_constructor,  sfmmu_scdcache_destructor,
1392 		    NULL, NULL, NULL, 0);
1393 	}
1394 
1395 	/*
1396 	 * Pre-allocate hrm_hashtab before enabling the collection of
1397 	 * refmod statistics.  Allocating on the fly would mean us
1398 	 * running the risk of suffering recursive mutex enters or
1399 	 * deadlocks.
1400 	 */
1401 	hrm_hashtab = kmem_zalloc(HRM_HASHSIZE * sizeof (struct hrmstat *),
1402 	    KM_SLEEP);
1403 
1404 	/* Allocate per-cpu pending freelist of hmeblks */
1405 	cpu_hme_pend = kmem_zalloc((NCPU * sizeof (cpu_hme_pend_t)) + 64,
1406 	    KM_SLEEP);
1407 	cpu_hme_pend = (cpu_hme_pend_t *)P2ROUNDUP(
1408 	    (uintptr_t)cpu_hme_pend, 64);
1409 
1410 	for (i = 0; i < NCPU; i++) {
1411 		mutex_init(&cpu_hme_pend[i].chp_mutex, NULL, MUTEX_DEFAULT,
1412 		    NULL);
1413 	}
1414 
1415 	if (cpu_hme_pend_thresh == 0) {
1416 		cpu_hme_pend_thresh = CPU_HME_PEND_THRESH;
1417 	}
1418 }
1419 
1420 /*
1421  * Initialize locking for the hat layer, called early during boot.
1422  */
1423 static void
1424 hat_lock_init()
1425 {
1426 	int i;
1427 
1428 	/*
1429 	 * initialize the array of mutexes protecting a page's mapping
1430 	 * list and p_nrm field.
1431 	 */
1432 	for (i = 0; i < MML_TABLE_SIZE; i++)
1433 		mutex_init(&mml_table[i].pad_mutex, NULL, MUTEX_DEFAULT, NULL);
1434 
1435 	if (kpm_enable) {
1436 		for (i = 0; i < kpmp_table_sz; i++) {
1437 			mutex_init(&kpmp_table[i].khl_mutex, NULL,
1438 			    MUTEX_DEFAULT, NULL);
1439 		}
1440 	}
1441 
1442 	/*
1443 	 * Initialize array of mutex locks that protects sfmmu fields and
1444 	 * TSB lists.
1445 	 */
1446 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
1447 		mutex_init(HATLOCK_MUTEXP(&hat_lock[i]), NULL, MUTEX_DEFAULT,
1448 		    NULL);
1449 }
1450 
1451 #define	SFMMU_KERNEL_MAXVA \
1452 	(kmem64_base ? (uintptr_t)kmem64_end : (SYSLIMIT))
1453 
1454 /*
1455  * Allocate a hat structure.
1456  * Called when an address space first uses a hat.
1457  */
1458 struct hat *
1459 hat_alloc(struct as *as)
1460 {
1461 	sfmmu_t *sfmmup;
1462 	int i;
1463 	uint64_t cnum;
1464 	extern uint_t get_color_start(struct as *);
1465 
1466 	ASSERT(AS_WRITE_HELD(as));
1467 	sfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
1468 	sfmmup->sfmmu_as = as;
1469 	sfmmup->sfmmu_flags = 0;
1470 	sfmmup->sfmmu_tteflags = 0;
1471 	sfmmup->sfmmu_rtteflags = 0;
1472 	LOCK_INIT_CLEAR(&sfmmup->sfmmu_ctx_lock);
1473 
1474 	if (as == &kas) {
1475 		ksfmmup = sfmmup;
1476 		sfmmup->sfmmu_cext = 0;
1477 		cnum = KCONTEXT;
1478 
1479 		sfmmup->sfmmu_clrstart = 0;
1480 		sfmmup->sfmmu_tsb = NULL;
1481 		/*
1482 		 * hat_kern_setup() will call sfmmu_init_ktsbinfo()
1483 		 * to setup tsb_info for ksfmmup.
1484 		 */
1485 	} else {
1486 
1487 		/*
1488 		 * Just set to invalid ctx. When it faults, it will
1489 		 * get a valid ctx. This would avoid the situation
1490 		 * where we get a ctx, but it gets stolen and then
1491 		 * we fault when we try to run and so have to get
1492 		 * another ctx.
1493 		 */
1494 		sfmmup->sfmmu_cext = 0;
1495 		cnum = INVALID_CONTEXT;
1496 
1497 		/* initialize original physical page coloring bin */
1498 		sfmmup->sfmmu_clrstart = get_color_start(as);
1499 #ifdef DEBUG
1500 		if (tsb_random_size) {
1501 			uint32_t randval = (uint32_t)gettick() >> 4;
1502 			int size = randval % (tsb_max_growsize + 1);
1503 
1504 			/* chose a random tsb size for stress testing */
1505 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb, size,
1506 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1507 		} else
1508 #endif /* DEBUG */
1509 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb,
1510 			    default_tsb_size,
1511 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1512 		sfmmup->sfmmu_flags = HAT_SWAPPED | HAT_ALLCTX_INVALID;
1513 		ASSERT(sfmmup->sfmmu_tsb != NULL);
1514 	}
1515 
1516 	ASSERT(max_mmu_ctxdoms > 0);
1517 	for (i = 0; i < max_mmu_ctxdoms; i++) {
1518 		sfmmup->sfmmu_ctxs[i].cnum = cnum;
1519 		sfmmup->sfmmu_ctxs[i].gnum = 0;
1520 	}
1521 
1522 	for (i = 0; i < max_mmu_page_sizes; i++) {
1523 		sfmmup->sfmmu_ttecnt[i] = 0;
1524 		sfmmup->sfmmu_scdrttecnt[i] = 0;
1525 		sfmmup->sfmmu_ismttecnt[i] = 0;
1526 		sfmmup->sfmmu_scdismttecnt[i] = 0;
1527 		sfmmup->sfmmu_pgsz[i] = TTE8K;
1528 	}
1529 	sfmmup->sfmmu_tsb0_4minflcnt = 0;
1530 	sfmmup->sfmmu_iblk = NULL;
1531 	sfmmup->sfmmu_ismhat = 0;
1532 	sfmmup->sfmmu_scdhat = 0;
1533 	sfmmup->sfmmu_ismblkpa = (uint64_t)-1;
1534 	if (sfmmup == ksfmmup) {
1535 		CPUSET_ALL(sfmmup->sfmmu_cpusran);
1536 	} else {
1537 		CPUSET_ZERO(sfmmup->sfmmu_cpusran);
1538 	}
1539 	sfmmup->sfmmu_free = 0;
1540 	sfmmup->sfmmu_rmstat = 0;
1541 	sfmmup->sfmmu_clrbin = sfmmup->sfmmu_clrstart;
1542 	sfmmup->sfmmu_xhat_provider = NULL;
1543 	cv_init(&sfmmup->sfmmu_tsb_cv, NULL, CV_DEFAULT, NULL);
1544 	sfmmup->sfmmu_srdp = NULL;
1545 	SF_RGNMAP_ZERO(sfmmup->sfmmu_region_map);
1546 	bzero(sfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
1547 	sfmmup->sfmmu_scdp = NULL;
1548 	sfmmup->sfmmu_scd_link.next = NULL;
1549 	sfmmup->sfmmu_scd_link.prev = NULL;
1550 	return (sfmmup);
1551 }
1552 
1553 /*
1554  * Create per-MMU context domain kstats for a given MMU ctx.
1555  */
1556 static void
1557 sfmmu_mmu_kstat_create(mmu_ctx_t *mmu_ctxp)
1558 {
1559 	mmu_ctx_stat_t	stat;
1560 	kstat_t		*mmu_kstat;
1561 
1562 	ASSERT(MUTEX_HELD(&cpu_lock));
1563 	ASSERT(mmu_ctxp->mmu_kstat == NULL);
1564 
1565 	mmu_kstat = kstat_create("unix", mmu_ctxp->mmu_idx, "mmu_ctx",
1566 	    "hat", KSTAT_TYPE_NAMED, MMU_CTX_NUM_STATS, KSTAT_FLAG_VIRTUAL);
1567 
1568 	if (mmu_kstat == NULL) {
1569 		cmn_err(CE_WARN, "kstat_create for MMU %d failed",
1570 		    mmu_ctxp->mmu_idx);
1571 	} else {
1572 		mmu_kstat->ks_data = mmu_ctxp->mmu_kstat_data;
1573 		for (stat = 0; stat < MMU_CTX_NUM_STATS; stat++)
1574 			kstat_named_init(&mmu_ctxp->mmu_kstat_data[stat],
1575 			    mmu_ctx_kstat_names[stat], KSTAT_DATA_INT64);
1576 		mmu_ctxp->mmu_kstat = mmu_kstat;
1577 		kstat_install(mmu_kstat);
1578 	}
1579 }
1580 
1581 /*
1582  * plat_cpuid_to_mmu_ctx_info() is a platform interface that returns MMU
1583  * context domain information for a given CPU. If a platform does not
1584  * specify that interface, then the function below is used instead to return
1585  * default information. The defaults are as follows:
1586  *
1587  *	- The number of MMU context IDs supported on any CPU in the
1588  *	  system is 8K.
1589  *	- There is one MMU context domain per CPU.
1590  */
1591 /*ARGSUSED*/
1592 static void
1593 sfmmu_cpuid_to_mmu_ctx_info(processorid_t cpuid, mmu_ctx_info_t *infop)
1594 {
1595 	infop->mmu_nctxs = nctxs;
1596 	infop->mmu_idx = cpu[cpuid]->cpu_seqid;
1597 }
1598 
1599 /*
1600  * Called during CPU initialization to set the MMU context-related information
1601  * for a CPU.
1602  *
1603  * cpu_lock serializes accesses to mmu_ctxs and mmu_saved_gnum.
1604  */
1605 void
1606 sfmmu_cpu_init(cpu_t *cp)
1607 {
1608 	mmu_ctx_info_t	info;
1609 	mmu_ctx_t	*mmu_ctxp;
1610 
1611 	ASSERT(MUTEX_HELD(&cpu_lock));
1612 
1613 	if (&plat_cpuid_to_mmu_ctx_info == NULL)
1614 		sfmmu_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1615 	else
1616 		plat_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1617 
1618 	ASSERT(info.mmu_idx < max_mmu_ctxdoms);
1619 
1620 	if ((mmu_ctxp = mmu_ctxs_tbl[info.mmu_idx]) == NULL) {
1621 		/* Each mmu_ctx is cacheline aligned. */
1622 		mmu_ctxp = kmem_cache_alloc(mmuctxdom_cache, KM_SLEEP);
1623 		bzero(mmu_ctxp, sizeof (mmu_ctx_t));
1624 
1625 		mutex_init(&mmu_ctxp->mmu_lock, NULL, MUTEX_SPIN,
1626 		    (void *)ipltospl(DISP_LEVEL));
1627 		mmu_ctxp->mmu_idx = info.mmu_idx;
1628 		mmu_ctxp->mmu_nctxs = info.mmu_nctxs;
1629 		/*
1630 		 * Globally for lifetime of a system,
1631 		 * gnum must always increase.
1632 		 * mmu_saved_gnum is protected by the cpu_lock.
1633 		 */
1634 		mmu_ctxp->mmu_gnum = mmu_saved_gnum + 1;
1635 		mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
1636 
1637 		sfmmu_mmu_kstat_create(mmu_ctxp);
1638 
1639 		mmu_ctxs_tbl[info.mmu_idx] = mmu_ctxp;
1640 	} else {
1641 		ASSERT(mmu_ctxp->mmu_idx == info.mmu_idx);
1642 		ASSERT(mmu_ctxp->mmu_nctxs <= info.mmu_nctxs);
1643 	}
1644 
1645 	/*
1646 	 * The mmu_lock is acquired here to prevent races with
1647 	 * the wrap-around code.
1648 	 */
1649 	mutex_enter(&mmu_ctxp->mmu_lock);
1650 
1651 
1652 	mmu_ctxp->mmu_ncpus++;
1653 	CPUSET_ADD(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1654 	CPU_MMU_IDX(cp) = info.mmu_idx;
1655 	CPU_MMU_CTXP(cp) = mmu_ctxp;
1656 
1657 	mutex_exit(&mmu_ctxp->mmu_lock);
1658 }
1659 
1660 static void
1661 sfmmu_ctxdom_free(mmu_ctx_t *mmu_ctxp)
1662 {
1663 	ASSERT(MUTEX_HELD(&cpu_lock));
1664 	ASSERT(!MUTEX_HELD(&mmu_ctxp->mmu_lock));
1665 
1666 	mutex_destroy(&mmu_ctxp->mmu_lock);
1667 
1668 	if (mmu_ctxp->mmu_kstat)
1669 		kstat_delete(mmu_ctxp->mmu_kstat);
1670 
1671 	/* mmu_saved_gnum is protected by the cpu_lock. */
1672 	if (mmu_saved_gnum < mmu_ctxp->mmu_gnum)
1673 		mmu_saved_gnum = mmu_ctxp->mmu_gnum;
1674 
1675 	kmem_cache_free(mmuctxdom_cache, mmu_ctxp);
1676 }
1677 
1678 /*
1679  * Called to perform MMU context-related cleanup for a CPU.
1680  */
1681 void
1682 sfmmu_cpu_cleanup(cpu_t *cp)
1683 {
1684 	mmu_ctx_t	*mmu_ctxp;
1685 
1686 	ASSERT(MUTEX_HELD(&cpu_lock));
1687 
1688 	mmu_ctxp = CPU_MMU_CTXP(cp);
1689 	ASSERT(mmu_ctxp != NULL);
1690 
1691 	/*
1692 	 * The mmu_lock is acquired here to prevent races with
1693 	 * the wrap-around code.
1694 	 */
1695 	mutex_enter(&mmu_ctxp->mmu_lock);
1696 
1697 	CPU_MMU_CTXP(cp) = NULL;
1698 
1699 	CPUSET_DEL(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1700 	if (--mmu_ctxp->mmu_ncpus == 0) {
1701 		mmu_ctxs_tbl[mmu_ctxp->mmu_idx] = NULL;
1702 		mutex_exit(&mmu_ctxp->mmu_lock);
1703 		sfmmu_ctxdom_free(mmu_ctxp);
1704 		return;
1705 	}
1706 
1707 	mutex_exit(&mmu_ctxp->mmu_lock);
1708 }
1709 
1710 uint_t
1711 sfmmu_ctxdom_nctxs(int idx)
1712 {
1713 	return (mmu_ctxs_tbl[idx]->mmu_nctxs);
1714 }
1715 
1716 #ifdef sun4v
1717 /*
1718  * sfmmu_ctxdoms_* is an interface provided to help keep context domains
1719  * consistant after suspend/resume on system that can resume on a different
1720  * hardware than it was suspended.
1721  *
1722  * sfmmu_ctxdom_lock(void) locks all context domains and prevents new contexts
1723  * from being allocated.  It acquires all hat_locks, which blocks most access to
1724  * context data, except for a few cases that are handled separately or are
1725  * harmless.  It wraps each domain to increment gnum and invalidate on-CPU
1726  * contexts, and forces cnum to its max.  As a result of this call all user
1727  * threads that are running on CPUs trap and try to perform wrap around but
1728  * can't because hat_locks are taken.  Threads that were not on CPUs but started
1729  * by scheduler go to sfmmu_alloc_ctx() to aquire context without checking
1730  * hat_lock, but fail, because cnum == nctxs, and therefore also trap and block
1731  * on hat_lock trying to wrap.  sfmmu_ctxdom_lock() must be called before CPUs
1732  * are paused, else it could deadlock acquiring locks held by paused CPUs.
1733  *
1734  * sfmmu_ctxdoms_remove() removes context domains from every CPUs and records
1735  * the CPUs that had them.  It must be called after CPUs have been paused. This
1736  * ensures that no threads are in sfmmu_alloc_ctx() accessing domain data,
1737  * because pause_cpus sends a mondo interrupt to every CPU, and sfmmu_alloc_ctx
1738  * runs with interrupts disabled.  When CPUs are later resumed, they may enter
1739  * sfmmu_alloc_ctx, but it will check for CPU_MMU_CTXP = NULL and immediately
1740  * return failure.  Or, they will be blocked trying to acquire hat_lock. Thus
1741  * after sfmmu_ctxdoms_remove returns, we are guaranteed that no one is
1742  * accessing the old context domains.
1743  *
1744  * sfmmu_ctxdoms_update(void) frees space used by old context domains and
1745  * allocates new context domains based on hardware layout.  It initializes
1746  * every CPU that had context domain before migration to have one again.
1747  * sfmmu_ctxdoms_update must be called after CPUs are resumed, else it
1748  * could deadlock acquiring locks held by paused CPUs.
1749  *
1750  * sfmmu_ctxdoms_unlock(void) releases all hat_locks after which user threads
1751  * acquire new context ids and continue execution.
1752  *
1753  * Therefore functions should be called in the following order:
1754  *       suspend_routine()
1755  *		sfmmu_ctxdom_lock()
1756  *		pause_cpus()
1757  *		suspend()
1758  *			if (suspend failed)
1759  *				sfmmu_ctxdom_unlock()
1760  *		...
1761  *		sfmmu_ctxdom_remove()
1762  *		resume_cpus()
1763  *		sfmmu_ctxdom_update()
1764  *		sfmmu_ctxdom_unlock()
1765  */
1766 static cpuset_t sfmmu_ctxdoms_pset;
1767 
1768 void
1769 sfmmu_ctxdoms_remove()
1770 {
1771 	processorid_t	id;
1772 	cpu_t		*cp;
1773 
1774 	/*
1775 	 * Record the CPUs that have domains in sfmmu_ctxdoms_pset, so they can
1776 	 * be restored post-migration. A CPU may be powered off and not have a
1777 	 * domain, for example.
1778 	 */
1779 	CPUSET_ZERO(sfmmu_ctxdoms_pset);
1780 
1781 	for (id = 0; id < NCPU; id++) {
1782 		if ((cp = cpu[id]) != NULL && CPU_MMU_CTXP(cp) != NULL) {
1783 			CPUSET_ADD(sfmmu_ctxdoms_pset, id);
1784 			CPU_MMU_CTXP(cp) = NULL;
1785 		}
1786 	}
1787 }
1788 
1789 void
1790 sfmmu_ctxdoms_lock(void)
1791 {
1792 	int		idx;
1793 	mmu_ctx_t	*mmu_ctxp;
1794 
1795 	sfmmu_hat_lock_all();
1796 
1797 	/*
1798 	 * At this point, no thread can be in sfmmu_ctx_wrap_around, because
1799 	 * hat_lock is always taken before calling it.
1800 	 *
1801 	 * For each domain, set mmu_cnum to max so no more contexts can be
1802 	 * allocated, and wrap to flush on-CPU contexts and force threads to
1803 	 * acquire a new context when we later drop hat_lock after migration.
1804 	 * Setting mmu_cnum may race with sfmmu_alloc_ctx which also sets cnum,
1805 	 * but the latter uses CAS and will miscompare and not overwrite it.
1806 	 */
1807 	kpreempt_disable(); /* required by sfmmu_ctx_wrap_around */
1808 	for (idx = 0; idx < max_mmu_ctxdoms; idx++) {
1809 		if ((mmu_ctxp = mmu_ctxs_tbl[idx]) != NULL) {
1810 			mutex_enter(&mmu_ctxp->mmu_lock);
1811 			mmu_ctxp->mmu_cnum = mmu_ctxp->mmu_nctxs;
1812 			/* make sure updated cnum visible */
1813 			membar_enter();
1814 			mutex_exit(&mmu_ctxp->mmu_lock);
1815 			sfmmu_ctx_wrap_around(mmu_ctxp, B_FALSE);
1816 		}
1817 	}
1818 	kpreempt_enable();
1819 }
1820 
1821 void
1822 sfmmu_ctxdoms_unlock(void)
1823 {
1824 	sfmmu_hat_unlock_all();
1825 }
1826 
1827 void
1828 sfmmu_ctxdoms_update(void)
1829 {
1830 	processorid_t	id;
1831 	cpu_t		*cp;
1832 	uint_t		idx;
1833 	mmu_ctx_t	*mmu_ctxp;
1834 
1835 	/*
1836 	 * Free all context domains.  As side effect, this increases
1837 	 * mmu_saved_gnum to the maximum gnum over all domains, which is used to
1838 	 * init gnum in the new domains, which therefore will be larger than the
1839 	 * sfmmu gnum for any process, guaranteeing that every process will see
1840 	 * a new generation and allocate a new context regardless of what new
1841 	 * domain it runs in.
1842 	 */
1843 	mutex_enter(&cpu_lock);
1844 
1845 	for (idx = 0; idx < max_mmu_ctxdoms; idx++) {
1846 		if (mmu_ctxs_tbl[idx] != NULL) {
1847 			mmu_ctxp = mmu_ctxs_tbl[idx];
1848 			mmu_ctxs_tbl[idx] = NULL;
1849 			sfmmu_ctxdom_free(mmu_ctxp);
1850 		}
1851 	}
1852 
1853 	for (id = 0; id < NCPU; id++) {
1854 		if (CPU_IN_SET(sfmmu_ctxdoms_pset, id) &&
1855 		    (cp = cpu[id]) != NULL)
1856 			sfmmu_cpu_init(cp);
1857 	}
1858 	mutex_exit(&cpu_lock);
1859 }
1860 #endif
1861 
1862 /*
1863  * Hat_setup, makes an address space context the current active one.
1864  * In sfmmu this translates to setting the secondary context with the
1865  * corresponding context.
1866  */
1867 void
1868 hat_setup(struct hat *sfmmup, int allocflag)
1869 {
1870 	hatlock_t *hatlockp;
1871 
1872 	/* Init needs some special treatment. */
1873 	if (allocflag == HAT_INIT) {
1874 		/*
1875 		 * Make sure that we have
1876 		 * 1. a TSB
1877 		 * 2. a valid ctx that doesn't get stolen after this point.
1878 		 */
1879 		hatlockp = sfmmu_hat_enter(sfmmup);
1880 
1881 		/*
1882 		 * Swap in the TSB.  hat_init() allocates tsbinfos without
1883 		 * TSBs, but we need one for init, since the kernel does some
1884 		 * special things to set up its stack and needs the TSB to
1885 		 * resolve page faults.
1886 		 */
1887 		sfmmu_tsb_swapin(sfmmup, hatlockp);
1888 
1889 		sfmmu_get_ctx(sfmmup);
1890 
1891 		sfmmu_hat_exit(hatlockp);
1892 	} else {
1893 		ASSERT(allocflag == HAT_ALLOC);
1894 
1895 		hatlockp = sfmmu_hat_enter(sfmmup);
1896 		kpreempt_disable();
1897 
1898 		CPUSET_ADD(sfmmup->sfmmu_cpusran, CPU->cpu_id);
1899 		/*
1900 		 * sfmmu_setctx_sec takes <pgsz|cnum> as a parameter,
1901 		 * pagesize bits don't matter in this case since we are passing
1902 		 * INVALID_CONTEXT to it.
1903 		 * Compatibility Note: hw takes care of MMU_SCONTEXT1
1904 		 */
1905 		sfmmu_setctx_sec(INVALID_CONTEXT);
1906 		sfmmu_clear_utsbinfo();
1907 
1908 		kpreempt_enable();
1909 		sfmmu_hat_exit(hatlockp);
1910 	}
1911 }
1912 
1913 /*
1914  * Free all the translation resources for the specified address space.
1915  * Called from as_free when an address space is being destroyed.
1916  */
1917 void
1918 hat_free_start(struct hat *sfmmup)
1919 {
1920 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as));
1921 	ASSERT(sfmmup != ksfmmup);
1922 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1923 
1924 	sfmmup->sfmmu_free = 1;
1925 	if (sfmmup->sfmmu_scdp != NULL) {
1926 		sfmmu_leave_scd(sfmmup, 0);
1927 	}
1928 
1929 	ASSERT(sfmmup->sfmmu_scdp == NULL);
1930 }
1931 
1932 void
1933 hat_free_end(struct hat *sfmmup)
1934 {
1935 	int i;
1936 
1937 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1938 	ASSERT(sfmmup->sfmmu_free == 1);
1939 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
1940 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
1941 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
1942 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
1943 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
1944 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
1945 
1946 	if (sfmmup->sfmmu_rmstat) {
1947 		hat_freestat(sfmmup->sfmmu_as, NULL);
1948 	}
1949 
1950 	while (sfmmup->sfmmu_tsb != NULL) {
1951 		struct tsb_info *next = sfmmup->sfmmu_tsb->tsb_next;
1952 		sfmmu_tsbinfo_free(sfmmup->sfmmu_tsb);
1953 		sfmmup->sfmmu_tsb = next;
1954 	}
1955 
1956 	if (sfmmup->sfmmu_srdp != NULL) {
1957 		sfmmu_leave_srd(sfmmup);
1958 		ASSERT(sfmmup->sfmmu_srdp == NULL);
1959 		for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1960 			if (sfmmup->sfmmu_hmeregion_links[i] != NULL) {
1961 				kmem_free(sfmmup->sfmmu_hmeregion_links[i],
1962 				    SFMMU_L2_HMERLINKS_SIZE);
1963 				sfmmup->sfmmu_hmeregion_links[i] = NULL;
1964 			}
1965 		}
1966 	}
1967 	sfmmu_free_sfmmu(sfmmup);
1968 
1969 #ifdef DEBUG
1970 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1971 		ASSERT(sfmmup->sfmmu_hmeregion_links[i] == NULL);
1972 	}
1973 #endif
1974 
1975 	kmem_cache_free(sfmmuid_cache, sfmmup);
1976 }
1977 
1978 /*
1979  * Set up any translation structures, for the specified address space,
1980  * that are needed or preferred when the process is being swapped in.
1981  */
1982 /* ARGSUSED */
1983 void
1984 hat_swapin(struct hat *hat)
1985 {
1986 	ASSERT(hat->sfmmu_xhat_provider == NULL);
1987 }
1988 
1989 /*
1990  * Free all of the translation resources, for the specified address space,
1991  * that can be freed while the process is swapped out. Called from as_swapout.
1992  * Also, free up the ctx that this process was using.
1993  */
1994 void
1995 hat_swapout(struct hat *sfmmup)
1996 {
1997 	struct hmehash_bucket *hmebp;
1998 	struct hme_blk *hmeblkp;
1999 	struct hme_blk *pr_hblk = NULL;
2000 	struct hme_blk *nx_hblk;
2001 	int i;
2002 	struct hme_blk *list = NULL;
2003 	hatlock_t *hatlockp;
2004 	struct tsb_info *tsbinfop;
2005 	struct free_tsb {
2006 		struct free_tsb *next;
2007 		struct tsb_info *tsbinfop;
2008 	};			/* free list of TSBs */
2009 	struct free_tsb *freelist, *last, *next;
2010 
2011 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
2012 	SFMMU_STAT(sf_swapout);
2013 
2014 	/*
2015 	 * There is no way to go from an as to all its translations in sfmmu.
2016 	 * Here is one of the times when we take the big hit and traverse
2017 	 * the hash looking for hme_blks to free up.  Not only do we free up
2018 	 * this as hme_blks but all those that are free.  We are obviously
2019 	 * swapping because we need memory so let's free up as much
2020 	 * as we can.
2021 	 *
2022 	 * Note that we don't flush TLB/TSB here -- it's not necessary
2023 	 * because:
2024 	 *  1) we free the ctx we're using and throw away the TSB(s);
2025 	 *  2) processes aren't runnable while being swapped out.
2026 	 */
2027 	ASSERT(sfmmup != KHATID);
2028 	for (i = 0; i <= UHMEHASH_SZ; i++) {
2029 		hmebp = &uhme_hash[i];
2030 		SFMMU_HASH_LOCK(hmebp);
2031 		hmeblkp = hmebp->hmeblkp;
2032 		pr_hblk = NULL;
2033 		while (hmeblkp) {
2034 
2035 			ASSERT(!hmeblkp->hblk_xhat_bit);
2036 
2037 			if ((hmeblkp->hblk_tag.htag_id == sfmmup) &&
2038 			    !hmeblkp->hblk_shw_bit && !hmeblkp->hblk_lckcnt) {
2039 				ASSERT(!hmeblkp->hblk_shared);
2040 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
2041 				    (caddr_t)get_hblk_base(hmeblkp),
2042 				    get_hblk_endaddr(hmeblkp),
2043 				    NULL, HAT_UNLOAD);
2044 			}
2045 			nx_hblk = hmeblkp->hblk_next;
2046 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
2047 				ASSERT(!hmeblkp->hblk_lckcnt);
2048 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
2049 				    &list, 0);
2050 			} else {
2051 				pr_hblk = hmeblkp;
2052 			}
2053 			hmeblkp = nx_hblk;
2054 		}
2055 		SFMMU_HASH_UNLOCK(hmebp);
2056 	}
2057 
2058 	sfmmu_hblks_list_purge(&list, 0);
2059 
2060 	/*
2061 	 * Now free up the ctx so that others can reuse it.
2062 	 */
2063 	hatlockp = sfmmu_hat_enter(sfmmup);
2064 
2065 	sfmmu_invalidate_ctx(sfmmup);
2066 
2067 	/*
2068 	 * Free TSBs, but not tsbinfos, and set SWAPPED flag.
2069 	 * If TSBs were never swapped in, just return.
2070 	 * This implies that we don't support partial swapping
2071 	 * of TSBs -- either all are swapped out, or none are.
2072 	 *
2073 	 * We must hold the HAT lock here to prevent racing with another
2074 	 * thread trying to unmap TTEs from the TSB or running the post-
2075 	 * relocator after relocating the TSB's memory.  Unfortunately, we
2076 	 * can't free memory while holding the HAT lock or we could
2077 	 * deadlock, so we build a list of TSBs to be freed after marking
2078 	 * the tsbinfos as swapped out and free them after dropping the
2079 	 * lock.
2080 	 */
2081 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
2082 		sfmmu_hat_exit(hatlockp);
2083 		return;
2084 	}
2085 
2086 	SFMMU_FLAGS_SET(sfmmup, HAT_SWAPPED);
2087 	last = freelist = NULL;
2088 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
2089 	    tsbinfop = tsbinfop->tsb_next) {
2090 		ASSERT((tsbinfop->tsb_flags & TSB_SWAPPED) == 0);
2091 
2092 		/*
2093 		 * Cast the TSB into a struct free_tsb and put it on the free
2094 		 * list.
2095 		 */
2096 		if (freelist == NULL) {
2097 			last = freelist = (struct free_tsb *)tsbinfop->tsb_va;
2098 		} else {
2099 			last->next = (struct free_tsb *)tsbinfop->tsb_va;
2100 			last = last->next;
2101 		}
2102 		last->next = NULL;
2103 		last->tsbinfop = tsbinfop;
2104 		tsbinfop->tsb_flags |= TSB_SWAPPED;
2105 		/*
2106 		 * Zero out the TTE to clear the valid bit.
2107 		 * Note we can't use a value like 0xbad because we want to
2108 		 * ensure diagnostic bits are NEVER set on TTEs that might
2109 		 * be loaded.  The intent is to catch any invalid access
2110 		 * to the swapped TSB, such as a thread running with a valid
2111 		 * context without first calling sfmmu_tsb_swapin() to
2112 		 * allocate TSB memory.
2113 		 */
2114 		tsbinfop->tsb_tte.ll = 0;
2115 	}
2116 
2117 	/* Now we can drop the lock and free the TSB memory. */
2118 	sfmmu_hat_exit(hatlockp);
2119 	for (; freelist != NULL; freelist = next) {
2120 		next = freelist->next;
2121 		sfmmu_tsb_free(freelist->tsbinfop);
2122 	}
2123 }
2124 
2125 /*
2126  * Duplicate the translations of an as into another newas
2127  */
2128 /* ARGSUSED */
2129 int
2130 hat_dup(struct hat *hat, struct hat *newhat, caddr_t addr, size_t len,
2131 	uint_t flag)
2132 {
2133 	sf_srd_t *srdp;
2134 	sf_scd_t *scdp;
2135 	int i;
2136 	extern uint_t get_color_start(struct as *);
2137 
2138 	ASSERT(hat->sfmmu_xhat_provider == NULL);
2139 	ASSERT((flag == 0) || (flag == HAT_DUP_ALL) || (flag == HAT_DUP_COW) ||
2140 	    (flag == HAT_DUP_SRD));
2141 	ASSERT(hat != ksfmmup);
2142 	ASSERT(newhat != ksfmmup);
2143 	ASSERT(flag != HAT_DUP_ALL || hat->sfmmu_srdp == newhat->sfmmu_srdp);
2144 
2145 	if (flag == HAT_DUP_COW) {
2146 		panic("hat_dup: HAT_DUP_COW not supported");
2147 	}
2148 
2149 	if (flag == HAT_DUP_SRD && ((srdp = hat->sfmmu_srdp) != NULL)) {
2150 		ASSERT(srdp->srd_evp != NULL);
2151 		VN_HOLD(srdp->srd_evp);
2152 		ASSERT(srdp->srd_refcnt > 0);
2153 		newhat->sfmmu_srdp = srdp;
2154 		atomic_inc_32((volatile uint_t *)&srdp->srd_refcnt);
2155 	}
2156 
2157 	/*
2158 	 * HAT_DUP_ALL flag is used after as duplication is done.
2159 	 */
2160 	if (flag == HAT_DUP_ALL && ((srdp = newhat->sfmmu_srdp) != NULL)) {
2161 		ASSERT(newhat->sfmmu_srdp->srd_refcnt >= 2);
2162 		newhat->sfmmu_rtteflags = hat->sfmmu_rtteflags;
2163 		if (hat->sfmmu_flags & HAT_4MTEXT_FLAG) {
2164 			newhat->sfmmu_flags |= HAT_4MTEXT_FLAG;
2165 		}
2166 
2167 		/* check if need to join scd */
2168 		if ((scdp = hat->sfmmu_scdp) != NULL &&
2169 		    newhat->sfmmu_scdp != scdp) {
2170 			int ret;
2171 			SF_RGNMAP_IS_SUBSET(&newhat->sfmmu_region_map,
2172 			    &scdp->scd_region_map, ret);
2173 			ASSERT(ret);
2174 			sfmmu_join_scd(scdp, newhat);
2175 			ASSERT(newhat->sfmmu_scdp == scdp &&
2176 			    scdp->scd_refcnt >= 2);
2177 			for (i = 0; i < max_mmu_page_sizes; i++) {
2178 				newhat->sfmmu_ismttecnt[i] =
2179 				    hat->sfmmu_ismttecnt[i];
2180 				newhat->sfmmu_scdismttecnt[i] =
2181 				    hat->sfmmu_scdismttecnt[i];
2182 			}
2183 		}
2184 
2185 		sfmmu_check_page_sizes(newhat, 1);
2186 	}
2187 
2188 	if (flag == HAT_DUP_ALL && consistent_coloring == 0 &&
2189 	    update_proc_pgcolorbase_after_fork != 0) {
2190 		hat->sfmmu_clrbin = get_color_start(hat->sfmmu_as);
2191 	}
2192 	return (0);
2193 }
2194 
2195 void
2196 hat_memload(struct hat *hat, caddr_t addr, struct page *pp,
2197 	uint_t attr, uint_t flags)
2198 {
2199 	hat_do_memload(hat, addr, pp, attr, flags,
2200 	    SFMMU_INVALID_SHMERID);
2201 }
2202 
2203 void
2204 hat_memload_region(struct hat *hat, caddr_t addr, struct page *pp,
2205 	uint_t attr, uint_t flags, hat_region_cookie_t rcookie)
2206 {
2207 	uint_t rid;
2208 	if (rcookie == HAT_INVALID_REGION_COOKIE ||
2209 	    hat->sfmmu_xhat_provider != NULL) {
2210 		hat_do_memload(hat, addr, pp, attr, flags,
2211 		    SFMMU_INVALID_SHMERID);
2212 		return;
2213 	}
2214 	rid = (uint_t)((uint64_t)rcookie);
2215 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2216 	hat_do_memload(hat, addr, pp, attr, flags, rid);
2217 }
2218 
2219 /*
2220  * Set up addr to map to page pp with protection prot.
2221  * As an optimization we also load the TSB with the
2222  * corresponding tte but it is no big deal if  the tte gets kicked out.
2223  */
2224 static void
2225 hat_do_memload(struct hat *hat, caddr_t addr, struct page *pp,
2226 	uint_t attr, uint_t flags, uint_t rid)
2227 {
2228 	tte_t tte;
2229 
2230 
2231 	ASSERT(hat != NULL);
2232 	ASSERT(PAGE_LOCKED(pp));
2233 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2234 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2235 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2236 	SFMMU_VALIDATE_HMERID(hat, rid, addr, MMU_PAGESIZE);
2237 
2238 	if (PP_ISFREE(pp)) {
2239 		panic("hat_memload: loading a mapping to free page %p",
2240 		    (void *)pp);
2241 	}
2242 
2243 	if (hat->sfmmu_xhat_provider) {
2244 		/* no regions for xhats */
2245 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
2246 		XHAT_MEMLOAD(hat, addr, pp, attr, flags);
2247 		return;
2248 	}
2249 
2250 	ASSERT((hat == ksfmmup) || AS_LOCK_HELD(hat->sfmmu_as));
2251 
2252 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2253 		cmn_err(CE_NOTE, "hat_memload: unsupported flags %d",
2254 		    flags & ~SFMMU_LOAD_ALLFLAG);
2255 
2256 	if (hat->sfmmu_rmstat)
2257 		hat_resvstat(MMU_PAGESIZE, hat->sfmmu_as, addr);
2258 
2259 #if defined(SF_ERRATA_57)
2260 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2261 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2262 	    !(flags & HAT_LOAD_SHARE)) {
2263 		cmn_err(CE_WARN, "hat_memload: illegal attempt to make user "
2264 		    " page executable");
2265 		attr &= ~PROT_EXEC;
2266 	}
2267 #endif
2268 
2269 	sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2270 	(void) sfmmu_tteload_array(hat, &tte, addr, &pp, flags, rid);
2271 
2272 	/*
2273 	 * Check TSB and TLB page sizes.
2274 	 */
2275 	if ((flags & HAT_LOAD_SHARE) == 0) {
2276 		sfmmu_check_page_sizes(hat, 1);
2277 	}
2278 }
2279 
2280 /*
2281  * hat_devload can be called to map real memory (e.g.
2282  * /dev/kmem) and even though hat_devload will determine pf is
2283  * for memory, it will be unable to get a shared lock on the
2284  * page (because someone else has it exclusively) and will
2285  * pass dp = NULL.  If tteload doesn't get a non-NULL
2286  * page pointer it can't cache memory.
2287  */
2288 void
2289 hat_devload(struct hat *hat, caddr_t addr, size_t len, pfn_t pfn,
2290 	uint_t attr, int flags)
2291 {
2292 	tte_t tte;
2293 	struct page *pp = NULL;
2294 	int use_lgpg = 0;
2295 
2296 	ASSERT(hat != NULL);
2297 
2298 	if (hat->sfmmu_xhat_provider) {
2299 		XHAT_DEVLOAD(hat, addr, len, pfn, attr, flags);
2300 		return;
2301 	}
2302 
2303 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2304 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2305 	ASSERT((hat == ksfmmup) || AS_LOCK_HELD(hat->sfmmu_as));
2306 	if (len == 0)
2307 		panic("hat_devload: zero len");
2308 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2309 		cmn_err(CE_NOTE, "hat_devload: unsupported flags %d",
2310 		    flags & ~SFMMU_LOAD_ALLFLAG);
2311 
2312 #if defined(SF_ERRATA_57)
2313 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2314 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2315 	    !(flags & HAT_LOAD_SHARE)) {
2316 		cmn_err(CE_WARN, "hat_devload: illegal attempt to make user "
2317 		    " page executable");
2318 		attr &= ~PROT_EXEC;
2319 	}
2320 #endif
2321 
2322 	/*
2323 	 * If it's a memory page find its pp
2324 	 */
2325 	if (!(flags & HAT_LOAD_NOCONSIST) && pf_is_memory(pfn)) {
2326 		pp = page_numtopp_nolock(pfn);
2327 		if (pp == NULL) {
2328 			flags |= HAT_LOAD_NOCONSIST;
2329 		} else {
2330 			if (PP_ISFREE(pp)) {
2331 				panic("hat_memload: loading "
2332 				    "a mapping to free page %p",
2333 				    (void *)pp);
2334 			}
2335 			if (!PAGE_LOCKED(pp) && !PP_ISNORELOC(pp)) {
2336 				panic("hat_memload: loading a mapping "
2337 				    "to unlocked relocatable page %p",
2338 				    (void *)pp);
2339 			}
2340 			ASSERT(len == MMU_PAGESIZE);
2341 		}
2342 	}
2343 
2344 	if (hat->sfmmu_rmstat)
2345 		hat_resvstat(len, hat->sfmmu_as, addr);
2346 
2347 	if (flags & HAT_LOAD_NOCONSIST) {
2348 		attr |= SFMMU_UNCACHEVTTE;
2349 		use_lgpg = 1;
2350 	}
2351 	if (!pf_is_memory(pfn)) {
2352 		attr |= SFMMU_UNCACHEPTTE | HAT_NOSYNC;
2353 		use_lgpg = 1;
2354 		switch (attr & HAT_ORDER_MASK) {
2355 			case HAT_STRICTORDER:
2356 			case HAT_UNORDERED_OK:
2357 				/*
2358 				 * we set the side effect bit for all non
2359 				 * memory mappings unless merging is ok
2360 				 */
2361 				attr |= SFMMU_SIDEFFECT;
2362 				break;
2363 			case HAT_MERGING_OK:
2364 			case HAT_LOADCACHING_OK:
2365 			case HAT_STORECACHING_OK:
2366 				break;
2367 			default:
2368 				panic("hat_devload: bad attr");
2369 				break;
2370 		}
2371 	}
2372 	while (len) {
2373 		if (!use_lgpg) {
2374 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2375 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2376 			    flags, SFMMU_INVALID_SHMERID);
2377 			len -= MMU_PAGESIZE;
2378 			addr += MMU_PAGESIZE;
2379 			pfn++;
2380 			continue;
2381 		}
2382 		/*
2383 		 *  try to use large pages, check va/pa alignments
2384 		 *  Note that 32M/256M page sizes are not (yet) supported.
2385 		 */
2386 		if ((len >= MMU_PAGESIZE4M) &&
2387 		    !((uintptr_t)addr & MMU_PAGEOFFSET4M) &&
2388 		    !(disable_large_pages & (1 << TTE4M)) &&
2389 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET4M)) {
2390 			sfmmu_memtte(&tte, pfn, attr, TTE4M);
2391 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2392 			    flags, SFMMU_INVALID_SHMERID);
2393 			len -= MMU_PAGESIZE4M;
2394 			addr += MMU_PAGESIZE4M;
2395 			pfn += MMU_PAGESIZE4M / MMU_PAGESIZE;
2396 		} else if ((len >= MMU_PAGESIZE512K) &&
2397 		    !((uintptr_t)addr & MMU_PAGEOFFSET512K) &&
2398 		    !(disable_large_pages & (1 << TTE512K)) &&
2399 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET512K)) {
2400 			sfmmu_memtte(&tte, pfn, attr, TTE512K);
2401 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2402 			    flags, SFMMU_INVALID_SHMERID);
2403 			len -= MMU_PAGESIZE512K;
2404 			addr += MMU_PAGESIZE512K;
2405 			pfn += MMU_PAGESIZE512K / MMU_PAGESIZE;
2406 		} else if ((len >= MMU_PAGESIZE64K) &&
2407 		    !((uintptr_t)addr & MMU_PAGEOFFSET64K) &&
2408 		    !(disable_large_pages & (1 << TTE64K)) &&
2409 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET64K)) {
2410 			sfmmu_memtte(&tte, pfn, attr, TTE64K);
2411 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2412 			    flags, SFMMU_INVALID_SHMERID);
2413 			len -= MMU_PAGESIZE64K;
2414 			addr += MMU_PAGESIZE64K;
2415 			pfn += MMU_PAGESIZE64K / MMU_PAGESIZE;
2416 		} else {
2417 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2418 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2419 			    flags, SFMMU_INVALID_SHMERID);
2420 			len -= MMU_PAGESIZE;
2421 			addr += MMU_PAGESIZE;
2422 			pfn++;
2423 		}
2424 	}
2425 
2426 	/*
2427 	 * Check TSB and TLB page sizes.
2428 	 */
2429 	if ((flags & HAT_LOAD_SHARE) == 0) {
2430 		sfmmu_check_page_sizes(hat, 1);
2431 	}
2432 }
2433 
2434 void
2435 hat_memload_array(struct hat *hat, caddr_t addr, size_t len,
2436 	struct page **pps, uint_t attr, uint_t flags)
2437 {
2438 	hat_do_memload_array(hat, addr, len, pps, attr, flags,
2439 	    SFMMU_INVALID_SHMERID);
2440 }
2441 
2442 void
2443 hat_memload_array_region(struct hat *hat, caddr_t addr, size_t len,
2444 	struct page **pps, uint_t attr, uint_t flags,
2445 	hat_region_cookie_t rcookie)
2446 {
2447 	uint_t rid;
2448 	if (rcookie == HAT_INVALID_REGION_COOKIE ||
2449 	    hat->sfmmu_xhat_provider != NULL) {
2450 		hat_do_memload_array(hat, addr, len, pps, attr, flags,
2451 		    SFMMU_INVALID_SHMERID);
2452 		return;
2453 	}
2454 	rid = (uint_t)((uint64_t)rcookie);
2455 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2456 	hat_do_memload_array(hat, addr, len, pps, attr, flags, rid);
2457 }
2458 
2459 /*
2460  * Map the largest extend possible out of the page array. The array may NOT
2461  * be in order.  The largest possible mapping a page can have
2462  * is specified in the p_szc field.  The p_szc field
2463  * cannot change as long as there any mappings (large or small)
2464  * to any of the pages that make up the large page. (ie. any
2465  * promotion/demotion of page size is not up to the hat but up to
2466  * the page free list manager).  The array
2467  * should consist of properly aligned contigous pages that are
2468  * part of a big page for a large mapping to be created.
2469  */
2470 static void
2471 hat_do_memload_array(struct hat *hat, caddr_t addr, size_t len,
2472 	struct page **pps, uint_t attr, uint_t flags, uint_t rid)
2473 {
2474 	int  ttesz;
2475 	size_t mapsz;
2476 	pgcnt_t	numpg, npgs;
2477 	tte_t tte;
2478 	page_t *pp;
2479 	uint_t large_pages_disable;
2480 
2481 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2482 	SFMMU_VALIDATE_HMERID(hat, rid, addr, len);
2483 
2484 	if (hat->sfmmu_xhat_provider) {
2485 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
2486 		XHAT_MEMLOAD_ARRAY(hat, addr, len, pps, attr, flags);
2487 		return;
2488 	}
2489 
2490 	if (hat->sfmmu_rmstat)
2491 		hat_resvstat(len, hat->sfmmu_as, addr);
2492 
2493 #if defined(SF_ERRATA_57)
2494 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2495 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2496 	    !(flags & HAT_LOAD_SHARE)) {
2497 		cmn_err(CE_WARN, "hat_memload_array: illegal attempt to make "
2498 		    "user page executable");
2499 		attr &= ~PROT_EXEC;
2500 	}
2501 #endif
2502 
2503 	/* Get number of pages */
2504 	npgs = len >> MMU_PAGESHIFT;
2505 
2506 	if (flags & HAT_LOAD_SHARE) {
2507 		large_pages_disable = disable_ism_large_pages;
2508 	} else {
2509 		large_pages_disable = disable_large_pages;
2510 	}
2511 
2512 	if (npgs < NHMENTS || large_pages_disable == LARGE_PAGES_OFF) {
2513 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2514 		    rid);
2515 		return;
2516 	}
2517 
2518 	while (npgs >= NHMENTS) {
2519 		pp = *pps;
2520 		for (ttesz = pp->p_szc; ttesz != TTE8K; ttesz--) {
2521 			/*
2522 			 * Check if this page size is disabled.
2523 			 */
2524 			if (large_pages_disable & (1 << ttesz))
2525 				continue;
2526 
2527 			numpg = TTEPAGES(ttesz);
2528 			mapsz = numpg << MMU_PAGESHIFT;
2529 			if ((npgs >= numpg) &&
2530 			    IS_P2ALIGNED(addr, mapsz) &&
2531 			    IS_P2ALIGNED(pp->p_pagenum, numpg)) {
2532 				/*
2533 				 * At this point we have enough pages and
2534 				 * we know the virtual address and the pfn
2535 				 * are properly aligned.  We still need
2536 				 * to check for physical contiguity but since
2537 				 * it is very likely that this is the case
2538 				 * we will assume they are so and undo
2539 				 * the request if necessary.  It would
2540 				 * be great if we could get a hint flag
2541 				 * like HAT_CONTIG which would tell us
2542 				 * the pages are contigous for sure.
2543 				 */
2544 				sfmmu_memtte(&tte, (*pps)->p_pagenum,
2545 				    attr, ttesz);
2546 				if (!sfmmu_tteload_array(hat, &tte, addr,
2547 				    pps, flags, rid)) {
2548 					break;
2549 				}
2550 			}
2551 		}
2552 		if (ttesz == TTE8K) {
2553 			/*
2554 			 * We were not able to map array using a large page
2555 			 * batch a hmeblk or fraction at a time.
2556 			 */
2557 			numpg = ((uintptr_t)addr >> MMU_PAGESHIFT)
2558 			    & (NHMENTS-1);
2559 			numpg = NHMENTS - numpg;
2560 			ASSERT(numpg <= npgs);
2561 			mapsz = numpg * MMU_PAGESIZE;
2562 			sfmmu_memload_batchsmall(hat, addr, pps, attr, flags,
2563 			    numpg, rid);
2564 		}
2565 		addr += mapsz;
2566 		npgs -= numpg;
2567 		pps += numpg;
2568 	}
2569 
2570 	if (npgs) {
2571 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2572 		    rid);
2573 	}
2574 
2575 	/*
2576 	 * Check TSB and TLB page sizes.
2577 	 */
2578 	if ((flags & HAT_LOAD_SHARE) == 0) {
2579 		sfmmu_check_page_sizes(hat, 1);
2580 	}
2581 }
2582 
2583 /*
2584  * Function tries to batch 8K pages into the same hme blk.
2585  */
2586 static void
2587 sfmmu_memload_batchsmall(struct hat *hat, caddr_t vaddr, page_t **pps,
2588 		    uint_t attr, uint_t flags, pgcnt_t npgs, uint_t rid)
2589 {
2590 	tte_t	tte;
2591 	page_t *pp;
2592 	struct hmehash_bucket *hmebp;
2593 	struct hme_blk *hmeblkp;
2594 	int	index;
2595 
2596 	while (npgs) {
2597 		/*
2598 		 * Acquire the hash bucket.
2599 		 */
2600 		hmebp = sfmmu_tteload_acquire_hashbucket(hat, vaddr, TTE8K,
2601 		    rid);
2602 		ASSERT(hmebp);
2603 
2604 		/*
2605 		 * Find the hment block.
2606 		 */
2607 		hmeblkp = sfmmu_tteload_find_hmeblk(hat, hmebp, vaddr,
2608 		    TTE8K, flags, rid);
2609 		ASSERT(hmeblkp);
2610 
2611 		do {
2612 			/*
2613 			 * Make the tte.
2614 			 */
2615 			pp = *pps;
2616 			sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2617 
2618 			/*
2619 			 * Add the translation.
2620 			 */
2621 			(void) sfmmu_tteload_addentry(hat, hmeblkp, &tte,
2622 			    vaddr, pps, flags, rid);
2623 
2624 			/*
2625 			 * Goto next page.
2626 			 */
2627 			pps++;
2628 			npgs--;
2629 
2630 			/*
2631 			 * Goto next address.
2632 			 */
2633 			vaddr += MMU_PAGESIZE;
2634 
2635 			/*
2636 			 * Don't crossover into a different hmentblk.
2637 			 */
2638 			index = (int)(((uintptr_t)vaddr >> MMU_PAGESHIFT) &
2639 			    (NHMENTS-1));
2640 
2641 		} while (index != 0 && npgs != 0);
2642 
2643 		/*
2644 		 * Release the hash bucket.
2645 		 */
2646 
2647 		sfmmu_tteload_release_hashbucket(hmebp);
2648 	}
2649 }
2650 
2651 /*
2652  * Construct a tte for a page:
2653  *
2654  * tte_valid = 1
2655  * tte_size2 = size & TTE_SZ2_BITS (Panther and Olympus-C only)
2656  * tte_size = size
2657  * tte_nfo = attr & HAT_NOFAULT
2658  * tte_ie = attr & HAT_STRUCTURE_LE
2659  * tte_hmenum = hmenum
2660  * tte_pahi = pp->p_pagenum >> TTE_PASHIFT;
2661  * tte_palo = pp->p_pagenum & TTE_PALOMASK;
2662  * tte_ref = 1 (optimization)
2663  * tte_wr_perm = attr & PROT_WRITE;
2664  * tte_no_sync = attr & HAT_NOSYNC
2665  * tte_lock = attr & SFMMU_LOCKTTE
2666  * tte_cp = !(attr & SFMMU_UNCACHEPTTE)
2667  * tte_cv = !(attr & SFMMU_UNCACHEVTTE)
2668  * tte_e = attr & SFMMU_SIDEFFECT
2669  * tte_priv = !(attr & PROT_USER)
2670  * tte_hwwr = if nosync is set and it is writable we set the mod bit (opt)
2671  * tte_glb = 0
2672  */
2673 void
2674 sfmmu_memtte(tte_t *ttep, pfn_t pfn, uint_t attr, int tte_sz)
2675 {
2676 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2677 
2678 	ttep->tte_inthi = MAKE_TTE_INTHI(pfn, attr, tte_sz, 0 /* hmenum */);
2679 	ttep->tte_intlo = MAKE_TTE_INTLO(pfn, attr, tte_sz, 0 /* hmenum */);
2680 
2681 	if (TTE_IS_NOSYNC(ttep)) {
2682 		TTE_SET_REF(ttep);
2683 		if (TTE_IS_WRITABLE(ttep)) {
2684 			TTE_SET_MOD(ttep);
2685 		}
2686 	}
2687 	if (TTE_IS_NFO(ttep) && TTE_IS_EXECUTABLE(ttep)) {
2688 		panic("sfmmu_memtte: can't set both NFO and EXEC bits");
2689 	}
2690 }
2691 
2692 /*
2693  * This function will add a translation to the hme_blk and allocate the
2694  * hme_blk if one does not exist.
2695  * If a page structure is specified then it will add the
2696  * corresponding hment to the mapping list.
2697  * It will also update the hmenum field for the tte.
2698  *
2699  * Currently this function is only used for kernel mappings.
2700  * So pass invalid region to sfmmu_tteload_array().
2701  */
2702 void
2703 sfmmu_tteload(struct hat *sfmmup, tte_t *ttep, caddr_t vaddr, page_t *pp,
2704 	uint_t flags)
2705 {
2706 	ASSERT(sfmmup == ksfmmup);
2707 	(void) sfmmu_tteload_array(sfmmup, ttep, vaddr, &pp, flags,
2708 	    SFMMU_INVALID_SHMERID);
2709 }
2710 
2711 /*
2712  * Load (ttep != NULL) or unload (ttep == NULL) one entry in the TSB.
2713  * Assumes that a particular page size may only be resident in one TSB.
2714  */
2715 static void
2716 sfmmu_mod_tsb(sfmmu_t *sfmmup, caddr_t vaddr, tte_t *ttep, int ttesz)
2717 {
2718 	struct tsb_info *tsbinfop = NULL;
2719 	uint64_t tag;
2720 	struct tsbe *tsbe_addr;
2721 	uint64_t tsb_base;
2722 	uint_t tsb_size;
2723 	int vpshift = MMU_PAGESHIFT;
2724 	int phys = 0;
2725 
2726 	if (sfmmup == ksfmmup) { /* No support for 32/256M ksfmmu pages */
2727 		phys = ktsb_phys;
2728 		if (ttesz >= TTE4M) {
2729 #ifndef sun4v
2730 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2731 #endif
2732 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2733 			tsb_size = ktsb4m_szcode;
2734 		} else {
2735 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2736 			tsb_size = ktsb_szcode;
2737 		}
2738 	} else {
2739 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2740 
2741 		/*
2742 		 * If there isn't a TSB for this page size, or the TSB is
2743 		 * swapped out, there is nothing to do.  Note that the latter
2744 		 * case seems impossible but can occur if hat_pageunload()
2745 		 * is called on an ISM mapping while the process is swapped
2746 		 * out.
2747 		 */
2748 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2749 			return;
2750 
2751 		/*
2752 		 * If another thread is in the middle of relocating a TSB
2753 		 * we can't unload the entry so set a flag so that the
2754 		 * TSB will be flushed before it can be accessed by the
2755 		 * process.
2756 		 */
2757 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2758 			if (ttep == NULL)
2759 				tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2760 			return;
2761 		}
2762 #if defined(UTSB_PHYS)
2763 		phys = 1;
2764 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2765 #else
2766 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2767 #endif
2768 		tsb_size = tsbinfop->tsb_szc;
2769 	}
2770 	if (ttesz >= TTE4M)
2771 		vpshift = MMU_PAGESHIFT4M;
2772 
2773 	tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2774 	tag = sfmmu_make_tsbtag(vaddr);
2775 
2776 	if (ttep == NULL) {
2777 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2778 	} else {
2779 		if (ttesz >= TTE4M) {
2780 			SFMMU_STAT(sf_tsb_load4m);
2781 		} else {
2782 			SFMMU_STAT(sf_tsb_load8k);
2783 		}
2784 
2785 		sfmmu_load_tsbe(tsbe_addr, tag, ttep, phys);
2786 	}
2787 }
2788 
2789 /*
2790  * Unmap all entries from [start, end) matching the given page size.
2791  *
2792  * This function is used primarily to unmap replicated 64K or 512K entries
2793  * from the TSB that are inserted using the base page size TSB pointer, but
2794  * it may also be called to unmap a range of addresses from the TSB.
2795  */
2796 void
2797 sfmmu_unload_tsb_range(sfmmu_t *sfmmup, caddr_t start, caddr_t end, int ttesz)
2798 {
2799 	struct tsb_info *tsbinfop;
2800 	uint64_t tag;
2801 	struct tsbe *tsbe_addr;
2802 	caddr_t vaddr;
2803 	uint64_t tsb_base;
2804 	int vpshift, vpgsz;
2805 	uint_t tsb_size;
2806 	int phys = 0;
2807 
2808 	/*
2809 	 * Assumptions:
2810 	 *  If ttesz == 8K, 64K or 512K, we walk through the range 8K
2811 	 *  at a time shooting down any valid entries we encounter.
2812 	 *
2813 	 *  If ttesz >= 4M we walk the range 4M at a time shooting
2814 	 *  down any valid mappings we find.
2815 	 */
2816 	if (sfmmup == ksfmmup) {
2817 		phys = ktsb_phys;
2818 		if (ttesz >= TTE4M) {
2819 #ifndef sun4v
2820 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2821 #endif
2822 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2823 			tsb_size = ktsb4m_szcode;
2824 		} else {
2825 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2826 			tsb_size = ktsb_szcode;
2827 		}
2828 	} else {
2829 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2830 
2831 		/*
2832 		 * If there isn't a TSB for this page size, or the TSB is
2833 		 * swapped out, there is nothing to do.  Note that the latter
2834 		 * case seems impossible but can occur if hat_pageunload()
2835 		 * is called on an ISM mapping while the process is swapped
2836 		 * out.
2837 		 */
2838 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2839 			return;
2840 
2841 		/*
2842 		 * If another thread is in the middle of relocating a TSB
2843 		 * we can't unload the entry so set a flag so that the
2844 		 * TSB will be flushed before it can be accessed by the
2845 		 * process.
2846 		 */
2847 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2848 			tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2849 			return;
2850 		}
2851 #if defined(UTSB_PHYS)
2852 		phys = 1;
2853 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2854 #else
2855 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2856 #endif
2857 		tsb_size = tsbinfop->tsb_szc;
2858 	}
2859 	if (ttesz >= TTE4M) {
2860 		vpshift = MMU_PAGESHIFT4M;
2861 		vpgsz = MMU_PAGESIZE4M;
2862 	} else {
2863 		vpshift = MMU_PAGESHIFT;
2864 		vpgsz = MMU_PAGESIZE;
2865 	}
2866 
2867 	for (vaddr = start; vaddr < end; vaddr += vpgsz) {
2868 		tag = sfmmu_make_tsbtag(vaddr);
2869 		tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2870 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2871 	}
2872 }
2873 
2874 /*
2875  * Select the optimum TSB size given the number of mappings
2876  * that need to be cached.
2877  */
2878 static int
2879 sfmmu_select_tsb_szc(pgcnt_t pgcnt)
2880 {
2881 	int szc = 0;
2882 
2883 #ifdef DEBUG
2884 	if (tsb_grow_stress) {
2885 		uint32_t randval = (uint32_t)gettick() >> 4;
2886 		return (randval % (tsb_max_growsize + 1));
2887 	}
2888 #endif	/* DEBUG */
2889 
2890 	while ((szc < tsb_max_growsize) && (pgcnt > SFMMU_RSS_TSBSIZE(szc)))
2891 		szc++;
2892 	return (szc);
2893 }
2894 
2895 /*
2896  * This function will add a translation to the hme_blk and allocate the
2897  * hme_blk if one does not exist.
2898  * If a page structure is specified then it will add the
2899  * corresponding hment to the mapping list.
2900  * It will also update the hmenum field for the tte.
2901  * Furthermore, it attempts to create a large page translation
2902  * for <addr,hat> at page array pps.  It assumes addr and first
2903  * pp is correctly aligned.  It returns 0 if successful and 1 otherwise.
2904  */
2905 static int
2906 sfmmu_tteload_array(sfmmu_t *sfmmup, tte_t *ttep, caddr_t vaddr,
2907 	page_t **pps, uint_t flags, uint_t rid)
2908 {
2909 	struct hmehash_bucket *hmebp;
2910 	struct hme_blk *hmeblkp;
2911 	int 	ret;
2912 	uint_t	size;
2913 
2914 	/*
2915 	 * Get mapping size.
2916 	 */
2917 	size = TTE_CSZ(ttep);
2918 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
2919 
2920 	/*
2921 	 * Acquire the hash bucket.
2922 	 */
2923 	hmebp = sfmmu_tteload_acquire_hashbucket(sfmmup, vaddr, size, rid);
2924 	ASSERT(hmebp);
2925 
2926 	/*
2927 	 * Find the hment block.
2928 	 */
2929 	hmeblkp = sfmmu_tteload_find_hmeblk(sfmmup, hmebp, vaddr, size, flags,
2930 	    rid);
2931 	ASSERT(hmeblkp);
2932 
2933 	/*
2934 	 * Add the translation.
2935 	 */
2936 	ret = sfmmu_tteload_addentry(sfmmup, hmeblkp, ttep, vaddr, pps, flags,
2937 	    rid);
2938 
2939 	/*
2940 	 * Release the hash bucket.
2941 	 */
2942 	sfmmu_tteload_release_hashbucket(hmebp);
2943 
2944 	return (ret);
2945 }
2946 
2947 /*
2948  * Function locks and returns a pointer to the hash bucket for vaddr and size.
2949  */
2950 static struct hmehash_bucket *
2951 sfmmu_tteload_acquire_hashbucket(sfmmu_t *sfmmup, caddr_t vaddr, int size,
2952     uint_t rid)
2953 {
2954 	struct hmehash_bucket *hmebp;
2955 	int hmeshift;
2956 	void *htagid = sfmmutohtagid(sfmmup, rid);
2957 
2958 	ASSERT(htagid != NULL);
2959 
2960 	hmeshift = HME_HASH_SHIFT(size);
2961 
2962 	hmebp = HME_HASH_FUNCTION(htagid, vaddr, hmeshift);
2963 
2964 	SFMMU_HASH_LOCK(hmebp);
2965 
2966 	return (hmebp);
2967 }
2968 
2969 /*
2970  * Function returns a pointer to an hmeblk in the hash bucket, hmebp. If the
2971  * hmeblk doesn't exists for the [sfmmup, vaddr & size] signature, a hmeblk is
2972  * allocated.
2973  */
2974 static struct hme_blk *
2975 sfmmu_tteload_find_hmeblk(sfmmu_t *sfmmup, struct hmehash_bucket *hmebp,
2976 	caddr_t vaddr, uint_t size, uint_t flags, uint_t rid)
2977 {
2978 	hmeblk_tag hblktag;
2979 	int hmeshift;
2980 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
2981 
2982 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
2983 
2984 	hblktag.htag_id = sfmmutohtagid(sfmmup, rid);
2985 	ASSERT(hblktag.htag_id != NULL);
2986 	hmeshift = HME_HASH_SHIFT(size);
2987 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
2988 	hblktag.htag_rehash = HME_HASH_REHASH(size);
2989 	hblktag.htag_rid = rid;
2990 
2991 ttearray_realloc:
2992 
2993 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
2994 
2995 	/*
2996 	 * We block until hblk_reserve_lock is released; it's held by
2997 	 * the thread, temporarily using hblk_reserve, until hblk_reserve is
2998 	 * replaced by a hblk from sfmmu8_cache.
2999 	 */
3000 	if (hmeblkp == (struct hme_blk *)hblk_reserve &&
3001 	    hblk_reserve_thread != curthread) {
3002 		SFMMU_HASH_UNLOCK(hmebp);
3003 		mutex_enter(&hblk_reserve_lock);
3004 		mutex_exit(&hblk_reserve_lock);
3005 		SFMMU_STAT(sf_hblk_reserve_hit);
3006 		SFMMU_HASH_LOCK(hmebp);
3007 		goto ttearray_realloc;
3008 	}
3009 
3010 	if (hmeblkp == NULL) {
3011 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
3012 		    hblktag, flags, rid);
3013 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3014 		ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3015 	} else {
3016 		/*
3017 		 * It is possible for 8k and 64k hblks to collide since they
3018 		 * have the same rehash value. This is because we
3019 		 * lazily free hblks and 8K/64K blks could be lingering.
3020 		 * If we find size mismatch we free the block and & try again.
3021 		 */
3022 		if (get_hblk_ttesz(hmeblkp) != size) {
3023 			ASSERT(!hmeblkp->hblk_vcnt);
3024 			ASSERT(!hmeblkp->hblk_hmecnt);
3025 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3026 			    &list, 0);
3027 			goto ttearray_realloc;
3028 		}
3029 		if (hmeblkp->hblk_shw_bit) {
3030 			/*
3031 			 * if the hblk was previously used as a shadow hblk then
3032 			 * we will change it to a normal hblk
3033 			 */
3034 			ASSERT(!hmeblkp->hblk_shared);
3035 			if (hmeblkp->hblk_shw_mask) {
3036 				sfmmu_shadow_hcleanup(sfmmup, hmeblkp, hmebp);
3037 				ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3038 				goto ttearray_realloc;
3039 			} else {
3040 				hmeblkp->hblk_shw_bit = 0;
3041 			}
3042 		}
3043 		SFMMU_STAT(sf_hblk_hit);
3044 	}
3045 
3046 	/*
3047 	 * hat_memload() should never call kmem_cache_free() for kernel hmeblks;
3048 	 * see block comment showing the stacktrace in sfmmu_hblk_alloc();
3049 	 * set the flag parameter to 1 so that sfmmu_hblks_list_purge() will
3050 	 * just add these hmeblks to the per-cpu pending queue.
3051 	 */
3052 	sfmmu_hblks_list_purge(&list, 1);
3053 
3054 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
3055 	ASSERT(!hmeblkp->hblk_shw_bit);
3056 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3057 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3058 	ASSERT(hmeblkp->hblk_tag.htag_rid == rid);
3059 
3060 	return (hmeblkp);
3061 }
3062 
3063 /*
3064  * Function adds a tte entry into the hmeblk. It returns 0 if successful and 1
3065  * otherwise.
3066  */
3067 static int
3068 sfmmu_tteload_addentry(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, tte_t *ttep,
3069 	caddr_t vaddr, page_t **pps, uint_t flags, uint_t rid)
3070 {
3071 	page_t *pp = *pps;
3072 	int hmenum, size, remap;
3073 	tte_t tteold, flush_tte;
3074 #ifdef DEBUG
3075 	tte_t orig_old;
3076 #endif /* DEBUG */
3077 	struct sf_hment *sfhme;
3078 	kmutex_t *pml, *pmtx;
3079 	hatlock_t *hatlockp;
3080 	int myflt;
3081 
3082 	/*
3083 	 * remove this panic when we decide to let user virtual address
3084 	 * space be >= USERLIMIT.
3085 	 */
3086 	if (!TTE_IS_PRIVILEGED(ttep) && vaddr >= (caddr_t)USERLIMIT)
3087 		panic("user addr %p in kernel space", (void *)vaddr);
3088 #if defined(TTE_IS_GLOBAL)
3089 	if (TTE_IS_GLOBAL(ttep))
3090 		panic("sfmmu_tteload: creating global tte");
3091 #endif
3092 
3093 #ifdef DEBUG
3094 	if (pf_is_memory(sfmmu_ttetopfn(ttep, vaddr)) &&
3095 	    !TTE_IS_PCACHEABLE(ttep) && !sfmmu_allow_nc_trans)
3096 		panic("sfmmu_tteload: non cacheable memory tte");
3097 #endif /* DEBUG */
3098 
3099 	/* don't simulate dirty bit for writeable ISM/DISM mappings */
3100 	if ((flags & HAT_LOAD_SHARE) && TTE_IS_WRITABLE(ttep)) {
3101 		TTE_SET_REF(ttep);
3102 		TTE_SET_MOD(ttep);
3103 	}
3104 
3105 	if ((flags & HAT_LOAD_SHARE) || !TTE_IS_REF(ttep) ||
3106 	    !TTE_IS_MOD(ttep)) {
3107 		/*
3108 		 * Don't load TSB for dummy as in ISM.  Also don't preload
3109 		 * the TSB if the TTE isn't writable since we're likely to
3110 		 * fault on it again -- preloading can be fairly expensive.
3111 		 */
3112 		flags |= SFMMU_NO_TSBLOAD;
3113 	}
3114 
3115 	size = TTE_CSZ(ttep);
3116 	switch (size) {
3117 	case TTE8K:
3118 		SFMMU_STAT(sf_tteload8k);
3119 		break;
3120 	case TTE64K:
3121 		SFMMU_STAT(sf_tteload64k);
3122 		break;
3123 	case TTE512K:
3124 		SFMMU_STAT(sf_tteload512k);
3125 		break;
3126 	case TTE4M:
3127 		SFMMU_STAT(sf_tteload4m);
3128 		break;
3129 	case (TTE32M):
3130 		SFMMU_STAT(sf_tteload32m);
3131 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3132 		break;
3133 	case (TTE256M):
3134 		SFMMU_STAT(sf_tteload256m);
3135 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3136 		break;
3137 	}
3138 
3139 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
3140 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
3141 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3142 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3143 
3144 	HBLKTOHME_IDX(sfhme, hmeblkp, vaddr, hmenum);
3145 
3146 	/*
3147 	 * Need to grab mlist lock here so that pageunload
3148 	 * will not change tte behind us.
3149 	 */
3150 	if (pp) {
3151 		pml = sfmmu_mlist_enter(pp);
3152 	}
3153 
3154 	sfmmu_copytte(&sfhme->hme_tte, &tteold);
3155 	/*
3156 	 * Look for corresponding hment and if valid verify
3157 	 * pfns are equal.
3158 	 */
3159 	remap = TTE_IS_VALID(&tteold);
3160 	if (remap) {
3161 		pfn_t	new_pfn, old_pfn;
3162 
3163 		old_pfn = TTE_TO_PFN(vaddr, &tteold);
3164 		new_pfn = TTE_TO_PFN(vaddr, ttep);
3165 
3166 		if (flags & HAT_LOAD_REMAP) {
3167 			/* make sure we are remapping same type of pages */
3168 			if (pf_is_memory(old_pfn) != pf_is_memory(new_pfn)) {
3169 				panic("sfmmu_tteload - tte remap io<->memory");
3170 			}
3171 			if (old_pfn != new_pfn &&
3172 			    (pp != NULL || sfhme->hme_page != NULL)) {
3173 				panic("sfmmu_tteload - tte remap pp != NULL");
3174 			}
3175 		} else if (old_pfn != new_pfn) {
3176 			panic("sfmmu_tteload - tte remap, hmeblkp 0x%p",
3177 			    (void *)hmeblkp);
3178 		}
3179 		ASSERT(TTE_CSZ(&tteold) == TTE_CSZ(ttep));
3180 	}
3181 
3182 	if (pp) {
3183 		if (size == TTE8K) {
3184 #ifdef VAC
3185 			/*
3186 			 * Handle VAC consistency
3187 			 */
3188 			if (!remap && (cache & CACHE_VAC) && !PP_ISNC(pp)) {
3189 				sfmmu_vac_conflict(sfmmup, vaddr, pp);
3190 			}
3191 #endif
3192 
3193 			if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3194 				pmtx = sfmmu_page_enter(pp);
3195 				PP_CLRRO(pp);
3196 				sfmmu_page_exit(pmtx);
3197 			} else if (!PP_ISMAPPED(pp) &&
3198 			    (!TTE_IS_WRITABLE(ttep)) && !(PP_ISMOD(pp))) {
3199 				pmtx = sfmmu_page_enter(pp);
3200 				if (!(PP_ISMOD(pp))) {
3201 					PP_SETRO(pp);
3202 				}
3203 				sfmmu_page_exit(pmtx);
3204 			}
3205 
3206 		} else if (sfmmu_pagearray_setup(vaddr, pps, ttep, remap)) {
3207 			/*
3208 			 * sfmmu_pagearray_setup failed so return
3209 			 */
3210 			sfmmu_mlist_exit(pml);
3211 			return (1);
3212 		}
3213 	}
3214 
3215 	/*
3216 	 * Make sure hment is not on a mapping list.
3217 	 */
3218 	ASSERT(remap || (sfhme->hme_page == NULL));
3219 
3220 	/* if it is not a remap then hme->next better be NULL */
3221 	ASSERT((!remap) ? sfhme->hme_next == NULL : 1);
3222 
3223 	if (flags & HAT_LOAD_LOCK) {
3224 		if ((hmeblkp->hblk_lckcnt + 1) >= MAX_HBLK_LCKCNT) {
3225 			panic("too high lckcnt-hmeblk %p",
3226 			    (void *)hmeblkp);
3227 		}
3228 		atomic_inc_32(&hmeblkp->hblk_lckcnt);
3229 
3230 		HBLK_STACK_TRACE(hmeblkp, HBLK_LOCK);
3231 	}
3232 
3233 #ifdef VAC
3234 	if (pp && PP_ISNC(pp)) {
3235 		/*
3236 		 * If the physical page is marked to be uncacheable, like
3237 		 * by a vac conflict, make sure the new mapping is also
3238 		 * uncacheable.
3239 		 */
3240 		TTE_CLR_VCACHEABLE(ttep);
3241 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
3242 	}
3243 #endif
3244 	ttep->tte_hmenum = hmenum;
3245 
3246 #ifdef DEBUG
3247 	orig_old = tteold;
3248 #endif /* DEBUG */
3249 
3250 	while (sfmmu_modifytte_try(&tteold, ttep, &sfhme->hme_tte) < 0) {
3251 		if ((sfmmup == KHATID) &&
3252 		    (flags & (HAT_LOAD_LOCK | HAT_LOAD_REMAP))) {
3253 			sfmmu_copytte(&sfhme->hme_tte, &tteold);
3254 		}
3255 #ifdef DEBUG
3256 		chk_tte(&orig_old, &tteold, ttep, hmeblkp);
3257 #endif /* DEBUG */
3258 	}
3259 	ASSERT(TTE_IS_VALID(&sfhme->hme_tte));
3260 
3261 	if (!TTE_IS_VALID(&tteold)) {
3262 
3263 		atomic_inc_16(&hmeblkp->hblk_vcnt);
3264 		if (rid == SFMMU_INVALID_SHMERID) {
3265 			atomic_inc_ulong(&sfmmup->sfmmu_ttecnt[size]);
3266 		} else {
3267 			sf_srd_t *srdp = sfmmup->sfmmu_srdp;
3268 			sf_region_t *rgnp = srdp->srd_hmergnp[rid];
3269 			/*
3270 			 * We already accounted for region ttecnt's in sfmmu
3271 			 * during hat_join_region() processing. Here we
3272 			 * only update ttecnt's in region struture.
3273 			 */
3274 			atomic_inc_ulong(&rgnp->rgn_ttecnt[size]);
3275 		}
3276 	}
3277 
3278 	myflt = (astosfmmu(curthread->t_procp->p_as) == sfmmup);
3279 	if (size > TTE8K && (flags & HAT_LOAD_SHARE) == 0 &&
3280 	    sfmmup != ksfmmup) {
3281 		uchar_t tteflag = 1 << size;
3282 		if (rid == SFMMU_INVALID_SHMERID) {
3283 			if (!(sfmmup->sfmmu_tteflags & tteflag)) {
3284 				hatlockp = sfmmu_hat_enter(sfmmup);
3285 				sfmmup->sfmmu_tteflags |= tteflag;
3286 				sfmmu_hat_exit(hatlockp);
3287 			}
3288 		} else if (!(sfmmup->sfmmu_rtteflags & tteflag)) {
3289 			hatlockp = sfmmu_hat_enter(sfmmup);
3290 			sfmmup->sfmmu_rtteflags |= tteflag;
3291 			sfmmu_hat_exit(hatlockp);
3292 		}
3293 		/*
3294 		 * Update the current CPU tsbmiss area, so the current thread
3295 		 * won't need to take the tsbmiss for the new pagesize.
3296 		 * The other threads in the process will update their tsb
3297 		 * miss area lazily in sfmmu_tsbmiss_exception() when they
3298 		 * fail to find the translation for a newly added pagesize.
3299 		 */
3300 		if (size > TTE64K && myflt) {
3301 			struct tsbmiss *tsbmp;
3302 			kpreempt_disable();
3303 			tsbmp = &tsbmiss_area[CPU->cpu_id];
3304 			if (rid == SFMMU_INVALID_SHMERID) {
3305 				if (!(tsbmp->uhat_tteflags & tteflag)) {
3306 					tsbmp->uhat_tteflags |= tteflag;
3307 				}
3308 			} else {
3309 				if (!(tsbmp->uhat_rtteflags & tteflag)) {
3310 					tsbmp->uhat_rtteflags |= tteflag;
3311 				}
3312 			}
3313 			kpreempt_enable();
3314 		}
3315 	}
3316 
3317 	if (size >= TTE4M && (flags & HAT_LOAD_TEXT) &&
3318 	    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
3319 		hatlockp = sfmmu_hat_enter(sfmmup);
3320 		SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
3321 		sfmmu_hat_exit(hatlockp);
3322 	}
3323 
3324 	flush_tte.tte_intlo = (tteold.tte_intlo ^ ttep->tte_intlo) &
3325 	    hw_tte.tte_intlo;
3326 	flush_tte.tte_inthi = (tteold.tte_inthi ^ ttep->tte_inthi) &
3327 	    hw_tte.tte_inthi;
3328 
3329 	if (remap && (flush_tte.tte_inthi || flush_tte.tte_intlo)) {
3330 		/*
3331 		 * If remap and new tte differs from old tte we need
3332 		 * to sync the mod bit and flush TLB/TSB.  We don't
3333 		 * need to sync ref bit because we currently always set
3334 		 * ref bit in tteload.
3335 		 */
3336 		ASSERT(TTE_IS_REF(ttep));
3337 		if (TTE_IS_MOD(&tteold)) {
3338 			sfmmu_ttesync(sfmmup, vaddr, &tteold, pp);
3339 		}
3340 		/*
3341 		 * hwtte bits shouldn't change for SRD hmeblks as long as SRD
3342 		 * hmes are only used for read only text. Adding this code for
3343 		 * completeness and future use of shared hmeblks with writable
3344 		 * mappings of VMODSORT vnodes.
3345 		 */
3346 		if (hmeblkp->hblk_shared) {
3347 			cpuset_t cpuset = sfmmu_rgntlb_demap(vaddr,
3348 			    sfmmup->sfmmu_srdp->srd_hmergnp[rid], hmeblkp, 1);
3349 			xt_sync(cpuset);
3350 			SFMMU_STAT_ADD(sf_region_remap_demap, 1);
3351 		} else {
3352 			sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 0);
3353 			xt_sync(sfmmup->sfmmu_cpusran);
3354 		}
3355 	}
3356 
3357 	if ((flags & SFMMU_NO_TSBLOAD) == 0) {
3358 		/*
3359 		 * We only preload 8K and 4M mappings into the TSB, since
3360 		 * 64K and 512K mappings are replicated and hence don't
3361 		 * have a single, unique TSB entry. Ditto for 32M/256M.
3362 		 */
3363 		if (size == TTE8K || size == TTE4M) {
3364 			sf_scd_t *scdp;
3365 			hatlockp = sfmmu_hat_enter(sfmmup);
3366 			/*
3367 			 * Don't preload private TSB if the mapping is used
3368 			 * by the shctx in the SCD.
3369 			 */
3370 			scdp = sfmmup->sfmmu_scdp;
3371 			if (rid == SFMMU_INVALID_SHMERID || scdp == NULL ||
3372 			    !SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
3373 				sfmmu_load_tsb(sfmmup, vaddr, &sfhme->hme_tte,
3374 				    size);
3375 			}
3376 			sfmmu_hat_exit(hatlockp);
3377 		}
3378 	}
3379 	if (pp) {
3380 		if (!remap) {
3381 			HME_ADD(sfhme, pp);
3382 			atomic_inc_16(&hmeblkp->hblk_hmecnt);
3383 			ASSERT(hmeblkp->hblk_hmecnt > 0);
3384 
3385 			/*
3386 			 * Cannot ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
3387 			 * see pageunload() for comment.
3388 			 */
3389 		}
3390 		sfmmu_mlist_exit(pml);
3391 	}
3392 
3393 	return (0);
3394 }
3395 /*
3396  * Function unlocks hash bucket.
3397  */
3398 static void
3399 sfmmu_tteload_release_hashbucket(struct hmehash_bucket *hmebp)
3400 {
3401 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3402 	SFMMU_HASH_UNLOCK(hmebp);
3403 }
3404 
3405 /*
3406  * function which checks and sets up page array for a large
3407  * translation.  Will set p_vcolor, p_index, p_ro fields.
3408  * Assumes addr and pfnum of first page are properly aligned.
3409  * Will check for physical contiguity. If check fails it return
3410  * non null.
3411  */
3412 static int
3413 sfmmu_pagearray_setup(caddr_t addr, page_t **pps, tte_t *ttep, int remap)
3414 {
3415 	int 	i, index, ttesz;
3416 	pfn_t	pfnum;
3417 	pgcnt_t	npgs;
3418 	page_t *pp, *pp1;
3419 	kmutex_t *pmtx;
3420 #ifdef VAC
3421 	int osz;
3422 	int cflags = 0;
3423 	int vac_err = 0;
3424 #endif
3425 	int newidx = 0;
3426 
3427 	ttesz = TTE_CSZ(ttep);
3428 
3429 	ASSERT(ttesz > TTE8K);
3430 
3431 	npgs = TTEPAGES(ttesz);
3432 	index = PAGESZ_TO_INDEX(ttesz);
3433 
3434 	pfnum = (*pps)->p_pagenum;
3435 	ASSERT(IS_P2ALIGNED(pfnum, npgs));
3436 
3437 	/*
3438 	 * Save the first pp so we can do HAT_TMPNC at the end.
3439 	 */
3440 	pp1 = *pps;
3441 #ifdef VAC
3442 	osz = fnd_mapping_sz(pp1);
3443 #endif
3444 
3445 	for (i = 0; i < npgs; i++, pps++) {
3446 		pp = *pps;
3447 		ASSERT(PAGE_LOCKED(pp));
3448 		ASSERT(pp->p_szc >= ttesz);
3449 		ASSERT(pp->p_szc == pp1->p_szc);
3450 		ASSERT(sfmmu_mlist_held(pp));
3451 
3452 		/*
3453 		 * XXX is it possible to maintain P_RO on the root only?
3454 		 */
3455 		if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3456 			pmtx = sfmmu_page_enter(pp);
3457 			PP_CLRRO(pp);
3458 			sfmmu_page_exit(pmtx);
3459 		} else if (!PP_ISMAPPED(pp) && !TTE_IS_WRITABLE(ttep) &&
3460 		    !PP_ISMOD(pp)) {
3461 			pmtx = sfmmu_page_enter(pp);
3462 			if (!(PP_ISMOD(pp))) {
3463 				PP_SETRO(pp);
3464 			}
3465 			sfmmu_page_exit(pmtx);
3466 		}
3467 
3468 		/*
3469 		 * If this is a remap we skip vac & contiguity checks.
3470 		 */
3471 		if (remap)
3472 			continue;
3473 
3474 		/*
3475 		 * set p_vcolor and detect any vac conflicts.
3476 		 */
3477 #ifdef VAC
3478 		if (vac_err == 0) {
3479 			vac_err = sfmmu_vacconflict_array(addr, pp, &cflags);
3480 
3481 		}
3482 #endif
3483 
3484 		/*
3485 		 * Save current index in case we need to undo it.
3486 		 * Note: "PAGESZ_TO_INDEX(sz)	(1 << (sz))"
3487 		 *	"SFMMU_INDEX_SHIFT	6"
3488 		 *	 "SFMMU_INDEX_MASK	((1 << SFMMU_INDEX_SHIFT) - 1)"
3489 		 *	 "PP_MAPINDEX(p_index)	(p_index & SFMMU_INDEX_MASK)"
3490 		 *
3491 		 * So:	index = PAGESZ_TO_INDEX(ttesz);
3492 		 *	if ttesz == 1 then index = 0x2
3493 		 *		    2 then index = 0x4
3494 		 *		    3 then index = 0x8
3495 		 *		    4 then index = 0x10
3496 		 *		    5 then index = 0x20
3497 		 * The code below checks if it's a new pagesize (ie, newidx)
3498 		 * in case we need to take it back out of p_index,
3499 		 * and then or's the new index into the existing index.
3500 		 */
3501 		if ((PP_MAPINDEX(pp) & index) == 0)
3502 			newidx = 1;
3503 		pp->p_index = (PP_MAPINDEX(pp) | index);
3504 
3505 		/*
3506 		 * contiguity check
3507 		 */
3508 		if (pp->p_pagenum != pfnum) {
3509 			/*
3510 			 * If we fail the contiguity test then
3511 			 * the only thing we need to fix is the p_index field.
3512 			 * We might get a few extra flushes but since this
3513 			 * path is rare that is ok.  The p_ro field will
3514 			 * get automatically fixed on the next tteload to
3515 			 * the page.  NO TNC bit is set yet.
3516 			 */
3517 			while (i >= 0) {
3518 				pp = *pps;
3519 				if (newidx)
3520 					pp->p_index = (PP_MAPINDEX(pp) &
3521 					    ~index);
3522 				pps--;
3523 				i--;
3524 			}
3525 			return (1);
3526 		}
3527 		pfnum++;
3528 		addr += MMU_PAGESIZE;
3529 	}
3530 
3531 #ifdef VAC
3532 	if (vac_err) {
3533 		if (ttesz > osz) {
3534 			/*
3535 			 * There are some smaller mappings that causes vac
3536 			 * conflicts. Convert all existing small mappings to
3537 			 * TNC.
3538 			 */
3539 			SFMMU_STAT_ADD(sf_uncache_conflict, npgs);
3540 			sfmmu_page_cache_array(pp1, HAT_TMPNC, CACHE_FLUSH,
3541 			    npgs);
3542 		} else {
3543 			/* EMPTY */
3544 			/*
3545 			 * If there exists an big page mapping,
3546 			 * that means the whole existing big page
3547 			 * has TNC setting already. No need to covert to
3548 			 * TNC again.
3549 			 */
3550 			ASSERT(PP_ISTNC(pp1));
3551 		}
3552 	}
3553 #endif	/* VAC */
3554 
3555 	return (0);
3556 }
3557 
3558 #ifdef VAC
3559 /*
3560  * Routine that detects vac consistency for a large page. It also
3561  * sets virtual color for all pp's for this big mapping.
3562  */
3563 static int
3564 sfmmu_vacconflict_array(caddr_t addr, page_t *pp, int *cflags)
3565 {
3566 	int vcolor, ocolor;
3567 
3568 	ASSERT(sfmmu_mlist_held(pp));
3569 
3570 	if (PP_ISNC(pp)) {
3571 		return (HAT_TMPNC);
3572 	}
3573 
3574 	vcolor = addr_to_vcolor(addr);
3575 	if (PP_NEWPAGE(pp)) {
3576 		PP_SET_VCOLOR(pp, vcolor);
3577 		return (0);
3578 	}
3579 
3580 	ocolor = PP_GET_VCOLOR(pp);
3581 	if (ocolor == vcolor) {
3582 		return (0);
3583 	}
3584 
3585 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
3586 		/*
3587 		 * Previous user of page had a differnet color
3588 		 * but since there are no current users
3589 		 * we just flush the cache and change the color.
3590 		 * As an optimization for large pages we flush the
3591 		 * entire cache of that color and set a flag.
3592 		 */
3593 		SFMMU_STAT(sf_pgcolor_conflict);
3594 		if (!CacheColor_IsFlushed(*cflags, ocolor)) {
3595 			CacheColor_SetFlushed(*cflags, ocolor);
3596 			sfmmu_cache_flushcolor(ocolor, pp->p_pagenum);
3597 		}
3598 		PP_SET_VCOLOR(pp, vcolor);
3599 		return (0);
3600 	}
3601 
3602 	/*
3603 	 * We got a real conflict with a current mapping.
3604 	 * set flags to start unencaching all mappings
3605 	 * and return failure so we restart looping
3606 	 * the pp array from the beginning.
3607 	 */
3608 	return (HAT_TMPNC);
3609 }
3610 #endif	/* VAC */
3611 
3612 /*
3613  * creates a large page shadow hmeblk for a tte.
3614  * The purpose of this routine is to allow us to do quick unloads because
3615  * the vm layer can easily pass a very large but sparsely populated range.
3616  */
3617 static struct hme_blk *
3618 sfmmu_shadow_hcreate(sfmmu_t *sfmmup, caddr_t vaddr, int ttesz, uint_t flags)
3619 {
3620 	struct hmehash_bucket *hmebp;
3621 	hmeblk_tag hblktag;
3622 	int hmeshift, size, vshift;
3623 	uint_t shw_mask, newshw_mask;
3624 	struct hme_blk *hmeblkp;
3625 
3626 	ASSERT(sfmmup != KHATID);
3627 	if (mmu_page_sizes == max_mmu_page_sizes) {
3628 		ASSERT(ttesz < TTE256M);
3629 	} else {
3630 		ASSERT(ttesz < TTE4M);
3631 		ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
3632 		ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
3633 	}
3634 
3635 	if (ttesz == TTE8K) {
3636 		size = TTE512K;
3637 	} else {
3638 		size = ++ttesz;
3639 	}
3640 
3641 	hblktag.htag_id = sfmmup;
3642 	hmeshift = HME_HASH_SHIFT(size);
3643 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
3644 	hblktag.htag_rehash = HME_HASH_REHASH(size);
3645 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3646 	hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
3647 
3648 	SFMMU_HASH_LOCK(hmebp);
3649 
3650 	HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
3651 	ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
3652 	if (hmeblkp == NULL) {
3653 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
3654 		    hblktag, flags, SFMMU_INVALID_SHMERID);
3655 	}
3656 	ASSERT(hmeblkp);
3657 	if (!hmeblkp->hblk_shw_mask) {
3658 		/*
3659 		 * if this is a unused hblk it was just allocated or could
3660 		 * potentially be a previous large page hblk so we need to
3661 		 * set the shadow bit.
3662 		 */
3663 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3664 		hmeblkp->hblk_shw_bit = 1;
3665 	} else if (hmeblkp->hblk_shw_bit == 0) {
3666 		panic("sfmmu_shadow_hcreate: shw bit not set in hmeblkp 0x%p",
3667 		    (void *)hmeblkp);
3668 	}
3669 	ASSERT(hmeblkp->hblk_shw_bit == 1);
3670 	ASSERT(!hmeblkp->hblk_shared);
3671 	vshift = vaddr_to_vshift(hblktag, vaddr, size);
3672 	ASSERT(vshift < 8);
3673 	/*
3674 	 * Atomically set shw mask bit
3675 	 */
3676 	do {
3677 		shw_mask = hmeblkp->hblk_shw_mask;
3678 		newshw_mask = shw_mask | (1 << vshift);
3679 		newshw_mask = atomic_cas_32(&hmeblkp->hblk_shw_mask, shw_mask,
3680 		    newshw_mask);
3681 	} while (newshw_mask != shw_mask);
3682 
3683 	SFMMU_HASH_UNLOCK(hmebp);
3684 
3685 	return (hmeblkp);
3686 }
3687 
3688 /*
3689  * This routine cleanup a previous shadow hmeblk and changes it to
3690  * a regular hblk.  This happens rarely but it is possible
3691  * when a process wants to use large pages and there are hblks still
3692  * lying around from the previous as that used these hmeblks.
3693  * The alternative was to cleanup the shadow hblks at unload time
3694  * but since so few user processes actually use large pages, it is
3695  * better to be lazy and cleanup at this time.
3696  */
3697 static void
3698 sfmmu_shadow_hcleanup(sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
3699 	struct hmehash_bucket *hmebp)
3700 {
3701 	caddr_t addr, endaddr;
3702 	int hashno, size;
3703 
3704 	ASSERT(hmeblkp->hblk_shw_bit);
3705 	ASSERT(!hmeblkp->hblk_shared);
3706 
3707 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3708 
3709 	if (!hmeblkp->hblk_shw_mask) {
3710 		hmeblkp->hblk_shw_bit = 0;
3711 		return;
3712 	}
3713 	addr = (caddr_t)get_hblk_base(hmeblkp);
3714 	endaddr = get_hblk_endaddr(hmeblkp);
3715 	size = get_hblk_ttesz(hmeblkp);
3716 	hashno = size - 1;
3717 	ASSERT(hashno > 0);
3718 	SFMMU_HASH_UNLOCK(hmebp);
3719 
3720 	sfmmu_free_hblks(sfmmup, addr, endaddr, hashno);
3721 
3722 	SFMMU_HASH_LOCK(hmebp);
3723 }
3724 
3725 static void
3726 sfmmu_free_hblks(sfmmu_t *sfmmup, caddr_t addr, caddr_t endaddr,
3727 	int hashno)
3728 {
3729 	int hmeshift, shadow = 0;
3730 	hmeblk_tag hblktag;
3731 	struct hmehash_bucket *hmebp;
3732 	struct hme_blk *hmeblkp;
3733 	struct hme_blk *nx_hblk, *pr_hblk, *list = NULL;
3734 
3735 	ASSERT(hashno > 0);
3736 	hblktag.htag_id = sfmmup;
3737 	hblktag.htag_rehash = hashno;
3738 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3739 
3740 	hmeshift = HME_HASH_SHIFT(hashno);
3741 
3742 	while (addr < endaddr) {
3743 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3744 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3745 		SFMMU_HASH_LOCK(hmebp);
3746 		/* inline HME_HASH_SEARCH */
3747 		hmeblkp = hmebp->hmeblkp;
3748 		pr_hblk = NULL;
3749 		while (hmeblkp) {
3750 			if (HTAGS_EQ(hmeblkp->hblk_tag, hblktag)) {
3751 				/* found hme_blk */
3752 				ASSERT(!hmeblkp->hblk_shared);
3753 				if (hmeblkp->hblk_shw_bit) {
3754 					if (hmeblkp->hblk_shw_mask) {
3755 						shadow = 1;
3756 						sfmmu_shadow_hcleanup(sfmmup,
3757 						    hmeblkp, hmebp);
3758 						break;
3759 					} else {
3760 						hmeblkp->hblk_shw_bit = 0;
3761 					}
3762 				}
3763 
3764 				/*
3765 				 * Hblk_hmecnt and hblk_vcnt could be non zero
3766 				 * since hblk_unload() does not gurantee that.
3767 				 *
3768 				 * XXX - this could cause tteload() to spin
3769 				 * where sfmmu_shadow_hcleanup() is called.
3770 				 */
3771 			}
3772 
3773 			nx_hblk = hmeblkp->hblk_next;
3774 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
3775 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3776 				    &list, 0);
3777 			} else {
3778 				pr_hblk = hmeblkp;
3779 			}
3780 			hmeblkp = nx_hblk;
3781 		}
3782 
3783 		SFMMU_HASH_UNLOCK(hmebp);
3784 
3785 		if (shadow) {
3786 			/*
3787 			 * We found another shadow hblk so cleaned its
3788 			 * children.  We need to go back and cleanup
3789 			 * the original hblk so we don't change the
3790 			 * addr.
3791 			 */
3792 			shadow = 0;
3793 		} else {
3794 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
3795 			    (1 << hmeshift));
3796 		}
3797 	}
3798 	sfmmu_hblks_list_purge(&list, 0);
3799 }
3800 
3801 /*
3802  * This routine's job is to delete stale invalid shared hmeregions hmeblks that
3803  * may still linger on after pageunload.
3804  */
3805 static void
3806 sfmmu_cleanup_rhblk(sf_srd_t *srdp, caddr_t addr, uint_t rid, int ttesz)
3807 {
3808 	int hmeshift;
3809 	hmeblk_tag hblktag;
3810 	struct hmehash_bucket *hmebp;
3811 	struct hme_blk *hmeblkp;
3812 	struct hme_blk *pr_hblk;
3813 	struct hme_blk *list = NULL;
3814 
3815 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3816 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3817 
3818 	hmeshift = HME_HASH_SHIFT(ttesz);
3819 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3820 	hblktag.htag_rehash = ttesz;
3821 	hblktag.htag_rid = rid;
3822 	hblktag.htag_id = srdp;
3823 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3824 
3825 	SFMMU_HASH_LOCK(hmebp);
3826 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3827 	if (hmeblkp != NULL) {
3828 		ASSERT(hmeblkp->hblk_shared);
3829 		ASSERT(!hmeblkp->hblk_shw_bit);
3830 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3831 			panic("sfmmu_cleanup_rhblk: valid hmeblk");
3832 		}
3833 		ASSERT(!hmeblkp->hblk_lckcnt);
3834 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3835 		    &list, 0);
3836 	}
3837 	SFMMU_HASH_UNLOCK(hmebp);
3838 	sfmmu_hblks_list_purge(&list, 0);
3839 }
3840 
3841 /* ARGSUSED */
3842 static void
3843 sfmmu_rgn_cb_noop(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
3844     size_t r_size, void *r_obj, u_offset_t r_objoff)
3845 {
3846 }
3847 
3848 /*
3849  * Searches for an hmeblk which maps addr, then unloads this mapping
3850  * and updates *eaddrp, if the hmeblk is found.
3851  */
3852 static void
3853 sfmmu_unload_hmeregion_va(sf_srd_t *srdp, uint_t rid, caddr_t addr,
3854     caddr_t eaddr, int ttesz, caddr_t *eaddrp)
3855 {
3856 	int hmeshift;
3857 	hmeblk_tag hblktag;
3858 	struct hmehash_bucket *hmebp;
3859 	struct hme_blk *hmeblkp;
3860 	struct hme_blk *pr_hblk;
3861 	struct hme_blk *list = NULL;
3862 
3863 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3864 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3865 	ASSERT(ttesz >= HBLK_MIN_TTESZ);
3866 
3867 	hmeshift = HME_HASH_SHIFT(ttesz);
3868 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3869 	hblktag.htag_rehash = ttesz;
3870 	hblktag.htag_rid = rid;
3871 	hblktag.htag_id = srdp;
3872 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3873 
3874 	SFMMU_HASH_LOCK(hmebp);
3875 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3876 	if (hmeblkp != NULL) {
3877 		ASSERT(hmeblkp->hblk_shared);
3878 		ASSERT(!hmeblkp->hblk_lckcnt);
3879 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3880 			*eaddrp = sfmmu_hblk_unload(NULL, hmeblkp, addr,
3881 			    eaddr, NULL, HAT_UNLOAD);
3882 			ASSERT(*eaddrp > addr);
3883 		}
3884 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3885 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3886 		    &list, 0);
3887 	}
3888 	SFMMU_HASH_UNLOCK(hmebp);
3889 	sfmmu_hblks_list_purge(&list, 0);
3890 }
3891 
3892 static void
3893 sfmmu_unload_hmeregion(sf_srd_t *srdp, sf_region_t *rgnp)
3894 {
3895 	int ttesz = rgnp->rgn_pgszc;
3896 	size_t rsz = rgnp->rgn_size;
3897 	caddr_t rsaddr = rgnp->rgn_saddr;
3898 	caddr_t readdr = rsaddr + rsz;
3899 	caddr_t rhsaddr;
3900 	caddr_t va;
3901 	uint_t rid = rgnp->rgn_id;
3902 	caddr_t cbsaddr;
3903 	caddr_t cbeaddr;
3904 	hat_rgn_cb_func_t rcbfunc;
3905 	ulong_t cnt;
3906 
3907 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3908 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3909 
3910 	ASSERT(IS_P2ALIGNED(rsaddr, TTEBYTES(ttesz)));
3911 	ASSERT(IS_P2ALIGNED(rsz, TTEBYTES(ttesz)));
3912 	if (ttesz < HBLK_MIN_TTESZ) {
3913 		ttesz = HBLK_MIN_TTESZ;
3914 		rhsaddr = (caddr_t)P2ALIGN((uintptr_t)rsaddr, HBLK_MIN_BYTES);
3915 	} else {
3916 		rhsaddr = rsaddr;
3917 	}
3918 
3919 	if ((rcbfunc = rgnp->rgn_cb_function) == NULL) {
3920 		rcbfunc = sfmmu_rgn_cb_noop;
3921 	}
3922 
3923 	while (ttesz >= HBLK_MIN_TTESZ) {
3924 		cbsaddr = rsaddr;
3925 		cbeaddr = rsaddr;
3926 		if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
3927 			ttesz--;
3928 			continue;
3929 		}
3930 		cnt = 0;
3931 		va = rsaddr;
3932 		while (va < readdr) {
3933 			ASSERT(va >= rhsaddr);
3934 			if (va != cbeaddr) {
3935 				if (cbeaddr != cbsaddr) {
3936 					ASSERT(cbeaddr > cbsaddr);
3937 					(*rcbfunc)(cbsaddr, cbeaddr,
3938 					    rsaddr, rsz, rgnp->rgn_obj,
3939 					    rgnp->rgn_objoff);
3940 				}
3941 				cbsaddr = va;
3942 				cbeaddr = va;
3943 			}
3944 			sfmmu_unload_hmeregion_va(srdp, rid, va, readdr,
3945 			    ttesz, &cbeaddr);
3946 			cnt++;
3947 			va = rhsaddr + (cnt << TTE_PAGE_SHIFT(ttesz));
3948 		}
3949 		if (cbeaddr != cbsaddr) {
3950 			ASSERT(cbeaddr > cbsaddr);
3951 			(*rcbfunc)(cbsaddr, cbeaddr, rsaddr,
3952 			    rsz, rgnp->rgn_obj,
3953 			    rgnp->rgn_objoff);
3954 		}
3955 		ttesz--;
3956 	}
3957 }
3958 
3959 /*
3960  * Release one hardware address translation lock on the given address range.
3961  */
3962 void
3963 hat_unlock(struct hat *sfmmup, caddr_t addr, size_t len)
3964 {
3965 	struct hmehash_bucket *hmebp;
3966 	hmeblk_tag hblktag;
3967 	int hmeshift, hashno = 1;
3968 	struct hme_blk *hmeblkp, *list = NULL;
3969 	caddr_t endaddr;
3970 
3971 	ASSERT(sfmmup != NULL);
3972 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3973 
3974 	ASSERT((sfmmup == ksfmmup) || AS_LOCK_HELD(sfmmup->sfmmu_as));
3975 	ASSERT((len & MMU_PAGEOFFSET) == 0);
3976 	endaddr = addr + len;
3977 	hblktag.htag_id = sfmmup;
3978 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3979 
3980 	/*
3981 	 * Spitfire supports 4 page sizes.
3982 	 * Most pages are expected to be of the smallest page size (8K) and
3983 	 * these will not need to be rehashed. 64K pages also don't need to be
3984 	 * rehashed because an hmeblk spans 64K of address space. 512K pages
3985 	 * might need 1 rehash and and 4M pages might need 2 rehashes.
3986 	 */
3987 	while (addr < endaddr) {
3988 		hmeshift = HME_HASH_SHIFT(hashno);
3989 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3990 		hblktag.htag_rehash = hashno;
3991 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3992 
3993 		SFMMU_HASH_LOCK(hmebp);
3994 
3995 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
3996 		if (hmeblkp != NULL) {
3997 			ASSERT(!hmeblkp->hblk_shared);
3998 			/*
3999 			 * If we encounter a shadow hmeblk then
4000 			 * we know there are no valid hmeblks mapping
4001 			 * this address at this size or larger.
4002 			 * Just increment address by the smallest
4003 			 * page size.
4004 			 */
4005 			if (hmeblkp->hblk_shw_bit) {
4006 				addr += MMU_PAGESIZE;
4007 			} else {
4008 				addr = sfmmu_hblk_unlock(hmeblkp, addr,
4009 				    endaddr);
4010 			}
4011 			SFMMU_HASH_UNLOCK(hmebp);
4012 			hashno = 1;
4013 			continue;
4014 		}
4015 		SFMMU_HASH_UNLOCK(hmebp);
4016 
4017 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
4018 			/*
4019 			 * We have traversed the whole list and rehashed
4020 			 * if necessary without finding the address to unlock
4021 			 * which should never happen.
4022 			 */
4023 			panic("sfmmu_unlock: addr not found. "
4024 			    "addr %p hat %p", (void *)addr, (void *)sfmmup);
4025 		} else {
4026 			hashno++;
4027 		}
4028 	}
4029 
4030 	sfmmu_hblks_list_purge(&list, 0);
4031 }
4032 
4033 void
4034 hat_unlock_region(struct hat *sfmmup, caddr_t addr, size_t len,
4035     hat_region_cookie_t rcookie)
4036 {
4037 	sf_srd_t *srdp;
4038 	sf_region_t *rgnp;
4039 	int ttesz;
4040 	uint_t rid;
4041 	caddr_t eaddr;
4042 	caddr_t va;
4043 	int hmeshift;
4044 	hmeblk_tag hblktag;
4045 	struct hmehash_bucket *hmebp;
4046 	struct hme_blk *hmeblkp;
4047 	struct hme_blk *pr_hblk;
4048 	struct hme_blk *list;
4049 
4050 	if (rcookie == HAT_INVALID_REGION_COOKIE) {
4051 		hat_unlock(sfmmup, addr, len);
4052 		return;
4053 	}
4054 
4055 	ASSERT(sfmmup != NULL);
4056 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4057 	ASSERT(sfmmup != ksfmmup);
4058 
4059 	srdp = sfmmup->sfmmu_srdp;
4060 	rid = (uint_t)((uint64_t)rcookie);
4061 	VERIFY3U(rid, <, SFMMU_MAX_HME_REGIONS);
4062 	eaddr = addr + len;
4063 	va = addr;
4064 	list = NULL;
4065 	rgnp = srdp->srd_hmergnp[rid];
4066 	SFMMU_VALIDATE_HMERID(sfmmup, rid, addr, len);
4067 
4068 	ASSERT(IS_P2ALIGNED(addr, TTEBYTES(rgnp->rgn_pgszc)));
4069 	ASSERT(IS_P2ALIGNED(len, TTEBYTES(rgnp->rgn_pgszc)));
4070 	if (rgnp->rgn_pgszc < HBLK_MIN_TTESZ) {
4071 		ttesz = HBLK_MIN_TTESZ;
4072 	} else {
4073 		ttesz = rgnp->rgn_pgszc;
4074 	}
4075 	while (va < eaddr) {
4076 		while (ttesz < rgnp->rgn_pgszc &&
4077 		    IS_P2ALIGNED(va, TTEBYTES(ttesz + 1))) {
4078 			ttesz++;
4079 		}
4080 		while (ttesz >= HBLK_MIN_TTESZ) {
4081 			if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
4082 				ttesz--;
4083 				continue;
4084 			}
4085 			hmeshift = HME_HASH_SHIFT(ttesz);
4086 			hblktag.htag_bspage = HME_HASH_BSPAGE(va, hmeshift);
4087 			hblktag.htag_rehash = ttesz;
4088 			hblktag.htag_rid = rid;
4089 			hblktag.htag_id = srdp;
4090 			hmebp = HME_HASH_FUNCTION(srdp, va, hmeshift);
4091 			SFMMU_HASH_LOCK(hmebp);
4092 			HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk,
4093 			    &list);
4094 			if (hmeblkp == NULL) {
4095 				SFMMU_HASH_UNLOCK(hmebp);
4096 				ttesz--;
4097 				continue;
4098 			}
4099 			ASSERT(hmeblkp->hblk_shared);
4100 			va = sfmmu_hblk_unlock(hmeblkp, va, eaddr);
4101 			ASSERT(va >= eaddr ||
4102 			    IS_P2ALIGNED((uintptr_t)va, TTEBYTES(ttesz)));
4103 			SFMMU_HASH_UNLOCK(hmebp);
4104 			break;
4105 		}
4106 		if (ttesz < HBLK_MIN_TTESZ) {
4107 			panic("hat_unlock_region: addr not found "
4108 			    "addr %p hat %p", (void *)va, (void *)sfmmup);
4109 		}
4110 	}
4111 	sfmmu_hblks_list_purge(&list, 0);
4112 }
4113 
4114 /*
4115  * Function to unlock a range of addresses in an hmeblk.  It returns the
4116  * next address that needs to be unlocked.
4117  * Should be called with the hash lock held.
4118  */
4119 static caddr_t
4120 sfmmu_hblk_unlock(struct hme_blk *hmeblkp, caddr_t addr, caddr_t endaddr)
4121 {
4122 	struct sf_hment *sfhme;
4123 	tte_t tteold, ttemod;
4124 	int ttesz, ret;
4125 
4126 	ASSERT(in_hblk_range(hmeblkp, addr));
4127 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4128 
4129 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4130 	ttesz = get_hblk_ttesz(hmeblkp);
4131 
4132 	HBLKTOHME(sfhme, hmeblkp, addr);
4133 	while (addr < endaddr) {
4134 readtte:
4135 		sfmmu_copytte(&sfhme->hme_tte, &tteold);
4136 		if (TTE_IS_VALID(&tteold)) {
4137 
4138 			ttemod = tteold;
4139 
4140 			ret = sfmmu_modifytte_try(&tteold, &ttemod,
4141 			    &sfhme->hme_tte);
4142 
4143 			if (ret < 0)
4144 				goto readtte;
4145 
4146 			if (hmeblkp->hblk_lckcnt == 0)
4147 				panic("zero hblk lckcnt");
4148 
4149 			if (((uintptr_t)addr + TTEBYTES(ttesz)) >
4150 			    (uintptr_t)endaddr)
4151 				panic("can't unlock large tte");
4152 
4153 			ASSERT(hmeblkp->hblk_lckcnt > 0);
4154 			atomic_dec_32(&hmeblkp->hblk_lckcnt);
4155 			HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
4156 		} else {
4157 			panic("sfmmu_hblk_unlock: invalid tte");
4158 		}
4159 		addr += TTEBYTES(ttesz);
4160 		sfhme++;
4161 	}
4162 	return (addr);
4163 }
4164 
4165 /*
4166  * Physical Address Mapping Framework
4167  *
4168  * General rules:
4169  *
4170  * (1) Applies only to seg_kmem memory pages. To make things easier,
4171  *     seg_kpm addresses are also accepted by the routines, but nothing
4172  *     is done with them since by definition their PA mappings are static.
4173  * (2) hat_add_callback() may only be called while holding the page lock
4174  *     SE_SHARED or SE_EXCL of the underlying page (e.g., as_pagelock()),
4175  *     or passing HAC_PAGELOCK flag.
4176  * (3) prehandler() and posthandler() may not call hat_add_callback() or
4177  *     hat_delete_callback(), nor should they allocate memory. Post quiesce
4178  *     callbacks may not sleep or acquire adaptive mutex locks.
4179  * (4) Either prehandler() or posthandler() (but not both) may be specified
4180  *     as being NULL.  Specifying an errhandler() is optional.
4181  *
4182  * Details of using the framework:
4183  *
4184  * registering a callback (hat_register_callback())
4185  *
4186  *	Pass prehandler, posthandler, errhandler addresses
4187  *	as described below. If capture_cpus argument is nonzero,
4188  *	suspend callback to the prehandler will occur with CPUs
4189  *	captured and executing xc_loop() and CPUs will remain
4190  *	captured until after the posthandler suspend callback
4191  *	occurs.
4192  *
4193  * adding a callback (hat_add_callback())
4194  *
4195  *      as_pagelock();
4196  *	hat_add_callback();
4197  *      save returned pfn in private data structures or program registers;
4198  *      as_pageunlock();
4199  *
4200  * prehandler()
4201  *
4202  *	Stop all accesses by physical address to this memory page.
4203  *	Called twice: the first, PRESUSPEND, is a context safe to acquire
4204  *	adaptive locks. The second, SUSPEND, is called at high PIL with
4205  *	CPUs captured so adaptive locks may NOT be acquired (and all spin
4206  *	locks must be XCALL_PIL or higher locks).
4207  *
4208  *	May return the following errors:
4209  *		EIO:	A fatal error has occurred. This will result in panic.
4210  *		EAGAIN:	The page cannot be suspended. This will fail the
4211  *			relocation.
4212  *		0:	Success.
4213  *
4214  * posthandler()
4215  *
4216  *      Save new pfn in private data structures or program registers;
4217  *	not allowed to fail (non-zero return values will result in panic).
4218  *
4219  * errhandler()
4220  *
4221  *	called when an error occurs related to the callback.  Currently
4222  *	the only such error is HAT_CB_ERR_LEAKED which indicates that
4223  *	a page is being freed, but there are still outstanding callback(s)
4224  *	registered on the page.
4225  *
4226  * removing a callback (hat_delete_callback(); e.g., prior to freeing memory)
4227  *
4228  *	stop using physical address
4229  *	hat_delete_callback();
4230  *
4231  */
4232 
4233 /*
4234  * Register a callback class.  Each subsystem should do this once and
4235  * cache the id_t returned for use in setting up and tearing down callbacks.
4236  *
4237  * There is no facility for removing callback IDs once they are created;
4238  * the "key" should be unique for each module, so in case a module is unloaded
4239  * and subsequently re-loaded, we can recycle the module's previous entry.
4240  */
4241 id_t
4242 hat_register_callback(int key,
4243 	int (*prehandler)(caddr_t, uint_t, uint_t, void *),
4244 	int (*posthandler)(caddr_t, uint_t, uint_t, void *, pfn_t),
4245 	int (*errhandler)(caddr_t, uint_t, uint_t, void *),
4246 	int capture_cpus)
4247 {
4248 	id_t id;
4249 
4250 	/*
4251 	 * Search the table for a pre-existing callback associated with
4252 	 * the identifier "key".  If one exists, we re-use that entry in
4253 	 * the table for this instance, otherwise we assign the next
4254 	 * available table slot.
4255 	 */
4256 	for (id = 0; id < sfmmu_max_cb_id; id++) {
4257 		if (sfmmu_cb_table[id].key == key)
4258 			break;
4259 	}
4260 
4261 	if (id == sfmmu_max_cb_id) {
4262 		id = sfmmu_cb_nextid++;
4263 		if (id >= sfmmu_max_cb_id)
4264 			panic("hat_register_callback: out of callback IDs");
4265 	}
4266 
4267 	ASSERT(prehandler != NULL || posthandler != NULL);
4268 
4269 	sfmmu_cb_table[id].key = key;
4270 	sfmmu_cb_table[id].prehandler = prehandler;
4271 	sfmmu_cb_table[id].posthandler = posthandler;
4272 	sfmmu_cb_table[id].errhandler = errhandler;
4273 	sfmmu_cb_table[id].capture_cpus = capture_cpus;
4274 
4275 	return (id);
4276 }
4277 
4278 #define	HAC_COOKIE_NONE	(void *)-1
4279 
4280 /*
4281  * Add relocation callbacks to the specified addr/len which will be called
4282  * when relocating the associated page. See the description of pre and
4283  * posthandler above for more details.
4284  *
4285  * If HAC_PAGELOCK is included in flags, the underlying memory page is
4286  * locked internally so the caller must be able to deal with the callback
4287  * running even before this function has returned.  If HAC_PAGELOCK is not
4288  * set, it is assumed that the underlying memory pages are locked.
4289  *
4290  * Since the caller must track the individual page boundaries anyway,
4291  * we only allow a callback to be added to a single page (large
4292  * or small).  Thus [addr, addr + len) MUST be contained within a single
4293  * page.
4294  *
4295  * Registering multiple callbacks on the same [addr, addr+len) is supported,
4296  * _provided_that_ a unique parameter is specified for each callback.
4297  * If multiple callbacks are registered on the same range the callback will
4298  * be invoked with each unique parameter. Registering the same callback with
4299  * the same argument more than once will result in corrupted kernel state.
4300  *
4301  * Returns the pfn of the underlying kernel page in *rpfn
4302  * on success, or PFN_INVALID on failure.
4303  *
4304  * cookiep (if passed) provides storage space for an opaque cookie
4305  * to return later to hat_delete_callback(). This cookie makes the callback
4306  * deletion significantly quicker by avoiding a potentially lengthy hash
4307  * search.
4308  *
4309  * Returns values:
4310  *    0:      success
4311  *    ENOMEM: memory allocation failure (e.g. flags was passed as HAC_NOSLEEP)
4312  *    EINVAL: callback ID is not valid
4313  *    ENXIO:  ["vaddr", "vaddr" + len) is not mapped in the kernel's address
4314  *            space
4315  *    ERANGE: ["vaddr", "vaddr" + len) crosses a page boundary
4316  */
4317 int
4318 hat_add_callback(id_t callback_id, caddr_t vaddr, uint_t len, uint_t flags,
4319 	void *pvt, pfn_t *rpfn, void **cookiep)
4320 {
4321 	struct 		hmehash_bucket *hmebp;
4322 	hmeblk_tag 	hblktag;
4323 	struct hme_blk	*hmeblkp;
4324 	int 		hmeshift, hashno;
4325 	caddr_t 	saddr, eaddr, baseaddr;
4326 	struct pa_hment *pahmep;
4327 	struct sf_hment *sfhmep, *osfhmep;
4328 	kmutex_t	*pml;
4329 	tte_t   	tte;
4330 	page_t		*pp;
4331 	vnode_t		*vp;
4332 	u_offset_t	off;
4333 	pfn_t		pfn;
4334 	int		kmflags = (flags & HAC_SLEEP)? KM_SLEEP : KM_NOSLEEP;
4335 	int		locked = 0;
4336 
4337 	/*
4338 	 * For KPM mappings, just return the physical address since we
4339 	 * don't need to register any callbacks.
4340 	 */
4341 	if (IS_KPM_ADDR(vaddr)) {
4342 		uint64_t paddr;
4343 		SFMMU_KPM_VTOP(vaddr, paddr);
4344 		*rpfn = btop(paddr);
4345 		if (cookiep != NULL)
4346 			*cookiep = HAC_COOKIE_NONE;
4347 		return (0);
4348 	}
4349 
4350 	if (callback_id < (id_t)0 || callback_id >= sfmmu_cb_nextid) {
4351 		*rpfn = PFN_INVALID;
4352 		return (EINVAL);
4353 	}
4354 
4355 	if ((pahmep = kmem_cache_alloc(pa_hment_cache, kmflags)) == NULL) {
4356 		*rpfn = PFN_INVALID;
4357 		return (ENOMEM);
4358 	}
4359 
4360 	sfhmep = &pahmep->sfment;
4361 
4362 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4363 	eaddr = saddr + len;
4364 
4365 rehash:
4366 	/* Find the mapping(s) for this page */
4367 	for (hashno = TTE64K, hmeblkp = NULL;
4368 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4369 	    hashno++) {
4370 		hmeshift = HME_HASH_SHIFT(hashno);
4371 		hblktag.htag_id = ksfmmup;
4372 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4373 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4374 		hblktag.htag_rehash = hashno;
4375 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4376 
4377 		SFMMU_HASH_LOCK(hmebp);
4378 
4379 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4380 
4381 		if (hmeblkp == NULL)
4382 			SFMMU_HASH_UNLOCK(hmebp);
4383 	}
4384 
4385 	if (hmeblkp == NULL) {
4386 		kmem_cache_free(pa_hment_cache, pahmep);
4387 		*rpfn = PFN_INVALID;
4388 		return (ENXIO);
4389 	}
4390 
4391 	ASSERT(!hmeblkp->hblk_shared);
4392 
4393 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4394 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4395 
4396 	if (!TTE_IS_VALID(&tte)) {
4397 		SFMMU_HASH_UNLOCK(hmebp);
4398 		kmem_cache_free(pa_hment_cache, pahmep);
4399 		*rpfn = PFN_INVALID;
4400 		return (ENXIO);
4401 	}
4402 
4403 	/*
4404 	 * Make sure the boundaries for the callback fall within this
4405 	 * single mapping.
4406 	 */
4407 	baseaddr = (caddr_t)get_hblk_base(hmeblkp);
4408 	ASSERT(saddr >= baseaddr);
4409 	if (eaddr > saddr + TTEBYTES(TTE_CSZ(&tte))) {
4410 		SFMMU_HASH_UNLOCK(hmebp);
4411 		kmem_cache_free(pa_hment_cache, pahmep);
4412 		*rpfn = PFN_INVALID;
4413 		return (ERANGE);
4414 	}
4415 
4416 	pfn = sfmmu_ttetopfn(&tte, vaddr);
4417 
4418 	/*
4419 	 * The pfn may not have a page_t underneath in which case we
4420 	 * just return it. This can happen if we are doing I/O to a
4421 	 * static portion of the kernel's address space, for instance.
4422 	 */
4423 	pp = osfhmep->hme_page;
4424 	if (pp == NULL) {
4425 		SFMMU_HASH_UNLOCK(hmebp);
4426 		kmem_cache_free(pa_hment_cache, pahmep);
4427 		*rpfn = pfn;
4428 		if (cookiep)
4429 			*cookiep = HAC_COOKIE_NONE;
4430 		return (0);
4431 	}
4432 	ASSERT(pp == PP_PAGEROOT(pp));
4433 
4434 	vp = pp->p_vnode;
4435 	off = pp->p_offset;
4436 
4437 	pml = sfmmu_mlist_enter(pp);
4438 
4439 	if (flags & HAC_PAGELOCK) {
4440 		if (!page_trylock(pp, SE_SHARED)) {
4441 			/*
4442 			 * Somebody is holding SE_EXCL lock. Might
4443 			 * even be hat_page_relocate(). Drop all
4444 			 * our locks, lookup the page in &kvp, and
4445 			 * retry. If it doesn't exist in &kvp and &zvp,
4446 			 * then we must be dealing with a kernel mapped
4447 			 * page which doesn't actually belong to
4448 			 * segkmem so we punt.
4449 			 */
4450 			sfmmu_mlist_exit(pml);
4451 			SFMMU_HASH_UNLOCK(hmebp);
4452 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4453 
4454 			/* check zvp before giving up */
4455 			if (pp == NULL)
4456 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4457 				    SE_SHARED);
4458 
4459 			/* Okay, we didn't find it, give up */
4460 			if (pp == NULL) {
4461 				kmem_cache_free(pa_hment_cache, pahmep);
4462 				*rpfn = pfn;
4463 				if (cookiep)
4464 					*cookiep = HAC_COOKIE_NONE;
4465 				return (0);
4466 			}
4467 			page_unlock(pp);
4468 			goto rehash;
4469 		}
4470 		locked = 1;
4471 	}
4472 
4473 	if (!PAGE_LOCKED(pp) && !panicstr)
4474 		panic("hat_add_callback: page 0x%p not locked", (void *)pp);
4475 
4476 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4477 	    pp->p_offset != off) {
4478 		/*
4479 		 * The page moved before we got our hands on it.  Drop
4480 		 * all the locks and try again.
4481 		 */
4482 		ASSERT((flags & HAC_PAGELOCK) != 0);
4483 		sfmmu_mlist_exit(pml);
4484 		SFMMU_HASH_UNLOCK(hmebp);
4485 		page_unlock(pp);
4486 		locked = 0;
4487 		goto rehash;
4488 	}
4489 
4490 	if (!VN_ISKAS(vp)) {
4491 		/*
4492 		 * This is not a segkmem page but another page which
4493 		 * has been kernel mapped. It had better have at least
4494 		 * a share lock on it. Return the pfn.
4495 		 */
4496 		sfmmu_mlist_exit(pml);
4497 		SFMMU_HASH_UNLOCK(hmebp);
4498 		if (locked)
4499 			page_unlock(pp);
4500 		kmem_cache_free(pa_hment_cache, pahmep);
4501 		ASSERT(PAGE_LOCKED(pp));
4502 		*rpfn = pfn;
4503 		if (cookiep)
4504 			*cookiep = HAC_COOKIE_NONE;
4505 		return (0);
4506 	}
4507 
4508 	/*
4509 	 * Setup this pa_hment and link its embedded dummy sf_hment into
4510 	 * the mapping list.
4511 	 */
4512 	pp->p_share++;
4513 	pahmep->cb_id = callback_id;
4514 	pahmep->addr = vaddr;
4515 	pahmep->len = len;
4516 	pahmep->refcnt = 1;
4517 	pahmep->flags = 0;
4518 	pahmep->pvt = pvt;
4519 
4520 	sfhmep->hme_tte.ll = 0;
4521 	sfhmep->hme_data = pahmep;
4522 	sfhmep->hme_prev = osfhmep;
4523 	sfhmep->hme_next = osfhmep->hme_next;
4524 
4525 	if (osfhmep->hme_next)
4526 		osfhmep->hme_next->hme_prev = sfhmep;
4527 
4528 	osfhmep->hme_next = sfhmep;
4529 
4530 	sfmmu_mlist_exit(pml);
4531 	SFMMU_HASH_UNLOCK(hmebp);
4532 
4533 	if (locked)
4534 		page_unlock(pp);
4535 
4536 	*rpfn = pfn;
4537 	if (cookiep)
4538 		*cookiep = (void *)pahmep;
4539 
4540 	return (0);
4541 }
4542 
4543 /*
4544  * Remove the relocation callbacks from the specified addr/len.
4545  */
4546 void
4547 hat_delete_callback(caddr_t vaddr, uint_t len, void *pvt, uint_t flags,
4548 	void *cookie)
4549 {
4550 	struct		hmehash_bucket *hmebp;
4551 	hmeblk_tag	hblktag;
4552 	struct hme_blk	*hmeblkp;
4553 	int		hmeshift, hashno;
4554 	caddr_t		saddr;
4555 	struct pa_hment	*pahmep;
4556 	struct sf_hment	*sfhmep, *osfhmep;
4557 	kmutex_t	*pml;
4558 	tte_t		tte;
4559 	page_t		*pp;
4560 	vnode_t		*vp;
4561 	u_offset_t	off;
4562 	int		locked = 0;
4563 
4564 	/*
4565 	 * If the cookie is HAC_COOKIE_NONE then there is no pa_hment to
4566 	 * remove so just return.
4567 	 */
4568 	if (cookie == HAC_COOKIE_NONE || IS_KPM_ADDR(vaddr))
4569 		return;
4570 
4571 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4572 
4573 rehash:
4574 	/* Find the mapping(s) for this page */
4575 	for (hashno = TTE64K, hmeblkp = NULL;
4576 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4577 	    hashno++) {
4578 		hmeshift = HME_HASH_SHIFT(hashno);
4579 		hblktag.htag_id = ksfmmup;
4580 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4581 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4582 		hblktag.htag_rehash = hashno;
4583 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4584 
4585 		SFMMU_HASH_LOCK(hmebp);
4586 
4587 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4588 
4589 		if (hmeblkp == NULL)
4590 			SFMMU_HASH_UNLOCK(hmebp);
4591 	}
4592 
4593 	if (hmeblkp == NULL)
4594 		return;
4595 
4596 	ASSERT(!hmeblkp->hblk_shared);
4597 
4598 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4599 
4600 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4601 	if (!TTE_IS_VALID(&tte)) {
4602 		SFMMU_HASH_UNLOCK(hmebp);
4603 		return;
4604 	}
4605 
4606 	pp = osfhmep->hme_page;
4607 	if (pp == NULL) {
4608 		SFMMU_HASH_UNLOCK(hmebp);
4609 		ASSERT(cookie == NULL);
4610 		return;
4611 	}
4612 
4613 	vp = pp->p_vnode;
4614 	off = pp->p_offset;
4615 
4616 	pml = sfmmu_mlist_enter(pp);
4617 
4618 	if (flags & HAC_PAGELOCK) {
4619 		if (!page_trylock(pp, SE_SHARED)) {
4620 			/*
4621 			 * Somebody is holding SE_EXCL lock. Might
4622 			 * even be hat_page_relocate(). Drop all
4623 			 * our locks, lookup the page in &kvp, and
4624 			 * retry. If it doesn't exist in &kvp and &zvp,
4625 			 * then we must be dealing with a kernel mapped
4626 			 * page which doesn't actually belong to
4627 			 * segkmem so we punt.
4628 			 */
4629 			sfmmu_mlist_exit(pml);
4630 			SFMMU_HASH_UNLOCK(hmebp);
4631 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4632 			/* check zvp before giving up */
4633 			if (pp == NULL)
4634 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4635 				    SE_SHARED);
4636 
4637 			if (pp == NULL) {
4638 				ASSERT(cookie == NULL);
4639 				return;
4640 			}
4641 			page_unlock(pp);
4642 			goto rehash;
4643 		}
4644 		locked = 1;
4645 	}
4646 
4647 	ASSERT(PAGE_LOCKED(pp));
4648 
4649 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4650 	    pp->p_offset != off) {
4651 		/*
4652 		 * The page moved before we got our hands on it.  Drop
4653 		 * all the locks and try again.
4654 		 */
4655 		ASSERT((flags & HAC_PAGELOCK) != 0);
4656 		sfmmu_mlist_exit(pml);
4657 		SFMMU_HASH_UNLOCK(hmebp);
4658 		page_unlock(pp);
4659 		locked = 0;
4660 		goto rehash;
4661 	}
4662 
4663 	if (!VN_ISKAS(vp)) {
4664 		/*
4665 		 * This is not a segkmem page but another page which
4666 		 * has been kernel mapped.
4667 		 */
4668 		sfmmu_mlist_exit(pml);
4669 		SFMMU_HASH_UNLOCK(hmebp);
4670 		if (locked)
4671 			page_unlock(pp);
4672 		ASSERT(cookie == NULL);
4673 		return;
4674 	}
4675 
4676 	if (cookie != NULL) {
4677 		pahmep = (struct pa_hment *)cookie;
4678 		sfhmep = &pahmep->sfment;
4679 	} else {
4680 		for (sfhmep = pp->p_mapping; sfhmep != NULL;
4681 		    sfhmep = sfhmep->hme_next) {
4682 
4683 			/*
4684 			 * skip va<->pa mappings
4685 			 */
4686 			if (!IS_PAHME(sfhmep))
4687 				continue;
4688 
4689 			pahmep = sfhmep->hme_data;
4690 			ASSERT(pahmep != NULL);
4691 
4692 			/*
4693 			 * if pa_hment matches, remove it
4694 			 */
4695 			if ((pahmep->pvt == pvt) &&
4696 			    (pahmep->addr == vaddr) &&
4697 			    (pahmep->len == len)) {
4698 				break;
4699 			}
4700 		}
4701 	}
4702 
4703 	if (sfhmep == NULL) {
4704 		if (!panicstr) {
4705 			panic("hat_delete_callback: pa_hment not found, pp %p",
4706 			    (void *)pp);
4707 		}
4708 		return;
4709 	}
4710 
4711 	/*
4712 	 * Note: at this point a valid kernel mapping must still be
4713 	 * present on this page.
4714 	 */
4715 	pp->p_share--;
4716 	if (pp->p_share <= 0)
4717 		panic("hat_delete_callback: zero p_share");
4718 
4719 	if (--pahmep->refcnt == 0) {
4720 		if (pahmep->flags != 0)
4721 			panic("hat_delete_callback: pa_hment is busy");
4722 
4723 		/*
4724 		 * Remove sfhmep from the mapping list for the page.
4725 		 */
4726 		if (sfhmep->hme_prev) {
4727 			sfhmep->hme_prev->hme_next = sfhmep->hme_next;
4728 		} else {
4729 			pp->p_mapping = sfhmep->hme_next;
4730 		}
4731 
4732 		if (sfhmep->hme_next)
4733 			sfhmep->hme_next->hme_prev = sfhmep->hme_prev;
4734 
4735 		sfmmu_mlist_exit(pml);
4736 		SFMMU_HASH_UNLOCK(hmebp);
4737 
4738 		if (locked)
4739 			page_unlock(pp);
4740 
4741 		kmem_cache_free(pa_hment_cache, pahmep);
4742 		return;
4743 	}
4744 
4745 	sfmmu_mlist_exit(pml);
4746 	SFMMU_HASH_UNLOCK(hmebp);
4747 	if (locked)
4748 		page_unlock(pp);
4749 }
4750 
4751 /*
4752  * hat_probe returns 1 if the translation for the address 'addr' is
4753  * loaded, zero otherwise.
4754  *
4755  * hat_probe should be used only for advisorary purposes because it may
4756  * occasionally return the wrong value. The implementation must guarantee that
4757  * returning the wrong value is a very rare event. hat_probe is used
4758  * to implement optimizations in the segment drivers.
4759  *
4760  */
4761 int
4762 hat_probe(struct hat *sfmmup, caddr_t addr)
4763 {
4764 	pfn_t pfn;
4765 	tte_t tte;
4766 
4767 	ASSERT(sfmmup != NULL);
4768 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4769 
4770 	ASSERT((sfmmup == ksfmmup) || AS_LOCK_HELD(sfmmup->sfmmu_as));
4771 
4772 	if (sfmmup == ksfmmup) {
4773 		while ((pfn = sfmmu_vatopfn(addr, sfmmup, &tte))
4774 		    == PFN_SUSPENDED) {
4775 			sfmmu_vatopfn_suspended(addr, sfmmup, &tte);
4776 		}
4777 	} else {
4778 		pfn = sfmmu_uvatopfn(addr, sfmmup, NULL);
4779 	}
4780 
4781 	if (pfn != PFN_INVALID)
4782 		return (1);
4783 	else
4784 		return (0);
4785 }
4786 
4787 ssize_t
4788 hat_getpagesize(struct hat *sfmmup, caddr_t addr)
4789 {
4790 	tte_t tte;
4791 
4792 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4793 
4794 	if (sfmmup == ksfmmup) {
4795 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4796 			return (-1);
4797 		}
4798 	} else {
4799 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4800 			return (-1);
4801 		}
4802 	}
4803 
4804 	ASSERT(TTE_IS_VALID(&tte));
4805 	return (TTEBYTES(TTE_CSZ(&tte)));
4806 }
4807 
4808 uint_t
4809 hat_getattr(struct hat *sfmmup, caddr_t addr, uint_t *attr)
4810 {
4811 	tte_t tte;
4812 
4813 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4814 
4815 	if (sfmmup == ksfmmup) {
4816 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4817 			tte.ll = 0;
4818 		}
4819 	} else {
4820 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4821 			tte.ll = 0;
4822 		}
4823 	}
4824 	if (TTE_IS_VALID(&tte)) {
4825 		*attr = sfmmu_ptov_attr(&tte);
4826 		return (0);
4827 	}
4828 	*attr = 0;
4829 	return ((uint_t)0xffffffff);
4830 }
4831 
4832 /*
4833  * Enables more attributes on specified address range (ie. logical OR)
4834  */
4835 void
4836 hat_setattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4837 {
4838 	if (hat->sfmmu_xhat_provider) {
4839 		XHAT_SETATTR(hat, addr, len, attr);
4840 		return;
4841 	} else {
4842 		/*
4843 		 * This must be a CPU HAT. If the address space has
4844 		 * XHATs attached, change attributes for all of them,
4845 		 * just in case
4846 		 */
4847 		ASSERT(hat->sfmmu_as != NULL);
4848 		if (hat->sfmmu_as->a_xhat != NULL)
4849 			xhat_setattr_all(hat->sfmmu_as, addr, len, attr);
4850 	}
4851 
4852 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_SETATTR);
4853 }
4854 
4855 /*
4856  * Assigns attributes to the specified address range.  All the attributes
4857  * are specified.
4858  */
4859 void
4860 hat_chgattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4861 {
4862 	if (hat->sfmmu_xhat_provider) {
4863 		XHAT_CHGATTR(hat, addr, len, attr);
4864 		return;
4865 	} else {
4866 		/*
4867 		 * This must be a CPU HAT. If the address space has
4868 		 * XHATs attached, change attributes for all of them,
4869 		 * just in case
4870 		 */
4871 		ASSERT(hat->sfmmu_as != NULL);
4872 		if (hat->sfmmu_as->a_xhat != NULL)
4873 			xhat_chgattr_all(hat->sfmmu_as, addr, len, attr);
4874 	}
4875 
4876 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CHGATTR);
4877 }
4878 
4879 /*
4880  * Remove attributes on the specified address range (ie. loginal NAND)
4881  */
4882 void
4883 hat_clrattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4884 {
4885 	if (hat->sfmmu_xhat_provider) {
4886 		XHAT_CLRATTR(hat, addr, len, attr);
4887 		return;
4888 	} else {
4889 		/*
4890 		 * This must be a CPU HAT. If the address space has
4891 		 * XHATs attached, change attributes for all of them,
4892 		 * just in case
4893 		 */
4894 		ASSERT(hat->sfmmu_as != NULL);
4895 		if (hat->sfmmu_as->a_xhat != NULL)
4896 			xhat_clrattr_all(hat->sfmmu_as, addr, len, attr);
4897 	}
4898 
4899 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CLRATTR);
4900 }
4901 
4902 /*
4903  * Change attributes on an address range to that specified by attr and mode.
4904  */
4905 static void
4906 sfmmu_chgattr(struct hat *sfmmup, caddr_t addr, size_t len, uint_t attr,
4907 	int mode)
4908 {
4909 	struct hmehash_bucket *hmebp;
4910 	hmeblk_tag hblktag;
4911 	int hmeshift, hashno = 1;
4912 	struct hme_blk *hmeblkp, *list = NULL;
4913 	caddr_t endaddr;
4914 	cpuset_t cpuset;
4915 	demap_range_t dmr;
4916 
4917 	CPUSET_ZERO(cpuset);
4918 
4919 	ASSERT((sfmmup == ksfmmup) || AS_LOCK_HELD(sfmmup->sfmmu_as));
4920 	ASSERT((len & MMU_PAGEOFFSET) == 0);
4921 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
4922 
4923 	if ((attr & PROT_USER) && (mode != SFMMU_CLRATTR) &&
4924 	    ((addr + len) > (caddr_t)USERLIMIT)) {
4925 		panic("user addr %p in kernel space",
4926 		    (void *)addr);
4927 	}
4928 
4929 	endaddr = addr + len;
4930 	hblktag.htag_id = sfmmup;
4931 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4932 	DEMAP_RANGE_INIT(sfmmup, &dmr);
4933 
4934 	while (addr < endaddr) {
4935 		hmeshift = HME_HASH_SHIFT(hashno);
4936 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
4937 		hblktag.htag_rehash = hashno;
4938 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
4939 
4940 		SFMMU_HASH_LOCK(hmebp);
4941 
4942 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
4943 		if (hmeblkp != NULL) {
4944 			ASSERT(!hmeblkp->hblk_shared);
4945 			/*
4946 			 * We've encountered a shadow hmeblk so skip the range
4947 			 * of the next smaller mapping size.
4948 			 */
4949 			if (hmeblkp->hblk_shw_bit) {
4950 				ASSERT(sfmmup != ksfmmup);
4951 				ASSERT(hashno > 1);
4952 				addr = (caddr_t)P2END((uintptr_t)addr,
4953 				    TTEBYTES(hashno - 1));
4954 			} else {
4955 				addr = sfmmu_hblk_chgattr(sfmmup,
4956 				    hmeblkp, addr, endaddr, &dmr, attr, mode);
4957 			}
4958 			SFMMU_HASH_UNLOCK(hmebp);
4959 			hashno = 1;
4960 			continue;
4961 		}
4962 		SFMMU_HASH_UNLOCK(hmebp);
4963 
4964 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
4965 			/*
4966 			 * We have traversed the whole list and rehashed
4967 			 * if necessary without finding the address to chgattr.
4968 			 * This is ok, so we increment the address by the
4969 			 * smallest hmeblk range for kernel mappings or for
4970 			 * user mappings with no large pages, and the largest
4971 			 * hmeblk range, to account for shadow hmeblks, for
4972 			 * user mappings with large pages and continue.
4973 			 */
4974 			if (sfmmup == ksfmmup)
4975 				addr = (caddr_t)P2END((uintptr_t)addr,
4976 				    TTEBYTES(1));
4977 			else
4978 				addr = (caddr_t)P2END((uintptr_t)addr,
4979 				    TTEBYTES(hashno));
4980 			hashno = 1;
4981 		} else {
4982 			hashno++;
4983 		}
4984 	}
4985 
4986 	sfmmu_hblks_list_purge(&list, 0);
4987 	DEMAP_RANGE_FLUSH(&dmr);
4988 	cpuset = sfmmup->sfmmu_cpusran;
4989 	xt_sync(cpuset);
4990 }
4991 
4992 /*
4993  * This function chgattr on a range of addresses in an hmeblk.  It returns the
4994  * next addres that needs to be chgattr.
4995  * It should be called with the hash lock held.
4996  * XXX It should be possible to optimize chgattr by not flushing every time but
4997  * on the other hand:
4998  * 1. do one flush crosscall.
4999  * 2. only flush if we are increasing permissions (make sure this will work)
5000  */
5001 static caddr_t
5002 sfmmu_hblk_chgattr(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5003 	caddr_t endaddr, demap_range_t *dmrp, uint_t attr, int mode)
5004 {
5005 	tte_t tte, tteattr, tteflags, ttemod;
5006 	struct sf_hment *sfhmep;
5007 	int ttesz;
5008 	struct page *pp = NULL;
5009 	kmutex_t *pml, *pmtx;
5010 	int ret;
5011 	int use_demap_range;
5012 #if defined(SF_ERRATA_57)
5013 	int check_exec;
5014 #endif
5015 
5016 	ASSERT(in_hblk_range(hmeblkp, addr));
5017 	ASSERT(hmeblkp->hblk_shw_bit == 0);
5018 	ASSERT(!hmeblkp->hblk_shared);
5019 
5020 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5021 	ttesz = get_hblk_ttesz(hmeblkp);
5022 
5023 	/*
5024 	 * Flush the current demap region if addresses have been
5025 	 * skipped or the page size doesn't match.
5026 	 */
5027 	use_demap_range = (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp));
5028 	if (use_demap_range) {
5029 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5030 	} else if (dmrp != NULL) {
5031 		DEMAP_RANGE_FLUSH(dmrp);
5032 	}
5033 
5034 	tteattr.ll = sfmmu_vtop_attr(attr, mode, &tteflags);
5035 #if defined(SF_ERRATA_57)
5036 	check_exec = (sfmmup != ksfmmup) &&
5037 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
5038 	    TTE_IS_EXECUTABLE(&tteattr);
5039 #endif
5040 	HBLKTOHME(sfhmep, hmeblkp, addr);
5041 	while (addr < endaddr) {
5042 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5043 		if (TTE_IS_VALID(&tte)) {
5044 			if ((tte.ll & tteflags.ll) == tteattr.ll) {
5045 				/*
5046 				 * if the new attr is the same as old
5047 				 * continue
5048 				 */
5049 				goto next_addr;
5050 			}
5051 			if (!TTE_IS_WRITABLE(&tteattr)) {
5052 				/*
5053 				 * make sure we clear hw modify bit if we
5054 				 * removing write protections
5055 				 */
5056 				tteflags.tte_intlo |= TTE_HWWR_INT;
5057 			}
5058 
5059 			pml = NULL;
5060 			pp = sfhmep->hme_page;
5061 			if (pp) {
5062 				pml = sfmmu_mlist_enter(pp);
5063 			}
5064 
5065 			if (pp != sfhmep->hme_page) {
5066 				/*
5067 				 * tte must have been unloaded.
5068 				 */
5069 				ASSERT(pml);
5070 				sfmmu_mlist_exit(pml);
5071 				continue;
5072 			}
5073 
5074 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5075 
5076 			ttemod = tte;
5077 			ttemod.ll = (ttemod.ll & ~tteflags.ll) | tteattr.ll;
5078 			ASSERT(TTE_TO_TTEPFN(&ttemod) == TTE_TO_TTEPFN(&tte));
5079 
5080 #if defined(SF_ERRATA_57)
5081 			if (check_exec && addr < errata57_limit)
5082 				ttemod.tte_exec_perm = 0;
5083 #endif
5084 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5085 			    &sfhmep->hme_tte);
5086 
5087 			if (ret < 0) {
5088 				/* tte changed underneath us */
5089 				if (pml) {
5090 					sfmmu_mlist_exit(pml);
5091 				}
5092 				continue;
5093 			}
5094 
5095 			if (tteflags.tte_intlo & TTE_HWWR_INT) {
5096 				/*
5097 				 * need to sync if we are clearing modify bit.
5098 				 */
5099 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5100 			}
5101 
5102 			if (pp && PP_ISRO(pp)) {
5103 				if (tteattr.tte_intlo & TTE_WRPRM_INT) {
5104 					pmtx = sfmmu_page_enter(pp);
5105 					PP_CLRRO(pp);
5106 					sfmmu_page_exit(pmtx);
5107 				}
5108 			}
5109 
5110 			if (ret > 0 && use_demap_range) {
5111 				DEMAP_RANGE_MARKPG(dmrp, addr);
5112 			} else if (ret > 0) {
5113 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5114 			}
5115 
5116 			if (pml) {
5117 				sfmmu_mlist_exit(pml);
5118 			}
5119 		}
5120 next_addr:
5121 		addr += TTEBYTES(ttesz);
5122 		sfhmep++;
5123 		DEMAP_RANGE_NEXTPG(dmrp);
5124 	}
5125 	return (addr);
5126 }
5127 
5128 /*
5129  * This routine converts virtual attributes to physical ones.  It will
5130  * update the tteflags field with the tte mask corresponding to the attributes
5131  * affected and it returns the new attributes.  It will also clear the modify
5132  * bit if we are taking away write permission.  This is necessary since the
5133  * modify bit is the hardware permission bit and we need to clear it in order
5134  * to detect write faults.
5135  */
5136 static uint64_t
5137 sfmmu_vtop_attr(uint_t attr, int mode, tte_t *ttemaskp)
5138 {
5139 	tte_t ttevalue;
5140 
5141 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
5142 
5143 	switch (mode) {
5144 	case SFMMU_CHGATTR:
5145 		/* all attributes specified */
5146 		ttevalue.tte_inthi = MAKE_TTEATTR_INTHI(attr);
5147 		ttevalue.tte_intlo = MAKE_TTEATTR_INTLO(attr);
5148 		ttemaskp->tte_inthi = TTEINTHI_ATTR;
5149 		ttemaskp->tte_intlo = TTEINTLO_ATTR;
5150 		break;
5151 	case SFMMU_SETATTR:
5152 		ASSERT(!(attr & ~HAT_PROT_MASK));
5153 		ttemaskp->ll = 0;
5154 		ttevalue.ll = 0;
5155 		/*
5156 		 * a valid tte implies exec and read for sfmmu
5157 		 * so no need to do anything about them.
5158 		 * since priviledged access implies user access
5159 		 * PROT_USER doesn't make sense either.
5160 		 */
5161 		if (attr & PROT_WRITE) {
5162 			ttemaskp->tte_intlo |= TTE_WRPRM_INT;
5163 			ttevalue.tte_intlo |= TTE_WRPRM_INT;
5164 		}
5165 		break;
5166 	case SFMMU_CLRATTR:
5167 		/* attributes will be nand with current ones */
5168 		if (attr & ~(PROT_WRITE | PROT_USER)) {
5169 			panic("sfmmu: attr %x not supported", attr);
5170 		}
5171 		ttemaskp->ll = 0;
5172 		ttevalue.ll = 0;
5173 		if (attr & PROT_WRITE) {
5174 			/* clear both writable and modify bit */
5175 			ttemaskp->tte_intlo |= TTE_WRPRM_INT | TTE_HWWR_INT;
5176 		}
5177 		if (attr & PROT_USER) {
5178 			ttemaskp->tte_intlo |= TTE_PRIV_INT;
5179 			ttevalue.tte_intlo |= TTE_PRIV_INT;
5180 		}
5181 		break;
5182 	default:
5183 		panic("sfmmu_vtop_attr: bad mode %x", mode);
5184 	}
5185 	ASSERT(TTE_TO_TTEPFN(&ttevalue) == 0);
5186 	return (ttevalue.ll);
5187 }
5188 
5189 static uint_t
5190 sfmmu_ptov_attr(tte_t *ttep)
5191 {
5192 	uint_t attr;
5193 
5194 	ASSERT(TTE_IS_VALID(ttep));
5195 
5196 	attr = PROT_READ;
5197 
5198 	if (TTE_IS_WRITABLE(ttep)) {
5199 		attr |= PROT_WRITE;
5200 	}
5201 	if (TTE_IS_EXECUTABLE(ttep)) {
5202 		attr |= PROT_EXEC;
5203 	}
5204 	if (!TTE_IS_PRIVILEGED(ttep)) {
5205 		attr |= PROT_USER;
5206 	}
5207 	if (TTE_IS_NFO(ttep)) {
5208 		attr |= HAT_NOFAULT;
5209 	}
5210 	if (TTE_IS_NOSYNC(ttep)) {
5211 		attr |= HAT_NOSYNC;
5212 	}
5213 	if (TTE_IS_SIDEFFECT(ttep)) {
5214 		attr |= SFMMU_SIDEFFECT;
5215 	}
5216 	if (!TTE_IS_VCACHEABLE(ttep)) {
5217 		attr |= SFMMU_UNCACHEVTTE;
5218 	}
5219 	if (!TTE_IS_PCACHEABLE(ttep)) {
5220 		attr |= SFMMU_UNCACHEPTTE;
5221 	}
5222 	return (attr);
5223 }
5224 
5225 /*
5226  * hat_chgprot is a deprecated hat call.  New segment drivers
5227  * should store all attributes and use hat_*attr calls.
5228  *
5229  * Change the protections in the virtual address range
5230  * given to the specified virtual protection.  If vprot is ~PROT_WRITE,
5231  * then remove write permission, leaving the other
5232  * permissions unchanged.  If vprot is ~PROT_USER, remove user permissions.
5233  *
5234  */
5235 void
5236 hat_chgprot(struct hat *sfmmup, caddr_t addr, size_t len, uint_t vprot)
5237 {
5238 	struct hmehash_bucket *hmebp;
5239 	hmeblk_tag hblktag;
5240 	int hmeshift, hashno = 1;
5241 	struct hme_blk *hmeblkp, *list = NULL;
5242 	caddr_t endaddr;
5243 	cpuset_t cpuset;
5244 	demap_range_t dmr;
5245 
5246 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5247 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
5248 
5249 	if (sfmmup->sfmmu_xhat_provider) {
5250 		XHAT_CHGPROT(sfmmup, addr, len, vprot);
5251 		return;
5252 	} else {
5253 		/*
5254 		 * This must be a CPU HAT. If the address space has
5255 		 * XHATs attached, change attributes for all of them,
5256 		 * just in case
5257 		 */
5258 		ASSERT(sfmmup->sfmmu_as != NULL);
5259 		if (sfmmup->sfmmu_as->a_xhat != NULL)
5260 			xhat_chgprot_all(sfmmup->sfmmu_as, addr, len, vprot);
5261 	}
5262 
5263 	CPUSET_ZERO(cpuset);
5264 
5265 	if ((vprot != (uint_t)~PROT_WRITE) && (vprot & PROT_USER) &&
5266 	    ((addr + len) > (caddr_t)USERLIMIT)) {
5267 		panic("user addr %p vprot %x in kernel space",
5268 		    (void *)addr, vprot);
5269 	}
5270 	endaddr = addr + len;
5271 	hblktag.htag_id = sfmmup;
5272 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5273 	DEMAP_RANGE_INIT(sfmmup, &dmr);
5274 
5275 	while (addr < endaddr) {
5276 		hmeshift = HME_HASH_SHIFT(hashno);
5277 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5278 		hblktag.htag_rehash = hashno;
5279 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5280 
5281 		SFMMU_HASH_LOCK(hmebp);
5282 
5283 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
5284 		if (hmeblkp != NULL) {
5285 			ASSERT(!hmeblkp->hblk_shared);
5286 			/*
5287 			 * We've encountered a shadow hmeblk so skip the range
5288 			 * of the next smaller mapping size.
5289 			 */
5290 			if (hmeblkp->hblk_shw_bit) {
5291 				ASSERT(sfmmup != ksfmmup);
5292 				ASSERT(hashno > 1);
5293 				addr = (caddr_t)P2END((uintptr_t)addr,
5294 				    TTEBYTES(hashno - 1));
5295 			} else {
5296 				addr = sfmmu_hblk_chgprot(sfmmup, hmeblkp,
5297 				    addr, endaddr, &dmr, vprot);
5298 			}
5299 			SFMMU_HASH_UNLOCK(hmebp);
5300 			hashno = 1;
5301 			continue;
5302 		}
5303 		SFMMU_HASH_UNLOCK(hmebp);
5304 
5305 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
5306 			/*
5307 			 * We have traversed the whole list and rehashed
5308 			 * if necessary without finding the address to chgprot.
5309 			 * This is ok so we increment the address by the
5310 			 * smallest hmeblk range for kernel mappings and the
5311 			 * largest hmeblk range, to account for shadow hmeblks,
5312 			 * for user mappings and continue.
5313 			 */
5314 			if (sfmmup == ksfmmup)
5315 				addr = (caddr_t)P2END((uintptr_t)addr,
5316 				    TTEBYTES(1));
5317 			else
5318 				addr = (caddr_t)P2END((uintptr_t)addr,
5319 				    TTEBYTES(hashno));
5320 			hashno = 1;
5321 		} else {
5322 			hashno++;
5323 		}
5324 	}
5325 
5326 	sfmmu_hblks_list_purge(&list, 0);
5327 	DEMAP_RANGE_FLUSH(&dmr);
5328 	cpuset = sfmmup->sfmmu_cpusran;
5329 	xt_sync(cpuset);
5330 }
5331 
5332 /*
5333  * This function chgprots a range of addresses in an hmeblk.  It returns the
5334  * next addres that needs to be chgprot.
5335  * It should be called with the hash lock held.
5336  * XXX It shold be possible to optimize chgprot by not flushing every time but
5337  * on the other hand:
5338  * 1. do one flush crosscall.
5339  * 2. only flush if we are increasing permissions (make sure this will work)
5340  */
5341 static caddr_t
5342 sfmmu_hblk_chgprot(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5343 	caddr_t endaddr, demap_range_t *dmrp, uint_t vprot)
5344 {
5345 	uint_t pprot;
5346 	tte_t tte, ttemod;
5347 	struct sf_hment *sfhmep;
5348 	uint_t tteflags;
5349 	int ttesz;
5350 	struct page *pp = NULL;
5351 	kmutex_t *pml, *pmtx;
5352 	int ret;
5353 	int use_demap_range;
5354 #if defined(SF_ERRATA_57)
5355 	int check_exec;
5356 #endif
5357 
5358 	ASSERT(in_hblk_range(hmeblkp, addr));
5359 	ASSERT(hmeblkp->hblk_shw_bit == 0);
5360 	ASSERT(!hmeblkp->hblk_shared);
5361 
5362 #ifdef DEBUG
5363 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5364 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5365 		panic("sfmmu_hblk_chgprot: partial chgprot of large page");
5366 	}
5367 #endif /* DEBUG */
5368 
5369 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5370 	ttesz = get_hblk_ttesz(hmeblkp);
5371 
5372 	pprot = sfmmu_vtop_prot(vprot, &tteflags);
5373 #if defined(SF_ERRATA_57)
5374 	check_exec = (sfmmup != ksfmmup) &&
5375 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
5376 	    ((vprot & PROT_EXEC) == PROT_EXEC);
5377 #endif
5378 	HBLKTOHME(sfhmep, hmeblkp, addr);
5379 
5380 	/*
5381 	 * Flush the current demap region if addresses have been
5382 	 * skipped or the page size doesn't match.
5383 	 */
5384 	use_demap_range = (TTEBYTES(ttesz) == MMU_PAGESIZE);
5385 	if (use_demap_range) {
5386 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5387 	} else if (dmrp != NULL) {
5388 		DEMAP_RANGE_FLUSH(dmrp);
5389 	}
5390 
5391 	while (addr < endaddr) {
5392 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5393 		if (TTE_IS_VALID(&tte)) {
5394 			if (TTE_GET_LOFLAGS(&tte, tteflags) == pprot) {
5395 				/*
5396 				 * if the new protection is the same as old
5397 				 * continue
5398 				 */
5399 				goto next_addr;
5400 			}
5401 			pml = NULL;
5402 			pp = sfhmep->hme_page;
5403 			if (pp) {
5404 				pml = sfmmu_mlist_enter(pp);
5405 			}
5406 			if (pp != sfhmep->hme_page) {
5407 				/*
5408 				 * tte most have been unloaded
5409 				 * underneath us.  Recheck
5410 				 */
5411 				ASSERT(pml);
5412 				sfmmu_mlist_exit(pml);
5413 				continue;
5414 			}
5415 
5416 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5417 
5418 			ttemod = tte;
5419 			TTE_SET_LOFLAGS(&ttemod, tteflags, pprot);
5420 #if defined(SF_ERRATA_57)
5421 			if (check_exec && addr < errata57_limit)
5422 				ttemod.tte_exec_perm = 0;
5423 #endif
5424 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5425 			    &sfhmep->hme_tte);
5426 
5427 			if (ret < 0) {
5428 				/* tte changed underneath us */
5429 				if (pml) {
5430 					sfmmu_mlist_exit(pml);
5431 				}
5432 				continue;
5433 			}
5434 
5435 			if (tteflags & TTE_HWWR_INT) {
5436 				/*
5437 				 * need to sync if we are clearing modify bit.
5438 				 */
5439 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5440 			}
5441 
5442 			if (pp && PP_ISRO(pp)) {
5443 				if (pprot & TTE_WRPRM_INT) {
5444 					pmtx = sfmmu_page_enter(pp);
5445 					PP_CLRRO(pp);
5446 					sfmmu_page_exit(pmtx);
5447 				}
5448 			}
5449 
5450 			if (ret > 0 && use_demap_range) {
5451 				DEMAP_RANGE_MARKPG(dmrp, addr);
5452 			} else if (ret > 0) {
5453 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5454 			}
5455 
5456 			if (pml) {
5457 				sfmmu_mlist_exit(pml);
5458 			}
5459 		}
5460 next_addr:
5461 		addr += TTEBYTES(ttesz);
5462 		sfhmep++;
5463 		DEMAP_RANGE_NEXTPG(dmrp);
5464 	}
5465 	return (addr);
5466 }
5467 
5468 /*
5469  * This routine is deprecated and should only be used by hat_chgprot.
5470  * The correct routine is sfmmu_vtop_attr.
5471  * This routine converts virtual page protections to physical ones.  It will
5472  * update the tteflags field with the tte mask corresponding to the protections
5473  * affected and it returns the new protections.  It will also clear the modify
5474  * bit if we are taking away write permission.  This is necessary since the
5475  * modify bit is the hardware permission bit and we need to clear it in order
5476  * to detect write faults.
5477  * It accepts the following special protections:
5478  * ~PROT_WRITE = remove write permissions.
5479  * ~PROT_USER = remove user permissions.
5480  */
5481 static uint_t
5482 sfmmu_vtop_prot(uint_t vprot, uint_t *tteflagsp)
5483 {
5484 	if (vprot == (uint_t)~PROT_WRITE) {
5485 		*tteflagsp = TTE_WRPRM_INT | TTE_HWWR_INT;
5486 		return (0);		/* will cause wrprm to be cleared */
5487 	}
5488 	if (vprot == (uint_t)~PROT_USER) {
5489 		*tteflagsp = TTE_PRIV_INT;
5490 		return (0);		/* will cause privprm to be cleared */
5491 	}
5492 	if ((vprot == 0) || (vprot == PROT_USER) ||
5493 	    ((vprot & PROT_ALL) != vprot)) {
5494 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5495 	}
5496 
5497 	switch (vprot) {
5498 	case (PROT_READ):
5499 	case (PROT_EXEC):
5500 	case (PROT_EXEC | PROT_READ):
5501 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5502 		return (TTE_PRIV_INT); 		/* set prv and clr wrt */
5503 	case (PROT_WRITE):
5504 	case (PROT_WRITE | PROT_READ):
5505 	case (PROT_EXEC | PROT_WRITE):
5506 	case (PROT_EXEC | PROT_WRITE | PROT_READ):
5507 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5508 		return (TTE_PRIV_INT | TTE_WRPRM_INT); 	/* set prv and wrt */
5509 	case (PROT_USER | PROT_READ):
5510 	case (PROT_USER | PROT_EXEC):
5511 	case (PROT_USER | PROT_EXEC | PROT_READ):
5512 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5513 		return (0); 			/* clr prv and wrt */
5514 	case (PROT_USER | PROT_WRITE):
5515 	case (PROT_USER | PROT_WRITE | PROT_READ):
5516 	case (PROT_USER | PROT_EXEC | PROT_WRITE):
5517 	case (PROT_USER | PROT_EXEC | PROT_WRITE | PROT_READ):
5518 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5519 		return (TTE_WRPRM_INT); 	/* clr prv and set wrt */
5520 	default:
5521 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5522 	}
5523 	return (0);
5524 }
5525 
5526 /*
5527  * Alternate unload for very large virtual ranges. With a true 64 bit VA,
5528  * the normal algorithm would take too long for a very large VA range with
5529  * few real mappings. This routine just walks thru all HMEs in the global
5530  * hash table to find and remove mappings.
5531  */
5532 static void
5533 hat_unload_large_virtual(
5534 	struct hat		*sfmmup,
5535 	caddr_t			startaddr,
5536 	size_t			len,
5537 	uint_t			flags,
5538 	hat_callback_t		*callback)
5539 {
5540 	struct hmehash_bucket *hmebp;
5541 	struct hme_blk *hmeblkp;
5542 	struct hme_blk *pr_hblk = NULL;
5543 	struct hme_blk *nx_hblk;
5544 	struct hme_blk *list = NULL;
5545 	int i;
5546 	demap_range_t dmr, *dmrp;
5547 	cpuset_t cpuset;
5548 	caddr_t	endaddr = startaddr + len;
5549 	caddr_t	sa;
5550 	caddr_t	ea;
5551 	caddr_t	cb_sa[MAX_CB_ADDR];
5552 	caddr_t	cb_ea[MAX_CB_ADDR];
5553 	int	addr_cnt = 0;
5554 	int	a = 0;
5555 
5556 	if (sfmmup->sfmmu_free) {
5557 		dmrp = NULL;
5558 	} else {
5559 		dmrp = &dmr;
5560 		DEMAP_RANGE_INIT(sfmmup, dmrp);
5561 	}
5562 
5563 	/*
5564 	 * Loop through all the hash buckets of HME blocks looking for matches.
5565 	 */
5566 	for (i = 0; i <= UHMEHASH_SZ; i++) {
5567 		hmebp = &uhme_hash[i];
5568 		SFMMU_HASH_LOCK(hmebp);
5569 		hmeblkp = hmebp->hmeblkp;
5570 		pr_hblk = NULL;
5571 		while (hmeblkp) {
5572 			nx_hblk = hmeblkp->hblk_next;
5573 
5574 			/*
5575 			 * skip if not this context, if a shadow block or
5576 			 * if the mapping is not in the requested range
5577 			 */
5578 			if (hmeblkp->hblk_tag.htag_id != sfmmup ||
5579 			    hmeblkp->hblk_shw_bit ||
5580 			    (sa = (caddr_t)get_hblk_base(hmeblkp)) >= endaddr ||
5581 			    (ea = get_hblk_endaddr(hmeblkp)) <= startaddr) {
5582 				pr_hblk = hmeblkp;
5583 				goto next_block;
5584 			}
5585 
5586 			ASSERT(!hmeblkp->hblk_shared);
5587 			/*
5588 			 * unload if there are any current valid mappings
5589 			 */
5590 			if (hmeblkp->hblk_vcnt != 0 ||
5591 			    hmeblkp->hblk_hmecnt != 0)
5592 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
5593 				    sa, ea, dmrp, flags);
5594 
5595 			/*
5596 			 * on unmap we also release the HME block itself, once
5597 			 * all mappings are gone.
5598 			 */
5599 			if ((flags & HAT_UNLOAD_UNMAP) != 0 &&
5600 			    !hmeblkp->hblk_vcnt &&
5601 			    !hmeblkp->hblk_hmecnt) {
5602 				ASSERT(!hmeblkp->hblk_lckcnt);
5603 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
5604 				    &list, 0);
5605 			} else {
5606 				pr_hblk = hmeblkp;
5607 			}
5608 
5609 			if (callback == NULL)
5610 				goto next_block;
5611 
5612 			/*
5613 			 * HME blocks may span more than one page, but we may be
5614 			 * unmapping only one page, so check for a smaller range
5615 			 * for the callback
5616 			 */
5617 			if (sa < startaddr)
5618 				sa = startaddr;
5619 			if (--ea > endaddr)
5620 				ea = endaddr - 1;
5621 
5622 			cb_sa[addr_cnt] = sa;
5623 			cb_ea[addr_cnt] = ea;
5624 			if (++addr_cnt == MAX_CB_ADDR) {
5625 				if (dmrp != NULL) {
5626 					DEMAP_RANGE_FLUSH(dmrp);
5627 					cpuset = sfmmup->sfmmu_cpusran;
5628 					xt_sync(cpuset);
5629 				}
5630 
5631 				for (a = 0; a < MAX_CB_ADDR; ++a) {
5632 					callback->hcb_start_addr = cb_sa[a];
5633 					callback->hcb_end_addr = cb_ea[a];
5634 					callback->hcb_function(callback);
5635 				}
5636 				addr_cnt = 0;
5637 			}
5638 
5639 next_block:
5640 			hmeblkp = nx_hblk;
5641 		}
5642 		SFMMU_HASH_UNLOCK(hmebp);
5643 	}
5644 
5645 	sfmmu_hblks_list_purge(&list, 0);
5646 	if (dmrp != NULL) {
5647 		DEMAP_RANGE_FLUSH(dmrp);
5648 		cpuset = sfmmup->sfmmu_cpusran;
5649 		xt_sync(cpuset);
5650 	}
5651 
5652 	for (a = 0; a < addr_cnt; ++a) {
5653 		callback->hcb_start_addr = cb_sa[a];
5654 		callback->hcb_end_addr = cb_ea[a];
5655 		callback->hcb_function(callback);
5656 	}
5657 
5658 	/*
5659 	 * Check TSB and TLB page sizes if the process isn't exiting.
5660 	 */
5661 	if (!sfmmup->sfmmu_free)
5662 		sfmmu_check_page_sizes(sfmmup, 0);
5663 }
5664 
5665 /*
5666  * Unload all the mappings in the range [addr..addr+len). addr and len must
5667  * be MMU_PAGESIZE aligned.
5668  */
5669 
5670 extern struct seg *segkmap;
5671 #define	ISSEGKMAP(sfmmup, addr) (sfmmup == ksfmmup && \
5672 segkmap->s_base <= (addr) && (addr) < (segkmap->s_base + segkmap->s_size))
5673 
5674 
5675 void
5676 hat_unload_callback(
5677 	struct hat *sfmmup,
5678 	caddr_t addr,
5679 	size_t len,
5680 	uint_t flags,
5681 	hat_callback_t *callback)
5682 {
5683 	struct hmehash_bucket *hmebp;
5684 	hmeblk_tag hblktag;
5685 	int hmeshift, hashno, iskernel;
5686 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
5687 	caddr_t endaddr;
5688 	cpuset_t cpuset;
5689 	int addr_count = 0;
5690 	int a;
5691 	caddr_t cb_start_addr[MAX_CB_ADDR];
5692 	caddr_t cb_end_addr[MAX_CB_ADDR];
5693 	int issegkmap = ISSEGKMAP(sfmmup, addr);
5694 	demap_range_t dmr, *dmrp;
5695 
5696 	if (sfmmup->sfmmu_xhat_provider) {
5697 		XHAT_UNLOAD_CALLBACK(sfmmup, addr, len, flags, callback);
5698 		return;
5699 	} else {
5700 		/*
5701 		 * This must be a CPU HAT. If the address space has
5702 		 * XHATs attached, unload the mappings for all of them,
5703 		 * just in case
5704 		 */
5705 		ASSERT(sfmmup->sfmmu_as != NULL);
5706 		if (sfmmup->sfmmu_as->a_xhat != NULL)
5707 			xhat_unload_callback_all(sfmmup->sfmmu_as, addr,
5708 			    len, flags, callback);
5709 	}
5710 
5711 	ASSERT((sfmmup == ksfmmup) || (flags & HAT_UNLOAD_OTHER) || \
5712 	    AS_LOCK_HELD(sfmmup->sfmmu_as));
5713 
5714 	ASSERT(sfmmup != NULL);
5715 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5716 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
5717 
5718 	/*
5719 	 * Probing through a large VA range (say 63 bits) will be slow, even
5720 	 * at 4 Meg steps between the probes. So, when the virtual address range
5721 	 * is very large, search the HME entries for what to unload.
5722 	 *
5723 	 *	len >> TTE_PAGE_SHIFT(TTE4M) is the # of 4Meg probes we'd need
5724 	 *
5725 	 *	UHMEHASH_SZ is number of hash buckets to examine
5726 	 *
5727 	 */
5728 	if (sfmmup != KHATID && (len >> TTE_PAGE_SHIFT(TTE4M)) > UHMEHASH_SZ) {
5729 		hat_unload_large_virtual(sfmmup, addr, len, flags, callback);
5730 		return;
5731 	}
5732 
5733 	CPUSET_ZERO(cpuset);
5734 
5735 	/*
5736 	 * If the process is exiting, we can save a lot of fuss since
5737 	 * we'll flush the TLB when we free the ctx anyway.
5738 	 */
5739 	if (sfmmup->sfmmu_free) {
5740 		dmrp = NULL;
5741 	} else {
5742 		dmrp = &dmr;
5743 		DEMAP_RANGE_INIT(sfmmup, dmrp);
5744 	}
5745 
5746 	endaddr = addr + len;
5747 	hblktag.htag_id = sfmmup;
5748 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5749 
5750 	/*
5751 	 * It is likely for the vm to call unload over a wide range of
5752 	 * addresses that are actually very sparsely populated by
5753 	 * translations.  In order to speed this up the sfmmu hat supports
5754 	 * the concept of shadow hmeblks. Dummy large page hmeblks that
5755 	 * correspond to actual small translations are allocated at tteload
5756 	 * time and are referred to as shadow hmeblks.  Now, during unload
5757 	 * time, we first check if we have a shadow hmeblk for that
5758 	 * translation.  The absence of one means the corresponding address
5759 	 * range is empty and can be skipped.
5760 	 *
5761 	 * The kernel is an exception to above statement and that is why
5762 	 * we don't use shadow hmeblks and hash starting from the smallest
5763 	 * page size.
5764 	 */
5765 	if (sfmmup == KHATID) {
5766 		iskernel = 1;
5767 		hashno = TTE64K;
5768 	} else {
5769 		iskernel = 0;
5770 		if (mmu_page_sizes == max_mmu_page_sizes) {
5771 			hashno = TTE256M;
5772 		} else {
5773 			hashno = TTE4M;
5774 		}
5775 	}
5776 	while (addr < endaddr) {
5777 		hmeshift = HME_HASH_SHIFT(hashno);
5778 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5779 		hblktag.htag_rehash = hashno;
5780 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5781 
5782 		SFMMU_HASH_LOCK(hmebp);
5783 
5784 		HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
5785 		if (hmeblkp == NULL) {
5786 			/*
5787 			 * didn't find an hmeblk. skip the appropiate
5788 			 * address range.
5789 			 */
5790 			SFMMU_HASH_UNLOCK(hmebp);
5791 			if (iskernel) {
5792 				if (hashno < mmu_hashcnt) {
5793 					hashno++;
5794 					continue;
5795 				} else {
5796 					hashno = TTE64K;
5797 					addr = (caddr_t)roundup((uintptr_t)addr
5798 					    + 1, MMU_PAGESIZE64K);
5799 					continue;
5800 				}
5801 			}
5802 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5803 			    (1 << hmeshift));
5804 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5805 				ASSERT(hashno == TTE64K);
5806 				continue;
5807 			}
5808 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5809 				hashno = TTE512K;
5810 				continue;
5811 			}
5812 			if (mmu_page_sizes == max_mmu_page_sizes) {
5813 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5814 					hashno = TTE4M;
5815 					continue;
5816 				}
5817 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5818 					hashno = TTE32M;
5819 					continue;
5820 				}
5821 				hashno = TTE256M;
5822 				continue;
5823 			} else {
5824 				hashno = TTE4M;
5825 				continue;
5826 			}
5827 		}
5828 		ASSERT(hmeblkp);
5829 		ASSERT(!hmeblkp->hblk_shared);
5830 		if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5831 			/*
5832 			 * If the valid count is zero we can skip the range
5833 			 * mapped by this hmeblk.
5834 			 * We free hblks in the case of HAT_UNMAP.  HAT_UNMAP
5835 			 * is used by segment drivers as a hint
5836 			 * that the mapping resource won't be used any longer.
5837 			 * The best example of this is during exit().
5838 			 */
5839 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5840 			    get_hblk_span(hmeblkp));
5841 			if ((flags & HAT_UNLOAD_UNMAP) ||
5842 			    (iskernel && !issegkmap)) {
5843 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
5844 				    &list, 0);
5845 			}
5846 			SFMMU_HASH_UNLOCK(hmebp);
5847 
5848 			if (iskernel) {
5849 				hashno = TTE64K;
5850 				continue;
5851 			}
5852 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5853 				ASSERT(hashno == TTE64K);
5854 				continue;
5855 			}
5856 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5857 				hashno = TTE512K;
5858 				continue;
5859 			}
5860 			if (mmu_page_sizes == max_mmu_page_sizes) {
5861 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5862 					hashno = TTE4M;
5863 					continue;
5864 				}
5865 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5866 					hashno = TTE32M;
5867 					continue;
5868 				}
5869 				hashno = TTE256M;
5870 				continue;
5871 			} else {
5872 				hashno = TTE4M;
5873 				continue;
5874 			}
5875 		}
5876 		if (hmeblkp->hblk_shw_bit) {
5877 			/*
5878 			 * If we encounter a shadow hmeblk we know there is
5879 			 * smaller sized hmeblks mapping the same address space.
5880 			 * Decrement the hash size and rehash.
5881 			 */
5882 			ASSERT(sfmmup != KHATID);
5883 			hashno--;
5884 			SFMMU_HASH_UNLOCK(hmebp);
5885 			continue;
5886 		}
5887 
5888 		/*
5889 		 * track callback address ranges.
5890 		 * only start a new range when it's not contiguous
5891 		 */
5892 		if (callback != NULL) {
5893 			if (addr_count > 0 &&
5894 			    addr == cb_end_addr[addr_count - 1])
5895 				--addr_count;
5896 			else
5897 				cb_start_addr[addr_count] = addr;
5898 		}
5899 
5900 		addr = sfmmu_hblk_unload(sfmmup, hmeblkp, addr, endaddr,
5901 		    dmrp, flags);
5902 
5903 		if (callback != NULL)
5904 			cb_end_addr[addr_count++] = addr;
5905 
5906 		if (((flags & HAT_UNLOAD_UNMAP) || (iskernel && !issegkmap)) &&
5907 		    !hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5908 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 0);
5909 		}
5910 		SFMMU_HASH_UNLOCK(hmebp);
5911 
5912 		/*
5913 		 * Notify our caller as to exactly which pages
5914 		 * have been unloaded. We do these in clumps,
5915 		 * to minimize the number of xt_sync()s that need to occur.
5916 		 */
5917 		if (callback != NULL && addr_count == MAX_CB_ADDR) {
5918 			if (dmrp != NULL) {
5919 				DEMAP_RANGE_FLUSH(dmrp);
5920 				cpuset = sfmmup->sfmmu_cpusran;
5921 				xt_sync(cpuset);
5922 			}
5923 
5924 			for (a = 0; a < MAX_CB_ADDR; ++a) {
5925 				callback->hcb_start_addr = cb_start_addr[a];
5926 				callback->hcb_end_addr = cb_end_addr[a];
5927 				callback->hcb_function(callback);
5928 			}
5929 			addr_count = 0;
5930 		}
5931 		if (iskernel) {
5932 			hashno = TTE64K;
5933 			continue;
5934 		}
5935 		if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5936 			ASSERT(hashno == TTE64K);
5937 			continue;
5938 		}
5939 		if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5940 			hashno = TTE512K;
5941 			continue;
5942 		}
5943 		if (mmu_page_sizes == max_mmu_page_sizes) {
5944 			if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5945 				hashno = TTE4M;
5946 				continue;
5947 			}
5948 			if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5949 				hashno = TTE32M;
5950 				continue;
5951 			}
5952 			hashno = TTE256M;
5953 		} else {
5954 			hashno = TTE4M;
5955 		}
5956 	}
5957 
5958 	sfmmu_hblks_list_purge(&list, 0);
5959 	if (dmrp != NULL) {
5960 		DEMAP_RANGE_FLUSH(dmrp);
5961 		cpuset = sfmmup->sfmmu_cpusran;
5962 		xt_sync(cpuset);
5963 	}
5964 	if (callback && addr_count != 0) {
5965 		for (a = 0; a < addr_count; ++a) {
5966 			callback->hcb_start_addr = cb_start_addr[a];
5967 			callback->hcb_end_addr = cb_end_addr[a];
5968 			callback->hcb_function(callback);
5969 		}
5970 	}
5971 
5972 	/*
5973 	 * Check TSB and TLB page sizes if the process isn't exiting.
5974 	 */
5975 	if (!sfmmup->sfmmu_free)
5976 		sfmmu_check_page_sizes(sfmmup, 0);
5977 }
5978 
5979 /*
5980  * Unload all the mappings in the range [addr..addr+len). addr and len must
5981  * be MMU_PAGESIZE aligned.
5982  */
5983 void
5984 hat_unload(struct hat *sfmmup, caddr_t addr, size_t len, uint_t flags)
5985 {
5986 	if (sfmmup->sfmmu_xhat_provider) {
5987 		XHAT_UNLOAD(sfmmup, addr, len, flags);
5988 		return;
5989 	}
5990 	hat_unload_callback(sfmmup, addr, len, flags, NULL);
5991 }
5992 
5993 
5994 /*
5995  * Find the largest mapping size for this page.
5996  */
5997 int
5998 fnd_mapping_sz(page_t *pp)
5999 {
6000 	int sz;
6001 	int p_index;
6002 
6003 	p_index = PP_MAPINDEX(pp);
6004 
6005 	sz = 0;
6006 	p_index >>= 1;	/* don't care about 8K bit */
6007 	for (; p_index; p_index >>= 1) {
6008 		sz++;
6009 	}
6010 
6011 	return (sz);
6012 }
6013 
6014 /*
6015  * This function unloads a range of addresses for an hmeblk.
6016  * It returns the next address to be unloaded.
6017  * It should be called with the hash lock held.
6018  */
6019 static caddr_t
6020 sfmmu_hblk_unload(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
6021 	caddr_t endaddr, demap_range_t *dmrp, uint_t flags)
6022 {
6023 	tte_t	tte, ttemod;
6024 	struct	sf_hment *sfhmep;
6025 	int	ttesz;
6026 	long	ttecnt;
6027 	page_t *pp;
6028 	kmutex_t *pml;
6029 	int ret;
6030 	int use_demap_range;
6031 
6032 	ASSERT(in_hblk_range(hmeblkp, addr));
6033 	ASSERT(!hmeblkp->hblk_shw_bit);
6034 	ASSERT(sfmmup != NULL || hmeblkp->hblk_shared);
6035 	ASSERT(sfmmup == NULL || !hmeblkp->hblk_shared);
6036 	ASSERT(dmrp == NULL || !hmeblkp->hblk_shared);
6037 
6038 #ifdef DEBUG
6039 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
6040 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
6041 		panic("sfmmu_hblk_unload: partial unload of large page");
6042 	}
6043 #endif /* DEBUG */
6044 
6045 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
6046 	ttesz = get_hblk_ttesz(hmeblkp);
6047 
6048 	use_demap_range = ((dmrp == NULL) ||
6049 	    (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp)));
6050 
6051 	if (use_demap_range) {
6052 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
6053 	} else if (dmrp != NULL) {
6054 		DEMAP_RANGE_FLUSH(dmrp);
6055 	}
6056 	ttecnt = 0;
6057 	HBLKTOHME(sfhmep, hmeblkp, addr);
6058 
6059 	while (addr < endaddr) {
6060 		pml = NULL;
6061 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6062 		if (TTE_IS_VALID(&tte)) {
6063 			pp = sfhmep->hme_page;
6064 			if (pp != NULL) {
6065 				pml = sfmmu_mlist_enter(pp);
6066 			}
6067 
6068 			/*
6069 			 * Verify if hme still points to 'pp' now that
6070 			 * we have p_mapping lock.
6071 			 */
6072 			if (sfhmep->hme_page != pp) {
6073 				if (pp != NULL && sfhmep->hme_page != NULL) {
6074 					ASSERT(pml != NULL);
6075 					sfmmu_mlist_exit(pml);
6076 					/* Re-start this iteration. */
6077 					continue;
6078 				}
6079 				ASSERT((pp != NULL) &&
6080 				    (sfhmep->hme_page == NULL));
6081 				goto tte_unloaded;
6082 			}
6083 
6084 			/*
6085 			 * This point on we have both HASH and p_mapping
6086 			 * lock.
6087 			 */
6088 			ASSERT(pp == sfhmep->hme_page);
6089 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
6090 
6091 			/*
6092 			 * We need to loop on modify tte because it is
6093 			 * possible for pagesync to come along and
6094 			 * change the software bits beneath us.
6095 			 *
6096 			 * Page_unload can also invalidate the tte after
6097 			 * we read tte outside of p_mapping lock.
6098 			 */
6099 again:
6100 			ttemod = tte;
6101 
6102 			TTE_SET_INVALID(&ttemod);
6103 			ret = sfmmu_modifytte_try(&tte, &ttemod,
6104 			    &sfhmep->hme_tte);
6105 
6106 			if (ret <= 0) {
6107 				if (TTE_IS_VALID(&tte)) {
6108 					ASSERT(ret < 0);
6109 					goto again;
6110 				}
6111 				if (pp != NULL) {
6112 					panic("sfmmu_hblk_unload: pp = 0x%p "
6113 					    "tte became invalid under mlist"
6114 					    " lock = 0x%p", (void *)pp,
6115 					    (void *)pml);
6116 				}
6117 				continue;
6118 			}
6119 
6120 			if (!(flags & HAT_UNLOAD_NOSYNC)) {
6121 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
6122 			}
6123 
6124 			/*
6125 			 * Ok- we invalidated the tte. Do the rest of the job.
6126 			 */
6127 			ttecnt++;
6128 
6129 			if (flags & HAT_UNLOAD_UNLOCK) {
6130 				ASSERT(hmeblkp->hblk_lckcnt > 0);
6131 				atomic_dec_32(&hmeblkp->hblk_lckcnt);
6132 				HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
6133 			}
6134 
6135 			/*
6136 			 * Normally we would need to flush the page
6137 			 * from the virtual cache at this point in
6138 			 * order to prevent a potential cache alias
6139 			 * inconsistency.
6140 			 * The particular scenario we need to worry
6141 			 * about is:
6142 			 * Given:  va1 and va2 are two virtual address
6143 			 * that alias and map the same physical
6144 			 * address.
6145 			 * 1.   mapping exists from va1 to pa and data
6146 			 * has been read into the cache.
6147 			 * 2.   unload va1.
6148 			 * 3.   load va2 and modify data using va2.
6149 			 * 4    unload va2.
6150 			 * 5.   load va1 and reference data.  Unless we
6151 			 * flush the data cache when we unload we will
6152 			 * get stale data.
6153 			 * Fortunately, page coloring eliminates the
6154 			 * above scenario by remembering the color a
6155 			 * physical page was last or is currently
6156 			 * mapped to.  Now, we delay the flush until
6157 			 * the loading of translations.  Only when the
6158 			 * new translation is of a different color
6159 			 * are we forced to flush.
6160 			 */
6161 			if (use_demap_range) {
6162 				/*
6163 				 * Mark this page as needing a demap.
6164 				 */
6165 				DEMAP_RANGE_MARKPG(dmrp, addr);
6166 			} else {
6167 				ASSERT(sfmmup != NULL);
6168 				ASSERT(!hmeblkp->hblk_shared);
6169 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
6170 				    sfmmup->sfmmu_free, 0);
6171 			}
6172 
6173 			if (pp) {
6174 				/*
6175 				 * Remove the hment from the mapping list
6176 				 */
6177 				ASSERT(hmeblkp->hblk_hmecnt > 0);
6178 
6179 				/*
6180 				 * Again, we cannot
6181 				 * ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS);
6182 				 */
6183 				HME_SUB(sfhmep, pp);
6184 				membar_stst();
6185 				atomic_dec_16(&hmeblkp->hblk_hmecnt);
6186 			}
6187 
6188 			ASSERT(hmeblkp->hblk_vcnt > 0);
6189 			atomic_dec_16(&hmeblkp->hblk_vcnt);
6190 
6191 			ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
6192 			    !hmeblkp->hblk_lckcnt);
6193 
6194 #ifdef VAC
6195 			if (pp && (pp->p_nrm & (P_KPMC | P_KPMS | P_TNC))) {
6196 				if (PP_ISTNC(pp)) {
6197 					/*
6198 					 * If page was temporary
6199 					 * uncached, try to recache
6200 					 * it. Note that HME_SUB() was
6201 					 * called above so p_index and
6202 					 * mlist had been updated.
6203 					 */
6204 					conv_tnc(pp, ttesz);
6205 				} else if (pp->p_mapping == NULL) {
6206 					ASSERT(kpm_enable);
6207 					/*
6208 					 * Page is marked to be in VAC conflict
6209 					 * to an existing kpm mapping and/or is
6210 					 * kpm mapped using only the regular
6211 					 * pagesize.
6212 					 */
6213 					sfmmu_kpm_hme_unload(pp);
6214 				}
6215 			}
6216 #endif	/* VAC */
6217 		} else if ((pp = sfhmep->hme_page) != NULL) {
6218 				/*
6219 				 * TTE is invalid but the hme
6220 				 * still exists. let pageunload
6221 				 * complete its job.
6222 				 */
6223 				ASSERT(pml == NULL);
6224 				pml = sfmmu_mlist_enter(pp);
6225 				if (sfhmep->hme_page != NULL) {
6226 					sfmmu_mlist_exit(pml);
6227 					continue;
6228 				}
6229 				ASSERT(sfhmep->hme_page == NULL);
6230 		} else if (hmeblkp->hblk_hmecnt != 0) {
6231 			/*
6232 			 * pageunload may have not finished decrementing
6233 			 * hblk_vcnt and hblk_hmecnt. Find page_t if any and
6234 			 * wait for pageunload to finish. Rely on pageunload
6235 			 * to decrement hblk_hmecnt after hblk_vcnt.
6236 			 */
6237 			pfn_t pfn = TTE_TO_TTEPFN(&tte);
6238 			ASSERT(pml == NULL);
6239 			if (pf_is_memory(pfn)) {
6240 				pp = page_numtopp_nolock(pfn);
6241 				if (pp != NULL) {
6242 					pml = sfmmu_mlist_enter(pp);
6243 					sfmmu_mlist_exit(pml);
6244 					pml = NULL;
6245 				}
6246 			}
6247 		}
6248 
6249 tte_unloaded:
6250 		/*
6251 		 * At this point, the tte we are looking at
6252 		 * should be unloaded, and hme has been unlinked
6253 		 * from page too. This is important because in
6254 		 * pageunload, it does ttesync() then HME_SUB.
6255 		 * We need to make sure HME_SUB has been completed
6256 		 * so we know ttesync() has been completed. Otherwise,
6257 		 * at exit time, after return from hat layer, VM will
6258 		 * release as structure which hat_setstat() (called
6259 		 * by ttesync()) needs.
6260 		 */
6261 #ifdef DEBUG
6262 		{
6263 			tte_t	dtte;
6264 
6265 			ASSERT(sfhmep->hme_page == NULL);
6266 
6267 			sfmmu_copytte(&sfhmep->hme_tte, &dtte);
6268 			ASSERT(!TTE_IS_VALID(&dtte));
6269 		}
6270 #endif
6271 
6272 		if (pml) {
6273 			sfmmu_mlist_exit(pml);
6274 		}
6275 
6276 		addr += TTEBYTES(ttesz);
6277 		sfhmep++;
6278 		DEMAP_RANGE_NEXTPG(dmrp);
6279 	}
6280 	/*
6281 	 * For shared hmeblks this routine is only called when region is freed
6282 	 * and no longer referenced.  So no need to decrement ttecnt
6283 	 * in the region structure here.
6284 	 */
6285 	if (ttecnt > 0 && sfmmup != NULL) {
6286 		atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -ttecnt);
6287 	}
6288 	return (addr);
6289 }
6290 
6291 /*
6292  * Invalidate a virtual address range for the local CPU.
6293  * For best performance ensure that the va range is completely
6294  * mapped, otherwise the entire TLB will be flushed.
6295  */
6296 void
6297 hat_flush_range(struct hat *sfmmup, caddr_t va, size_t size)
6298 {
6299 	ssize_t sz;
6300 	caddr_t endva = va + size;
6301 
6302 	while (va < endva) {
6303 		sz = hat_getpagesize(sfmmup, va);
6304 		if (sz < 0) {
6305 			vtag_flushall();
6306 			break;
6307 		}
6308 		vtag_flushpage(va, (uint64_t)sfmmup);
6309 		va += sz;
6310 	}
6311 }
6312 
6313 /*
6314  * Synchronize all the mappings in the range [addr..addr+len).
6315  * Can be called with clearflag having two states:
6316  * HAT_SYNC_DONTZERO means just return the rm stats
6317  * HAT_SYNC_ZERORM means zero rm bits in the tte and return the stats
6318  */
6319 void
6320 hat_sync(struct hat *sfmmup, caddr_t addr, size_t len, uint_t clearflag)
6321 {
6322 	struct hmehash_bucket *hmebp;
6323 	hmeblk_tag hblktag;
6324 	int hmeshift, hashno = 1;
6325 	struct hme_blk *hmeblkp, *list = NULL;
6326 	caddr_t endaddr;
6327 	cpuset_t cpuset;
6328 
6329 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
6330 	ASSERT((sfmmup == ksfmmup) || AS_LOCK_HELD(sfmmup->sfmmu_as));
6331 	ASSERT((len & MMU_PAGEOFFSET) == 0);
6332 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
6333 	    (clearflag == HAT_SYNC_ZERORM));
6334 
6335 	CPUSET_ZERO(cpuset);
6336 
6337 	endaddr = addr + len;
6338 	hblktag.htag_id = sfmmup;
6339 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
6340 
6341 	/*
6342 	 * Spitfire supports 4 page sizes.
6343 	 * Most pages are expected to be of the smallest page
6344 	 * size (8K) and these will not need to be rehashed. 64K
6345 	 * pages also don't need to be rehashed because the an hmeblk
6346 	 * spans 64K of address space. 512K pages might need 1 rehash and
6347 	 * and 4M pages 2 rehashes.
6348 	 */
6349 	while (addr < endaddr) {
6350 		hmeshift = HME_HASH_SHIFT(hashno);
6351 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
6352 		hblktag.htag_rehash = hashno;
6353 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
6354 
6355 		SFMMU_HASH_LOCK(hmebp);
6356 
6357 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
6358 		if (hmeblkp != NULL) {
6359 			ASSERT(!hmeblkp->hblk_shared);
6360 			/*
6361 			 * We've encountered a shadow hmeblk so skip the range
6362 			 * of the next smaller mapping size.
6363 			 */
6364 			if (hmeblkp->hblk_shw_bit) {
6365 				ASSERT(sfmmup != ksfmmup);
6366 				ASSERT(hashno > 1);
6367 				addr = (caddr_t)P2END((uintptr_t)addr,
6368 				    TTEBYTES(hashno - 1));
6369 			} else {
6370 				addr = sfmmu_hblk_sync(sfmmup, hmeblkp,
6371 				    addr, endaddr, clearflag);
6372 			}
6373 			SFMMU_HASH_UNLOCK(hmebp);
6374 			hashno = 1;
6375 			continue;
6376 		}
6377 		SFMMU_HASH_UNLOCK(hmebp);
6378 
6379 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
6380 			/*
6381 			 * We have traversed the whole list and rehashed
6382 			 * if necessary without finding the address to sync.
6383 			 * This is ok so we increment the address by the
6384 			 * smallest hmeblk range for kernel mappings and the
6385 			 * largest hmeblk range, to account for shadow hmeblks,
6386 			 * for user mappings and continue.
6387 			 */
6388 			if (sfmmup == ksfmmup)
6389 				addr = (caddr_t)P2END((uintptr_t)addr,
6390 				    TTEBYTES(1));
6391 			else
6392 				addr = (caddr_t)P2END((uintptr_t)addr,
6393 				    TTEBYTES(hashno));
6394 			hashno = 1;
6395 		} else {
6396 			hashno++;
6397 		}
6398 	}
6399 	sfmmu_hblks_list_purge(&list, 0);
6400 	cpuset = sfmmup->sfmmu_cpusran;
6401 	xt_sync(cpuset);
6402 }
6403 
6404 static caddr_t
6405 sfmmu_hblk_sync(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
6406 	caddr_t endaddr, int clearflag)
6407 {
6408 	tte_t	tte, ttemod;
6409 	struct sf_hment *sfhmep;
6410 	int ttesz;
6411 	struct page *pp;
6412 	kmutex_t *pml;
6413 	int ret;
6414 
6415 	ASSERT(hmeblkp->hblk_shw_bit == 0);
6416 	ASSERT(!hmeblkp->hblk_shared);
6417 
6418 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
6419 
6420 	ttesz = get_hblk_ttesz(hmeblkp);
6421 	HBLKTOHME(sfhmep, hmeblkp, addr);
6422 
6423 	while (addr < endaddr) {
6424 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6425 		if (TTE_IS_VALID(&tte)) {
6426 			pml = NULL;
6427 			pp = sfhmep->hme_page;
6428 			if (pp) {
6429 				pml = sfmmu_mlist_enter(pp);
6430 			}
6431 			if (pp != sfhmep->hme_page) {
6432 				/*
6433 				 * tte most have been unloaded
6434 				 * underneath us.  Recheck
6435 				 */
6436 				ASSERT(pml);
6437 				sfmmu_mlist_exit(pml);
6438 				continue;
6439 			}
6440 
6441 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
6442 
6443 			if (clearflag == HAT_SYNC_ZERORM) {
6444 				ttemod = tte;
6445 				TTE_CLR_RM(&ttemod);
6446 				ret = sfmmu_modifytte_try(&tte, &ttemod,
6447 				    &sfhmep->hme_tte);
6448 				if (ret < 0) {
6449 					if (pml) {
6450 						sfmmu_mlist_exit(pml);
6451 					}
6452 					continue;
6453 				}
6454 
6455 				if (ret > 0) {
6456 					sfmmu_tlb_demap(addr, sfmmup,
6457 					    hmeblkp, 0, 0);
6458 				}
6459 			}
6460 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
6461 			if (pml) {
6462 				sfmmu_mlist_exit(pml);
6463 			}
6464 		}
6465 		addr += TTEBYTES(ttesz);
6466 		sfhmep++;
6467 	}
6468 	return (addr);
6469 }
6470 
6471 /*
6472  * This function will sync a tte to the page struct and it will
6473  * update the hat stats. Currently it allows us to pass a NULL pp
6474  * and we will simply update the stats.  We may want to change this
6475  * so we only keep stats for pages backed by pp's.
6476  */
6477 static void
6478 sfmmu_ttesync(struct hat *sfmmup, caddr_t addr, tte_t *ttep, page_t *pp)
6479 {
6480 	uint_t rm = 0;
6481 	int   	sz;
6482 	pgcnt_t	npgs;
6483 
6484 	ASSERT(TTE_IS_VALID(ttep));
6485 
6486 	if (TTE_IS_NOSYNC(ttep)) {
6487 		return;
6488 	}
6489 
6490 	if (TTE_IS_REF(ttep))  {
6491 		rm = P_REF;
6492 	}
6493 	if (TTE_IS_MOD(ttep))  {
6494 		rm |= P_MOD;
6495 	}
6496 
6497 	if (rm == 0) {
6498 		return;
6499 	}
6500 
6501 	sz = TTE_CSZ(ttep);
6502 	if (sfmmup != NULL && sfmmup->sfmmu_rmstat) {
6503 		int i;
6504 		caddr_t	vaddr = addr;
6505 
6506 		for (i = 0; i < TTEPAGES(sz); i++, vaddr += MMU_PAGESIZE) {
6507 			hat_setstat(sfmmup->sfmmu_as, vaddr, MMU_PAGESIZE, rm);
6508 		}
6509 
6510 	}
6511 
6512 	/*
6513 	 * XXX I want to use cas to update nrm bits but they
6514 	 * currently belong in common/vm and not in hat where
6515 	 * they should be.
6516 	 * The nrm bits are protected by the same mutex as
6517 	 * the one that protects the page's mapping list.
6518 	 */
6519 	if (!pp)
6520 		return;
6521 	ASSERT(sfmmu_mlist_held(pp));
6522 	/*
6523 	 * If the tte is for a large page, we need to sync all the
6524 	 * pages covered by the tte.
6525 	 */
6526 	if (sz != TTE8K) {
6527 		ASSERT(pp->p_szc != 0);
6528 		pp = PP_GROUPLEADER(pp, sz);
6529 		ASSERT(sfmmu_mlist_held(pp));
6530 	}
6531 
6532 	/* Get number of pages from tte size. */
6533 	npgs = TTEPAGES(sz);
6534 
6535 	do {
6536 		ASSERT(pp);
6537 		ASSERT(sfmmu_mlist_held(pp));
6538 		if (((rm & P_REF) != 0 && !PP_ISREF(pp)) ||
6539 		    ((rm & P_MOD) != 0 && !PP_ISMOD(pp)))
6540 			hat_page_setattr(pp, rm);
6541 
6542 		/*
6543 		 * Are we done? If not, we must have a large mapping.
6544 		 * For large mappings we need to sync the rest of the pages
6545 		 * covered by this tte; goto the next page.
6546 		 */
6547 	} while (--npgs > 0 && (pp = PP_PAGENEXT(pp)));
6548 }
6549 
6550 /*
6551  * Execute pre-callback handler of each pa_hment linked to pp
6552  *
6553  * Inputs:
6554  *   flag: either HAT_PRESUSPEND or HAT_SUSPEND.
6555  *   capture_cpus: pointer to return value (below)
6556  *
6557  * Returns:
6558  *   Propagates the subsystem callback return values back to the caller;
6559  *   returns 0 on success.  If capture_cpus is non-NULL, the value returned
6560  *   is zero if all of the pa_hments are of a type that do not require
6561  *   capturing CPUs prior to suspending the mapping, else it is 1.
6562  */
6563 static int
6564 hat_pageprocess_precallbacks(struct page *pp, uint_t flag, int *capture_cpus)
6565 {
6566 	struct sf_hment	*sfhmep;
6567 	struct pa_hment *pahmep;
6568 	int (*f)(caddr_t, uint_t, uint_t, void *);
6569 	int		ret;
6570 	id_t		id;
6571 	int		locked = 0;
6572 	kmutex_t	*pml;
6573 
6574 	ASSERT(PAGE_EXCL(pp));
6575 	if (!sfmmu_mlist_held(pp)) {
6576 		pml = sfmmu_mlist_enter(pp);
6577 		locked = 1;
6578 	}
6579 
6580 	if (capture_cpus)
6581 		*capture_cpus = 0;
6582 
6583 top:
6584 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6585 		/*
6586 		 * skip sf_hments corresponding to VA<->PA mappings;
6587 		 * for pa_hment's, hme_tte.ll is zero
6588 		 */
6589 		if (!IS_PAHME(sfhmep))
6590 			continue;
6591 
6592 		pahmep = sfhmep->hme_data;
6593 		ASSERT(pahmep != NULL);
6594 
6595 		/*
6596 		 * skip if pre-handler has been called earlier in this loop
6597 		 */
6598 		if (pahmep->flags & flag)
6599 			continue;
6600 
6601 		id = pahmep->cb_id;
6602 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6603 		if (capture_cpus && sfmmu_cb_table[id].capture_cpus != 0)
6604 			*capture_cpus = 1;
6605 		if ((f = sfmmu_cb_table[id].prehandler) == NULL) {
6606 			pahmep->flags |= flag;
6607 			continue;
6608 		}
6609 
6610 		/*
6611 		 * Drop the mapping list lock to avoid locking order issues.
6612 		 */
6613 		if (locked)
6614 			sfmmu_mlist_exit(pml);
6615 
6616 		ret = f(pahmep->addr, pahmep->len, flag, pahmep->pvt);
6617 		if (ret != 0)
6618 			return (ret);	/* caller must do the cleanup */
6619 
6620 		if (locked) {
6621 			pml = sfmmu_mlist_enter(pp);
6622 			pahmep->flags |= flag;
6623 			goto top;
6624 		}
6625 
6626 		pahmep->flags |= flag;
6627 	}
6628 
6629 	if (locked)
6630 		sfmmu_mlist_exit(pml);
6631 
6632 	return (0);
6633 }
6634 
6635 /*
6636  * Execute post-callback handler of each pa_hment linked to pp
6637  *
6638  * Same overall assumptions and restrictions apply as for
6639  * hat_pageprocess_precallbacks().
6640  */
6641 static void
6642 hat_pageprocess_postcallbacks(struct page *pp, uint_t flag)
6643 {
6644 	pfn_t pgpfn = pp->p_pagenum;
6645 	pfn_t pgmask = btop(page_get_pagesize(pp->p_szc)) - 1;
6646 	pfn_t newpfn;
6647 	struct sf_hment *sfhmep;
6648 	struct pa_hment *pahmep;
6649 	int (*f)(caddr_t, uint_t, uint_t, void *, pfn_t);
6650 	id_t	id;
6651 	int	locked = 0;
6652 	kmutex_t *pml;
6653 
6654 	ASSERT(PAGE_EXCL(pp));
6655 	if (!sfmmu_mlist_held(pp)) {
6656 		pml = sfmmu_mlist_enter(pp);
6657 		locked = 1;
6658 	}
6659 
6660 top:
6661 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6662 		/*
6663 		 * skip sf_hments corresponding to VA<->PA mappings;
6664 		 * for pa_hment's, hme_tte.ll is zero
6665 		 */
6666 		if (!IS_PAHME(sfhmep))
6667 			continue;
6668 
6669 		pahmep = sfhmep->hme_data;
6670 		ASSERT(pahmep != NULL);
6671 
6672 		if ((pahmep->flags & flag) == 0)
6673 			continue;
6674 
6675 		pahmep->flags &= ~flag;
6676 
6677 		id = pahmep->cb_id;
6678 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6679 		if ((f = sfmmu_cb_table[id].posthandler) == NULL)
6680 			continue;
6681 
6682 		/*
6683 		 * Convert the base page PFN into the constituent PFN
6684 		 * which is needed by the callback handler.
6685 		 */
6686 		newpfn = pgpfn | (btop((uintptr_t)pahmep->addr) & pgmask);
6687 
6688 		/*
6689 		 * Drop the mapping list lock to avoid locking order issues.
6690 		 */
6691 		if (locked)
6692 			sfmmu_mlist_exit(pml);
6693 
6694 		if (f(pahmep->addr, pahmep->len, flag, pahmep->pvt, newpfn)
6695 		    != 0)
6696 			panic("sfmmu: posthandler failed");
6697 
6698 		if (locked) {
6699 			pml = sfmmu_mlist_enter(pp);
6700 			goto top;
6701 		}
6702 	}
6703 
6704 	if (locked)
6705 		sfmmu_mlist_exit(pml);
6706 }
6707 
6708 /*
6709  * Suspend locked kernel mapping
6710  */
6711 void
6712 hat_pagesuspend(struct page *pp)
6713 {
6714 	struct sf_hment *sfhmep;
6715 	sfmmu_t *sfmmup;
6716 	tte_t tte, ttemod;
6717 	struct hme_blk *hmeblkp;
6718 	caddr_t addr;
6719 	int index, cons;
6720 	cpuset_t cpuset;
6721 
6722 	ASSERT(PAGE_EXCL(pp));
6723 	ASSERT(sfmmu_mlist_held(pp));
6724 
6725 	mutex_enter(&kpr_suspendlock);
6726 
6727 	/*
6728 	 * We're about to suspend a kernel mapping so mark this thread as
6729 	 * non-traceable by DTrace. This prevents us from running into issues
6730 	 * with probe context trying to touch a suspended page
6731 	 * in the relocation codepath itself.
6732 	 */
6733 	curthread->t_flag |= T_DONTDTRACE;
6734 
6735 	index = PP_MAPINDEX(pp);
6736 	cons = TTE8K;
6737 
6738 retry:
6739 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6740 
6741 		if (IS_PAHME(sfhmep))
6742 			continue;
6743 
6744 		if (get_hblk_ttesz(sfmmu_hmetohblk(sfhmep)) != cons)
6745 			continue;
6746 
6747 		/*
6748 		 * Loop until we successfully set the suspend bit in
6749 		 * the TTE.
6750 		 */
6751 again:
6752 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6753 		ASSERT(TTE_IS_VALID(&tte));
6754 
6755 		ttemod = tte;
6756 		TTE_SET_SUSPEND(&ttemod);
6757 		if (sfmmu_modifytte_try(&tte, &ttemod,
6758 		    &sfhmep->hme_tte) < 0)
6759 			goto again;
6760 
6761 		/*
6762 		 * Invalidate TSB entry
6763 		 */
6764 		hmeblkp = sfmmu_hmetohblk(sfhmep);
6765 
6766 		sfmmup = hblktosfmmu(hmeblkp);
6767 		ASSERT(sfmmup == ksfmmup);
6768 		ASSERT(!hmeblkp->hblk_shared);
6769 
6770 		addr = tte_to_vaddr(hmeblkp, tte);
6771 
6772 		/*
6773 		 * No need to make sure that the TSB for this sfmmu is
6774 		 * not being relocated since it is ksfmmup and thus it
6775 		 * will never be relocated.
6776 		 */
6777 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
6778 
6779 		/*
6780 		 * Update xcall stats
6781 		 */
6782 		cpuset = cpu_ready_set;
6783 		CPUSET_DEL(cpuset, CPU->cpu_id);
6784 
6785 		/* LINTED: constant in conditional context */
6786 		SFMMU_XCALL_STATS(ksfmmup);
6787 
6788 		/*
6789 		 * Flush TLB entry on remote CPU's
6790 		 */
6791 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
6792 		    (uint64_t)ksfmmup);
6793 		xt_sync(cpuset);
6794 
6795 		/*
6796 		 * Flush TLB entry on local CPU
6797 		 */
6798 		vtag_flushpage(addr, (uint64_t)ksfmmup);
6799 	}
6800 
6801 	while (index != 0) {
6802 		index = index >> 1;
6803 		if (index != 0)
6804 			cons++;
6805 		if (index & 0x1) {
6806 			pp = PP_GROUPLEADER(pp, cons);
6807 			goto retry;
6808 		}
6809 	}
6810 }
6811 
6812 #ifdef	DEBUG
6813 
6814 #define	N_PRLE	1024
6815 struct prle {
6816 	page_t *targ;
6817 	page_t *repl;
6818 	int status;
6819 	int pausecpus;
6820 	hrtime_t whence;
6821 };
6822 
6823 static struct prle page_relocate_log[N_PRLE];
6824 static int prl_entry;
6825 static kmutex_t prl_mutex;
6826 
6827 #define	PAGE_RELOCATE_LOG(t, r, s, p)					\
6828 	mutex_enter(&prl_mutex);					\
6829 	page_relocate_log[prl_entry].targ = *(t);			\
6830 	page_relocate_log[prl_entry].repl = *(r);			\
6831 	page_relocate_log[prl_entry].status = (s);			\
6832 	page_relocate_log[prl_entry].pausecpus = (p);			\
6833 	page_relocate_log[prl_entry].whence = gethrtime();		\
6834 	prl_entry = (prl_entry == (N_PRLE - 1))? 0 : prl_entry + 1;	\
6835 	mutex_exit(&prl_mutex);
6836 
6837 #else	/* !DEBUG */
6838 #define	PAGE_RELOCATE_LOG(t, r, s, p)
6839 #endif
6840 
6841 /*
6842  * Core Kernel Page Relocation Algorithm
6843  *
6844  * Input:
6845  *
6846  * target : 	constituent pages are SE_EXCL locked.
6847  * replacement:	constituent pages are SE_EXCL locked.
6848  *
6849  * Output:
6850  *
6851  * nrelocp:	number of pages relocated
6852  */
6853 int
6854 hat_page_relocate(page_t **target, page_t **replacement, spgcnt_t *nrelocp)
6855 {
6856 	page_t		*targ, *repl;
6857 	page_t		*tpp, *rpp;
6858 	kmutex_t	*low, *high;
6859 	spgcnt_t	npages, i;
6860 	page_t		*pl = NULL;
6861 	int		old_pil;
6862 	cpuset_t	cpuset;
6863 	int		cap_cpus;
6864 	int		ret;
6865 #ifdef VAC
6866 	int		cflags = 0;
6867 #endif
6868 
6869 	if (!kcage_on || PP_ISNORELOC(*target)) {
6870 		PAGE_RELOCATE_LOG(target, replacement, EAGAIN, -1);
6871 		return (EAGAIN);
6872 	}
6873 
6874 	mutex_enter(&kpr_mutex);
6875 	kreloc_thread = curthread;
6876 
6877 	targ = *target;
6878 	repl = *replacement;
6879 	ASSERT(repl != NULL);
6880 	ASSERT(targ->p_szc == repl->p_szc);
6881 
6882 	npages = page_get_pagecnt(targ->p_szc);
6883 
6884 	/*
6885 	 * unload VA<->PA mappings that are not locked
6886 	 */
6887 	tpp = targ;
6888 	for (i = 0; i < npages; i++) {
6889 		(void) hat_pageunload(tpp, SFMMU_KERNEL_RELOC);
6890 		tpp++;
6891 	}
6892 
6893 	/*
6894 	 * Do "presuspend" callbacks, in a context from which we can still
6895 	 * block as needed. Note that we don't hold the mapping list lock
6896 	 * of "targ" at this point due to potential locking order issues;
6897 	 * we assume that between the hat_pageunload() above and holding
6898 	 * the SE_EXCL lock that the mapping list *cannot* change at this
6899 	 * point.
6900 	 */
6901 	ret = hat_pageprocess_precallbacks(targ, HAT_PRESUSPEND, &cap_cpus);
6902 	if (ret != 0) {
6903 		/*
6904 		 * EIO translates to fatal error, for all others cleanup
6905 		 * and return EAGAIN.
6906 		 */
6907 		ASSERT(ret != EIO);
6908 		hat_pageprocess_postcallbacks(targ, HAT_POSTUNSUSPEND);
6909 		PAGE_RELOCATE_LOG(target, replacement, ret, -1);
6910 		kreloc_thread = NULL;
6911 		mutex_exit(&kpr_mutex);
6912 		return (EAGAIN);
6913 	}
6914 
6915 	/*
6916 	 * acquire p_mapping list lock for both the target and replacement
6917 	 * root pages.
6918 	 *
6919 	 * low and high refer to the need to grab the mlist locks in a
6920 	 * specific order in order to prevent race conditions.  Thus the
6921 	 * lower lock must be grabbed before the higher lock.
6922 	 *
6923 	 * This will block hat_unload's accessing p_mapping list.  Since
6924 	 * we have SE_EXCL lock, hat_memload and hat_pageunload will be
6925 	 * blocked.  Thus, no one else will be accessing the p_mapping list
6926 	 * while we suspend and reload the locked mapping below.
6927 	 */
6928 	tpp = targ;
6929 	rpp = repl;
6930 	sfmmu_mlist_reloc_enter(tpp, rpp, &low, &high);
6931 
6932 	kpreempt_disable();
6933 
6934 	/*
6935 	 * We raise our PIL to 13 so that we don't get captured by
6936 	 * another CPU or pinned by an interrupt thread.  We can't go to
6937 	 * PIL 14 since the nexus driver(s) may need to interrupt at
6938 	 * that level in the case of IOMMU pseudo mappings.
6939 	 */
6940 	cpuset = cpu_ready_set;
6941 	CPUSET_DEL(cpuset, CPU->cpu_id);
6942 	if (!cap_cpus || CPUSET_ISNULL(cpuset)) {
6943 		old_pil = splr(XCALL_PIL);
6944 	} else {
6945 		old_pil = -1;
6946 		xc_attention(cpuset);
6947 	}
6948 	ASSERT(getpil() == XCALL_PIL);
6949 
6950 	/*
6951 	 * Now do suspend callbacks. In the case of an IOMMU mapping
6952 	 * this will suspend all DMA activity to the page while it is
6953 	 * being relocated. Since we are well above LOCK_LEVEL and CPUs
6954 	 * may be captured at this point we should have acquired any needed
6955 	 * locks in the presuspend callback.
6956 	 */
6957 	ret = hat_pageprocess_precallbacks(targ, HAT_SUSPEND, NULL);
6958 	if (ret != 0) {
6959 		repl = targ;
6960 		goto suspend_fail;
6961 	}
6962 
6963 	/*
6964 	 * Raise the PIL yet again, this time to block all high-level
6965 	 * interrupts on this CPU. This is necessary to prevent an
6966 	 * interrupt routine from pinning the thread which holds the
6967 	 * mapping suspended and then touching the suspended page.
6968 	 *
6969 	 * Once the page is suspended we also need to be careful to
6970 	 * avoid calling any functions which touch any seg_kmem memory
6971 	 * since that memory may be backed by the very page we are
6972 	 * relocating in here!
6973 	 */
6974 	hat_pagesuspend(targ);
6975 
6976 	/*
6977 	 * Now that we are confident everybody has stopped using this page,
6978 	 * copy the page contents.  Note we use a physical copy to prevent
6979 	 * locking issues and to avoid fpRAS because we can't handle it in
6980 	 * this context.
6981 	 */
6982 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6983 #ifdef VAC
6984 		/*
6985 		 * If the replacement has a different vcolor than
6986 		 * the one being replacd, we need to handle VAC
6987 		 * consistency for it just as we were setting up
6988 		 * a new mapping to it.
6989 		 */
6990 		if ((PP_GET_VCOLOR(rpp) != NO_VCOLOR) &&
6991 		    (tpp->p_vcolor != rpp->p_vcolor) &&
6992 		    !CacheColor_IsFlushed(cflags, PP_GET_VCOLOR(rpp))) {
6993 			CacheColor_SetFlushed(cflags, PP_GET_VCOLOR(rpp));
6994 			sfmmu_cache_flushcolor(PP_GET_VCOLOR(rpp),
6995 			    rpp->p_pagenum);
6996 		}
6997 #endif
6998 		/*
6999 		 * Copy the contents of the page.
7000 		 */
7001 		ppcopy_kernel(tpp, rpp);
7002 	}
7003 
7004 	tpp = targ;
7005 	rpp = repl;
7006 	for (i = 0; i < npages; i++, tpp++, rpp++) {
7007 		/*
7008 		 * Copy attributes.  VAC consistency was handled above,
7009 		 * if required.
7010 		 */
7011 		rpp->p_nrm = tpp->p_nrm;
7012 		tpp->p_nrm = 0;
7013 		rpp->p_index = tpp->p_index;
7014 		tpp->p_index = 0;
7015 #ifdef VAC
7016 		rpp->p_vcolor = tpp->p_vcolor;
7017 #endif
7018 	}
7019 
7020 	/*
7021 	 * First, unsuspend the page, if we set the suspend bit, and transfer
7022 	 * the mapping list from the target page to the replacement page.
7023 	 * Next process postcallbacks; since pa_hment's are linked only to the
7024 	 * p_mapping list of root page, we don't iterate over the constituent
7025 	 * pages.
7026 	 */
7027 	hat_pagereload(targ, repl);
7028 
7029 suspend_fail:
7030 	hat_pageprocess_postcallbacks(repl, HAT_UNSUSPEND);
7031 
7032 	/*
7033 	 * Now lower our PIL and release any captured CPUs since we
7034 	 * are out of the "danger zone".  After this it will again be
7035 	 * safe to acquire adaptive mutex locks, or to drop them...
7036 	 */
7037 	if (old_pil != -1) {
7038 		splx(old_pil);
7039 	} else {
7040 		xc_dismissed(cpuset);
7041 	}
7042 
7043 	kpreempt_enable();
7044 
7045 	sfmmu_mlist_reloc_exit(low, high);
7046 
7047 	/*
7048 	 * Postsuspend callbacks should drop any locks held across
7049 	 * the suspend callbacks.  As before, we don't hold the mapping
7050 	 * list lock at this point.. our assumption is that the mapping
7051 	 * list still can't change due to our holding SE_EXCL lock and
7052 	 * there being no unlocked mappings left. Hence the restriction
7053 	 * on calling context to hat_delete_callback()
7054 	 */
7055 	hat_pageprocess_postcallbacks(repl, HAT_POSTUNSUSPEND);
7056 	if (ret != 0) {
7057 		/*
7058 		 * The second presuspend call failed: we got here through
7059 		 * the suspend_fail label above.
7060 		 */
7061 		ASSERT(ret != EIO);
7062 		PAGE_RELOCATE_LOG(target, replacement, ret, cap_cpus);
7063 		kreloc_thread = NULL;
7064 		mutex_exit(&kpr_mutex);
7065 		return (EAGAIN);
7066 	}
7067 
7068 	/*
7069 	 * Now that we're out of the performance critical section we can
7070 	 * take care of updating the hash table, since we still
7071 	 * hold all the pages locked SE_EXCL at this point we
7072 	 * needn't worry about things changing out from under us.
7073 	 */
7074 	tpp = targ;
7075 	rpp = repl;
7076 	for (i = 0; i < npages; i++, tpp++, rpp++) {
7077 
7078 		/*
7079 		 * replace targ with replacement in page_hash table
7080 		 */
7081 		targ = tpp;
7082 		page_relocate_hash(rpp, targ);
7083 
7084 		/*
7085 		 * concatenate target; caller of platform_page_relocate()
7086 		 * expects target to be concatenated after returning.
7087 		 */
7088 		ASSERT(targ->p_next == targ);
7089 		ASSERT(targ->p_prev == targ);
7090 		page_list_concat(&pl, &targ);
7091 	}
7092 
7093 	ASSERT(*target == pl);
7094 	*nrelocp = npages;
7095 	PAGE_RELOCATE_LOG(target, replacement, 0, cap_cpus);
7096 	kreloc_thread = NULL;
7097 	mutex_exit(&kpr_mutex);
7098 	return (0);
7099 }
7100 
7101 /*
7102  * Called when stray pa_hments are found attached to a page which is
7103  * being freed.  Notify the subsystem which attached the pa_hment of
7104  * the error if it registered a suitable handler, else panic.
7105  */
7106 static void
7107 sfmmu_pahment_leaked(struct pa_hment *pahmep)
7108 {
7109 	id_t cb_id = pahmep->cb_id;
7110 
7111 	ASSERT(cb_id >= (id_t)0 && cb_id < sfmmu_cb_nextid);
7112 	if (sfmmu_cb_table[cb_id].errhandler != NULL) {
7113 		if (sfmmu_cb_table[cb_id].errhandler(pahmep->addr, pahmep->len,
7114 		    HAT_CB_ERR_LEAKED, pahmep->pvt) == 0)
7115 			return;		/* non-fatal */
7116 	}
7117 	panic("pa_hment leaked: 0x%p", (void *)pahmep);
7118 }
7119 
7120 /*
7121  * Remove all mappings to page 'pp'.
7122  */
7123 int
7124 hat_pageunload(struct page *pp, uint_t forceflag)
7125 {
7126 	struct page *origpp = pp;
7127 	struct sf_hment *sfhme, *tmphme;
7128 	struct hme_blk *hmeblkp;
7129 	kmutex_t *pml;
7130 #ifdef VAC
7131 	kmutex_t *pmtx;
7132 #endif
7133 	cpuset_t cpuset, tset;
7134 	int index, cons;
7135 	int xhme_blks;
7136 	int pa_hments;
7137 
7138 	ASSERT(PAGE_EXCL(pp));
7139 
7140 retry_xhat:
7141 	tmphme = NULL;
7142 	xhme_blks = 0;
7143 	pa_hments = 0;
7144 	CPUSET_ZERO(cpuset);
7145 
7146 	pml = sfmmu_mlist_enter(pp);
7147 
7148 #ifdef VAC
7149 	if (pp->p_kpmref)
7150 		sfmmu_kpm_pageunload(pp);
7151 	ASSERT(!PP_ISMAPPED_KPM(pp));
7152 #endif
7153 	/*
7154 	 * Clear vpm reference. Since the page is exclusively locked
7155 	 * vpm cannot be referencing it.
7156 	 */
7157 	if (vpm_enable) {
7158 		pp->p_vpmref = 0;
7159 	}
7160 
7161 	index = PP_MAPINDEX(pp);
7162 	cons = TTE8K;
7163 retry:
7164 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7165 		tmphme = sfhme->hme_next;
7166 
7167 		if (IS_PAHME(sfhme)) {
7168 			ASSERT(sfhme->hme_data != NULL);
7169 			pa_hments++;
7170 			continue;
7171 		}
7172 
7173 		hmeblkp = sfmmu_hmetohblk(sfhme);
7174 		if (hmeblkp->hblk_xhat_bit) {
7175 			struct xhat_hme_blk *xblk =
7176 			    (struct xhat_hme_blk *)hmeblkp;
7177 
7178 			(void) XHAT_PAGEUNLOAD(xblk->xhat_hme_blk_hat,
7179 			    pp, forceflag, XBLK2PROVBLK(xblk));
7180 
7181 			xhme_blks = 1;
7182 			continue;
7183 		}
7184 
7185 		/*
7186 		 * If there are kernel mappings don't unload them, they will
7187 		 * be suspended.
7188 		 */
7189 		if (forceflag == SFMMU_KERNEL_RELOC && hmeblkp->hblk_lckcnt &&
7190 		    hmeblkp->hblk_tag.htag_id == ksfmmup)
7191 			continue;
7192 
7193 		tset = sfmmu_pageunload(pp, sfhme, cons);
7194 		CPUSET_OR(cpuset, tset);
7195 	}
7196 
7197 	while (index != 0) {
7198 		index = index >> 1;
7199 		if (index != 0)
7200 			cons++;
7201 		if (index & 0x1) {
7202 			/* Go to leading page */
7203 			pp = PP_GROUPLEADER(pp, cons);
7204 			ASSERT(sfmmu_mlist_held(pp));
7205 			goto retry;
7206 		}
7207 	}
7208 
7209 	/*
7210 	 * cpuset may be empty if the page was only mapped by segkpm,
7211 	 * in which case we won't actually cross-trap.
7212 	 */
7213 	xt_sync(cpuset);
7214 
7215 	/*
7216 	 * The page should have no mappings at this point, unless
7217 	 * we were called from hat_page_relocate() in which case we
7218 	 * leave the locked mappings which will be suspended later.
7219 	 */
7220 	ASSERT(!PP_ISMAPPED(origpp) || xhme_blks || pa_hments ||
7221 	    (forceflag == SFMMU_KERNEL_RELOC));
7222 
7223 #ifdef VAC
7224 	if (PP_ISTNC(pp)) {
7225 		if (cons == TTE8K) {
7226 			pmtx = sfmmu_page_enter(pp);
7227 			PP_CLRTNC(pp);
7228 			sfmmu_page_exit(pmtx);
7229 		} else {
7230 			conv_tnc(pp, cons);
7231 		}
7232 	}
7233 #endif	/* VAC */
7234 
7235 	if (pa_hments && forceflag != SFMMU_KERNEL_RELOC) {
7236 		/*
7237 		 * Unlink any pa_hments and free them, calling back
7238 		 * the responsible subsystem to notify it of the error.
7239 		 * This can occur in situations such as drivers leaking
7240 		 * DMA handles: naughty, but common enough that we'd like
7241 		 * to keep the system running rather than bringing it
7242 		 * down with an obscure error like "pa_hment leaked"
7243 		 * which doesn't aid the user in debugging their driver.
7244 		 */
7245 		for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7246 			tmphme = sfhme->hme_next;
7247 			if (IS_PAHME(sfhme)) {
7248 				struct pa_hment *pahmep = sfhme->hme_data;
7249 				sfmmu_pahment_leaked(pahmep);
7250 				HME_SUB(sfhme, pp);
7251 				kmem_cache_free(pa_hment_cache, pahmep);
7252 			}
7253 		}
7254 
7255 		ASSERT(!PP_ISMAPPED(origpp) || xhme_blks);
7256 	}
7257 
7258 	sfmmu_mlist_exit(pml);
7259 
7260 	/*
7261 	 * XHAT may not have finished unloading pages
7262 	 * because some other thread was waiting for
7263 	 * mlist lock and XHAT_PAGEUNLOAD let it do
7264 	 * the job.
7265 	 */
7266 	if (xhme_blks) {
7267 		pp = origpp;
7268 		goto retry_xhat;
7269 	}
7270 
7271 	return (0);
7272 }
7273 
7274 cpuset_t
7275 sfmmu_pageunload(page_t *pp, struct sf_hment *sfhme, int cons)
7276 {
7277 	struct hme_blk *hmeblkp;
7278 	sfmmu_t *sfmmup;
7279 	tte_t tte, ttemod;
7280 #ifdef DEBUG
7281 	tte_t orig_old;
7282 #endif /* DEBUG */
7283 	caddr_t addr;
7284 	int ttesz;
7285 	int ret;
7286 	cpuset_t cpuset;
7287 
7288 	ASSERT(pp != NULL);
7289 	ASSERT(sfmmu_mlist_held(pp));
7290 	ASSERT(!PP_ISKAS(pp));
7291 
7292 	CPUSET_ZERO(cpuset);
7293 
7294 	hmeblkp = sfmmu_hmetohblk(sfhme);
7295 
7296 readtte:
7297 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7298 	if (TTE_IS_VALID(&tte)) {
7299 		sfmmup = hblktosfmmu(hmeblkp);
7300 		ttesz = get_hblk_ttesz(hmeblkp);
7301 		/*
7302 		 * Only unload mappings of 'cons' size.
7303 		 */
7304 		if (ttesz != cons)
7305 			return (cpuset);
7306 
7307 		/*
7308 		 * Note that we have p_mapping lock, but no hash lock here.
7309 		 * hblk_unload() has to have both hash lock AND p_mapping
7310 		 * lock before it tries to modify tte. So, the tte could
7311 		 * not become invalid in the sfmmu_modifytte_try() below.
7312 		 */
7313 		ttemod = tte;
7314 #ifdef DEBUG
7315 		orig_old = tte;
7316 #endif /* DEBUG */
7317 
7318 		TTE_SET_INVALID(&ttemod);
7319 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7320 		if (ret < 0) {
7321 #ifdef DEBUG
7322 			/* only R/M bits can change. */
7323 			chk_tte(&orig_old, &tte, &ttemod, hmeblkp);
7324 #endif /* DEBUG */
7325 			goto readtte;
7326 		}
7327 
7328 		if (ret == 0) {
7329 			panic("pageunload: cas failed?");
7330 		}
7331 
7332 		addr = tte_to_vaddr(hmeblkp, tte);
7333 
7334 		if (hmeblkp->hblk_shared) {
7335 			sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7336 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
7337 			sf_region_t *rgnp;
7338 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7339 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7340 			ASSERT(srdp != NULL);
7341 			rgnp = srdp->srd_hmergnp[rid];
7342 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
7343 			cpuset = sfmmu_rgntlb_demap(addr, rgnp, hmeblkp, 1);
7344 			sfmmu_ttesync(NULL, addr, &tte, pp);
7345 			ASSERT(rgnp->rgn_ttecnt[ttesz] > 0);
7346 			atomic_dec_ulong(&rgnp->rgn_ttecnt[ttesz]);
7347 		} else {
7348 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
7349 			atomic_dec_ulong(&sfmmup->sfmmu_ttecnt[ttesz]);
7350 
7351 			/*
7352 			 * We need to flush the page from the virtual cache
7353 			 * in order to prevent a virtual cache alias
7354 			 * inconsistency. The particular scenario we need
7355 			 * to worry about is:
7356 			 * Given:  va1 and va2 are two virtual address that
7357 			 * alias and will map the same physical address.
7358 			 * 1.   mapping exists from va1 to pa and data has
7359 			 *	been read into the cache.
7360 			 * 2.   unload va1.
7361 			 * 3.   load va2 and modify data using va2.
7362 			 * 4    unload va2.
7363 			 * 5.   load va1 and reference data.  Unless we flush
7364 			 *	the data cache when we unload we will get
7365 			 *	stale data.
7366 			 * This scenario is taken care of by using virtual
7367 			 * page coloring.
7368 			 */
7369 			if (sfmmup->sfmmu_ismhat) {
7370 				/*
7371 				 * Flush TSBs, TLBs and caches
7372 				 * of every process
7373 				 * sharing this ism segment.
7374 				 */
7375 				sfmmu_hat_lock_all();
7376 				mutex_enter(&ism_mlist_lock);
7377 				kpreempt_disable();
7378 				sfmmu_ismtlbcache_demap(addr, sfmmup, hmeblkp,
7379 				    pp->p_pagenum, CACHE_NO_FLUSH);
7380 				kpreempt_enable();
7381 				mutex_exit(&ism_mlist_lock);
7382 				sfmmu_hat_unlock_all();
7383 				cpuset = cpu_ready_set;
7384 			} else {
7385 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7386 				cpuset = sfmmup->sfmmu_cpusran;
7387 			}
7388 		}
7389 
7390 		/*
7391 		 * Hme_sub has to run after ttesync() and a_rss update.
7392 		 * See hblk_unload().
7393 		 */
7394 		HME_SUB(sfhme, pp);
7395 		membar_stst();
7396 
7397 		/*
7398 		 * We can not make ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
7399 		 * since pteload may have done a HME_ADD() right after
7400 		 * we did the HME_SUB() above. Hmecnt is now maintained
7401 		 * by cas only. no lock guranteed its value. The only
7402 		 * gurantee we have is the hmecnt should not be less than
7403 		 * what it should be so the hblk will not be taken away.
7404 		 * It's also important that we decremented the hmecnt after
7405 		 * we are done with hmeblkp so that this hmeblk won't be
7406 		 * stolen.
7407 		 */
7408 		ASSERT(hmeblkp->hblk_hmecnt > 0);
7409 		ASSERT(hmeblkp->hblk_vcnt > 0);
7410 		atomic_dec_16(&hmeblkp->hblk_vcnt);
7411 		atomic_dec_16(&hmeblkp->hblk_hmecnt);
7412 		/*
7413 		 * This is bug 4063182.
7414 		 * XXX: fixme
7415 		 * ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
7416 		 *	!hmeblkp->hblk_lckcnt);
7417 		 */
7418 	} else {
7419 		panic("invalid tte? pp %p &tte %p",
7420 		    (void *)pp, (void *)&tte);
7421 	}
7422 
7423 	return (cpuset);
7424 }
7425 
7426 /*
7427  * While relocating a kernel page, this function will move the mappings
7428  * from tpp to dpp and modify any associated data with these mappings.
7429  * It also unsuspends the suspended kernel mapping.
7430  */
7431 static void
7432 hat_pagereload(struct page *tpp, struct page *dpp)
7433 {
7434 	struct sf_hment *sfhme;
7435 	tte_t tte, ttemod;
7436 	int index, cons;
7437 
7438 	ASSERT(getpil() == PIL_MAX);
7439 	ASSERT(sfmmu_mlist_held(tpp));
7440 	ASSERT(sfmmu_mlist_held(dpp));
7441 
7442 	index = PP_MAPINDEX(tpp);
7443 	cons = TTE8K;
7444 
7445 	/* Update real mappings to the page */
7446 retry:
7447 	for (sfhme = tpp->p_mapping; sfhme != NULL; sfhme = sfhme->hme_next) {
7448 		if (IS_PAHME(sfhme))
7449 			continue;
7450 		sfmmu_copytte(&sfhme->hme_tte, &tte);
7451 		ttemod = tte;
7452 
7453 		/*
7454 		 * replace old pfn with new pfn in TTE
7455 		 */
7456 		PFN_TO_TTE(ttemod, dpp->p_pagenum);
7457 
7458 		/*
7459 		 * clear suspend bit
7460 		 */
7461 		ASSERT(TTE_IS_SUSPEND(&ttemod));
7462 		TTE_CLR_SUSPEND(&ttemod);
7463 
7464 		if (sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte) < 0)
7465 			panic("hat_pagereload(): sfmmu_modifytte_try() failed");
7466 
7467 		/*
7468 		 * set hme_page point to new page
7469 		 */
7470 		sfhme->hme_page = dpp;
7471 	}
7472 
7473 	/*
7474 	 * move p_mapping list from old page to new page
7475 	 */
7476 	dpp->p_mapping = tpp->p_mapping;
7477 	tpp->p_mapping = NULL;
7478 	dpp->p_share = tpp->p_share;
7479 	tpp->p_share = 0;
7480 
7481 	while (index != 0) {
7482 		index = index >> 1;
7483 		if (index != 0)
7484 			cons++;
7485 		if (index & 0x1) {
7486 			tpp = PP_GROUPLEADER(tpp, cons);
7487 			dpp = PP_GROUPLEADER(dpp, cons);
7488 			goto retry;
7489 		}
7490 	}
7491 
7492 	curthread->t_flag &= ~T_DONTDTRACE;
7493 	mutex_exit(&kpr_suspendlock);
7494 }
7495 
7496 uint_t
7497 hat_pagesync(struct page *pp, uint_t clearflag)
7498 {
7499 	struct sf_hment *sfhme, *tmphme = NULL;
7500 	struct hme_blk *hmeblkp;
7501 	kmutex_t *pml;
7502 	cpuset_t cpuset, tset;
7503 	int	index, cons;
7504 	extern	ulong_t po_share;
7505 	page_t	*save_pp = pp;
7506 	int	stop_on_sh = 0;
7507 	uint_t	shcnt;
7508 
7509 	CPUSET_ZERO(cpuset);
7510 
7511 	if (PP_ISRO(pp) && (clearflag & HAT_SYNC_STOPON_MOD)) {
7512 		return (PP_GENERIC_ATTR(pp));
7513 	}
7514 
7515 	if ((clearflag & HAT_SYNC_ZERORM) == 0) {
7516 		if ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(pp)) {
7517 			return (PP_GENERIC_ATTR(pp));
7518 		}
7519 		if ((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(pp)) {
7520 			return (PP_GENERIC_ATTR(pp));
7521 		}
7522 		if (clearflag & HAT_SYNC_STOPON_SHARED) {
7523 			if (pp->p_share > po_share) {
7524 				hat_page_setattr(pp, P_REF);
7525 				return (PP_GENERIC_ATTR(pp));
7526 			}
7527 			stop_on_sh = 1;
7528 			shcnt = 0;
7529 		}
7530 	}
7531 
7532 	clearflag &= ~HAT_SYNC_STOPON_SHARED;
7533 	pml = sfmmu_mlist_enter(pp);
7534 	index = PP_MAPINDEX(pp);
7535 	cons = TTE8K;
7536 retry:
7537 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7538 		/*
7539 		 * We need to save the next hment on the list since
7540 		 * it is possible for pagesync to remove an invalid hment
7541 		 * from the list.
7542 		 */
7543 		tmphme = sfhme->hme_next;
7544 		if (IS_PAHME(sfhme))
7545 			continue;
7546 		/*
7547 		 * If we are looking for large mappings and this hme doesn't
7548 		 * reach the range we are seeking, just ignore it.
7549 		 */
7550 		hmeblkp = sfmmu_hmetohblk(sfhme);
7551 		if (hmeblkp->hblk_xhat_bit)
7552 			continue;
7553 
7554 		if (hme_size(sfhme) < cons)
7555 			continue;
7556 
7557 		if (stop_on_sh) {
7558 			if (hmeblkp->hblk_shared) {
7559 				sf_srd_t *srdp = hblktosrd(hmeblkp);
7560 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7561 				sf_region_t *rgnp;
7562 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7563 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7564 				ASSERT(srdp != NULL);
7565 				rgnp = srdp->srd_hmergnp[rid];
7566 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
7567 				    rgnp, rid);
7568 				shcnt += rgnp->rgn_refcnt;
7569 			} else {
7570 				shcnt++;
7571 			}
7572 			if (shcnt > po_share) {
7573 				/*
7574 				 * tell the pager to spare the page this time
7575 				 * around.
7576 				 */
7577 				hat_page_setattr(save_pp, P_REF);
7578 				index = 0;
7579 				break;
7580 			}
7581 		}
7582 		tset = sfmmu_pagesync(pp, sfhme,
7583 		    clearflag & ~HAT_SYNC_STOPON_RM);
7584 		CPUSET_OR(cpuset, tset);
7585 
7586 		/*
7587 		 * If clearflag is HAT_SYNC_DONTZERO, break out as soon
7588 		 * as the "ref" or "mod" is set or share cnt exceeds po_share.
7589 		 */
7590 		if ((clearflag & ~HAT_SYNC_STOPON_RM) == HAT_SYNC_DONTZERO &&
7591 		    (((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(save_pp)) ||
7592 		    ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(save_pp)))) {
7593 			index = 0;
7594 			break;
7595 		}
7596 	}
7597 
7598 	while (index) {
7599 		index = index >> 1;
7600 		cons++;
7601 		if (index & 0x1) {
7602 			/* Go to leading page */
7603 			pp = PP_GROUPLEADER(pp, cons);
7604 			goto retry;
7605 		}
7606 	}
7607 
7608 	xt_sync(cpuset);
7609 	sfmmu_mlist_exit(pml);
7610 	return (PP_GENERIC_ATTR(save_pp));
7611 }
7612 
7613 /*
7614  * Get all the hardware dependent attributes for a page struct
7615  */
7616 static cpuset_t
7617 sfmmu_pagesync(struct page *pp, struct sf_hment *sfhme,
7618 	uint_t clearflag)
7619 {
7620 	caddr_t addr;
7621 	tte_t tte, ttemod;
7622 	struct hme_blk *hmeblkp;
7623 	int ret;
7624 	sfmmu_t *sfmmup;
7625 	cpuset_t cpuset;
7626 
7627 	ASSERT(pp != NULL);
7628 	ASSERT(sfmmu_mlist_held(pp));
7629 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
7630 	    (clearflag == HAT_SYNC_ZERORM));
7631 
7632 	SFMMU_STAT(sf_pagesync);
7633 
7634 	CPUSET_ZERO(cpuset);
7635 
7636 sfmmu_pagesync_retry:
7637 
7638 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7639 	if (TTE_IS_VALID(&tte)) {
7640 		hmeblkp = sfmmu_hmetohblk(sfhme);
7641 		sfmmup = hblktosfmmu(hmeblkp);
7642 		addr = tte_to_vaddr(hmeblkp, tte);
7643 		if (clearflag == HAT_SYNC_ZERORM) {
7644 			ttemod = tte;
7645 			TTE_CLR_RM(&ttemod);
7646 			ret = sfmmu_modifytte_try(&tte, &ttemod,
7647 			    &sfhme->hme_tte);
7648 			if (ret < 0) {
7649 				/*
7650 				 * cas failed and the new value is not what
7651 				 * we want.
7652 				 */
7653 				goto sfmmu_pagesync_retry;
7654 			}
7655 
7656 			if (ret > 0) {
7657 				/* we win the cas */
7658 				if (hmeblkp->hblk_shared) {
7659 					sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7660 					uint_t rid =
7661 					    hmeblkp->hblk_tag.htag_rid;
7662 					sf_region_t *rgnp;
7663 					ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7664 					ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7665 					ASSERT(srdp != NULL);
7666 					rgnp = srdp->srd_hmergnp[rid];
7667 					SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7668 					    srdp, rgnp, rid);
7669 					cpuset = sfmmu_rgntlb_demap(addr,
7670 					    rgnp, hmeblkp, 1);
7671 				} else {
7672 					sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
7673 					    0, 0);
7674 					cpuset = sfmmup->sfmmu_cpusran;
7675 				}
7676 			}
7677 		}
7678 		sfmmu_ttesync(hmeblkp->hblk_shared ? NULL : sfmmup, addr,
7679 		    &tte, pp);
7680 	}
7681 	return (cpuset);
7682 }
7683 
7684 /*
7685  * Remove write permission from a mappings to a page, so that
7686  * we can detect the next modification of it. This requires modifying
7687  * the TTE then invalidating (demap) any TLB entry using that TTE.
7688  * This code is similar to sfmmu_pagesync().
7689  */
7690 static cpuset_t
7691 sfmmu_pageclrwrt(struct page *pp, struct sf_hment *sfhme)
7692 {
7693 	caddr_t addr;
7694 	tte_t tte;
7695 	tte_t ttemod;
7696 	struct hme_blk *hmeblkp;
7697 	int ret;
7698 	sfmmu_t *sfmmup;
7699 	cpuset_t cpuset;
7700 
7701 	ASSERT(pp != NULL);
7702 	ASSERT(sfmmu_mlist_held(pp));
7703 
7704 	CPUSET_ZERO(cpuset);
7705 	SFMMU_STAT(sf_clrwrt);
7706 
7707 retry:
7708 
7709 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7710 	if (TTE_IS_VALID(&tte) && TTE_IS_WRITABLE(&tte)) {
7711 		hmeblkp = sfmmu_hmetohblk(sfhme);
7712 
7713 		/*
7714 		 * xhat mappings should never be to a VMODSORT page.
7715 		 */
7716 		ASSERT(hmeblkp->hblk_xhat_bit == 0);
7717 
7718 		sfmmup = hblktosfmmu(hmeblkp);
7719 		addr = tte_to_vaddr(hmeblkp, tte);
7720 
7721 		ttemod = tte;
7722 		TTE_CLR_WRT(&ttemod);
7723 		TTE_CLR_MOD(&ttemod);
7724 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7725 
7726 		/*
7727 		 * if cas failed and the new value is not what
7728 		 * we want retry
7729 		 */
7730 		if (ret < 0)
7731 			goto retry;
7732 
7733 		/* we win the cas */
7734 		if (ret > 0) {
7735 			if (hmeblkp->hblk_shared) {
7736 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7737 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7738 				sf_region_t *rgnp;
7739 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7740 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7741 				ASSERT(srdp != NULL);
7742 				rgnp = srdp->srd_hmergnp[rid];
7743 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7744 				    srdp, rgnp, rid);
7745 				cpuset = sfmmu_rgntlb_demap(addr,
7746 				    rgnp, hmeblkp, 1);
7747 			} else {
7748 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7749 				cpuset = sfmmup->sfmmu_cpusran;
7750 			}
7751 		}
7752 	}
7753 
7754 	return (cpuset);
7755 }
7756 
7757 /*
7758  * Walk all mappings of a page, removing write permission and clearing the
7759  * ref/mod bits. This code is similar to hat_pagesync()
7760  */
7761 static void
7762 hat_page_clrwrt(page_t *pp)
7763 {
7764 	struct sf_hment *sfhme;
7765 	struct sf_hment *tmphme = NULL;
7766 	kmutex_t *pml;
7767 	cpuset_t cpuset;
7768 	cpuset_t tset;
7769 	int	index;
7770 	int	 cons;
7771 
7772 	CPUSET_ZERO(cpuset);
7773 
7774 	pml = sfmmu_mlist_enter(pp);
7775 	index = PP_MAPINDEX(pp);
7776 	cons = TTE8K;
7777 retry:
7778 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7779 		tmphme = sfhme->hme_next;
7780 
7781 		/*
7782 		 * If we are looking for large mappings and this hme doesn't
7783 		 * reach the range we are seeking, just ignore its.
7784 		 */
7785 
7786 		if (hme_size(sfhme) < cons)
7787 			continue;
7788 
7789 		tset = sfmmu_pageclrwrt(pp, sfhme);
7790 		CPUSET_OR(cpuset, tset);
7791 	}
7792 
7793 	while (index) {
7794 		index = index >> 1;
7795 		cons++;
7796 		if (index & 0x1) {
7797 			/* Go to leading page */
7798 			pp = PP_GROUPLEADER(pp, cons);
7799 			goto retry;
7800 		}
7801 	}
7802 
7803 	xt_sync(cpuset);
7804 	sfmmu_mlist_exit(pml);
7805 }
7806 
7807 /*
7808  * Set the given REF/MOD/RO bits for the given page.
7809  * For a vnode with a sorted v_pages list, we need to change
7810  * the attributes and the v_pages list together under page_vnode_mutex.
7811  */
7812 void
7813 hat_page_setattr(page_t *pp, uint_t flag)
7814 {
7815 	vnode_t		*vp = pp->p_vnode;
7816 	page_t		**listp;
7817 	kmutex_t	*pmtx;
7818 	kmutex_t	*vphm = NULL;
7819 	int		noshuffle;
7820 
7821 	noshuffle = flag & P_NSH;
7822 	flag &= ~P_NSH;
7823 
7824 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7825 
7826 	/*
7827 	 * nothing to do if attribute already set
7828 	 */
7829 	if ((pp->p_nrm & flag) == flag)
7830 		return;
7831 
7832 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp) &&
7833 	    !noshuffle) {
7834 		vphm = page_vnode_mutex(vp);
7835 		mutex_enter(vphm);
7836 	}
7837 
7838 	pmtx = sfmmu_page_enter(pp);
7839 	pp->p_nrm |= flag;
7840 	sfmmu_page_exit(pmtx);
7841 
7842 	if (vphm != NULL) {
7843 		/*
7844 		 * Some File Systems examine v_pages for NULL w/o
7845 		 * grabbing the vphm mutex. Must not let it become NULL when
7846 		 * pp is the only page on the list.
7847 		 */
7848 		if (pp->p_vpnext != pp) {
7849 			page_vpsub(&vp->v_pages, pp);
7850 			if (vp->v_pages != NULL)
7851 				listp = &vp->v_pages->p_vpprev->p_vpnext;
7852 			else
7853 				listp = &vp->v_pages;
7854 			page_vpadd(listp, pp);
7855 		}
7856 		mutex_exit(vphm);
7857 	}
7858 }
7859 
7860 void
7861 hat_page_clrattr(page_t *pp, uint_t flag)
7862 {
7863 	vnode_t		*vp = pp->p_vnode;
7864 	kmutex_t	*pmtx;
7865 
7866 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7867 
7868 	pmtx = sfmmu_page_enter(pp);
7869 
7870 	/*
7871 	 * Caller is expected to hold page's io lock for VMODSORT to work
7872 	 * correctly with pvn_vplist_dirty() and pvn_getdirty() when mod
7873 	 * bit is cleared.
7874 	 * We don't have assert to avoid tripping some existing third party
7875 	 * code. The dirty page is moved back to top of the v_page list
7876 	 * after IO is done in pvn_write_done().
7877 	 */
7878 	pp->p_nrm &= ~flag;
7879 	sfmmu_page_exit(pmtx);
7880 
7881 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) {
7882 
7883 		/*
7884 		 * VMODSORT works by removing write permissions and getting
7885 		 * a fault when a page is made dirty. At this point
7886 		 * we need to remove write permission from all mappings
7887 		 * to this page.
7888 		 */
7889 		hat_page_clrwrt(pp);
7890 	}
7891 }
7892 
7893 uint_t
7894 hat_page_getattr(page_t *pp, uint_t flag)
7895 {
7896 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7897 	return ((uint_t)(pp->p_nrm & flag));
7898 }
7899 
7900 /*
7901  * DEBUG kernels: verify that a kernel va<->pa translation
7902  * is safe by checking the underlying page_t is in a page
7903  * relocation-safe state.
7904  */
7905 #ifdef	DEBUG
7906 void
7907 sfmmu_check_kpfn(pfn_t pfn)
7908 {
7909 	page_t *pp;
7910 	int index, cons;
7911 
7912 	if (hat_check_vtop == 0)
7913 		return;
7914 
7915 	if (kvseg.s_base == NULL || panicstr)
7916 		return;
7917 
7918 	pp = page_numtopp_nolock(pfn);
7919 	if (!pp)
7920 		return;
7921 
7922 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7923 		return;
7924 
7925 	/*
7926 	 * Handed a large kernel page, we dig up the root page since we
7927 	 * know the root page might have the lock also.
7928 	 */
7929 	if (pp->p_szc != 0) {
7930 		index = PP_MAPINDEX(pp);
7931 		cons = TTE8K;
7932 again:
7933 		while (index != 0) {
7934 			index >>= 1;
7935 			if (index != 0)
7936 				cons++;
7937 			if (index & 0x1) {
7938 				pp = PP_GROUPLEADER(pp, cons);
7939 				goto again;
7940 			}
7941 		}
7942 	}
7943 
7944 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7945 		return;
7946 
7947 	/*
7948 	 * Pages need to be locked or allocated "permanent" (either from
7949 	 * static_arena arena or explicitly setting PG_NORELOC when calling
7950 	 * page_create_va()) for VA->PA translations to be valid.
7951 	 */
7952 	if (!PP_ISNORELOC(pp))
7953 		panic("Illegal VA->PA translation, pp 0x%p not permanent",
7954 		    (void *)pp);
7955 	else
7956 		panic("Illegal VA->PA translation, pp 0x%p not locked",
7957 		    (void *)pp);
7958 }
7959 #endif	/* DEBUG */
7960 
7961 /*
7962  * Returns a page frame number for a given virtual address.
7963  * Returns PFN_INVALID to indicate an invalid mapping
7964  */
7965 pfn_t
7966 hat_getpfnum(struct hat *hat, caddr_t addr)
7967 {
7968 	pfn_t pfn;
7969 	tte_t tte;
7970 
7971 	/*
7972 	 * We would like to
7973 	 * ASSERT(AS_LOCK_HELD(as));
7974 	 * but we can't because the iommu driver will call this
7975 	 * routine at interrupt time and it can't grab the as lock
7976 	 * or it will deadlock: A thread could have the as lock
7977 	 * and be waiting for io.  The io can't complete
7978 	 * because the interrupt thread is blocked trying to grab
7979 	 * the as lock.
7980 	 */
7981 
7982 	ASSERT(hat->sfmmu_xhat_provider == NULL);
7983 
7984 	if (hat == ksfmmup) {
7985 		if (IS_KMEM_VA_LARGEPAGE(addr)) {
7986 			ASSERT(segkmem_lpszc > 0);
7987 			pfn = sfmmu_kvaszc2pfn(addr, segkmem_lpszc);
7988 			if (pfn != PFN_INVALID) {
7989 				sfmmu_check_kpfn(pfn);
7990 				return (pfn);
7991 			}
7992 		} else if (segkpm && IS_KPM_ADDR(addr)) {
7993 			return (sfmmu_kpm_vatopfn(addr));
7994 		}
7995 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
7996 		    == PFN_SUSPENDED) {
7997 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
7998 		}
7999 		sfmmu_check_kpfn(pfn);
8000 		return (pfn);
8001 	} else {
8002 		return (sfmmu_uvatopfn(addr, hat, NULL));
8003 	}
8004 }
8005 
8006 /*
8007  * This routine will return both pfn and tte for the vaddr.
8008  */
8009 static pfn_t
8010 sfmmu_uvatopfn(caddr_t vaddr, struct hat *sfmmup, tte_t *ttep)
8011 {
8012 	struct hmehash_bucket *hmebp;
8013 	hmeblk_tag hblktag;
8014 	int hmeshift, hashno = 1;
8015 	struct hme_blk *hmeblkp = NULL;
8016 	tte_t tte;
8017 
8018 	struct sf_hment *sfhmep;
8019 	pfn_t pfn;
8020 
8021 	/* support for ISM */
8022 	ism_map_t	*ism_map;
8023 	ism_blk_t	*ism_blkp;
8024 	int		i;
8025 	sfmmu_t *ism_hatid = NULL;
8026 	sfmmu_t *locked_hatid = NULL;
8027 	sfmmu_t	*sv_sfmmup = sfmmup;
8028 	caddr_t	sv_vaddr = vaddr;
8029 	sf_srd_t *srdp;
8030 
8031 	if (ttep == NULL) {
8032 		ttep = &tte;
8033 	} else {
8034 		ttep->ll = 0;
8035 	}
8036 
8037 	ASSERT(sfmmup != ksfmmup);
8038 	SFMMU_STAT(sf_user_vtop);
8039 	/*
8040 	 * Set ism_hatid if vaddr falls in a ISM segment.
8041 	 */
8042 	ism_blkp = sfmmup->sfmmu_iblk;
8043 	if (ism_blkp != NULL) {
8044 		sfmmu_ismhat_enter(sfmmup, 0);
8045 		locked_hatid = sfmmup;
8046 	}
8047 	while (ism_blkp != NULL && ism_hatid == NULL) {
8048 		ism_map = ism_blkp->iblk_maps;
8049 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
8050 			if (vaddr >= ism_start(ism_map[i]) &&
8051 			    vaddr < ism_end(ism_map[i])) {
8052 				sfmmup = ism_hatid = ism_map[i].imap_ismhat;
8053 				vaddr = (caddr_t)(vaddr -
8054 				    ism_start(ism_map[i]));
8055 				break;
8056 			}
8057 		}
8058 		ism_blkp = ism_blkp->iblk_next;
8059 	}
8060 	if (locked_hatid) {
8061 		sfmmu_ismhat_exit(locked_hatid, 0);
8062 	}
8063 
8064 	hblktag.htag_id = sfmmup;
8065 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
8066 	do {
8067 		hmeshift = HME_HASH_SHIFT(hashno);
8068 		hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
8069 		hblktag.htag_rehash = hashno;
8070 		hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
8071 
8072 		SFMMU_HASH_LOCK(hmebp);
8073 
8074 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
8075 		if (hmeblkp != NULL) {
8076 			ASSERT(!hmeblkp->hblk_shared);
8077 			HBLKTOHME(sfhmep, hmeblkp, vaddr);
8078 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
8079 			SFMMU_HASH_UNLOCK(hmebp);
8080 			if (TTE_IS_VALID(ttep)) {
8081 				pfn = TTE_TO_PFN(vaddr, ttep);
8082 				return (pfn);
8083 			}
8084 			break;
8085 		}
8086 		SFMMU_HASH_UNLOCK(hmebp);
8087 		hashno++;
8088 	} while (HME_REHASH(sfmmup) && (hashno <= mmu_hashcnt));
8089 
8090 	if (SF_HMERGNMAP_ISNULL(sv_sfmmup)) {
8091 		return (PFN_INVALID);
8092 	}
8093 	srdp = sv_sfmmup->sfmmu_srdp;
8094 	ASSERT(srdp != NULL);
8095 	ASSERT(srdp->srd_refcnt != 0);
8096 	hblktag.htag_id = srdp;
8097 	hashno = 1;
8098 	do {
8099 		hmeshift = HME_HASH_SHIFT(hashno);
8100 		hblktag.htag_bspage = HME_HASH_BSPAGE(sv_vaddr, hmeshift);
8101 		hblktag.htag_rehash = hashno;
8102 		hmebp = HME_HASH_FUNCTION(srdp, sv_vaddr, hmeshift);
8103 
8104 		SFMMU_HASH_LOCK(hmebp);
8105 		for (hmeblkp = hmebp->hmeblkp; hmeblkp != NULL;
8106 		    hmeblkp = hmeblkp->hblk_next) {
8107 			uint_t rid;
8108 			sf_region_t *rgnp;
8109 			caddr_t rsaddr;
8110 			caddr_t readdr;
8111 
8112 			if (!HTAGS_EQ_SHME(hmeblkp->hblk_tag, hblktag,
8113 			    sv_sfmmup->sfmmu_hmeregion_map)) {
8114 				continue;
8115 			}
8116 			ASSERT(hmeblkp->hblk_shared);
8117 			rid = hmeblkp->hblk_tag.htag_rid;
8118 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8119 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8120 			rgnp = srdp->srd_hmergnp[rid];
8121 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
8122 			HBLKTOHME(sfhmep, hmeblkp, sv_vaddr);
8123 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
8124 			rsaddr = rgnp->rgn_saddr;
8125 			readdr = rsaddr + rgnp->rgn_size;
8126 #ifdef DEBUG
8127 			if (TTE_IS_VALID(ttep) ||
8128 			    get_hblk_ttesz(hmeblkp) > TTE8K) {
8129 				caddr_t eva = tte_to_evaddr(hmeblkp, ttep);
8130 				ASSERT(eva > sv_vaddr);
8131 				ASSERT(sv_vaddr >= rsaddr);
8132 				ASSERT(sv_vaddr < readdr);
8133 				ASSERT(eva <= readdr);
8134 			}
8135 #endif /* DEBUG */
8136 			/*
8137 			 * Continue the search if we
8138 			 * found an invalid 8K tte outside of the area
8139 			 * covered by this hmeblk's region.
8140 			 */
8141 			if (TTE_IS_VALID(ttep)) {
8142 				SFMMU_HASH_UNLOCK(hmebp);
8143 				pfn = TTE_TO_PFN(sv_vaddr, ttep);
8144 				return (pfn);
8145 			} else if (get_hblk_ttesz(hmeblkp) > TTE8K ||
8146 			    (sv_vaddr >= rsaddr && sv_vaddr < readdr)) {
8147 				SFMMU_HASH_UNLOCK(hmebp);
8148 				pfn = PFN_INVALID;
8149 				return (pfn);
8150 			}
8151 		}
8152 		SFMMU_HASH_UNLOCK(hmebp);
8153 		hashno++;
8154 	} while (hashno <= mmu_hashcnt);
8155 	return (PFN_INVALID);
8156 }
8157 
8158 
8159 /*
8160  * For compatability with AT&T and later optimizations
8161  */
8162 /* ARGSUSED */
8163 void
8164 hat_map(struct hat *hat, caddr_t addr, size_t len, uint_t flags)
8165 {
8166 	ASSERT(hat != NULL);
8167 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8168 }
8169 
8170 /*
8171  * Return the number of mappings to a particular page.  This number is an
8172  * approximation of the number of people sharing the page.
8173  *
8174  * shared hmeblks or ism hmeblks are counted as 1 mapping here.
8175  * hat_page_checkshare() can be used to compare threshold to share
8176  * count that reflects the number of region sharers albeit at higher cost.
8177  */
8178 ulong_t
8179 hat_page_getshare(page_t *pp)
8180 {
8181 	page_t *spp = pp;	/* start page */
8182 	kmutex_t *pml;
8183 	ulong_t	cnt;
8184 	int index, sz = TTE64K;
8185 
8186 	/*
8187 	 * We need to grab the mlist lock to make sure any outstanding
8188 	 * load/unloads complete.  Otherwise we could return zero
8189 	 * even though the unload(s) hasn't finished yet.
8190 	 */
8191 	pml = sfmmu_mlist_enter(spp);
8192 	cnt = spp->p_share;
8193 
8194 #ifdef VAC
8195 	if (kpm_enable)
8196 		cnt += spp->p_kpmref;
8197 #endif
8198 	if (vpm_enable && pp->p_vpmref) {
8199 		cnt += 1;
8200 	}
8201 
8202 	/*
8203 	 * If we have any large mappings, we count the number of
8204 	 * mappings that this large page is part of.
8205 	 */
8206 	index = PP_MAPINDEX(spp);
8207 	index >>= 1;
8208 	while (index) {
8209 		pp = PP_GROUPLEADER(spp, sz);
8210 		if ((index & 0x1) && pp != spp) {
8211 			cnt += pp->p_share;
8212 			spp = pp;
8213 		}
8214 		index >>= 1;
8215 		sz++;
8216 	}
8217 	sfmmu_mlist_exit(pml);
8218 	return (cnt);
8219 }
8220 
8221 /*
8222  * Return 1 if the number of mappings exceeds sh_thresh. Return 0
8223  * otherwise. Count shared hmeblks by region's refcnt.
8224  */
8225 int
8226 hat_page_checkshare(page_t *pp, ulong_t sh_thresh)
8227 {
8228 	kmutex_t *pml;
8229 	ulong_t	cnt = 0;
8230 	int index, sz = TTE8K;
8231 	struct sf_hment *sfhme, *tmphme = NULL;
8232 	struct hme_blk *hmeblkp;
8233 
8234 	pml = sfmmu_mlist_enter(pp);
8235 
8236 #ifdef VAC
8237 	if (kpm_enable)
8238 		cnt = pp->p_kpmref;
8239 #endif
8240 
8241 	if (vpm_enable && pp->p_vpmref) {
8242 		cnt += 1;
8243 	}
8244 
8245 	if (pp->p_share + cnt > sh_thresh) {
8246 		sfmmu_mlist_exit(pml);
8247 		return (1);
8248 	}
8249 
8250 	index = PP_MAPINDEX(pp);
8251 
8252 again:
8253 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
8254 		tmphme = sfhme->hme_next;
8255 		if (IS_PAHME(sfhme)) {
8256 			continue;
8257 		}
8258 
8259 		hmeblkp = sfmmu_hmetohblk(sfhme);
8260 		if (hmeblkp->hblk_xhat_bit) {
8261 			cnt++;
8262 			if (cnt > sh_thresh) {
8263 				sfmmu_mlist_exit(pml);
8264 				return (1);
8265 			}
8266 			continue;
8267 		}
8268 		if (hme_size(sfhme) != sz) {
8269 			continue;
8270 		}
8271 
8272 		if (hmeblkp->hblk_shared) {
8273 			sf_srd_t *srdp = hblktosrd(hmeblkp);
8274 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
8275 			sf_region_t *rgnp;
8276 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8277 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8278 			ASSERT(srdp != NULL);
8279 			rgnp = srdp->srd_hmergnp[rid];
8280 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
8281 			    rgnp, rid);
8282 			cnt += rgnp->rgn_refcnt;
8283 		} else {
8284 			cnt++;
8285 		}
8286 		if (cnt > sh_thresh) {
8287 			sfmmu_mlist_exit(pml);
8288 			return (1);
8289 		}
8290 	}
8291 
8292 	index >>= 1;
8293 	sz++;
8294 	while (index) {
8295 		pp = PP_GROUPLEADER(pp, sz);
8296 		ASSERT(sfmmu_mlist_held(pp));
8297 		if (index & 0x1) {
8298 			goto again;
8299 		}
8300 		index >>= 1;
8301 		sz++;
8302 	}
8303 	sfmmu_mlist_exit(pml);
8304 	return (0);
8305 }
8306 
8307 /*
8308  * Unload all large mappings to the pp and reset the p_szc field of every
8309  * constituent page according to the remaining mappings.
8310  *
8311  * pp must be locked SE_EXCL. Even though no other constituent pages are
8312  * locked it's legal to unload the large mappings to the pp because all
8313  * constituent pages of large locked mappings have to be locked SE_SHARED.
8314  * This means if we have SE_EXCL lock on one of constituent pages none of the
8315  * large mappings to pp are locked.
8316  *
8317  * Decrease p_szc field starting from the last constituent page and ending
8318  * with the root page. This method is used because other threads rely on the
8319  * root's p_szc to find the lock to syncronize on. After a root page_t's p_szc
8320  * is demoted then other threads will succeed in sfmmu_mlspl_enter(). This
8321  * ensures that p_szc changes of the constituent pages appears atomic for all
8322  * threads that use sfmmu_mlspl_enter() to examine p_szc field.
8323  *
8324  * This mechanism is only used for file system pages where it's not always
8325  * possible to get SE_EXCL locks on all constituent pages to demote the size
8326  * code (as is done for anonymous or kernel large pages).
8327  *
8328  * See more comments in front of sfmmu_mlspl_enter().
8329  */
8330 void
8331 hat_page_demote(page_t *pp)
8332 {
8333 	int index;
8334 	int sz;
8335 	cpuset_t cpuset;
8336 	int sync = 0;
8337 	page_t *rootpp;
8338 	struct sf_hment *sfhme;
8339 	struct sf_hment *tmphme = NULL;
8340 	struct hme_blk *hmeblkp;
8341 	uint_t pszc;
8342 	page_t *lastpp;
8343 	cpuset_t tset;
8344 	pgcnt_t npgs;
8345 	kmutex_t *pml;
8346 	kmutex_t *pmtx = NULL;
8347 
8348 	ASSERT(PAGE_EXCL(pp));
8349 	ASSERT(!PP_ISFREE(pp));
8350 	ASSERT(!PP_ISKAS(pp));
8351 	ASSERT(page_szc_lock_assert(pp));
8352 	pml = sfmmu_mlist_enter(pp);
8353 
8354 	pszc = pp->p_szc;
8355 	if (pszc == 0) {
8356 		goto out;
8357 	}
8358 
8359 	index = PP_MAPINDEX(pp) >> 1;
8360 
8361 	if (index) {
8362 		CPUSET_ZERO(cpuset);
8363 		sz = TTE64K;
8364 		sync = 1;
8365 	}
8366 
8367 	while (index) {
8368 		if (!(index & 0x1)) {
8369 			index >>= 1;
8370 			sz++;
8371 			continue;
8372 		}
8373 		ASSERT(sz <= pszc);
8374 		rootpp = PP_GROUPLEADER(pp, sz);
8375 		for (sfhme = rootpp->p_mapping; sfhme; sfhme = tmphme) {
8376 			tmphme = sfhme->hme_next;
8377 			ASSERT(!IS_PAHME(sfhme));
8378 			hmeblkp = sfmmu_hmetohblk(sfhme);
8379 			if (hme_size(sfhme) != sz) {
8380 				continue;
8381 			}
8382 			if (hmeblkp->hblk_xhat_bit) {
8383 				cmn_err(CE_PANIC,
8384 				    "hat_page_demote: xhat hmeblk");
8385 			}
8386 			tset = sfmmu_pageunload(rootpp, sfhme, sz);
8387 			CPUSET_OR(cpuset, tset);
8388 		}
8389 		if (index >>= 1) {
8390 			sz++;
8391 		}
8392 	}
8393 
8394 	ASSERT(!PP_ISMAPPED_LARGE(pp));
8395 
8396 	if (sync) {
8397 		xt_sync(cpuset);
8398 #ifdef VAC
8399 		if (PP_ISTNC(pp)) {
8400 			conv_tnc(rootpp, sz);
8401 		}
8402 #endif	/* VAC */
8403 	}
8404 
8405 	pmtx = sfmmu_page_enter(pp);
8406 
8407 	ASSERT(pp->p_szc == pszc);
8408 	rootpp = PP_PAGEROOT(pp);
8409 	ASSERT(rootpp->p_szc == pszc);
8410 	lastpp = PP_PAGENEXT_N(rootpp, TTEPAGES(pszc) - 1);
8411 
8412 	while (lastpp != rootpp) {
8413 		sz = PP_MAPINDEX(lastpp) ? fnd_mapping_sz(lastpp) : 0;
8414 		ASSERT(sz < pszc);
8415 		npgs = (sz == 0) ? 1 : TTEPAGES(sz);
8416 		ASSERT(P2PHASE(lastpp->p_pagenum, npgs) == npgs - 1);
8417 		while (--npgs > 0) {
8418 			lastpp->p_szc = (uchar_t)sz;
8419 			lastpp = PP_PAGEPREV(lastpp);
8420 		}
8421 		if (sz) {
8422 			/*
8423 			 * make sure before current root's pszc
8424 			 * is updated all updates to constituent pages pszc
8425 			 * fields are globally visible.
8426 			 */
8427 			membar_producer();
8428 		}
8429 		lastpp->p_szc = sz;
8430 		ASSERT(IS_P2ALIGNED(lastpp->p_pagenum, TTEPAGES(sz)));
8431 		if (lastpp != rootpp) {
8432 			lastpp = PP_PAGEPREV(lastpp);
8433 		}
8434 	}
8435 	if (sz == 0) {
8436 		/* the loop above doesn't cover this case */
8437 		rootpp->p_szc = 0;
8438 	}
8439 out:
8440 	ASSERT(pp->p_szc == 0);
8441 	if (pmtx != NULL) {
8442 		sfmmu_page_exit(pmtx);
8443 	}
8444 	sfmmu_mlist_exit(pml);
8445 }
8446 
8447 /*
8448  * Refresh the HAT ismttecnt[] element for size szc.
8449  * Caller must have set ISM busy flag to prevent mapping
8450  * lists from changing while we're traversing them.
8451  */
8452 pgcnt_t
8453 ism_tsb_entries(sfmmu_t *sfmmup, int szc)
8454 {
8455 	ism_blk_t	*ism_blkp = sfmmup->sfmmu_iblk;
8456 	ism_map_t	*ism_map;
8457 	pgcnt_t		npgs = 0;
8458 	pgcnt_t		npgs_scd = 0;
8459 	int		j;
8460 	sf_scd_t	*scdp;
8461 	uchar_t		rid;
8462 
8463 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
8464 	scdp = sfmmup->sfmmu_scdp;
8465 
8466 	for (; ism_blkp != NULL; ism_blkp = ism_blkp->iblk_next) {
8467 		ism_map = ism_blkp->iblk_maps;
8468 		for (j = 0; ism_map[j].imap_ismhat && j < ISM_MAP_SLOTS; j++) {
8469 			rid = ism_map[j].imap_rid;
8470 			ASSERT(rid == SFMMU_INVALID_ISMRID ||
8471 			    rid < sfmmup->sfmmu_srdp->srd_next_ismrid);
8472 
8473 			if (scdp != NULL && rid != SFMMU_INVALID_ISMRID &&
8474 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
8475 				/* ISM is in sfmmup's SCD */
8476 				npgs_scd +=
8477 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8478 			} else {
8479 				/* ISMs is not in SCD */
8480 				npgs +=
8481 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8482 			}
8483 		}
8484 	}
8485 	sfmmup->sfmmu_ismttecnt[szc] = npgs;
8486 	sfmmup->sfmmu_scdismttecnt[szc] = npgs_scd;
8487 	return (npgs);
8488 }
8489 
8490 /*
8491  * Yield the memory claim requirement for an address space.
8492  *
8493  * This is currently implemented as the number of bytes that have active
8494  * hardware translations that have page structures.  Therefore, it can
8495  * underestimate the traditional resident set size, eg, if the
8496  * physical page is present and the hardware translation is missing;
8497  * and it can overestimate the rss, eg, if there are active
8498  * translations to a frame buffer with page structs.
8499  * Also, it does not take sharing into account.
8500  *
8501  * Note that we don't acquire locks here since this function is most often
8502  * called from the clock thread.
8503  */
8504 size_t
8505 hat_get_mapped_size(struct hat *hat)
8506 {
8507 	size_t		assize = 0;
8508 	int 		i;
8509 
8510 	if (hat == NULL)
8511 		return (0);
8512 
8513 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8514 
8515 	for (i = 0; i < mmu_page_sizes; i++)
8516 		assize += ((pgcnt_t)hat->sfmmu_ttecnt[i] +
8517 		    (pgcnt_t)hat->sfmmu_scdrttecnt[i]) * TTEBYTES(i);
8518 
8519 	if (hat->sfmmu_iblk == NULL)
8520 		return (assize);
8521 
8522 	for (i = 0; i < mmu_page_sizes; i++)
8523 		assize += ((pgcnt_t)hat->sfmmu_ismttecnt[i] +
8524 		    (pgcnt_t)hat->sfmmu_scdismttecnt[i]) * TTEBYTES(i);
8525 
8526 	return (assize);
8527 }
8528 
8529 int
8530 hat_stats_enable(struct hat *hat)
8531 {
8532 	hatlock_t	*hatlockp;
8533 
8534 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8535 
8536 	hatlockp = sfmmu_hat_enter(hat);
8537 	hat->sfmmu_rmstat++;
8538 	sfmmu_hat_exit(hatlockp);
8539 	return (1);
8540 }
8541 
8542 void
8543 hat_stats_disable(struct hat *hat)
8544 {
8545 	hatlock_t	*hatlockp;
8546 
8547 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8548 
8549 	hatlockp = sfmmu_hat_enter(hat);
8550 	hat->sfmmu_rmstat--;
8551 	sfmmu_hat_exit(hatlockp);
8552 }
8553 
8554 /*
8555  * Routines for entering or removing  ourselves from the
8556  * ism_hat's mapping list. This is used for both private and
8557  * SCD hats.
8558  */
8559 static void
8560 iment_add(struct ism_ment *iment,  struct hat *ism_hat)
8561 {
8562 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8563 
8564 	iment->iment_prev = NULL;
8565 	iment->iment_next = ism_hat->sfmmu_iment;
8566 	if (ism_hat->sfmmu_iment) {
8567 		ism_hat->sfmmu_iment->iment_prev = iment;
8568 	}
8569 	ism_hat->sfmmu_iment = iment;
8570 }
8571 
8572 static void
8573 iment_sub(struct ism_ment *iment, struct hat *ism_hat)
8574 {
8575 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8576 
8577 	if (ism_hat->sfmmu_iment == NULL) {
8578 		panic("ism map entry remove - no entries");
8579 	}
8580 
8581 	if (iment->iment_prev) {
8582 		ASSERT(ism_hat->sfmmu_iment != iment);
8583 		iment->iment_prev->iment_next = iment->iment_next;
8584 	} else {
8585 		ASSERT(ism_hat->sfmmu_iment == iment);
8586 		ism_hat->sfmmu_iment = iment->iment_next;
8587 	}
8588 
8589 	if (iment->iment_next) {
8590 		iment->iment_next->iment_prev = iment->iment_prev;
8591 	}
8592 
8593 	/*
8594 	 * zero out the entry
8595 	 */
8596 	iment->iment_next = NULL;
8597 	iment->iment_prev = NULL;
8598 	iment->iment_hat =  NULL;
8599 	iment->iment_base_va = 0;
8600 }
8601 
8602 /*
8603  * Hat_share()/unshare() return an (non-zero) error
8604  * when saddr and daddr are not properly aligned.
8605  *
8606  * The top level mapping element determines the alignment
8607  * requirement for saddr and daddr, depending on different
8608  * architectures.
8609  *
8610  * When hat_share()/unshare() are not supported,
8611  * HATOP_SHARE()/UNSHARE() return 0
8612  */
8613 int
8614 hat_share(struct hat *sfmmup, caddr_t addr,
8615 	struct hat *ism_hatid, caddr_t sptaddr, size_t len, uint_t ismszc)
8616 {
8617 	ism_blk_t	*ism_blkp;
8618 	ism_blk_t	*new_iblk;
8619 	ism_map_t 	*ism_map;
8620 	ism_ment_t	*ism_ment;
8621 	int		i, added;
8622 	hatlock_t	*hatlockp;
8623 	int		reload_mmu = 0;
8624 	uint_t		ismshift = page_get_shift(ismszc);
8625 	size_t		ismpgsz = page_get_pagesize(ismszc);
8626 	uint_t		ismmask = (uint_t)ismpgsz - 1;
8627 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8628 	ushort_t	ismhatflag;
8629 	hat_region_cookie_t rcookie;
8630 	sf_scd_t	*old_scdp;
8631 
8632 #ifdef DEBUG
8633 	caddr_t		eaddr = addr + len;
8634 #endif /* DEBUG */
8635 
8636 	ASSERT(ism_hatid != NULL && sfmmup != NULL);
8637 	ASSERT(sptaddr == ISMID_STARTADDR);
8638 	/*
8639 	 * Check the alignment.
8640 	 */
8641 	if (!ISM_ALIGNED(ismshift, addr) || !ISM_ALIGNED(ismshift, sptaddr))
8642 		return (EINVAL);
8643 
8644 	/*
8645 	 * Check size alignment.
8646 	 */
8647 	if (!ISM_ALIGNED(ismshift, len))
8648 		return (EINVAL);
8649 
8650 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
8651 
8652 	/*
8653 	 * Allocate ism_ment for the ism_hat's mapping list, and an
8654 	 * ism map blk in case we need one.  We must do our
8655 	 * allocations before acquiring locks to prevent a deadlock
8656 	 * in the kmem allocator on the mapping list lock.
8657 	 */
8658 	new_iblk = kmem_cache_alloc(ism_blk_cache, KM_SLEEP);
8659 	ism_ment = kmem_cache_alloc(ism_ment_cache, KM_SLEEP);
8660 
8661 	/*
8662 	 * Serialize ISM mappings with the ISM busy flag, and also the
8663 	 * trap handlers.
8664 	 */
8665 	sfmmu_ismhat_enter(sfmmup, 0);
8666 
8667 	/*
8668 	 * Allocate an ism map blk if necessary.
8669 	 */
8670 	if (sfmmup->sfmmu_iblk == NULL) {
8671 		sfmmup->sfmmu_iblk = new_iblk;
8672 		bzero(new_iblk, sizeof (*new_iblk));
8673 		new_iblk->iblk_nextpa = (uint64_t)-1;
8674 		membar_stst();	/* make sure next ptr visible to all CPUs */
8675 		sfmmup->sfmmu_ismblkpa = va_to_pa((caddr_t)new_iblk);
8676 		reload_mmu = 1;
8677 		new_iblk = NULL;
8678 	}
8679 
8680 #ifdef DEBUG
8681 	/*
8682 	 * Make sure mapping does not already exist.
8683 	 */
8684 	ism_blkp = sfmmup->sfmmu_iblk;
8685 	while (ism_blkp != NULL) {
8686 		ism_map = ism_blkp->iblk_maps;
8687 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
8688 			if ((addr >= ism_start(ism_map[i]) &&
8689 			    addr < ism_end(ism_map[i])) ||
8690 			    eaddr > ism_start(ism_map[i]) &&
8691 			    eaddr <= ism_end(ism_map[i])) {
8692 				panic("sfmmu_share: Already mapped!");
8693 			}
8694 		}
8695 		ism_blkp = ism_blkp->iblk_next;
8696 	}
8697 #endif /* DEBUG */
8698 
8699 	ASSERT(ismszc >= TTE4M);
8700 	if (ismszc == TTE4M) {
8701 		ismhatflag = HAT_4M_FLAG;
8702 	} else if (ismszc == TTE32M) {
8703 		ismhatflag = HAT_32M_FLAG;
8704 	} else if (ismszc == TTE256M) {
8705 		ismhatflag = HAT_256M_FLAG;
8706 	}
8707 	/*
8708 	 * Add mapping to first available mapping slot.
8709 	 */
8710 	ism_blkp = sfmmup->sfmmu_iblk;
8711 	added = 0;
8712 	while (!added) {
8713 		ism_map = ism_blkp->iblk_maps;
8714 		for (i = 0; i < ISM_MAP_SLOTS; i++)  {
8715 			if (ism_map[i].imap_ismhat == NULL) {
8716 
8717 				ism_map[i].imap_ismhat = ism_hatid;
8718 				ism_map[i].imap_vb_shift = (uchar_t)ismshift;
8719 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8720 				ism_map[i].imap_hatflags = ismhatflag;
8721 				ism_map[i].imap_sz_mask = ismmask;
8722 				/*
8723 				 * imap_seg is checked in ISM_CHECK to see if
8724 				 * non-NULL, then other info assumed valid.
8725 				 */
8726 				membar_stst();
8727 				ism_map[i].imap_seg = (uintptr_t)addr | sh_size;
8728 				ism_map[i].imap_ment = ism_ment;
8729 
8730 				/*
8731 				 * Now add ourselves to the ism_hat's
8732 				 * mapping list.
8733 				 */
8734 				ism_ment->iment_hat = sfmmup;
8735 				ism_ment->iment_base_va = addr;
8736 				ism_hatid->sfmmu_ismhat = 1;
8737 				mutex_enter(&ism_mlist_lock);
8738 				iment_add(ism_ment, ism_hatid);
8739 				mutex_exit(&ism_mlist_lock);
8740 				added = 1;
8741 				break;
8742 			}
8743 		}
8744 		if (!added && ism_blkp->iblk_next == NULL) {
8745 			ism_blkp->iblk_next = new_iblk;
8746 			new_iblk = NULL;
8747 			bzero(ism_blkp->iblk_next,
8748 			    sizeof (*ism_blkp->iblk_next));
8749 			ism_blkp->iblk_next->iblk_nextpa = (uint64_t)-1;
8750 			membar_stst();
8751 			ism_blkp->iblk_nextpa =
8752 			    va_to_pa((caddr_t)ism_blkp->iblk_next);
8753 		}
8754 		ism_blkp = ism_blkp->iblk_next;
8755 	}
8756 
8757 	/*
8758 	 * After calling hat_join_region, sfmmup may join a new SCD or
8759 	 * move from the old scd to a new scd, in which case, we want to
8760 	 * shrink the sfmmup's private tsb size, i.e., pass shrink to
8761 	 * sfmmu_check_page_sizes at the end of this routine.
8762 	 */
8763 	old_scdp = sfmmup->sfmmu_scdp;
8764 
8765 	rcookie = hat_join_region(sfmmup, addr, len, (void *)ism_hatid, 0,
8766 	    PROT_ALL, ismszc, NULL, HAT_REGION_ISM);
8767 	if (rcookie != HAT_INVALID_REGION_COOKIE) {
8768 		ism_map[i].imap_rid = (uchar_t)((uint64_t)rcookie);
8769 	}
8770 	/*
8771 	 * Update our counters for this sfmmup's ism mappings.
8772 	 */
8773 	for (i = 0; i <= ismszc; i++) {
8774 		if (!(disable_ism_large_pages & (1 << i)))
8775 			(void) ism_tsb_entries(sfmmup, i);
8776 	}
8777 
8778 	/*
8779 	 * For ISM and DISM we do not support 512K pages, so we only only
8780 	 * search the 4M and 8K/64K hashes for 4 pagesize cpus, and search the
8781 	 * 256M or 32M, and 4M and 8K/64K hashes for 6 pagesize cpus.
8782 	 *
8783 	 * Need to set 32M/256M ISM flags to make sure
8784 	 * sfmmu_check_page_sizes() enables them on Panther.
8785 	 */
8786 	ASSERT((disable_ism_large_pages & (1 << TTE512K)) != 0);
8787 
8788 	switch (ismszc) {
8789 	case TTE256M:
8790 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_256M_ISM)) {
8791 			hatlockp = sfmmu_hat_enter(sfmmup);
8792 			SFMMU_FLAGS_SET(sfmmup, HAT_256M_ISM);
8793 			sfmmu_hat_exit(hatlockp);
8794 		}
8795 		break;
8796 	case TTE32M:
8797 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_32M_ISM)) {
8798 			hatlockp = sfmmu_hat_enter(sfmmup);
8799 			SFMMU_FLAGS_SET(sfmmup, HAT_32M_ISM);
8800 			sfmmu_hat_exit(hatlockp);
8801 		}
8802 		break;
8803 	default:
8804 		break;
8805 	}
8806 
8807 	/*
8808 	 * If we updated the ismblkpa for this HAT we must make
8809 	 * sure all CPUs running this process reload their tsbmiss area.
8810 	 * Otherwise they will fail to load the mappings in the tsbmiss
8811 	 * handler and will loop calling pagefault().
8812 	 */
8813 	if (reload_mmu) {
8814 		hatlockp = sfmmu_hat_enter(sfmmup);
8815 		sfmmu_sync_mmustate(sfmmup);
8816 		sfmmu_hat_exit(hatlockp);
8817 	}
8818 
8819 	sfmmu_ismhat_exit(sfmmup, 0);
8820 
8821 	/*
8822 	 * Free up ismblk if we didn't use it.
8823 	 */
8824 	if (new_iblk != NULL)
8825 		kmem_cache_free(ism_blk_cache, new_iblk);
8826 
8827 	/*
8828 	 * Check TSB and TLB page sizes.
8829 	 */
8830 	if (sfmmup->sfmmu_scdp != NULL && old_scdp != sfmmup->sfmmu_scdp) {
8831 		sfmmu_check_page_sizes(sfmmup, 0);
8832 	} else {
8833 		sfmmu_check_page_sizes(sfmmup, 1);
8834 	}
8835 	return (0);
8836 }
8837 
8838 /*
8839  * hat_unshare removes exactly one ism_map from
8840  * this process's as.  It expects multiple calls
8841  * to hat_unshare for multiple shm segments.
8842  */
8843 void
8844 hat_unshare(struct hat *sfmmup, caddr_t addr, size_t len, uint_t ismszc)
8845 {
8846 	ism_map_t 	*ism_map;
8847 	ism_ment_t	*free_ment = NULL;
8848 	ism_blk_t	*ism_blkp;
8849 	struct hat	*ism_hatid;
8850 	int 		found, i;
8851 	hatlock_t	*hatlockp;
8852 	struct tsb_info	*tsbinfo;
8853 	uint_t		ismshift = page_get_shift(ismszc);
8854 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8855 	uchar_t		ism_rid;
8856 	sf_scd_t	*old_scdp;
8857 
8858 	ASSERT(ISM_ALIGNED(ismshift, addr));
8859 	ASSERT(ISM_ALIGNED(ismshift, len));
8860 	ASSERT(sfmmup != NULL);
8861 	ASSERT(sfmmup != ksfmmup);
8862 
8863 	if (sfmmup->sfmmu_xhat_provider) {
8864 		XHAT_UNSHARE(sfmmup, addr, len);
8865 		return;
8866 	} else {
8867 		/*
8868 		 * This must be a CPU HAT. If the address space has
8869 		 * XHATs attached, inform all XHATs that ISM segment
8870 		 * is going away
8871 		 */
8872 		ASSERT(sfmmup->sfmmu_as != NULL);
8873 		if (sfmmup->sfmmu_as->a_xhat != NULL)
8874 			xhat_unshare_all(sfmmup->sfmmu_as, addr, len);
8875 	}
8876 
8877 	/*
8878 	 * Make sure that during the entire time ISM mappings are removed,
8879 	 * the trap handlers serialize behind us, and that no one else
8880 	 * can be mucking with ISM mappings.  This also lets us get away
8881 	 * with not doing expensive cross calls to flush the TLB -- we
8882 	 * just discard the context, flush the entire TSB, and call it
8883 	 * a day.
8884 	 */
8885 	sfmmu_ismhat_enter(sfmmup, 0);
8886 
8887 	/*
8888 	 * Remove the mapping.
8889 	 *
8890 	 * We can't have any holes in the ism map.
8891 	 * The tsb miss code while searching the ism map will
8892 	 * stop on an empty map slot.  So we must move
8893 	 * everyone past the hole up 1 if any.
8894 	 *
8895 	 * Also empty ism map blks are not freed until the
8896 	 * process exits. This is to prevent a MT race condition
8897 	 * between sfmmu_unshare() and sfmmu_tsbmiss_exception().
8898 	 */
8899 	found = 0;
8900 	ism_blkp = sfmmup->sfmmu_iblk;
8901 	while (!found && ism_blkp != NULL) {
8902 		ism_map = ism_blkp->iblk_maps;
8903 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
8904 			if (addr == ism_start(ism_map[i]) &&
8905 			    sh_size == (size_t)(ism_size(ism_map[i]))) {
8906 				found = 1;
8907 				break;
8908 			}
8909 		}
8910 		if (!found)
8911 			ism_blkp = ism_blkp->iblk_next;
8912 	}
8913 
8914 	if (found) {
8915 		ism_hatid = ism_map[i].imap_ismhat;
8916 		ism_rid = ism_map[i].imap_rid;
8917 		ASSERT(ism_hatid != NULL);
8918 		ASSERT(ism_hatid->sfmmu_ismhat == 1);
8919 
8920 		/*
8921 		 * After hat_leave_region, the sfmmup may leave SCD,
8922 		 * in which case, we want to grow the private tsb size when
8923 		 * calling sfmmu_check_page_sizes at the end of the routine.
8924 		 */
8925 		old_scdp = sfmmup->sfmmu_scdp;
8926 		/*
8927 		 * Then remove ourselves from the region.
8928 		 */
8929 		if (ism_rid != SFMMU_INVALID_ISMRID) {
8930 			hat_leave_region(sfmmup, (void *)((uint64_t)ism_rid),
8931 			    HAT_REGION_ISM);
8932 		}
8933 
8934 		/*
8935 		 * And now guarantee that any other cpu
8936 		 * that tries to process an ISM miss
8937 		 * will go to tl=0.
8938 		 */
8939 		hatlockp = sfmmu_hat_enter(sfmmup);
8940 		sfmmu_invalidate_ctx(sfmmup);
8941 		sfmmu_hat_exit(hatlockp);
8942 
8943 		/*
8944 		 * Remove ourselves from the ism mapping list.
8945 		 */
8946 		mutex_enter(&ism_mlist_lock);
8947 		iment_sub(ism_map[i].imap_ment, ism_hatid);
8948 		mutex_exit(&ism_mlist_lock);
8949 		free_ment = ism_map[i].imap_ment;
8950 
8951 		/*
8952 		 * We delete the ism map by copying
8953 		 * the next map over the current one.
8954 		 * We will take the next one in the maps
8955 		 * array or from the next ism_blk.
8956 		 */
8957 		while (ism_blkp != NULL) {
8958 			ism_map = ism_blkp->iblk_maps;
8959 			while (i < (ISM_MAP_SLOTS - 1)) {
8960 				ism_map[i] = ism_map[i + 1];
8961 				i++;
8962 			}
8963 			/* i == (ISM_MAP_SLOTS - 1) */
8964 			ism_blkp = ism_blkp->iblk_next;
8965 			if (ism_blkp != NULL) {
8966 				ism_map[i] = ism_blkp->iblk_maps[0];
8967 				i = 0;
8968 			} else {
8969 				ism_map[i].imap_seg = 0;
8970 				ism_map[i].imap_vb_shift = 0;
8971 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8972 				ism_map[i].imap_hatflags = 0;
8973 				ism_map[i].imap_sz_mask = 0;
8974 				ism_map[i].imap_ismhat = NULL;
8975 				ism_map[i].imap_ment = NULL;
8976 			}
8977 		}
8978 
8979 		/*
8980 		 * Now flush entire TSB for the process, since
8981 		 * demapping page by page can be too expensive.
8982 		 * We don't have to flush the TLB here anymore
8983 		 * since we switch to a new TLB ctx instead.
8984 		 * Also, there is no need to flush if the process
8985 		 * is exiting since the TSB will be freed later.
8986 		 */
8987 		if (!sfmmup->sfmmu_free) {
8988 			hatlockp = sfmmu_hat_enter(sfmmup);
8989 			for (tsbinfo = sfmmup->sfmmu_tsb; tsbinfo != NULL;
8990 			    tsbinfo = tsbinfo->tsb_next) {
8991 				if (tsbinfo->tsb_flags & TSB_SWAPPED)
8992 					continue;
8993 				if (tsbinfo->tsb_flags & TSB_RELOC_FLAG) {
8994 					tsbinfo->tsb_flags |=
8995 					    TSB_FLUSH_NEEDED;
8996 					continue;
8997 				}
8998 
8999 				sfmmu_inv_tsb(tsbinfo->tsb_va,
9000 				    TSB_BYTES(tsbinfo->tsb_szc));
9001 			}
9002 			sfmmu_hat_exit(hatlockp);
9003 		}
9004 	}
9005 
9006 	/*
9007 	 * Update our counters for this sfmmup's ism mappings.
9008 	 */
9009 	for (i = 0; i <= ismszc; i++) {
9010 		if (!(disable_ism_large_pages & (1 << i)))
9011 			(void) ism_tsb_entries(sfmmup, i);
9012 	}
9013 
9014 	sfmmu_ismhat_exit(sfmmup, 0);
9015 
9016 	/*
9017 	 * We must do our freeing here after dropping locks
9018 	 * to prevent a deadlock in the kmem allocator on the
9019 	 * mapping list lock.
9020 	 */
9021 	if (free_ment != NULL)
9022 		kmem_cache_free(ism_ment_cache, free_ment);
9023 
9024 	/*
9025 	 * Check TSB and TLB page sizes if the process isn't exiting.
9026 	 */
9027 	if (!sfmmup->sfmmu_free) {
9028 		if (found && old_scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
9029 			sfmmu_check_page_sizes(sfmmup, 1);
9030 		} else {
9031 			sfmmu_check_page_sizes(sfmmup, 0);
9032 		}
9033 	}
9034 }
9035 
9036 /* ARGSUSED */
9037 static int
9038 sfmmu_idcache_constructor(void *buf, void *cdrarg, int kmflags)
9039 {
9040 	/* void *buf is sfmmu_t pointer */
9041 	bzero(buf, sizeof (sfmmu_t));
9042 
9043 	return (0);
9044 }
9045 
9046 /* ARGSUSED */
9047 static void
9048 sfmmu_idcache_destructor(void *buf, void *cdrarg)
9049 {
9050 	/* void *buf is sfmmu_t pointer */
9051 }
9052 
9053 /*
9054  * setup kmem hmeblks by bzeroing all members and initializing the nextpa
9055  * field to be the pa of this hmeblk
9056  */
9057 /* ARGSUSED */
9058 static int
9059 sfmmu_hblkcache_constructor(void *buf, void *cdrarg, int kmflags)
9060 {
9061 	struct hme_blk *hmeblkp;
9062 
9063 	bzero(buf, (size_t)cdrarg);
9064 	hmeblkp = (struct hme_blk *)buf;
9065 	hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
9066 
9067 #ifdef	HBLK_TRACE
9068 	mutex_init(&hmeblkp->hblk_audit_lock, NULL, MUTEX_DEFAULT, NULL);
9069 #endif	/* HBLK_TRACE */
9070 
9071 	return (0);
9072 }
9073 
9074 /* ARGSUSED */
9075 static void
9076 sfmmu_hblkcache_destructor(void *buf, void *cdrarg)
9077 {
9078 
9079 #ifdef	HBLK_TRACE
9080 
9081 	struct hme_blk *hmeblkp;
9082 
9083 	hmeblkp = (struct hme_blk *)buf;
9084 	mutex_destroy(&hmeblkp->hblk_audit_lock);
9085 
9086 #endif	/* HBLK_TRACE */
9087 }
9088 
9089 #define	SFMMU_CACHE_RECLAIM_SCAN_RATIO 8
9090 static int sfmmu_cache_reclaim_scan_ratio = SFMMU_CACHE_RECLAIM_SCAN_RATIO;
9091 /*
9092  * The kmem allocator will callback into our reclaim routine when the system
9093  * is running low in memory.  We traverse the hash and free up all unused but
9094  * still cached hme_blks.  We also traverse the free list and free them up
9095  * as well.
9096  */
9097 /*ARGSUSED*/
9098 static void
9099 sfmmu_hblkcache_reclaim(void *cdrarg)
9100 {
9101 	int i;
9102 	struct hmehash_bucket *hmebp;
9103 	struct hme_blk *hmeblkp, *nx_hblk, *pr_hblk = NULL;
9104 	static struct hmehash_bucket *uhmehash_reclaim_hand;
9105 	static struct hmehash_bucket *khmehash_reclaim_hand;
9106 	struct hme_blk *list = NULL, *last_hmeblkp;
9107 	cpuset_t cpuset = cpu_ready_set;
9108 	cpu_hme_pend_t *cpuhp;
9109 
9110 	/* Free up hmeblks on the cpu pending lists */
9111 	for (i = 0; i < NCPU; i++) {
9112 		cpuhp = &cpu_hme_pend[i];
9113 		if (cpuhp->chp_listp != NULL)  {
9114 			mutex_enter(&cpuhp->chp_mutex);
9115 			if (cpuhp->chp_listp == NULL) {
9116 				mutex_exit(&cpuhp->chp_mutex);
9117 				continue;
9118 			}
9119 			for (last_hmeblkp = cpuhp->chp_listp;
9120 			    last_hmeblkp->hblk_next != NULL;
9121 			    last_hmeblkp = last_hmeblkp->hblk_next)
9122 				;
9123 			last_hmeblkp->hblk_next = list;
9124 			list = cpuhp->chp_listp;
9125 			cpuhp->chp_listp = NULL;
9126 			cpuhp->chp_count = 0;
9127 			mutex_exit(&cpuhp->chp_mutex);
9128 		}
9129 
9130 	}
9131 
9132 	if (list != NULL) {
9133 		kpreempt_disable();
9134 		CPUSET_DEL(cpuset, CPU->cpu_id);
9135 		xt_sync(cpuset);
9136 		xt_sync(cpuset);
9137 		kpreempt_enable();
9138 		sfmmu_hblk_free(&list);
9139 		list = NULL;
9140 	}
9141 
9142 	hmebp = uhmehash_reclaim_hand;
9143 	if (hmebp == NULL || hmebp > &uhme_hash[UHMEHASH_SZ])
9144 		uhmehash_reclaim_hand = hmebp = uhme_hash;
9145 	uhmehash_reclaim_hand += UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
9146 
9147 	for (i = UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
9148 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
9149 			hmeblkp = hmebp->hmeblkp;
9150 			pr_hblk = NULL;
9151 			while (hmeblkp) {
9152 				nx_hblk = hmeblkp->hblk_next;
9153 				if (!hmeblkp->hblk_vcnt &&
9154 				    !hmeblkp->hblk_hmecnt) {
9155 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9156 					    pr_hblk, &list, 0);
9157 				} else {
9158 					pr_hblk = hmeblkp;
9159 				}
9160 				hmeblkp = nx_hblk;
9161 			}
9162 			SFMMU_HASH_UNLOCK(hmebp);
9163 		}
9164 		if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
9165 			hmebp = uhme_hash;
9166 	}
9167 
9168 	hmebp = khmehash_reclaim_hand;
9169 	if (hmebp == NULL || hmebp > &khme_hash[KHMEHASH_SZ])
9170 		khmehash_reclaim_hand = hmebp = khme_hash;
9171 	khmehash_reclaim_hand += KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
9172 
9173 	for (i = KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
9174 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
9175 			hmeblkp = hmebp->hmeblkp;
9176 			pr_hblk = NULL;
9177 			while (hmeblkp) {
9178 				nx_hblk = hmeblkp->hblk_next;
9179 				if (!hmeblkp->hblk_vcnt &&
9180 				    !hmeblkp->hblk_hmecnt) {
9181 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9182 					    pr_hblk, &list, 0);
9183 				} else {
9184 					pr_hblk = hmeblkp;
9185 				}
9186 				hmeblkp = nx_hblk;
9187 			}
9188 			SFMMU_HASH_UNLOCK(hmebp);
9189 		}
9190 		if (hmebp++ == &khme_hash[KHMEHASH_SZ])
9191 			hmebp = khme_hash;
9192 	}
9193 	sfmmu_hblks_list_purge(&list, 0);
9194 }
9195 
9196 /*
9197  * sfmmu_get_ppvcolor should become a vm_machdep or hatop interface.
9198  * same goes for sfmmu_get_addrvcolor().
9199  *
9200  * This function will return the virtual color for the specified page. The
9201  * virtual color corresponds to this page current mapping or its last mapping.
9202  * It is used by memory allocators to choose addresses with the correct
9203  * alignment so vac consistency is automatically maintained.  If the page
9204  * has no color it returns -1.
9205  */
9206 /*ARGSUSED*/
9207 int
9208 sfmmu_get_ppvcolor(struct page *pp)
9209 {
9210 #ifdef VAC
9211 	int color;
9212 
9213 	if (!(cache & CACHE_VAC) || PP_NEWPAGE(pp)) {
9214 		return (-1);
9215 	}
9216 	color = PP_GET_VCOLOR(pp);
9217 	ASSERT(color < mmu_btop(shm_alignment));
9218 	return (color);
9219 #else
9220 	return (-1);
9221 #endif	/* VAC */
9222 }
9223 
9224 /*
9225  * This function will return the desired alignment for vac consistency
9226  * (vac color) given a virtual address.  If no vac is present it returns -1.
9227  */
9228 /*ARGSUSED*/
9229 int
9230 sfmmu_get_addrvcolor(caddr_t vaddr)
9231 {
9232 #ifdef VAC
9233 	if (cache & CACHE_VAC) {
9234 		return (addr_to_vcolor(vaddr));
9235 	} else {
9236 		return (-1);
9237 	}
9238 #else
9239 	return (-1);
9240 #endif	/* VAC */
9241 }
9242 
9243 #ifdef VAC
9244 /*
9245  * Check for conflicts.
9246  * A conflict exists if the new and existent mappings do not match in
9247  * their "shm_alignment fields. If conflicts exist, the existant mappings
9248  * are flushed unless one of them is locked. If one of them is locked, then
9249  * the mappings are flushed and converted to non-cacheable mappings.
9250  */
9251 static void
9252 sfmmu_vac_conflict(struct hat *hat, caddr_t addr, page_t *pp)
9253 {
9254 	struct hat *tmphat;
9255 	struct sf_hment *sfhmep, *tmphme = NULL;
9256 	struct hme_blk *hmeblkp;
9257 	int vcolor;
9258 	tte_t tte;
9259 
9260 	ASSERT(sfmmu_mlist_held(pp));
9261 	ASSERT(!PP_ISNC(pp));		/* page better be cacheable */
9262 
9263 	vcolor = addr_to_vcolor(addr);
9264 	if (PP_NEWPAGE(pp)) {
9265 		PP_SET_VCOLOR(pp, vcolor);
9266 		return;
9267 	}
9268 
9269 	if (PP_GET_VCOLOR(pp) == vcolor) {
9270 		return;
9271 	}
9272 
9273 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
9274 		/*
9275 		 * Previous user of page had a different color
9276 		 * but since there are no current users
9277 		 * we just flush the cache and change the color.
9278 		 */
9279 		SFMMU_STAT(sf_pgcolor_conflict);
9280 		sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9281 		PP_SET_VCOLOR(pp, vcolor);
9282 		return;
9283 	}
9284 
9285 	/*
9286 	 * If we get here we have a vac conflict with a current
9287 	 * mapping.  VAC conflict policy is as follows.
9288 	 * - The default is to unload the other mappings unless:
9289 	 * - If we have a large mapping we uncache the page.
9290 	 * We need to uncache the rest of the large page too.
9291 	 * - If any of the mappings are locked we uncache the page.
9292 	 * - If the requested mapping is inconsistent
9293 	 * with another mapping and that mapping
9294 	 * is in the same address space we have to
9295 	 * make it non-cached.  The default thing
9296 	 * to do is unload the inconsistent mapping
9297 	 * but if they are in the same address space
9298 	 * we run the risk of unmapping the pc or the
9299 	 * stack which we will use as we return to the user,
9300 	 * in which case we can then fault on the thing
9301 	 * we just unloaded and get into an infinite loop.
9302 	 */
9303 	if (PP_ISMAPPED_LARGE(pp)) {
9304 		int sz;
9305 
9306 		/*
9307 		 * Existing mapping is for big pages. We don't unload
9308 		 * existing big mappings to satisfy new mappings.
9309 		 * Always convert all mappings to TNC.
9310 		 */
9311 		sz = fnd_mapping_sz(pp);
9312 		pp = PP_GROUPLEADER(pp, sz);
9313 		SFMMU_STAT_ADD(sf_uncache_conflict, TTEPAGES(sz));
9314 		sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH,
9315 		    TTEPAGES(sz));
9316 
9317 		return;
9318 	}
9319 
9320 	/*
9321 	 * check if any mapping is in same as or if it is locked
9322 	 * since in that case we need to uncache.
9323 	 */
9324 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9325 		tmphme = sfhmep->hme_next;
9326 		if (IS_PAHME(sfhmep))
9327 			continue;
9328 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9329 		if (hmeblkp->hblk_xhat_bit)
9330 			continue;
9331 		tmphat = hblktosfmmu(hmeblkp);
9332 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
9333 		ASSERT(TTE_IS_VALID(&tte));
9334 		if (hmeblkp->hblk_shared || tmphat == hat ||
9335 		    hmeblkp->hblk_lckcnt) {
9336 			/*
9337 			 * We have an uncache conflict
9338 			 */
9339 			SFMMU_STAT(sf_uncache_conflict);
9340 			sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH, 1);
9341 			return;
9342 		}
9343 	}
9344 
9345 	/*
9346 	 * We have an unload conflict
9347 	 * We have already checked for LARGE mappings, therefore
9348 	 * the remaining mapping(s) must be TTE8K.
9349 	 */
9350 	SFMMU_STAT(sf_unload_conflict);
9351 
9352 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9353 		tmphme = sfhmep->hme_next;
9354 		if (IS_PAHME(sfhmep))
9355 			continue;
9356 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9357 		if (hmeblkp->hblk_xhat_bit)
9358 			continue;
9359 		ASSERT(!hmeblkp->hblk_shared);
9360 		(void) sfmmu_pageunload(pp, sfhmep, TTE8K);
9361 	}
9362 
9363 	if (PP_ISMAPPED_KPM(pp))
9364 		sfmmu_kpm_vac_unload(pp, addr);
9365 
9366 	/*
9367 	 * Unloads only do TLB flushes so we need to flush the
9368 	 * cache here.
9369 	 */
9370 	sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9371 	PP_SET_VCOLOR(pp, vcolor);
9372 }
9373 
9374 /*
9375  * Whenever a mapping is unloaded and the page is in TNC state,
9376  * we see if the page can be made cacheable again. 'pp' is
9377  * the page that we just unloaded a mapping from, the size
9378  * of mapping that was unloaded is 'ottesz'.
9379  * Remark:
9380  * The recache policy for mpss pages can leave a performance problem
9381  * under the following circumstances:
9382  * . A large page in uncached mode has just been unmapped.
9383  * . All constituent pages are TNC due to a conflicting small mapping.
9384  * . There are many other, non conflicting, small mappings around for
9385  *   a lot of the constituent pages.
9386  * . We're called w/ the "old" groupleader page and the old ottesz,
9387  *   but this is irrelevant, since we're no more "PP_ISMAPPED_LARGE", so
9388  *   we end up w/ TTE8K or npages == 1.
9389  * . We call tst_tnc w/ the old groupleader only, and if there is no
9390  *   conflict, we re-cache only this page.
9391  * . All other small mappings are not checked and will be left in TNC mode.
9392  * The problem is not very serious because:
9393  * . mpss is actually only defined for heap and stack, so the probability
9394  *   is not very high that a large page mapping exists in parallel to a small
9395  *   one (this is possible, but seems to be bad programming style in the
9396  *   appl).
9397  * . The problem gets a little bit more serious, when those TNC pages
9398  *   have to be mapped into kernel space, e.g. for networking.
9399  * . When VAC alias conflicts occur in applications, this is regarded
9400  *   as an application bug. So if kstat's show them, the appl should
9401  *   be changed anyway.
9402  */
9403 void
9404 conv_tnc(page_t *pp, int ottesz)
9405 {
9406 	int cursz, dosz;
9407 	pgcnt_t curnpgs, dopgs;
9408 	pgcnt_t pg64k;
9409 	page_t *pp2;
9410 
9411 	/*
9412 	 * Determine how big a range we check for TNC and find
9413 	 * leader page. cursz is the size of the biggest
9414 	 * mapping that still exist on 'pp'.
9415 	 */
9416 	if (PP_ISMAPPED_LARGE(pp)) {
9417 		cursz = fnd_mapping_sz(pp);
9418 	} else {
9419 		cursz = TTE8K;
9420 	}
9421 
9422 	if (ottesz >= cursz) {
9423 		dosz = ottesz;
9424 		pp2 = pp;
9425 	} else {
9426 		dosz = cursz;
9427 		pp2 = PP_GROUPLEADER(pp, dosz);
9428 	}
9429 
9430 	pg64k = TTEPAGES(TTE64K);
9431 	dopgs = TTEPAGES(dosz);
9432 
9433 	ASSERT(dopgs == 1 || ((dopgs & (pg64k - 1)) == 0));
9434 
9435 	while (dopgs != 0) {
9436 		curnpgs = TTEPAGES(cursz);
9437 		if (tst_tnc(pp2, curnpgs)) {
9438 			SFMMU_STAT_ADD(sf_recache, curnpgs);
9439 			sfmmu_page_cache_array(pp2, HAT_CACHE, CACHE_NO_FLUSH,
9440 			    curnpgs);
9441 		}
9442 
9443 		ASSERT(dopgs >= curnpgs);
9444 		dopgs -= curnpgs;
9445 
9446 		if (dopgs == 0) {
9447 			break;
9448 		}
9449 
9450 		pp2 = PP_PAGENEXT_N(pp2, curnpgs);
9451 		if (((dopgs & (pg64k - 1)) == 0) && PP_ISMAPPED_LARGE(pp2)) {
9452 			cursz = fnd_mapping_sz(pp2);
9453 		} else {
9454 			cursz = TTE8K;
9455 		}
9456 	}
9457 }
9458 
9459 /*
9460  * Returns 1 if page(s) can be converted from TNC to cacheable setting,
9461  * returns 0 otherwise. Note that oaddr argument is valid for only
9462  * 8k pages.
9463  */
9464 int
9465 tst_tnc(page_t *pp, pgcnt_t npages)
9466 {
9467 	struct	sf_hment *sfhme;
9468 	struct	hme_blk *hmeblkp;
9469 	tte_t	tte;
9470 	caddr_t	vaddr;
9471 	int	clr_valid = 0;
9472 	int 	color, color1, bcolor;
9473 	int	i, ncolors;
9474 
9475 	ASSERT(pp != NULL);
9476 	ASSERT(!(cache & CACHE_WRITEBACK));
9477 
9478 	if (npages > 1) {
9479 		ncolors = CACHE_NUM_COLOR;
9480 	}
9481 
9482 	for (i = 0; i < npages; i++) {
9483 		ASSERT(sfmmu_mlist_held(pp));
9484 		ASSERT(PP_ISTNC(pp));
9485 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
9486 
9487 		if (PP_ISPNC(pp)) {
9488 			return (0);
9489 		}
9490 
9491 		clr_valid = 0;
9492 		if (PP_ISMAPPED_KPM(pp)) {
9493 			caddr_t kpmvaddr;
9494 
9495 			ASSERT(kpm_enable);
9496 			kpmvaddr = hat_kpm_page2va(pp, 1);
9497 			ASSERT(!(npages > 1 && IS_KPM_ALIAS_RANGE(kpmvaddr)));
9498 			color1 = addr_to_vcolor(kpmvaddr);
9499 			clr_valid = 1;
9500 		}
9501 
9502 		for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9503 			if (IS_PAHME(sfhme))
9504 				continue;
9505 			hmeblkp = sfmmu_hmetohblk(sfhme);
9506 			if (hmeblkp->hblk_xhat_bit)
9507 				continue;
9508 
9509 			sfmmu_copytte(&sfhme->hme_tte, &tte);
9510 			ASSERT(TTE_IS_VALID(&tte));
9511 
9512 			vaddr = tte_to_vaddr(hmeblkp, tte);
9513 			color = addr_to_vcolor(vaddr);
9514 
9515 			if (npages > 1) {
9516 				/*
9517 				 * If there is a big mapping, make sure
9518 				 * 8K mapping is consistent with the big
9519 				 * mapping.
9520 				 */
9521 				bcolor = i % ncolors;
9522 				if (color != bcolor) {
9523 					return (0);
9524 				}
9525 			}
9526 			if (!clr_valid) {
9527 				clr_valid = 1;
9528 				color1 = color;
9529 			}
9530 
9531 			if (color1 != color) {
9532 				return (0);
9533 			}
9534 		}
9535 
9536 		pp = PP_PAGENEXT(pp);
9537 	}
9538 
9539 	return (1);
9540 }
9541 
9542 void
9543 sfmmu_page_cache_array(page_t *pp, int flags, int cache_flush_flag,
9544 	pgcnt_t npages)
9545 {
9546 	kmutex_t *pmtx;
9547 	int i, ncolors, bcolor;
9548 	kpm_hlk_t *kpmp;
9549 	cpuset_t cpuset;
9550 
9551 	ASSERT(pp != NULL);
9552 	ASSERT(!(cache & CACHE_WRITEBACK));
9553 
9554 	kpmp = sfmmu_kpm_kpmp_enter(pp, npages);
9555 	pmtx = sfmmu_page_enter(pp);
9556 
9557 	/*
9558 	 * Fast path caching single unmapped page
9559 	 */
9560 	if (npages == 1 && !PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp) &&
9561 	    flags == HAT_CACHE) {
9562 		PP_CLRTNC(pp);
9563 		PP_CLRPNC(pp);
9564 		sfmmu_page_exit(pmtx);
9565 		sfmmu_kpm_kpmp_exit(kpmp);
9566 		return;
9567 	}
9568 
9569 	/*
9570 	 * We need to capture all cpus in order to change cacheability
9571 	 * because we can't allow one cpu to access the same physical
9572 	 * page using a cacheable and a non-cachebale mapping at the same
9573 	 * time. Since we may end up walking the ism mapping list
9574 	 * have to grab it's lock now since we can't after all the
9575 	 * cpus have been captured.
9576 	 */
9577 	sfmmu_hat_lock_all();
9578 	mutex_enter(&ism_mlist_lock);
9579 	kpreempt_disable();
9580 	cpuset = cpu_ready_set;
9581 	xc_attention(cpuset);
9582 
9583 	if (npages > 1) {
9584 		/*
9585 		 * Make sure all colors are flushed since the
9586 		 * sfmmu_page_cache() only flushes one color-
9587 		 * it does not know big pages.
9588 		 */
9589 		ncolors = CACHE_NUM_COLOR;
9590 		if (flags & HAT_TMPNC) {
9591 			for (i = 0; i < ncolors; i++) {
9592 				sfmmu_cache_flushcolor(i, pp->p_pagenum);
9593 			}
9594 			cache_flush_flag = CACHE_NO_FLUSH;
9595 		}
9596 	}
9597 
9598 	for (i = 0; i < npages; i++) {
9599 
9600 		ASSERT(sfmmu_mlist_held(pp));
9601 
9602 		if (!(flags == HAT_TMPNC && PP_ISTNC(pp))) {
9603 
9604 			if (npages > 1) {
9605 				bcolor = i % ncolors;
9606 			} else {
9607 				bcolor = NO_VCOLOR;
9608 			}
9609 
9610 			sfmmu_page_cache(pp, flags, cache_flush_flag,
9611 			    bcolor);
9612 		}
9613 
9614 		pp = PP_PAGENEXT(pp);
9615 	}
9616 
9617 	xt_sync(cpuset);
9618 	xc_dismissed(cpuset);
9619 	mutex_exit(&ism_mlist_lock);
9620 	sfmmu_hat_unlock_all();
9621 	sfmmu_page_exit(pmtx);
9622 	sfmmu_kpm_kpmp_exit(kpmp);
9623 	kpreempt_enable();
9624 }
9625 
9626 /*
9627  * This function changes the virtual cacheability of all mappings to a
9628  * particular page.  When changing from uncache to cacheable the mappings will
9629  * only be changed if all of them have the same virtual color.
9630  * We need to flush the cache in all cpus.  It is possible that
9631  * a process referenced a page as cacheable but has sinced exited
9632  * and cleared the mapping list.  We still to flush it but have no
9633  * state so all cpus is the only alternative.
9634  */
9635 static void
9636 sfmmu_page_cache(page_t *pp, int flags, int cache_flush_flag, int bcolor)
9637 {
9638 	struct	sf_hment *sfhme;
9639 	struct	hme_blk *hmeblkp;
9640 	sfmmu_t *sfmmup;
9641 	tte_t	tte, ttemod;
9642 	caddr_t	vaddr;
9643 	int	ret, color;
9644 	pfn_t	pfn;
9645 
9646 	color = bcolor;
9647 	pfn = pp->p_pagenum;
9648 
9649 	for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9650 
9651 		if (IS_PAHME(sfhme))
9652 			continue;
9653 		hmeblkp = sfmmu_hmetohblk(sfhme);
9654 
9655 		if (hmeblkp->hblk_xhat_bit)
9656 			continue;
9657 
9658 		sfmmu_copytte(&sfhme->hme_tte, &tte);
9659 		ASSERT(TTE_IS_VALID(&tte));
9660 		vaddr = tte_to_vaddr(hmeblkp, tte);
9661 		color = addr_to_vcolor(vaddr);
9662 
9663 #ifdef DEBUG
9664 		if ((flags & HAT_CACHE) && bcolor != NO_VCOLOR) {
9665 			ASSERT(color == bcolor);
9666 		}
9667 #endif
9668 
9669 		ASSERT(flags != HAT_TMPNC || color == PP_GET_VCOLOR(pp));
9670 
9671 		ttemod = tte;
9672 		if (flags & (HAT_UNCACHE | HAT_TMPNC)) {
9673 			TTE_CLR_VCACHEABLE(&ttemod);
9674 		} else {	/* flags & HAT_CACHE */
9675 			TTE_SET_VCACHEABLE(&ttemod);
9676 		}
9677 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
9678 		if (ret < 0) {
9679 			/*
9680 			 * Since all cpus are captured modifytte should not
9681 			 * fail.
9682 			 */
9683 			panic("sfmmu_page_cache: write to tte failed");
9684 		}
9685 
9686 		sfmmup = hblktosfmmu(hmeblkp);
9687 		if (cache_flush_flag == CACHE_FLUSH) {
9688 			/*
9689 			 * Flush TSBs, TLBs and caches
9690 			 */
9691 			if (hmeblkp->hblk_shared) {
9692 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9693 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9694 				sf_region_t *rgnp;
9695 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9696 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9697 				ASSERT(srdp != NULL);
9698 				rgnp = srdp->srd_hmergnp[rid];
9699 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9700 				    srdp, rgnp, rid);
9701 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9702 				    hmeblkp, 0);
9703 				sfmmu_cache_flush(pfn, addr_to_vcolor(vaddr));
9704 			} else if (sfmmup->sfmmu_ismhat) {
9705 				if (flags & HAT_CACHE) {
9706 					SFMMU_STAT(sf_ism_recache);
9707 				} else {
9708 					SFMMU_STAT(sf_ism_uncache);
9709 				}
9710 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9711 				    pfn, CACHE_FLUSH);
9712 			} else {
9713 				sfmmu_tlbcache_demap(vaddr, sfmmup, hmeblkp,
9714 				    pfn, 0, FLUSH_ALL_CPUS, CACHE_FLUSH, 1);
9715 			}
9716 
9717 			/*
9718 			 * all cache entries belonging to this pfn are
9719 			 * now flushed.
9720 			 */
9721 			cache_flush_flag = CACHE_NO_FLUSH;
9722 		} else {
9723 			/*
9724 			 * Flush only TSBs and TLBs.
9725 			 */
9726 			if (hmeblkp->hblk_shared) {
9727 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9728 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9729 				sf_region_t *rgnp;
9730 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9731 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9732 				ASSERT(srdp != NULL);
9733 				rgnp = srdp->srd_hmergnp[rid];
9734 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9735 				    srdp, rgnp, rid);
9736 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9737 				    hmeblkp, 0);
9738 			} else if (sfmmup->sfmmu_ismhat) {
9739 				if (flags & HAT_CACHE) {
9740 					SFMMU_STAT(sf_ism_recache);
9741 				} else {
9742 					SFMMU_STAT(sf_ism_uncache);
9743 				}
9744 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9745 				    pfn, CACHE_NO_FLUSH);
9746 			} else {
9747 				sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 1);
9748 			}
9749 		}
9750 	}
9751 
9752 	if (PP_ISMAPPED_KPM(pp))
9753 		sfmmu_kpm_page_cache(pp, flags, cache_flush_flag);
9754 
9755 	switch (flags) {
9756 
9757 		default:
9758 			panic("sfmmu_pagecache: unknown flags");
9759 			break;
9760 
9761 		case HAT_CACHE:
9762 			PP_CLRTNC(pp);
9763 			PP_CLRPNC(pp);
9764 			PP_SET_VCOLOR(pp, color);
9765 			break;
9766 
9767 		case HAT_TMPNC:
9768 			PP_SETTNC(pp);
9769 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9770 			break;
9771 
9772 		case HAT_UNCACHE:
9773 			PP_SETPNC(pp);
9774 			PP_CLRTNC(pp);
9775 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9776 			break;
9777 	}
9778 }
9779 #endif	/* VAC */
9780 
9781 
9782 /*
9783  * Wrapper routine used to return a context.
9784  *
9785  * It's the responsibility of the caller to guarantee that the
9786  * process serializes on calls here by taking the HAT lock for
9787  * the hat.
9788  *
9789  */
9790 static void
9791 sfmmu_get_ctx(sfmmu_t *sfmmup)
9792 {
9793 	mmu_ctx_t *mmu_ctxp;
9794 	uint_t pstate_save;
9795 	int ret;
9796 
9797 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9798 	ASSERT(sfmmup != ksfmmup);
9799 
9800 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID)) {
9801 		sfmmu_setup_tsbinfo(sfmmup);
9802 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ALLCTX_INVALID);
9803 	}
9804 
9805 	kpreempt_disable();
9806 
9807 	mmu_ctxp = CPU_MMU_CTXP(CPU);
9808 	ASSERT(mmu_ctxp);
9809 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
9810 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
9811 
9812 	/*
9813 	 * Do a wrap-around if cnum reaches the max # cnum supported by a MMU.
9814 	 */
9815 	if (mmu_ctxp->mmu_cnum == mmu_ctxp->mmu_nctxs)
9816 		sfmmu_ctx_wrap_around(mmu_ctxp, B_TRUE);
9817 
9818 	/*
9819 	 * Let the MMU set up the page sizes to use for
9820 	 * this context in the TLB. Don't program 2nd dtlb for ism hat.
9821 	 */
9822 	if ((&mmu_set_ctx_page_sizes) && (sfmmup->sfmmu_ismhat == 0)) {
9823 		mmu_set_ctx_page_sizes(sfmmup);
9824 	}
9825 
9826 	/*
9827 	 * sfmmu_alloc_ctx and sfmmu_load_mmustate will be performed with
9828 	 * interrupts disabled to prevent race condition with wrap-around
9829 	 * ctx invalidatation. In sun4v, ctx invalidation also involves
9830 	 * a HV call to set the number of TSBs to 0. If interrupts are not
9831 	 * disabled until after sfmmu_load_mmustate is complete TSBs may
9832 	 * become assigned to INVALID_CONTEXT. This is not allowed.
9833 	 */
9834 	pstate_save = sfmmu_disable_intrs();
9835 
9836 	if (sfmmu_alloc_ctx(sfmmup, 1, CPU, SFMMU_PRIVATE) &&
9837 	    sfmmup->sfmmu_scdp != NULL) {
9838 		sf_scd_t *scdp = sfmmup->sfmmu_scdp;
9839 		sfmmu_t *scsfmmup = scdp->scd_sfmmup;
9840 		ret = sfmmu_alloc_ctx(scsfmmup, 1, CPU, SFMMU_SHARED);
9841 		/* debug purpose only */
9842 		ASSERT(!ret || scsfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
9843 		    != INVALID_CONTEXT);
9844 	}
9845 	sfmmu_load_mmustate(sfmmup);
9846 
9847 	sfmmu_enable_intrs(pstate_save);
9848 
9849 	kpreempt_enable();
9850 }
9851 
9852 /*
9853  * When all cnums are used up in a MMU, cnum will wrap around to the
9854  * next generation and start from 2.
9855  */
9856 static void
9857 sfmmu_ctx_wrap_around(mmu_ctx_t *mmu_ctxp, boolean_t reset_cnum)
9858 {
9859 
9860 	/* caller must have disabled the preemption */
9861 	ASSERT(curthread->t_preempt >= 1);
9862 	ASSERT(mmu_ctxp != NULL);
9863 
9864 	/* acquire Per-MMU (PM) spin lock */
9865 	mutex_enter(&mmu_ctxp->mmu_lock);
9866 
9867 	/* re-check to see if wrap-around is needed */
9868 	if (mmu_ctxp->mmu_cnum < mmu_ctxp->mmu_nctxs)
9869 		goto done;
9870 
9871 	SFMMU_MMU_STAT(mmu_wrap_around);
9872 
9873 	/* update gnum */
9874 	ASSERT(mmu_ctxp->mmu_gnum != 0);
9875 	mmu_ctxp->mmu_gnum++;
9876 	if (mmu_ctxp->mmu_gnum == 0 ||
9877 	    mmu_ctxp->mmu_gnum > MAX_SFMMU_GNUM_VAL) {
9878 		cmn_err(CE_PANIC, "mmu_gnum of mmu_ctx 0x%p is out of bound.",
9879 		    (void *)mmu_ctxp);
9880 	}
9881 
9882 	if (mmu_ctxp->mmu_ncpus > 1) {
9883 		cpuset_t cpuset;
9884 
9885 		membar_enter(); /* make sure updated gnum visible */
9886 
9887 		SFMMU_XCALL_STATS(NULL);
9888 
9889 		/* xcall to others on the same MMU to invalidate ctx */
9890 		cpuset = mmu_ctxp->mmu_cpuset;
9891 		ASSERT(CPU_IN_SET(cpuset, CPU->cpu_id) || !reset_cnum);
9892 		CPUSET_DEL(cpuset, CPU->cpu_id);
9893 		CPUSET_AND(cpuset, cpu_ready_set);
9894 
9895 		/*
9896 		 * Pass in INVALID_CONTEXT as the first parameter to
9897 		 * sfmmu_raise_tsb_exception, which invalidates the context
9898 		 * of any process running on the CPUs in the MMU.
9899 		 */
9900 		xt_some(cpuset, sfmmu_raise_tsb_exception,
9901 		    INVALID_CONTEXT, INVALID_CONTEXT);
9902 		xt_sync(cpuset);
9903 
9904 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
9905 	}
9906 
9907 	if (sfmmu_getctx_sec() != INVALID_CONTEXT) {
9908 		sfmmu_setctx_sec(INVALID_CONTEXT);
9909 		sfmmu_clear_utsbinfo();
9910 	}
9911 
9912 	/*
9913 	 * No xcall is needed here. For sun4u systems all CPUs in context
9914 	 * domain share a single physical MMU therefore it's enough to flush
9915 	 * TLB on local CPU. On sun4v systems we use 1 global context
9916 	 * domain and flush all remote TLBs in sfmmu_raise_tsb_exception
9917 	 * handler. Note that vtag_flushall_uctxs() is called
9918 	 * for Ultra II machine, where the equivalent flushall functionality
9919 	 * is implemented in SW, and only user ctx TLB entries are flushed.
9920 	 */
9921 	if (&vtag_flushall_uctxs != NULL) {
9922 		vtag_flushall_uctxs();
9923 	} else {
9924 		vtag_flushall();
9925 	}
9926 
9927 	/* reset mmu cnum, skips cnum 0 and 1 */
9928 	if (reset_cnum == B_TRUE)
9929 		mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
9930 
9931 done:
9932 	mutex_exit(&mmu_ctxp->mmu_lock);
9933 }
9934 
9935 
9936 /*
9937  * For multi-threaded process, set the process context to INVALID_CONTEXT
9938  * so that it faults and reloads the MMU state from TL=0. For single-threaded
9939  * process, we can just load the MMU state directly without having to
9940  * set context invalid. Caller must hold the hat lock since we don't
9941  * acquire it here.
9942  */
9943 static void
9944 sfmmu_sync_mmustate(sfmmu_t *sfmmup)
9945 {
9946 	uint_t cnum;
9947 	uint_t pstate_save;
9948 
9949 	ASSERT(sfmmup != ksfmmup);
9950 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9951 
9952 	kpreempt_disable();
9953 
9954 	/*
9955 	 * We check whether the pass'ed-in sfmmup is the same as the
9956 	 * current running proc. This is to makes sure the current proc
9957 	 * stays single-threaded if it already is.
9958 	 */
9959 	if ((sfmmup == curthread->t_procp->p_as->a_hat) &&
9960 	    (curthread->t_procp->p_lwpcnt == 1)) {
9961 		/* single-thread */
9962 		cnum = sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum;
9963 		if (cnum != INVALID_CONTEXT) {
9964 			uint_t curcnum;
9965 			/*
9966 			 * Disable interrupts to prevent race condition
9967 			 * with sfmmu_ctx_wrap_around ctx invalidation.
9968 			 * In sun4v, ctx invalidation involves setting
9969 			 * TSB to NULL, hence, interrupts should be disabled
9970 			 * untill after sfmmu_load_mmustate is completed.
9971 			 */
9972 			pstate_save = sfmmu_disable_intrs();
9973 			curcnum = sfmmu_getctx_sec();
9974 			if (curcnum == cnum)
9975 				sfmmu_load_mmustate(sfmmup);
9976 			sfmmu_enable_intrs(pstate_save);
9977 			ASSERT(curcnum == cnum || curcnum == INVALID_CONTEXT);
9978 		}
9979 	} else {
9980 		/*
9981 		 * multi-thread
9982 		 * or when sfmmup is not the same as the curproc.
9983 		 */
9984 		sfmmu_invalidate_ctx(sfmmup);
9985 	}
9986 
9987 	kpreempt_enable();
9988 }
9989 
9990 
9991 /*
9992  * Replace the specified TSB with a new TSB.  This function gets called when
9993  * we grow, shrink or swapin a TSB.  When swapping in a TSB (TSB_SWAPIN), the
9994  * TSB_FORCEALLOC flag may be used to force allocation of a minimum-sized TSB
9995  * (8K).
9996  *
9997  * Caller must hold the HAT lock, but should assume any tsb_info
9998  * pointers it has are no longer valid after calling this function.
9999  *
10000  * Return values:
10001  *	TSB_ALLOCFAIL	Failed to allocate a TSB, due to memory constraints
10002  *	TSB_LOSTRACE	HAT is busy, i.e. another thread is already doing
10003  *			something to this tsbinfo/TSB
10004  *	TSB_SUCCESS	Operation succeeded
10005  */
10006 static tsb_replace_rc_t
10007 sfmmu_replace_tsb(sfmmu_t *sfmmup, struct tsb_info *old_tsbinfo, uint_t szc,
10008     hatlock_t *hatlockp, uint_t flags)
10009 {
10010 	struct tsb_info *new_tsbinfo = NULL;
10011 	struct tsb_info *curtsb, *prevtsb;
10012 	uint_t tte_sz_mask;
10013 	int i;
10014 
10015 	ASSERT(sfmmup != ksfmmup);
10016 	ASSERT(sfmmup->sfmmu_ismhat == 0);
10017 	ASSERT(sfmmu_hat_lock_held(sfmmup));
10018 	ASSERT(szc <= tsb_max_growsize);
10019 
10020 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_BUSY))
10021 		return (TSB_LOSTRACE);
10022 
10023 	/*
10024 	 * Find the tsb_info ahead of this one in the list, and
10025 	 * also make sure that the tsb_info passed in really
10026 	 * exists!
10027 	 */
10028 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
10029 	    curtsb != old_tsbinfo && curtsb != NULL;
10030 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
10031 		;
10032 	ASSERT(curtsb != NULL);
10033 
10034 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
10035 		/*
10036 		 * The process is swapped out, so just set the new size
10037 		 * code.  When it swaps back in, we'll allocate a new one
10038 		 * of the new chosen size.
10039 		 */
10040 		curtsb->tsb_szc = szc;
10041 		return (TSB_SUCCESS);
10042 	}
10043 	SFMMU_FLAGS_SET(sfmmup, HAT_BUSY);
10044 
10045 	tte_sz_mask = old_tsbinfo->tsb_ttesz_mask;
10046 
10047 	/*
10048 	 * All initialization is done inside of sfmmu_tsbinfo_alloc().
10049 	 * If we fail to allocate a TSB, exit.
10050 	 *
10051 	 * If tsb grows with new tsb size > 4M and old tsb size < 4M,
10052 	 * then try 4M slab after the initial alloc fails.
10053 	 *
10054 	 * If tsb swapin with tsb size > 4M, then try 4M after the
10055 	 * initial alloc fails.
10056 	 */
10057 	sfmmu_hat_exit(hatlockp);
10058 	if (sfmmu_tsbinfo_alloc(&new_tsbinfo, szc,
10059 	    tte_sz_mask, flags, sfmmup) &&
10060 	    (!(flags & (TSB_GROW | TSB_SWAPIN)) || (szc <= TSB_4M_SZCODE) ||
10061 	    (!(flags & TSB_SWAPIN) &&
10062 	    (old_tsbinfo->tsb_szc >= TSB_4M_SZCODE)) ||
10063 	    sfmmu_tsbinfo_alloc(&new_tsbinfo, TSB_4M_SZCODE,
10064 	    tte_sz_mask, flags, sfmmup))) {
10065 		(void) sfmmu_hat_enter(sfmmup);
10066 		if (!(flags & TSB_SWAPIN))
10067 			SFMMU_STAT(sf_tsb_resize_failures);
10068 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10069 		return (TSB_ALLOCFAIL);
10070 	}
10071 	(void) sfmmu_hat_enter(sfmmup);
10072 
10073 	/*
10074 	 * Re-check to make sure somebody else didn't muck with us while we
10075 	 * didn't hold the HAT lock.  If the process swapped out, fine, just
10076 	 * exit; this can happen if we try to shrink the TSB from the context
10077 	 * of another process (such as on an ISM unmap), though it is rare.
10078 	 */
10079 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
10080 		SFMMU_STAT(sf_tsb_resize_failures);
10081 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10082 		sfmmu_hat_exit(hatlockp);
10083 		sfmmu_tsbinfo_free(new_tsbinfo);
10084 		(void) sfmmu_hat_enter(sfmmup);
10085 		return (TSB_LOSTRACE);
10086 	}
10087 
10088 #ifdef	DEBUG
10089 	/* Reverify that the tsb_info still exists.. for debugging only */
10090 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
10091 	    curtsb != old_tsbinfo && curtsb != NULL;
10092 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
10093 		;
10094 	ASSERT(curtsb != NULL);
10095 #endif	/* DEBUG */
10096 
10097 	/*
10098 	 * Quiesce any CPUs running this process on their next TLB miss
10099 	 * so they atomically see the new tsb_info.  We temporarily set the
10100 	 * context to invalid context so new threads that come on processor
10101 	 * after we do the xcall to cpusran will also serialize behind the
10102 	 * HAT lock on TLB miss and will see the new TSB.  Since this short
10103 	 * race with a new thread coming on processor is relatively rare,
10104 	 * this synchronization mechanism should be cheaper than always
10105 	 * pausing all CPUs for the duration of the setup, which is what
10106 	 * the old implementation did.  This is particuarly true if we are
10107 	 * copying a huge chunk of memory around during that window.
10108 	 *
10109 	 * The memory barriers are to make sure things stay consistent
10110 	 * with resume() since it does not hold the HAT lock while
10111 	 * walking the list of tsb_info structures.
10112 	 */
10113 	if ((flags & TSB_SWAPIN) != TSB_SWAPIN) {
10114 		/* The TSB is either growing or shrinking. */
10115 		sfmmu_invalidate_ctx(sfmmup);
10116 	} else {
10117 		/*
10118 		 * It is illegal to swap in TSBs from a process other
10119 		 * than a process being swapped in.  This in turn
10120 		 * implies we do not have a valid MMU context here
10121 		 * since a process needs one to resolve translation
10122 		 * misses.
10123 		 */
10124 		ASSERT(curthread->t_procp->p_as->a_hat == sfmmup);
10125 	}
10126 
10127 #ifdef DEBUG
10128 	ASSERT(max_mmu_ctxdoms > 0);
10129 
10130 	/*
10131 	 * Process should have INVALID_CONTEXT on all MMUs
10132 	 */
10133 	for (i = 0; i < max_mmu_ctxdoms; i++) {
10134 
10135 		ASSERT(sfmmup->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
10136 	}
10137 #endif
10138 
10139 	new_tsbinfo->tsb_next = old_tsbinfo->tsb_next;
10140 	membar_stst();	/* strict ordering required */
10141 	if (prevtsb)
10142 		prevtsb->tsb_next = new_tsbinfo;
10143 	else
10144 		sfmmup->sfmmu_tsb = new_tsbinfo;
10145 	membar_enter();	/* make sure new TSB globally visible */
10146 
10147 	/*
10148 	 * We need to migrate TSB entries from the old TSB to the new TSB
10149 	 * if tsb_remap_ttes is set and the TSB is growing.
10150 	 */
10151 	if (tsb_remap_ttes && ((flags & TSB_GROW) == TSB_GROW))
10152 		sfmmu_copy_tsb(old_tsbinfo, new_tsbinfo);
10153 
10154 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10155 
10156 	/*
10157 	 * Drop the HAT lock to free our old tsb_info.
10158 	 */
10159 	sfmmu_hat_exit(hatlockp);
10160 
10161 	if ((flags & TSB_GROW) == TSB_GROW) {
10162 		SFMMU_STAT(sf_tsb_grow);
10163 	} else if ((flags & TSB_SHRINK) == TSB_SHRINK) {
10164 		SFMMU_STAT(sf_tsb_shrink);
10165 	}
10166 
10167 	sfmmu_tsbinfo_free(old_tsbinfo);
10168 
10169 	(void) sfmmu_hat_enter(sfmmup);
10170 	return (TSB_SUCCESS);
10171 }
10172 
10173 /*
10174  * This function will re-program hat pgsz array, and invalidate the
10175  * process' context, forcing the process to switch to another
10176  * context on the next TLB miss, and therefore start using the
10177  * TLB that is reprogrammed for the new page sizes.
10178  */
10179 void
10180 sfmmu_reprog_pgsz_arr(sfmmu_t *sfmmup, uint8_t *tmp_pgsz)
10181 {
10182 	int i;
10183 	hatlock_t *hatlockp = NULL;
10184 
10185 	hatlockp = sfmmu_hat_enter(sfmmup);
10186 	/* USIII+-IV+ optimization, requires hat lock */
10187 	if (tmp_pgsz) {
10188 		for (i = 0; i < mmu_page_sizes; i++)
10189 			sfmmup->sfmmu_pgsz[i] = tmp_pgsz[i];
10190 	}
10191 	SFMMU_STAT(sf_tlb_reprog_pgsz);
10192 
10193 	sfmmu_invalidate_ctx(sfmmup);
10194 
10195 	sfmmu_hat_exit(hatlockp);
10196 }
10197 
10198 /*
10199  * The scd_rttecnt field in the SCD must be updated to take account of the
10200  * regions which it contains.
10201  */
10202 static void
10203 sfmmu_set_scd_rttecnt(sf_srd_t *srdp, sf_scd_t *scdp)
10204 {
10205 	uint_t rid;
10206 	uint_t i, j;
10207 	ulong_t w;
10208 	sf_region_t *rgnp;
10209 
10210 	ASSERT(srdp != NULL);
10211 
10212 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
10213 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
10214 			continue;
10215 		}
10216 
10217 		j = 0;
10218 		while (w) {
10219 			if (!(w & 0x1)) {
10220 				j++;
10221 				w >>= 1;
10222 				continue;
10223 			}
10224 			rid = (i << BT_ULSHIFT) | j;
10225 			j++;
10226 			w >>= 1;
10227 
10228 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
10229 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
10230 			rgnp = srdp->srd_hmergnp[rid];
10231 			ASSERT(rgnp->rgn_refcnt > 0);
10232 			ASSERT(rgnp->rgn_id == rid);
10233 
10234 			scdp->scd_rttecnt[rgnp->rgn_pgszc] +=
10235 			    rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
10236 
10237 			/*
10238 			 * Maintain the tsb0 inflation cnt for the regions
10239 			 * in the SCD.
10240 			 */
10241 			if (rgnp->rgn_pgszc >= TTE4M) {
10242 				scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt +=
10243 				    rgnp->rgn_size >>
10244 				    (TTE_PAGE_SHIFT(TTE8K) + 2);
10245 			}
10246 		}
10247 	}
10248 }
10249 
10250 /*
10251  * This function assumes that there are either four or six supported page
10252  * sizes and at most two programmable TLBs, so we need to decide which
10253  * page sizes are most important and then tell the MMU layer so it
10254  * can adjust the TLB page sizes accordingly (if supported).
10255  *
10256  * If these assumptions change, this function will need to be
10257  * updated to support whatever the new limits are.
10258  *
10259  * The growing flag is nonzero if we are growing the address space,
10260  * and zero if it is shrinking.  This allows us to decide whether
10261  * to grow or shrink our TSB, depending upon available memory
10262  * conditions.
10263  */
10264 static void
10265 sfmmu_check_page_sizes(sfmmu_t *sfmmup, int growing)
10266 {
10267 	uint64_t ttecnt[MMU_PAGE_SIZES];
10268 	uint64_t tte8k_cnt, tte4m_cnt;
10269 	uint8_t i;
10270 	int sectsb_thresh;
10271 
10272 	/*
10273 	 * Kernel threads, processes with small address spaces not using
10274 	 * large pages, and dummy ISM HATs need not apply.
10275 	 */
10276 	if (sfmmup == ksfmmup || sfmmup->sfmmu_ismhat != NULL)
10277 		return;
10278 
10279 	if (!SFMMU_LGPGS_INUSE(sfmmup) &&
10280 	    sfmmup->sfmmu_ttecnt[TTE8K] <= tsb_rss_factor)
10281 		return;
10282 
10283 	for (i = 0; i < mmu_page_sizes; i++) {
10284 		ttecnt[i] = sfmmup->sfmmu_ttecnt[i] +
10285 		    sfmmup->sfmmu_ismttecnt[i];
10286 	}
10287 
10288 	/* Check pagesizes in use, and possibly reprogram DTLB. */
10289 	if (&mmu_check_page_sizes)
10290 		mmu_check_page_sizes(sfmmup, ttecnt);
10291 
10292 	/*
10293 	 * Calculate the number of 8k ttes to represent the span of these
10294 	 * pages.
10295 	 */
10296 	tte8k_cnt = ttecnt[TTE8K] +
10297 	    (ttecnt[TTE64K] << (MMU_PAGESHIFT64K - MMU_PAGESHIFT)) +
10298 	    (ttecnt[TTE512K] << (MMU_PAGESHIFT512K - MMU_PAGESHIFT));
10299 	if (mmu_page_sizes == max_mmu_page_sizes) {
10300 		tte4m_cnt = ttecnt[TTE4M] +
10301 		    (ttecnt[TTE32M] << (MMU_PAGESHIFT32M - MMU_PAGESHIFT4M)) +
10302 		    (ttecnt[TTE256M] << (MMU_PAGESHIFT256M - MMU_PAGESHIFT4M));
10303 	} else {
10304 		tte4m_cnt = ttecnt[TTE4M];
10305 	}
10306 
10307 	/*
10308 	 * Inflate tte8k_cnt to allow for region large page allocation failure.
10309 	 */
10310 	tte8k_cnt += sfmmup->sfmmu_tsb0_4minflcnt;
10311 
10312 	/*
10313 	 * Inflate TSB sizes by a factor of 2 if this process
10314 	 * uses 4M text pages to minimize extra conflict misses
10315 	 * in the first TSB since without counting text pages
10316 	 * 8K TSB may become too small.
10317 	 *
10318 	 * Also double the size of the second TSB to minimize
10319 	 * extra conflict misses due to competition between 4M text pages
10320 	 * and data pages.
10321 	 *
10322 	 * We need to adjust the second TSB allocation threshold by the
10323 	 * inflation factor, since there is no point in creating a second
10324 	 * TSB when we know all the mappings can fit in the I/D TLBs.
10325 	 */
10326 	sectsb_thresh = tsb_sectsb_threshold;
10327 	if (sfmmup->sfmmu_flags & HAT_4MTEXT_FLAG) {
10328 		tte8k_cnt <<= 1;
10329 		tte4m_cnt <<= 1;
10330 		sectsb_thresh <<= 1;
10331 	}
10332 
10333 	/*
10334 	 * Check to see if our TSB is the right size; we may need to
10335 	 * grow or shrink it.  If the process is small, our work is
10336 	 * finished at this point.
10337 	 */
10338 	if (tte8k_cnt <= tsb_rss_factor && tte4m_cnt <= sectsb_thresh) {
10339 		return;
10340 	}
10341 	sfmmu_size_tsb(sfmmup, growing, tte8k_cnt, tte4m_cnt, sectsb_thresh);
10342 }
10343 
10344 static void
10345 sfmmu_size_tsb(sfmmu_t *sfmmup, int growing, uint64_t tte8k_cnt,
10346 	uint64_t tte4m_cnt, int sectsb_thresh)
10347 {
10348 	int tsb_bits;
10349 	uint_t tsb_szc;
10350 	struct tsb_info *tsbinfop;
10351 	hatlock_t *hatlockp = NULL;
10352 
10353 	hatlockp = sfmmu_hat_enter(sfmmup);
10354 	ASSERT(hatlockp != NULL);
10355 	tsbinfop = sfmmup->sfmmu_tsb;
10356 	ASSERT(tsbinfop != NULL);
10357 
10358 	/*
10359 	 * If we're growing, select the size based on RSS.  If we're
10360 	 * shrinking, leave some room so we don't have to turn around and
10361 	 * grow again immediately.
10362 	 */
10363 	if (growing)
10364 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
10365 	else
10366 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt << 1);
10367 
10368 	if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10369 	    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10370 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10371 		    hatlockp, TSB_SHRINK);
10372 	} else if (growing && tsb_szc > tsbinfop->tsb_szc && TSB_OK_GROW()) {
10373 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10374 		    hatlockp, TSB_GROW);
10375 	}
10376 	tsbinfop = sfmmup->sfmmu_tsb;
10377 
10378 	/*
10379 	 * With the TLB and first TSB out of the way, we need to see if
10380 	 * we need a second TSB for 4M pages.  If we managed to reprogram
10381 	 * the TLB page sizes above, the process will start using this new
10382 	 * TSB right away; otherwise, it will start using it on the next
10383 	 * context switch.  Either way, it's no big deal so there's no
10384 	 * synchronization with the trap handlers here unless we grow the
10385 	 * TSB (in which case it's required to prevent using the old one
10386 	 * after it's freed). Note: second tsb is required for 32M/256M
10387 	 * page sizes.
10388 	 */
10389 	if (tte4m_cnt > sectsb_thresh) {
10390 		/*
10391 		 * If we're growing, select the size based on RSS.  If we're
10392 		 * shrinking, leave some room so we don't have to turn
10393 		 * around and grow again immediately.
10394 		 */
10395 		if (growing)
10396 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
10397 		else
10398 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt << 1);
10399 		if (tsbinfop->tsb_next == NULL) {
10400 			struct tsb_info *newtsb;
10401 			int allocflags = SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)?
10402 			    0 : TSB_ALLOC;
10403 
10404 			sfmmu_hat_exit(hatlockp);
10405 
10406 			/*
10407 			 * Try to allocate a TSB for 4[32|256]M pages.  If we
10408 			 * can't get the size we want, retry w/a minimum sized
10409 			 * TSB.  If that still didn't work, give up; we can
10410 			 * still run without one.
10411 			 */
10412 			tsb_bits = (mmu_page_sizes == max_mmu_page_sizes)?
10413 			    TSB4M|TSB32M|TSB256M:TSB4M;
10414 			if ((sfmmu_tsbinfo_alloc(&newtsb, tsb_szc, tsb_bits,
10415 			    allocflags, sfmmup)) &&
10416 			    (tsb_szc <= TSB_4M_SZCODE ||
10417 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
10418 			    tsb_bits, allocflags, sfmmup)) &&
10419 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_MIN_SZCODE,
10420 			    tsb_bits, allocflags, sfmmup)) {
10421 				return;
10422 			}
10423 
10424 			hatlockp = sfmmu_hat_enter(sfmmup);
10425 
10426 			sfmmu_invalidate_ctx(sfmmup);
10427 
10428 			if (sfmmup->sfmmu_tsb->tsb_next == NULL) {
10429 				sfmmup->sfmmu_tsb->tsb_next = newtsb;
10430 				SFMMU_STAT(sf_tsb_sectsb_create);
10431 				sfmmu_hat_exit(hatlockp);
10432 				return;
10433 			} else {
10434 				/*
10435 				 * It's annoying, but possible for us
10436 				 * to get here.. we dropped the HAT lock
10437 				 * because of locking order in the kmem
10438 				 * allocator, and while we were off getting
10439 				 * our memory, some other thread decided to
10440 				 * do us a favor and won the race to get a
10441 				 * second TSB for this process.  Sigh.
10442 				 */
10443 				sfmmu_hat_exit(hatlockp);
10444 				sfmmu_tsbinfo_free(newtsb);
10445 				return;
10446 			}
10447 		}
10448 
10449 		/*
10450 		 * We have a second TSB, see if it's big enough.
10451 		 */
10452 		tsbinfop = tsbinfop->tsb_next;
10453 
10454 		/*
10455 		 * Check to see if our second TSB is the right size;
10456 		 * we may need to grow or shrink it.
10457 		 * To prevent thrashing (e.g. growing the TSB on a
10458 		 * subsequent map operation), only try to shrink if
10459 		 * the TSB reach exceeds twice the virtual address
10460 		 * space size.
10461 		 */
10462 		if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10463 		    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10464 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10465 			    tsb_szc, hatlockp, TSB_SHRINK);
10466 		} else if (growing && tsb_szc > tsbinfop->tsb_szc &&
10467 		    TSB_OK_GROW()) {
10468 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10469 			    tsb_szc, hatlockp, TSB_GROW);
10470 		}
10471 	}
10472 
10473 	sfmmu_hat_exit(hatlockp);
10474 }
10475 
10476 /*
10477  * Free up a sfmmu
10478  * Since the sfmmu is currently embedded in the hat struct we simply zero
10479  * out our fields and free up the ism map blk list if any.
10480  */
10481 static void
10482 sfmmu_free_sfmmu(sfmmu_t *sfmmup)
10483 {
10484 	ism_blk_t	*blkp, *nx_blkp;
10485 #ifdef	DEBUG
10486 	ism_map_t	*map;
10487 	int 		i;
10488 #endif
10489 
10490 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
10491 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
10492 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
10493 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
10494 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
10495 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
10496 	ASSERT(SF_RGNMAP_ISNULL(sfmmup));
10497 
10498 	sfmmup->sfmmu_free = 0;
10499 	sfmmup->sfmmu_ismhat = 0;
10500 
10501 	blkp = sfmmup->sfmmu_iblk;
10502 	sfmmup->sfmmu_iblk = NULL;
10503 
10504 	while (blkp) {
10505 #ifdef	DEBUG
10506 		map = blkp->iblk_maps;
10507 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
10508 			ASSERT(map[i].imap_seg == 0);
10509 			ASSERT(map[i].imap_ismhat == NULL);
10510 			ASSERT(map[i].imap_ment == NULL);
10511 		}
10512 #endif
10513 		nx_blkp = blkp->iblk_next;
10514 		blkp->iblk_next = NULL;
10515 		blkp->iblk_nextpa = (uint64_t)-1;
10516 		kmem_cache_free(ism_blk_cache, blkp);
10517 		blkp = nx_blkp;
10518 	}
10519 }
10520 
10521 /*
10522  * Locking primitves accessed by HATLOCK macros
10523  */
10524 
10525 #define	SFMMU_SPL_MTX	(0x0)
10526 #define	SFMMU_ML_MTX	(0x1)
10527 
10528 #define	SFMMU_MLSPL_MTX(type, pg)	(((type) == SFMMU_SPL_MTX) ? \
10529 					    SPL_HASH(pg) : MLIST_HASH(pg))
10530 
10531 kmutex_t *
10532 sfmmu_page_enter(struct page *pp)
10533 {
10534 	return (sfmmu_mlspl_enter(pp, SFMMU_SPL_MTX));
10535 }
10536 
10537 void
10538 sfmmu_page_exit(kmutex_t *spl)
10539 {
10540 	mutex_exit(spl);
10541 }
10542 
10543 int
10544 sfmmu_page_spl_held(struct page *pp)
10545 {
10546 	return (sfmmu_mlspl_held(pp, SFMMU_SPL_MTX));
10547 }
10548 
10549 kmutex_t *
10550 sfmmu_mlist_enter(struct page *pp)
10551 {
10552 	return (sfmmu_mlspl_enter(pp, SFMMU_ML_MTX));
10553 }
10554 
10555 void
10556 sfmmu_mlist_exit(kmutex_t *mml)
10557 {
10558 	mutex_exit(mml);
10559 }
10560 
10561 int
10562 sfmmu_mlist_held(struct page *pp)
10563 {
10564 
10565 	return (sfmmu_mlspl_held(pp, SFMMU_ML_MTX));
10566 }
10567 
10568 /*
10569  * Common code for sfmmu_mlist_enter() and sfmmu_page_enter().  For
10570  * sfmmu_mlist_enter() case mml_table lock array is used and for
10571  * sfmmu_page_enter() sfmmu_page_lock lock array is used.
10572  *
10573  * The lock is taken on a root page so that it protects an operation on all
10574  * constituent pages of a large page pp belongs to.
10575  *
10576  * The routine takes a lock from the appropriate array. The lock is determined
10577  * by hashing the root page. After taking the lock this routine checks if the
10578  * root page has the same size code that was used to determine the root (i.e
10579  * that root hasn't changed).  If root page has the expected p_szc field we
10580  * have the right lock and it's returned to the caller. If root's p_szc
10581  * decreased we release the lock and retry from the beginning.  This case can
10582  * happen due to hat_page_demote() decreasing p_szc between our load of p_szc
10583  * value and taking the lock. The number of retries due to p_szc decrease is
10584  * limited by the maximum p_szc value. If p_szc is 0 we return the lock
10585  * determined by hashing pp itself.
10586  *
10587  * If our caller doesn't hold a SE_SHARED or SE_EXCL lock on pp it's also
10588  * possible that p_szc can increase. To increase p_szc a thread has to lock
10589  * all constituent pages EXCL and do hat_pageunload() on all of them. All the
10590  * callers that don't hold a page locked recheck if hmeblk through which pp
10591  * was found still maps this pp.  If it doesn't map it anymore returned lock
10592  * is immediately dropped. Therefore if sfmmu_mlspl_enter() hits the case of
10593  * p_szc increase after taking the lock it returns this lock without further
10594  * retries because in this case the caller doesn't care about which lock was
10595  * taken. The caller will drop it right away.
10596  *
10597  * After the routine returns it's guaranteed that hat_page_demote() can't
10598  * change p_szc field of any of constituent pages of a large page pp belongs
10599  * to as long as pp was either locked at least SHARED prior to this call or
10600  * the caller finds that hment that pointed to this pp still references this
10601  * pp (this also assumes that the caller holds hme hash bucket lock so that
10602  * the same pp can't be remapped into the same hmeblk after it was unmapped by
10603  * hat_pageunload()).
10604  */
10605 static kmutex_t *
10606 sfmmu_mlspl_enter(struct page *pp, int type)
10607 {
10608 	kmutex_t	*mtx;
10609 	uint_t		prev_rszc = UINT_MAX;
10610 	page_t		*rootpp;
10611 	uint_t		szc;
10612 	uint_t		rszc;
10613 	uint_t		pszc = pp->p_szc;
10614 
10615 	ASSERT(pp != NULL);
10616 
10617 again:
10618 	if (pszc == 0) {
10619 		mtx = SFMMU_MLSPL_MTX(type, pp);
10620 		mutex_enter(mtx);
10621 		return (mtx);
10622 	}
10623 
10624 	/* The lock lives in the root page */
10625 	rootpp = PP_GROUPLEADER(pp, pszc);
10626 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10627 	mutex_enter(mtx);
10628 
10629 	/*
10630 	 * Return mml in the following 3 cases:
10631 	 *
10632 	 * 1) If pp itself is root since if its p_szc decreased before we took
10633 	 * the lock pp is still the root of smaller szc page. And if its p_szc
10634 	 * increased it doesn't matter what lock we return (see comment in
10635 	 * front of this routine).
10636 	 *
10637 	 * 2) If pp's not root but rootpp is the root of a rootpp->p_szc size
10638 	 * large page we have the right lock since any previous potential
10639 	 * hat_page_demote() is done demoting from greater than current root's
10640 	 * p_szc because hat_page_demote() changes root's p_szc last. No
10641 	 * further hat_page_demote() can start or be in progress since it
10642 	 * would need the same lock we currently hold.
10643 	 *
10644 	 * 3) If rootpp's p_szc increased since previous iteration it doesn't
10645 	 * matter what lock we return (see comment in front of this routine).
10646 	 */
10647 	if (pp == rootpp || (rszc = rootpp->p_szc) == pszc ||
10648 	    rszc >= prev_rszc) {
10649 		return (mtx);
10650 	}
10651 
10652 	/*
10653 	 * hat_page_demote() could have decreased root's p_szc.
10654 	 * In this case pp's p_szc must also be smaller than pszc.
10655 	 * Retry.
10656 	 */
10657 	if (rszc < pszc) {
10658 		szc = pp->p_szc;
10659 		if (szc < pszc) {
10660 			mutex_exit(mtx);
10661 			pszc = szc;
10662 			goto again;
10663 		}
10664 		/*
10665 		 * pp's p_szc increased after it was decreased.
10666 		 * page cannot be mapped. Return current lock. The caller
10667 		 * will drop it right away.
10668 		 */
10669 		return (mtx);
10670 	}
10671 
10672 	/*
10673 	 * root's p_szc is greater than pp's p_szc.
10674 	 * hat_page_demote() is not done with all pages
10675 	 * yet. Wait for it to complete.
10676 	 */
10677 	mutex_exit(mtx);
10678 	rootpp = PP_GROUPLEADER(rootpp, rszc);
10679 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10680 	mutex_enter(mtx);
10681 	mutex_exit(mtx);
10682 	prev_rszc = rszc;
10683 	goto again;
10684 }
10685 
10686 static int
10687 sfmmu_mlspl_held(struct page *pp, int type)
10688 {
10689 	kmutex_t	*mtx;
10690 
10691 	ASSERT(pp != NULL);
10692 	/* The lock lives in the root page */
10693 	pp = PP_PAGEROOT(pp);
10694 	ASSERT(pp != NULL);
10695 
10696 	mtx = SFMMU_MLSPL_MTX(type, pp);
10697 	return (MUTEX_HELD(mtx));
10698 }
10699 
10700 static uint_t
10701 sfmmu_get_free_hblk(struct hme_blk **hmeblkpp, uint_t critical)
10702 {
10703 	struct  hme_blk *hblkp;
10704 
10705 
10706 	if (freehblkp != NULL) {
10707 		mutex_enter(&freehblkp_lock);
10708 		if (freehblkp != NULL) {
10709 			/*
10710 			 * If the current thread is owning hblk_reserve OR
10711 			 * critical request from sfmmu_hblk_steal()
10712 			 * let it succeed even if freehblkcnt is really low.
10713 			 */
10714 			if (freehblkcnt <= HBLK_RESERVE_MIN && !critical) {
10715 				SFMMU_STAT(sf_get_free_throttle);
10716 				mutex_exit(&freehblkp_lock);
10717 				return (0);
10718 			}
10719 			freehblkcnt--;
10720 			*hmeblkpp = freehblkp;
10721 			hblkp = *hmeblkpp;
10722 			freehblkp = hblkp->hblk_next;
10723 			mutex_exit(&freehblkp_lock);
10724 			hblkp->hblk_next = NULL;
10725 			SFMMU_STAT(sf_get_free_success);
10726 
10727 			ASSERT(hblkp->hblk_hmecnt == 0);
10728 			ASSERT(hblkp->hblk_vcnt == 0);
10729 			ASSERT(hblkp->hblk_nextpa == va_to_pa((caddr_t)hblkp));
10730 
10731 			return (1);
10732 		}
10733 		mutex_exit(&freehblkp_lock);
10734 	}
10735 
10736 	/* Check cpu hblk pending queues */
10737 	if ((*hmeblkpp = sfmmu_check_pending_hblks(TTE8K)) != NULL) {
10738 		hblkp = *hmeblkpp;
10739 		hblkp->hblk_next = NULL;
10740 		hblkp->hblk_nextpa = va_to_pa((caddr_t)hblkp);
10741 
10742 		ASSERT(hblkp->hblk_hmecnt == 0);
10743 		ASSERT(hblkp->hblk_vcnt == 0);
10744 
10745 		return (1);
10746 	}
10747 
10748 	SFMMU_STAT(sf_get_free_fail);
10749 	return (0);
10750 }
10751 
10752 static uint_t
10753 sfmmu_put_free_hblk(struct hme_blk *hmeblkp, uint_t critical)
10754 {
10755 	struct  hme_blk *hblkp;
10756 
10757 	ASSERT(hmeblkp->hblk_hmecnt == 0);
10758 	ASSERT(hmeblkp->hblk_vcnt == 0);
10759 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
10760 
10761 	/*
10762 	 * If the current thread is mapping into kernel space,
10763 	 * let it succede even if freehblkcnt is max
10764 	 * so that it will avoid freeing it to kmem.
10765 	 * This will prevent stack overflow due to
10766 	 * possible recursion since kmem_cache_free()
10767 	 * might require creation of a slab which
10768 	 * in turn needs an hmeblk to map that slab;
10769 	 * let's break this vicious chain at the first
10770 	 * opportunity.
10771 	 */
10772 	if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10773 		mutex_enter(&freehblkp_lock);
10774 		if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10775 			SFMMU_STAT(sf_put_free_success);
10776 			freehblkcnt++;
10777 			hmeblkp->hblk_next = freehblkp;
10778 			freehblkp = hmeblkp;
10779 			mutex_exit(&freehblkp_lock);
10780 			return (1);
10781 		}
10782 		mutex_exit(&freehblkp_lock);
10783 	}
10784 
10785 	/*
10786 	 * Bring down freehblkcnt to HBLK_RESERVE_CNT. We are here
10787 	 * only if freehblkcnt is at least HBLK_RESERVE_CNT *and*
10788 	 * we are not in the process of mapping into kernel space.
10789 	 */
10790 	ASSERT(!critical);
10791 	while (freehblkcnt > HBLK_RESERVE_CNT) {
10792 		mutex_enter(&freehblkp_lock);
10793 		if (freehblkcnt > HBLK_RESERVE_CNT) {
10794 			freehblkcnt--;
10795 			hblkp = freehblkp;
10796 			freehblkp = hblkp->hblk_next;
10797 			mutex_exit(&freehblkp_lock);
10798 			ASSERT(get_hblk_cache(hblkp) == sfmmu8_cache);
10799 			kmem_cache_free(sfmmu8_cache, hblkp);
10800 			continue;
10801 		}
10802 		mutex_exit(&freehblkp_lock);
10803 	}
10804 	SFMMU_STAT(sf_put_free_fail);
10805 	return (0);
10806 }
10807 
10808 static void
10809 sfmmu_hblk_swap(struct hme_blk *new)
10810 {
10811 	struct hme_blk *old, *hblkp, *prev;
10812 	uint64_t newpa;
10813 	caddr_t	base, vaddr, endaddr;
10814 	struct hmehash_bucket *hmebp;
10815 	struct sf_hment *osfhme, *nsfhme;
10816 	page_t *pp;
10817 	kmutex_t *pml;
10818 	tte_t tte;
10819 	struct hme_blk *list = NULL;
10820 
10821 #ifdef	DEBUG
10822 	hmeblk_tag		hblktag;
10823 	struct hme_blk		*found;
10824 #endif
10825 	old = HBLK_RESERVE;
10826 	ASSERT(!old->hblk_shared);
10827 
10828 	/*
10829 	 * save pa before bcopy clobbers it
10830 	 */
10831 	newpa = new->hblk_nextpa;
10832 
10833 	base = (caddr_t)get_hblk_base(old);
10834 	endaddr = base + get_hblk_span(old);
10835 
10836 	/*
10837 	 * acquire hash bucket lock.
10838 	 */
10839 	hmebp = sfmmu_tteload_acquire_hashbucket(ksfmmup, base, TTE8K,
10840 	    SFMMU_INVALID_SHMERID);
10841 
10842 	/*
10843 	 * copy contents from old to new
10844 	 */
10845 	bcopy((void *)old, (void *)new, HME8BLK_SZ);
10846 
10847 	/*
10848 	 * add new to hash chain
10849 	 */
10850 	sfmmu_hblk_hash_add(hmebp, new, newpa);
10851 
10852 	/*
10853 	 * search hash chain for hblk_reserve; this needs to be performed
10854 	 * after adding new, otherwise prev won't correspond to the hblk which
10855 	 * is prior to old in hash chain when we call sfmmu_hblk_hash_rm to
10856 	 * remove old later.
10857 	 */
10858 	for (prev = NULL,
10859 	    hblkp = hmebp->hmeblkp; hblkp != NULL && hblkp != old;
10860 	    prev = hblkp, hblkp = hblkp->hblk_next)
10861 		;
10862 
10863 	if (hblkp != old)
10864 		panic("sfmmu_hblk_swap: hblk_reserve not found");
10865 
10866 	/*
10867 	 * p_mapping list is still pointing to hments in hblk_reserve;
10868 	 * fix up p_mapping list so that they point to hments in new.
10869 	 *
10870 	 * Since all these mappings are created by hblk_reserve_thread
10871 	 * on the way and it's using at least one of the buffers from each of
10872 	 * the newly minted slabs, there is no danger of any of these
10873 	 * mappings getting unloaded by another thread.
10874 	 *
10875 	 * tsbmiss could only modify ref/mod bits of hments in old/new.
10876 	 * Since all of these hments hold mappings established by segkmem
10877 	 * and mappings in segkmem are setup with HAT_NOSYNC, ref/mod bits
10878 	 * have no meaning for the mappings in hblk_reserve.  hments in
10879 	 * old and new are identical except for ref/mod bits.
10880 	 */
10881 	for (vaddr = base; vaddr < endaddr; vaddr += TTEBYTES(TTE8K)) {
10882 
10883 		HBLKTOHME(osfhme, old, vaddr);
10884 		sfmmu_copytte(&osfhme->hme_tte, &tte);
10885 
10886 		if (TTE_IS_VALID(&tte)) {
10887 			if ((pp = osfhme->hme_page) == NULL)
10888 				panic("sfmmu_hblk_swap: page not mapped");
10889 
10890 			pml = sfmmu_mlist_enter(pp);
10891 
10892 			if (pp != osfhme->hme_page)
10893 				panic("sfmmu_hblk_swap: mapping changed");
10894 
10895 			HBLKTOHME(nsfhme, new, vaddr);
10896 
10897 			HME_ADD(nsfhme, pp);
10898 			HME_SUB(osfhme, pp);
10899 
10900 			sfmmu_mlist_exit(pml);
10901 		}
10902 	}
10903 
10904 	/*
10905 	 * remove old from hash chain
10906 	 */
10907 	sfmmu_hblk_hash_rm(hmebp, old, prev, &list, 1);
10908 
10909 #ifdef	DEBUG
10910 
10911 	hblktag.htag_id = ksfmmup;
10912 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
10913 	hblktag.htag_bspage = HME_HASH_BSPAGE(base, HME_HASH_SHIFT(TTE8K));
10914 	hblktag.htag_rehash = HME_HASH_REHASH(TTE8K);
10915 	HME_HASH_FAST_SEARCH(hmebp, hblktag, found);
10916 
10917 	if (found != new)
10918 		panic("sfmmu_hblk_swap: new hblk not found");
10919 #endif
10920 
10921 	SFMMU_HASH_UNLOCK(hmebp);
10922 
10923 	/*
10924 	 * Reset hblk_reserve
10925 	 */
10926 	bzero((void *)old, HME8BLK_SZ);
10927 	old->hblk_nextpa = va_to_pa((caddr_t)old);
10928 }
10929 
10930 /*
10931  * Grab the mlist mutex for both pages passed in.
10932  *
10933  * low and high will be returned as pointers to the mutexes for these pages.
10934  * low refers to the mutex residing in the lower bin of the mlist hash, while
10935  * high refers to the mutex residing in the higher bin of the mlist hash.  This
10936  * is due to the locking order restrictions on the same thread grabbing
10937  * multiple mlist mutexes.  The low lock must be acquired before the high lock.
10938  *
10939  * If both pages hash to the same mutex, only grab that single mutex, and
10940  * high will be returned as NULL
10941  * If the pages hash to different bins in the hash, grab the lower addressed
10942  * lock first and then the higher addressed lock in order to follow the locking
10943  * rules involved with the same thread grabbing multiple mlist mutexes.
10944  * low and high will both have non-NULL values.
10945  */
10946 static void
10947 sfmmu_mlist_reloc_enter(struct page *targ, struct page *repl,
10948     kmutex_t **low, kmutex_t **high)
10949 {
10950 	kmutex_t	*mml_targ, *mml_repl;
10951 
10952 	/*
10953 	 * no need to do the dance around szc as in sfmmu_mlist_enter()
10954 	 * because this routine is only called by hat_page_relocate() and all
10955 	 * targ and repl pages are already locked EXCL so szc can't change.
10956 	 */
10957 
10958 	mml_targ = MLIST_HASH(PP_PAGEROOT(targ));
10959 	mml_repl = MLIST_HASH(PP_PAGEROOT(repl));
10960 
10961 	if (mml_targ == mml_repl) {
10962 		*low = mml_targ;
10963 		*high = NULL;
10964 	} else {
10965 		if (mml_targ < mml_repl) {
10966 			*low = mml_targ;
10967 			*high = mml_repl;
10968 		} else {
10969 			*low = mml_repl;
10970 			*high = mml_targ;
10971 		}
10972 	}
10973 
10974 	mutex_enter(*low);
10975 	if (*high)
10976 		mutex_enter(*high);
10977 }
10978 
10979 static void
10980 sfmmu_mlist_reloc_exit(kmutex_t *low, kmutex_t *high)
10981 {
10982 	if (high)
10983 		mutex_exit(high);
10984 	mutex_exit(low);
10985 }
10986 
10987 static hatlock_t *
10988 sfmmu_hat_enter(sfmmu_t *sfmmup)
10989 {
10990 	hatlock_t	*hatlockp;
10991 
10992 	if (sfmmup != ksfmmup) {
10993 		hatlockp = TSB_HASH(sfmmup);
10994 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
10995 		return (hatlockp);
10996 	}
10997 	return (NULL);
10998 }
10999 
11000 static hatlock_t *
11001 sfmmu_hat_tryenter(sfmmu_t *sfmmup)
11002 {
11003 	hatlock_t	*hatlockp;
11004 
11005 	if (sfmmup != ksfmmup) {
11006 		hatlockp = TSB_HASH(sfmmup);
11007 		if (mutex_tryenter(HATLOCK_MUTEXP(hatlockp)) == 0)
11008 			return (NULL);
11009 		return (hatlockp);
11010 	}
11011 	return (NULL);
11012 }
11013 
11014 static void
11015 sfmmu_hat_exit(hatlock_t *hatlockp)
11016 {
11017 	if (hatlockp != NULL)
11018 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
11019 }
11020 
11021 static void
11022 sfmmu_hat_lock_all(void)
11023 {
11024 	int i;
11025 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
11026 		mutex_enter(HATLOCK_MUTEXP(&hat_lock[i]));
11027 }
11028 
11029 static void
11030 sfmmu_hat_unlock_all(void)
11031 {
11032 	int i;
11033 	for (i = SFMMU_NUM_LOCK - 1; i >= 0; i--)
11034 		mutex_exit(HATLOCK_MUTEXP(&hat_lock[i]));
11035 }
11036 
11037 int
11038 sfmmu_hat_lock_held(sfmmu_t *sfmmup)
11039 {
11040 	ASSERT(sfmmup != ksfmmup);
11041 	return (MUTEX_HELD(HATLOCK_MUTEXP(TSB_HASH(sfmmup))));
11042 }
11043 
11044 /*
11045  * Locking primitives to provide consistency between ISM unmap
11046  * and other operations.  Since ISM unmap can take a long time, we
11047  * use HAT_ISMBUSY flag (protected by the hatlock) to avoid creating
11048  * contention on the hatlock buckets while ISM segments are being
11049  * unmapped.  The tradeoff is that the flags don't prevent priority
11050  * inversion from occurring, so we must request kernel priority in
11051  * case we have to sleep to keep from getting buried while holding
11052  * the HAT_ISMBUSY flag set, which in turn could block other kernel
11053  * threads from running (for example, in sfmmu_uvatopfn()).
11054  */
11055 static void
11056 sfmmu_ismhat_enter(sfmmu_t *sfmmup, int hatlock_held)
11057 {
11058 	hatlock_t *hatlockp;
11059 
11060 	THREAD_KPRI_REQUEST();
11061 	if (!hatlock_held)
11062 		hatlockp = sfmmu_hat_enter(sfmmup);
11063 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY))
11064 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
11065 	SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
11066 	if (!hatlock_held)
11067 		sfmmu_hat_exit(hatlockp);
11068 }
11069 
11070 static void
11071 sfmmu_ismhat_exit(sfmmu_t *sfmmup, int hatlock_held)
11072 {
11073 	hatlock_t *hatlockp;
11074 
11075 	if (!hatlock_held)
11076 		hatlockp = sfmmu_hat_enter(sfmmup);
11077 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
11078 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
11079 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11080 	if (!hatlock_held)
11081 		sfmmu_hat_exit(hatlockp);
11082 	THREAD_KPRI_RELEASE();
11083 }
11084 
11085 /*
11086  *
11087  * Algorithm:
11088  *
11089  * (1) if segkmem is not ready, allocate hblk from an array of pre-alloc'ed
11090  *	hblks.
11091  *
11092  * (2) if we are allocating an hblk for mapping a slab in sfmmu_cache,
11093  *
11094  * 		(a) try to return an hblk from reserve pool of free hblks;
11095  *		(b) if the reserve pool is empty, acquire hblk_reserve_lock
11096  *		    and return hblk_reserve.
11097  *
11098  * (3) call kmem_cache_alloc() to allocate hblk;
11099  *
11100  *		(a) if hblk_reserve_lock is held by the current thread,
11101  *		    atomically replace hblk_reserve by the hblk that is
11102  *		    returned by kmem_cache_alloc; release hblk_reserve_lock
11103  *		    and call kmem_cache_alloc() again.
11104  *		(b) if reserve pool is not full, add the hblk that is
11105  *		    returned by kmem_cache_alloc to reserve pool and
11106  *		    call kmem_cache_alloc again.
11107  *
11108  */
11109 static struct hme_blk *
11110 sfmmu_hblk_alloc(sfmmu_t *sfmmup, caddr_t vaddr,
11111 	struct hmehash_bucket *hmebp, uint_t size, hmeblk_tag hblktag,
11112 	uint_t flags, uint_t rid)
11113 {
11114 	struct hme_blk *hmeblkp = NULL;
11115 	struct hme_blk *newhblkp;
11116 	struct hme_blk *shw_hblkp = NULL;
11117 	struct kmem_cache *sfmmu_cache = NULL;
11118 	uint64_t hblkpa;
11119 	ulong_t index;
11120 	uint_t owner;		/* set to 1 if using hblk_reserve */
11121 	uint_t forcefree;
11122 	int sleep;
11123 	sf_srd_t *srdp;
11124 	sf_region_t *rgnp;
11125 
11126 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11127 	ASSERT(hblktag.htag_rid == rid);
11128 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
11129 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
11130 	    IS_P2ALIGNED(vaddr, TTEBYTES(size)));
11131 
11132 	/*
11133 	 * If segkmem is not created yet, allocate from static hmeblks
11134 	 * created at the end of startup_modules().  See the block comment
11135 	 * in startup_modules() describing how we estimate the number of
11136 	 * static hmeblks that will be needed during re-map.
11137 	 */
11138 	if (!hblk_alloc_dynamic) {
11139 
11140 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11141 
11142 		if (size == TTE8K) {
11143 			index = nucleus_hblk8.index;
11144 			if (index >= nucleus_hblk8.len) {
11145 				/*
11146 				 * If we panic here, see startup_modules() to
11147 				 * make sure that we are calculating the
11148 				 * number of hblk8's that we need correctly.
11149 				 */
11150 				prom_panic("no nucleus hblk8 to allocate");
11151 			}
11152 			hmeblkp =
11153 			    (struct hme_blk *)&nucleus_hblk8.list[index];
11154 			nucleus_hblk8.index++;
11155 			SFMMU_STAT(sf_hblk8_nalloc);
11156 		} else {
11157 			index = nucleus_hblk1.index;
11158 			if (nucleus_hblk1.index >= nucleus_hblk1.len) {
11159 				/*
11160 				 * If we panic here, see startup_modules().
11161 				 * Most likely you need to update the
11162 				 * calculation of the number of hblk1 elements
11163 				 * that the kernel needs to boot.
11164 				 */
11165 				prom_panic("no nucleus hblk1 to allocate");
11166 			}
11167 			hmeblkp =
11168 			    (struct hme_blk *)&nucleus_hblk1.list[index];
11169 			nucleus_hblk1.index++;
11170 			SFMMU_STAT(sf_hblk1_nalloc);
11171 		}
11172 
11173 		goto hblk_init;
11174 	}
11175 
11176 	SFMMU_HASH_UNLOCK(hmebp);
11177 
11178 	if (sfmmup != KHATID && !SFMMU_IS_SHMERID_VALID(rid)) {
11179 		if (mmu_page_sizes == max_mmu_page_sizes) {
11180 			if (size < TTE256M)
11181 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
11182 				    size, flags);
11183 		} else {
11184 			if (size < TTE4M)
11185 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
11186 				    size, flags);
11187 		}
11188 	} else if (SFMMU_IS_SHMERID_VALID(rid)) {
11189 		/*
11190 		 * Shared hmes use per region bitmaps in rgn_hmeflag
11191 		 * rather than shadow hmeblks to keep track of the
11192 		 * mapping sizes which have been allocated for the region.
11193 		 * Here we cleanup old invalid hmeblks with this rid,
11194 		 * which may be left around by pageunload().
11195 		 */
11196 		int ttesz;
11197 		caddr_t va;
11198 		caddr_t	eva = vaddr + TTEBYTES(size);
11199 
11200 		ASSERT(sfmmup != KHATID);
11201 
11202 		srdp = sfmmup->sfmmu_srdp;
11203 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11204 		rgnp = srdp->srd_hmergnp[rid];
11205 		ASSERT(rgnp != NULL && rgnp->rgn_id == rid);
11206 		ASSERT(rgnp->rgn_refcnt != 0);
11207 		ASSERT(size <= rgnp->rgn_pgszc);
11208 
11209 		ttesz = HBLK_MIN_TTESZ;
11210 		do {
11211 			if (!(rgnp->rgn_hmeflags & (0x1 << ttesz))) {
11212 				continue;
11213 			}
11214 
11215 			if (ttesz > size && ttesz != HBLK_MIN_TTESZ) {
11216 				sfmmu_cleanup_rhblk(srdp, vaddr, rid, ttesz);
11217 			} else if (ttesz < size) {
11218 				for (va = vaddr; va < eva;
11219 				    va += TTEBYTES(ttesz)) {
11220 					sfmmu_cleanup_rhblk(srdp, va, rid,
11221 					    ttesz);
11222 				}
11223 			}
11224 		} while (++ttesz <= rgnp->rgn_pgszc);
11225 	}
11226 
11227 fill_hblk:
11228 	owner = (hblk_reserve_thread == curthread) ? 1 : 0;
11229 
11230 	if (owner && size == TTE8K) {
11231 
11232 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11233 		/*
11234 		 * We are really in a tight spot. We already own
11235 		 * hblk_reserve and we need another hblk.  In anticipation
11236 		 * of this kind of scenario, we specifically set aside
11237 		 * HBLK_RESERVE_MIN number of hblks to be used exclusively
11238 		 * by owner of hblk_reserve.
11239 		 */
11240 		SFMMU_STAT(sf_hblk_recurse_cnt);
11241 
11242 		if (!sfmmu_get_free_hblk(&hmeblkp, 1))
11243 			panic("sfmmu_hblk_alloc: reserve list is empty");
11244 
11245 		goto hblk_verify;
11246 	}
11247 
11248 	ASSERT(!owner);
11249 
11250 	if ((flags & HAT_NO_KALLOC) == 0) {
11251 
11252 		sfmmu_cache = ((size == TTE8K) ? sfmmu8_cache : sfmmu1_cache);
11253 		sleep = ((sfmmup == KHATID) ? KM_NOSLEEP : KM_SLEEP);
11254 
11255 		if ((hmeblkp = kmem_cache_alloc(sfmmu_cache, sleep)) == NULL) {
11256 			hmeblkp = sfmmu_hblk_steal(size);
11257 		} else {
11258 			/*
11259 			 * if we are the owner of hblk_reserve,
11260 			 * swap hblk_reserve with hmeblkp and
11261 			 * start a fresh life.  Hope things go
11262 			 * better this time.
11263 			 */
11264 			if (hblk_reserve_thread == curthread) {
11265 				ASSERT(sfmmu_cache == sfmmu8_cache);
11266 				sfmmu_hblk_swap(hmeblkp);
11267 				hblk_reserve_thread = NULL;
11268 				mutex_exit(&hblk_reserve_lock);
11269 				goto fill_hblk;
11270 			}
11271 			/*
11272 			 * let's donate this hblk to our reserve list if
11273 			 * we are not mapping kernel range
11274 			 */
11275 			if (size == TTE8K && sfmmup != KHATID) {
11276 				if (sfmmu_put_free_hblk(hmeblkp, 0))
11277 					goto fill_hblk;
11278 			}
11279 		}
11280 	} else {
11281 		/*
11282 		 * We are here to map the slab in sfmmu8_cache; let's
11283 		 * check if we could tap our reserve list; if successful,
11284 		 * this will avoid the pain of going thru sfmmu_hblk_swap
11285 		 */
11286 		SFMMU_STAT(sf_hblk_slab_cnt);
11287 		if (!sfmmu_get_free_hblk(&hmeblkp, 0)) {
11288 			/*
11289 			 * let's start hblk_reserve dance
11290 			 */
11291 			SFMMU_STAT(sf_hblk_reserve_cnt);
11292 			owner = 1;
11293 			mutex_enter(&hblk_reserve_lock);
11294 			hmeblkp = HBLK_RESERVE;
11295 			hblk_reserve_thread = curthread;
11296 		}
11297 	}
11298 
11299 hblk_verify:
11300 	ASSERT(hmeblkp != NULL);
11301 	set_hblk_sz(hmeblkp, size);
11302 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
11303 	SFMMU_HASH_LOCK(hmebp);
11304 	HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11305 	if (newhblkp != NULL) {
11306 		SFMMU_HASH_UNLOCK(hmebp);
11307 		if (hmeblkp != HBLK_RESERVE) {
11308 			/*
11309 			 * This is really tricky!
11310 			 *
11311 			 * vmem_alloc(vmem_seg_arena)
11312 			 *  vmem_alloc(vmem_internal_arena)
11313 			 *   segkmem_alloc(heap_arena)
11314 			 *    vmem_alloc(heap_arena)
11315 			 *    page_create()
11316 			 *    hat_memload()
11317 			 *	kmem_cache_free()
11318 			 *	 kmem_cache_alloc()
11319 			 *	  kmem_slab_create()
11320 			 *	   vmem_alloc(kmem_internal_arena)
11321 			 *	    segkmem_alloc(heap_arena)
11322 			 *		vmem_alloc(heap_arena)
11323 			 *		page_create()
11324 			 *		hat_memload()
11325 			 *		  kmem_cache_free()
11326 			 *		...
11327 			 *
11328 			 * Thus, hat_memload() could call kmem_cache_free
11329 			 * for enough number of times that we could easily
11330 			 * hit the bottom of the stack or run out of reserve
11331 			 * list of vmem_seg structs.  So, we must donate
11332 			 * this hblk to reserve list if it's allocated
11333 			 * from sfmmu8_cache *and* mapping kernel range.
11334 			 * We don't need to worry about freeing hmeblk1's
11335 			 * to kmem since they don't map any kmem slabs.
11336 			 *
11337 			 * Note: When segkmem supports largepages, we must
11338 			 * free hmeblk1's to reserve list as well.
11339 			 */
11340 			forcefree = (sfmmup == KHATID) ? 1 : 0;
11341 			if (size == TTE8K &&
11342 			    sfmmu_put_free_hblk(hmeblkp, forcefree)) {
11343 				goto re_verify;
11344 			}
11345 			ASSERT(sfmmup != KHATID);
11346 			kmem_cache_free(get_hblk_cache(hmeblkp), hmeblkp);
11347 		} else {
11348 			/*
11349 			 * Hey! we don't need hblk_reserve any more.
11350 			 */
11351 			ASSERT(owner);
11352 			hblk_reserve_thread = NULL;
11353 			mutex_exit(&hblk_reserve_lock);
11354 			owner = 0;
11355 		}
11356 re_verify:
11357 		/*
11358 		 * let's check if the goodies are still present
11359 		 */
11360 		SFMMU_HASH_LOCK(hmebp);
11361 		HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11362 		if (newhblkp != NULL) {
11363 			/*
11364 			 * return newhblkp if it's not hblk_reserve;
11365 			 * if newhblkp is hblk_reserve, return it
11366 			 * _only if_ we are the owner of hblk_reserve.
11367 			 */
11368 			if (newhblkp != HBLK_RESERVE || owner) {
11369 				ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
11370 				    newhblkp->hblk_shared);
11371 				ASSERT(SFMMU_IS_SHMERID_VALID(rid) ||
11372 				    !newhblkp->hblk_shared);
11373 				return (newhblkp);
11374 			} else {
11375 				/*
11376 				 * we just hit hblk_reserve in the hash and
11377 				 * we are not the owner of that;
11378 				 *
11379 				 * block until hblk_reserve_thread completes
11380 				 * swapping hblk_reserve and try the dance
11381 				 * once again.
11382 				 */
11383 				SFMMU_HASH_UNLOCK(hmebp);
11384 				mutex_enter(&hblk_reserve_lock);
11385 				mutex_exit(&hblk_reserve_lock);
11386 				SFMMU_STAT(sf_hblk_reserve_hit);
11387 				goto fill_hblk;
11388 			}
11389 		} else {
11390 			/*
11391 			 * it's no more! try the dance once again.
11392 			 */
11393 			SFMMU_HASH_UNLOCK(hmebp);
11394 			goto fill_hblk;
11395 		}
11396 	}
11397 
11398 hblk_init:
11399 	if (SFMMU_IS_SHMERID_VALID(rid)) {
11400 		uint16_t tteflag = 0x1 <<
11401 		    ((size < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ : size);
11402 
11403 		if (!(rgnp->rgn_hmeflags & tteflag)) {
11404 			atomic_or_16(&rgnp->rgn_hmeflags, tteflag);
11405 		}
11406 		hmeblkp->hblk_shared = 1;
11407 	} else {
11408 		hmeblkp->hblk_shared = 0;
11409 	}
11410 	set_hblk_sz(hmeblkp, size);
11411 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11412 	hmeblkp->hblk_next = (struct hme_blk *)NULL;
11413 	hmeblkp->hblk_tag = hblktag;
11414 	hmeblkp->hblk_shadow = shw_hblkp;
11415 	hblkpa = hmeblkp->hblk_nextpa;
11416 	hmeblkp->hblk_nextpa = HMEBLK_ENDPA;
11417 
11418 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
11419 	ASSERT(get_hblk_span(hmeblkp) == HMEBLK_SPAN(size));
11420 	ASSERT(hmeblkp->hblk_hmecnt == 0);
11421 	ASSERT(hmeblkp->hblk_vcnt == 0);
11422 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11423 	ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
11424 	sfmmu_hblk_hash_add(hmebp, hmeblkp, hblkpa);
11425 	return (hmeblkp);
11426 }
11427 
11428 /*
11429  * This function cleans up the hme_blk and returns it to the free list.
11430  */
11431 /* ARGSUSED */
11432 static void
11433 sfmmu_hblk_free(struct hme_blk **listp)
11434 {
11435 	struct hme_blk *hmeblkp, *next_hmeblkp;
11436 	int		size;
11437 	uint_t		critical;
11438 	uint64_t	hblkpa;
11439 
11440 	ASSERT(*listp != NULL);
11441 
11442 	hmeblkp = *listp;
11443 	while (hmeblkp != NULL) {
11444 		next_hmeblkp = hmeblkp->hblk_next;
11445 		ASSERT(!hmeblkp->hblk_hmecnt);
11446 		ASSERT(!hmeblkp->hblk_vcnt);
11447 		ASSERT(!hmeblkp->hblk_lckcnt);
11448 		ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
11449 		ASSERT(hmeblkp->hblk_shared == 0);
11450 		ASSERT(hmeblkp->hblk_shw_bit == 0);
11451 		ASSERT(hmeblkp->hblk_shadow == NULL);
11452 
11453 		hblkpa = va_to_pa((caddr_t)hmeblkp);
11454 		ASSERT(hblkpa != (uint64_t)-1);
11455 		critical = (hblktosfmmu(hmeblkp) == KHATID) ? 1 : 0;
11456 
11457 		size = get_hblk_ttesz(hmeblkp);
11458 		hmeblkp->hblk_next = NULL;
11459 		hmeblkp->hblk_nextpa = hblkpa;
11460 
11461 		if (hmeblkp->hblk_nuc_bit == 0) {
11462 
11463 			if (size != TTE8K ||
11464 			    !sfmmu_put_free_hblk(hmeblkp, critical))
11465 				kmem_cache_free(get_hblk_cache(hmeblkp),
11466 				    hmeblkp);
11467 		}
11468 		hmeblkp = next_hmeblkp;
11469 	}
11470 }
11471 
11472 #define	BUCKETS_TO_SEARCH_BEFORE_UNLOAD	30
11473 #define	SFMMU_HBLK_STEAL_THRESHOLD 5
11474 
11475 static uint_t sfmmu_hblk_steal_twice;
11476 static uint_t sfmmu_hblk_steal_count, sfmmu_hblk_steal_unload_count;
11477 
11478 /*
11479  * Steal a hmeblk from user or kernel hme hash lists.
11480  * For 8K tte grab one from reserve pool (freehblkp) before proceeding to
11481  * steal and if we fail to steal after SFMMU_HBLK_STEAL_THRESHOLD attempts
11482  * tap into critical reserve of freehblkp.
11483  * Note: We remain looping in this routine until we find one.
11484  */
11485 static struct hme_blk *
11486 sfmmu_hblk_steal(int size)
11487 {
11488 	static struct hmehash_bucket *uhmehash_steal_hand = NULL;
11489 	struct hmehash_bucket *hmebp;
11490 	struct hme_blk *hmeblkp = NULL, *pr_hblk;
11491 	uint64_t hblkpa;
11492 	int i;
11493 	uint_t loop_cnt = 0, critical;
11494 
11495 	for (;;) {
11496 		/* Check cpu hblk pending queues */
11497 		if ((hmeblkp = sfmmu_check_pending_hblks(size)) != NULL) {
11498 			hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
11499 			ASSERT(hmeblkp->hblk_hmecnt == 0);
11500 			ASSERT(hmeblkp->hblk_vcnt == 0);
11501 			return (hmeblkp);
11502 		}
11503 
11504 		if (size == TTE8K) {
11505 			critical =
11506 			    (++loop_cnt > SFMMU_HBLK_STEAL_THRESHOLD) ? 1 : 0;
11507 			if (sfmmu_get_free_hblk(&hmeblkp, critical))
11508 				return (hmeblkp);
11509 		}
11510 
11511 		hmebp = (uhmehash_steal_hand == NULL) ? uhme_hash :
11512 		    uhmehash_steal_hand;
11513 		ASSERT(hmebp >= uhme_hash && hmebp <= &uhme_hash[UHMEHASH_SZ]);
11514 
11515 		for (i = 0; hmeblkp == NULL && i <= UHMEHASH_SZ +
11516 		    BUCKETS_TO_SEARCH_BEFORE_UNLOAD; i++) {
11517 			SFMMU_HASH_LOCK(hmebp);
11518 			hmeblkp = hmebp->hmeblkp;
11519 			hblkpa = hmebp->hmeh_nextpa;
11520 			pr_hblk = NULL;
11521 			while (hmeblkp) {
11522 				/*
11523 				 * check if it is a hmeblk that is not locked
11524 				 * and not shared. skip shadow hmeblks with
11525 				 * shadow_mask set i.e valid count non zero.
11526 				 */
11527 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11528 				    (hmeblkp->hblk_shw_bit == 0 ||
11529 				    hmeblkp->hblk_vcnt == 0) &&
11530 				    (hmeblkp->hblk_lckcnt == 0)) {
11531 					/*
11532 					 * there is a high probability that we
11533 					 * will find a free one. search some
11534 					 * buckets for a free hmeblk initially
11535 					 * before unloading a valid hmeblk.
11536 					 */
11537 					if ((hmeblkp->hblk_vcnt == 0 &&
11538 					    hmeblkp->hblk_hmecnt == 0) || (i >=
11539 					    BUCKETS_TO_SEARCH_BEFORE_UNLOAD)) {
11540 						if (sfmmu_steal_this_hblk(hmebp,
11541 						    hmeblkp, hblkpa, pr_hblk)) {
11542 							/*
11543 							 * Hblk is unloaded
11544 							 * successfully
11545 							 */
11546 							break;
11547 						}
11548 					}
11549 				}
11550 				pr_hblk = hmeblkp;
11551 				hblkpa = hmeblkp->hblk_nextpa;
11552 				hmeblkp = hmeblkp->hblk_next;
11553 			}
11554 
11555 			SFMMU_HASH_UNLOCK(hmebp);
11556 			if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
11557 				hmebp = uhme_hash;
11558 		}
11559 		uhmehash_steal_hand = hmebp;
11560 
11561 		if (hmeblkp != NULL)
11562 			break;
11563 
11564 		/*
11565 		 * in the worst case, look for a free one in the kernel
11566 		 * hash table.
11567 		 */
11568 		for (i = 0, hmebp = khme_hash; i <= KHMEHASH_SZ; i++) {
11569 			SFMMU_HASH_LOCK(hmebp);
11570 			hmeblkp = hmebp->hmeblkp;
11571 			hblkpa = hmebp->hmeh_nextpa;
11572 			pr_hblk = NULL;
11573 			while (hmeblkp) {
11574 				/*
11575 				 * check if it is free hmeblk
11576 				 */
11577 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11578 				    (hmeblkp->hblk_lckcnt == 0) &&
11579 				    (hmeblkp->hblk_vcnt == 0) &&
11580 				    (hmeblkp->hblk_hmecnt == 0)) {
11581 					if (sfmmu_steal_this_hblk(hmebp,
11582 					    hmeblkp, hblkpa, pr_hblk)) {
11583 						break;
11584 					} else {
11585 						/*
11586 						 * Cannot fail since we have
11587 						 * hash lock.
11588 						 */
11589 						panic("fail to steal?");
11590 					}
11591 				}
11592 
11593 				pr_hblk = hmeblkp;
11594 				hblkpa = hmeblkp->hblk_nextpa;
11595 				hmeblkp = hmeblkp->hblk_next;
11596 			}
11597 
11598 			SFMMU_HASH_UNLOCK(hmebp);
11599 			if (hmebp++ == &khme_hash[KHMEHASH_SZ])
11600 				hmebp = khme_hash;
11601 		}
11602 
11603 		if (hmeblkp != NULL)
11604 			break;
11605 		sfmmu_hblk_steal_twice++;
11606 	}
11607 	return (hmeblkp);
11608 }
11609 
11610 /*
11611  * This routine does real work to prepare a hblk to be "stolen" by
11612  * unloading the mappings, updating shadow counts ....
11613  * It returns 1 if the block is ready to be reused (stolen), or 0
11614  * means the block cannot be stolen yet- pageunload is still working
11615  * on this hblk.
11616  */
11617 static int
11618 sfmmu_steal_this_hblk(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
11619 	uint64_t hblkpa, struct hme_blk *pr_hblk)
11620 {
11621 	int shw_size, vshift;
11622 	struct hme_blk *shw_hblkp;
11623 	caddr_t vaddr;
11624 	uint_t shw_mask, newshw_mask;
11625 	struct hme_blk *list = NULL;
11626 
11627 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11628 
11629 	/*
11630 	 * check if the hmeblk is free, unload if necessary
11631 	 */
11632 	if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11633 		sfmmu_t *sfmmup;
11634 		demap_range_t dmr;
11635 
11636 		sfmmup = hblktosfmmu(hmeblkp);
11637 		if (hmeblkp->hblk_shared || sfmmup->sfmmu_ismhat) {
11638 			return (0);
11639 		}
11640 		DEMAP_RANGE_INIT(sfmmup, &dmr);
11641 		(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
11642 		    (caddr_t)get_hblk_base(hmeblkp),
11643 		    get_hblk_endaddr(hmeblkp), &dmr, HAT_UNLOAD);
11644 		DEMAP_RANGE_FLUSH(&dmr);
11645 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11646 			/*
11647 			 * Pageunload is working on the same hblk.
11648 			 */
11649 			return (0);
11650 		}
11651 
11652 		sfmmu_hblk_steal_unload_count++;
11653 	}
11654 
11655 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11656 	ASSERT(hmeblkp->hblk_vcnt == 0 && hmeblkp->hblk_hmecnt == 0);
11657 
11658 	sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 1);
11659 	hmeblkp->hblk_nextpa = hblkpa;
11660 
11661 	shw_hblkp = hmeblkp->hblk_shadow;
11662 	if (shw_hblkp) {
11663 		ASSERT(!hmeblkp->hblk_shared);
11664 		shw_size = get_hblk_ttesz(shw_hblkp);
11665 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
11666 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
11667 		ASSERT(vshift < 8);
11668 		/*
11669 		 * Atomically clear shadow mask bit
11670 		 */
11671 		do {
11672 			shw_mask = shw_hblkp->hblk_shw_mask;
11673 			ASSERT(shw_mask & (1 << vshift));
11674 			newshw_mask = shw_mask & ~(1 << vshift);
11675 			newshw_mask = atomic_cas_32(&shw_hblkp->hblk_shw_mask,
11676 			    shw_mask, newshw_mask);
11677 		} while (newshw_mask != shw_mask);
11678 		hmeblkp->hblk_shadow = NULL;
11679 	}
11680 
11681 	/*
11682 	 * remove shadow bit if we are stealing an unused shadow hmeblk.
11683 	 * sfmmu_hblk_alloc needs it that way, will set shadow bit later if
11684 	 * we are indeed allocating a shadow hmeblk.
11685 	 */
11686 	hmeblkp->hblk_shw_bit = 0;
11687 
11688 	if (hmeblkp->hblk_shared) {
11689 		sf_srd_t	*srdp;
11690 		sf_region_t	*rgnp;
11691 		uint_t		rid;
11692 
11693 		srdp = hblktosrd(hmeblkp);
11694 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11695 		rid = hmeblkp->hblk_tag.htag_rid;
11696 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
11697 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
11698 		rgnp = srdp->srd_hmergnp[rid];
11699 		ASSERT(rgnp != NULL);
11700 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
11701 		hmeblkp->hblk_shared = 0;
11702 	}
11703 
11704 	sfmmu_hblk_steal_count++;
11705 	SFMMU_STAT(sf_steal_count);
11706 
11707 	return (1);
11708 }
11709 
11710 struct hme_blk *
11711 sfmmu_hmetohblk(struct sf_hment *sfhme)
11712 {
11713 	struct hme_blk *hmeblkp;
11714 	struct sf_hment *sfhme0;
11715 	struct hme_blk *hblk_dummy = 0;
11716 
11717 	/*
11718 	 * No dummy sf_hments, please.
11719 	 */
11720 	ASSERT(sfhme->hme_tte.ll != 0);
11721 
11722 	sfhme0 = sfhme - sfhme->hme_tte.tte_hmenum;
11723 	hmeblkp = (struct hme_blk *)((uintptr_t)sfhme0 -
11724 	    (uintptr_t)&hblk_dummy->hblk_hme[0]);
11725 
11726 	return (hmeblkp);
11727 }
11728 
11729 /*
11730  * On swapin, get appropriately sized TSB(s) and clear the HAT_SWAPPED flag.
11731  * If we can't get appropriately sized TSB(s), try for 8K TSB(s) using
11732  * KM_SLEEP allocation.
11733  *
11734  * Return 0 on success, -1 otherwise.
11735  */
11736 static void
11737 sfmmu_tsb_swapin(sfmmu_t *sfmmup, hatlock_t *hatlockp)
11738 {
11739 	struct tsb_info *tsbinfop, *next;
11740 	tsb_replace_rc_t rc;
11741 	boolean_t gotfirst = B_FALSE;
11742 
11743 	ASSERT(sfmmup != ksfmmup);
11744 	ASSERT(sfmmu_hat_lock_held(sfmmup));
11745 
11746 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPIN)) {
11747 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
11748 	}
11749 
11750 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
11751 		SFMMU_FLAGS_SET(sfmmup, HAT_SWAPIN);
11752 	} else {
11753 		return;
11754 	}
11755 
11756 	ASSERT(sfmmup->sfmmu_tsb != NULL);
11757 
11758 	/*
11759 	 * Loop over all tsbinfo's replacing them with ones that actually have
11760 	 * a TSB.  If any of the replacements ever fail, bail out of the loop.
11761 	 */
11762 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; tsbinfop = next) {
11763 		ASSERT(tsbinfop->tsb_flags & TSB_SWAPPED);
11764 		next = tsbinfop->tsb_next;
11765 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, tsbinfop->tsb_szc,
11766 		    hatlockp, TSB_SWAPIN);
11767 		if (rc != TSB_SUCCESS) {
11768 			break;
11769 		}
11770 		gotfirst = B_TRUE;
11771 	}
11772 
11773 	switch (rc) {
11774 	case TSB_SUCCESS:
11775 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11776 		cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11777 		return;
11778 	case TSB_LOSTRACE:
11779 		break;
11780 	case TSB_ALLOCFAIL:
11781 		break;
11782 	default:
11783 		panic("sfmmu_replace_tsb returned unrecognized failure code "
11784 		    "%d", rc);
11785 	}
11786 
11787 	/*
11788 	 * In this case, we failed to get one of our TSBs.  If we failed to
11789 	 * get the first TSB, get one of minimum size (8KB).  Walk the list
11790 	 * and throw away the tsbinfos, starting where the allocation failed;
11791 	 * we can get by with just one TSB as long as we don't leave the
11792 	 * SWAPPED tsbinfo structures lying around.
11793 	 */
11794 	tsbinfop = sfmmup->sfmmu_tsb;
11795 	next = tsbinfop->tsb_next;
11796 	tsbinfop->tsb_next = NULL;
11797 
11798 	sfmmu_hat_exit(hatlockp);
11799 	for (tsbinfop = next; tsbinfop != NULL; tsbinfop = next) {
11800 		next = tsbinfop->tsb_next;
11801 		sfmmu_tsbinfo_free(tsbinfop);
11802 	}
11803 	hatlockp = sfmmu_hat_enter(sfmmup);
11804 
11805 	/*
11806 	 * If we don't have any TSBs, get a single 8K TSB for 8K, 64K and 512K
11807 	 * pages.
11808 	 */
11809 	if (!gotfirst) {
11810 		tsbinfop = sfmmup->sfmmu_tsb;
11811 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, TSB_MIN_SZCODE,
11812 		    hatlockp, TSB_SWAPIN | TSB_FORCEALLOC);
11813 		ASSERT(rc == TSB_SUCCESS);
11814 	}
11815 
11816 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11817 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11818 }
11819 
11820 static int
11821 sfmmu_is_rgnva(sf_srd_t *srdp, caddr_t addr, ulong_t w, ulong_t bmw)
11822 {
11823 	ulong_t bix = 0;
11824 	uint_t rid;
11825 	sf_region_t *rgnp;
11826 
11827 	ASSERT(srdp != NULL);
11828 	ASSERT(srdp->srd_refcnt != 0);
11829 
11830 	w <<= BT_ULSHIFT;
11831 	while (bmw) {
11832 		if (!(bmw & 0x1)) {
11833 			bix++;
11834 			bmw >>= 1;
11835 			continue;
11836 		}
11837 		rid = w | bix;
11838 		rgnp = srdp->srd_hmergnp[rid];
11839 		ASSERT(rgnp->rgn_refcnt > 0);
11840 		ASSERT(rgnp->rgn_id == rid);
11841 		if (addr < rgnp->rgn_saddr ||
11842 		    addr >= (rgnp->rgn_saddr + rgnp->rgn_size)) {
11843 			bix++;
11844 			bmw >>= 1;
11845 		} else {
11846 			return (1);
11847 		}
11848 	}
11849 	return (0);
11850 }
11851 
11852 /*
11853  * Handle exceptions for low level tsb_handler.
11854  *
11855  * There are many scenarios that could land us here:
11856  *
11857  * If the context is invalid we land here. The context can be invalid
11858  * for 3 reasons: 1) we couldn't allocate a new context and now need to
11859  * perform a wrap around operation in order to allocate a new context.
11860  * 2) Context was invalidated to change pagesize programming 3) ISMs or
11861  * TSBs configuration is changeing for this process and we are forced into
11862  * here to do a syncronization operation. If the context is valid we can
11863  * be here from window trap hanlder. In this case just call trap to handle
11864  * the fault.
11865  *
11866  * Note that the process will run in INVALID_CONTEXT before
11867  * faulting into here and subsequently loading the MMU registers
11868  * (including the TSB base register) associated with this process.
11869  * For this reason, the trap handlers must all test for
11870  * INVALID_CONTEXT before attempting to access any registers other
11871  * than the context registers.
11872  */
11873 void
11874 sfmmu_tsbmiss_exception(struct regs *rp, uintptr_t tagaccess, uint_t traptype)
11875 {
11876 	sfmmu_t *sfmmup, *shsfmmup;
11877 	uint_t ctxtype;
11878 	klwp_id_t lwp;
11879 	char lwp_save_state;
11880 	hatlock_t *hatlockp, *shatlockp;
11881 	struct tsb_info *tsbinfop;
11882 	struct tsbmiss *tsbmp;
11883 	sf_scd_t *scdp;
11884 
11885 	SFMMU_STAT(sf_tsb_exceptions);
11886 	SFMMU_MMU_STAT(mmu_tsb_exceptions);
11887 	sfmmup = astosfmmu(curthread->t_procp->p_as);
11888 	/*
11889 	 * note that in sun4u, tagacces register contains ctxnum
11890 	 * while sun4v passes ctxtype in the tagaccess register.
11891 	 */
11892 	ctxtype = tagaccess & TAGACC_CTX_MASK;
11893 
11894 	ASSERT(sfmmup != ksfmmup && ctxtype != KCONTEXT);
11895 	ASSERT(sfmmup->sfmmu_ismhat == 0);
11896 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED) ||
11897 	    ctxtype == INVALID_CONTEXT);
11898 
11899 	if (ctxtype != INVALID_CONTEXT && traptype != T_DATA_PROT) {
11900 		/*
11901 		 * We may land here because shme bitmap and pagesize
11902 		 * flags are updated lazily in tsbmiss area on other cpus.
11903 		 * If we detect here that tsbmiss area is out of sync with
11904 		 * sfmmu update it and retry the trapped instruction.
11905 		 * Otherwise call trap().
11906 		 */
11907 		int ret = 0;
11908 		uchar_t tteflag_mask = (1 << TTE64K) | (1 << TTE8K);
11909 		caddr_t addr = (caddr_t)(tagaccess & TAGACC_VADDR_MASK);
11910 
11911 		/*
11912 		 * Must set lwp state to LWP_SYS before
11913 		 * trying to acquire any adaptive lock
11914 		 */
11915 		lwp = ttolwp(curthread);
11916 		ASSERT(lwp);
11917 		lwp_save_state = lwp->lwp_state;
11918 		lwp->lwp_state = LWP_SYS;
11919 
11920 		hatlockp = sfmmu_hat_enter(sfmmup);
11921 		kpreempt_disable();
11922 		tsbmp = &tsbmiss_area[CPU->cpu_id];
11923 		ASSERT(sfmmup == tsbmp->usfmmup);
11924 		if (((tsbmp->uhat_tteflags ^ sfmmup->sfmmu_tteflags) &
11925 		    ~tteflag_mask) ||
11926 		    ((tsbmp->uhat_rtteflags ^  sfmmup->sfmmu_rtteflags) &
11927 		    ~tteflag_mask)) {
11928 			tsbmp->uhat_tteflags = sfmmup->sfmmu_tteflags;
11929 			tsbmp->uhat_rtteflags = sfmmup->sfmmu_rtteflags;
11930 			ret = 1;
11931 		}
11932 		if (sfmmup->sfmmu_srdp != NULL) {
11933 			ulong_t *sm = sfmmup->sfmmu_hmeregion_map.bitmap;
11934 			ulong_t *tm = tsbmp->shmermap;
11935 			ulong_t i;
11936 			for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
11937 				ulong_t d = tm[i] ^ sm[i];
11938 				if (d) {
11939 					if (d & sm[i]) {
11940 						if (!ret && sfmmu_is_rgnva(
11941 						    sfmmup->sfmmu_srdp,
11942 						    addr, i, d & sm[i])) {
11943 							ret = 1;
11944 						}
11945 					}
11946 					tm[i] = sm[i];
11947 				}
11948 			}
11949 		}
11950 		kpreempt_enable();
11951 		sfmmu_hat_exit(hatlockp);
11952 		lwp->lwp_state = lwp_save_state;
11953 		if (ret) {
11954 			return;
11955 		}
11956 	} else if (ctxtype == INVALID_CONTEXT) {
11957 		/*
11958 		 * First, make sure we come out of here with a valid ctx,
11959 		 * since if we don't get one we'll simply loop on the
11960 		 * faulting instruction.
11961 		 *
11962 		 * If the ISM mappings are changing, the TSB is relocated,
11963 		 * the process is swapped, the process is joining SCD or
11964 		 * leaving SCD or shared regions we serialize behind the
11965 		 * controlling thread with hat lock, sfmmu_flags and
11966 		 * sfmmu_tsb_cv condition variable.
11967 		 */
11968 
11969 		/*
11970 		 * Must set lwp state to LWP_SYS before
11971 		 * trying to acquire any adaptive lock
11972 		 */
11973 		lwp = ttolwp(curthread);
11974 		ASSERT(lwp);
11975 		lwp_save_state = lwp->lwp_state;
11976 		lwp->lwp_state = LWP_SYS;
11977 
11978 		hatlockp = sfmmu_hat_enter(sfmmup);
11979 retry:
11980 		if ((scdp = sfmmup->sfmmu_scdp) != NULL) {
11981 			shsfmmup = scdp->scd_sfmmup;
11982 			ASSERT(shsfmmup != NULL);
11983 
11984 			for (tsbinfop = shsfmmup->sfmmu_tsb; tsbinfop != NULL;
11985 			    tsbinfop = tsbinfop->tsb_next) {
11986 				if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
11987 					/* drop the private hat lock */
11988 					sfmmu_hat_exit(hatlockp);
11989 					/* acquire the shared hat lock */
11990 					shatlockp = sfmmu_hat_enter(shsfmmup);
11991 					/*
11992 					 * recheck to see if anything changed
11993 					 * after we drop the private hat lock.
11994 					 */
11995 					if (sfmmup->sfmmu_scdp == scdp &&
11996 					    shsfmmup == scdp->scd_sfmmup) {
11997 						sfmmu_tsb_chk_reloc(shsfmmup,
11998 						    shatlockp);
11999 					}
12000 					sfmmu_hat_exit(shatlockp);
12001 					hatlockp = sfmmu_hat_enter(sfmmup);
12002 					goto retry;
12003 				}
12004 			}
12005 		}
12006 
12007 		for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
12008 		    tsbinfop = tsbinfop->tsb_next) {
12009 			if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
12010 				cv_wait(&sfmmup->sfmmu_tsb_cv,
12011 				    HATLOCK_MUTEXP(hatlockp));
12012 				goto retry;
12013 			}
12014 		}
12015 
12016 		/*
12017 		 * Wait for ISM maps to be updated.
12018 		 */
12019 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
12020 			cv_wait(&sfmmup->sfmmu_tsb_cv,
12021 			    HATLOCK_MUTEXP(hatlockp));
12022 			goto retry;
12023 		}
12024 
12025 		/* Is this process joining an SCD? */
12026 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
12027 			/*
12028 			 * Flush private TSB and setup shared TSB.
12029 			 * sfmmu_finish_join_scd() does not drop the
12030 			 * hat lock.
12031 			 */
12032 			sfmmu_finish_join_scd(sfmmup);
12033 			SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
12034 		}
12035 
12036 		/*
12037 		 * If we're swapping in, get TSB(s).  Note that we must do
12038 		 * this before we get a ctx or load the MMU state.  Once
12039 		 * we swap in we have to recheck to make sure the TSB(s) and
12040 		 * ISM mappings didn't change while we slept.
12041 		 */
12042 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
12043 			sfmmu_tsb_swapin(sfmmup, hatlockp);
12044 			goto retry;
12045 		}
12046 
12047 		sfmmu_get_ctx(sfmmup);
12048 
12049 		sfmmu_hat_exit(hatlockp);
12050 		/*
12051 		 * Must restore lwp_state if not calling
12052 		 * trap() for further processing. Restore
12053 		 * it anyway.
12054 		 */
12055 		lwp->lwp_state = lwp_save_state;
12056 		return;
12057 	}
12058 	trap(rp, (caddr_t)tagaccess, traptype, 0);
12059 }
12060 
12061 static void
12062 sfmmu_tsb_chk_reloc(sfmmu_t *sfmmup, hatlock_t *hatlockp)
12063 {
12064 	struct tsb_info *tp;
12065 
12066 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12067 
12068 	for (tp = sfmmup->sfmmu_tsb; tp != NULL; tp = tp->tsb_next) {
12069 		if (tp->tsb_flags & TSB_RELOC_FLAG) {
12070 			cv_wait(&sfmmup->sfmmu_tsb_cv,
12071 			    HATLOCK_MUTEXP(hatlockp));
12072 			break;
12073 		}
12074 	}
12075 }
12076 
12077 /*
12078  * sfmmu_vatopfn_suspended is called from GET_TTE when TL=0 and
12079  * TTE_SUSPENDED bit set in tte we block on aquiring a page lock
12080  * rather than spinning to avoid send mondo timeouts with
12081  * interrupts enabled. When the lock is acquired it is immediately
12082  * released and we return back to sfmmu_vatopfn just after
12083  * the GET_TTE call.
12084  */
12085 void
12086 sfmmu_vatopfn_suspended(caddr_t vaddr, sfmmu_t *sfmmu, tte_t *ttep)
12087 {
12088 	struct page	**pp;
12089 
12090 	(void) as_pagelock(sfmmu->sfmmu_as, &pp, vaddr, TTE_CSZ(ttep), S_WRITE);
12091 	as_pageunlock(sfmmu->sfmmu_as, pp, vaddr, TTE_CSZ(ttep), S_WRITE);
12092 }
12093 
12094 /*
12095  * sfmmu_tsbmiss_suspended is called from GET_TTE when TL>0 and
12096  * TTE_SUSPENDED bit set in tte. We do this so that we can handle
12097  * cross traps which cannot be handled while spinning in the
12098  * trap handlers. Simply enter and exit the kpr_suspendlock spin
12099  * mutex, which is held by the holder of the suspend bit, and then
12100  * retry the trapped instruction after unwinding.
12101  */
12102 /*ARGSUSED*/
12103 void
12104 sfmmu_tsbmiss_suspended(struct regs *rp, uintptr_t tagacc, uint_t traptype)
12105 {
12106 	ASSERT(curthread != kreloc_thread);
12107 	mutex_enter(&kpr_suspendlock);
12108 	mutex_exit(&kpr_suspendlock);
12109 }
12110 
12111 /*
12112  * This routine could be optimized to reduce the number of xcalls by flushing
12113  * the entire TLBs if region reference count is above some threshold but the
12114  * tradeoff will depend on the size of the TLB. So for now flush the specific
12115  * page a context at a time.
12116  *
12117  * If uselocks is 0 then it's called after all cpus were captured and all the
12118  * hat locks were taken. In this case don't take the region lock by relying on
12119  * the order of list region update operations in hat_join_region(),
12120  * hat_leave_region() and hat_dup_region(). The ordering in those routines
12121  * guarantees that list is always forward walkable and reaches active sfmmus
12122  * regardless of where xc_attention() captures a cpu.
12123  */
12124 cpuset_t
12125 sfmmu_rgntlb_demap(caddr_t addr, sf_region_t *rgnp,
12126     struct hme_blk *hmeblkp, int uselocks)
12127 {
12128 	sfmmu_t	*sfmmup;
12129 	cpuset_t cpuset;
12130 	cpuset_t rcpuset;
12131 	hatlock_t *hatlockp;
12132 	uint_t rid = rgnp->rgn_id;
12133 	sf_rgn_link_t *rlink;
12134 	sf_scd_t *scdp;
12135 
12136 	ASSERT(hmeblkp->hblk_shared);
12137 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
12138 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
12139 
12140 	CPUSET_ZERO(rcpuset);
12141 	if (uselocks) {
12142 		mutex_enter(&rgnp->rgn_mutex);
12143 	}
12144 	sfmmup = rgnp->rgn_sfmmu_head;
12145 	while (sfmmup != NULL) {
12146 		if (uselocks) {
12147 			hatlockp = sfmmu_hat_enter(sfmmup);
12148 		}
12149 
12150 		/*
12151 		 * When an SCD is created the SCD hat is linked on the sfmmu
12152 		 * region lists for each hme region which is part of the
12153 		 * SCD. If we find an SCD hat, when walking these lists,
12154 		 * then we flush the shared TSBs, if we find a private hat,
12155 		 * which is part of an SCD, but where the region
12156 		 * is not part of the SCD then we flush the private TSBs.
12157 		 */
12158 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12159 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
12160 			scdp = sfmmup->sfmmu_scdp;
12161 			if (SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
12162 				if (uselocks) {
12163 					sfmmu_hat_exit(hatlockp);
12164 				}
12165 				goto next;
12166 			}
12167 		}
12168 
12169 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12170 
12171 		kpreempt_disable();
12172 		cpuset = sfmmup->sfmmu_cpusran;
12173 		CPUSET_AND(cpuset, cpu_ready_set);
12174 		CPUSET_DEL(cpuset, CPU->cpu_id);
12175 		SFMMU_XCALL_STATS(sfmmup);
12176 		xt_some(cpuset, vtag_flushpage_tl1,
12177 		    (uint64_t)addr, (uint64_t)sfmmup);
12178 		vtag_flushpage(addr, (uint64_t)sfmmup);
12179 		if (uselocks) {
12180 			sfmmu_hat_exit(hatlockp);
12181 		}
12182 		kpreempt_enable();
12183 		CPUSET_OR(rcpuset, cpuset);
12184 
12185 next:
12186 		/* LINTED: constant in conditional context */
12187 		SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
12188 		ASSERT(rlink != NULL);
12189 		sfmmup = rlink->next;
12190 	}
12191 	if (uselocks) {
12192 		mutex_exit(&rgnp->rgn_mutex);
12193 	}
12194 	return (rcpuset);
12195 }
12196 
12197 /*
12198  * This routine takes an sfmmu pointer and the va for an adddress in an
12199  * ISM region as input and returns the corresponding region id in ism_rid.
12200  * The return value of 1 indicates that a region has been found and ism_rid
12201  * is valid, otherwise 0 is returned.
12202  */
12203 static int
12204 find_ism_rid(sfmmu_t *sfmmup, sfmmu_t *ism_sfmmup, caddr_t va, uint_t *ism_rid)
12205 {
12206 	ism_blk_t	*ism_blkp;
12207 	int		i;
12208 	ism_map_t	*ism_map;
12209 #ifdef DEBUG
12210 	struct hat	*ism_hatid;
12211 #endif
12212 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12213 
12214 	ism_blkp = sfmmup->sfmmu_iblk;
12215 	while (ism_blkp != NULL) {
12216 		ism_map = ism_blkp->iblk_maps;
12217 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
12218 			if ((va >= ism_start(ism_map[i])) &&
12219 			    (va < ism_end(ism_map[i]))) {
12220 
12221 				*ism_rid = ism_map[i].imap_rid;
12222 #ifdef DEBUG
12223 				ism_hatid = ism_map[i].imap_ismhat;
12224 				ASSERT(ism_hatid == ism_sfmmup);
12225 				ASSERT(ism_hatid->sfmmu_ismhat);
12226 #endif
12227 				return (1);
12228 			}
12229 		}
12230 		ism_blkp = ism_blkp->iblk_next;
12231 	}
12232 	return (0);
12233 }
12234 
12235 /*
12236  * Special routine to flush out ism mappings- TSBs, TLBs and D-caches.
12237  * This routine may be called with all cpu's captured. Therefore, the
12238  * caller is responsible for holding all locks and disabling kernel
12239  * preemption.
12240  */
12241 /* ARGSUSED */
12242 static void
12243 sfmmu_ismtlbcache_demap(caddr_t addr, sfmmu_t *ism_sfmmup,
12244 	struct hme_blk *hmeblkp, pfn_t pfnum, int cache_flush_flag)
12245 {
12246 	cpuset_t 	cpuset;
12247 	caddr_t 	va;
12248 	ism_ment_t	*ment;
12249 	sfmmu_t		*sfmmup;
12250 #ifdef VAC
12251 	int 		vcolor;
12252 #endif
12253 
12254 	sf_scd_t	*scdp;
12255 	uint_t		ism_rid;
12256 
12257 	ASSERT(!hmeblkp->hblk_shared);
12258 	/*
12259 	 * Walk the ism_hat's mapping list and flush the page
12260 	 * from every hat sharing this ism_hat. This routine
12261 	 * may be called while all cpu's have been captured.
12262 	 * Therefore we can't attempt to grab any locks. For now
12263 	 * this means we will protect the ism mapping list under
12264 	 * a single lock which will be grabbed by the caller.
12265 	 * If hat_share/unshare scalibility becomes a performance
12266 	 * problem then we may need to re-think ism mapping list locking.
12267 	 */
12268 	ASSERT(ism_sfmmup->sfmmu_ismhat);
12269 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
12270 	addr = addr - ISMID_STARTADDR;
12271 
12272 	for (ment = ism_sfmmup->sfmmu_iment; ment; ment = ment->iment_next) {
12273 
12274 		sfmmup = ment->iment_hat;
12275 
12276 		va = ment->iment_base_va;
12277 		va = (caddr_t)((uintptr_t)va  + (uintptr_t)addr);
12278 
12279 		/*
12280 		 * When an SCD is created the SCD hat is linked on the ism
12281 		 * mapping lists for each ISM segment which is part of the
12282 		 * SCD. If we find an SCD hat, when walking these lists,
12283 		 * then we flush the shared TSBs, if we find a private hat,
12284 		 * which is part of an SCD, but where the region
12285 		 * corresponding to this va is not part of the SCD then we
12286 		 * flush the private TSBs.
12287 		 */
12288 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12289 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD) &&
12290 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
12291 			if (!find_ism_rid(sfmmup, ism_sfmmup, va,
12292 			    &ism_rid)) {
12293 				cmn_err(CE_PANIC,
12294 				    "can't find matching ISM rid!");
12295 			}
12296 
12297 			scdp = sfmmup->sfmmu_scdp;
12298 			if (SFMMU_IS_ISMRID_VALID(ism_rid) &&
12299 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map,
12300 			    ism_rid)) {
12301 				continue;
12302 			}
12303 		}
12304 		SFMMU_UNLOAD_TSB(va, sfmmup, hmeblkp, 1);
12305 
12306 		cpuset = sfmmup->sfmmu_cpusran;
12307 		CPUSET_AND(cpuset, cpu_ready_set);
12308 		CPUSET_DEL(cpuset, CPU->cpu_id);
12309 		SFMMU_XCALL_STATS(sfmmup);
12310 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)va,
12311 		    (uint64_t)sfmmup);
12312 		vtag_flushpage(va, (uint64_t)sfmmup);
12313 
12314 #ifdef VAC
12315 		/*
12316 		 * Flush D$
12317 		 * When flushing D$ we must flush all
12318 		 * cpu's. See sfmmu_cache_flush().
12319 		 */
12320 		if (cache_flush_flag == CACHE_FLUSH) {
12321 			cpuset = cpu_ready_set;
12322 			CPUSET_DEL(cpuset, CPU->cpu_id);
12323 
12324 			SFMMU_XCALL_STATS(sfmmup);
12325 			vcolor = addr_to_vcolor(va);
12326 			xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12327 			vac_flushpage(pfnum, vcolor);
12328 		}
12329 #endif	/* VAC */
12330 	}
12331 }
12332 
12333 /*
12334  * Demaps the TSB, CPU caches, and flushes all TLBs on all CPUs of
12335  * a particular virtual address and ctx.  If noflush is set we do not
12336  * flush the TLB/TSB.  This function may or may not be called with the
12337  * HAT lock held.
12338  */
12339 static void
12340 sfmmu_tlbcache_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12341 	pfn_t pfnum, int tlb_noflush, int cpu_flag, int cache_flush_flag,
12342 	int hat_lock_held)
12343 {
12344 #ifdef VAC
12345 	int vcolor;
12346 #endif
12347 	cpuset_t cpuset;
12348 	hatlock_t *hatlockp;
12349 
12350 	ASSERT(!hmeblkp->hblk_shared);
12351 
12352 #if defined(lint) && !defined(VAC)
12353 	pfnum = pfnum;
12354 	cpu_flag = cpu_flag;
12355 	cache_flush_flag = cache_flush_flag;
12356 #endif
12357 
12358 	/*
12359 	 * There is no longer a need to protect against ctx being
12360 	 * stolen here since we don't store the ctx in the TSB anymore.
12361 	 */
12362 #ifdef VAC
12363 	vcolor = addr_to_vcolor(addr);
12364 #endif
12365 
12366 	/*
12367 	 * We must hold the hat lock during the flush of TLB,
12368 	 * to avoid a race with sfmmu_invalidate_ctx(), where
12369 	 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12370 	 * causing TLB demap routine to skip flush on that MMU.
12371 	 * If the context on a MMU has already been set to
12372 	 * INVALID_CONTEXT, we just get an extra flush on
12373 	 * that MMU.
12374 	 */
12375 	if (!hat_lock_held && !tlb_noflush)
12376 		hatlockp = sfmmu_hat_enter(sfmmup);
12377 
12378 	kpreempt_disable();
12379 	if (!tlb_noflush) {
12380 		/*
12381 		 * Flush the TSB and TLB.
12382 		 */
12383 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12384 
12385 		cpuset = sfmmup->sfmmu_cpusran;
12386 		CPUSET_AND(cpuset, cpu_ready_set);
12387 		CPUSET_DEL(cpuset, CPU->cpu_id);
12388 
12389 		SFMMU_XCALL_STATS(sfmmup);
12390 
12391 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
12392 		    (uint64_t)sfmmup);
12393 
12394 		vtag_flushpage(addr, (uint64_t)sfmmup);
12395 	}
12396 
12397 	if (!hat_lock_held && !tlb_noflush)
12398 		sfmmu_hat_exit(hatlockp);
12399 
12400 #ifdef VAC
12401 	/*
12402 	 * Flush the D$
12403 	 *
12404 	 * Even if the ctx is stolen, we need to flush the
12405 	 * cache. Our ctx stealer only flushes the TLBs.
12406 	 */
12407 	if (cache_flush_flag == CACHE_FLUSH) {
12408 		if (cpu_flag & FLUSH_ALL_CPUS) {
12409 			cpuset = cpu_ready_set;
12410 		} else {
12411 			cpuset = sfmmup->sfmmu_cpusran;
12412 			CPUSET_AND(cpuset, cpu_ready_set);
12413 		}
12414 		CPUSET_DEL(cpuset, CPU->cpu_id);
12415 		SFMMU_XCALL_STATS(sfmmup);
12416 		xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12417 		vac_flushpage(pfnum, vcolor);
12418 	}
12419 #endif	/* VAC */
12420 	kpreempt_enable();
12421 }
12422 
12423 /*
12424  * Demaps the TSB and flushes all TLBs on all cpus for a particular virtual
12425  * address and ctx.  If noflush is set we do not currently do anything.
12426  * This function may or may not be called with the HAT lock held.
12427  */
12428 static void
12429 sfmmu_tlb_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12430 	int tlb_noflush, int hat_lock_held)
12431 {
12432 	cpuset_t cpuset;
12433 	hatlock_t *hatlockp;
12434 
12435 	ASSERT(!hmeblkp->hblk_shared);
12436 
12437 	/*
12438 	 * If the process is exiting we have nothing to do.
12439 	 */
12440 	if (tlb_noflush)
12441 		return;
12442 
12443 	/*
12444 	 * Flush TSB.
12445 	 */
12446 	if (!hat_lock_held)
12447 		hatlockp = sfmmu_hat_enter(sfmmup);
12448 	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12449 
12450 	kpreempt_disable();
12451 
12452 	cpuset = sfmmup->sfmmu_cpusran;
12453 	CPUSET_AND(cpuset, cpu_ready_set);
12454 	CPUSET_DEL(cpuset, CPU->cpu_id);
12455 
12456 	SFMMU_XCALL_STATS(sfmmup);
12457 	xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, (uint64_t)sfmmup);
12458 
12459 	vtag_flushpage(addr, (uint64_t)sfmmup);
12460 
12461 	if (!hat_lock_held)
12462 		sfmmu_hat_exit(hatlockp);
12463 
12464 	kpreempt_enable();
12465 
12466 }
12467 
12468 /*
12469  * Special case of sfmmu_tlb_demap for MMU_PAGESIZE hblks. Use the xcall
12470  * call handler that can flush a range of pages to save on xcalls.
12471  */
12472 static int sfmmu_xcall_save;
12473 
12474 /*
12475  * this routine is never used for demaping addresses backed by SRD hmeblks.
12476  */
12477 static void
12478 sfmmu_tlb_range_demap(demap_range_t *dmrp)
12479 {
12480 	sfmmu_t *sfmmup = dmrp->dmr_sfmmup;
12481 	hatlock_t *hatlockp;
12482 	cpuset_t cpuset;
12483 	uint64_t sfmmu_pgcnt;
12484 	pgcnt_t pgcnt = 0;
12485 	int pgunload = 0;
12486 	int dirtypg = 0;
12487 	caddr_t addr = dmrp->dmr_addr;
12488 	caddr_t eaddr;
12489 	uint64_t bitvec = dmrp->dmr_bitvec;
12490 
12491 	ASSERT(bitvec & 1);
12492 
12493 	/*
12494 	 * Flush TSB and calculate number of pages to flush.
12495 	 */
12496 	while (bitvec != 0) {
12497 		dirtypg = 0;
12498 		/*
12499 		 * Find the first page to flush and then count how many
12500 		 * pages there are after it that also need to be flushed.
12501 		 * This way the number of TSB flushes is minimized.
12502 		 */
12503 		while ((bitvec & 1) == 0) {
12504 			pgcnt++;
12505 			addr += MMU_PAGESIZE;
12506 			bitvec >>= 1;
12507 		}
12508 		while (bitvec & 1) {
12509 			dirtypg++;
12510 			bitvec >>= 1;
12511 		}
12512 		eaddr = addr + ptob(dirtypg);
12513 		hatlockp = sfmmu_hat_enter(sfmmup);
12514 		sfmmu_unload_tsb_range(sfmmup, addr, eaddr, TTE8K);
12515 		sfmmu_hat_exit(hatlockp);
12516 		pgunload += dirtypg;
12517 		addr = eaddr;
12518 		pgcnt += dirtypg;
12519 	}
12520 
12521 	ASSERT((pgcnt<<MMU_PAGESHIFT) <= dmrp->dmr_endaddr - dmrp->dmr_addr);
12522 	if (sfmmup->sfmmu_free == 0) {
12523 		addr = dmrp->dmr_addr;
12524 		bitvec = dmrp->dmr_bitvec;
12525 
12526 		/*
12527 		 * make sure it has SFMMU_PGCNT_SHIFT bits only,
12528 		 * as it will be used to pack argument for xt_some
12529 		 */
12530 		ASSERT((pgcnt > 0) &&
12531 		    (pgcnt <= (1 << SFMMU_PGCNT_SHIFT)));
12532 
12533 		/*
12534 		 * Encode pgcnt as (pgcnt -1 ), and pass (pgcnt - 1) in
12535 		 * the low 6 bits of sfmmup. This is doable since pgcnt
12536 		 * always >= 1.
12537 		 */
12538 		ASSERT(!((uint64_t)sfmmup & SFMMU_PGCNT_MASK));
12539 		sfmmu_pgcnt = (uint64_t)sfmmup |
12540 		    ((pgcnt - 1) & SFMMU_PGCNT_MASK);
12541 
12542 		/*
12543 		 * We must hold the hat lock during the flush of TLB,
12544 		 * to avoid a race with sfmmu_invalidate_ctx(), where
12545 		 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12546 		 * causing TLB demap routine to skip flush on that MMU.
12547 		 * If the context on a MMU has already been set to
12548 		 * INVALID_CONTEXT, we just get an extra flush on
12549 		 * that MMU.
12550 		 */
12551 		hatlockp = sfmmu_hat_enter(sfmmup);
12552 		kpreempt_disable();
12553 
12554 		cpuset = sfmmup->sfmmu_cpusran;
12555 		CPUSET_AND(cpuset, cpu_ready_set);
12556 		CPUSET_DEL(cpuset, CPU->cpu_id);
12557 
12558 		SFMMU_XCALL_STATS(sfmmup);
12559 		xt_some(cpuset, vtag_flush_pgcnt_tl1, (uint64_t)addr,
12560 		    sfmmu_pgcnt);
12561 
12562 		for (; bitvec != 0; bitvec >>= 1) {
12563 			if (bitvec & 1)
12564 				vtag_flushpage(addr, (uint64_t)sfmmup);
12565 			addr += MMU_PAGESIZE;
12566 		}
12567 		kpreempt_enable();
12568 		sfmmu_hat_exit(hatlockp);
12569 
12570 		sfmmu_xcall_save += (pgunload-1);
12571 	}
12572 	dmrp->dmr_bitvec = 0;
12573 }
12574 
12575 /*
12576  * In cases where we need to synchronize with TLB/TSB miss trap
12577  * handlers, _and_ need to flush the TLB, it's a lot easier to
12578  * throw away the context from the process than to do a
12579  * special song and dance to keep things consistent for the
12580  * handlers.
12581  *
12582  * Since the process suddenly ends up without a context and our caller
12583  * holds the hat lock, threads that fault after this function is called
12584  * will pile up on the lock.  We can then do whatever we need to
12585  * atomically from the context of the caller.  The first blocked thread
12586  * to resume executing will get the process a new context, and the
12587  * process will resume executing.
12588  *
12589  * One added advantage of this approach is that on MMUs that
12590  * support a "flush all" operation, we will delay the flush until
12591  * cnum wrap-around, and then flush the TLB one time.  This
12592  * is rather rare, so it's a lot less expensive than making 8000
12593  * x-calls to flush the TLB 8000 times.
12594  *
12595  * A per-process (PP) lock is used to synchronize ctx allocations in
12596  * resume() and ctx invalidations here.
12597  */
12598 static void
12599 sfmmu_invalidate_ctx(sfmmu_t *sfmmup)
12600 {
12601 	cpuset_t cpuset;
12602 	int cnum, currcnum;
12603 	mmu_ctx_t *mmu_ctxp;
12604 	int i;
12605 	uint_t pstate_save;
12606 
12607 	SFMMU_STAT(sf_ctx_inv);
12608 
12609 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12610 	ASSERT(sfmmup != ksfmmup);
12611 
12612 	kpreempt_disable();
12613 
12614 	mmu_ctxp = CPU_MMU_CTXP(CPU);
12615 	ASSERT(mmu_ctxp);
12616 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
12617 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
12618 
12619 	currcnum = sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum;
12620 
12621 	pstate_save = sfmmu_disable_intrs();
12622 
12623 	lock_set(&sfmmup->sfmmu_ctx_lock);	/* acquire PP lock */
12624 	/* set HAT cnum invalid across all context domains. */
12625 	for (i = 0; i < max_mmu_ctxdoms; i++) {
12626 
12627 		cnum = 	sfmmup->sfmmu_ctxs[i].cnum;
12628 		if (cnum == INVALID_CONTEXT) {
12629 			continue;
12630 		}
12631 
12632 		sfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
12633 	}
12634 	membar_enter();	/* make sure globally visible to all CPUs */
12635 	lock_clear(&sfmmup->sfmmu_ctx_lock);	/* release PP lock */
12636 
12637 	sfmmu_enable_intrs(pstate_save);
12638 
12639 	cpuset = sfmmup->sfmmu_cpusran;
12640 	CPUSET_DEL(cpuset, CPU->cpu_id);
12641 	CPUSET_AND(cpuset, cpu_ready_set);
12642 	if (!CPUSET_ISNULL(cpuset)) {
12643 		SFMMU_XCALL_STATS(sfmmup);
12644 		xt_some(cpuset, sfmmu_raise_tsb_exception,
12645 		    (uint64_t)sfmmup, INVALID_CONTEXT);
12646 		xt_sync(cpuset);
12647 		SFMMU_STAT(sf_tsb_raise_exception);
12648 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
12649 	}
12650 
12651 	/*
12652 	 * If the hat to-be-invalidated is the same as the current
12653 	 * process on local CPU we need to invalidate
12654 	 * this CPU context as well.
12655 	 */
12656 	if ((sfmmu_getctx_sec() == currcnum) &&
12657 	    (currcnum != INVALID_CONTEXT)) {
12658 		/* sets shared context to INVALID too */
12659 		sfmmu_setctx_sec(INVALID_CONTEXT);
12660 		sfmmu_clear_utsbinfo();
12661 	}
12662 
12663 	SFMMU_FLAGS_SET(sfmmup, HAT_ALLCTX_INVALID);
12664 
12665 	kpreempt_enable();
12666 
12667 	/*
12668 	 * we hold the hat lock, so nobody should allocate a context
12669 	 * for us yet
12670 	 */
12671 	ASSERT(sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum == INVALID_CONTEXT);
12672 }
12673 
12674 #ifdef VAC
12675 /*
12676  * We need to flush the cache in all cpus.  It is possible that
12677  * a process referenced a page as cacheable but has sinced exited
12678  * and cleared the mapping list.  We still to flush it but have no
12679  * state so all cpus is the only alternative.
12680  */
12681 void
12682 sfmmu_cache_flush(pfn_t pfnum, int vcolor)
12683 {
12684 	cpuset_t cpuset;
12685 
12686 	kpreempt_disable();
12687 	cpuset = cpu_ready_set;
12688 	CPUSET_DEL(cpuset, CPU->cpu_id);
12689 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12690 	xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12691 	xt_sync(cpuset);
12692 	vac_flushpage(pfnum, vcolor);
12693 	kpreempt_enable();
12694 }
12695 
12696 void
12697 sfmmu_cache_flushcolor(int vcolor, pfn_t pfnum)
12698 {
12699 	cpuset_t cpuset;
12700 
12701 	ASSERT(vcolor >= 0);
12702 
12703 	kpreempt_disable();
12704 	cpuset = cpu_ready_set;
12705 	CPUSET_DEL(cpuset, CPU->cpu_id);
12706 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12707 	xt_some(cpuset, vac_flushcolor_tl1, vcolor, pfnum);
12708 	xt_sync(cpuset);
12709 	vac_flushcolor(vcolor, pfnum);
12710 	kpreempt_enable();
12711 }
12712 #endif	/* VAC */
12713 
12714 /*
12715  * We need to prevent processes from accessing the TSB using a cached physical
12716  * address.  It's alright if they try to access the TSB via virtual address
12717  * since they will just fault on that virtual address once the mapping has
12718  * been suspended.
12719  */
12720 #pragma weak sendmondo_in_recover
12721 
12722 /* ARGSUSED */
12723 static int
12724 sfmmu_tsb_pre_relocator(caddr_t va, uint_t tsbsz, uint_t flags, void *tsbinfo)
12725 {
12726 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12727 	sfmmu_t *sfmmup = tsbinfop->tsb_sfmmu;
12728 	hatlock_t *hatlockp;
12729 	sf_scd_t *scdp;
12730 
12731 	if (flags != HAT_PRESUSPEND)
12732 		return (0);
12733 
12734 	/*
12735 	 * If tsb is a shared TSB with TSB_SHAREDCTX set, sfmmup must
12736 	 * be a shared hat, then set SCD's tsbinfo's flag.
12737 	 * If tsb is not shared, sfmmup is a private hat, then set
12738 	 * its private tsbinfo's flag.
12739 	 */
12740 	hatlockp = sfmmu_hat_enter(sfmmup);
12741 	tsbinfop->tsb_flags |= TSB_RELOC_FLAG;
12742 
12743 	if (!(tsbinfop->tsb_flags & TSB_SHAREDCTX)) {
12744 		sfmmu_tsb_inv_ctx(sfmmup);
12745 		sfmmu_hat_exit(hatlockp);
12746 	} else {
12747 		/* release lock on the shared hat */
12748 		sfmmu_hat_exit(hatlockp);
12749 		/* sfmmup is a shared hat */
12750 		ASSERT(sfmmup->sfmmu_scdhat);
12751 		scdp = sfmmup->sfmmu_scdp;
12752 		ASSERT(scdp != NULL);
12753 		/* get private hat from the scd list */
12754 		mutex_enter(&scdp->scd_mutex);
12755 		sfmmup = scdp->scd_sf_list;
12756 		while (sfmmup != NULL) {
12757 			hatlockp = sfmmu_hat_enter(sfmmup);
12758 			/*
12759 			 * We do not call sfmmu_tsb_inv_ctx here because
12760 			 * sendmondo_in_recover check is only needed for
12761 			 * sun4u.
12762 			 */
12763 			sfmmu_invalidate_ctx(sfmmup);
12764 			sfmmu_hat_exit(hatlockp);
12765 			sfmmup = sfmmup->sfmmu_scd_link.next;
12766 
12767 		}
12768 		mutex_exit(&scdp->scd_mutex);
12769 	}
12770 	return (0);
12771 }
12772 
12773 static void
12774 sfmmu_tsb_inv_ctx(sfmmu_t *sfmmup)
12775 {
12776 	extern uint32_t sendmondo_in_recover;
12777 
12778 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12779 
12780 	/*
12781 	 * For Cheetah+ Erratum 25:
12782 	 * Wait for any active recovery to finish.  We can't risk
12783 	 * relocating the TSB of the thread running mondo_recover_proc()
12784 	 * since, if we did that, we would deadlock.  The scenario we are
12785 	 * trying to avoid is as follows:
12786 	 *
12787 	 * THIS CPU			RECOVER CPU
12788 	 * --------			-----------
12789 	 *				Begins recovery, walking through TSB
12790 	 * hat_pagesuspend() TSB TTE
12791 	 *				TLB miss on TSB TTE, spins at TL1
12792 	 * xt_sync()
12793 	 *	send_mondo_timeout()
12794 	 *	mondo_recover_proc()
12795 	 *	((deadlocked))
12796 	 *
12797 	 * The second half of the workaround is that mondo_recover_proc()
12798 	 * checks to see if the tsb_info has the RELOC flag set, and if it
12799 	 * does, it skips over that TSB without ever touching tsbinfop->tsb_va
12800 	 * and hence avoiding the TLB miss that could result in a deadlock.
12801 	 */
12802 	if (&sendmondo_in_recover) {
12803 		membar_enter();	/* make sure RELOC flag visible */
12804 		while (sendmondo_in_recover) {
12805 			drv_usecwait(1);
12806 			membar_consumer();
12807 		}
12808 	}
12809 
12810 	sfmmu_invalidate_ctx(sfmmup);
12811 }
12812 
12813 /* ARGSUSED */
12814 static int
12815 sfmmu_tsb_post_relocator(caddr_t va, uint_t tsbsz, uint_t flags,
12816 	void *tsbinfo, pfn_t newpfn)
12817 {
12818 	hatlock_t *hatlockp;
12819 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12820 	sfmmu_t	*sfmmup = tsbinfop->tsb_sfmmu;
12821 
12822 	if (flags != HAT_POSTUNSUSPEND)
12823 		return (0);
12824 
12825 	hatlockp = sfmmu_hat_enter(sfmmup);
12826 
12827 	SFMMU_STAT(sf_tsb_reloc);
12828 
12829 	/*
12830 	 * The process may have swapped out while we were relocating one
12831 	 * of its TSBs.  If so, don't bother doing the setup since the
12832 	 * process can't be using the memory anymore.
12833 	 */
12834 	if ((tsbinfop->tsb_flags & TSB_SWAPPED) == 0) {
12835 		ASSERT(va == tsbinfop->tsb_va);
12836 		sfmmu_tsbinfo_setup_phys(tsbinfop, newpfn);
12837 
12838 		if (tsbinfop->tsb_flags & TSB_FLUSH_NEEDED) {
12839 			sfmmu_inv_tsb(tsbinfop->tsb_va,
12840 			    TSB_BYTES(tsbinfop->tsb_szc));
12841 			tsbinfop->tsb_flags &= ~TSB_FLUSH_NEEDED;
12842 		}
12843 	}
12844 
12845 	membar_exit();
12846 	tsbinfop->tsb_flags &= ~TSB_RELOC_FLAG;
12847 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
12848 
12849 	sfmmu_hat_exit(hatlockp);
12850 
12851 	return (0);
12852 }
12853 
12854 /*
12855  * Allocate and initialize a tsb_info structure.  Note that we may or may not
12856  * allocate a TSB here, depending on the flags passed in.
12857  */
12858 static int
12859 sfmmu_tsbinfo_alloc(struct tsb_info **tsbinfopp, int tsb_szc, int tte_sz_mask,
12860 	uint_t flags, sfmmu_t *sfmmup)
12861 {
12862 	int err;
12863 
12864 	*tsbinfopp = (struct tsb_info *)kmem_cache_alloc(
12865 	    sfmmu_tsbinfo_cache, KM_SLEEP);
12866 
12867 	if ((err = sfmmu_init_tsbinfo(*tsbinfopp, tte_sz_mask,
12868 	    tsb_szc, flags, sfmmup)) != 0) {
12869 		kmem_cache_free(sfmmu_tsbinfo_cache, *tsbinfopp);
12870 		SFMMU_STAT(sf_tsb_allocfail);
12871 		*tsbinfopp = NULL;
12872 		return (err);
12873 	}
12874 	SFMMU_STAT(sf_tsb_alloc);
12875 
12876 	/*
12877 	 * Bump the TSB size counters for this TSB size.
12878 	 */
12879 	(*(((int *)&sfmmu_tsbsize_stat) + tsb_szc))++;
12880 	return (0);
12881 }
12882 
12883 static void
12884 sfmmu_tsb_free(struct tsb_info *tsbinfo)
12885 {
12886 	caddr_t tsbva = tsbinfo->tsb_va;
12887 	uint_t tsb_size = TSB_BYTES(tsbinfo->tsb_szc);
12888 	struct kmem_cache *kmem_cachep = tsbinfo->tsb_cache;
12889 	vmem_t	*vmp = tsbinfo->tsb_vmp;
12890 
12891 	/*
12892 	 * If we allocated this TSB from relocatable kernel memory, then we
12893 	 * need to uninstall the callback handler.
12894 	 */
12895 	if (tsbinfo->tsb_cache != sfmmu_tsb8k_cache) {
12896 		uintptr_t slab_mask;
12897 		caddr_t slab_vaddr;
12898 		page_t **ppl;
12899 		int ret;
12900 
12901 		ASSERT(tsb_size <= MMU_PAGESIZE4M || use_bigtsb_arena);
12902 		if (tsb_size > MMU_PAGESIZE4M)
12903 			slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
12904 		else
12905 			slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
12906 		slab_vaddr = (caddr_t)((uintptr_t)tsbva & slab_mask);
12907 
12908 		ret = as_pagelock(&kas, &ppl, slab_vaddr, PAGESIZE, S_WRITE);
12909 		ASSERT(ret == 0);
12910 		hat_delete_callback(tsbva, (uint_t)tsb_size, (void *)tsbinfo,
12911 		    0, NULL);
12912 		as_pageunlock(&kas, ppl, slab_vaddr, PAGESIZE, S_WRITE);
12913 	}
12914 
12915 	if (kmem_cachep != NULL) {
12916 		kmem_cache_free(kmem_cachep, tsbva);
12917 	} else {
12918 		vmem_xfree(vmp, (void *)tsbva, tsb_size);
12919 	}
12920 	tsbinfo->tsb_va = (caddr_t)0xbad00bad;
12921 	atomic_add_64(&tsb_alloc_bytes, -(int64_t)tsb_size);
12922 }
12923 
12924 static void
12925 sfmmu_tsbinfo_free(struct tsb_info *tsbinfo)
12926 {
12927 	if ((tsbinfo->tsb_flags & TSB_SWAPPED) == 0) {
12928 		sfmmu_tsb_free(tsbinfo);
12929 	}
12930 	kmem_cache_free(sfmmu_tsbinfo_cache, tsbinfo);
12931 
12932 }
12933 
12934 /*
12935  * Setup all the references to physical memory for this tsbinfo.
12936  * The underlying page(s) must be locked.
12937  */
12938 static void
12939 sfmmu_tsbinfo_setup_phys(struct tsb_info *tsbinfo, pfn_t pfn)
12940 {
12941 	ASSERT(pfn != PFN_INVALID);
12942 	ASSERT(pfn == va_to_pfn(tsbinfo->tsb_va));
12943 
12944 #ifndef sun4v
12945 	if (tsbinfo->tsb_szc == 0) {
12946 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn,
12947 		    PROT_WRITE|PROT_READ, TTE8K);
12948 	} else {
12949 		/*
12950 		 * Round down PA and use a large mapping; the handlers will
12951 		 * compute the TSB pointer at the correct offset into the
12952 		 * big virtual page.  NOTE: this assumes all TSBs larger
12953 		 * than 8K must come from physically contiguous slabs of
12954 		 * size tsb_slab_size.
12955 		 */
12956 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn & ~tsb_slab_mask,
12957 		    PROT_WRITE|PROT_READ, tsb_slab_ttesz);
12958 	}
12959 	tsbinfo->tsb_pa = ptob(pfn);
12960 
12961 	TTE_SET_LOCKED(&tsbinfo->tsb_tte); /* lock the tte into dtlb */
12962 	TTE_SET_MOD(&tsbinfo->tsb_tte);    /* enable writes */
12963 
12964 	ASSERT(TTE_IS_PRIVILEGED(&tsbinfo->tsb_tte));
12965 	ASSERT(TTE_IS_LOCKED(&tsbinfo->tsb_tte));
12966 #else /* sun4v */
12967 	tsbinfo->tsb_pa = ptob(pfn);
12968 #endif /* sun4v */
12969 }
12970 
12971 
12972 /*
12973  * Returns zero on success, ENOMEM if over the high water mark,
12974  * or EAGAIN if the caller needs to retry with a smaller TSB
12975  * size (or specify TSB_FORCEALLOC if the allocation can't fail).
12976  *
12977  * This call cannot fail to allocate a TSB if TSB_FORCEALLOC
12978  * is specified and the TSB requested is PAGESIZE, though it
12979  * may sleep waiting for memory if sufficient memory is not
12980  * available.
12981  */
12982 static int
12983 sfmmu_init_tsbinfo(struct tsb_info *tsbinfo, int tteszmask,
12984     int tsbcode, uint_t flags, sfmmu_t *sfmmup)
12985 {
12986 	caddr_t vaddr = NULL;
12987 	caddr_t slab_vaddr;
12988 	uintptr_t slab_mask;
12989 	int tsbbytes = TSB_BYTES(tsbcode);
12990 	int lowmem = 0;
12991 	struct kmem_cache *kmem_cachep = NULL;
12992 	vmem_t *vmp = NULL;
12993 	lgrp_id_t lgrpid = LGRP_NONE;
12994 	pfn_t pfn;
12995 	uint_t cbflags = HAC_SLEEP;
12996 	page_t **pplist;
12997 	int ret;
12998 
12999 	ASSERT(tsbbytes <= MMU_PAGESIZE4M || use_bigtsb_arena);
13000 	if (tsbbytes > MMU_PAGESIZE4M)
13001 		slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
13002 	else
13003 		slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
13004 
13005 	if (flags & (TSB_FORCEALLOC | TSB_SWAPIN | TSB_GROW | TSB_SHRINK))
13006 		flags |= TSB_ALLOC;
13007 
13008 	ASSERT((flags & TSB_FORCEALLOC) == 0 || tsbcode == TSB_MIN_SZCODE);
13009 
13010 	tsbinfo->tsb_sfmmu = sfmmup;
13011 
13012 	/*
13013 	 * If not allocating a TSB, set up the tsbinfo, set TSB_SWAPPED, and
13014 	 * return.
13015 	 */
13016 	if ((flags & TSB_ALLOC) == 0) {
13017 		tsbinfo->tsb_szc = tsbcode;
13018 		tsbinfo->tsb_ttesz_mask = tteszmask;
13019 		tsbinfo->tsb_va = (caddr_t)0xbadbadbeef;
13020 		tsbinfo->tsb_pa = -1;
13021 		tsbinfo->tsb_tte.ll = 0;
13022 		tsbinfo->tsb_next = NULL;
13023 		tsbinfo->tsb_flags = TSB_SWAPPED;
13024 		tsbinfo->tsb_cache = NULL;
13025 		tsbinfo->tsb_vmp = NULL;
13026 		return (0);
13027 	}
13028 
13029 #ifdef DEBUG
13030 	/*
13031 	 * For debugging:
13032 	 * Randomly force allocation failures every tsb_alloc_mtbf
13033 	 * tries if TSB_FORCEALLOC is not specified.  This will
13034 	 * return ENOMEM if tsb_alloc_mtbf is odd, or EAGAIN if
13035 	 * it is even, to allow testing of both failure paths...
13036 	 */
13037 	if (tsb_alloc_mtbf && ((flags & TSB_FORCEALLOC) == 0) &&
13038 	    (tsb_alloc_count++ == tsb_alloc_mtbf)) {
13039 		tsb_alloc_count = 0;
13040 		tsb_alloc_fail_mtbf++;
13041 		return ((tsb_alloc_mtbf & 1)? ENOMEM : EAGAIN);
13042 	}
13043 #endif	/* DEBUG */
13044 
13045 	/*
13046 	 * Enforce high water mark if we are not doing a forced allocation
13047 	 * and are not shrinking a process' TSB.
13048 	 */
13049 	if ((flags & TSB_SHRINK) == 0 &&
13050 	    (tsbbytes + tsb_alloc_bytes) > tsb_alloc_hiwater) {
13051 		if ((flags & TSB_FORCEALLOC) == 0)
13052 			return (ENOMEM);
13053 		lowmem = 1;
13054 	}
13055 
13056 	/*
13057 	 * Allocate from the correct location based upon the size of the TSB
13058 	 * compared to the base page size, and what memory conditions dictate.
13059 	 * Note we always do nonblocking allocations from the TSB arena since
13060 	 * we don't want memory fragmentation to cause processes to block
13061 	 * indefinitely waiting for memory; until the kernel algorithms that
13062 	 * coalesce large pages are improved this is our best option.
13063 	 *
13064 	 * Algorithm:
13065 	 *	If allocating a "large" TSB (>8K), allocate from the
13066 	 *		appropriate kmem_tsb_default_arena vmem arena
13067 	 *	else if low on memory or the TSB_FORCEALLOC flag is set or
13068 	 *	tsb_forceheap is set
13069 	 *		Allocate from kernel heap via sfmmu_tsb8k_cache with
13070 	 *		KM_SLEEP (never fails)
13071 	 *	else
13072 	 *		Allocate from appropriate sfmmu_tsb_cache with
13073 	 *		KM_NOSLEEP
13074 	 *	endif
13075 	 */
13076 	if (tsb_lgrp_affinity)
13077 		lgrpid = lgrp_home_id(curthread);
13078 	if (lgrpid == LGRP_NONE)
13079 		lgrpid = 0;	/* use lgrp of boot CPU */
13080 
13081 	if (tsbbytes > MMU_PAGESIZE) {
13082 		if (tsbbytes > MMU_PAGESIZE4M) {
13083 			vmp = kmem_bigtsb_default_arena[lgrpid];
13084 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
13085 			    0, 0, NULL, NULL, VM_NOSLEEP);
13086 		} else {
13087 			vmp = kmem_tsb_default_arena[lgrpid];
13088 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
13089 			    0, 0, NULL, NULL, VM_NOSLEEP);
13090 		}
13091 #ifdef	DEBUG
13092 	} else if (lowmem || (flags & TSB_FORCEALLOC) || tsb_forceheap) {
13093 #else	/* !DEBUG */
13094 	} else if (lowmem || (flags & TSB_FORCEALLOC)) {
13095 #endif	/* DEBUG */
13096 		kmem_cachep = sfmmu_tsb8k_cache;
13097 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_SLEEP);
13098 		ASSERT(vaddr != NULL);
13099 	} else {
13100 		kmem_cachep = sfmmu_tsb_cache[lgrpid];
13101 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_NOSLEEP);
13102 	}
13103 
13104 	tsbinfo->tsb_cache = kmem_cachep;
13105 	tsbinfo->tsb_vmp = vmp;
13106 
13107 	if (vaddr == NULL) {
13108 		return (EAGAIN);
13109 	}
13110 
13111 	atomic_add_64(&tsb_alloc_bytes, (int64_t)tsbbytes);
13112 	kmem_cachep = tsbinfo->tsb_cache;
13113 
13114 	/*
13115 	 * If we are allocating from outside the cage, then we need to
13116 	 * register a relocation callback handler.  Note that for now
13117 	 * since pseudo mappings always hang off of the slab's root page,
13118 	 * we need only lock the first 8K of the TSB slab.  This is a bit
13119 	 * hacky but it is good for performance.
13120 	 */
13121 	if (kmem_cachep != sfmmu_tsb8k_cache) {
13122 		slab_vaddr = (caddr_t)((uintptr_t)vaddr & slab_mask);
13123 		ret = as_pagelock(&kas, &pplist, slab_vaddr, PAGESIZE, S_WRITE);
13124 		ASSERT(ret == 0);
13125 		ret = hat_add_callback(sfmmu_tsb_cb_id, vaddr, (uint_t)tsbbytes,
13126 		    cbflags, (void *)tsbinfo, &pfn, NULL);
13127 
13128 		/*
13129 		 * Need to free up resources if we could not successfully
13130 		 * add the callback function and return an error condition.
13131 		 */
13132 		if (ret != 0) {
13133 			if (kmem_cachep) {
13134 				kmem_cache_free(kmem_cachep, vaddr);
13135 			} else {
13136 				vmem_xfree(vmp, (void *)vaddr, tsbbytes);
13137 			}
13138 			as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE,
13139 			    S_WRITE);
13140 			return (EAGAIN);
13141 		}
13142 	} else {
13143 		/*
13144 		 * Since allocation of 8K TSBs from heap is rare and occurs
13145 		 * during memory pressure we allocate them from permanent
13146 		 * memory rather than using callbacks to get the PFN.
13147 		 */
13148 		pfn = hat_getpfnum(kas.a_hat, vaddr);
13149 	}
13150 
13151 	tsbinfo->tsb_va = vaddr;
13152 	tsbinfo->tsb_szc = tsbcode;
13153 	tsbinfo->tsb_ttesz_mask = tteszmask;
13154 	tsbinfo->tsb_next = NULL;
13155 	tsbinfo->tsb_flags = 0;
13156 
13157 	sfmmu_tsbinfo_setup_phys(tsbinfo, pfn);
13158 
13159 	sfmmu_inv_tsb(vaddr, tsbbytes);
13160 
13161 	if (kmem_cachep != sfmmu_tsb8k_cache) {
13162 		as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE, S_WRITE);
13163 	}
13164 
13165 	return (0);
13166 }
13167 
13168 /*
13169  * Initialize per cpu tsb and per cpu tsbmiss_area
13170  */
13171 void
13172 sfmmu_init_tsbs(void)
13173 {
13174 	int i;
13175 	struct tsbmiss	*tsbmissp;
13176 	struct kpmtsbm	*kpmtsbmp;
13177 #ifndef sun4v
13178 	extern int	dcache_line_mask;
13179 #endif /* sun4v */
13180 	extern uint_t	vac_colors;
13181 
13182 	/*
13183 	 * Init. tsb miss area.
13184 	 */
13185 	tsbmissp = tsbmiss_area;
13186 
13187 	for (i = 0; i < NCPU; tsbmissp++, i++) {
13188 		/*
13189 		 * initialize the tsbmiss area.
13190 		 * Do this for all possible CPUs as some may be added
13191 		 * while the system is running. There is no cost to this.
13192 		 */
13193 		tsbmissp->ksfmmup = ksfmmup;
13194 #ifndef sun4v
13195 		tsbmissp->dcache_line_mask = (uint16_t)dcache_line_mask;
13196 #endif /* sun4v */
13197 		tsbmissp->khashstart =
13198 		    (struct hmehash_bucket *)va_to_pa((caddr_t)khme_hash);
13199 		tsbmissp->uhashstart =
13200 		    (struct hmehash_bucket *)va_to_pa((caddr_t)uhme_hash);
13201 		tsbmissp->khashsz = khmehash_num;
13202 		tsbmissp->uhashsz = uhmehash_num;
13203 	}
13204 
13205 	sfmmu_tsb_cb_id = hat_register_callback('T'<<16 | 'S' << 8 | 'B',
13206 	    sfmmu_tsb_pre_relocator, sfmmu_tsb_post_relocator, NULL, 0);
13207 
13208 	if (kpm_enable == 0)
13209 		return;
13210 
13211 	/* -- Begin KPM specific init -- */
13212 
13213 	if (kpm_smallpages) {
13214 		/*
13215 		 * If we're using base pagesize pages for seg_kpm
13216 		 * mappings, we use the kernel TSB since we can't afford
13217 		 * to allocate a second huge TSB for these mappings.
13218 		 */
13219 		kpm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13220 		kpm_tsbsz = ktsb_szcode;
13221 		kpmsm_tsbbase = kpm_tsbbase;
13222 		kpmsm_tsbsz = kpm_tsbsz;
13223 	} else {
13224 		/*
13225 		 * In VAC conflict case, just put the entries in the
13226 		 * kernel 8K indexed TSB for now so we can find them.
13227 		 * This could really be changed in the future if we feel
13228 		 * the need...
13229 		 */
13230 		kpmsm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13231 		kpmsm_tsbsz = ktsb_szcode;
13232 		kpm_tsbbase = ktsb_phys? ktsb4m_pbase : (uint64_t)ktsb4m_base;
13233 		kpm_tsbsz = ktsb4m_szcode;
13234 	}
13235 
13236 	kpmtsbmp = kpmtsbm_area;
13237 	for (i = 0; i < NCPU; kpmtsbmp++, i++) {
13238 		/*
13239 		 * Initialize the kpmtsbm area.
13240 		 * Do this for all possible CPUs as some may be added
13241 		 * while the system is running. There is no cost to this.
13242 		 */
13243 		kpmtsbmp->vbase = kpm_vbase;
13244 		kpmtsbmp->vend = kpm_vbase + kpm_size * vac_colors;
13245 		kpmtsbmp->sz_shift = kpm_size_shift;
13246 		kpmtsbmp->kpmp_shift = kpmp_shift;
13247 		kpmtsbmp->kpmp2pshft = (uchar_t)kpmp2pshft;
13248 		if (kpm_smallpages == 0) {
13249 			kpmtsbmp->kpmp_table_sz = kpmp_table_sz;
13250 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_table);
13251 		} else {
13252 			kpmtsbmp->kpmp_table_sz = kpmp_stable_sz;
13253 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_stable);
13254 		}
13255 		kpmtsbmp->msegphashpa = va_to_pa(memseg_phash);
13256 		kpmtsbmp->flags = KPMTSBM_ENABLE_FLAG;
13257 #ifdef	DEBUG
13258 		kpmtsbmp->flags |= (kpm_tsbmtl) ?  KPMTSBM_TLTSBM_FLAG : 0;
13259 #endif	/* DEBUG */
13260 		if (ktsb_phys)
13261 			kpmtsbmp->flags |= KPMTSBM_TSBPHYS_FLAG;
13262 	}
13263 
13264 	/* -- End KPM specific init -- */
13265 }
13266 
13267 /* Avoid using sfmmu_tsbinfo_alloc() to avoid kmem_alloc - no real reason */
13268 struct tsb_info ktsb_info[2];
13269 
13270 /*
13271  * Called from hat_kern_setup() to setup the tsb_info for ksfmmup.
13272  */
13273 void
13274 sfmmu_init_ktsbinfo()
13275 {
13276 	ASSERT(ksfmmup != NULL);
13277 	ASSERT(ksfmmup->sfmmu_tsb == NULL);
13278 	/*
13279 	 * Allocate tsbinfos for kernel and copy in data
13280 	 * to make debug easier and sun4v setup easier.
13281 	 */
13282 	ktsb_info[0].tsb_sfmmu = ksfmmup;
13283 	ktsb_info[0].tsb_szc = ktsb_szcode;
13284 	ktsb_info[0].tsb_ttesz_mask = TSB8K|TSB64K|TSB512K;
13285 	ktsb_info[0].tsb_va = ktsb_base;
13286 	ktsb_info[0].tsb_pa = ktsb_pbase;
13287 	ktsb_info[0].tsb_flags = 0;
13288 	ktsb_info[0].tsb_tte.ll = 0;
13289 	ktsb_info[0].tsb_cache = NULL;
13290 
13291 	ktsb_info[1].tsb_sfmmu = ksfmmup;
13292 	ktsb_info[1].tsb_szc = ktsb4m_szcode;
13293 	ktsb_info[1].tsb_ttesz_mask = TSB4M;
13294 	ktsb_info[1].tsb_va = ktsb4m_base;
13295 	ktsb_info[1].tsb_pa = ktsb4m_pbase;
13296 	ktsb_info[1].tsb_flags = 0;
13297 	ktsb_info[1].tsb_tte.ll = 0;
13298 	ktsb_info[1].tsb_cache = NULL;
13299 
13300 	/* Link them into ksfmmup. */
13301 	ktsb_info[0].tsb_next = &ktsb_info[1];
13302 	ktsb_info[1].tsb_next = NULL;
13303 	ksfmmup->sfmmu_tsb = &ktsb_info[0];
13304 
13305 	sfmmu_setup_tsbinfo(ksfmmup);
13306 }
13307 
13308 /*
13309  * Cache the last value returned from va_to_pa().  If the VA specified
13310  * in the current call to cached_va_to_pa() maps to the same Page (as the
13311  * previous call to cached_va_to_pa()), then compute the PA using
13312  * cached info, else call va_to_pa().
13313  *
13314  * Note: this function is neither MT-safe nor consistent in the presence
13315  * of multiple, interleaved threads.  This function was created to enable
13316  * an optimization used during boot (at a point when there's only one thread
13317  * executing on the "boot CPU", and before startup_vm() has been called).
13318  */
13319 static uint64_t
13320 cached_va_to_pa(void *vaddr)
13321 {
13322 	static uint64_t prev_vaddr_base = 0;
13323 	static uint64_t prev_pfn = 0;
13324 
13325 	if ((((uint64_t)vaddr) & MMU_PAGEMASK) == prev_vaddr_base) {
13326 		return (prev_pfn | ((uint64_t)vaddr & MMU_PAGEOFFSET));
13327 	} else {
13328 		uint64_t pa = va_to_pa(vaddr);
13329 
13330 		if (pa != ((uint64_t)-1)) {
13331 			/*
13332 			 * Computed physical address is valid.  Cache its
13333 			 * related info for the next cached_va_to_pa() call.
13334 			 */
13335 			prev_pfn = pa & MMU_PAGEMASK;
13336 			prev_vaddr_base = ((uint64_t)vaddr) & MMU_PAGEMASK;
13337 		}
13338 
13339 		return (pa);
13340 	}
13341 }
13342 
13343 /*
13344  * Carve up our nucleus hblk region.  We may allocate more hblks than
13345  * asked due to rounding errors but we are guaranteed to have at least
13346  * enough space to allocate the requested number of hblk8's and hblk1's.
13347  */
13348 void
13349 sfmmu_init_nucleus_hblks(caddr_t addr, size_t size, int nhblk8, int nhblk1)
13350 {
13351 	struct hme_blk *hmeblkp;
13352 	size_t hme8blk_sz, hme1blk_sz;
13353 	size_t i;
13354 	size_t hblk8_bound;
13355 	ulong_t j = 0, k = 0;
13356 
13357 	ASSERT(addr != NULL && size != 0);
13358 
13359 	/* Need to use proper structure alignment */
13360 	hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t));
13361 	hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t));
13362 
13363 	nucleus_hblk8.list = (void *)addr;
13364 	nucleus_hblk8.index = 0;
13365 
13366 	/*
13367 	 * Use as much memory as possible for hblk8's since we
13368 	 * expect all bop_alloc'ed memory to be allocated in 8k chunks.
13369 	 * We need to hold back enough space for the hblk1's which
13370 	 * we'll allocate next.
13371 	 */
13372 	hblk8_bound = size - (nhblk1 * hme1blk_sz) - hme8blk_sz;
13373 	for (i = 0; i <= hblk8_bound; i += hme8blk_sz, j++) {
13374 		hmeblkp = (struct hme_blk *)addr;
13375 		addr += hme8blk_sz;
13376 		hmeblkp->hblk_nuc_bit = 1;
13377 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13378 	}
13379 	nucleus_hblk8.len = j;
13380 	ASSERT(j >= nhblk8);
13381 	SFMMU_STAT_ADD(sf_hblk8_ncreate, j);
13382 
13383 	nucleus_hblk1.list = (void *)addr;
13384 	nucleus_hblk1.index = 0;
13385 	for (; i <= (size - hme1blk_sz); i += hme1blk_sz, k++) {
13386 		hmeblkp = (struct hme_blk *)addr;
13387 		addr += hme1blk_sz;
13388 		hmeblkp->hblk_nuc_bit = 1;
13389 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13390 	}
13391 	ASSERT(k >= nhblk1);
13392 	nucleus_hblk1.len = k;
13393 	SFMMU_STAT_ADD(sf_hblk1_ncreate, k);
13394 }
13395 
13396 /*
13397  * This function is currently not supported on this platform. For what
13398  * it's supposed to do, see hat.c and hat_srmmu.c
13399  */
13400 /* ARGSUSED */
13401 faultcode_t
13402 hat_softlock(struct hat *hat, caddr_t addr, size_t *lenp, page_t **ppp,
13403     uint_t flags)
13404 {
13405 	ASSERT(hat->sfmmu_xhat_provider == NULL);
13406 	return (FC_NOSUPPORT);
13407 }
13408 
13409 /*
13410  * Searchs the mapping list of the page for a mapping of the same size. If not
13411  * found the corresponding bit is cleared in the p_index field. When large
13412  * pages are more prevalent in the system, we can maintain the mapping list
13413  * in order and we don't have to traverse the list each time. Just check the
13414  * next and prev entries, and if both are of different size, we clear the bit.
13415  */
13416 static void
13417 sfmmu_rm_large_mappings(page_t *pp, int ttesz)
13418 {
13419 	struct sf_hment *sfhmep;
13420 	struct hme_blk *hmeblkp;
13421 	int	index;
13422 	pgcnt_t	npgs;
13423 
13424 	ASSERT(ttesz > TTE8K);
13425 
13426 	ASSERT(sfmmu_mlist_held(pp));
13427 
13428 	ASSERT(PP_ISMAPPED_LARGE(pp));
13429 
13430 	/*
13431 	 * Traverse mapping list looking for another mapping of same size.
13432 	 * since we only want to clear index field if all mappings of
13433 	 * that size are gone.
13434 	 */
13435 
13436 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
13437 		if (IS_PAHME(sfhmep))
13438 			continue;
13439 		hmeblkp = sfmmu_hmetohblk(sfhmep);
13440 		if (hmeblkp->hblk_xhat_bit)
13441 			continue;
13442 		if (hme_size(sfhmep) == ttesz) {
13443 			/*
13444 			 * another mapping of the same size. don't clear index.
13445 			 */
13446 			return;
13447 		}
13448 	}
13449 
13450 	/*
13451 	 * Clear the p_index bit for large page.
13452 	 */
13453 	index = PAGESZ_TO_INDEX(ttesz);
13454 	npgs = TTEPAGES(ttesz);
13455 	while (npgs-- > 0) {
13456 		ASSERT(pp->p_index & index);
13457 		pp->p_index &= ~index;
13458 		pp = PP_PAGENEXT(pp);
13459 	}
13460 }
13461 
13462 /*
13463  * return supported features
13464  */
13465 /* ARGSUSED */
13466 int
13467 hat_supported(enum hat_features feature, void *arg)
13468 {
13469 	switch (feature) {
13470 	case    HAT_SHARED_PT:
13471 	case	HAT_DYNAMIC_ISM_UNMAP:
13472 	case	HAT_VMODSORT:
13473 		return (1);
13474 	case	HAT_SHARED_REGIONS:
13475 		if (shctx_on)
13476 			return (1);
13477 		else
13478 			return (0);
13479 	default:
13480 		return (0);
13481 	}
13482 }
13483 
13484 void
13485 hat_enter(struct hat *hat)
13486 {
13487 	hatlock_t	*hatlockp;
13488 
13489 	if (hat != ksfmmup) {
13490 		hatlockp = TSB_HASH(hat);
13491 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
13492 	}
13493 }
13494 
13495 void
13496 hat_exit(struct hat *hat)
13497 {
13498 	hatlock_t	*hatlockp;
13499 
13500 	if (hat != ksfmmup) {
13501 		hatlockp = TSB_HASH(hat);
13502 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
13503 	}
13504 }
13505 
13506 /*ARGSUSED*/
13507 void
13508 hat_reserve(struct as *as, caddr_t addr, size_t len)
13509 {
13510 }
13511 
13512 static void
13513 hat_kstat_init(void)
13514 {
13515 	kstat_t *ksp;
13516 
13517 	ksp = kstat_create("unix", 0, "sfmmu_global_stat", "hat",
13518 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_global_stat),
13519 	    KSTAT_FLAG_VIRTUAL);
13520 	if (ksp) {
13521 		ksp->ks_data = (void *) &sfmmu_global_stat;
13522 		kstat_install(ksp);
13523 	}
13524 	ksp = kstat_create("unix", 0, "sfmmu_tsbsize_stat", "hat",
13525 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_tsbsize_stat),
13526 	    KSTAT_FLAG_VIRTUAL);
13527 	if (ksp) {
13528 		ksp->ks_data = (void *) &sfmmu_tsbsize_stat;
13529 		kstat_install(ksp);
13530 	}
13531 	ksp = kstat_create("unix", 0, "sfmmu_percpu_stat", "hat",
13532 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_percpu_stat) * NCPU,
13533 	    KSTAT_FLAG_WRITABLE);
13534 	if (ksp) {
13535 		ksp->ks_update = sfmmu_kstat_percpu_update;
13536 		kstat_install(ksp);
13537 	}
13538 }
13539 
13540 /* ARGSUSED */
13541 static int
13542 sfmmu_kstat_percpu_update(kstat_t *ksp, int rw)
13543 {
13544 	struct sfmmu_percpu_stat *cpu_kstat = ksp->ks_data;
13545 	struct tsbmiss *tsbm = tsbmiss_area;
13546 	struct kpmtsbm *kpmtsbm = kpmtsbm_area;
13547 	int i;
13548 
13549 	ASSERT(cpu_kstat);
13550 	if (rw == KSTAT_READ) {
13551 		for (i = 0; i < NCPU; cpu_kstat++, tsbm++, kpmtsbm++, i++) {
13552 			cpu_kstat->sf_itlb_misses = 0;
13553 			cpu_kstat->sf_dtlb_misses = 0;
13554 			cpu_kstat->sf_utsb_misses = tsbm->utsb_misses -
13555 			    tsbm->uprot_traps;
13556 			cpu_kstat->sf_ktsb_misses = tsbm->ktsb_misses +
13557 			    kpmtsbm->kpm_tsb_misses - tsbm->kprot_traps;
13558 			cpu_kstat->sf_tsb_hits = 0;
13559 			cpu_kstat->sf_umod_faults = tsbm->uprot_traps;
13560 			cpu_kstat->sf_kmod_faults = tsbm->kprot_traps;
13561 		}
13562 	} else {
13563 		/* KSTAT_WRITE is used to clear stats */
13564 		for (i = 0; i < NCPU; tsbm++, kpmtsbm++, i++) {
13565 			tsbm->utsb_misses = 0;
13566 			tsbm->ktsb_misses = 0;
13567 			tsbm->uprot_traps = 0;
13568 			tsbm->kprot_traps = 0;
13569 			kpmtsbm->kpm_dtlb_misses = 0;
13570 			kpmtsbm->kpm_tsb_misses = 0;
13571 		}
13572 	}
13573 	return (0);
13574 }
13575 
13576 #ifdef	DEBUG
13577 
13578 tte_t  *gorig[NCPU], *gcur[NCPU], *gnew[NCPU];
13579 
13580 /*
13581  * A tte checker. *orig_old is the value we read before cas.
13582  *	*cur is the value returned by cas.
13583  *	*new is the desired value when we do the cas.
13584  *
13585  *	*hmeblkp is currently unused.
13586  */
13587 
13588 /* ARGSUSED */
13589 void
13590 chk_tte(tte_t *orig_old, tte_t *cur, tte_t *new, struct hme_blk *hmeblkp)
13591 {
13592 	pfn_t i, j, k;
13593 	int cpuid = CPU->cpu_id;
13594 
13595 	gorig[cpuid] = orig_old;
13596 	gcur[cpuid] = cur;
13597 	gnew[cpuid] = new;
13598 
13599 #ifdef lint
13600 	hmeblkp = hmeblkp;
13601 #endif
13602 
13603 	if (TTE_IS_VALID(orig_old)) {
13604 		if (TTE_IS_VALID(cur)) {
13605 			i = TTE_TO_TTEPFN(orig_old);
13606 			j = TTE_TO_TTEPFN(cur);
13607 			k = TTE_TO_TTEPFN(new);
13608 			if (i != j) {
13609 				/* remap error? */
13610 				panic("chk_tte: bad pfn, 0x%lx, 0x%lx", i, j);
13611 			}
13612 
13613 			if (i != k) {
13614 				/* remap error? */
13615 				panic("chk_tte: bad pfn2, 0x%lx, 0x%lx", i, k);
13616 			}
13617 		} else {
13618 			if (TTE_IS_VALID(new)) {
13619 				panic("chk_tte: invalid cur? ");
13620 			}
13621 
13622 			i = TTE_TO_TTEPFN(orig_old);
13623 			k = TTE_TO_TTEPFN(new);
13624 			if (i != k) {
13625 				panic("chk_tte: bad pfn3, 0x%lx, 0x%lx", i, k);
13626 			}
13627 		}
13628 	} else {
13629 		if (TTE_IS_VALID(cur)) {
13630 			j = TTE_TO_TTEPFN(cur);
13631 			if (TTE_IS_VALID(new)) {
13632 				k = TTE_TO_TTEPFN(new);
13633 				if (j != k) {
13634 					panic("chk_tte: bad pfn4, 0x%lx, 0x%lx",
13635 					    j, k);
13636 				}
13637 			} else {
13638 				panic("chk_tte: why here?");
13639 			}
13640 		} else {
13641 			if (!TTE_IS_VALID(new)) {
13642 				panic("chk_tte: why here2 ?");
13643 			}
13644 		}
13645 	}
13646 }
13647 
13648 #endif /* DEBUG */
13649 
13650 extern void prefetch_tsbe_read(struct tsbe *);
13651 extern void prefetch_tsbe_write(struct tsbe *);
13652 
13653 
13654 /*
13655  * We want to prefetch 7 cache lines ahead for our read prefetch.  This gives
13656  * us optimal performance on Cheetah+.  You can only have 8 outstanding
13657  * prefetches at any one time, so we opted for 7 read prefetches and 1 write
13658  * prefetch to make the most utilization of the prefetch capability.
13659  */
13660 #define	TSBE_PREFETCH_STRIDE (7)
13661 
13662 void
13663 sfmmu_copy_tsb(struct tsb_info *old_tsbinfo, struct tsb_info *new_tsbinfo)
13664 {
13665 	int old_bytes = TSB_BYTES(old_tsbinfo->tsb_szc);
13666 	int new_bytes = TSB_BYTES(new_tsbinfo->tsb_szc);
13667 	int old_entries = TSB_ENTRIES(old_tsbinfo->tsb_szc);
13668 	int new_entries = TSB_ENTRIES(new_tsbinfo->tsb_szc);
13669 	struct tsbe *old;
13670 	struct tsbe *new;
13671 	struct tsbe *new_base = (struct tsbe *)new_tsbinfo->tsb_va;
13672 	uint64_t va;
13673 	int new_offset;
13674 	int i;
13675 	int vpshift;
13676 	int last_prefetch;
13677 
13678 	if (old_bytes == new_bytes) {
13679 		bcopy(old_tsbinfo->tsb_va, new_tsbinfo->tsb_va, new_bytes);
13680 	} else {
13681 
13682 		/*
13683 		 * A TSBE is 16 bytes which means there are four TSBE's per
13684 		 * P$ line (64 bytes), thus every 4 TSBE's we prefetch.
13685 		 */
13686 		old = (struct tsbe *)old_tsbinfo->tsb_va;
13687 		last_prefetch = old_entries - (4*(TSBE_PREFETCH_STRIDE+1));
13688 		for (i = 0; i < old_entries; i++, old++) {
13689 			if (((i & (4-1)) == 0) && (i < last_prefetch))
13690 				prefetch_tsbe_read(old);
13691 			if (!old->tte_tag.tag_invalid) {
13692 				/*
13693 				 * We have a valid TTE to remap.  Check the
13694 				 * size.  We won't remap 64K or 512K TTEs
13695 				 * because they span more than one TSB entry
13696 				 * and are indexed using an 8K virt. page.
13697 				 * Ditto for 32M and 256M TTEs.
13698 				 */
13699 				if (TTE_CSZ(&old->tte_data) == TTE64K ||
13700 				    TTE_CSZ(&old->tte_data) == TTE512K)
13701 					continue;
13702 				if (mmu_page_sizes == max_mmu_page_sizes) {
13703 					if (TTE_CSZ(&old->tte_data) == TTE32M ||
13704 					    TTE_CSZ(&old->tte_data) == TTE256M)
13705 						continue;
13706 				}
13707 
13708 				/* clear the lower 22 bits of the va */
13709 				va = *(uint64_t *)old << 22;
13710 				/* turn va into a virtual pfn */
13711 				va >>= 22 - TSB_START_SIZE;
13712 				/*
13713 				 * or in bits from the offset in the tsb
13714 				 * to get the real virtual pfn. These
13715 				 * correspond to bits [21:13] in the va
13716 				 */
13717 				vpshift =
13718 				    TTE_BSZS_SHIFT(TTE_CSZ(&old->tte_data)) &
13719 				    0x1ff;
13720 				va |= (i << vpshift);
13721 				va >>= vpshift;
13722 				new_offset = va & (new_entries - 1);
13723 				new = new_base + new_offset;
13724 				prefetch_tsbe_write(new);
13725 				*new = *old;
13726 			}
13727 		}
13728 	}
13729 }
13730 
13731 /*
13732  * unused in sfmmu
13733  */
13734 void
13735 hat_dump(void)
13736 {
13737 }
13738 
13739 /*
13740  * Called when a thread is exiting and we have switched to the kernel address
13741  * space.  Perform the same VM initialization resume() uses when switching
13742  * processes.
13743  *
13744  * Note that sfmmu_load_mmustate() is currently a no-op for kernel threads, but
13745  * we call it anyway in case the semantics change in the future.
13746  */
13747 /*ARGSUSED*/
13748 void
13749 hat_thread_exit(kthread_t *thd)
13750 {
13751 	uint_t pgsz_cnum;
13752 	uint_t pstate_save;
13753 
13754 	ASSERT(thd->t_procp->p_as == &kas);
13755 
13756 	pgsz_cnum = KCONTEXT;
13757 #ifdef sun4u
13758 	pgsz_cnum |= (ksfmmup->sfmmu_cext << CTXREG_EXT_SHIFT);
13759 #endif
13760 
13761 	/*
13762 	 * Note that sfmmu_load_mmustate() is currently a no-op for
13763 	 * kernel threads. We need to disable interrupts here,
13764 	 * simply because otherwise sfmmu_load_mmustate() would panic
13765 	 * if the caller does not disable interrupts.
13766 	 */
13767 	pstate_save = sfmmu_disable_intrs();
13768 
13769 	/* Compatibility Note: hw takes care of MMU_SCONTEXT1 */
13770 	sfmmu_setctx_sec(pgsz_cnum);
13771 	sfmmu_load_mmustate(ksfmmup);
13772 	sfmmu_enable_intrs(pstate_save);
13773 }
13774 
13775 
13776 /*
13777  * SRD support
13778  */
13779 #define	SRD_HASH_FUNCTION(vp)	(((((uintptr_t)(vp)) >> 4) ^ \
13780 				    (((uintptr_t)(vp)) >> 11)) & \
13781 				    srd_hashmask)
13782 
13783 /*
13784  * Attach the process to the srd struct associated with the exec vnode
13785  * from which the process is started.
13786  */
13787 void
13788 hat_join_srd(struct hat *sfmmup, vnode_t *evp)
13789 {
13790 	uint_t hash = SRD_HASH_FUNCTION(evp);
13791 	sf_srd_t *srdp;
13792 	sf_srd_t *newsrdp;
13793 
13794 	ASSERT(sfmmup != ksfmmup);
13795 	ASSERT(sfmmup->sfmmu_srdp == NULL);
13796 
13797 	if (!shctx_on) {
13798 		return;
13799 	}
13800 
13801 	VN_HOLD(evp);
13802 
13803 	if (srd_buckets[hash].srdb_srdp != NULL) {
13804 		mutex_enter(&srd_buckets[hash].srdb_lock);
13805 		for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13806 		    srdp = srdp->srd_hash) {
13807 			if (srdp->srd_evp == evp) {
13808 				ASSERT(srdp->srd_refcnt >= 0);
13809 				sfmmup->sfmmu_srdp = srdp;
13810 				atomic_inc_32(
13811 				    (volatile uint_t *)&srdp->srd_refcnt);
13812 				mutex_exit(&srd_buckets[hash].srdb_lock);
13813 				return;
13814 			}
13815 		}
13816 		mutex_exit(&srd_buckets[hash].srdb_lock);
13817 	}
13818 	newsrdp = kmem_cache_alloc(srd_cache, KM_SLEEP);
13819 	ASSERT(newsrdp->srd_next_ismrid == 0 && newsrdp->srd_next_hmerid == 0);
13820 
13821 	newsrdp->srd_evp = evp;
13822 	newsrdp->srd_refcnt = 1;
13823 	newsrdp->srd_hmergnfree = NULL;
13824 	newsrdp->srd_ismrgnfree = NULL;
13825 
13826 	mutex_enter(&srd_buckets[hash].srdb_lock);
13827 	for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13828 	    srdp = srdp->srd_hash) {
13829 		if (srdp->srd_evp == evp) {
13830 			ASSERT(srdp->srd_refcnt >= 0);
13831 			sfmmup->sfmmu_srdp = srdp;
13832 			atomic_inc_32((volatile uint_t *)&srdp->srd_refcnt);
13833 			mutex_exit(&srd_buckets[hash].srdb_lock);
13834 			kmem_cache_free(srd_cache, newsrdp);
13835 			return;
13836 		}
13837 	}
13838 	newsrdp->srd_hash = srd_buckets[hash].srdb_srdp;
13839 	srd_buckets[hash].srdb_srdp = newsrdp;
13840 	sfmmup->sfmmu_srdp = newsrdp;
13841 
13842 	mutex_exit(&srd_buckets[hash].srdb_lock);
13843 
13844 }
13845 
13846 static void
13847 sfmmu_leave_srd(sfmmu_t *sfmmup)
13848 {
13849 	vnode_t *evp;
13850 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
13851 	uint_t hash;
13852 	sf_srd_t **prev_srdpp;
13853 	sf_region_t *rgnp;
13854 	sf_region_t *nrgnp;
13855 #ifdef DEBUG
13856 	int rgns = 0;
13857 #endif
13858 	int i;
13859 
13860 	ASSERT(sfmmup != ksfmmup);
13861 	ASSERT(srdp != NULL);
13862 	ASSERT(srdp->srd_refcnt > 0);
13863 	ASSERT(sfmmup->sfmmu_scdp == NULL);
13864 	ASSERT(sfmmup->sfmmu_free == 1);
13865 
13866 	sfmmup->sfmmu_srdp = NULL;
13867 	evp = srdp->srd_evp;
13868 	ASSERT(evp != NULL);
13869 	if (atomic_dec_32_nv((volatile uint_t *)&srdp->srd_refcnt)) {
13870 		VN_RELE(evp);
13871 		return;
13872 	}
13873 
13874 	hash = SRD_HASH_FUNCTION(evp);
13875 	mutex_enter(&srd_buckets[hash].srdb_lock);
13876 	for (prev_srdpp = &srd_buckets[hash].srdb_srdp;
13877 	    (srdp = *prev_srdpp) != NULL; prev_srdpp = &srdp->srd_hash) {
13878 		if (srdp->srd_evp == evp) {
13879 			break;
13880 		}
13881 	}
13882 	if (srdp == NULL || srdp->srd_refcnt) {
13883 		mutex_exit(&srd_buckets[hash].srdb_lock);
13884 		VN_RELE(evp);
13885 		return;
13886 	}
13887 	*prev_srdpp = srdp->srd_hash;
13888 	mutex_exit(&srd_buckets[hash].srdb_lock);
13889 
13890 	ASSERT(srdp->srd_refcnt == 0);
13891 	VN_RELE(evp);
13892 
13893 #ifdef DEBUG
13894 	for (i = 0; i < SFMMU_MAX_REGION_BUCKETS; i++) {
13895 		ASSERT(srdp->srd_rgnhash[i] == NULL);
13896 	}
13897 #endif /* DEBUG */
13898 
13899 	/* free each hme regions in the srd */
13900 	for (rgnp = srdp->srd_hmergnfree; rgnp != NULL; rgnp = nrgnp) {
13901 		nrgnp = rgnp->rgn_next;
13902 		ASSERT(rgnp->rgn_id < srdp->srd_next_hmerid);
13903 		ASSERT(rgnp->rgn_refcnt == 0);
13904 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13905 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13906 		ASSERT(rgnp->rgn_hmeflags == 0);
13907 		ASSERT(srdp->srd_hmergnp[rgnp->rgn_id] == rgnp);
13908 #ifdef DEBUG
13909 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13910 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13911 		}
13912 		rgns++;
13913 #endif /* DEBUG */
13914 		kmem_cache_free(region_cache, rgnp);
13915 	}
13916 	ASSERT(rgns == srdp->srd_next_hmerid);
13917 
13918 #ifdef DEBUG
13919 	rgns = 0;
13920 #endif
13921 	/* free each ism rgns in the srd */
13922 	for (rgnp = srdp->srd_ismrgnfree; rgnp != NULL; rgnp = nrgnp) {
13923 		nrgnp = rgnp->rgn_next;
13924 		ASSERT(rgnp->rgn_id < srdp->srd_next_ismrid);
13925 		ASSERT(rgnp->rgn_refcnt == 0);
13926 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13927 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13928 		ASSERT(srdp->srd_ismrgnp[rgnp->rgn_id] == rgnp);
13929 #ifdef DEBUG
13930 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13931 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13932 		}
13933 		rgns++;
13934 #endif /* DEBUG */
13935 		kmem_cache_free(region_cache, rgnp);
13936 	}
13937 	ASSERT(rgns == srdp->srd_next_ismrid);
13938 	ASSERT(srdp->srd_ismbusyrgns == 0);
13939 	ASSERT(srdp->srd_hmebusyrgns == 0);
13940 
13941 	srdp->srd_next_ismrid = 0;
13942 	srdp->srd_next_hmerid = 0;
13943 
13944 	bzero((void *)srdp->srd_ismrgnp,
13945 	    sizeof (sf_region_t *) * SFMMU_MAX_ISM_REGIONS);
13946 	bzero((void *)srdp->srd_hmergnp,
13947 	    sizeof (sf_region_t *) * SFMMU_MAX_HME_REGIONS);
13948 
13949 	ASSERT(srdp->srd_scdp == NULL);
13950 	kmem_cache_free(srd_cache, srdp);
13951 }
13952 
13953 /* ARGSUSED */
13954 static int
13955 sfmmu_srdcache_constructor(void *buf, void *cdrarg, int kmflags)
13956 {
13957 	sf_srd_t *srdp = (sf_srd_t *)buf;
13958 	bzero(buf, sizeof (*srdp));
13959 
13960 	mutex_init(&srdp->srd_mutex, NULL, MUTEX_DEFAULT, NULL);
13961 	mutex_init(&srdp->srd_scd_mutex, NULL, MUTEX_DEFAULT, NULL);
13962 	return (0);
13963 }
13964 
13965 /* ARGSUSED */
13966 static void
13967 sfmmu_srdcache_destructor(void *buf, void *cdrarg)
13968 {
13969 	sf_srd_t *srdp = (sf_srd_t *)buf;
13970 
13971 	mutex_destroy(&srdp->srd_mutex);
13972 	mutex_destroy(&srdp->srd_scd_mutex);
13973 }
13974 
13975 /*
13976  * The caller makes sure hat_join_region()/hat_leave_region() can't be called
13977  * at the same time for the same process and address range. This is ensured by
13978  * the fact that address space is locked as writer when a process joins the
13979  * regions. Therefore there's no need to hold an srd lock during the entire
13980  * execution of hat_join_region()/hat_leave_region().
13981  */
13982 
13983 #define	RGN_HASH_FUNCTION(obj)	(((((uintptr_t)(obj)) >> 4) ^ \
13984 				    (((uintptr_t)(obj)) >> 11)) & \
13985 					srd_rgn_hashmask)
13986 /*
13987  * This routine implements the shared context functionality required when
13988  * attaching a segment to an address space. It must be called from
13989  * hat_share() for D(ISM) segments and from segvn_create() for segments
13990  * with the MAP_PRIVATE and MAP_TEXT flags set. It returns a region_cookie
13991  * which is saved in the private segment data for hme segments and
13992  * the ism_map structure for ism segments.
13993  */
13994 hat_region_cookie_t
13995 hat_join_region(struct hat *sfmmup,
13996 	caddr_t r_saddr,
13997 	size_t r_size,
13998 	void *r_obj,
13999 	u_offset_t r_objoff,
14000 	uchar_t r_perm,
14001 	uchar_t r_pgszc,
14002 	hat_rgn_cb_func_t r_cb_function,
14003 	uint_t flags)
14004 {
14005 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14006 	uint_t rhash;
14007 	uint_t rid;
14008 	hatlock_t *hatlockp;
14009 	sf_region_t *rgnp;
14010 	sf_region_t *new_rgnp = NULL;
14011 	int i;
14012 	uint16_t *nextidp;
14013 	sf_region_t **freelistp;
14014 	int maxids;
14015 	sf_region_t **rarrp;
14016 	uint16_t *busyrgnsp;
14017 	ulong_t rttecnt;
14018 	uchar_t tteflag;
14019 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
14020 	int text = (r_type == HAT_REGION_TEXT);
14021 
14022 	if (srdp == NULL || r_size == 0) {
14023 		return (HAT_INVALID_REGION_COOKIE);
14024 	}
14025 
14026 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
14027 	ASSERT(sfmmup != ksfmmup);
14028 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as));
14029 	ASSERT(srdp->srd_refcnt > 0);
14030 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
14031 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
14032 	ASSERT(r_pgszc < mmu_page_sizes);
14033 	if (!IS_P2ALIGNED(r_saddr, TTEBYTES(r_pgszc)) ||
14034 	    !IS_P2ALIGNED(r_size, TTEBYTES(r_pgszc))) {
14035 		panic("hat_join_region: region addr or size is not aligned\n");
14036 	}
14037 
14038 
14039 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
14040 	    SFMMU_REGION_HME;
14041 	/*
14042 	 * Currently only support shared hmes for the read only main text
14043 	 * region.
14044 	 */
14045 	if (r_type == SFMMU_REGION_HME && ((r_obj != srdp->srd_evp) ||
14046 	    (r_perm & PROT_WRITE))) {
14047 		return (HAT_INVALID_REGION_COOKIE);
14048 	}
14049 
14050 	rhash = RGN_HASH_FUNCTION(r_obj);
14051 
14052 	if (r_type == SFMMU_REGION_ISM) {
14053 		nextidp = &srdp->srd_next_ismrid;
14054 		freelistp = &srdp->srd_ismrgnfree;
14055 		maxids = SFMMU_MAX_ISM_REGIONS;
14056 		rarrp = srdp->srd_ismrgnp;
14057 		busyrgnsp = &srdp->srd_ismbusyrgns;
14058 	} else {
14059 		nextidp = &srdp->srd_next_hmerid;
14060 		freelistp = &srdp->srd_hmergnfree;
14061 		maxids = SFMMU_MAX_HME_REGIONS;
14062 		rarrp = srdp->srd_hmergnp;
14063 		busyrgnsp = &srdp->srd_hmebusyrgns;
14064 	}
14065 
14066 	mutex_enter(&srdp->srd_mutex);
14067 
14068 	for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
14069 	    rgnp = rgnp->rgn_hash) {
14070 		if (rgnp->rgn_saddr == r_saddr && rgnp->rgn_size == r_size &&
14071 		    rgnp->rgn_obj == r_obj && rgnp->rgn_objoff == r_objoff &&
14072 		    rgnp->rgn_perm == r_perm && rgnp->rgn_pgszc == r_pgszc) {
14073 			break;
14074 		}
14075 	}
14076 
14077 rfound:
14078 	if (rgnp != NULL) {
14079 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14080 		ASSERT(rgnp->rgn_cb_function == r_cb_function);
14081 		ASSERT(rgnp->rgn_refcnt >= 0);
14082 		rid = rgnp->rgn_id;
14083 		ASSERT(rid < maxids);
14084 		ASSERT(rarrp[rid] == rgnp);
14085 		ASSERT(rid < *nextidp);
14086 		atomic_inc_32((volatile uint_t *)&rgnp->rgn_refcnt);
14087 		mutex_exit(&srdp->srd_mutex);
14088 		if (new_rgnp != NULL) {
14089 			kmem_cache_free(region_cache, new_rgnp);
14090 		}
14091 		if (r_type == SFMMU_REGION_HME) {
14092 			int myjoin =
14093 			    (sfmmup == astosfmmu(curthread->t_procp->p_as));
14094 
14095 			sfmmu_link_to_hmeregion(sfmmup, rgnp);
14096 			/*
14097 			 * bitmap should be updated after linking sfmmu on
14098 			 * region list so that pageunload() doesn't skip
14099 			 * TSB/TLB flush. As soon as bitmap is updated another
14100 			 * thread in this process can already start accessing
14101 			 * this region.
14102 			 */
14103 			/*
14104 			 * Normally ttecnt accounting is done as part of
14105 			 * pagefault handling. But a process may not take any
14106 			 * pagefaults on shared hmeblks created by some other
14107 			 * process. To compensate for this assume that the
14108 			 * entire region will end up faulted in using
14109 			 * the region's pagesize.
14110 			 *
14111 			 */
14112 			if (r_pgszc > TTE8K) {
14113 				tteflag = 1 << r_pgszc;
14114 				if (disable_large_pages & tteflag) {
14115 					tteflag = 0;
14116 				}
14117 			} else {
14118 				tteflag = 0;
14119 			}
14120 			if (tteflag && !(sfmmup->sfmmu_rtteflags & tteflag)) {
14121 				hatlockp = sfmmu_hat_enter(sfmmup);
14122 				sfmmup->sfmmu_rtteflags |= tteflag;
14123 				sfmmu_hat_exit(hatlockp);
14124 			}
14125 			hatlockp = sfmmu_hat_enter(sfmmup);
14126 
14127 			/*
14128 			 * Preallocate 1/4 of ttecnt's in 8K TSB for >= 4M
14129 			 * region to allow for large page allocation failure.
14130 			 */
14131 			if (r_pgszc >= TTE4M) {
14132 				sfmmup->sfmmu_tsb0_4minflcnt +=
14133 				    r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14134 			}
14135 
14136 			/* update sfmmu_ttecnt with the shme rgn ttecnt */
14137 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14138 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14139 			    rttecnt);
14140 
14141 			if (text && r_pgszc >= TTE4M &&
14142 			    (tteflag || ((disable_large_pages >> TTE4M) &
14143 			    ((1 << (r_pgszc - TTE4M + 1)) - 1))) &&
14144 			    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
14145 				SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
14146 			}
14147 
14148 			sfmmu_hat_exit(hatlockp);
14149 			/*
14150 			 * On Panther we need to make sure TLB is programmed
14151 			 * to accept 32M/256M pages.  Call
14152 			 * sfmmu_check_page_sizes() now to make sure TLB is
14153 			 * setup before making hmeregions visible to other
14154 			 * threads.
14155 			 */
14156 			sfmmu_check_page_sizes(sfmmup, 1);
14157 			hatlockp = sfmmu_hat_enter(sfmmup);
14158 			SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14159 
14160 			/*
14161 			 * if context is invalid tsb miss exception code will
14162 			 * call sfmmu_check_page_sizes() and update tsbmiss
14163 			 * area later.
14164 			 */
14165 			kpreempt_disable();
14166 			if (myjoin &&
14167 			    (sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
14168 			    != INVALID_CONTEXT)) {
14169 				struct tsbmiss *tsbmp;
14170 
14171 				tsbmp = &tsbmiss_area[CPU->cpu_id];
14172 				ASSERT(sfmmup == tsbmp->usfmmup);
14173 				BT_SET(tsbmp->shmermap, rid);
14174 				if (r_pgszc > TTE64K) {
14175 					tsbmp->uhat_rtteflags |= tteflag;
14176 				}
14177 
14178 			}
14179 			kpreempt_enable();
14180 
14181 			sfmmu_hat_exit(hatlockp);
14182 			ASSERT((hat_region_cookie_t)((uint64_t)rid) !=
14183 			    HAT_INVALID_REGION_COOKIE);
14184 		} else {
14185 			hatlockp = sfmmu_hat_enter(sfmmup);
14186 			SF_RGNMAP_ADD(sfmmup->sfmmu_ismregion_map, rid);
14187 			sfmmu_hat_exit(hatlockp);
14188 		}
14189 		ASSERT(rid < maxids);
14190 
14191 		if (r_type == SFMMU_REGION_ISM) {
14192 			sfmmu_find_scd(sfmmup);
14193 		}
14194 		return ((hat_region_cookie_t)((uint64_t)rid));
14195 	}
14196 
14197 	ASSERT(new_rgnp == NULL);
14198 
14199 	if (*busyrgnsp >= maxids) {
14200 		mutex_exit(&srdp->srd_mutex);
14201 		return (HAT_INVALID_REGION_COOKIE);
14202 	}
14203 
14204 	ASSERT(MUTEX_HELD(&srdp->srd_mutex));
14205 	if (*freelistp != NULL) {
14206 		rgnp = *freelistp;
14207 		*freelistp = rgnp->rgn_next;
14208 		ASSERT(rgnp->rgn_id < *nextidp);
14209 		ASSERT(rgnp->rgn_id < maxids);
14210 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
14211 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK)
14212 		    == r_type);
14213 		ASSERT(rarrp[rgnp->rgn_id] == rgnp);
14214 		ASSERT(rgnp->rgn_hmeflags == 0);
14215 	} else {
14216 		/*
14217 		 * release local locks before memory allocation.
14218 		 */
14219 		mutex_exit(&srdp->srd_mutex);
14220 
14221 		new_rgnp = kmem_cache_alloc(region_cache, KM_SLEEP);
14222 
14223 		mutex_enter(&srdp->srd_mutex);
14224 		for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
14225 		    rgnp = rgnp->rgn_hash) {
14226 			if (rgnp->rgn_saddr == r_saddr &&
14227 			    rgnp->rgn_size == r_size &&
14228 			    rgnp->rgn_obj == r_obj &&
14229 			    rgnp->rgn_objoff == r_objoff &&
14230 			    rgnp->rgn_perm == r_perm &&
14231 			    rgnp->rgn_pgszc == r_pgszc) {
14232 				break;
14233 			}
14234 		}
14235 		if (rgnp != NULL) {
14236 			goto rfound;
14237 		}
14238 
14239 		if (*nextidp >= maxids) {
14240 			mutex_exit(&srdp->srd_mutex);
14241 			goto fail;
14242 		}
14243 		rgnp = new_rgnp;
14244 		new_rgnp = NULL;
14245 		rgnp->rgn_id = (*nextidp)++;
14246 		ASSERT(rgnp->rgn_id < maxids);
14247 		ASSERT(rarrp[rgnp->rgn_id] == NULL);
14248 		rarrp[rgnp->rgn_id] = rgnp;
14249 	}
14250 
14251 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14252 	ASSERT(rgnp->rgn_hmeflags == 0);
14253 #ifdef DEBUG
14254 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14255 		ASSERT(rgnp->rgn_ttecnt[i] == 0);
14256 	}
14257 #endif
14258 	rgnp->rgn_saddr = r_saddr;
14259 	rgnp->rgn_size = r_size;
14260 	rgnp->rgn_obj = r_obj;
14261 	rgnp->rgn_objoff = r_objoff;
14262 	rgnp->rgn_perm = r_perm;
14263 	rgnp->rgn_pgszc = r_pgszc;
14264 	rgnp->rgn_flags = r_type;
14265 	rgnp->rgn_refcnt = 0;
14266 	rgnp->rgn_cb_function = r_cb_function;
14267 	rgnp->rgn_hash = srdp->srd_rgnhash[rhash];
14268 	srdp->srd_rgnhash[rhash] = rgnp;
14269 	(*busyrgnsp)++;
14270 	ASSERT(*busyrgnsp <= maxids);
14271 	goto rfound;
14272 
14273 fail:
14274 	ASSERT(new_rgnp != NULL);
14275 	kmem_cache_free(region_cache, new_rgnp);
14276 	return (HAT_INVALID_REGION_COOKIE);
14277 }
14278 
14279 /*
14280  * This function implements the shared context functionality required
14281  * when detaching a segment from an address space. It must be called
14282  * from hat_unshare() for all D(ISM) segments and from segvn_unmap(),
14283  * for segments with a valid region_cookie.
14284  * It will also be called from all seg_vn routines which change a
14285  * segment's attributes such as segvn_setprot(), segvn_setpagesize(),
14286  * segvn_clrszc() & segvn_advise(), as well as in the case of COW fault
14287  * from segvn_fault().
14288  */
14289 void
14290 hat_leave_region(struct hat *sfmmup, hat_region_cookie_t rcookie, uint_t flags)
14291 {
14292 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14293 	sf_scd_t *scdp;
14294 	uint_t rhash;
14295 	uint_t rid = (uint_t)((uint64_t)rcookie);
14296 	hatlock_t *hatlockp = NULL;
14297 	sf_region_t *rgnp;
14298 	sf_region_t **prev_rgnpp;
14299 	sf_region_t *cur_rgnp;
14300 	void *r_obj;
14301 	int i;
14302 	caddr_t	r_saddr;
14303 	caddr_t r_eaddr;
14304 	size_t	r_size;
14305 	uchar_t	r_pgszc;
14306 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
14307 
14308 	ASSERT(sfmmup != ksfmmup);
14309 	ASSERT(srdp != NULL);
14310 	ASSERT(srdp->srd_refcnt > 0);
14311 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
14312 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
14313 	ASSERT(!sfmmup->sfmmu_free || sfmmup->sfmmu_scdp == NULL);
14314 
14315 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
14316 	    SFMMU_REGION_HME;
14317 
14318 	if (r_type == SFMMU_REGION_ISM) {
14319 		ASSERT(SFMMU_IS_ISMRID_VALID(rid));
14320 		ASSERT(rid < SFMMU_MAX_ISM_REGIONS);
14321 		rgnp = srdp->srd_ismrgnp[rid];
14322 	} else {
14323 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14324 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14325 		rgnp = srdp->srd_hmergnp[rid];
14326 	}
14327 	ASSERT(rgnp != NULL);
14328 	ASSERT(rgnp->rgn_id == rid);
14329 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14330 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14331 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as));
14332 
14333 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
14334 	if (r_type == SFMMU_REGION_HME && sfmmup->sfmmu_as->a_xhat != NULL) {
14335 		xhat_unload_callback_all(sfmmup->sfmmu_as, rgnp->rgn_saddr,
14336 		    rgnp->rgn_size, 0, NULL);
14337 	}
14338 
14339 	if (sfmmup->sfmmu_free) {
14340 		ulong_t rttecnt;
14341 		r_pgszc = rgnp->rgn_pgszc;
14342 		r_size = rgnp->rgn_size;
14343 
14344 		ASSERT(sfmmup->sfmmu_scdp == NULL);
14345 		if (r_type == SFMMU_REGION_ISM) {
14346 			SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14347 		} else {
14348 			/* update shme rgns ttecnt in sfmmu_ttecnt */
14349 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14350 			ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14351 
14352 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14353 			    -rttecnt);
14354 
14355 			SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14356 		}
14357 	} else if (r_type == SFMMU_REGION_ISM) {
14358 		hatlockp = sfmmu_hat_enter(sfmmup);
14359 		ASSERT(rid < srdp->srd_next_ismrid);
14360 		SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14361 		scdp = sfmmup->sfmmu_scdp;
14362 		if (scdp != NULL &&
14363 		    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
14364 			sfmmu_leave_scd(sfmmup, r_type);
14365 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14366 		}
14367 		sfmmu_hat_exit(hatlockp);
14368 	} else {
14369 		ulong_t rttecnt;
14370 		r_pgszc = rgnp->rgn_pgszc;
14371 		r_saddr = rgnp->rgn_saddr;
14372 		r_size = rgnp->rgn_size;
14373 		r_eaddr = r_saddr + r_size;
14374 
14375 		ASSERT(r_type == SFMMU_REGION_HME);
14376 		hatlockp = sfmmu_hat_enter(sfmmup);
14377 		ASSERT(rid < srdp->srd_next_hmerid);
14378 		SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14379 
14380 		/*
14381 		 * If region is part of an SCD call sfmmu_leave_scd().
14382 		 * Otherwise if process is not exiting and has valid context
14383 		 * just drop the context on the floor to lose stale TLB
14384 		 * entries and force the update of tsb miss area to reflect
14385 		 * the new region map. After that clean our TSB entries.
14386 		 */
14387 		scdp = sfmmup->sfmmu_scdp;
14388 		if (scdp != NULL &&
14389 		    SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
14390 			sfmmu_leave_scd(sfmmup, r_type);
14391 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14392 		}
14393 		sfmmu_invalidate_ctx(sfmmup);
14394 
14395 		i = TTE8K;
14396 		while (i < mmu_page_sizes) {
14397 			if (rgnp->rgn_ttecnt[i] != 0) {
14398 				sfmmu_unload_tsb_range(sfmmup, r_saddr,
14399 				    r_eaddr, i);
14400 				if (i < TTE4M) {
14401 					i = TTE4M;
14402 					continue;
14403 				} else {
14404 					break;
14405 				}
14406 			}
14407 			i++;
14408 		}
14409 		/* Remove the preallocated 1/4 8k ttecnt for 4M regions. */
14410 		if (r_pgszc >= TTE4M) {
14411 			rttecnt = r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14412 			ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
14413 			    rttecnt);
14414 			sfmmup->sfmmu_tsb0_4minflcnt -= rttecnt;
14415 		}
14416 
14417 		/* update shme rgns ttecnt in sfmmu_ttecnt */
14418 		rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14419 		ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14420 		atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc], -rttecnt);
14421 
14422 		sfmmu_hat_exit(hatlockp);
14423 		if (scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
14424 			/* sfmmup left the scd, grow private tsb */
14425 			sfmmu_check_page_sizes(sfmmup, 1);
14426 		} else {
14427 			sfmmu_check_page_sizes(sfmmup, 0);
14428 		}
14429 	}
14430 
14431 	if (r_type == SFMMU_REGION_HME) {
14432 		sfmmu_unlink_from_hmeregion(sfmmup, rgnp);
14433 	}
14434 
14435 	r_obj = rgnp->rgn_obj;
14436 	if (atomic_dec_32_nv((volatile uint_t *)&rgnp->rgn_refcnt)) {
14437 		return;
14438 	}
14439 
14440 	/*
14441 	 * looks like nobody uses this region anymore. Free it.
14442 	 */
14443 	rhash = RGN_HASH_FUNCTION(r_obj);
14444 	mutex_enter(&srdp->srd_mutex);
14445 	for (prev_rgnpp = &srdp->srd_rgnhash[rhash];
14446 	    (cur_rgnp = *prev_rgnpp) != NULL;
14447 	    prev_rgnpp = &cur_rgnp->rgn_hash) {
14448 		if (cur_rgnp == rgnp && cur_rgnp->rgn_refcnt == 0) {
14449 			break;
14450 		}
14451 	}
14452 
14453 	if (cur_rgnp == NULL) {
14454 		mutex_exit(&srdp->srd_mutex);
14455 		return;
14456 	}
14457 
14458 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14459 	*prev_rgnpp = rgnp->rgn_hash;
14460 	if (r_type == SFMMU_REGION_ISM) {
14461 		rgnp->rgn_flags |= SFMMU_REGION_FREE;
14462 		ASSERT(rid < srdp->srd_next_ismrid);
14463 		rgnp->rgn_next = srdp->srd_ismrgnfree;
14464 		srdp->srd_ismrgnfree = rgnp;
14465 		ASSERT(srdp->srd_ismbusyrgns > 0);
14466 		srdp->srd_ismbusyrgns--;
14467 		mutex_exit(&srdp->srd_mutex);
14468 		return;
14469 	}
14470 	mutex_exit(&srdp->srd_mutex);
14471 
14472 	/*
14473 	 * Destroy region's hmeblks.
14474 	 */
14475 	sfmmu_unload_hmeregion(srdp, rgnp);
14476 
14477 	rgnp->rgn_hmeflags = 0;
14478 
14479 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14480 	ASSERT(rgnp->rgn_id == rid);
14481 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14482 		rgnp->rgn_ttecnt[i] = 0;
14483 	}
14484 	rgnp->rgn_flags |= SFMMU_REGION_FREE;
14485 	mutex_enter(&srdp->srd_mutex);
14486 	ASSERT(rid < srdp->srd_next_hmerid);
14487 	rgnp->rgn_next = srdp->srd_hmergnfree;
14488 	srdp->srd_hmergnfree = rgnp;
14489 	ASSERT(srdp->srd_hmebusyrgns > 0);
14490 	srdp->srd_hmebusyrgns--;
14491 	mutex_exit(&srdp->srd_mutex);
14492 }
14493 
14494 /*
14495  * For now only called for hmeblk regions and not for ISM regions.
14496  */
14497 void
14498 hat_dup_region(struct hat *sfmmup, hat_region_cookie_t rcookie)
14499 {
14500 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14501 	uint_t rid = (uint_t)((uint64_t)rcookie);
14502 	sf_region_t *rgnp;
14503 	sf_rgn_link_t *rlink;
14504 	sf_rgn_link_t *hrlink;
14505 	ulong_t	rttecnt;
14506 
14507 	ASSERT(sfmmup != ksfmmup);
14508 	ASSERT(srdp != NULL);
14509 	ASSERT(srdp->srd_refcnt > 0);
14510 
14511 	ASSERT(rid < srdp->srd_next_hmerid);
14512 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14513 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14514 
14515 	rgnp = srdp->srd_hmergnp[rid];
14516 	ASSERT(rgnp->rgn_refcnt > 0);
14517 	ASSERT(rgnp->rgn_id == rid);
14518 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == SFMMU_REGION_HME);
14519 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14520 
14521 	atomic_inc_32((volatile uint_t *)&rgnp->rgn_refcnt);
14522 
14523 	/* LINTED: constant in conditional context */
14524 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 0);
14525 	ASSERT(rlink != NULL);
14526 	mutex_enter(&rgnp->rgn_mutex);
14527 	ASSERT(rgnp->rgn_sfmmu_head != NULL);
14528 	/* LINTED: constant in conditional context */
14529 	SFMMU_HMERID2RLINKP(rgnp->rgn_sfmmu_head, rid, hrlink, 0, 0);
14530 	ASSERT(hrlink != NULL);
14531 	ASSERT(hrlink->prev == NULL);
14532 	rlink->next = rgnp->rgn_sfmmu_head;
14533 	rlink->prev = NULL;
14534 	hrlink->prev = sfmmup;
14535 	/*
14536 	 * make sure rlink's next field is correct
14537 	 * before making this link visible.
14538 	 */
14539 	membar_stst();
14540 	rgnp->rgn_sfmmu_head = sfmmup;
14541 	mutex_exit(&rgnp->rgn_mutex);
14542 
14543 	/* update sfmmu_ttecnt with the shme rgn ttecnt */
14544 	rttecnt = rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
14545 	atomic_add_long(&sfmmup->sfmmu_ttecnt[rgnp->rgn_pgszc], rttecnt);
14546 	/* update tsb0 inflation count */
14547 	if (rgnp->rgn_pgszc >= TTE4M) {
14548 		sfmmup->sfmmu_tsb0_4minflcnt +=
14549 		    rgnp->rgn_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14550 	}
14551 	/*
14552 	 * Update regionid bitmask without hat lock since no other thread
14553 	 * can update this region bitmask right now.
14554 	 */
14555 	SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14556 }
14557 
14558 /* ARGSUSED */
14559 static int
14560 sfmmu_rgncache_constructor(void *buf, void *cdrarg, int kmflags)
14561 {
14562 	sf_region_t *rgnp = (sf_region_t *)buf;
14563 	bzero(buf, sizeof (*rgnp));
14564 
14565 	mutex_init(&rgnp->rgn_mutex, NULL, MUTEX_DEFAULT, NULL);
14566 
14567 	return (0);
14568 }
14569 
14570 /* ARGSUSED */
14571 static void
14572 sfmmu_rgncache_destructor(void *buf, void *cdrarg)
14573 {
14574 	sf_region_t *rgnp = (sf_region_t *)buf;
14575 	mutex_destroy(&rgnp->rgn_mutex);
14576 }
14577 
14578 static int
14579 sfrgnmap_isnull(sf_region_map_t *map)
14580 {
14581 	int i;
14582 
14583 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14584 		if (map->bitmap[i] != 0) {
14585 			return (0);
14586 		}
14587 	}
14588 	return (1);
14589 }
14590 
14591 static int
14592 sfhmergnmap_isnull(sf_hmeregion_map_t *map)
14593 {
14594 	int i;
14595 
14596 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
14597 		if (map->bitmap[i] != 0) {
14598 			return (0);
14599 		}
14600 	}
14601 	return (1);
14602 }
14603 
14604 #ifdef DEBUG
14605 static void
14606 check_scd_sfmmu_list(sfmmu_t **headp, sfmmu_t *sfmmup, int onlist)
14607 {
14608 	sfmmu_t *sp;
14609 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14610 
14611 	for (sp = *headp; sp != NULL; sp = sp->sfmmu_scd_link.next) {
14612 		ASSERT(srdp == sp->sfmmu_srdp);
14613 		if (sp == sfmmup) {
14614 			if (onlist) {
14615 				return;
14616 			} else {
14617 				panic("shctx: sfmmu 0x%p found on scd"
14618 				    "list 0x%p", (void *)sfmmup,
14619 				    (void *)*headp);
14620 			}
14621 		}
14622 	}
14623 	if (onlist) {
14624 		panic("shctx: sfmmu 0x%p not found on scd list 0x%p",
14625 		    (void *)sfmmup, (void *)*headp);
14626 	} else {
14627 		return;
14628 	}
14629 }
14630 #else /* DEBUG */
14631 #define	check_scd_sfmmu_list(headp, sfmmup, onlist)
14632 #endif /* DEBUG */
14633 
14634 /*
14635  * Removes an sfmmu from the SCD sfmmu list.
14636  */
14637 static void
14638 sfmmu_from_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14639 {
14640 	ASSERT(sfmmup->sfmmu_srdp != NULL);
14641 	check_scd_sfmmu_list(headp, sfmmup, 1);
14642 	if (sfmmup->sfmmu_scd_link.prev != NULL) {
14643 		ASSERT(*headp != sfmmup);
14644 		sfmmup->sfmmu_scd_link.prev->sfmmu_scd_link.next =
14645 		    sfmmup->sfmmu_scd_link.next;
14646 	} else {
14647 		ASSERT(*headp == sfmmup);
14648 		*headp = sfmmup->sfmmu_scd_link.next;
14649 	}
14650 	if (sfmmup->sfmmu_scd_link.next != NULL) {
14651 		sfmmup->sfmmu_scd_link.next->sfmmu_scd_link.prev =
14652 		    sfmmup->sfmmu_scd_link.prev;
14653 	}
14654 }
14655 
14656 
14657 /*
14658  * Adds an sfmmu to the start of the queue.
14659  */
14660 static void
14661 sfmmu_to_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14662 {
14663 	check_scd_sfmmu_list(headp, sfmmup, 0);
14664 	sfmmup->sfmmu_scd_link.prev = NULL;
14665 	sfmmup->sfmmu_scd_link.next = *headp;
14666 	if (*headp != NULL)
14667 		(*headp)->sfmmu_scd_link.prev = sfmmup;
14668 	*headp = sfmmup;
14669 }
14670 
14671 /*
14672  * Remove an scd from the start of the queue.
14673  */
14674 static void
14675 sfmmu_remove_scd(sf_scd_t **headp, sf_scd_t *scdp)
14676 {
14677 	if (scdp->scd_prev != NULL) {
14678 		ASSERT(*headp != scdp);
14679 		scdp->scd_prev->scd_next = scdp->scd_next;
14680 	} else {
14681 		ASSERT(*headp == scdp);
14682 		*headp = scdp->scd_next;
14683 	}
14684 
14685 	if (scdp->scd_next != NULL) {
14686 		scdp->scd_next->scd_prev = scdp->scd_prev;
14687 	}
14688 }
14689 
14690 /*
14691  * Add an scd to the start of the queue.
14692  */
14693 static void
14694 sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *scdp)
14695 {
14696 	scdp->scd_prev = NULL;
14697 	scdp->scd_next = *headp;
14698 	if (*headp != NULL) {
14699 		(*headp)->scd_prev = scdp;
14700 	}
14701 	*headp = scdp;
14702 }
14703 
14704 static int
14705 sfmmu_alloc_scd_tsbs(sf_srd_t *srdp, sf_scd_t *scdp)
14706 {
14707 	uint_t rid;
14708 	uint_t i;
14709 	uint_t j;
14710 	ulong_t w;
14711 	sf_region_t *rgnp;
14712 	ulong_t tte8k_cnt = 0;
14713 	ulong_t tte4m_cnt = 0;
14714 	uint_t tsb_szc;
14715 	sfmmu_t *scsfmmup = scdp->scd_sfmmup;
14716 	sfmmu_t	*ism_hatid;
14717 	struct tsb_info *newtsb;
14718 	int szc;
14719 
14720 	ASSERT(srdp != NULL);
14721 
14722 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14723 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14724 			continue;
14725 		}
14726 		j = 0;
14727 		while (w) {
14728 			if (!(w & 0x1)) {
14729 				j++;
14730 				w >>= 1;
14731 				continue;
14732 			}
14733 			rid = (i << BT_ULSHIFT) | j;
14734 			j++;
14735 			w >>= 1;
14736 
14737 			if (rid < SFMMU_MAX_HME_REGIONS) {
14738 				rgnp = srdp->srd_hmergnp[rid];
14739 				ASSERT(rgnp->rgn_id == rid);
14740 				ASSERT(rgnp->rgn_refcnt > 0);
14741 
14742 				if (rgnp->rgn_pgszc < TTE4M) {
14743 					tte8k_cnt += rgnp->rgn_size >>
14744 					    TTE_PAGE_SHIFT(TTE8K);
14745 				} else {
14746 					ASSERT(rgnp->rgn_pgszc >= TTE4M);
14747 					tte4m_cnt += rgnp->rgn_size >>
14748 					    TTE_PAGE_SHIFT(TTE4M);
14749 					/*
14750 					 * Inflate SCD tsb0 by preallocating
14751 					 * 1/4 8k ttecnt for 4M regions to
14752 					 * allow for lgpg alloc failure.
14753 					 */
14754 					tte8k_cnt += rgnp->rgn_size >>
14755 					    (TTE_PAGE_SHIFT(TTE8K) + 2);
14756 				}
14757 			} else {
14758 				rid -= SFMMU_MAX_HME_REGIONS;
14759 				rgnp = srdp->srd_ismrgnp[rid];
14760 				ASSERT(rgnp->rgn_id == rid);
14761 				ASSERT(rgnp->rgn_refcnt > 0);
14762 
14763 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14764 				ASSERT(ism_hatid->sfmmu_ismhat);
14765 
14766 				for (szc = 0; szc < TTE4M; szc++) {
14767 					tte8k_cnt +=
14768 					    ism_hatid->sfmmu_ttecnt[szc] <<
14769 					    TTE_BSZS_SHIFT(szc);
14770 				}
14771 
14772 				ASSERT(rgnp->rgn_pgszc >= TTE4M);
14773 				if (rgnp->rgn_pgszc >= TTE4M) {
14774 					tte4m_cnt += rgnp->rgn_size >>
14775 					    TTE_PAGE_SHIFT(TTE4M);
14776 				}
14777 			}
14778 		}
14779 	}
14780 
14781 	tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
14782 
14783 	/* Allocate both the SCD TSBs here. */
14784 	if (sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14785 	    tsb_szc, TSB8K|TSB64K|TSB512K, TSB_ALLOC, scsfmmup) &&
14786 	    (tsb_szc <= TSB_4M_SZCODE ||
14787 	    sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14788 	    TSB_4M_SZCODE, TSB8K|TSB64K|TSB512K,
14789 	    TSB_ALLOC, scsfmmup))) {
14790 
14791 		SFMMU_STAT(sf_scd_1sttsb_allocfail);
14792 		return (TSB_ALLOCFAIL);
14793 	} else {
14794 		scsfmmup->sfmmu_tsb->tsb_flags |= TSB_SHAREDCTX;
14795 
14796 		if (tte4m_cnt) {
14797 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
14798 			if (sfmmu_tsbinfo_alloc(&newtsb, tsb_szc,
14799 			    TSB4M|TSB32M|TSB256M, TSB_ALLOC, scsfmmup) &&
14800 			    (tsb_szc <= TSB_4M_SZCODE ||
14801 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
14802 			    TSB4M|TSB32M|TSB256M,
14803 			    TSB_ALLOC, scsfmmup))) {
14804 				/*
14805 				 * If we fail to allocate the 2nd shared tsb,
14806 				 * just free the 1st tsb, return failure.
14807 				 */
14808 				sfmmu_tsbinfo_free(scsfmmup->sfmmu_tsb);
14809 				SFMMU_STAT(sf_scd_2ndtsb_allocfail);
14810 				return (TSB_ALLOCFAIL);
14811 			} else {
14812 				ASSERT(scsfmmup->sfmmu_tsb->tsb_next == NULL);
14813 				newtsb->tsb_flags |= TSB_SHAREDCTX;
14814 				scsfmmup->sfmmu_tsb->tsb_next = newtsb;
14815 				SFMMU_STAT(sf_scd_2ndtsb_alloc);
14816 			}
14817 		}
14818 		SFMMU_STAT(sf_scd_1sttsb_alloc);
14819 	}
14820 	return (TSB_SUCCESS);
14821 }
14822 
14823 static void
14824 sfmmu_free_scd_tsbs(sfmmu_t *scd_sfmmu)
14825 {
14826 	while (scd_sfmmu->sfmmu_tsb != NULL) {
14827 		struct tsb_info *next = scd_sfmmu->sfmmu_tsb->tsb_next;
14828 		sfmmu_tsbinfo_free(scd_sfmmu->sfmmu_tsb);
14829 		scd_sfmmu->sfmmu_tsb = next;
14830 	}
14831 }
14832 
14833 /*
14834  * Link the sfmmu onto the hme region list.
14835  */
14836 void
14837 sfmmu_link_to_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14838 {
14839 	uint_t rid;
14840 	sf_rgn_link_t *rlink;
14841 	sfmmu_t *head;
14842 	sf_rgn_link_t *hrlink;
14843 
14844 	rid = rgnp->rgn_id;
14845 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14846 
14847 	/* LINTED: constant in conditional context */
14848 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 1);
14849 	ASSERT(rlink != NULL);
14850 	mutex_enter(&rgnp->rgn_mutex);
14851 	if ((head = rgnp->rgn_sfmmu_head) == NULL) {
14852 		rlink->next = NULL;
14853 		rlink->prev = NULL;
14854 		/*
14855 		 * make sure rlink's next field is NULL
14856 		 * before making this link visible.
14857 		 */
14858 		membar_stst();
14859 		rgnp->rgn_sfmmu_head = sfmmup;
14860 	} else {
14861 		/* LINTED: constant in conditional context */
14862 		SFMMU_HMERID2RLINKP(head, rid, hrlink, 0, 0);
14863 		ASSERT(hrlink != NULL);
14864 		ASSERT(hrlink->prev == NULL);
14865 		rlink->next = head;
14866 		rlink->prev = NULL;
14867 		hrlink->prev = sfmmup;
14868 		/*
14869 		 * make sure rlink's next field is correct
14870 		 * before making this link visible.
14871 		 */
14872 		membar_stst();
14873 		rgnp->rgn_sfmmu_head = sfmmup;
14874 	}
14875 	mutex_exit(&rgnp->rgn_mutex);
14876 }
14877 
14878 /*
14879  * Unlink the sfmmu from the hme region list.
14880  */
14881 void
14882 sfmmu_unlink_from_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14883 {
14884 	uint_t rid;
14885 	sf_rgn_link_t *rlink;
14886 
14887 	rid = rgnp->rgn_id;
14888 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14889 
14890 	/* LINTED: constant in conditional context */
14891 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
14892 	ASSERT(rlink != NULL);
14893 	mutex_enter(&rgnp->rgn_mutex);
14894 	if (rgnp->rgn_sfmmu_head == sfmmup) {
14895 		sfmmu_t *next = rlink->next;
14896 		rgnp->rgn_sfmmu_head = next;
14897 		/*
14898 		 * if we are stopped by xc_attention() after this
14899 		 * point the forward link walking in
14900 		 * sfmmu_rgntlb_demap() will work correctly since the
14901 		 * head correctly points to the next element.
14902 		 */
14903 		membar_stst();
14904 		rlink->next = NULL;
14905 		ASSERT(rlink->prev == NULL);
14906 		if (next != NULL) {
14907 			sf_rgn_link_t *nrlink;
14908 			/* LINTED: constant in conditional context */
14909 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14910 			ASSERT(nrlink != NULL);
14911 			ASSERT(nrlink->prev == sfmmup);
14912 			nrlink->prev = NULL;
14913 		}
14914 	} else {
14915 		sfmmu_t *next = rlink->next;
14916 		sfmmu_t *prev = rlink->prev;
14917 		sf_rgn_link_t *prlink;
14918 
14919 		ASSERT(prev != NULL);
14920 		/* LINTED: constant in conditional context */
14921 		SFMMU_HMERID2RLINKP(prev, rid, prlink, 0, 0);
14922 		ASSERT(prlink != NULL);
14923 		ASSERT(prlink->next == sfmmup);
14924 		prlink->next = next;
14925 		/*
14926 		 * if we are stopped by xc_attention()
14927 		 * after this point the forward link walking
14928 		 * will work correctly since the prev element
14929 		 * correctly points to the next element.
14930 		 */
14931 		membar_stst();
14932 		rlink->next = NULL;
14933 		rlink->prev = NULL;
14934 		if (next != NULL) {
14935 			sf_rgn_link_t *nrlink;
14936 			/* LINTED: constant in conditional context */
14937 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14938 			ASSERT(nrlink != NULL);
14939 			ASSERT(nrlink->prev == sfmmup);
14940 			nrlink->prev = prev;
14941 		}
14942 	}
14943 	mutex_exit(&rgnp->rgn_mutex);
14944 }
14945 
14946 /*
14947  * Link scd sfmmu onto ism or hme region list for each region in the
14948  * scd region map.
14949  */
14950 void
14951 sfmmu_link_scd_to_regions(sf_srd_t *srdp, sf_scd_t *scdp)
14952 {
14953 	uint_t rid;
14954 	uint_t i;
14955 	uint_t j;
14956 	ulong_t w;
14957 	sf_region_t *rgnp;
14958 	sfmmu_t *scsfmmup;
14959 
14960 	scsfmmup = scdp->scd_sfmmup;
14961 	ASSERT(scsfmmup->sfmmu_scdhat);
14962 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14963 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14964 			continue;
14965 		}
14966 		j = 0;
14967 		while (w) {
14968 			if (!(w & 0x1)) {
14969 				j++;
14970 				w >>= 1;
14971 				continue;
14972 			}
14973 			rid = (i << BT_ULSHIFT) | j;
14974 			j++;
14975 			w >>= 1;
14976 
14977 			if (rid < SFMMU_MAX_HME_REGIONS) {
14978 				rgnp = srdp->srd_hmergnp[rid];
14979 				ASSERT(rgnp->rgn_id == rid);
14980 				ASSERT(rgnp->rgn_refcnt > 0);
14981 				sfmmu_link_to_hmeregion(scsfmmup, rgnp);
14982 			} else {
14983 				sfmmu_t *ism_hatid = NULL;
14984 				ism_ment_t *ism_ment;
14985 				rid -= SFMMU_MAX_HME_REGIONS;
14986 				rgnp = srdp->srd_ismrgnp[rid];
14987 				ASSERT(rgnp->rgn_id == rid);
14988 				ASSERT(rgnp->rgn_refcnt > 0);
14989 
14990 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14991 				ASSERT(ism_hatid->sfmmu_ismhat);
14992 				ism_ment = &scdp->scd_ism_links[rid];
14993 				ism_ment->iment_hat = scsfmmup;
14994 				ism_ment->iment_base_va = rgnp->rgn_saddr;
14995 				mutex_enter(&ism_mlist_lock);
14996 				iment_add(ism_ment, ism_hatid);
14997 				mutex_exit(&ism_mlist_lock);
14998 
14999 			}
15000 		}
15001 	}
15002 }
15003 /*
15004  * Unlink scd sfmmu from ism or hme region list for each region in the
15005  * scd region map.
15006  */
15007 void
15008 sfmmu_unlink_scd_from_regions(sf_srd_t *srdp, sf_scd_t *scdp)
15009 {
15010 	uint_t rid;
15011 	uint_t i;
15012 	uint_t j;
15013 	ulong_t w;
15014 	sf_region_t *rgnp;
15015 	sfmmu_t *scsfmmup;
15016 
15017 	scsfmmup = scdp->scd_sfmmup;
15018 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
15019 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
15020 			continue;
15021 		}
15022 		j = 0;
15023 		while (w) {
15024 			if (!(w & 0x1)) {
15025 				j++;
15026 				w >>= 1;
15027 				continue;
15028 			}
15029 			rid = (i << BT_ULSHIFT) | j;
15030 			j++;
15031 			w >>= 1;
15032 
15033 			if (rid < SFMMU_MAX_HME_REGIONS) {
15034 				rgnp = srdp->srd_hmergnp[rid];
15035 				ASSERT(rgnp->rgn_id == rid);
15036 				ASSERT(rgnp->rgn_refcnt > 0);
15037 				sfmmu_unlink_from_hmeregion(scsfmmup,
15038 				    rgnp);
15039 
15040 			} else {
15041 				sfmmu_t *ism_hatid = NULL;
15042 				ism_ment_t *ism_ment;
15043 				rid -= SFMMU_MAX_HME_REGIONS;
15044 				rgnp = srdp->srd_ismrgnp[rid];
15045 				ASSERT(rgnp->rgn_id == rid);
15046 				ASSERT(rgnp->rgn_refcnt > 0);
15047 
15048 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
15049 				ASSERT(ism_hatid->sfmmu_ismhat);
15050 				ism_ment = &scdp->scd_ism_links[rid];
15051 				ASSERT(ism_ment->iment_hat == scdp->scd_sfmmup);
15052 				ASSERT(ism_ment->iment_base_va ==
15053 				    rgnp->rgn_saddr);
15054 				mutex_enter(&ism_mlist_lock);
15055 				iment_sub(ism_ment, ism_hatid);
15056 				mutex_exit(&ism_mlist_lock);
15057 
15058 			}
15059 		}
15060 	}
15061 }
15062 /*
15063  * Allocates and initialises a new SCD structure, this is called with
15064  * the srd_scd_mutex held and returns with the reference count
15065  * initialised to 1.
15066  */
15067 static sf_scd_t *
15068 sfmmu_alloc_scd(sf_srd_t *srdp, sf_region_map_t *new_map)
15069 {
15070 	sf_scd_t *new_scdp;
15071 	sfmmu_t *scsfmmup;
15072 	int i;
15073 
15074 	ASSERT(MUTEX_HELD(&srdp->srd_scd_mutex));
15075 	new_scdp = kmem_cache_alloc(scd_cache, KM_SLEEP);
15076 
15077 	scsfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
15078 	new_scdp->scd_sfmmup = scsfmmup;
15079 	scsfmmup->sfmmu_srdp = srdp;
15080 	scsfmmup->sfmmu_scdp = new_scdp;
15081 	scsfmmup->sfmmu_tsb0_4minflcnt = 0;
15082 	scsfmmup->sfmmu_scdhat = 1;
15083 	CPUSET_ALL(scsfmmup->sfmmu_cpusran);
15084 	bzero(scsfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
15085 
15086 	ASSERT(max_mmu_ctxdoms > 0);
15087 	for (i = 0; i < max_mmu_ctxdoms; i++) {
15088 		scsfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
15089 		scsfmmup->sfmmu_ctxs[i].gnum = 0;
15090 	}
15091 
15092 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
15093 		new_scdp->scd_rttecnt[i] = 0;
15094 	}
15095 
15096 	new_scdp->scd_region_map = *new_map;
15097 	new_scdp->scd_refcnt = 1;
15098 	if (sfmmu_alloc_scd_tsbs(srdp, new_scdp) != TSB_SUCCESS) {
15099 		kmem_cache_free(scd_cache, new_scdp);
15100 		kmem_cache_free(sfmmuid_cache, scsfmmup);
15101 		return (NULL);
15102 	}
15103 	if (&mmu_init_scd) {
15104 		mmu_init_scd(new_scdp);
15105 	}
15106 	return (new_scdp);
15107 }
15108 
15109 /*
15110  * The first phase of a process joining an SCD. The hat structure is
15111  * linked to the SCD queue and then the HAT_JOIN_SCD sfmmu flag is set
15112  * and a cross-call with context invalidation is used to cause the
15113  * remaining work to be carried out in the sfmmu_tsbmiss_exception()
15114  * routine.
15115  */
15116 static void
15117 sfmmu_join_scd(sf_scd_t *scdp, sfmmu_t *sfmmup)
15118 {
15119 	hatlock_t *hatlockp;
15120 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15121 	int i;
15122 	sf_scd_t *old_scdp;
15123 
15124 	ASSERT(srdp != NULL);
15125 	ASSERT(scdp != NULL);
15126 	ASSERT(scdp->scd_refcnt > 0);
15127 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as));
15128 
15129 	if ((old_scdp = sfmmup->sfmmu_scdp) != NULL) {
15130 		ASSERT(old_scdp != scdp);
15131 
15132 		mutex_enter(&old_scdp->scd_mutex);
15133 		sfmmu_from_scd_list(&old_scdp->scd_sf_list, sfmmup);
15134 		mutex_exit(&old_scdp->scd_mutex);
15135 		/*
15136 		 * sfmmup leaves the old scd. Update sfmmu_ttecnt to
15137 		 * include the shme rgn ttecnt for rgns that
15138 		 * were in the old SCD
15139 		 */
15140 		for (i = 0; i < mmu_page_sizes; i++) {
15141 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15142 			    old_scdp->scd_rttecnt[i]);
15143 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15144 			    sfmmup->sfmmu_scdrttecnt[i]);
15145 		}
15146 	}
15147 
15148 	/*
15149 	 * Move sfmmu to the scd lists.
15150 	 */
15151 	mutex_enter(&scdp->scd_mutex);
15152 	sfmmu_to_scd_list(&scdp->scd_sf_list, sfmmup);
15153 	mutex_exit(&scdp->scd_mutex);
15154 	SF_SCD_INCR_REF(scdp);
15155 
15156 	hatlockp = sfmmu_hat_enter(sfmmup);
15157 	/*
15158 	 * For a multi-thread process, we must stop
15159 	 * all the other threads before joining the scd.
15160 	 */
15161 
15162 	SFMMU_FLAGS_SET(sfmmup, HAT_JOIN_SCD);
15163 
15164 	sfmmu_invalidate_ctx(sfmmup);
15165 	sfmmup->sfmmu_scdp = scdp;
15166 
15167 	/*
15168 	 * Copy scd_rttecnt into sfmmup's sfmmu_scdrttecnt, and update
15169 	 * sfmmu_ttecnt to not include the rgn ttecnt just joined in SCD.
15170 	 */
15171 	for (i = 0; i < mmu_page_sizes; i++) {
15172 		sfmmup->sfmmu_scdrttecnt[i] = scdp->scd_rttecnt[i];
15173 		ASSERT(sfmmup->sfmmu_ttecnt[i] >= scdp->scd_rttecnt[i]);
15174 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15175 		    -sfmmup->sfmmu_scdrttecnt[i]);
15176 	}
15177 	/* update tsb0 inflation count */
15178 	if (old_scdp != NULL) {
15179 		sfmmup->sfmmu_tsb0_4minflcnt +=
15180 		    old_scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15181 	}
15182 	ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
15183 	    scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt);
15184 	sfmmup->sfmmu_tsb0_4minflcnt -= scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15185 
15186 	sfmmu_hat_exit(hatlockp);
15187 
15188 	if (old_scdp != NULL) {
15189 		SF_SCD_DECR_REF(srdp, old_scdp);
15190 	}
15191 
15192 }
15193 
15194 /*
15195  * This routine is called by a process to become part of an SCD. It is called
15196  * from sfmmu_tsbmiss_exception() once most of the initial work has been
15197  * done by sfmmu_join_scd(). This routine must not drop the hat lock.
15198  */
15199 static void
15200 sfmmu_finish_join_scd(sfmmu_t *sfmmup)
15201 {
15202 	struct tsb_info	*tsbinfop;
15203 
15204 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15205 	ASSERT(sfmmup->sfmmu_scdp != NULL);
15206 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD));
15207 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15208 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID));
15209 
15210 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
15211 	    tsbinfop = tsbinfop->tsb_next) {
15212 		if (tsbinfop->tsb_flags & TSB_SWAPPED) {
15213 			continue;
15214 		}
15215 		ASSERT(!(tsbinfop->tsb_flags & TSB_RELOC_FLAG));
15216 
15217 		sfmmu_inv_tsb(tsbinfop->tsb_va,
15218 		    TSB_BYTES(tsbinfop->tsb_szc));
15219 	}
15220 
15221 	/* Set HAT_CTX1_FLAG for all SCD ISMs */
15222 	sfmmu_ism_hatflags(sfmmup, 1);
15223 
15224 	SFMMU_STAT(sf_join_scd);
15225 }
15226 
15227 /*
15228  * This routine is called in order to check if there is an SCD which matches
15229  * the process's region map if not then a new SCD may be created.
15230  */
15231 static void
15232 sfmmu_find_scd(sfmmu_t *sfmmup)
15233 {
15234 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15235 	sf_scd_t *scdp, *new_scdp;
15236 	int ret;
15237 
15238 	ASSERT(srdp != NULL);
15239 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as));
15240 
15241 	mutex_enter(&srdp->srd_scd_mutex);
15242 	for (scdp = srdp->srd_scdp; scdp != NULL;
15243 	    scdp = scdp->scd_next) {
15244 		SF_RGNMAP_EQUAL(&scdp->scd_region_map,
15245 		    &sfmmup->sfmmu_region_map, ret);
15246 		if (ret == 1) {
15247 			SF_SCD_INCR_REF(scdp);
15248 			mutex_exit(&srdp->srd_scd_mutex);
15249 			sfmmu_join_scd(scdp, sfmmup);
15250 			ASSERT(scdp->scd_refcnt >= 2);
15251 			atomic_dec_32((volatile uint32_t *)&scdp->scd_refcnt);
15252 			return;
15253 		} else {
15254 			/*
15255 			 * If the sfmmu region map is a subset of the scd
15256 			 * region map, then the assumption is that this process
15257 			 * will continue attaching to ISM segments until the
15258 			 * region maps are equal.
15259 			 */
15260 			SF_RGNMAP_IS_SUBSET(&scdp->scd_region_map,
15261 			    &sfmmup->sfmmu_region_map, ret);
15262 			if (ret == 1) {
15263 				mutex_exit(&srdp->srd_scd_mutex);
15264 				return;
15265 			}
15266 		}
15267 	}
15268 
15269 	ASSERT(scdp == NULL);
15270 	/*
15271 	 * No matching SCD has been found, create a new one.
15272 	 */
15273 	if ((new_scdp = sfmmu_alloc_scd(srdp, &sfmmup->sfmmu_region_map)) ==
15274 	    NULL) {
15275 		mutex_exit(&srdp->srd_scd_mutex);
15276 		return;
15277 	}
15278 
15279 	/*
15280 	 * sfmmu_alloc_scd() returns with a ref count of 1 on the scd.
15281 	 */
15282 
15283 	/* Set scd_rttecnt for shme rgns in SCD */
15284 	sfmmu_set_scd_rttecnt(srdp, new_scdp);
15285 
15286 	/*
15287 	 * Link scd onto srd_scdp list and scd sfmmu onto region/iment lists.
15288 	 */
15289 	sfmmu_link_scd_to_regions(srdp, new_scdp);
15290 	sfmmu_add_scd(&srdp->srd_scdp, new_scdp);
15291 	SFMMU_STAT_ADD(sf_create_scd, 1);
15292 
15293 	mutex_exit(&srdp->srd_scd_mutex);
15294 	sfmmu_join_scd(new_scdp, sfmmup);
15295 	ASSERT(new_scdp->scd_refcnt >= 2);
15296 	atomic_dec_32((volatile uint32_t *)&new_scdp->scd_refcnt);
15297 }
15298 
15299 /*
15300  * This routine is called by a process to remove itself from an SCD. It is
15301  * either called when the processes has detached from a segment or from
15302  * hat_free_start() as a result of calling exit.
15303  */
15304 static void
15305 sfmmu_leave_scd(sfmmu_t *sfmmup, uchar_t r_type)
15306 {
15307 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15308 	sf_srd_t *srdp =  sfmmup->sfmmu_srdp;
15309 	hatlock_t *hatlockp = TSB_HASH(sfmmup);
15310 	int i;
15311 
15312 	ASSERT(scdp != NULL);
15313 	ASSERT(srdp != NULL);
15314 
15315 	if (sfmmup->sfmmu_free) {
15316 		/*
15317 		 * If the process is part of an SCD the sfmmu is unlinked
15318 		 * from scd_sf_list.
15319 		 */
15320 		mutex_enter(&scdp->scd_mutex);
15321 		sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15322 		mutex_exit(&scdp->scd_mutex);
15323 		/*
15324 		 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15325 		 * are about to leave the SCD
15326 		 */
15327 		for (i = 0; i < mmu_page_sizes; i++) {
15328 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15329 			    scdp->scd_rttecnt[i]);
15330 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15331 			    sfmmup->sfmmu_scdrttecnt[i]);
15332 			sfmmup->sfmmu_scdrttecnt[i] = 0;
15333 		}
15334 		sfmmup->sfmmu_scdp = NULL;
15335 
15336 		SF_SCD_DECR_REF(srdp, scdp);
15337 		return;
15338 	}
15339 
15340 	ASSERT(r_type != SFMMU_REGION_ISM ||
15341 	    SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15342 	ASSERT(scdp->scd_refcnt);
15343 	ASSERT(!sfmmup->sfmmu_free);
15344 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15345 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as));
15346 
15347 	/*
15348 	 * Wait for ISM maps to be updated.
15349 	 */
15350 	if (r_type != SFMMU_REGION_ISM) {
15351 		while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY) &&
15352 		    sfmmup->sfmmu_scdp != NULL) {
15353 			cv_wait(&sfmmup->sfmmu_tsb_cv,
15354 			    HATLOCK_MUTEXP(hatlockp));
15355 		}
15356 
15357 		if (sfmmup->sfmmu_scdp == NULL) {
15358 			sfmmu_hat_exit(hatlockp);
15359 			return;
15360 		}
15361 		SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
15362 	}
15363 
15364 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
15365 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
15366 		/*
15367 		 * Since HAT_JOIN_SCD was set our context
15368 		 * is still invalid.
15369 		 */
15370 	} else {
15371 		/*
15372 		 * For a multi-thread process, we must stop
15373 		 * all the other threads before leaving the scd.
15374 		 */
15375 
15376 		sfmmu_invalidate_ctx(sfmmup);
15377 	}
15378 
15379 	/* Clear all the rid's for ISM, delete flags, etc */
15380 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15381 	sfmmu_ism_hatflags(sfmmup, 0);
15382 
15383 	/*
15384 	 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15385 	 * are in SCD before this sfmmup leaves the SCD.
15386 	 */
15387 	for (i = 0; i < mmu_page_sizes; i++) {
15388 		ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15389 		    scdp->scd_rttecnt[i]);
15390 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15391 		    sfmmup->sfmmu_scdrttecnt[i]);
15392 		sfmmup->sfmmu_scdrttecnt[i] = 0;
15393 		/* update ismttecnt to include SCD ism before hat leaves SCD */
15394 		sfmmup->sfmmu_ismttecnt[i] += sfmmup->sfmmu_scdismttecnt[i];
15395 		sfmmup->sfmmu_scdismttecnt[i] = 0;
15396 	}
15397 	/* update tsb0 inflation count */
15398 	sfmmup->sfmmu_tsb0_4minflcnt += scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15399 
15400 	if (r_type != SFMMU_REGION_ISM) {
15401 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
15402 	}
15403 	sfmmup->sfmmu_scdp = NULL;
15404 
15405 	sfmmu_hat_exit(hatlockp);
15406 
15407 	/*
15408 	 * Unlink sfmmu from scd_sf_list this can be done without holding
15409 	 * the hat lock as we hold the sfmmu_as lock which prevents
15410 	 * hat_join_region from adding this thread to the scd again. Other
15411 	 * threads check if sfmmu_scdp is NULL under hat lock and if it's NULL
15412 	 * they won't get here, since sfmmu_leave_scd() clears sfmmu_scdp
15413 	 * while holding the hat lock.
15414 	 */
15415 	mutex_enter(&scdp->scd_mutex);
15416 	sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15417 	mutex_exit(&scdp->scd_mutex);
15418 	SFMMU_STAT(sf_leave_scd);
15419 
15420 	SF_SCD_DECR_REF(srdp, scdp);
15421 	hatlockp = sfmmu_hat_enter(sfmmup);
15422 
15423 }
15424 
15425 /*
15426  * Unlink and free up an SCD structure with a reference count of 0.
15427  */
15428 static void
15429 sfmmu_destroy_scd(sf_srd_t *srdp, sf_scd_t *scdp, sf_region_map_t *scd_rmap)
15430 {
15431 	sfmmu_t *scsfmmup;
15432 	sf_scd_t *sp;
15433 	hatlock_t *shatlockp;
15434 	int i, ret;
15435 
15436 	mutex_enter(&srdp->srd_scd_mutex);
15437 	for (sp = srdp->srd_scdp; sp != NULL; sp = sp->scd_next) {
15438 		if (sp == scdp)
15439 			break;
15440 	}
15441 	if (sp == NULL || sp->scd_refcnt) {
15442 		mutex_exit(&srdp->srd_scd_mutex);
15443 		return;
15444 	}
15445 
15446 	/*
15447 	 * It is possible that the scd has been freed and reallocated with a
15448 	 * different region map while we've been waiting for the srd_scd_mutex.
15449 	 */
15450 	SF_RGNMAP_EQUAL(scd_rmap, &sp->scd_region_map, ret);
15451 	if (ret != 1) {
15452 		mutex_exit(&srdp->srd_scd_mutex);
15453 		return;
15454 	}
15455 
15456 	ASSERT(scdp->scd_sf_list == NULL);
15457 	/*
15458 	 * Unlink scd from srd_scdp list.
15459 	 */
15460 	sfmmu_remove_scd(&srdp->srd_scdp, scdp);
15461 	mutex_exit(&srdp->srd_scd_mutex);
15462 
15463 	sfmmu_unlink_scd_from_regions(srdp, scdp);
15464 
15465 	/* Clear shared context tsb and release ctx */
15466 	scsfmmup = scdp->scd_sfmmup;
15467 
15468 	/*
15469 	 * create a barrier so that scd will not be destroyed
15470 	 * if other thread still holds the same shared hat lock.
15471 	 * E.g., sfmmu_tsbmiss_exception() needs to acquire the
15472 	 * shared hat lock before checking the shared tsb reloc flag.
15473 	 */
15474 	shatlockp = sfmmu_hat_enter(scsfmmup);
15475 	sfmmu_hat_exit(shatlockp);
15476 
15477 	sfmmu_free_scd_tsbs(scsfmmup);
15478 
15479 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
15480 		if (scsfmmup->sfmmu_hmeregion_links[i] != NULL) {
15481 			kmem_free(scsfmmup->sfmmu_hmeregion_links[i],
15482 			    SFMMU_L2_HMERLINKS_SIZE);
15483 			scsfmmup->sfmmu_hmeregion_links[i] = NULL;
15484 		}
15485 	}
15486 	kmem_cache_free(sfmmuid_cache, scsfmmup);
15487 	kmem_cache_free(scd_cache, scdp);
15488 	SFMMU_STAT(sf_destroy_scd);
15489 }
15490 
15491 /*
15492  * Modifies the HAT_CTX1_FLAG for each of the ISM segments which correspond to
15493  * bits which are set in the ism_region_map parameter. This flag indicates to
15494  * the tsbmiss handler that mapping for these segments should be loaded using
15495  * the shared context.
15496  */
15497 static void
15498 sfmmu_ism_hatflags(sfmmu_t *sfmmup, int addflag)
15499 {
15500 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15501 	ism_blk_t *ism_blkp;
15502 	ism_map_t *ism_map;
15503 	int i, rid;
15504 
15505 	ASSERT(sfmmup->sfmmu_iblk != NULL);
15506 	ASSERT(scdp != NULL);
15507 	/*
15508 	 * Note that the caller either set HAT_ISMBUSY flag or checked
15509 	 * under hat lock that HAT_ISMBUSY was not set by another thread.
15510 	 */
15511 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15512 
15513 	ism_blkp = sfmmup->sfmmu_iblk;
15514 	while (ism_blkp != NULL) {
15515 		ism_map = ism_blkp->iblk_maps;
15516 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
15517 			rid = ism_map[i].imap_rid;
15518 			if (rid == SFMMU_INVALID_ISMRID) {
15519 				continue;
15520 			}
15521 			ASSERT(rid >= 0 && rid < SFMMU_MAX_ISM_REGIONS);
15522 			if (SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid) &&
15523 			    addflag) {
15524 				ism_map[i].imap_hatflags |=
15525 				    HAT_CTX1_FLAG;
15526 			} else {
15527 				ism_map[i].imap_hatflags &=
15528 				    ~HAT_CTX1_FLAG;
15529 			}
15530 		}
15531 		ism_blkp = ism_blkp->iblk_next;
15532 	}
15533 }
15534 
15535 static int
15536 sfmmu_srd_lock_held(sf_srd_t *srdp)
15537 {
15538 	return (MUTEX_HELD(&srdp->srd_mutex));
15539 }
15540 
15541 /* ARGSUSED */
15542 static int
15543 sfmmu_scdcache_constructor(void *buf, void *cdrarg, int kmflags)
15544 {
15545 	sf_scd_t *scdp = (sf_scd_t *)buf;
15546 
15547 	bzero(buf, sizeof (sf_scd_t));
15548 	mutex_init(&scdp->scd_mutex, NULL, MUTEX_DEFAULT, NULL);
15549 	return (0);
15550 }
15551 
15552 /* ARGSUSED */
15553 static void
15554 sfmmu_scdcache_destructor(void *buf, void *cdrarg)
15555 {
15556 	sf_scd_t *scdp = (sf_scd_t *)buf;
15557 
15558 	mutex_destroy(&scdp->scd_mutex);
15559 }
15560 
15561 /*
15562  * The listp parameter is a pointer to a list of hmeblks which are partially
15563  * freed as result of calling sfmmu_hblk_hash_rm(), the last phase of the
15564  * freeing process is to cross-call all cpus to ensure that there are no
15565  * remaining cached references.
15566  *
15567  * If the local generation number is less than the global then we can free
15568  * hmeblks which are already on the pending queue as another cpu has completed
15569  * the cross-call.
15570  *
15571  * We cross-call to make sure that there are no threads on other cpus accessing
15572  * these hmblks and then complete the process of freeing them under the
15573  * following conditions:
15574  * 	The total number of pending hmeblks is greater than the threshold
15575  *	The reserve list has fewer than HBLK_RESERVE_CNT hmeblks
15576  *	It is at least 1 second since the last time we cross-called
15577  *
15578  * Otherwise, we add the hmeblks to the per-cpu pending queue.
15579  */
15580 static void
15581 sfmmu_hblks_list_purge(struct hme_blk **listp, int dontfree)
15582 {
15583 	struct hme_blk *hblkp, *pr_hblkp = NULL;
15584 	int		count = 0;
15585 	cpuset_t	cpuset = cpu_ready_set;
15586 	cpu_hme_pend_t	*cpuhp;
15587 	timestruc_t	now;
15588 	int		one_second_expired = 0;
15589 
15590 	gethrestime_lasttick(&now);
15591 
15592 	for (hblkp = *listp; hblkp != NULL; hblkp = hblkp->hblk_next) {
15593 		ASSERT(hblkp->hblk_shw_bit == 0);
15594 		ASSERT(hblkp->hblk_shared == 0);
15595 		count++;
15596 		pr_hblkp = hblkp;
15597 	}
15598 
15599 	cpuhp = &cpu_hme_pend[CPU->cpu_seqid];
15600 	mutex_enter(&cpuhp->chp_mutex);
15601 
15602 	if ((cpuhp->chp_count + count) == 0) {
15603 		mutex_exit(&cpuhp->chp_mutex);
15604 		return;
15605 	}
15606 
15607 	if ((now.tv_sec - cpuhp->chp_timestamp) > 1) {
15608 		one_second_expired  = 1;
15609 	}
15610 
15611 	if (!dontfree && (freehblkcnt < HBLK_RESERVE_CNT ||
15612 	    (cpuhp->chp_count + count) > cpu_hme_pend_thresh ||
15613 	    one_second_expired)) {
15614 		/* Append global list to local */
15615 		if (pr_hblkp == NULL) {
15616 			*listp = cpuhp->chp_listp;
15617 		} else {
15618 			pr_hblkp->hblk_next = cpuhp->chp_listp;
15619 		}
15620 		cpuhp->chp_listp = NULL;
15621 		cpuhp->chp_count = 0;
15622 		cpuhp->chp_timestamp = now.tv_sec;
15623 		mutex_exit(&cpuhp->chp_mutex);
15624 
15625 		kpreempt_disable();
15626 		CPUSET_DEL(cpuset, CPU->cpu_id);
15627 		xt_sync(cpuset);
15628 		xt_sync(cpuset);
15629 		kpreempt_enable();
15630 
15631 		/*
15632 		 * At this stage we know that no trap handlers on other
15633 		 * cpus can have references to hmeblks on the list.
15634 		 */
15635 		sfmmu_hblk_free(listp);
15636 	} else if (*listp != NULL) {
15637 		pr_hblkp->hblk_next = cpuhp->chp_listp;
15638 		cpuhp->chp_listp = *listp;
15639 		cpuhp->chp_count += count;
15640 		*listp = NULL;
15641 		mutex_exit(&cpuhp->chp_mutex);
15642 	} else {
15643 		mutex_exit(&cpuhp->chp_mutex);
15644 	}
15645 }
15646 
15647 /*
15648  * Add an hmeblk to the the hash list.
15649  */
15650 void
15651 sfmmu_hblk_hash_add(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
15652 	uint64_t hblkpa)
15653 {
15654 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
15655 #ifdef	DEBUG
15656 	if (hmebp->hmeblkp == NULL) {
15657 		ASSERT(hmebp->hmeh_nextpa == HMEBLK_ENDPA);
15658 	}
15659 #endif /* DEBUG */
15660 
15661 	hmeblkp->hblk_nextpa = hmebp->hmeh_nextpa;
15662 	/*
15663 	 * Since the TSB miss handler now does not lock the hash chain before
15664 	 * walking it, make sure that the hmeblks nextpa is globally visible
15665 	 * before we make the hmeblk globally visible by updating the chain root
15666 	 * pointer in the hash bucket.
15667 	 */
15668 	membar_producer();
15669 	hmebp->hmeh_nextpa = hblkpa;
15670 	hmeblkp->hblk_next = hmebp->hmeblkp;
15671 	hmebp->hmeblkp = hmeblkp;
15672 
15673 }
15674 
15675 /*
15676  * This function is the first part of a 2 part process to remove an hmeblk
15677  * from the hash chain. In this phase we unlink the hmeblk from the hash chain
15678  * but leave the next physical pointer unchanged. The hmeblk is then linked onto
15679  * a per-cpu pending list using the virtual address pointer.
15680  *
15681  * TSB miss trap handlers that start after this phase will no longer see
15682  * this hmeblk. TSB miss handlers that still cache this hmeblk in a register
15683  * can still use it for further chain traversal because we haven't yet modifed
15684  * the next physical pointer or freed it.
15685  *
15686  * In the second phase of hmeblk removal we'll issue a barrier xcall before
15687  * we reuse or free this hmeblk. This will make sure all lingering references to
15688  * the hmeblk after first phase disappear before we finally reclaim it.
15689  * This scheme eliminates the need for TSB miss handlers to lock hmeblk chains
15690  * during their traversal.
15691  *
15692  * The hmehash_mutex must be held when calling this function.
15693  *
15694  * Input:
15695  *	 hmebp - hme hash bucket pointer
15696  *	 hmeblkp - address of hmeblk to be removed
15697  *	 pr_hblk - virtual address of previous hmeblkp
15698  *	 listp - pointer to list of hmeblks linked by virtual address
15699  *	 free_now flag - indicates that a complete removal from the hash chains
15700  *			 is necessary.
15701  *
15702  * It is inefficient to use the free_now flag as a cross-call is required to
15703  * remove a single hmeblk from the hash chain but is necessary when hmeblks are
15704  * in short supply.
15705  */
15706 void
15707 sfmmu_hblk_hash_rm(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
15708     struct hme_blk *pr_hblk, struct hme_blk **listp,
15709     int free_now)
15710 {
15711 	int shw_size, vshift;
15712 	struct hme_blk *shw_hblkp;
15713 	uint_t		shw_mask, newshw_mask;
15714 	caddr_t		vaddr;
15715 	int		size;
15716 	cpuset_t cpuset = cpu_ready_set;
15717 
15718 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
15719 
15720 	if (hmebp->hmeblkp == hmeblkp) {
15721 		hmebp->hmeh_nextpa = hmeblkp->hblk_nextpa;
15722 		hmebp->hmeblkp = hmeblkp->hblk_next;
15723 	} else {
15724 		pr_hblk->hblk_nextpa = hmeblkp->hblk_nextpa;
15725 		pr_hblk->hblk_next = hmeblkp->hblk_next;
15726 	}
15727 
15728 	size = get_hblk_ttesz(hmeblkp);
15729 	shw_hblkp = hmeblkp->hblk_shadow;
15730 	if (shw_hblkp) {
15731 		ASSERT(hblktosfmmu(hmeblkp) != KHATID);
15732 		ASSERT(!hmeblkp->hblk_shared);
15733 #ifdef	DEBUG
15734 		if (mmu_page_sizes == max_mmu_page_sizes) {
15735 			ASSERT(size < TTE256M);
15736 		} else {
15737 			ASSERT(size < TTE4M);
15738 		}
15739 #endif /* DEBUG */
15740 
15741 		shw_size = get_hblk_ttesz(shw_hblkp);
15742 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
15743 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
15744 		ASSERT(vshift < 8);
15745 		/*
15746 		 * Atomically clear shadow mask bit
15747 		 */
15748 		do {
15749 			shw_mask = shw_hblkp->hblk_shw_mask;
15750 			ASSERT(shw_mask & (1 << vshift));
15751 			newshw_mask = shw_mask & ~(1 << vshift);
15752 			newshw_mask = atomic_cas_32(&shw_hblkp->hblk_shw_mask,
15753 			    shw_mask, newshw_mask);
15754 		} while (newshw_mask != shw_mask);
15755 		hmeblkp->hblk_shadow = NULL;
15756 	}
15757 	hmeblkp->hblk_shw_bit = 0;
15758 
15759 	if (hmeblkp->hblk_shared) {
15760 #ifdef	DEBUG
15761 		sf_srd_t	*srdp;
15762 		sf_region_t	*rgnp;
15763 		uint_t		rid;
15764 
15765 		srdp = hblktosrd(hmeblkp);
15766 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
15767 		rid = hmeblkp->hblk_tag.htag_rid;
15768 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
15769 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
15770 		rgnp = srdp->srd_hmergnp[rid];
15771 		ASSERT(rgnp != NULL);
15772 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
15773 #endif /* DEBUG */
15774 		hmeblkp->hblk_shared = 0;
15775 	}
15776 	if (free_now) {
15777 		kpreempt_disable();
15778 		CPUSET_DEL(cpuset, CPU->cpu_id);
15779 		xt_sync(cpuset);
15780 		xt_sync(cpuset);
15781 		kpreempt_enable();
15782 
15783 		hmeblkp->hblk_nextpa = HMEBLK_ENDPA;
15784 		hmeblkp->hblk_next = NULL;
15785 	} else {
15786 		/* Append hmeblkp to listp for processing later. */
15787 		hmeblkp->hblk_next = *listp;
15788 		*listp = hmeblkp;
15789 	}
15790 }
15791 
15792 /*
15793  * This routine is called when memory is in short supply and returns a free
15794  * hmeblk of the requested size from the cpu pending lists.
15795  */
15796 static struct hme_blk *
15797 sfmmu_check_pending_hblks(int size)
15798 {
15799 	int i;
15800 	struct hme_blk *hmeblkp = NULL, *last_hmeblkp;
15801 	int found_hmeblk;
15802 	cpuset_t cpuset = cpu_ready_set;
15803 	cpu_hme_pend_t *cpuhp;
15804 
15805 	/* Flush cpu hblk pending queues */
15806 	for (i = 0; i < NCPU; i++) {
15807 		cpuhp = &cpu_hme_pend[i];
15808 		if (cpuhp->chp_listp != NULL)  {
15809 			mutex_enter(&cpuhp->chp_mutex);
15810 			if (cpuhp->chp_listp == NULL)  {
15811 				mutex_exit(&cpuhp->chp_mutex);
15812 				continue;
15813 			}
15814 			found_hmeblk = 0;
15815 			last_hmeblkp = NULL;
15816 			for (hmeblkp = cpuhp->chp_listp; hmeblkp != NULL;
15817 			    hmeblkp = hmeblkp->hblk_next) {
15818 				if (get_hblk_ttesz(hmeblkp) == size) {
15819 					if (last_hmeblkp == NULL) {
15820 						cpuhp->chp_listp =
15821 						    hmeblkp->hblk_next;
15822 					} else {
15823 						last_hmeblkp->hblk_next =
15824 						    hmeblkp->hblk_next;
15825 					}
15826 					ASSERT(cpuhp->chp_count > 0);
15827 					cpuhp->chp_count--;
15828 					found_hmeblk = 1;
15829 					break;
15830 				} else {
15831 					last_hmeblkp = hmeblkp;
15832 				}
15833 			}
15834 			mutex_exit(&cpuhp->chp_mutex);
15835 
15836 			if (found_hmeblk) {
15837 				kpreempt_disable();
15838 				CPUSET_DEL(cpuset, CPU->cpu_id);
15839 				xt_sync(cpuset);
15840 				xt_sync(cpuset);
15841 				kpreempt_enable();
15842 				return (hmeblkp);
15843 			}
15844 		}
15845 	}
15846 	return (NULL);
15847 }
15848