xref: /titanic_52/usr/src/uts/sfmmu/vm/hat_sfmmu.c (revision 2c164fafa089aa352e513b095e1ecd0abd29c61f)
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  * Copyright 2019 Joyent, Inc.
27  */
28 
29 /*
30  * VM - Hardware Address Translation management for Spitfire MMU.
31  *
32  * This file implements the machine specific hardware translation
33  * needed by the VM system.  The machine independent interface is
34  * described in <vm/hat.h> while the machine dependent interface
35  * and data structures are described in <vm/hat_sfmmu.h>.
36  *
37  * The hat layer manages the address translation hardware as a cache
38  * driven by calls from the higher levels in the VM system.
39  */
40 
41 #include <sys/types.h>
42 #include <sys/kstat.h>
43 #include <vm/hat.h>
44 #include <vm/hat_sfmmu.h>
45 #include <vm/page.h>
46 #include <sys/pte.h>
47 #include <sys/systm.h>
48 #include <sys/mman.h>
49 #include <sys/sysmacros.h>
50 #include <sys/machparam.h>
51 #include <sys/vtrace.h>
52 #include <sys/kmem.h>
53 #include <sys/mmu.h>
54 #include <sys/cmn_err.h>
55 #include <sys/cpu.h>
56 #include <sys/cpuvar.h>
57 #include <sys/debug.h>
58 #include <sys/lgrp.h>
59 #include <sys/archsystm.h>
60 #include <sys/machsystm.h>
61 #include <sys/vmsystm.h>
62 #include <vm/as.h>
63 #include <vm/seg.h>
64 #include <vm/seg_kp.h>
65 #include <vm/seg_kmem.h>
66 #include <vm/seg_kpm.h>
67 #include <vm/rm.h>
68 #include <sys/t_lock.h>
69 #include <sys/obpdefs.h>
70 #include <sys/vm_machparam.h>
71 #include <sys/var.h>
72 #include <sys/trap.h>
73 #include <sys/machtrap.h>
74 #include <sys/scb.h>
75 #include <sys/bitmap.h>
76 #include <sys/machlock.h>
77 #include <sys/membar.h>
78 #include <sys/atomic.h>
79 #include <sys/cpu_module.h>
80 #include <sys/prom_debug.h>
81 #include <sys/ksynch.h>
82 #include <sys/mem_config.h>
83 #include <sys/mem_cage.h>
84 #include <vm/vm_dep.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 	uhme_hash_pa = va_to_pa(uhme_hash);
1354 	khme_hash_pa = va_to_pa(khme_hash);
1355 
1356 	/*
1357 	 * Initialize relocation locks. kpr_suspendlock is held
1358 	 * at PIL_MAX to prevent interrupts from pinning the holder
1359 	 * of a suspended TTE which may access it leading to a
1360 	 * deadlock condition.
1361 	 */
1362 	mutex_init(&kpr_mutex, NULL, MUTEX_DEFAULT, NULL);
1363 	mutex_init(&kpr_suspendlock, NULL, MUTEX_SPIN, (void *)PIL_MAX);
1364 
1365 	/*
1366 	 * If Shared context support is disabled via /etc/system
1367 	 * set shctx_on to 0 here if it was set to 1 earlier in boot
1368 	 * sequence by cpu module initialization code.
1369 	 */
1370 	if (shctx_on && disable_shctx) {
1371 		shctx_on = 0;
1372 	}
1373 
1374 	if (shctx_on) {
1375 		srd_buckets = kmem_zalloc(SFMMU_MAX_SRD_BUCKETS *
1376 		    sizeof (srd_buckets[0]), KM_SLEEP);
1377 		for (i = 0; i < SFMMU_MAX_SRD_BUCKETS; i++) {
1378 			mutex_init(&srd_buckets[i].srdb_lock, NULL,
1379 			    MUTEX_DEFAULT, NULL);
1380 		}
1381 
1382 		srd_cache = kmem_cache_create("srd_cache", sizeof (sf_srd_t),
1383 		    0, sfmmu_srdcache_constructor, sfmmu_srdcache_destructor,
1384 		    NULL, NULL, NULL, 0);
1385 		region_cache = kmem_cache_create("region_cache",
1386 		    sizeof (sf_region_t), 0, sfmmu_rgncache_constructor,
1387 		    sfmmu_rgncache_destructor, NULL, NULL, NULL, 0);
1388 		scd_cache = kmem_cache_create("scd_cache", sizeof (sf_scd_t),
1389 		    0, sfmmu_scdcache_constructor,  sfmmu_scdcache_destructor,
1390 		    NULL, NULL, NULL, 0);
1391 	}
1392 
1393 	/*
1394 	 * Pre-allocate hrm_hashtab before enabling the collection of
1395 	 * refmod statistics.  Allocating on the fly would mean us
1396 	 * running the risk of suffering recursive mutex enters or
1397 	 * deadlocks.
1398 	 */
1399 	hrm_hashtab = kmem_zalloc(HRM_HASHSIZE * sizeof (struct hrmstat *),
1400 	    KM_SLEEP);
1401 
1402 	/* Allocate per-cpu pending freelist of hmeblks */
1403 	cpu_hme_pend = kmem_zalloc((NCPU * sizeof (cpu_hme_pend_t)) + 64,
1404 	    KM_SLEEP);
1405 	cpu_hme_pend = (cpu_hme_pend_t *)P2ROUNDUP(
1406 	    (uintptr_t)cpu_hme_pend, 64);
1407 
1408 	for (i = 0; i < NCPU; i++) {
1409 		mutex_init(&cpu_hme_pend[i].chp_mutex, NULL, MUTEX_DEFAULT,
1410 		    NULL);
1411 	}
1412 
1413 	if (cpu_hme_pend_thresh == 0) {
1414 		cpu_hme_pend_thresh = CPU_HME_PEND_THRESH;
1415 	}
1416 }
1417 
1418 /*
1419  * Initialize locking for the hat layer, called early during boot.
1420  */
1421 static void
1422 hat_lock_init()
1423 {
1424 	int i;
1425 
1426 	/*
1427 	 * initialize the array of mutexes protecting a page's mapping
1428 	 * list and p_nrm field.
1429 	 */
1430 	for (i = 0; i < MML_TABLE_SIZE; i++)
1431 		mutex_init(&mml_table[i].pad_mutex, NULL, MUTEX_DEFAULT, NULL);
1432 
1433 	if (kpm_enable) {
1434 		for (i = 0; i < kpmp_table_sz; i++) {
1435 			mutex_init(&kpmp_table[i].khl_mutex, NULL,
1436 			    MUTEX_DEFAULT, NULL);
1437 		}
1438 	}
1439 
1440 	/*
1441 	 * Initialize array of mutex locks that protects sfmmu fields and
1442 	 * TSB lists.
1443 	 */
1444 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
1445 		mutex_init(HATLOCK_MUTEXP(&hat_lock[i]), NULL, MUTEX_DEFAULT,
1446 		    NULL);
1447 }
1448 
1449 #define	SFMMU_KERNEL_MAXVA \
1450 	(kmem64_base ? (uintptr_t)kmem64_end : (SYSLIMIT))
1451 
1452 /*
1453  * Allocate a hat structure.
1454  * Called when an address space first uses a hat.
1455  */
1456 struct hat *
1457 hat_alloc(struct as *as)
1458 {
1459 	sfmmu_t *sfmmup;
1460 	int i;
1461 	uint64_t cnum;
1462 	extern uint_t get_color_start(struct as *);
1463 
1464 	ASSERT(AS_WRITE_HELD(as));
1465 	sfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
1466 	sfmmup->sfmmu_as = as;
1467 	sfmmup->sfmmu_flags = 0;
1468 	sfmmup->sfmmu_tteflags = 0;
1469 	sfmmup->sfmmu_rtteflags = 0;
1470 	LOCK_INIT_CLEAR(&sfmmup->sfmmu_ctx_lock);
1471 
1472 	if (as == &kas) {
1473 		ksfmmup = sfmmup;
1474 		sfmmup->sfmmu_cext = 0;
1475 		cnum = KCONTEXT;
1476 
1477 		sfmmup->sfmmu_clrstart = 0;
1478 		sfmmup->sfmmu_tsb = NULL;
1479 		/*
1480 		 * hat_kern_setup() will call sfmmu_init_ktsbinfo()
1481 		 * to setup tsb_info for ksfmmup.
1482 		 */
1483 	} else {
1484 
1485 		/*
1486 		 * Just set to invalid ctx. When it faults, it will
1487 		 * get a valid ctx. This would avoid the situation
1488 		 * where we get a ctx, but it gets stolen and then
1489 		 * we fault when we try to run and so have to get
1490 		 * another ctx.
1491 		 */
1492 		sfmmup->sfmmu_cext = 0;
1493 		cnum = INVALID_CONTEXT;
1494 
1495 		/* initialize original physical page coloring bin */
1496 		sfmmup->sfmmu_clrstart = get_color_start(as);
1497 #ifdef DEBUG
1498 		if (tsb_random_size) {
1499 			uint32_t randval = (uint32_t)gettick() >> 4;
1500 			int size = randval % (tsb_max_growsize + 1);
1501 
1502 			/* chose a random tsb size for stress testing */
1503 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb, size,
1504 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1505 		} else
1506 #endif /* DEBUG */
1507 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb,
1508 			    default_tsb_size,
1509 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1510 		sfmmup->sfmmu_flags = HAT_SWAPPED | HAT_ALLCTX_INVALID;
1511 		ASSERT(sfmmup->sfmmu_tsb != NULL);
1512 	}
1513 
1514 	ASSERT(max_mmu_ctxdoms > 0);
1515 	for (i = 0; i < max_mmu_ctxdoms; i++) {
1516 		sfmmup->sfmmu_ctxs[i].cnum = cnum;
1517 		sfmmup->sfmmu_ctxs[i].gnum = 0;
1518 	}
1519 
1520 	for (i = 0; i < max_mmu_page_sizes; i++) {
1521 		sfmmup->sfmmu_ttecnt[i] = 0;
1522 		sfmmup->sfmmu_scdrttecnt[i] = 0;
1523 		sfmmup->sfmmu_ismttecnt[i] = 0;
1524 		sfmmup->sfmmu_scdismttecnt[i] = 0;
1525 		sfmmup->sfmmu_pgsz[i] = TTE8K;
1526 	}
1527 	sfmmup->sfmmu_tsb0_4minflcnt = 0;
1528 	sfmmup->sfmmu_iblk = NULL;
1529 	sfmmup->sfmmu_ismhat = 0;
1530 	sfmmup->sfmmu_scdhat = 0;
1531 	sfmmup->sfmmu_ismblkpa = (uint64_t)-1;
1532 	if (sfmmup == ksfmmup) {
1533 		CPUSET_ALL(sfmmup->sfmmu_cpusran);
1534 	} else {
1535 		CPUSET_ZERO(sfmmup->sfmmu_cpusran);
1536 	}
1537 	sfmmup->sfmmu_free = 0;
1538 	sfmmup->sfmmu_rmstat = 0;
1539 	sfmmup->sfmmu_clrbin = sfmmup->sfmmu_clrstart;
1540 	cv_init(&sfmmup->sfmmu_tsb_cv, NULL, CV_DEFAULT, NULL);
1541 	sfmmup->sfmmu_srdp = NULL;
1542 	SF_RGNMAP_ZERO(sfmmup->sfmmu_region_map);
1543 	bzero(sfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
1544 	sfmmup->sfmmu_scdp = NULL;
1545 	sfmmup->sfmmu_scd_link.next = NULL;
1546 	sfmmup->sfmmu_scd_link.prev = NULL;
1547 	return (sfmmup);
1548 }
1549 
1550 /*
1551  * Create per-MMU context domain kstats for a given MMU ctx.
1552  */
1553 static void
1554 sfmmu_mmu_kstat_create(mmu_ctx_t *mmu_ctxp)
1555 {
1556 	mmu_ctx_stat_t	stat;
1557 	kstat_t		*mmu_kstat;
1558 
1559 	ASSERT(MUTEX_HELD(&cpu_lock));
1560 	ASSERT(mmu_ctxp->mmu_kstat == NULL);
1561 
1562 	mmu_kstat = kstat_create("unix", mmu_ctxp->mmu_idx, "mmu_ctx",
1563 	    "hat", KSTAT_TYPE_NAMED, MMU_CTX_NUM_STATS, KSTAT_FLAG_VIRTUAL);
1564 
1565 	if (mmu_kstat == NULL) {
1566 		cmn_err(CE_WARN, "kstat_create for MMU %d failed",
1567 		    mmu_ctxp->mmu_idx);
1568 	} else {
1569 		mmu_kstat->ks_data = mmu_ctxp->mmu_kstat_data;
1570 		for (stat = 0; stat < MMU_CTX_NUM_STATS; stat++)
1571 			kstat_named_init(&mmu_ctxp->mmu_kstat_data[stat],
1572 			    mmu_ctx_kstat_names[stat], KSTAT_DATA_INT64);
1573 		mmu_ctxp->mmu_kstat = mmu_kstat;
1574 		kstat_install(mmu_kstat);
1575 	}
1576 }
1577 
1578 /*
1579  * plat_cpuid_to_mmu_ctx_info() is a platform interface that returns MMU
1580  * context domain information for a given CPU. If a platform does not
1581  * specify that interface, then the function below is used instead to return
1582  * default information. The defaults are as follows:
1583  *
1584  *	- The number of MMU context IDs supported on any CPU in the
1585  *	  system is 8K.
1586  *	- There is one MMU context domain per CPU.
1587  */
1588 /*ARGSUSED*/
1589 static void
1590 sfmmu_cpuid_to_mmu_ctx_info(processorid_t cpuid, mmu_ctx_info_t *infop)
1591 {
1592 	infop->mmu_nctxs = nctxs;
1593 	infop->mmu_idx = cpu[cpuid]->cpu_seqid;
1594 }
1595 
1596 /*
1597  * Called during CPU initialization to set the MMU context-related information
1598  * for a CPU.
1599  *
1600  * cpu_lock serializes accesses to mmu_ctxs and mmu_saved_gnum.
1601  */
1602 void
1603 sfmmu_cpu_init(cpu_t *cp)
1604 {
1605 	mmu_ctx_info_t	info;
1606 	mmu_ctx_t	*mmu_ctxp;
1607 
1608 	ASSERT(MUTEX_HELD(&cpu_lock));
1609 
1610 	if (&plat_cpuid_to_mmu_ctx_info == NULL)
1611 		sfmmu_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1612 	else
1613 		plat_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1614 
1615 	ASSERT(info.mmu_idx < max_mmu_ctxdoms);
1616 
1617 	if ((mmu_ctxp = mmu_ctxs_tbl[info.mmu_idx]) == NULL) {
1618 		/* Each mmu_ctx is cacheline aligned. */
1619 		mmu_ctxp = kmem_cache_alloc(mmuctxdom_cache, KM_SLEEP);
1620 		bzero(mmu_ctxp, sizeof (mmu_ctx_t));
1621 
1622 		mutex_init(&mmu_ctxp->mmu_lock, NULL, MUTEX_SPIN,
1623 		    (void *)ipltospl(DISP_LEVEL));
1624 		mmu_ctxp->mmu_idx = info.mmu_idx;
1625 		mmu_ctxp->mmu_nctxs = info.mmu_nctxs;
1626 		/*
1627 		 * Globally for lifetime of a system,
1628 		 * gnum must always increase.
1629 		 * mmu_saved_gnum is protected by the cpu_lock.
1630 		 */
1631 		mmu_ctxp->mmu_gnum = mmu_saved_gnum + 1;
1632 		mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
1633 
1634 		sfmmu_mmu_kstat_create(mmu_ctxp);
1635 
1636 		mmu_ctxs_tbl[info.mmu_idx] = mmu_ctxp;
1637 	} else {
1638 		ASSERT(mmu_ctxp->mmu_idx == info.mmu_idx);
1639 		ASSERT(mmu_ctxp->mmu_nctxs <= info.mmu_nctxs);
1640 	}
1641 
1642 	/*
1643 	 * The mmu_lock is acquired here to prevent races with
1644 	 * the wrap-around code.
1645 	 */
1646 	mutex_enter(&mmu_ctxp->mmu_lock);
1647 
1648 
1649 	mmu_ctxp->mmu_ncpus++;
1650 	CPUSET_ADD(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1651 	CPU_MMU_IDX(cp) = info.mmu_idx;
1652 	CPU_MMU_CTXP(cp) = mmu_ctxp;
1653 
1654 	mutex_exit(&mmu_ctxp->mmu_lock);
1655 }
1656 
1657 static void
1658 sfmmu_ctxdom_free(mmu_ctx_t *mmu_ctxp)
1659 {
1660 	ASSERT(MUTEX_HELD(&cpu_lock));
1661 	ASSERT(!MUTEX_HELD(&mmu_ctxp->mmu_lock));
1662 
1663 	mutex_destroy(&mmu_ctxp->mmu_lock);
1664 
1665 	if (mmu_ctxp->mmu_kstat)
1666 		kstat_delete(mmu_ctxp->mmu_kstat);
1667 
1668 	/* mmu_saved_gnum is protected by the cpu_lock. */
1669 	if (mmu_saved_gnum < mmu_ctxp->mmu_gnum)
1670 		mmu_saved_gnum = mmu_ctxp->mmu_gnum;
1671 
1672 	kmem_cache_free(mmuctxdom_cache, mmu_ctxp);
1673 }
1674 
1675 /*
1676  * Called to perform MMU context-related cleanup for a CPU.
1677  */
1678 void
1679 sfmmu_cpu_cleanup(cpu_t *cp)
1680 {
1681 	mmu_ctx_t	*mmu_ctxp;
1682 
1683 	ASSERT(MUTEX_HELD(&cpu_lock));
1684 
1685 	mmu_ctxp = CPU_MMU_CTXP(cp);
1686 	ASSERT(mmu_ctxp != NULL);
1687 
1688 	/*
1689 	 * The mmu_lock is acquired here to prevent races with
1690 	 * the wrap-around code.
1691 	 */
1692 	mutex_enter(&mmu_ctxp->mmu_lock);
1693 
1694 	CPU_MMU_CTXP(cp) = NULL;
1695 
1696 	CPUSET_DEL(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1697 	if (--mmu_ctxp->mmu_ncpus == 0) {
1698 		mmu_ctxs_tbl[mmu_ctxp->mmu_idx] = NULL;
1699 		mutex_exit(&mmu_ctxp->mmu_lock);
1700 		sfmmu_ctxdom_free(mmu_ctxp);
1701 		return;
1702 	}
1703 
1704 	mutex_exit(&mmu_ctxp->mmu_lock);
1705 }
1706 
1707 uint_t
1708 sfmmu_ctxdom_nctxs(int idx)
1709 {
1710 	return (mmu_ctxs_tbl[idx]->mmu_nctxs);
1711 }
1712 
1713 #ifdef sun4v
1714 /*
1715  * sfmmu_ctxdoms_* is an interface provided to help keep context domains
1716  * consistant after suspend/resume on system that can resume on a different
1717  * hardware than it was suspended.
1718  *
1719  * sfmmu_ctxdom_lock(void) locks all context domains and prevents new contexts
1720  * from being allocated.  It acquires all hat_locks, which blocks most access to
1721  * context data, except for a few cases that are handled separately or are
1722  * harmless.  It wraps each domain to increment gnum and invalidate on-CPU
1723  * contexts, and forces cnum to its max.  As a result of this call all user
1724  * threads that are running on CPUs trap and try to perform wrap around but
1725  * can't because hat_locks are taken.  Threads that were not on CPUs but started
1726  * by scheduler go to sfmmu_alloc_ctx() to aquire context without checking
1727  * hat_lock, but fail, because cnum == nctxs, and therefore also trap and block
1728  * on hat_lock trying to wrap.  sfmmu_ctxdom_lock() must be called before CPUs
1729  * are paused, else it could deadlock acquiring locks held by paused CPUs.
1730  *
1731  * sfmmu_ctxdoms_remove() removes context domains from every CPUs and records
1732  * the CPUs that had them.  It must be called after CPUs have been paused. This
1733  * ensures that no threads are in sfmmu_alloc_ctx() accessing domain data,
1734  * because pause_cpus sends a mondo interrupt to every CPU, and sfmmu_alloc_ctx
1735  * runs with interrupts disabled.  When CPUs are later resumed, they may enter
1736  * sfmmu_alloc_ctx, but it will check for CPU_MMU_CTXP = NULL and immediately
1737  * return failure.  Or, they will be blocked trying to acquire hat_lock. Thus
1738  * after sfmmu_ctxdoms_remove returns, we are guaranteed that no one is
1739  * accessing the old context domains.
1740  *
1741  * sfmmu_ctxdoms_update(void) frees space used by old context domains and
1742  * allocates new context domains based on hardware layout.  It initializes
1743  * every CPU that had context domain before migration to have one again.
1744  * sfmmu_ctxdoms_update must be called after CPUs are resumed, else it
1745  * could deadlock acquiring locks held by paused CPUs.
1746  *
1747  * sfmmu_ctxdoms_unlock(void) releases all hat_locks after which user threads
1748  * acquire new context ids and continue execution.
1749  *
1750  * Therefore functions should be called in the following order:
1751  *       suspend_routine()
1752  *		sfmmu_ctxdom_lock()
1753  *		pause_cpus()
1754  *		suspend()
1755  *			if (suspend failed)
1756  *				sfmmu_ctxdom_unlock()
1757  *		...
1758  *		sfmmu_ctxdom_remove()
1759  *		resume_cpus()
1760  *		sfmmu_ctxdom_update()
1761  *		sfmmu_ctxdom_unlock()
1762  */
1763 static cpuset_t sfmmu_ctxdoms_pset;
1764 
1765 void
1766 sfmmu_ctxdoms_remove()
1767 {
1768 	processorid_t	id;
1769 	cpu_t		*cp;
1770 
1771 	/*
1772 	 * Record the CPUs that have domains in sfmmu_ctxdoms_pset, so they can
1773 	 * be restored post-migration. A CPU may be powered off and not have a
1774 	 * domain, for example.
1775 	 */
1776 	CPUSET_ZERO(sfmmu_ctxdoms_pset);
1777 
1778 	for (id = 0; id < NCPU; id++) {
1779 		if ((cp = cpu[id]) != NULL && CPU_MMU_CTXP(cp) != NULL) {
1780 			CPUSET_ADD(sfmmu_ctxdoms_pset, id);
1781 			CPU_MMU_CTXP(cp) = NULL;
1782 		}
1783 	}
1784 }
1785 
1786 void
1787 sfmmu_ctxdoms_lock(void)
1788 {
1789 	int		idx;
1790 	mmu_ctx_t	*mmu_ctxp;
1791 
1792 	sfmmu_hat_lock_all();
1793 
1794 	/*
1795 	 * At this point, no thread can be in sfmmu_ctx_wrap_around, because
1796 	 * hat_lock is always taken before calling it.
1797 	 *
1798 	 * For each domain, set mmu_cnum to max so no more contexts can be
1799 	 * allocated, and wrap to flush on-CPU contexts and force threads to
1800 	 * acquire a new context when we later drop hat_lock after migration.
1801 	 * Setting mmu_cnum may race with sfmmu_alloc_ctx which also sets cnum,
1802 	 * but the latter uses CAS and will miscompare and not overwrite it.
1803 	 */
1804 	kpreempt_disable(); /* required by sfmmu_ctx_wrap_around */
1805 	for (idx = 0; idx < max_mmu_ctxdoms; idx++) {
1806 		if ((mmu_ctxp = mmu_ctxs_tbl[idx]) != NULL) {
1807 			mutex_enter(&mmu_ctxp->mmu_lock);
1808 			mmu_ctxp->mmu_cnum = mmu_ctxp->mmu_nctxs;
1809 			/* make sure updated cnum visible */
1810 			membar_enter();
1811 			mutex_exit(&mmu_ctxp->mmu_lock);
1812 			sfmmu_ctx_wrap_around(mmu_ctxp, B_FALSE);
1813 		}
1814 	}
1815 	kpreempt_enable();
1816 }
1817 
1818 void
1819 sfmmu_ctxdoms_unlock(void)
1820 {
1821 	sfmmu_hat_unlock_all();
1822 }
1823 
1824 void
1825 sfmmu_ctxdoms_update(void)
1826 {
1827 	processorid_t	id;
1828 	cpu_t		*cp;
1829 	uint_t		idx;
1830 	mmu_ctx_t	*mmu_ctxp;
1831 
1832 	/*
1833 	 * Free all context domains.  As side effect, this increases
1834 	 * mmu_saved_gnum to the maximum gnum over all domains, which is used to
1835 	 * init gnum in the new domains, which therefore will be larger than the
1836 	 * sfmmu gnum for any process, guaranteeing that every process will see
1837 	 * a new generation and allocate a new context regardless of what new
1838 	 * domain it runs in.
1839 	 */
1840 	mutex_enter(&cpu_lock);
1841 
1842 	for (idx = 0; idx < max_mmu_ctxdoms; idx++) {
1843 		if (mmu_ctxs_tbl[idx] != NULL) {
1844 			mmu_ctxp = mmu_ctxs_tbl[idx];
1845 			mmu_ctxs_tbl[idx] = NULL;
1846 			sfmmu_ctxdom_free(mmu_ctxp);
1847 		}
1848 	}
1849 
1850 	for (id = 0; id < NCPU; id++) {
1851 		if (CPU_IN_SET(sfmmu_ctxdoms_pset, id) &&
1852 		    (cp = cpu[id]) != NULL)
1853 			sfmmu_cpu_init(cp);
1854 	}
1855 	mutex_exit(&cpu_lock);
1856 }
1857 #endif
1858 
1859 /*
1860  * Hat_setup, makes an address space context the current active one.
1861  * In sfmmu this translates to setting the secondary context with the
1862  * corresponding context.
1863  */
1864 void
1865 hat_setup(struct hat *sfmmup, int allocflag)
1866 {
1867 	hatlock_t *hatlockp;
1868 
1869 	/* Init needs some special treatment. */
1870 	if (allocflag == HAT_INIT) {
1871 		/*
1872 		 * Make sure that we have
1873 		 * 1. a TSB
1874 		 * 2. a valid ctx that doesn't get stolen after this point.
1875 		 */
1876 		hatlockp = sfmmu_hat_enter(sfmmup);
1877 
1878 		/*
1879 		 * Swap in the TSB.  hat_init() allocates tsbinfos without
1880 		 * TSBs, but we need one for init, since the kernel does some
1881 		 * special things to set up its stack and needs the TSB to
1882 		 * resolve page faults.
1883 		 */
1884 		sfmmu_tsb_swapin(sfmmup, hatlockp);
1885 
1886 		sfmmu_get_ctx(sfmmup);
1887 
1888 		sfmmu_hat_exit(hatlockp);
1889 	} else {
1890 		ASSERT(allocflag == HAT_ALLOC);
1891 
1892 		hatlockp = sfmmu_hat_enter(sfmmup);
1893 		kpreempt_disable();
1894 
1895 		CPUSET_ADD(sfmmup->sfmmu_cpusran, CPU->cpu_id);
1896 		/*
1897 		 * sfmmu_setctx_sec takes <pgsz|cnum> as a parameter,
1898 		 * pagesize bits don't matter in this case since we are passing
1899 		 * INVALID_CONTEXT to it.
1900 		 * Compatibility Note: hw takes care of MMU_SCONTEXT1
1901 		 */
1902 		sfmmu_setctx_sec(INVALID_CONTEXT);
1903 		sfmmu_clear_utsbinfo();
1904 
1905 		kpreempt_enable();
1906 		sfmmu_hat_exit(hatlockp);
1907 	}
1908 }
1909 
1910 /*
1911  * Free all the translation resources for the specified address space.
1912  * Called from as_free when an address space is being destroyed.
1913  */
1914 void
1915 hat_free_start(struct hat *sfmmup)
1916 {
1917 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as));
1918 	ASSERT(sfmmup != ksfmmup);
1919 
1920 	sfmmup->sfmmu_free = 1;
1921 	if (sfmmup->sfmmu_scdp != NULL) {
1922 		sfmmu_leave_scd(sfmmup, 0);
1923 	}
1924 
1925 	ASSERT(sfmmup->sfmmu_scdp == NULL);
1926 }
1927 
1928 void
1929 hat_free_end(struct hat *sfmmup)
1930 {
1931 	int i;
1932 
1933 	ASSERT(sfmmup->sfmmu_free == 1);
1934 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
1935 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
1936 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
1937 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
1938 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
1939 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
1940 
1941 	if (sfmmup->sfmmu_rmstat) {
1942 		hat_freestat(sfmmup->sfmmu_as, NULL);
1943 	}
1944 
1945 	while (sfmmup->sfmmu_tsb != NULL) {
1946 		struct tsb_info *next = sfmmup->sfmmu_tsb->tsb_next;
1947 		sfmmu_tsbinfo_free(sfmmup->sfmmu_tsb);
1948 		sfmmup->sfmmu_tsb = next;
1949 	}
1950 
1951 	if (sfmmup->sfmmu_srdp != NULL) {
1952 		sfmmu_leave_srd(sfmmup);
1953 		ASSERT(sfmmup->sfmmu_srdp == NULL);
1954 		for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1955 			if (sfmmup->sfmmu_hmeregion_links[i] != NULL) {
1956 				kmem_free(sfmmup->sfmmu_hmeregion_links[i],
1957 				    SFMMU_L2_HMERLINKS_SIZE);
1958 				sfmmup->sfmmu_hmeregion_links[i] = NULL;
1959 			}
1960 		}
1961 	}
1962 	sfmmu_free_sfmmu(sfmmup);
1963 
1964 #ifdef DEBUG
1965 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1966 		ASSERT(sfmmup->sfmmu_hmeregion_links[i] == NULL);
1967 	}
1968 #endif
1969 
1970 	kmem_cache_free(sfmmuid_cache, sfmmup);
1971 }
1972 
1973 /*
1974  * Set up any translation structures, for the specified address space,
1975  * that are needed or preferred when the process is being swapped in.
1976  */
1977 /* ARGSUSED */
1978 void
1979 hat_swapin(struct hat *hat)
1980 {
1981 }
1982 
1983 /*
1984  * Free all of the translation resources, for the specified address space,
1985  * that can be freed while the process is swapped out. Called from as_swapout.
1986  * Also, free up the ctx that this process was using.
1987  */
1988 void
1989 hat_swapout(struct hat *sfmmup)
1990 {
1991 	struct hmehash_bucket *hmebp;
1992 	struct hme_blk *hmeblkp;
1993 	struct hme_blk *pr_hblk = NULL;
1994 	struct hme_blk *nx_hblk;
1995 	int i;
1996 	struct hme_blk *list = NULL;
1997 	hatlock_t *hatlockp;
1998 	struct tsb_info *tsbinfop;
1999 	struct free_tsb {
2000 		struct free_tsb *next;
2001 		struct tsb_info *tsbinfop;
2002 	};			/* free list of TSBs */
2003 	struct free_tsb *freelist, *last, *next;
2004 
2005 	SFMMU_STAT(sf_swapout);
2006 
2007 	/*
2008 	 * There is no way to go from an as to all its translations in sfmmu.
2009 	 * Here is one of the times when we take the big hit and traverse
2010 	 * the hash looking for hme_blks to free up.  Not only do we free up
2011 	 * this as hme_blks but all those that are free.  We are obviously
2012 	 * swapping because we need memory so let's free up as much
2013 	 * as we can.
2014 	 *
2015 	 * Note that we don't flush TLB/TSB here -- it's not necessary
2016 	 * because:
2017 	 *  1) we free the ctx we're using and throw away the TSB(s);
2018 	 *  2) processes aren't runnable while being swapped out.
2019 	 */
2020 	ASSERT(sfmmup != KHATID);
2021 	for (i = 0; i <= UHMEHASH_SZ; i++) {
2022 		hmebp = &uhme_hash[i];
2023 		SFMMU_HASH_LOCK(hmebp);
2024 		hmeblkp = hmebp->hmeblkp;
2025 		pr_hblk = NULL;
2026 		while (hmeblkp) {
2027 
2028 			if ((hmeblkp->hblk_tag.htag_id == sfmmup) &&
2029 			    !hmeblkp->hblk_shw_bit && !hmeblkp->hblk_lckcnt) {
2030 				ASSERT(!hmeblkp->hblk_shared);
2031 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
2032 				    (caddr_t)get_hblk_base(hmeblkp),
2033 				    get_hblk_endaddr(hmeblkp),
2034 				    NULL, HAT_UNLOAD);
2035 			}
2036 			nx_hblk = hmeblkp->hblk_next;
2037 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
2038 				ASSERT(!hmeblkp->hblk_lckcnt);
2039 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
2040 				    &list, 0);
2041 			} else {
2042 				pr_hblk = hmeblkp;
2043 			}
2044 			hmeblkp = nx_hblk;
2045 		}
2046 		SFMMU_HASH_UNLOCK(hmebp);
2047 	}
2048 
2049 	sfmmu_hblks_list_purge(&list, 0);
2050 
2051 	/*
2052 	 * Now free up the ctx so that others can reuse it.
2053 	 */
2054 	hatlockp = sfmmu_hat_enter(sfmmup);
2055 
2056 	sfmmu_invalidate_ctx(sfmmup);
2057 
2058 	/*
2059 	 * Free TSBs, but not tsbinfos, and set SWAPPED flag.
2060 	 * If TSBs were never swapped in, just return.
2061 	 * This implies that we don't support partial swapping
2062 	 * of TSBs -- either all are swapped out, or none are.
2063 	 *
2064 	 * We must hold the HAT lock here to prevent racing with another
2065 	 * thread trying to unmap TTEs from the TSB or running the post-
2066 	 * relocator after relocating the TSB's memory.  Unfortunately, we
2067 	 * can't free memory while holding the HAT lock or we could
2068 	 * deadlock, so we build a list of TSBs to be freed after marking
2069 	 * the tsbinfos as swapped out and free them after dropping the
2070 	 * lock.
2071 	 */
2072 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
2073 		sfmmu_hat_exit(hatlockp);
2074 		return;
2075 	}
2076 
2077 	SFMMU_FLAGS_SET(sfmmup, HAT_SWAPPED);
2078 	last = freelist = NULL;
2079 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
2080 	    tsbinfop = tsbinfop->tsb_next) {
2081 		ASSERT((tsbinfop->tsb_flags & TSB_SWAPPED) == 0);
2082 
2083 		/*
2084 		 * Cast the TSB into a struct free_tsb and put it on the free
2085 		 * list.
2086 		 */
2087 		if (freelist == NULL) {
2088 			last = freelist = (struct free_tsb *)tsbinfop->tsb_va;
2089 		} else {
2090 			last->next = (struct free_tsb *)tsbinfop->tsb_va;
2091 			last = last->next;
2092 		}
2093 		last->next = NULL;
2094 		last->tsbinfop = tsbinfop;
2095 		tsbinfop->tsb_flags |= TSB_SWAPPED;
2096 		/*
2097 		 * Zero out the TTE to clear the valid bit.
2098 		 * Note we can't use a value like 0xbad because we want to
2099 		 * ensure diagnostic bits are NEVER set on TTEs that might
2100 		 * be loaded.  The intent is to catch any invalid access
2101 		 * to the swapped TSB, such as a thread running with a valid
2102 		 * context without first calling sfmmu_tsb_swapin() to
2103 		 * allocate TSB memory.
2104 		 */
2105 		tsbinfop->tsb_tte.ll = 0;
2106 	}
2107 
2108 	/* Now we can drop the lock and free the TSB memory. */
2109 	sfmmu_hat_exit(hatlockp);
2110 	for (; freelist != NULL; freelist = next) {
2111 		next = freelist->next;
2112 		sfmmu_tsb_free(freelist->tsbinfop);
2113 	}
2114 }
2115 
2116 /*
2117  * Duplicate the translations of an as into another newas
2118  */
2119 /* ARGSUSED */
2120 int
2121 hat_dup(struct hat *hat, struct hat *newhat, caddr_t addr, size_t len,
2122     uint_t flag)
2123 {
2124 	sf_srd_t *srdp;
2125 	sf_scd_t *scdp;
2126 	int i;
2127 	extern uint_t get_color_start(struct as *);
2128 
2129 	ASSERT((flag == 0) || (flag == HAT_DUP_ALL) || (flag == HAT_DUP_COW) ||
2130 	    (flag == HAT_DUP_SRD));
2131 	ASSERT(hat != ksfmmup);
2132 	ASSERT(newhat != ksfmmup);
2133 	ASSERT(flag != HAT_DUP_ALL || hat->sfmmu_srdp == newhat->sfmmu_srdp);
2134 
2135 	if (flag == HAT_DUP_COW) {
2136 		panic("hat_dup: HAT_DUP_COW not supported");
2137 	}
2138 
2139 	if (flag == HAT_DUP_SRD && ((srdp = hat->sfmmu_srdp) != NULL)) {
2140 		ASSERT(srdp->srd_evp != NULL);
2141 		VN_HOLD(srdp->srd_evp);
2142 		ASSERT(srdp->srd_refcnt > 0);
2143 		newhat->sfmmu_srdp = srdp;
2144 		atomic_inc_32((volatile uint_t *)&srdp->srd_refcnt);
2145 	}
2146 
2147 	/*
2148 	 * HAT_DUP_ALL flag is used after as duplication is done.
2149 	 */
2150 	if (flag == HAT_DUP_ALL && ((srdp = newhat->sfmmu_srdp) != NULL)) {
2151 		ASSERT(newhat->sfmmu_srdp->srd_refcnt >= 2);
2152 		newhat->sfmmu_rtteflags = hat->sfmmu_rtteflags;
2153 		if (hat->sfmmu_flags & HAT_4MTEXT_FLAG) {
2154 			newhat->sfmmu_flags |= HAT_4MTEXT_FLAG;
2155 		}
2156 
2157 		/* check if need to join scd */
2158 		if ((scdp = hat->sfmmu_scdp) != NULL &&
2159 		    newhat->sfmmu_scdp != scdp) {
2160 			int ret;
2161 			SF_RGNMAP_IS_SUBSET(&newhat->sfmmu_region_map,
2162 			    &scdp->scd_region_map, ret);
2163 			ASSERT(ret);
2164 			sfmmu_join_scd(scdp, newhat);
2165 			ASSERT(newhat->sfmmu_scdp == scdp &&
2166 			    scdp->scd_refcnt >= 2);
2167 			for (i = 0; i < max_mmu_page_sizes; i++) {
2168 				newhat->sfmmu_ismttecnt[i] =
2169 				    hat->sfmmu_ismttecnt[i];
2170 				newhat->sfmmu_scdismttecnt[i] =
2171 				    hat->sfmmu_scdismttecnt[i];
2172 			}
2173 		}
2174 
2175 		sfmmu_check_page_sizes(newhat, 1);
2176 	}
2177 
2178 	if (flag == HAT_DUP_ALL && consistent_coloring == 0 &&
2179 	    update_proc_pgcolorbase_after_fork != 0) {
2180 		hat->sfmmu_clrbin = get_color_start(hat->sfmmu_as);
2181 	}
2182 	return (0);
2183 }
2184 
2185 void
2186 hat_memload(struct hat *hat, caddr_t addr, struct page *pp,
2187     uint_t attr, uint_t flags)
2188 {
2189 	hat_do_memload(hat, addr, pp, attr, flags,
2190 	    SFMMU_INVALID_SHMERID);
2191 }
2192 
2193 void
2194 hat_memload_region(struct hat *hat, caddr_t addr, struct page *pp,
2195     uint_t attr, uint_t flags, hat_region_cookie_t rcookie)
2196 {
2197 	uint_t rid;
2198 	if (rcookie == HAT_INVALID_REGION_COOKIE) {
2199 		hat_do_memload(hat, addr, pp, attr, flags,
2200 		    SFMMU_INVALID_SHMERID);
2201 		return;
2202 	}
2203 	rid = (uint_t)((uint64_t)rcookie);
2204 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2205 	hat_do_memload(hat, addr, pp, attr, flags, rid);
2206 }
2207 
2208 /*
2209  * Set up addr to map to page pp with protection prot.
2210  * As an optimization we also load the TSB with the
2211  * corresponding tte but it is no big deal if  the tte gets kicked out.
2212  */
2213 static void
2214 hat_do_memload(struct hat *hat, caddr_t addr, struct page *pp,
2215     uint_t attr, uint_t flags, uint_t rid)
2216 {
2217 	tte_t tte;
2218 
2219 
2220 	ASSERT(hat != NULL);
2221 	ASSERT(PAGE_LOCKED(pp));
2222 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2223 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2224 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2225 	SFMMU_VALIDATE_HMERID(hat, rid, addr, MMU_PAGESIZE);
2226 
2227 	if (PP_ISFREE(pp)) {
2228 		panic("hat_memload: loading a mapping to free page %p",
2229 		    (void *)pp);
2230 	}
2231 
2232 	ASSERT((hat == ksfmmup) || AS_LOCK_HELD(hat->sfmmu_as));
2233 
2234 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2235 		cmn_err(CE_NOTE, "hat_memload: unsupported flags %d",
2236 		    flags & ~SFMMU_LOAD_ALLFLAG);
2237 
2238 	if (hat->sfmmu_rmstat)
2239 		hat_resvstat(MMU_PAGESIZE, hat->sfmmu_as, addr);
2240 
2241 #if defined(SF_ERRATA_57)
2242 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2243 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2244 	    !(flags & HAT_LOAD_SHARE)) {
2245 		cmn_err(CE_WARN, "hat_memload: illegal attempt to make user "
2246 		    " page executable");
2247 		attr &= ~PROT_EXEC;
2248 	}
2249 #endif
2250 
2251 	sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2252 	(void) sfmmu_tteload_array(hat, &tte, addr, &pp, flags, rid);
2253 
2254 	/*
2255 	 * Check TSB and TLB page sizes.
2256 	 */
2257 	if ((flags & HAT_LOAD_SHARE) == 0) {
2258 		sfmmu_check_page_sizes(hat, 1);
2259 	}
2260 }
2261 
2262 /*
2263  * hat_devload can be called to map real memory (e.g.
2264  * /dev/kmem) and even though hat_devload will determine pf is
2265  * for memory, it will be unable to get a shared lock on the
2266  * page (because someone else has it exclusively) and will
2267  * pass dp = NULL.  If tteload doesn't get a non-NULL
2268  * page pointer it can't cache memory.
2269  */
2270 void
2271 hat_devload(struct hat *hat, caddr_t addr, size_t len, pfn_t pfn,
2272     uint_t attr, int flags)
2273 {
2274 	tte_t tte;
2275 	struct page *pp = NULL;
2276 	int use_lgpg = 0;
2277 
2278 	ASSERT(hat != NULL);
2279 
2280 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2281 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2282 	ASSERT((hat == ksfmmup) || AS_LOCK_HELD(hat->sfmmu_as));
2283 	if (len == 0)
2284 		panic("hat_devload: zero len");
2285 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2286 		cmn_err(CE_NOTE, "hat_devload: unsupported flags %d",
2287 		    flags & ~SFMMU_LOAD_ALLFLAG);
2288 
2289 #if defined(SF_ERRATA_57)
2290 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2291 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2292 	    !(flags & HAT_LOAD_SHARE)) {
2293 		cmn_err(CE_WARN, "hat_devload: illegal attempt to make user "
2294 		    " page executable");
2295 		attr &= ~PROT_EXEC;
2296 	}
2297 #endif
2298 
2299 	/*
2300 	 * If it's a memory page find its pp
2301 	 */
2302 	if (!(flags & HAT_LOAD_NOCONSIST) && pf_is_memory(pfn)) {
2303 		pp = page_numtopp_nolock(pfn);
2304 		if (pp == NULL) {
2305 			flags |= HAT_LOAD_NOCONSIST;
2306 		} else {
2307 			if (PP_ISFREE(pp)) {
2308 				panic("hat_memload: loading "
2309 				    "a mapping to free page %p",
2310 				    (void *)pp);
2311 			}
2312 			if (!PAGE_LOCKED(pp) && !PP_ISNORELOC(pp)) {
2313 				panic("hat_memload: loading a mapping "
2314 				    "to unlocked relocatable page %p",
2315 				    (void *)pp);
2316 			}
2317 			ASSERT(len == MMU_PAGESIZE);
2318 		}
2319 	}
2320 
2321 	if (hat->sfmmu_rmstat)
2322 		hat_resvstat(len, hat->sfmmu_as, addr);
2323 
2324 	if (flags & HAT_LOAD_NOCONSIST) {
2325 		attr |= SFMMU_UNCACHEVTTE;
2326 		use_lgpg = 1;
2327 	}
2328 	if (!pf_is_memory(pfn)) {
2329 		attr |= SFMMU_UNCACHEPTTE | HAT_NOSYNC;
2330 		use_lgpg = 1;
2331 		switch (attr & HAT_ORDER_MASK) {
2332 			case HAT_STRICTORDER:
2333 			case HAT_UNORDERED_OK:
2334 				/*
2335 				 * we set the side effect bit for all non
2336 				 * memory mappings unless merging is ok
2337 				 */
2338 				attr |= SFMMU_SIDEFFECT;
2339 				break;
2340 			case HAT_MERGING_OK:
2341 			case HAT_LOADCACHING_OK:
2342 			case HAT_STORECACHING_OK:
2343 				break;
2344 			default:
2345 				panic("hat_devload: bad attr");
2346 				break;
2347 		}
2348 	}
2349 	while (len) {
2350 		if (!use_lgpg) {
2351 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2352 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2353 			    flags, SFMMU_INVALID_SHMERID);
2354 			len -= MMU_PAGESIZE;
2355 			addr += MMU_PAGESIZE;
2356 			pfn++;
2357 			continue;
2358 		}
2359 		/*
2360 		 *  try to use large pages, check va/pa alignments
2361 		 *  Note that 32M/256M page sizes are not (yet) supported.
2362 		 */
2363 		if ((len >= MMU_PAGESIZE4M) &&
2364 		    !((uintptr_t)addr & MMU_PAGEOFFSET4M) &&
2365 		    !(disable_large_pages & (1 << TTE4M)) &&
2366 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET4M)) {
2367 			sfmmu_memtte(&tte, pfn, attr, TTE4M);
2368 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2369 			    flags, SFMMU_INVALID_SHMERID);
2370 			len -= MMU_PAGESIZE4M;
2371 			addr += MMU_PAGESIZE4M;
2372 			pfn += MMU_PAGESIZE4M / MMU_PAGESIZE;
2373 		} else if ((len >= MMU_PAGESIZE512K) &&
2374 		    !((uintptr_t)addr & MMU_PAGEOFFSET512K) &&
2375 		    !(disable_large_pages & (1 << TTE512K)) &&
2376 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET512K)) {
2377 			sfmmu_memtte(&tte, pfn, attr, TTE512K);
2378 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2379 			    flags, SFMMU_INVALID_SHMERID);
2380 			len -= MMU_PAGESIZE512K;
2381 			addr += MMU_PAGESIZE512K;
2382 			pfn += MMU_PAGESIZE512K / MMU_PAGESIZE;
2383 		} else if ((len >= MMU_PAGESIZE64K) &&
2384 		    !((uintptr_t)addr & MMU_PAGEOFFSET64K) &&
2385 		    !(disable_large_pages & (1 << TTE64K)) &&
2386 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET64K)) {
2387 			sfmmu_memtte(&tte, pfn, attr, TTE64K);
2388 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2389 			    flags, SFMMU_INVALID_SHMERID);
2390 			len -= MMU_PAGESIZE64K;
2391 			addr += MMU_PAGESIZE64K;
2392 			pfn += MMU_PAGESIZE64K / MMU_PAGESIZE;
2393 		} else {
2394 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2395 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2396 			    flags, SFMMU_INVALID_SHMERID);
2397 			len -= MMU_PAGESIZE;
2398 			addr += MMU_PAGESIZE;
2399 			pfn++;
2400 		}
2401 	}
2402 
2403 	/*
2404 	 * Check TSB and TLB page sizes.
2405 	 */
2406 	if ((flags & HAT_LOAD_SHARE) == 0) {
2407 		sfmmu_check_page_sizes(hat, 1);
2408 	}
2409 }
2410 
2411 void
2412 hat_memload_array(struct hat *hat, caddr_t addr, size_t len,
2413     struct page **pps, uint_t attr, uint_t flags)
2414 {
2415 	hat_do_memload_array(hat, addr, len, pps, attr, flags,
2416 	    SFMMU_INVALID_SHMERID);
2417 }
2418 
2419 void
2420 hat_memload_array_region(struct hat *hat, caddr_t addr, size_t len,
2421     struct page **pps, uint_t attr, uint_t flags,
2422     hat_region_cookie_t rcookie)
2423 {
2424 	uint_t rid;
2425 	if (rcookie == HAT_INVALID_REGION_COOKIE) {
2426 		hat_do_memload_array(hat, addr, len, pps, attr, flags,
2427 		    SFMMU_INVALID_SHMERID);
2428 		return;
2429 	}
2430 	rid = (uint_t)((uint64_t)rcookie);
2431 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2432 	hat_do_memload_array(hat, addr, len, pps, attr, flags, rid);
2433 }
2434 
2435 /*
2436  * Map the largest extend possible out of the page array. The array may NOT
2437  * be in order.  The largest possible mapping a page can have
2438  * is specified in the p_szc field.  The p_szc field
2439  * cannot change as long as there any mappings (large or small)
2440  * to any of the pages that make up the large page. (ie. any
2441  * promotion/demotion of page size is not up to the hat but up to
2442  * the page free list manager).  The array
2443  * should consist of properly aligned contigous pages that are
2444  * part of a big page for a large mapping to be created.
2445  */
2446 static void
2447 hat_do_memload_array(struct hat *hat, caddr_t addr, size_t len,
2448     struct page **pps, uint_t attr, uint_t flags, uint_t rid)
2449 {
2450 	int  ttesz;
2451 	size_t mapsz;
2452 	pgcnt_t	numpg, npgs;
2453 	tte_t tte;
2454 	page_t *pp;
2455 	uint_t large_pages_disable;
2456 
2457 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2458 	SFMMU_VALIDATE_HMERID(hat, rid, addr, len);
2459 
2460 	if (hat->sfmmu_rmstat)
2461 		hat_resvstat(len, hat->sfmmu_as, addr);
2462 
2463 #if defined(SF_ERRATA_57)
2464 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2465 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2466 	    !(flags & HAT_LOAD_SHARE)) {
2467 		cmn_err(CE_WARN, "hat_memload_array: illegal attempt to make "
2468 		    "user page executable");
2469 		attr &= ~PROT_EXEC;
2470 	}
2471 #endif
2472 
2473 	/* Get number of pages */
2474 	npgs = len >> MMU_PAGESHIFT;
2475 
2476 	if (flags & HAT_LOAD_SHARE) {
2477 		large_pages_disable = disable_ism_large_pages;
2478 	} else {
2479 		large_pages_disable = disable_large_pages;
2480 	}
2481 
2482 	if (npgs < NHMENTS || large_pages_disable == LARGE_PAGES_OFF) {
2483 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2484 		    rid);
2485 		return;
2486 	}
2487 
2488 	while (npgs >= NHMENTS) {
2489 		pp = *pps;
2490 		for (ttesz = pp->p_szc; ttesz != TTE8K; ttesz--) {
2491 			/*
2492 			 * Check if this page size is disabled.
2493 			 */
2494 			if (large_pages_disable & (1 << ttesz))
2495 				continue;
2496 
2497 			numpg = TTEPAGES(ttesz);
2498 			mapsz = numpg << MMU_PAGESHIFT;
2499 			if ((npgs >= numpg) &&
2500 			    IS_P2ALIGNED(addr, mapsz) &&
2501 			    IS_P2ALIGNED(pp->p_pagenum, numpg)) {
2502 				/*
2503 				 * At this point we have enough pages and
2504 				 * we know the virtual address and the pfn
2505 				 * are properly aligned.  We still need
2506 				 * to check for physical contiguity but since
2507 				 * it is very likely that this is the case
2508 				 * we will assume they are so and undo
2509 				 * the request if necessary.  It would
2510 				 * be great if we could get a hint flag
2511 				 * like HAT_CONTIG which would tell us
2512 				 * the pages are contigous for sure.
2513 				 */
2514 				sfmmu_memtte(&tte, (*pps)->p_pagenum,
2515 				    attr, ttesz);
2516 				if (!sfmmu_tteload_array(hat, &tte, addr,
2517 				    pps, flags, rid)) {
2518 					break;
2519 				}
2520 			}
2521 		}
2522 		if (ttesz == TTE8K) {
2523 			/*
2524 			 * We were not able to map array using a large page
2525 			 * batch a hmeblk or fraction at a time.
2526 			 */
2527 			numpg = ((uintptr_t)addr >> MMU_PAGESHIFT)
2528 			    & (NHMENTS-1);
2529 			numpg = NHMENTS - numpg;
2530 			ASSERT(numpg <= npgs);
2531 			mapsz = numpg * MMU_PAGESIZE;
2532 			sfmmu_memload_batchsmall(hat, addr, pps, attr, flags,
2533 			    numpg, rid);
2534 		}
2535 		addr += mapsz;
2536 		npgs -= numpg;
2537 		pps += numpg;
2538 	}
2539 
2540 	if (npgs) {
2541 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2542 		    rid);
2543 	}
2544 
2545 	/*
2546 	 * Check TSB and TLB page sizes.
2547 	 */
2548 	if ((flags & HAT_LOAD_SHARE) == 0) {
2549 		sfmmu_check_page_sizes(hat, 1);
2550 	}
2551 }
2552 
2553 /*
2554  * Function tries to batch 8K pages into the same hme blk.
2555  */
2556 static void
2557 sfmmu_memload_batchsmall(struct hat *hat, caddr_t vaddr, page_t **pps,
2558     uint_t attr, uint_t flags, pgcnt_t npgs, uint_t rid)
2559 {
2560 	tte_t	tte;
2561 	page_t *pp;
2562 	struct hmehash_bucket *hmebp;
2563 	struct hme_blk *hmeblkp;
2564 	int	index;
2565 
2566 	while (npgs) {
2567 		/*
2568 		 * Acquire the hash bucket.
2569 		 */
2570 		hmebp = sfmmu_tteload_acquire_hashbucket(hat, vaddr, TTE8K,
2571 		    rid);
2572 		ASSERT(hmebp);
2573 
2574 		/*
2575 		 * Find the hment block.
2576 		 */
2577 		hmeblkp = sfmmu_tteload_find_hmeblk(hat, hmebp, vaddr,
2578 		    TTE8K, flags, rid);
2579 		ASSERT(hmeblkp);
2580 
2581 		do {
2582 			/*
2583 			 * Make the tte.
2584 			 */
2585 			pp = *pps;
2586 			sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2587 
2588 			/*
2589 			 * Add the translation.
2590 			 */
2591 			(void) sfmmu_tteload_addentry(hat, hmeblkp, &tte,
2592 			    vaddr, pps, flags, rid);
2593 
2594 			/*
2595 			 * Goto next page.
2596 			 */
2597 			pps++;
2598 			npgs--;
2599 
2600 			/*
2601 			 * Goto next address.
2602 			 */
2603 			vaddr += MMU_PAGESIZE;
2604 
2605 			/*
2606 			 * Don't crossover into a different hmentblk.
2607 			 */
2608 			index = (int)(((uintptr_t)vaddr >> MMU_PAGESHIFT) &
2609 			    (NHMENTS-1));
2610 
2611 		} while (index != 0 && npgs != 0);
2612 
2613 		/*
2614 		 * Release the hash bucket.
2615 		 */
2616 
2617 		sfmmu_tteload_release_hashbucket(hmebp);
2618 	}
2619 }
2620 
2621 /*
2622  * Construct a tte for a page:
2623  *
2624  * tte_valid = 1
2625  * tte_size2 = size & TTE_SZ2_BITS (Panther and Olympus-C only)
2626  * tte_size = size
2627  * tte_nfo = attr & HAT_NOFAULT
2628  * tte_ie = attr & HAT_STRUCTURE_LE
2629  * tte_hmenum = hmenum
2630  * tte_pahi = pp->p_pagenum >> TTE_PASHIFT;
2631  * tte_palo = pp->p_pagenum & TTE_PALOMASK;
2632  * tte_ref = 1 (optimization)
2633  * tte_wr_perm = attr & PROT_WRITE;
2634  * tte_no_sync = attr & HAT_NOSYNC
2635  * tte_lock = attr & SFMMU_LOCKTTE
2636  * tte_cp = !(attr & SFMMU_UNCACHEPTTE)
2637  * tte_cv = !(attr & SFMMU_UNCACHEVTTE)
2638  * tte_e = attr & SFMMU_SIDEFFECT
2639  * tte_priv = !(attr & PROT_USER)
2640  * tte_hwwr = if nosync is set and it is writable we set the mod bit (opt)
2641  * tte_glb = 0
2642  */
2643 void
2644 sfmmu_memtte(tte_t *ttep, pfn_t pfn, uint_t attr, int tte_sz)
2645 {
2646 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2647 
2648 	ttep->tte_inthi = MAKE_TTE_INTHI(pfn, attr, tte_sz, 0 /* hmenum */);
2649 	ttep->tte_intlo = MAKE_TTE_INTLO(pfn, attr, tte_sz, 0 /* hmenum */);
2650 
2651 	if (TTE_IS_NOSYNC(ttep)) {
2652 		TTE_SET_REF(ttep);
2653 		if (TTE_IS_WRITABLE(ttep)) {
2654 			TTE_SET_MOD(ttep);
2655 		}
2656 	}
2657 	if (TTE_IS_NFO(ttep) && TTE_IS_EXECUTABLE(ttep)) {
2658 		panic("sfmmu_memtte: can't set both NFO and EXEC bits");
2659 	}
2660 }
2661 
2662 /*
2663  * This function will add a translation to the hme_blk and allocate the
2664  * hme_blk if one does not exist.
2665  * If a page structure is specified then it will add the
2666  * corresponding hment to the mapping list.
2667  * It will also update the hmenum field for the tte.
2668  *
2669  * Currently this function is only used for kernel mappings.
2670  * So pass invalid region to sfmmu_tteload_array().
2671  */
2672 void
2673 sfmmu_tteload(struct hat *sfmmup, tte_t *ttep, caddr_t vaddr, page_t *pp,
2674     uint_t flags)
2675 {
2676 	ASSERT(sfmmup == ksfmmup);
2677 	(void) sfmmu_tteload_array(sfmmup, ttep, vaddr, &pp, flags,
2678 	    SFMMU_INVALID_SHMERID);
2679 }
2680 
2681 /*
2682  * Load (ttep != NULL) or unload (ttep == NULL) one entry in the TSB.
2683  * Assumes that a particular page size may only be resident in one TSB.
2684  */
2685 static void
2686 sfmmu_mod_tsb(sfmmu_t *sfmmup, caddr_t vaddr, tte_t *ttep, int ttesz)
2687 {
2688 	struct tsb_info *tsbinfop = NULL;
2689 	uint64_t tag;
2690 	struct tsbe *tsbe_addr;
2691 	uint64_t tsb_base;
2692 	uint_t tsb_size;
2693 	int vpshift = MMU_PAGESHIFT;
2694 	int phys = 0;
2695 
2696 	if (sfmmup == ksfmmup) { /* No support for 32/256M ksfmmu pages */
2697 		phys = ktsb_phys;
2698 		if (ttesz >= TTE4M) {
2699 #ifndef sun4v
2700 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2701 #endif
2702 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2703 			tsb_size = ktsb4m_szcode;
2704 		} else {
2705 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2706 			tsb_size = ktsb_szcode;
2707 		}
2708 	} else {
2709 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2710 
2711 		/*
2712 		 * If there isn't a TSB for this page size, or the TSB is
2713 		 * swapped out, there is nothing to do.  Note that the latter
2714 		 * case seems impossible but can occur if hat_pageunload()
2715 		 * is called on an ISM mapping while the process is swapped
2716 		 * out.
2717 		 */
2718 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2719 			return;
2720 
2721 		/*
2722 		 * If another thread is in the middle of relocating a TSB
2723 		 * we can't unload the entry so set a flag so that the
2724 		 * TSB will be flushed before it can be accessed by the
2725 		 * process.
2726 		 */
2727 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2728 			if (ttep == NULL)
2729 				tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2730 			return;
2731 		}
2732 #if defined(UTSB_PHYS)
2733 		phys = 1;
2734 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2735 #else
2736 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2737 #endif
2738 		tsb_size = tsbinfop->tsb_szc;
2739 	}
2740 	if (ttesz >= TTE4M)
2741 		vpshift = MMU_PAGESHIFT4M;
2742 
2743 	tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2744 	tag = sfmmu_make_tsbtag(vaddr);
2745 
2746 	if (ttep == NULL) {
2747 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2748 	} else {
2749 		if (ttesz >= TTE4M) {
2750 			SFMMU_STAT(sf_tsb_load4m);
2751 		} else {
2752 			SFMMU_STAT(sf_tsb_load8k);
2753 		}
2754 
2755 		sfmmu_load_tsbe(tsbe_addr, tag, ttep, phys);
2756 	}
2757 }
2758 
2759 /*
2760  * Unmap all entries from [start, end) matching the given page size.
2761  *
2762  * This function is used primarily to unmap replicated 64K or 512K entries
2763  * from the TSB that are inserted using the base page size TSB pointer, but
2764  * it may also be called to unmap a range of addresses from the TSB.
2765  */
2766 void
2767 sfmmu_unload_tsb_range(sfmmu_t *sfmmup, caddr_t start, caddr_t end, int ttesz)
2768 {
2769 	struct tsb_info *tsbinfop;
2770 	uint64_t tag;
2771 	struct tsbe *tsbe_addr;
2772 	caddr_t vaddr;
2773 	uint64_t tsb_base;
2774 	int vpshift, vpgsz;
2775 	uint_t tsb_size;
2776 	int phys = 0;
2777 
2778 	/*
2779 	 * Assumptions:
2780 	 *  If ttesz == 8K, 64K or 512K, we walk through the range 8K
2781 	 *  at a time shooting down any valid entries we encounter.
2782 	 *
2783 	 *  If ttesz >= 4M we walk the range 4M at a time shooting
2784 	 *  down any valid mappings we find.
2785 	 */
2786 	if (sfmmup == ksfmmup) {
2787 		phys = ktsb_phys;
2788 		if (ttesz >= TTE4M) {
2789 #ifndef sun4v
2790 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2791 #endif
2792 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2793 			tsb_size = ktsb4m_szcode;
2794 		} else {
2795 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2796 			tsb_size = ktsb_szcode;
2797 		}
2798 	} else {
2799 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2800 
2801 		/*
2802 		 * If there isn't a TSB for this page size, or the TSB is
2803 		 * swapped out, there is nothing to do.  Note that the latter
2804 		 * case seems impossible but can occur if hat_pageunload()
2805 		 * is called on an ISM mapping while the process is swapped
2806 		 * out.
2807 		 */
2808 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2809 			return;
2810 
2811 		/*
2812 		 * If another thread is in the middle of relocating a TSB
2813 		 * we can't unload the entry so set a flag so that the
2814 		 * TSB will be flushed before it can be accessed by the
2815 		 * process.
2816 		 */
2817 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2818 			tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2819 			return;
2820 		}
2821 #if defined(UTSB_PHYS)
2822 		phys = 1;
2823 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2824 #else
2825 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2826 #endif
2827 		tsb_size = tsbinfop->tsb_szc;
2828 	}
2829 	if (ttesz >= TTE4M) {
2830 		vpshift = MMU_PAGESHIFT4M;
2831 		vpgsz = MMU_PAGESIZE4M;
2832 	} else {
2833 		vpshift = MMU_PAGESHIFT;
2834 		vpgsz = MMU_PAGESIZE;
2835 	}
2836 
2837 	for (vaddr = start; vaddr < end; vaddr += vpgsz) {
2838 		tag = sfmmu_make_tsbtag(vaddr);
2839 		tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2840 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2841 	}
2842 }
2843 
2844 /*
2845  * Select the optimum TSB size given the number of mappings
2846  * that need to be cached.
2847  */
2848 static int
2849 sfmmu_select_tsb_szc(pgcnt_t pgcnt)
2850 {
2851 	int szc = 0;
2852 
2853 #ifdef DEBUG
2854 	if (tsb_grow_stress) {
2855 		uint32_t randval = (uint32_t)gettick() >> 4;
2856 		return (randval % (tsb_max_growsize + 1));
2857 	}
2858 #endif	/* DEBUG */
2859 
2860 	while ((szc < tsb_max_growsize) && (pgcnt > SFMMU_RSS_TSBSIZE(szc)))
2861 		szc++;
2862 	return (szc);
2863 }
2864 
2865 /*
2866  * This function will add a translation to the hme_blk and allocate the
2867  * hme_blk if one does not exist.
2868  * If a page structure is specified then it will add the
2869  * corresponding hment to the mapping list.
2870  * It will also update the hmenum field for the tte.
2871  * Furthermore, it attempts to create a large page translation
2872  * for <addr,hat> at page array pps.  It assumes addr and first
2873  * pp is correctly aligned.  It returns 0 if successful and 1 otherwise.
2874  */
2875 static int
2876 sfmmu_tteload_array(sfmmu_t *sfmmup, tte_t *ttep, caddr_t vaddr,
2877     page_t **pps, uint_t flags, uint_t rid)
2878 {
2879 	struct hmehash_bucket *hmebp;
2880 	struct hme_blk *hmeblkp;
2881 	int	ret;
2882 	uint_t	size;
2883 
2884 	/*
2885 	 * Get mapping size.
2886 	 */
2887 	size = TTE_CSZ(ttep);
2888 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
2889 
2890 	/*
2891 	 * Acquire the hash bucket.
2892 	 */
2893 	hmebp = sfmmu_tteload_acquire_hashbucket(sfmmup, vaddr, size, rid);
2894 	ASSERT(hmebp);
2895 
2896 	/*
2897 	 * Find the hment block.
2898 	 */
2899 	hmeblkp = sfmmu_tteload_find_hmeblk(sfmmup, hmebp, vaddr, size, flags,
2900 	    rid);
2901 	ASSERT(hmeblkp);
2902 
2903 	/*
2904 	 * Add the translation.
2905 	 */
2906 	ret = sfmmu_tteload_addentry(sfmmup, hmeblkp, ttep, vaddr, pps, flags,
2907 	    rid);
2908 
2909 	/*
2910 	 * Release the hash bucket.
2911 	 */
2912 	sfmmu_tteload_release_hashbucket(hmebp);
2913 
2914 	return (ret);
2915 }
2916 
2917 /*
2918  * Function locks and returns a pointer to the hash bucket for vaddr and size.
2919  */
2920 static struct hmehash_bucket *
2921 sfmmu_tteload_acquire_hashbucket(sfmmu_t *sfmmup, caddr_t vaddr, int size,
2922     uint_t rid)
2923 {
2924 	struct hmehash_bucket *hmebp;
2925 	int hmeshift;
2926 	void *htagid = sfmmutohtagid(sfmmup, rid);
2927 
2928 	ASSERT(htagid != NULL);
2929 
2930 	hmeshift = HME_HASH_SHIFT(size);
2931 
2932 	hmebp = HME_HASH_FUNCTION(htagid, vaddr, hmeshift);
2933 
2934 	SFMMU_HASH_LOCK(hmebp);
2935 
2936 	return (hmebp);
2937 }
2938 
2939 /*
2940  * Function returns a pointer to an hmeblk in the hash bucket, hmebp. If the
2941  * hmeblk doesn't exists for the [sfmmup, vaddr & size] signature, a hmeblk is
2942  * allocated.
2943  */
2944 static struct hme_blk *
2945 sfmmu_tteload_find_hmeblk(sfmmu_t *sfmmup, struct hmehash_bucket *hmebp,
2946     caddr_t vaddr, uint_t size, uint_t flags, uint_t rid)
2947 {
2948 	hmeblk_tag hblktag;
2949 	int hmeshift;
2950 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
2951 
2952 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
2953 
2954 	hblktag.htag_id = sfmmutohtagid(sfmmup, rid);
2955 	ASSERT(hblktag.htag_id != NULL);
2956 	hmeshift = HME_HASH_SHIFT(size);
2957 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
2958 	hblktag.htag_rehash = HME_HASH_REHASH(size);
2959 	hblktag.htag_rid = rid;
2960 
2961 ttearray_realloc:
2962 
2963 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
2964 
2965 	/*
2966 	 * We block until hblk_reserve_lock is released; it's held by
2967 	 * the thread, temporarily using hblk_reserve, until hblk_reserve is
2968 	 * replaced by a hblk from sfmmu8_cache.
2969 	 */
2970 	if (hmeblkp == (struct hme_blk *)hblk_reserve &&
2971 	    hblk_reserve_thread != curthread) {
2972 		SFMMU_HASH_UNLOCK(hmebp);
2973 		mutex_enter(&hblk_reserve_lock);
2974 		mutex_exit(&hblk_reserve_lock);
2975 		SFMMU_STAT(sf_hblk_reserve_hit);
2976 		SFMMU_HASH_LOCK(hmebp);
2977 		goto ttearray_realloc;
2978 	}
2979 
2980 	if (hmeblkp == NULL) {
2981 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
2982 		    hblktag, flags, rid);
2983 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
2984 		ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
2985 	} else {
2986 		/*
2987 		 * It is possible for 8k and 64k hblks to collide since they
2988 		 * have the same rehash value. This is because we
2989 		 * lazily free hblks and 8K/64K blks could be lingering.
2990 		 * If we find size mismatch we free the block and & try again.
2991 		 */
2992 		if (get_hblk_ttesz(hmeblkp) != size) {
2993 			ASSERT(!hmeblkp->hblk_vcnt);
2994 			ASSERT(!hmeblkp->hblk_hmecnt);
2995 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
2996 			    &list, 0);
2997 			goto ttearray_realloc;
2998 		}
2999 		if (hmeblkp->hblk_shw_bit) {
3000 			/*
3001 			 * if the hblk was previously used as a shadow hblk then
3002 			 * we will change it to a normal hblk
3003 			 */
3004 			ASSERT(!hmeblkp->hblk_shared);
3005 			if (hmeblkp->hblk_shw_mask) {
3006 				sfmmu_shadow_hcleanup(sfmmup, hmeblkp, hmebp);
3007 				ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3008 				goto ttearray_realloc;
3009 			} else {
3010 				hmeblkp->hblk_shw_bit = 0;
3011 			}
3012 		}
3013 		SFMMU_STAT(sf_hblk_hit);
3014 	}
3015 
3016 	/*
3017 	 * hat_memload() should never call kmem_cache_free() for kernel hmeblks;
3018 	 * see block comment showing the stacktrace in sfmmu_hblk_alloc();
3019 	 * set the flag parameter to 1 so that sfmmu_hblks_list_purge() will
3020 	 * just add these hmeblks to the per-cpu pending queue.
3021 	 */
3022 	sfmmu_hblks_list_purge(&list, 1);
3023 
3024 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
3025 	ASSERT(!hmeblkp->hblk_shw_bit);
3026 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3027 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3028 	ASSERT(hmeblkp->hblk_tag.htag_rid == rid);
3029 
3030 	return (hmeblkp);
3031 }
3032 
3033 /*
3034  * Function adds a tte entry into the hmeblk. It returns 0 if successful and 1
3035  * otherwise.
3036  */
3037 static int
3038 sfmmu_tteload_addentry(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, tte_t *ttep,
3039     caddr_t vaddr, page_t **pps, uint_t flags, uint_t rid)
3040 {
3041 	page_t *pp = *pps;
3042 	int hmenum, size, remap;
3043 	tte_t tteold, flush_tte;
3044 #ifdef DEBUG
3045 	tte_t orig_old;
3046 #endif /* DEBUG */
3047 	struct sf_hment *sfhme;
3048 	kmutex_t *pml, *pmtx;
3049 	hatlock_t *hatlockp;
3050 	int myflt;
3051 
3052 	/*
3053 	 * remove this panic when we decide to let user virtual address
3054 	 * space be >= USERLIMIT.
3055 	 */
3056 	if (!TTE_IS_PRIVILEGED(ttep) && vaddr >= (caddr_t)USERLIMIT)
3057 		panic("user addr %p in kernel space", (void *)vaddr);
3058 #if defined(TTE_IS_GLOBAL)
3059 	if (TTE_IS_GLOBAL(ttep))
3060 		panic("sfmmu_tteload: creating global tte");
3061 #endif
3062 
3063 #ifdef DEBUG
3064 	if (pf_is_memory(sfmmu_ttetopfn(ttep, vaddr)) &&
3065 	    !TTE_IS_PCACHEABLE(ttep) && !sfmmu_allow_nc_trans)
3066 		panic("sfmmu_tteload: non cacheable memory tte");
3067 #endif /* DEBUG */
3068 
3069 	/* don't simulate dirty bit for writeable ISM/DISM mappings */
3070 	if ((flags & HAT_LOAD_SHARE) && TTE_IS_WRITABLE(ttep)) {
3071 		TTE_SET_REF(ttep);
3072 		TTE_SET_MOD(ttep);
3073 	}
3074 
3075 	if ((flags & HAT_LOAD_SHARE) || !TTE_IS_REF(ttep) ||
3076 	    !TTE_IS_MOD(ttep)) {
3077 		/*
3078 		 * Don't load TSB for dummy as in ISM.  Also don't preload
3079 		 * the TSB if the TTE isn't writable since we're likely to
3080 		 * fault on it again -- preloading can be fairly expensive.
3081 		 */
3082 		flags |= SFMMU_NO_TSBLOAD;
3083 	}
3084 
3085 	size = TTE_CSZ(ttep);
3086 	switch (size) {
3087 	case TTE8K:
3088 		SFMMU_STAT(sf_tteload8k);
3089 		break;
3090 	case TTE64K:
3091 		SFMMU_STAT(sf_tteload64k);
3092 		break;
3093 	case TTE512K:
3094 		SFMMU_STAT(sf_tteload512k);
3095 		break;
3096 	case TTE4M:
3097 		SFMMU_STAT(sf_tteload4m);
3098 		break;
3099 	case (TTE32M):
3100 		SFMMU_STAT(sf_tteload32m);
3101 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3102 		break;
3103 	case (TTE256M):
3104 		SFMMU_STAT(sf_tteload256m);
3105 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3106 		break;
3107 	}
3108 
3109 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
3110 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
3111 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3112 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3113 
3114 	HBLKTOHME_IDX(sfhme, hmeblkp, vaddr, hmenum);
3115 
3116 	/*
3117 	 * Need to grab mlist lock here so that pageunload
3118 	 * will not change tte behind us.
3119 	 */
3120 	if (pp) {
3121 		pml = sfmmu_mlist_enter(pp);
3122 	}
3123 
3124 	sfmmu_copytte(&sfhme->hme_tte, &tteold);
3125 	/*
3126 	 * Look for corresponding hment and if valid verify
3127 	 * pfns are equal.
3128 	 */
3129 	remap = TTE_IS_VALID(&tteold);
3130 	if (remap) {
3131 		pfn_t	new_pfn, old_pfn;
3132 
3133 		old_pfn = TTE_TO_PFN(vaddr, &tteold);
3134 		new_pfn = TTE_TO_PFN(vaddr, ttep);
3135 
3136 		if (flags & HAT_LOAD_REMAP) {
3137 			/* make sure we are remapping same type of pages */
3138 			if (pf_is_memory(old_pfn) != pf_is_memory(new_pfn)) {
3139 				panic("sfmmu_tteload - tte remap io<->memory");
3140 			}
3141 			if (old_pfn != new_pfn &&
3142 			    (pp != NULL || sfhme->hme_page != NULL)) {
3143 				panic("sfmmu_tteload - tte remap pp != NULL");
3144 			}
3145 		} else if (old_pfn != new_pfn) {
3146 			panic("sfmmu_tteload - tte remap, hmeblkp 0x%p",
3147 			    (void *)hmeblkp);
3148 		}
3149 		ASSERT(TTE_CSZ(&tteold) == TTE_CSZ(ttep));
3150 	}
3151 
3152 	if (pp) {
3153 		if (size == TTE8K) {
3154 #ifdef VAC
3155 			/*
3156 			 * Handle VAC consistency
3157 			 */
3158 			if (!remap && (cache & CACHE_VAC) && !PP_ISNC(pp)) {
3159 				sfmmu_vac_conflict(sfmmup, vaddr, pp);
3160 			}
3161 #endif
3162 
3163 			if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3164 				pmtx = sfmmu_page_enter(pp);
3165 				PP_CLRRO(pp);
3166 				sfmmu_page_exit(pmtx);
3167 			} else if (!PP_ISMAPPED(pp) &&
3168 			    (!TTE_IS_WRITABLE(ttep)) && !(PP_ISMOD(pp))) {
3169 				pmtx = sfmmu_page_enter(pp);
3170 				if (!(PP_ISMOD(pp))) {
3171 					PP_SETRO(pp);
3172 				}
3173 				sfmmu_page_exit(pmtx);
3174 			}
3175 
3176 		} else if (sfmmu_pagearray_setup(vaddr, pps, ttep, remap)) {
3177 			/*
3178 			 * sfmmu_pagearray_setup failed so return
3179 			 */
3180 			sfmmu_mlist_exit(pml);
3181 			return (1);
3182 		}
3183 	}
3184 
3185 	/*
3186 	 * Make sure hment is not on a mapping list.
3187 	 */
3188 	ASSERT(remap || (sfhme->hme_page == NULL));
3189 
3190 	/* if it is not a remap then hme->next better be NULL */
3191 	ASSERT((!remap) ? sfhme->hme_next == NULL : 1);
3192 
3193 	if (flags & HAT_LOAD_LOCK) {
3194 		if ((hmeblkp->hblk_lckcnt + 1) >= MAX_HBLK_LCKCNT) {
3195 			panic("too high lckcnt-hmeblk %p",
3196 			    (void *)hmeblkp);
3197 		}
3198 		atomic_inc_32(&hmeblkp->hblk_lckcnt);
3199 
3200 		HBLK_STACK_TRACE(hmeblkp, HBLK_LOCK);
3201 	}
3202 
3203 #ifdef VAC
3204 	if (pp && PP_ISNC(pp)) {
3205 		/*
3206 		 * If the physical page is marked to be uncacheable, like
3207 		 * by a vac conflict, make sure the new mapping is also
3208 		 * uncacheable.
3209 		 */
3210 		TTE_CLR_VCACHEABLE(ttep);
3211 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
3212 	}
3213 #endif
3214 	ttep->tte_hmenum = hmenum;
3215 
3216 #ifdef DEBUG
3217 	orig_old = tteold;
3218 #endif /* DEBUG */
3219 
3220 	while (sfmmu_modifytte_try(&tteold, ttep, &sfhme->hme_tte) < 0) {
3221 		if ((sfmmup == KHATID) &&
3222 		    (flags & (HAT_LOAD_LOCK | HAT_LOAD_REMAP))) {
3223 			sfmmu_copytte(&sfhme->hme_tte, &tteold);
3224 		}
3225 #ifdef DEBUG
3226 		chk_tte(&orig_old, &tteold, ttep, hmeblkp);
3227 #endif /* DEBUG */
3228 	}
3229 	ASSERT(TTE_IS_VALID(&sfhme->hme_tte));
3230 
3231 	if (!TTE_IS_VALID(&tteold)) {
3232 
3233 		atomic_inc_16(&hmeblkp->hblk_vcnt);
3234 		if (rid == SFMMU_INVALID_SHMERID) {
3235 			atomic_inc_ulong(&sfmmup->sfmmu_ttecnt[size]);
3236 		} else {
3237 			sf_srd_t *srdp = sfmmup->sfmmu_srdp;
3238 			sf_region_t *rgnp = srdp->srd_hmergnp[rid];
3239 			/*
3240 			 * We already accounted for region ttecnt's in sfmmu
3241 			 * during hat_join_region() processing. Here we
3242 			 * only update ttecnt's in region struture.
3243 			 */
3244 			atomic_inc_ulong(&rgnp->rgn_ttecnt[size]);
3245 		}
3246 	}
3247 
3248 	myflt = (astosfmmu(curthread->t_procp->p_as) == sfmmup);
3249 	if (size > TTE8K && (flags & HAT_LOAD_SHARE) == 0 &&
3250 	    sfmmup != ksfmmup) {
3251 		uchar_t tteflag = 1 << size;
3252 		if (rid == SFMMU_INVALID_SHMERID) {
3253 			if (!(sfmmup->sfmmu_tteflags & tteflag)) {
3254 				hatlockp = sfmmu_hat_enter(sfmmup);
3255 				sfmmup->sfmmu_tteflags |= tteflag;
3256 				sfmmu_hat_exit(hatlockp);
3257 			}
3258 		} else if (!(sfmmup->sfmmu_rtteflags & tteflag)) {
3259 			hatlockp = sfmmu_hat_enter(sfmmup);
3260 			sfmmup->sfmmu_rtteflags |= tteflag;
3261 			sfmmu_hat_exit(hatlockp);
3262 		}
3263 		/*
3264 		 * Update the current CPU tsbmiss area, so the current thread
3265 		 * won't need to take the tsbmiss for the new pagesize.
3266 		 * The other threads in the process will update their tsb
3267 		 * miss area lazily in sfmmu_tsbmiss_exception() when they
3268 		 * fail to find the translation for a newly added pagesize.
3269 		 */
3270 		if (size > TTE64K && myflt) {
3271 			struct tsbmiss *tsbmp;
3272 			kpreempt_disable();
3273 			tsbmp = &tsbmiss_area[CPU->cpu_id];
3274 			if (rid == SFMMU_INVALID_SHMERID) {
3275 				if (!(tsbmp->uhat_tteflags & tteflag)) {
3276 					tsbmp->uhat_tteflags |= tteflag;
3277 				}
3278 			} else {
3279 				if (!(tsbmp->uhat_rtteflags & tteflag)) {
3280 					tsbmp->uhat_rtteflags |= tteflag;
3281 				}
3282 			}
3283 			kpreempt_enable();
3284 		}
3285 	}
3286 
3287 	if (size >= TTE4M && (flags & HAT_LOAD_TEXT) &&
3288 	    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
3289 		hatlockp = sfmmu_hat_enter(sfmmup);
3290 		SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
3291 		sfmmu_hat_exit(hatlockp);
3292 	}
3293 
3294 	flush_tte.tte_intlo = (tteold.tte_intlo ^ ttep->tte_intlo) &
3295 	    hw_tte.tte_intlo;
3296 	flush_tte.tte_inthi = (tteold.tte_inthi ^ ttep->tte_inthi) &
3297 	    hw_tte.tte_inthi;
3298 
3299 	if (remap && (flush_tte.tte_inthi || flush_tte.tte_intlo)) {
3300 		/*
3301 		 * If remap and new tte differs from old tte we need
3302 		 * to sync the mod bit and flush TLB/TSB.  We don't
3303 		 * need to sync ref bit because we currently always set
3304 		 * ref bit in tteload.
3305 		 */
3306 		ASSERT(TTE_IS_REF(ttep));
3307 		if (TTE_IS_MOD(&tteold)) {
3308 			sfmmu_ttesync(sfmmup, vaddr, &tteold, pp);
3309 		}
3310 		/*
3311 		 * hwtte bits shouldn't change for SRD hmeblks as long as SRD
3312 		 * hmes are only used for read only text. Adding this code for
3313 		 * completeness and future use of shared hmeblks with writable
3314 		 * mappings of VMODSORT vnodes.
3315 		 */
3316 		if (hmeblkp->hblk_shared) {
3317 			cpuset_t cpuset = sfmmu_rgntlb_demap(vaddr,
3318 			    sfmmup->sfmmu_srdp->srd_hmergnp[rid], hmeblkp, 1);
3319 			xt_sync(cpuset);
3320 			SFMMU_STAT_ADD(sf_region_remap_demap, 1);
3321 		} else {
3322 			sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 0);
3323 			xt_sync(sfmmup->sfmmu_cpusran);
3324 		}
3325 	}
3326 
3327 	if ((flags & SFMMU_NO_TSBLOAD) == 0) {
3328 		/*
3329 		 * We only preload 8K and 4M mappings into the TSB, since
3330 		 * 64K and 512K mappings are replicated and hence don't
3331 		 * have a single, unique TSB entry. Ditto for 32M/256M.
3332 		 */
3333 		if (size == TTE8K || size == TTE4M) {
3334 			sf_scd_t *scdp;
3335 			hatlockp = sfmmu_hat_enter(sfmmup);
3336 			/*
3337 			 * Don't preload private TSB if the mapping is used
3338 			 * by the shctx in the SCD.
3339 			 */
3340 			scdp = sfmmup->sfmmu_scdp;
3341 			if (rid == SFMMU_INVALID_SHMERID || scdp == NULL ||
3342 			    !SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
3343 				sfmmu_load_tsb(sfmmup, vaddr, &sfhme->hme_tte,
3344 				    size);
3345 			}
3346 			sfmmu_hat_exit(hatlockp);
3347 		}
3348 	}
3349 	if (pp) {
3350 		if (!remap) {
3351 			HME_ADD(sfhme, pp);
3352 			atomic_inc_16(&hmeblkp->hblk_hmecnt);
3353 			ASSERT(hmeblkp->hblk_hmecnt > 0);
3354 
3355 			/*
3356 			 * Cannot ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
3357 			 * see pageunload() for comment.
3358 			 */
3359 		}
3360 		sfmmu_mlist_exit(pml);
3361 	}
3362 
3363 	return (0);
3364 }
3365 /*
3366  * Function unlocks hash bucket.
3367  */
3368 static void
3369 sfmmu_tteload_release_hashbucket(struct hmehash_bucket *hmebp)
3370 {
3371 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3372 	SFMMU_HASH_UNLOCK(hmebp);
3373 }
3374 
3375 /*
3376  * function which checks and sets up page array for a large
3377  * translation.  Will set p_vcolor, p_index, p_ro fields.
3378  * Assumes addr and pfnum of first page are properly aligned.
3379  * Will check for physical contiguity. If check fails it return
3380  * non null.
3381  */
3382 static int
3383 sfmmu_pagearray_setup(caddr_t addr, page_t **pps, tte_t *ttep, int remap)
3384 {
3385 	int	i, index, ttesz;
3386 	pfn_t	pfnum;
3387 	pgcnt_t	npgs;
3388 	page_t *pp, *pp1;
3389 	kmutex_t *pmtx;
3390 #ifdef VAC
3391 	int osz;
3392 	int cflags = 0;
3393 	int vac_err = 0;
3394 #endif
3395 	int newidx = 0;
3396 
3397 	ttesz = TTE_CSZ(ttep);
3398 
3399 	ASSERT(ttesz > TTE8K);
3400 
3401 	npgs = TTEPAGES(ttesz);
3402 	index = PAGESZ_TO_INDEX(ttesz);
3403 
3404 	pfnum = (*pps)->p_pagenum;
3405 	ASSERT(IS_P2ALIGNED(pfnum, npgs));
3406 
3407 	/*
3408 	 * Save the first pp so we can do HAT_TMPNC at the end.
3409 	 */
3410 	pp1 = *pps;
3411 #ifdef VAC
3412 	osz = fnd_mapping_sz(pp1);
3413 #endif
3414 
3415 	for (i = 0; i < npgs; i++, pps++) {
3416 		pp = *pps;
3417 		ASSERT(PAGE_LOCKED(pp));
3418 		ASSERT(pp->p_szc >= ttesz);
3419 		ASSERT(pp->p_szc == pp1->p_szc);
3420 		ASSERT(sfmmu_mlist_held(pp));
3421 
3422 		/*
3423 		 * XXX is it possible to maintain P_RO on the root only?
3424 		 */
3425 		if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3426 			pmtx = sfmmu_page_enter(pp);
3427 			PP_CLRRO(pp);
3428 			sfmmu_page_exit(pmtx);
3429 		} else if (!PP_ISMAPPED(pp) && !TTE_IS_WRITABLE(ttep) &&
3430 		    !PP_ISMOD(pp)) {
3431 			pmtx = sfmmu_page_enter(pp);
3432 			if (!(PP_ISMOD(pp))) {
3433 				PP_SETRO(pp);
3434 			}
3435 			sfmmu_page_exit(pmtx);
3436 		}
3437 
3438 		/*
3439 		 * If this is a remap we skip vac & contiguity checks.
3440 		 */
3441 		if (remap)
3442 			continue;
3443 
3444 		/*
3445 		 * set p_vcolor and detect any vac conflicts.
3446 		 */
3447 #ifdef VAC
3448 		if (vac_err == 0) {
3449 			vac_err = sfmmu_vacconflict_array(addr, pp, &cflags);
3450 
3451 		}
3452 #endif
3453 
3454 		/*
3455 		 * Save current index in case we need to undo it.
3456 		 * Note: "PAGESZ_TO_INDEX(sz)	(1 << (sz))"
3457 		 *	"SFMMU_INDEX_SHIFT	6"
3458 		 *	 "SFMMU_INDEX_MASK	((1 << SFMMU_INDEX_SHIFT) - 1)"
3459 		 *	 "PP_MAPINDEX(p_index)	(p_index & SFMMU_INDEX_MASK)"
3460 		 *
3461 		 * So:	index = PAGESZ_TO_INDEX(ttesz);
3462 		 *	if ttesz == 1 then index = 0x2
3463 		 *		    2 then index = 0x4
3464 		 *		    3 then index = 0x8
3465 		 *		    4 then index = 0x10
3466 		 *		    5 then index = 0x20
3467 		 * The code below checks if it's a new pagesize (ie, newidx)
3468 		 * in case we need to take it back out of p_index,
3469 		 * and then or's the new index into the existing index.
3470 		 */
3471 		if ((PP_MAPINDEX(pp) & index) == 0)
3472 			newidx = 1;
3473 		pp->p_index = (PP_MAPINDEX(pp) | index);
3474 
3475 		/*
3476 		 * contiguity check
3477 		 */
3478 		if (pp->p_pagenum != pfnum) {
3479 			/*
3480 			 * If we fail the contiguity test then
3481 			 * the only thing we need to fix is the p_index field.
3482 			 * We might get a few extra flushes but since this
3483 			 * path is rare that is ok.  The p_ro field will
3484 			 * get automatically fixed on the next tteload to
3485 			 * the page.  NO TNC bit is set yet.
3486 			 */
3487 			while (i >= 0) {
3488 				pp = *pps;
3489 				if (newidx)
3490 					pp->p_index = (PP_MAPINDEX(pp) &
3491 					    ~index);
3492 				pps--;
3493 				i--;
3494 			}
3495 			return (1);
3496 		}
3497 		pfnum++;
3498 		addr += MMU_PAGESIZE;
3499 	}
3500 
3501 #ifdef VAC
3502 	if (vac_err) {
3503 		if (ttesz > osz) {
3504 			/*
3505 			 * There are some smaller mappings that causes vac
3506 			 * conflicts. Convert all existing small mappings to
3507 			 * TNC.
3508 			 */
3509 			SFMMU_STAT_ADD(sf_uncache_conflict, npgs);
3510 			sfmmu_page_cache_array(pp1, HAT_TMPNC, CACHE_FLUSH,
3511 			    npgs);
3512 		} else {
3513 			/* EMPTY */
3514 			/*
3515 			 * If there exists an big page mapping,
3516 			 * that means the whole existing big page
3517 			 * has TNC setting already. No need to covert to
3518 			 * TNC again.
3519 			 */
3520 			ASSERT(PP_ISTNC(pp1));
3521 		}
3522 	}
3523 #endif	/* VAC */
3524 
3525 	return (0);
3526 }
3527 
3528 #ifdef VAC
3529 /*
3530  * Routine that detects vac consistency for a large page. It also
3531  * sets virtual color for all pp's for this big mapping.
3532  */
3533 static int
3534 sfmmu_vacconflict_array(caddr_t addr, page_t *pp, int *cflags)
3535 {
3536 	int vcolor, ocolor;
3537 
3538 	ASSERT(sfmmu_mlist_held(pp));
3539 
3540 	if (PP_ISNC(pp)) {
3541 		return (HAT_TMPNC);
3542 	}
3543 
3544 	vcolor = addr_to_vcolor(addr);
3545 	if (PP_NEWPAGE(pp)) {
3546 		PP_SET_VCOLOR(pp, vcolor);
3547 		return (0);
3548 	}
3549 
3550 	ocolor = PP_GET_VCOLOR(pp);
3551 	if (ocolor == vcolor) {
3552 		return (0);
3553 	}
3554 
3555 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
3556 		/*
3557 		 * Previous user of page had a differnet color
3558 		 * but since there are no current users
3559 		 * we just flush the cache and change the color.
3560 		 * As an optimization for large pages we flush the
3561 		 * entire cache of that color and set a flag.
3562 		 */
3563 		SFMMU_STAT(sf_pgcolor_conflict);
3564 		if (!CacheColor_IsFlushed(*cflags, ocolor)) {
3565 			CacheColor_SetFlushed(*cflags, ocolor);
3566 			sfmmu_cache_flushcolor(ocolor, pp->p_pagenum);
3567 		}
3568 		PP_SET_VCOLOR(pp, vcolor);
3569 		return (0);
3570 	}
3571 
3572 	/*
3573 	 * We got a real conflict with a current mapping.
3574 	 * set flags to start unencaching all mappings
3575 	 * and return failure so we restart looping
3576 	 * the pp array from the beginning.
3577 	 */
3578 	return (HAT_TMPNC);
3579 }
3580 #endif	/* VAC */
3581 
3582 /*
3583  * creates a large page shadow hmeblk for a tte.
3584  * The purpose of this routine is to allow us to do quick unloads because
3585  * the vm layer can easily pass a very large but sparsely populated range.
3586  */
3587 static struct hme_blk *
3588 sfmmu_shadow_hcreate(sfmmu_t *sfmmup, caddr_t vaddr, int ttesz, uint_t flags)
3589 {
3590 	struct hmehash_bucket *hmebp;
3591 	hmeblk_tag hblktag;
3592 	int hmeshift, size, vshift;
3593 	uint_t shw_mask, newshw_mask;
3594 	struct hme_blk *hmeblkp;
3595 
3596 	ASSERT(sfmmup != KHATID);
3597 	if (mmu_page_sizes == max_mmu_page_sizes) {
3598 		ASSERT(ttesz < TTE256M);
3599 	} else {
3600 		ASSERT(ttesz < TTE4M);
3601 		ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
3602 		ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
3603 	}
3604 
3605 	if (ttesz == TTE8K) {
3606 		size = TTE512K;
3607 	} else {
3608 		size = ++ttesz;
3609 	}
3610 
3611 	hblktag.htag_id = sfmmup;
3612 	hmeshift = HME_HASH_SHIFT(size);
3613 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
3614 	hblktag.htag_rehash = HME_HASH_REHASH(size);
3615 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3616 	hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
3617 
3618 	SFMMU_HASH_LOCK(hmebp);
3619 
3620 	HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
3621 	ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
3622 	if (hmeblkp == NULL) {
3623 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
3624 		    hblktag, flags, SFMMU_INVALID_SHMERID);
3625 	}
3626 	ASSERT(hmeblkp);
3627 	if (!hmeblkp->hblk_shw_mask) {
3628 		/*
3629 		 * if this is a unused hblk it was just allocated or could
3630 		 * potentially be a previous large page hblk so we need to
3631 		 * set the shadow bit.
3632 		 */
3633 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3634 		hmeblkp->hblk_shw_bit = 1;
3635 	} else if (hmeblkp->hblk_shw_bit == 0) {
3636 		panic("sfmmu_shadow_hcreate: shw bit not set in hmeblkp 0x%p",
3637 		    (void *)hmeblkp);
3638 	}
3639 	ASSERT(hmeblkp->hblk_shw_bit == 1);
3640 	ASSERT(!hmeblkp->hblk_shared);
3641 	vshift = vaddr_to_vshift(hblktag, vaddr, size);
3642 	ASSERT(vshift < 8);
3643 	/*
3644 	 * Atomically set shw mask bit
3645 	 */
3646 	do {
3647 		shw_mask = hmeblkp->hblk_shw_mask;
3648 		newshw_mask = shw_mask | (1 << vshift);
3649 		newshw_mask = atomic_cas_32(&hmeblkp->hblk_shw_mask, shw_mask,
3650 		    newshw_mask);
3651 	} while (newshw_mask != shw_mask);
3652 
3653 	SFMMU_HASH_UNLOCK(hmebp);
3654 
3655 	return (hmeblkp);
3656 }
3657 
3658 /*
3659  * This routine cleanup a previous shadow hmeblk and changes it to
3660  * a regular hblk.  This happens rarely but it is possible
3661  * when a process wants to use large pages and there are hblks still
3662  * lying around from the previous as that used these hmeblks.
3663  * The alternative was to cleanup the shadow hblks at unload time
3664  * but since so few user processes actually use large pages, it is
3665  * better to be lazy and cleanup at this time.
3666  */
3667 static void
3668 sfmmu_shadow_hcleanup(sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
3669     struct hmehash_bucket *hmebp)
3670 {
3671 	caddr_t addr, endaddr;
3672 	int hashno, size;
3673 
3674 	ASSERT(hmeblkp->hblk_shw_bit);
3675 	ASSERT(!hmeblkp->hblk_shared);
3676 
3677 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3678 
3679 	if (!hmeblkp->hblk_shw_mask) {
3680 		hmeblkp->hblk_shw_bit = 0;
3681 		return;
3682 	}
3683 	addr = (caddr_t)get_hblk_base(hmeblkp);
3684 	endaddr = get_hblk_endaddr(hmeblkp);
3685 	size = get_hblk_ttesz(hmeblkp);
3686 	hashno = size - 1;
3687 	ASSERT(hashno > 0);
3688 	SFMMU_HASH_UNLOCK(hmebp);
3689 
3690 	sfmmu_free_hblks(sfmmup, addr, endaddr, hashno);
3691 
3692 	SFMMU_HASH_LOCK(hmebp);
3693 }
3694 
3695 static void
3696 sfmmu_free_hblks(sfmmu_t *sfmmup, caddr_t addr, caddr_t endaddr,
3697     int hashno)
3698 {
3699 	int hmeshift, shadow = 0;
3700 	hmeblk_tag hblktag;
3701 	struct hmehash_bucket *hmebp;
3702 	struct hme_blk *hmeblkp;
3703 	struct hme_blk *nx_hblk, *pr_hblk, *list = NULL;
3704 
3705 	ASSERT(hashno > 0);
3706 	hblktag.htag_id = sfmmup;
3707 	hblktag.htag_rehash = hashno;
3708 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3709 
3710 	hmeshift = HME_HASH_SHIFT(hashno);
3711 
3712 	while (addr < endaddr) {
3713 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3714 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3715 		SFMMU_HASH_LOCK(hmebp);
3716 		/* inline HME_HASH_SEARCH */
3717 		hmeblkp = hmebp->hmeblkp;
3718 		pr_hblk = NULL;
3719 		while (hmeblkp) {
3720 			if (HTAGS_EQ(hmeblkp->hblk_tag, hblktag)) {
3721 				/* found hme_blk */
3722 				ASSERT(!hmeblkp->hblk_shared);
3723 				if (hmeblkp->hblk_shw_bit) {
3724 					if (hmeblkp->hblk_shw_mask) {
3725 						shadow = 1;
3726 						sfmmu_shadow_hcleanup(sfmmup,
3727 						    hmeblkp, hmebp);
3728 						break;
3729 					} else {
3730 						hmeblkp->hblk_shw_bit = 0;
3731 					}
3732 				}
3733 
3734 				/*
3735 				 * Hblk_hmecnt and hblk_vcnt could be non zero
3736 				 * since hblk_unload() does not gurantee that.
3737 				 *
3738 				 * XXX - this could cause tteload() to spin
3739 				 * where sfmmu_shadow_hcleanup() is called.
3740 				 */
3741 			}
3742 
3743 			nx_hblk = hmeblkp->hblk_next;
3744 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
3745 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3746 				    &list, 0);
3747 			} else {
3748 				pr_hblk = hmeblkp;
3749 			}
3750 			hmeblkp = nx_hblk;
3751 		}
3752 
3753 		SFMMU_HASH_UNLOCK(hmebp);
3754 
3755 		if (shadow) {
3756 			/*
3757 			 * We found another shadow hblk so cleaned its
3758 			 * children.  We need to go back and cleanup
3759 			 * the original hblk so we don't change the
3760 			 * addr.
3761 			 */
3762 			shadow = 0;
3763 		} else {
3764 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
3765 			    (1 << hmeshift));
3766 		}
3767 	}
3768 	sfmmu_hblks_list_purge(&list, 0);
3769 }
3770 
3771 /*
3772  * This routine's job is to delete stale invalid shared hmeregions hmeblks that
3773  * may still linger on after pageunload.
3774  */
3775 static void
3776 sfmmu_cleanup_rhblk(sf_srd_t *srdp, caddr_t addr, uint_t rid, int ttesz)
3777 {
3778 	int hmeshift;
3779 	hmeblk_tag hblktag;
3780 	struct hmehash_bucket *hmebp;
3781 	struct hme_blk *hmeblkp;
3782 	struct hme_blk *pr_hblk;
3783 	struct hme_blk *list = NULL;
3784 
3785 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3786 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3787 
3788 	hmeshift = HME_HASH_SHIFT(ttesz);
3789 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3790 	hblktag.htag_rehash = ttesz;
3791 	hblktag.htag_rid = rid;
3792 	hblktag.htag_id = srdp;
3793 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3794 
3795 	SFMMU_HASH_LOCK(hmebp);
3796 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3797 	if (hmeblkp != NULL) {
3798 		ASSERT(hmeblkp->hblk_shared);
3799 		ASSERT(!hmeblkp->hblk_shw_bit);
3800 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3801 			panic("sfmmu_cleanup_rhblk: valid hmeblk");
3802 		}
3803 		ASSERT(!hmeblkp->hblk_lckcnt);
3804 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3805 		    &list, 0);
3806 	}
3807 	SFMMU_HASH_UNLOCK(hmebp);
3808 	sfmmu_hblks_list_purge(&list, 0);
3809 }
3810 
3811 /* ARGSUSED */
3812 static void
3813 sfmmu_rgn_cb_noop(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
3814     size_t r_size, void *r_obj, u_offset_t r_objoff)
3815 {
3816 }
3817 
3818 /*
3819  * Searches for an hmeblk which maps addr, then unloads this mapping
3820  * and updates *eaddrp, if the hmeblk is found.
3821  */
3822 static void
3823 sfmmu_unload_hmeregion_va(sf_srd_t *srdp, uint_t rid, caddr_t addr,
3824     caddr_t eaddr, int ttesz, caddr_t *eaddrp)
3825 {
3826 	int hmeshift;
3827 	hmeblk_tag hblktag;
3828 	struct hmehash_bucket *hmebp;
3829 	struct hme_blk *hmeblkp;
3830 	struct hme_blk *pr_hblk;
3831 	struct hme_blk *list = NULL;
3832 
3833 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3834 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3835 	ASSERT(ttesz >= HBLK_MIN_TTESZ);
3836 
3837 	hmeshift = HME_HASH_SHIFT(ttesz);
3838 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3839 	hblktag.htag_rehash = ttesz;
3840 	hblktag.htag_rid = rid;
3841 	hblktag.htag_id = srdp;
3842 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3843 
3844 	SFMMU_HASH_LOCK(hmebp);
3845 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
3846 	if (hmeblkp != NULL) {
3847 		ASSERT(hmeblkp->hblk_shared);
3848 		ASSERT(!hmeblkp->hblk_lckcnt);
3849 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3850 			*eaddrp = sfmmu_hblk_unload(NULL, hmeblkp, addr,
3851 			    eaddr, NULL, HAT_UNLOAD);
3852 			ASSERT(*eaddrp > addr);
3853 		}
3854 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3855 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
3856 		    &list, 0);
3857 	}
3858 	SFMMU_HASH_UNLOCK(hmebp);
3859 	sfmmu_hblks_list_purge(&list, 0);
3860 }
3861 
3862 static void
3863 sfmmu_unload_hmeregion(sf_srd_t *srdp, sf_region_t *rgnp)
3864 {
3865 	int ttesz = rgnp->rgn_pgszc;
3866 	size_t rsz = rgnp->rgn_size;
3867 	caddr_t rsaddr = rgnp->rgn_saddr;
3868 	caddr_t readdr = rsaddr + rsz;
3869 	caddr_t rhsaddr;
3870 	caddr_t va;
3871 	uint_t rid = rgnp->rgn_id;
3872 	caddr_t cbsaddr;
3873 	caddr_t cbeaddr;
3874 	hat_rgn_cb_func_t rcbfunc;
3875 	ulong_t cnt;
3876 
3877 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3878 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3879 
3880 	ASSERT(IS_P2ALIGNED(rsaddr, TTEBYTES(ttesz)));
3881 	ASSERT(IS_P2ALIGNED(rsz, TTEBYTES(ttesz)));
3882 	if (ttesz < HBLK_MIN_TTESZ) {
3883 		ttesz = HBLK_MIN_TTESZ;
3884 		rhsaddr = (caddr_t)P2ALIGN((uintptr_t)rsaddr, HBLK_MIN_BYTES);
3885 	} else {
3886 		rhsaddr = rsaddr;
3887 	}
3888 
3889 	if ((rcbfunc = rgnp->rgn_cb_function) == NULL) {
3890 		rcbfunc = sfmmu_rgn_cb_noop;
3891 	}
3892 
3893 	while (ttesz >= HBLK_MIN_TTESZ) {
3894 		cbsaddr = rsaddr;
3895 		cbeaddr = rsaddr;
3896 		if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
3897 			ttesz--;
3898 			continue;
3899 		}
3900 		cnt = 0;
3901 		va = rsaddr;
3902 		while (va < readdr) {
3903 			ASSERT(va >= rhsaddr);
3904 			if (va != cbeaddr) {
3905 				if (cbeaddr != cbsaddr) {
3906 					ASSERT(cbeaddr > cbsaddr);
3907 					(*rcbfunc)(cbsaddr, cbeaddr,
3908 					    rsaddr, rsz, rgnp->rgn_obj,
3909 					    rgnp->rgn_objoff);
3910 				}
3911 				cbsaddr = va;
3912 				cbeaddr = va;
3913 			}
3914 			sfmmu_unload_hmeregion_va(srdp, rid, va, readdr,
3915 			    ttesz, &cbeaddr);
3916 			cnt++;
3917 			va = rhsaddr + (cnt << TTE_PAGE_SHIFT(ttesz));
3918 		}
3919 		if (cbeaddr != cbsaddr) {
3920 			ASSERT(cbeaddr > cbsaddr);
3921 			(*rcbfunc)(cbsaddr, cbeaddr, rsaddr,
3922 			    rsz, rgnp->rgn_obj,
3923 			    rgnp->rgn_objoff);
3924 		}
3925 		ttesz--;
3926 	}
3927 }
3928 
3929 /*
3930  * Release one hardware address translation lock on the given address range.
3931  */
3932 void
3933 hat_unlock(struct hat *sfmmup, caddr_t addr, size_t len)
3934 {
3935 	struct hmehash_bucket *hmebp;
3936 	hmeblk_tag hblktag;
3937 	int hmeshift, hashno = 1;
3938 	struct hme_blk *hmeblkp, *list = NULL;
3939 	caddr_t endaddr;
3940 
3941 	ASSERT(sfmmup != NULL);
3942 
3943 	ASSERT((sfmmup == ksfmmup) || AS_LOCK_HELD(sfmmup->sfmmu_as));
3944 	ASSERT((len & MMU_PAGEOFFSET) == 0);
3945 	endaddr = addr + len;
3946 	hblktag.htag_id = sfmmup;
3947 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3948 
3949 	/*
3950 	 * Spitfire supports 4 page sizes.
3951 	 * Most pages are expected to be of the smallest page size (8K) and
3952 	 * these will not need to be rehashed. 64K pages also don't need to be
3953 	 * rehashed because an hmeblk spans 64K of address space. 512K pages
3954 	 * might need 1 rehash and and 4M pages might need 2 rehashes.
3955 	 */
3956 	while (addr < endaddr) {
3957 		hmeshift = HME_HASH_SHIFT(hashno);
3958 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3959 		hblktag.htag_rehash = hashno;
3960 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3961 
3962 		SFMMU_HASH_LOCK(hmebp);
3963 
3964 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
3965 		if (hmeblkp != NULL) {
3966 			ASSERT(!hmeblkp->hblk_shared);
3967 			/*
3968 			 * If we encounter a shadow hmeblk then
3969 			 * we know there are no valid hmeblks mapping
3970 			 * this address at this size or larger.
3971 			 * Just increment address by the smallest
3972 			 * page size.
3973 			 */
3974 			if (hmeblkp->hblk_shw_bit) {
3975 				addr += MMU_PAGESIZE;
3976 			} else {
3977 				addr = sfmmu_hblk_unlock(hmeblkp, addr,
3978 				    endaddr);
3979 			}
3980 			SFMMU_HASH_UNLOCK(hmebp);
3981 			hashno = 1;
3982 			continue;
3983 		}
3984 		SFMMU_HASH_UNLOCK(hmebp);
3985 
3986 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
3987 			/*
3988 			 * We have traversed the whole list and rehashed
3989 			 * if necessary without finding the address to unlock
3990 			 * which should never happen.
3991 			 */
3992 			panic("sfmmu_unlock: addr not found. "
3993 			    "addr %p hat %p", (void *)addr, (void *)sfmmup);
3994 		} else {
3995 			hashno++;
3996 		}
3997 	}
3998 
3999 	sfmmu_hblks_list_purge(&list, 0);
4000 }
4001 
4002 void
4003 hat_unlock_region(struct hat *sfmmup, caddr_t addr, size_t len,
4004     hat_region_cookie_t rcookie)
4005 {
4006 	sf_srd_t *srdp;
4007 	sf_region_t *rgnp;
4008 	int ttesz;
4009 	uint_t rid;
4010 	caddr_t eaddr;
4011 	caddr_t va;
4012 	int hmeshift;
4013 	hmeblk_tag hblktag;
4014 	struct hmehash_bucket *hmebp;
4015 	struct hme_blk *hmeblkp;
4016 	struct hme_blk *pr_hblk;
4017 	struct hme_blk *list;
4018 
4019 	if (rcookie == HAT_INVALID_REGION_COOKIE) {
4020 		hat_unlock(sfmmup, addr, len);
4021 		return;
4022 	}
4023 
4024 	ASSERT(sfmmup != NULL);
4025 	ASSERT(sfmmup != ksfmmup);
4026 
4027 	srdp = sfmmup->sfmmu_srdp;
4028 	rid = (uint_t)((uint64_t)rcookie);
4029 	VERIFY3U(rid, <, SFMMU_MAX_HME_REGIONS);
4030 	eaddr = addr + len;
4031 	va = addr;
4032 	list = NULL;
4033 	rgnp = srdp->srd_hmergnp[rid];
4034 	SFMMU_VALIDATE_HMERID(sfmmup, rid, addr, len);
4035 
4036 	ASSERT(IS_P2ALIGNED(addr, TTEBYTES(rgnp->rgn_pgszc)));
4037 	ASSERT(IS_P2ALIGNED(len, TTEBYTES(rgnp->rgn_pgszc)));
4038 	if (rgnp->rgn_pgszc < HBLK_MIN_TTESZ) {
4039 		ttesz = HBLK_MIN_TTESZ;
4040 	} else {
4041 		ttesz = rgnp->rgn_pgszc;
4042 	}
4043 	while (va < eaddr) {
4044 		while (ttesz < rgnp->rgn_pgszc &&
4045 		    IS_P2ALIGNED(va, TTEBYTES(ttesz + 1))) {
4046 			ttesz++;
4047 		}
4048 		while (ttesz >= HBLK_MIN_TTESZ) {
4049 			if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
4050 				ttesz--;
4051 				continue;
4052 			}
4053 			hmeshift = HME_HASH_SHIFT(ttesz);
4054 			hblktag.htag_bspage = HME_HASH_BSPAGE(va, hmeshift);
4055 			hblktag.htag_rehash = ttesz;
4056 			hblktag.htag_rid = rid;
4057 			hblktag.htag_id = srdp;
4058 			hmebp = HME_HASH_FUNCTION(srdp, va, hmeshift);
4059 			SFMMU_HASH_LOCK(hmebp);
4060 			HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk,
4061 			    &list);
4062 			if (hmeblkp == NULL) {
4063 				SFMMU_HASH_UNLOCK(hmebp);
4064 				ttesz--;
4065 				continue;
4066 			}
4067 			ASSERT(hmeblkp->hblk_shared);
4068 			va = sfmmu_hblk_unlock(hmeblkp, va, eaddr);
4069 			ASSERT(va >= eaddr ||
4070 			    IS_P2ALIGNED((uintptr_t)va, TTEBYTES(ttesz)));
4071 			SFMMU_HASH_UNLOCK(hmebp);
4072 			break;
4073 		}
4074 		if (ttesz < HBLK_MIN_TTESZ) {
4075 			panic("hat_unlock_region: addr not found "
4076 			    "addr %p hat %p", (void *)va, (void *)sfmmup);
4077 		}
4078 	}
4079 	sfmmu_hblks_list_purge(&list, 0);
4080 }
4081 
4082 /*
4083  * Function to unlock a range of addresses in an hmeblk.  It returns the
4084  * next address that needs to be unlocked.
4085  * Should be called with the hash lock held.
4086  */
4087 static caddr_t
4088 sfmmu_hblk_unlock(struct hme_blk *hmeblkp, caddr_t addr, caddr_t endaddr)
4089 {
4090 	struct sf_hment *sfhme;
4091 	tte_t tteold, ttemod;
4092 	int ttesz, ret;
4093 
4094 	ASSERT(in_hblk_range(hmeblkp, addr));
4095 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4096 
4097 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4098 	ttesz = get_hblk_ttesz(hmeblkp);
4099 
4100 	HBLKTOHME(sfhme, hmeblkp, addr);
4101 	while (addr < endaddr) {
4102 readtte:
4103 		sfmmu_copytte(&sfhme->hme_tte, &tteold);
4104 		if (TTE_IS_VALID(&tteold)) {
4105 
4106 			ttemod = tteold;
4107 
4108 			ret = sfmmu_modifytte_try(&tteold, &ttemod,
4109 			    &sfhme->hme_tte);
4110 
4111 			if (ret < 0)
4112 				goto readtte;
4113 
4114 			if (hmeblkp->hblk_lckcnt == 0)
4115 				panic("zero hblk lckcnt");
4116 
4117 			if (((uintptr_t)addr + TTEBYTES(ttesz)) >
4118 			    (uintptr_t)endaddr)
4119 				panic("can't unlock large tte");
4120 
4121 			ASSERT(hmeblkp->hblk_lckcnt > 0);
4122 			atomic_dec_32(&hmeblkp->hblk_lckcnt);
4123 			HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
4124 		} else {
4125 			panic("sfmmu_hblk_unlock: invalid tte");
4126 		}
4127 		addr += TTEBYTES(ttesz);
4128 		sfhme++;
4129 	}
4130 	return (addr);
4131 }
4132 
4133 /*
4134  * Physical Address Mapping Framework
4135  *
4136  * General rules:
4137  *
4138  * (1) Applies only to seg_kmem memory pages. To make things easier,
4139  *     seg_kpm addresses are also accepted by the routines, but nothing
4140  *     is done with them since by definition their PA mappings are static.
4141  * (2) hat_add_callback() may only be called while holding the page lock
4142  *     SE_SHARED or SE_EXCL of the underlying page (e.g., as_pagelock()),
4143  *     or passing HAC_PAGELOCK flag.
4144  * (3) prehandler() and posthandler() may not call hat_add_callback() or
4145  *     hat_delete_callback(), nor should they allocate memory. Post quiesce
4146  *     callbacks may not sleep or acquire adaptive mutex locks.
4147  * (4) Either prehandler() or posthandler() (but not both) may be specified
4148  *     as being NULL.  Specifying an errhandler() is optional.
4149  *
4150  * Details of using the framework:
4151  *
4152  * registering a callback (hat_register_callback())
4153  *
4154  *	Pass prehandler, posthandler, errhandler addresses
4155  *	as described below. If capture_cpus argument is nonzero,
4156  *	suspend callback to the prehandler will occur with CPUs
4157  *	captured and executing xc_loop() and CPUs will remain
4158  *	captured until after the posthandler suspend callback
4159  *	occurs.
4160  *
4161  * adding a callback (hat_add_callback())
4162  *
4163  *      as_pagelock();
4164  *	hat_add_callback();
4165  *      save returned pfn in private data structures or program registers;
4166  *      as_pageunlock();
4167  *
4168  * prehandler()
4169  *
4170  *	Stop all accesses by physical address to this memory page.
4171  *	Called twice: the first, PRESUSPEND, is a context safe to acquire
4172  *	adaptive locks. The second, SUSPEND, is called at high PIL with
4173  *	CPUs captured so adaptive locks may NOT be acquired (and all spin
4174  *	locks must be XCALL_PIL or higher locks).
4175  *
4176  *	May return the following errors:
4177  *		EIO:	A fatal error has occurred. This will result in panic.
4178  *		EAGAIN:	The page cannot be suspended. This will fail the
4179  *			relocation.
4180  *		0:	Success.
4181  *
4182  * posthandler()
4183  *
4184  *      Save new pfn in private data structures or program registers;
4185  *	not allowed to fail (non-zero return values will result in panic).
4186  *
4187  * errhandler()
4188  *
4189  *	called when an error occurs related to the callback.  Currently
4190  *	the only such error is HAT_CB_ERR_LEAKED which indicates that
4191  *	a page is being freed, but there are still outstanding callback(s)
4192  *	registered on the page.
4193  *
4194  * removing a callback (hat_delete_callback(); e.g., prior to freeing memory)
4195  *
4196  *	stop using physical address
4197  *	hat_delete_callback();
4198  *
4199  */
4200 
4201 /*
4202  * Register a callback class.  Each subsystem should do this once and
4203  * cache the id_t returned for use in setting up and tearing down callbacks.
4204  *
4205  * There is no facility for removing callback IDs once they are created;
4206  * the "key" should be unique for each module, so in case a module is unloaded
4207  * and subsequently re-loaded, we can recycle the module's previous entry.
4208  */
4209 id_t
4210 hat_register_callback(int key,
4211     int (*prehandler)(caddr_t, uint_t, uint_t, void *),
4212     int (*posthandler)(caddr_t, uint_t, uint_t, void *, pfn_t),
4213     int (*errhandler)(caddr_t, uint_t, uint_t, void *),
4214     int capture_cpus)
4215 {
4216 	id_t id;
4217 
4218 	/*
4219 	 * Search the table for a pre-existing callback associated with
4220 	 * the identifier "key".  If one exists, we re-use that entry in
4221 	 * the table for this instance, otherwise we assign the next
4222 	 * available table slot.
4223 	 */
4224 	for (id = 0; id < sfmmu_max_cb_id; id++) {
4225 		if (sfmmu_cb_table[id].key == key)
4226 			break;
4227 	}
4228 
4229 	if (id == sfmmu_max_cb_id) {
4230 		id = sfmmu_cb_nextid++;
4231 		if (id >= sfmmu_max_cb_id)
4232 			panic("hat_register_callback: out of callback IDs");
4233 	}
4234 
4235 	ASSERT(prehandler != NULL || posthandler != NULL);
4236 
4237 	sfmmu_cb_table[id].key = key;
4238 	sfmmu_cb_table[id].prehandler = prehandler;
4239 	sfmmu_cb_table[id].posthandler = posthandler;
4240 	sfmmu_cb_table[id].errhandler = errhandler;
4241 	sfmmu_cb_table[id].capture_cpus = capture_cpus;
4242 
4243 	return (id);
4244 }
4245 
4246 #define	HAC_COOKIE_NONE	(void *)-1
4247 
4248 /*
4249  * Add relocation callbacks to the specified addr/len which will be called
4250  * when relocating the associated page. See the description of pre and
4251  * posthandler above for more details.
4252  *
4253  * If HAC_PAGELOCK is included in flags, the underlying memory page is
4254  * locked internally so the caller must be able to deal with the callback
4255  * running even before this function has returned.  If HAC_PAGELOCK is not
4256  * set, it is assumed that the underlying memory pages are locked.
4257  *
4258  * Since the caller must track the individual page boundaries anyway,
4259  * we only allow a callback to be added to a single page (large
4260  * or small).  Thus [addr, addr + len) MUST be contained within a single
4261  * page.
4262  *
4263  * Registering multiple callbacks on the same [addr, addr+len) is supported,
4264  * _provided_that_ a unique parameter is specified for each callback.
4265  * If multiple callbacks are registered on the same range the callback will
4266  * be invoked with each unique parameter. Registering the same callback with
4267  * the same argument more than once will result in corrupted kernel state.
4268  *
4269  * Returns the pfn of the underlying kernel page in *rpfn
4270  * on success, or PFN_INVALID on failure.
4271  *
4272  * cookiep (if passed) provides storage space for an opaque cookie
4273  * to return later to hat_delete_callback(). This cookie makes the callback
4274  * deletion significantly quicker by avoiding a potentially lengthy hash
4275  * search.
4276  *
4277  * Returns values:
4278  *    0:      success
4279  *    ENOMEM: memory allocation failure (e.g. flags was passed as HAC_NOSLEEP)
4280  *    EINVAL: callback ID is not valid
4281  *    ENXIO:  ["vaddr", "vaddr" + len) is not mapped in the kernel's address
4282  *            space
4283  *    ERANGE: ["vaddr", "vaddr" + len) crosses a page boundary
4284  */
4285 int
4286 hat_add_callback(id_t callback_id, caddr_t vaddr, uint_t len, uint_t flags,
4287     void *pvt, pfn_t *rpfn, void **cookiep)
4288 {
4289 	struct		hmehash_bucket *hmebp;
4290 	hmeblk_tag	hblktag;
4291 	struct hme_blk	*hmeblkp;
4292 	int		hmeshift, hashno;
4293 	caddr_t		saddr, eaddr, baseaddr;
4294 	struct pa_hment *pahmep;
4295 	struct sf_hment *sfhmep, *osfhmep;
4296 	kmutex_t	*pml;
4297 	tte_t		tte;
4298 	page_t		*pp;
4299 	vnode_t		*vp;
4300 	u_offset_t	off;
4301 	pfn_t		pfn;
4302 	int		kmflags = (flags & HAC_SLEEP)? KM_SLEEP : KM_NOSLEEP;
4303 	int		locked = 0;
4304 
4305 	/*
4306 	 * For KPM mappings, just return the physical address since we
4307 	 * don't need to register any callbacks.
4308 	 */
4309 	if (IS_KPM_ADDR(vaddr)) {
4310 		uint64_t paddr;
4311 		SFMMU_KPM_VTOP(vaddr, paddr);
4312 		*rpfn = btop(paddr);
4313 		if (cookiep != NULL)
4314 			*cookiep = HAC_COOKIE_NONE;
4315 		return (0);
4316 	}
4317 
4318 	if (callback_id < (id_t)0 || callback_id >= sfmmu_cb_nextid) {
4319 		*rpfn = PFN_INVALID;
4320 		return (EINVAL);
4321 	}
4322 
4323 	if ((pahmep = kmem_cache_alloc(pa_hment_cache, kmflags)) == NULL) {
4324 		*rpfn = PFN_INVALID;
4325 		return (ENOMEM);
4326 	}
4327 
4328 	sfhmep = &pahmep->sfment;
4329 
4330 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4331 	eaddr = saddr + len;
4332 
4333 rehash:
4334 	/* Find the mapping(s) for this page */
4335 	for (hashno = TTE64K, hmeblkp = NULL;
4336 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4337 	    hashno++) {
4338 		hmeshift = HME_HASH_SHIFT(hashno);
4339 		hblktag.htag_id = ksfmmup;
4340 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4341 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4342 		hblktag.htag_rehash = hashno;
4343 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4344 
4345 		SFMMU_HASH_LOCK(hmebp);
4346 
4347 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4348 
4349 		if (hmeblkp == NULL)
4350 			SFMMU_HASH_UNLOCK(hmebp);
4351 	}
4352 
4353 	if (hmeblkp == NULL) {
4354 		kmem_cache_free(pa_hment_cache, pahmep);
4355 		*rpfn = PFN_INVALID;
4356 		return (ENXIO);
4357 	}
4358 
4359 	ASSERT(!hmeblkp->hblk_shared);
4360 
4361 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4362 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4363 
4364 	if (!TTE_IS_VALID(&tte)) {
4365 		SFMMU_HASH_UNLOCK(hmebp);
4366 		kmem_cache_free(pa_hment_cache, pahmep);
4367 		*rpfn = PFN_INVALID;
4368 		return (ENXIO);
4369 	}
4370 
4371 	/*
4372 	 * Make sure the boundaries for the callback fall within this
4373 	 * single mapping.
4374 	 */
4375 	baseaddr = (caddr_t)get_hblk_base(hmeblkp);
4376 	ASSERT(saddr >= baseaddr);
4377 	if (eaddr > saddr + TTEBYTES(TTE_CSZ(&tte))) {
4378 		SFMMU_HASH_UNLOCK(hmebp);
4379 		kmem_cache_free(pa_hment_cache, pahmep);
4380 		*rpfn = PFN_INVALID;
4381 		return (ERANGE);
4382 	}
4383 
4384 	pfn = sfmmu_ttetopfn(&tte, vaddr);
4385 
4386 	/*
4387 	 * The pfn may not have a page_t underneath in which case we
4388 	 * just return it. This can happen if we are doing I/O to a
4389 	 * static portion of the kernel's address space, for instance.
4390 	 */
4391 	pp = osfhmep->hme_page;
4392 	if (pp == NULL) {
4393 		SFMMU_HASH_UNLOCK(hmebp);
4394 		kmem_cache_free(pa_hment_cache, pahmep);
4395 		*rpfn = pfn;
4396 		if (cookiep)
4397 			*cookiep = HAC_COOKIE_NONE;
4398 		return (0);
4399 	}
4400 	ASSERT(pp == PP_PAGEROOT(pp));
4401 
4402 	vp = pp->p_vnode;
4403 	off = pp->p_offset;
4404 
4405 	pml = sfmmu_mlist_enter(pp);
4406 
4407 	if (flags & HAC_PAGELOCK) {
4408 		if (!page_trylock(pp, SE_SHARED)) {
4409 			/*
4410 			 * Somebody is holding SE_EXCL lock. Might
4411 			 * even be hat_page_relocate(). Drop all
4412 			 * our locks, lookup the page in &kvp, and
4413 			 * retry. If it doesn't exist in &kvp and &zvp,
4414 			 * then we must be dealing with a kernel mapped
4415 			 * page which doesn't actually belong to
4416 			 * segkmem so we punt.
4417 			 */
4418 			sfmmu_mlist_exit(pml);
4419 			SFMMU_HASH_UNLOCK(hmebp);
4420 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4421 
4422 			/* check zvp before giving up */
4423 			if (pp == NULL)
4424 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4425 				    SE_SHARED);
4426 
4427 			/* Okay, we didn't find it, give up */
4428 			if (pp == NULL) {
4429 				kmem_cache_free(pa_hment_cache, pahmep);
4430 				*rpfn = pfn;
4431 				if (cookiep)
4432 					*cookiep = HAC_COOKIE_NONE;
4433 				return (0);
4434 			}
4435 			page_unlock(pp);
4436 			goto rehash;
4437 		}
4438 		locked = 1;
4439 	}
4440 
4441 	if (!PAGE_LOCKED(pp) && !panicstr)
4442 		panic("hat_add_callback: page 0x%p not locked", (void *)pp);
4443 
4444 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4445 	    pp->p_offset != off) {
4446 		/*
4447 		 * The page moved before we got our hands on it.  Drop
4448 		 * all the locks and try again.
4449 		 */
4450 		ASSERT((flags & HAC_PAGELOCK) != 0);
4451 		sfmmu_mlist_exit(pml);
4452 		SFMMU_HASH_UNLOCK(hmebp);
4453 		page_unlock(pp);
4454 		locked = 0;
4455 		goto rehash;
4456 	}
4457 
4458 	if (!VN_ISKAS(vp)) {
4459 		/*
4460 		 * This is not a segkmem page but another page which
4461 		 * has been kernel mapped. It had better have at least
4462 		 * a share lock on it. Return the pfn.
4463 		 */
4464 		sfmmu_mlist_exit(pml);
4465 		SFMMU_HASH_UNLOCK(hmebp);
4466 		if (locked)
4467 			page_unlock(pp);
4468 		kmem_cache_free(pa_hment_cache, pahmep);
4469 		ASSERT(PAGE_LOCKED(pp));
4470 		*rpfn = pfn;
4471 		if (cookiep)
4472 			*cookiep = HAC_COOKIE_NONE;
4473 		return (0);
4474 	}
4475 
4476 	/*
4477 	 * Setup this pa_hment and link its embedded dummy sf_hment into
4478 	 * the mapping list.
4479 	 */
4480 	pp->p_share++;
4481 	pahmep->cb_id = callback_id;
4482 	pahmep->addr = vaddr;
4483 	pahmep->len = len;
4484 	pahmep->refcnt = 1;
4485 	pahmep->flags = 0;
4486 	pahmep->pvt = pvt;
4487 
4488 	sfhmep->hme_tte.ll = 0;
4489 	sfhmep->hme_data = pahmep;
4490 	sfhmep->hme_prev = osfhmep;
4491 	sfhmep->hme_next = osfhmep->hme_next;
4492 
4493 	if (osfhmep->hme_next)
4494 		osfhmep->hme_next->hme_prev = sfhmep;
4495 
4496 	osfhmep->hme_next = sfhmep;
4497 
4498 	sfmmu_mlist_exit(pml);
4499 	SFMMU_HASH_UNLOCK(hmebp);
4500 
4501 	if (locked)
4502 		page_unlock(pp);
4503 
4504 	*rpfn = pfn;
4505 	if (cookiep)
4506 		*cookiep = (void *)pahmep;
4507 
4508 	return (0);
4509 }
4510 
4511 /*
4512  * Remove the relocation callbacks from the specified addr/len.
4513  */
4514 void
4515 hat_delete_callback(caddr_t vaddr, uint_t len, void *pvt, uint_t flags,
4516     void *cookie)
4517 {
4518 	struct		hmehash_bucket *hmebp;
4519 	hmeblk_tag	hblktag;
4520 	struct hme_blk	*hmeblkp;
4521 	int		hmeshift, hashno;
4522 	caddr_t		saddr;
4523 	struct pa_hment	*pahmep;
4524 	struct sf_hment	*sfhmep, *osfhmep;
4525 	kmutex_t	*pml;
4526 	tte_t		tte;
4527 	page_t		*pp;
4528 	vnode_t		*vp;
4529 	u_offset_t	off;
4530 	int		locked = 0;
4531 
4532 	/*
4533 	 * If the cookie is HAC_COOKIE_NONE then there is no pa_hment to
4534 	 * remove so just return.
4535 	 */
4536 	if (cookie == HAC_COOKIE_NONE || IS_KPM_ADDR(vaddr))
4537 		return;
4538 
4539 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4540 
4541 rehash:
4542 	/* Find the mapping(s) for this page */
4543 	for (hashno = TTE64K, hmeblkp = NULL;
4544 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4545 	    hashno++) {
4546 		hmeshift = HME_HASH_SHIFT(hashno);
4547 		hblktag.htag_id = ksfmmup;
4548 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4549 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4550 		hblktag.htag_rehash = hashno;
4551 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4552 
4553 		SFMMU_HASH_LOCK(hmebp);
4554 
4555 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4556 
4557 		if (hmeblkp == NULL)
4558 			SFMMU_HASH_UNLOCK(hmebp);
4559 	}
4560 
4561 	if (hmeblkp == NULL)
4562 		return;
4563 
4564 	ASSERT(!hmeblkp->hblk_shared);
4565 
4566 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4567 
4568 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4569 	if (!TTE_IS_VALID(&tte)) {
4570 		SFMMU_HASH_UNLOCK(hmebp);
4571 		return;
4572 	}
4573 
4574 	pp = osfhmep->hme_page;
4575 	if (pp == NULL) {
4576 		SFMMU_HASH_UNLOCK(hmebp);
4577 		ASSERT(cookie == NULL);
4578 		return;
4579 	}
4580 
4581 	vp = pp->p_vnode;
4582 	off = pp->p_offset;
4583 
4584 	pml = sfmmu_mlist_enter(pp);
4585 
4586 	if (flags & HAC_PAGELOCK) {
4587 		if (!page_trylock(pp, SE_SHARED)) {
4588 			/*
4589 			 * Somebody is holding SE_EXCL lock. Might
4590 			 * even be hat_page_relocate(). Drop all
4591 			 * our locks, lookup the page in &kvp, and
4592 			 * retry. If it doesn't exist in &kvp and &zvp,
4593 			 * then we must be dealing with a kernel mapped
4594 			 * page which doesn't actually belong to
4595 			 * segkmem so we punt.
4596 			 */
4597 			sfmmu_mlist_exit(pml);
4598 			SFMMU_HASH_UNLOCK(hmebp);
4599 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4600 			/* check zvp before giving up */
4601 			if (pp == NULL)
4602 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4603 				    SE_SHARED);
4604 
4605 			if (pp == NULL) {
4606 				ASSERT(cookie == NULL);
4607 				return;
4608 			}
4609 			page_unlock(pp);
4610 			goto rehash;
4611 		}
4612 		locked = 1;
4613 	}
4614 
4615 	ASSERT(PAGE_LOCKED(pp));
4616 
4617 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4618 	    pp->p_offset != off) {
4619 		/*
4620 		 * The page moved before we got our hands on it.  Drop
4621 		 * all the locks and try again.
4622 		 */
4623 		ASSERT((flags & HAC_PAGELOCK) != 0);
4624 		sfmmu_mlist_exit(pml);
4625 		SFMMU_HASH_UNLOCK(hmebp);
4626 		page_unlock(pp);
4627 		locked = 0;
4628 		goto rehash;
4629 	}
4630 
4631 	if (!VN_ISKAS(vp)) {
4632 		/*
4633 		 * This is not a segkmem page but another page which
4634 		 * has been kernel mapped.
4635 		 */
4636 		sfmmu_mlist_exit(pml);
4637 		SFMMU_HASH_UNLOCK(hmebp);
4638 		if (locked)
4639 			page_unlock(pp);
4640 		ASSERT(cookie == NULL);
4641 		return;
4642 	}
4643 
4644 	if (cookie != NULL) {
4645 		pahmep = (struct pa_hment *)cookie;
4646 		sfhmep = &pahmep->sfment;
4647 	} else {
4648 		for (sfhmep = pp->p_mapping; sfhmep != NULL;
4649 		    sfhmep = sfhmep->hme_next) {
4650 
4651 			/*
4652 			 * skip va<->pa mappings
4653 			 */
4654 			if (!IS_PAHME(sfhmep))
4655 				continue;
4656 
4657 			pahmep = sfhmep->hme_data;
4658 			ASSERT(pahmep != NULL);
4659 
4660 			/*
4661 			 * if pa_hment matches, remove it
4662 			 */
4663 			if ((pahmep->pvt == pvt) &&
4664 			    (pahmep->addr == vaddr) &&
4665 			    (pahmep->len == len)) {
4666 				break;
4667 			}
4668 		}
4669 	}
4670 
4671 	if (sfhmep == NULL) {
4672 		if (!panicstr) {
4673 			panic("hat_delete_callback: pa_hment not found, pp %p",
4674 			    (void *)pp);
4675 		}
4676 		return;
4677 	}
4678 
4679 	/*
4680 	 * Note: at this point a valid kernel mapping must still be
4681 	 * present on this page.
4682 	 */
4683 	pp->p_share--;
4684 	if (pp->p_share <= 0)
4685 		panic("hat_delete_callback: zero p_share");
4686 
4687 	if (--pahmep->refcnt == 0) {
4688 		if (pahmep->flags != 0)
4689 			panic("hat_delete_callback: pa_hment is busy");
4690 
4691 		/*
4692 		 * Remove sfhmep from the mapping list for the page.
4693 		 */
4694 		if (sfhmep->hme_prev) {
4695 			sfhmep->hme_prev->hme_next = sfhmep->hme_next;
4696 		} else {
4697 			pp->p_mapping = sfhmep->hme_next;
4698 		}
4699 
4700 		if (sfhmep->hme_next)
4701 			sfhmep->hme_next->hme_prev = sfhmep->hme_prev;
4702 
4703 		sfmmu_mlist_exit(pml);
4704 		SFMMU_HASH_UNLOCK(hmebp);
4705 
4706 		if (locked)
4707 			page_unlock(pp);
4708 
4709 		kmem_cache_free(pa_hment_cache, pahmep);
4710 		return;
4711 	}
4712 
4713 	sfmmu_mlist_exit(pml);
4714 	SFMMU_HASH_UNLOCK(hmebp);
4715 	if (locked)
4716 		page_unlock(pp);
4717 }
4718 
4719 /*
4720  * hat_probe returns 1 if the translation for the address 'addr' is
4721  * loaded, zero otherwise.
4722  *
4723  * hat_probe should be used only for advisorary purposes because it may
4724  * occasionally return the wrong value. The implementation must guarantee that
4725  * returning the wrong value is a very rare event. hat_probe is used
4726  * to implement optimizations in the segment drivers.
4727  *
4728  */
4729 int
4730 hat_probe(struct hat *sfmmup, caddr_t addr)
4731 {
4732 	pfn_t pfn;
4733 	tte_t tte;
4734 
4735 	ASSERT(sfmmup != NULL);
4736 
4737 	ASSERT((sfmmup == ksfmmup) || AS_LOCK_HELD(sfmmup->sfmmu_as));
4738 
4739 	if (sfmmup == ksfmmup) {
4740 		while ((pfn = sfmmu_vatopfn(addr, sfmmup, &tte))
4741 		    == PFN_SUSPENDED) {
4742 			sfmmu_vatopfn_suspended(addr, sfmmup, &tte);
4743 		}
4744 	} else {
4745 		pfn = sfmmu_uvatopfn(addr, sfmmup, NULL);
4746 	}
4747 
4748 	if (pfn != PFN_INVALID)
4749 		return (1);
4750 	else
4751 		return (0);
4752 }
4753 
4754 ssize_t
4755 hat_getpagesize(struct hat *sfmmup, caddr_t addr)
4756 {
4757 	tte_t tte;
4758 
4759 	if (sfmmup == ksfmmup) {
4760 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4761 			return (-1);
4762 		}
4763 	} else {
4764 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4765 			return (-1);
4766 		}
4767 	}
4768 
4769 	ASSERT(TTE_IS_VALID(&tte));
4770 	return (TTEBYTES(TTE_CSZ(&tte)));
4771 }
4772 
4773 uint_t
4774 hat_getattr(struct hat *sfmmup, caddr_t addr, uint_t *attr)
4775 {
4776 	tte_t tte;
4777 
4778 	if (sfmmup == ksfmmup) {
4779 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4780 			tte.ll = 0;
4781 		}
4782 	} else {
4783 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4784 			tte.ll = 0;
4785 		}
4786 	}
4787 	if (TTE_IS_VALID(&tte)) {
4788 		*attr = sfmmu_ptov_attr(&tte);
4789 		return (0);
4790 	}
4791 	*attr = 0;
4792 	return ((uint_t)0xffffffff);
4793 }
4794 
4795 /*
4796  * Enables more attributes on specified address range (ie. logical OR)
4797  */
4798 void
4799 hat_setattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4800 {
4801 	ASSERT(hat->sfmmu_as != NULL);
4802 
4803 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_SETATTR);
4804 }
4805 
4806 /*
4807  * Assigns attributes to the specified address range.  All the attributes
4808  * are specified.
4809  */
4810 void
4811 hat_chgattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4812 {
4813 	ASSERT(hat->sfmmu_as != NULL);
4814 
4815 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CHGATTR);
4816 }
4817 
4818 /*
4819  * Remove attributes on the specified address range (ie. loginal NAND)
4820  */
4821 void
4822 hat_clrattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4823 {
4824 	ASSERT(hat->sfmmu_as != NULL);
4825 
4826 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CLRATTR);
4827 }
4828 
4829 /*
4830  * Change attributes on an address range to that specified by attr and mode.
4831  */
4832 static void
4833 sfmmu_chgattr(struct hat *sfmmup, caddr_t addr, size_t len, uint_t attr,
4834     int mode)
4835 {
4836 	struct hmehash_bucket *hmebp;
4837 	hmeblk_tag hblktag;
4838 	int hmeshift, hashno = 1;
4839 	struct hme_blk *hmeblkp, *list = NULL;
4840 	caddr_t endaddr;
4841 	cpuset_t cpuset;
4842 	demap_range_t dmr;
4843 
4844 	CPUSET_ZERO(cpuset);
4845 
4846 	ASSERT((sfmmup == ksfmmup) || AS_LOCK_HELD(sfmmup->sfmmu_as));
4847 	ASSERT((len & MMU_PAGEOFFSET) == 0);
4848 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
4849 
4850 	if ((attr & PROT_USER) && (mode != SFMMU_CLRATTR) &&
4851 	    ((addr + len) > (caddr_t)USERLIMIT)) {
4852 		panic("user addr %p in kernel space",
4853 		    (void *)addr);
4854 	}
4855 
4856 	endaddr = addr + len;
4857 	hblktag.htag_id = sfmmup;
4858 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4859 	DEMAP_RANGE_INIT(sfmmup, &dmr);
4860 
4861 	while (addr < endaddr) {
4862 		hmeshift = HME_HASH_SHIFT(hashno);
4863 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
4864 		hblktag.htag_rehash = hashno;
4865 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
4866 
4867 		SFMMU_HASH_LOCK(hmebp);
4868 
4869 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
4870 		if (hmeblkp != NULL) {
4871 			ASSERT(!hmeblkp->hblk_shared);
4872 			/*
4873 			 * We've encountered a shadow hmeblk so skip the range
4874 			 * of the next smaller mapping size.
4875 			 */
4876 			if (hmeblkp->hblk_shw_bit) {
4877 				ASSERT(sfmmup != ksfmmup);
4878 				ASSERT(hashno > 1);
4879 				addr = (caddr_t)P2END((uintptr_t)addr,
4880 				    TTEBYTES(hashno - 1));
4881 			} else {
4882 				addr = sfmmu_hblk_chgattr(sfmmup,
4883 				    hmeblkp, addr, endaddr, &dmr, attr, mode);
4884 			}
4885 			SFMMU_HASH_UNLOCK(hmebp);
4886 			hashno = 1;
4887 			continue;
4888 		}
4889 		SFMMU_HASH_UNLOCK(hmebp);
4890 
4891 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
4892 			/*
4893 			 * We have traversed the whole list and rehashed
4894 			 * if necessary without finding the address to chgattr.
4895 			 * This is ok, so we increment the address by the
4896 			 * smallest hmeblk range for kernel mappings or for
4897 			 * user mappings with no large pages, and the largest
4898 			 * hmeblk range, to account for shadow hmeblks, for
4899 			 * user mappings with large pages and continue.
4900 			 */
4901 			if (sfmmup == ksfmmup)
4902 				addr = (caddr_t)P2END((uintptr_t)addr,
4903 				    TTEBYTES(1));
4904 			else
4905 				addr = (caddr_t)P2END((uintptr_t)addr,
4906 				    TTEBYTES(hashno));
4907 			hashno = 1;
4908 		} else {
4909 			hashno++;
4910 		}
4911 	}
4912 
4913 	sfmmu_hblks_list_purge(&list, 0);
4914 	DEMAP_RANGE_FLUSH(&dmr);
4915 	cpuset = sfmmup->sfmmu_cpusran;
4916 	xt_sync(cpuset);
4917 }
4918 
4919 /*
4920  * This function chgattr on a range of addresses in an hmeblk.  It returns the
4921  * next addres that needs to be chgattr.
4922  * It should be called with the hash lock held.
4923  * XXX It should be possible to optimize chgattr by not flushing every time but
4924  * on the other hand:
4925  * 1. do one flush crosscall.
4926  * 2. only flush if we are increasing permissions (make sure this will work)
4927  */
4928 static caddr_t
4929 sfmmu_hblk_chgattr(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
4930     caddr_t endaddr, demap_range_t *dmrp, uint_t attr, int mode)
4931 {
4932 	tte_t tte, tteattr, tteflags, ttemod;
4933 	struct sf_hment *sfhmep;
4934 	int ttesz;
4935 	struct page *pp = NULL;
4936 	kmutex_t *pml, *pmtx;
4937 	int ret;
4938 	int use_demap_range;
4939 #if defined(SF_ERRATA_57)
4940 	int check_exec;
4941 #endif
4942 
4943 	ASSERT(in_hblk_range(hmeblkp, addr));
4944 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4945 	ASSERT(!hmeblkp->hblk_shared);
4946 
4947 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4948 	ttesz = get_hblk_ttesz(hmeblkp);
4949 
4950 	/*
4951 	 * Flush the current demap region if addresses have been
4952 	 * skipped or the page size doesn't match.
4953 	 */
4954 	use_demap_range = (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp));
4955 	if (use_demap_range) {
4956 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
4957 	} else if (dmrp != NULL) {
4958 		DEMAP_RANGE_FLUSH(dmrp);
4959 	}
4960 
4961 	tteattr.ll = sfmmu_vtop_attr(attr, mode, &tteflags);
4962 #if defined(SF_ERRATA_57)
4963 	check_exec = (sfmmup != ksfmmup) &&
4964 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
4965 	    TTE_IS_EXECUTABLE(&tteattr);
4966 #endif
4967 	HBLKTOHME(sfhmep, hmeblkp, addr);
4968 	while (addr < endaddr) {
4969 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
4970 		if (TTE_IS_VALID(&tte)) {
4971 			if ((tte.ll & tteflags.ll) == tteattr.ll) {
4972 				/*
4973 				 * if the new attr is the same as old
4974 				 * continue
4975 				 */
4976 				goto next_addr;
4977 			}
4978 			if (!TTE_IS_WRITABLE(&tteattr)) {
4979 				/*
4980 				 * make sure we clear hw modify bit if we
4981 				 * removing write protections
4982 				 */
4983 				tteflags.tte_intlo |= TTE_HWWR_INT;
4984 			}
4985 
4986 			pml = NULL;
4987 			pp = sfhmep->hme_page;
4988 			if (pp) {
4989 				pml = sfmmu_mlist_enter(pp);
4990 			}
4991 
4992 			if (pp != sfhmep->hme_page) {
4993 				/*
4994 				 * tte must have been unloaded.
4995 				 */
4996 				ASSERT(pml);
4997 				sfmmu_mlist_exit(pml);
4998 				continue;
4999 			}
5000 
5001 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5002 
5003 			ttemod = tte;
5004 			ttemod.ll = (ttemod.ll & ~tteflags.ll) | tteattr.ll;
5005 			ASSERT(TTE_TO_TTEPFN(&ttemod) == TTE_TO_TTEPFN(&tte));
5006 
5007 #if defined(SF_ERRATA_57)
5008 			if (check_exec && addr < errata57_limit)
5009 				ttemod.tte_exec_perm = 0;
5010 #endif
5011 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5012 			    &sfhmep->hme_tte);
5013 
5014 			if (ret < 0) {
5015 				/* tte changed underneath us */
5016 				if (pml) {
5017 					sfmmu_mlist_exit(pml);
5018 				}
5019 				continue;
5020 			}
5021 
5022 			if (tteflags.tte_intlo & TTE_HWWR_INT) {
5023 				/*
5024 				 * need to sync if we are clearing modify bit.
5025 				 */
5026 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5027 			}
5028 
5029 			if (pp && PP_ISRO(pp)) {
5030 				if (tteattr.tte_intlo & TTE_WRPRM_INT) {
5031 					pmtx = sfmmu_page_enter(pp);
5032 					PP_CLRRO(pp);
5033 					sfmmu_page_exit(pmtx);
5034 				}
5035 			}
5036 
5037 			if (ret > 0 && use_demap_range) {
5038 				DEMAP_RANGE_MARKPG(dmrp, addr);
5039 			} else if (ret > 0) {
5040 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5041 			}
5042 
5043 			if (pml) {
5044 				sfmmu_mlist_exit(pml);
5045 			}
5046 		}
5047 next_addr:
5048 		addr += TTEBYTES(ttesz);
5049 		sfhmep++;
5050 		DEMAP_RANGE_NEXTPG(dmrp);
5051 	}
5052 	return (addr);
5053 }
5054 
5055 /*
5056  * This routine converts virtual attributes to physical ones.  It will
5057  * update the tteflags field with the tte mask corresponding to the attributes
5058  * affected and it returns the new attributes.  It will also clear the modify
5059  * bit if we are taking away write permission.  This is necessary since the
5060  * modify bit is the hardware permission bit and we need to clear it in order
5061  * to detect write faults.
5062  */
5063 static uint64_t
5064 sfmmu_vtop_attr(uint_t attr, int mode, tte_t *ttemaskp)
5065 {
5066 	tte_t ttevalue;
5067 
5068 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
5069 
5070 	switch (mode) {
5071 	case SFMMU_CHGATTR:
5072 		/* all attributes specified */
5073 		ttevalue.tte_inthi = MAKE_TTEATTR_INTHI(attr);
5074 		ttevalue.tte_intlo = MAKE_TTEATTR_INTLO(attr);
5075 		ttemaskp->tte_inthi = TTEINTHI_ATTR;
5076 		ttemaskp->tte_intlo = TTEINTLO_ATTR;
5077 		break;
5078 	case SFMMU_SETATTR:
5079 		ASSERT(!(attr & ~HAT_PROT_MASK));
5080 		ttemaskp->ll = 0;
5081 		ttevalue.ll = 0;
5082 		/*
5083 		 * a valid tte implies exec and read for sfmmu
5084 		 * so no need to do anything about them.
5085 		 * since priviledged access implies user access
5086 		 * PROT_USER doesn't make sense either.
5087 		 */
5088 		if (attr & PROT_WRITE) {
5089 			ttemaskp->tte_intlo |= TTE_WRPRM_INT;
5090 			ttevalue.tte_intlo |= TTE_WRPRM_INT;
5091 		}
5092 		break;
5093 	case SFMMU_CLRATTR:
5094 		/* attributes will be nand with current ones */
5095 		if (attr & ~(PROT_WRITE | PROT_USER)) {
5096 			panic("sfmmu: attr %x not supported", attr);
5097 		}
5098 		ttemaskp->ll = 0;
5099 		ttevalue.ll = 0;
5100 		if (attr & PROT_WRITE) {
5101 			/* clear both writable and modify bit */
5102 			ttemaskp->tte_intlo |= TTE_WRPRM_INT | TTE_HWWR_INT;
5103 		}
5104 		if (attr & PROT_USER) {
5105 			ttemaskp->tte_intlo |= TTE_PRIV_INT;
5106 			ttevalue.tte_intlo |= TTE_PRIV_INT;
5107 		}
5108 		break;
5109 	default:
5110 		panic("sfmmu_vtop_attr: bad mode %x", mode);
5111 	}
5112 	ASSERT(TTE_TO_TTEPFN(&ttevalue) == 0);
5113 	return (ttevalue.ll);
5114 }
5115 
5116 static uint_t
5117 sfmmu_ptov_attr(tte_t *ttep)
5118 {
5119 	uint_t attr;
5120 
5121 	ASSERT(TTE_IS_VALID(ttep));
5122 
5123 	attr = PROT_READ;
5124 
5125 	if (TTE_IS_WRITABLE(ttep)) {
5126 		attr |= PROT_WRITE;
5127 	}
5128 	if (TTE_IS_EXECUTABLE(ttep)) {
5129 		attr |= PROT_EXEC;
5130 	}
5131 	if (!TTE_IS_PRIVILEGED(ttep)) {
5132 		attr |= PROT_USER;
5133 	}
5134 	if (TTE_IS_NFO(ttep)) {
5135 		attr |= HAT_NOFAULT;
5136 	}
5137 	if (TTE_IS_NOSYNC(ttep)) {
5138 		attr |= HAT_NOSYNC;
5139 	}
5140 	if (TTE_IS_SIDEFFECT(ttep)) {
5141 		attr |= SFMMU_SIDEFFECT;
5142 	}
5143 	if (!TTE_IS_VCACHEABLE(ttep)) {
5144 		attr |= SFMMU_UNCACHEVTTE;
5145 	}
5146 	if (!TTE_IS_PCACHEABLE(ttep)) {
5147 		attr |= SFMMU_UNCACHEPTTE;
5148 	}
5149 	return (attr);
5150 }
5151 
5152 /*
5153  * hat_chgprot is a deprecated hat call.  New segment drivers
5154  * should store all attributes and use hat_*attr calls.
5155  *
5156  * Change the protections in the virtual address range
5157  * given to the specified virtual protection.  If vprot is ~PROT_WRITE,
5158  * then remove write permission, leaving the other
5159  * permissions unchanged.  If vprot is ~PROT_USER, remove user permissions.
5160  *
5161  */
5162 void
5163 hat_chgprot(struct hat *sfmmup, caddr_t addr, size_t len, uint_t vprot)
5164 {
5165 	struct hmehash_bucket *hmebp;
5166 	hmeblk_tag hblktag;
5167 	int hmeshift, hashno = 1;
5168 	struct hme_blk *hmeblkp, *list = NULL;
5169 	caddr_t endaddr;
5170 	cpuset_t cpuset;
5171 	demap_range_t dmr;
5172 
5173 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5174 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
5175 
5176 	ASSERT(sfmmup->sfmmu_as != NULL);
5177 
5178 	CPUSET_ZERO(cpuset);
5179 
5180 	if ((vprot != (uint_t)~PROT_WRITE) && (vprot & PROT_USER) &&
5181 	    ((addr + len) > (caddr_t)USERLIMIT)) {
5182 		panic("user addr %p vprot %x in kernel space",
5183 		    (void *)addr, vprot);
5184 	}
5185 	endaddr = addr + len;
5186 	hblktag.htag_id = sfmmup;
5187 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5188 	DEMAP_RANGE_INIT(sfmmup, &dmr);
5189 
5190 	while (addr < endaddr) {
5191 		hmeshift = HME_HASH_SHIFT(hashno);
5192 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5193 		hblktag.htag_rehash = hashno;
5194 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5195 
5196 		SFMMU_HASH_LOCK(hmebp);
5197 
5198 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
5199 		if (hmeblkp != NULL) {
5200 			ASSERT(!hmeblkp->hblk_shared);
5201 			/*
5202 			 * We've encountered a shadow hmeblk so skip the range
5203 			 * of the next smaller mapping size.
5204 			 */
5205 			if (hmeblkp->hblk_shw_bit) {
5206 				ASSERT(sfmmup != ksfmmup);
5207 				ASSERT(hashno > 1);
5208 				addr = (caddr_t)P2END((uintptr_t)addr,
5209 				    TTEBYTES(hashno - 1));
5210 			} else {
5211 				addr = sfmmu_hblk_chgprot(sfmmup, hmeblkp,
5212 				    addr, endaddr, &dmr, vprot);
5213 			}
5214 			SFMMU_HASH_UNLOCK(hmebp);
5215 			hashno = 1;
5216 			continue;
5217 		}
5218 		SFMMU_HASH_UNLOCK(hmebp);
5219 
5220 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
5221 			/*
5222 			 * We have traversed the whole list and rehashed
5223 			 * if necessary without finding the address to chgprot.
5224 			 * This is ok so we increment the address by the
5225 			 * smallest hmeblk range for kernel mappings and the
5226 			 * largest hmeblk range, to account for shadow hmeblks,
5227 			 * for user mappings and continue.
5228 			 */
5229 			if (sfmmup == ksfmmup)
5230 				addr = (caddr_t)P2END((uintptr_t)addr,
5231 				    TTEBYTES(1));
5232 			else
5233 				addr = (caddr_t)P2END((uintptr_t)addr,
5234 				    TTEBYTES(hashno));
5235 			hashno = 1;
5236 		} else {
5237 			hashno++;
5238 		}
5239 	}
5240 
5241 	sfmmu_hblks_list_purge(&list, 0);
5242 	DEMAP_RANGE_FLUSH(&dmr);
5243 	cpuset = sfmmup->sfmmu_cpusran;
5244 	xt_sync(cpuset);
5245 }
5246 
5247 /*
5248  * This function chgprots a range of addresses in an hmeblk.  It returns the
5249  * next addres that needs to be chgprot.
5250  * It should be called with the hash lock held.
5251  * XXX It shold be possible to optimize chgprot by not flushing every time but
5252  * on the other hand:
5253  * 1. do one flush crosscall.
5254  * 2. only flush if we are increasing permissions (make sure this will work)
5255  */
5256 static caddr_t
5257 sfmmu_hblk_chgprot(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5258     caddr_t endaddr, demap_range_t *dmrp, uint_t vprot)
5259 {
5260 	uint_t pprot;
5261 	tte_t tte, ttemod;
5262 	struct sf_hment *sfhmep;
5263 	uint_t tteflags;
5264 	int ttesz;
5265 	struct page *pp = NULL;
5266 	kmutex_t *pml, *pmtx;
5267 	int ret;
5268 	int use_demap_range;
5269 #if defined(SF_ERRATA_57)
5270 	int check_exec;
5271 #endif
5272 
5273 	ASSERT(in_hblk_range(hmeblkp, addr));
5274 	ASSERT(hmeblkp->hblk_shw_bit == 0);
5275 	ASSERT(!hmeblkp->hblk_shared);
5276 
5277 #ifdef DEBUG
5278 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5279 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5280 		panic("sfmmu_hblk_chgprot: partial chgprot of large page");
5281 	}
5282 #endif /* DEBUG */
5283 
5284 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5285 	ttesz = get_hblk_ttesz(hmeblkp);
5286 
5287 	pprot = sfmmu_vtop_prot(vprot, &tteflags);
5288 #if defined(SF_ERRATA_57)
5289 	check_exec = (sfmmup != ksfmmup) &&
5290 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
5291 	    ((vprot & PROT_EXEC) == PROT_EXEC);
5292 #endif
5293 	HBLKTOHME(sfhmep, hmeblkp, addr);
5294 
5295 	/*
5296 	 * Flush the current demap region if addresses have been
5297 	 * skipped or the page size doesn't match.
5298 	 */
5299 	use_demap_range = (TTEBYTES(ttesz) == MMU_PAGESIZE);
5300 	if (use_demap_range) {
5301 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5302 	} else if (dmrp != NULL) {
5303 		DEMAP_RANGE_FLUSH(dmrp);
5304 	}
5305 
5306 	while (addr < endaddr) {
5307 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5308 		if (TTE_IS_VALID(&tte)) {
5309 			if (TTE_GET_LOFLAGS(&tte, tteflags) == pprot) {
5310 				/*
5311 				 * if the new protection is the same as old
5312 				 * continue
5313 				 */
5314 				goto next_addr;
5315 			}
5316 			pml = NULL;
5317 			pp = sfhmep->hme_page;
5318 			if (pp) {
5319 				pml = sfmmu_mlist_enter(pp);
5320 			}
5321 			if (pp != sfhmep->hme_page) {
5322 				/*
5323 				 * tte most have been unloaded
5324 				 * underneath us.  Recheck
5325 				 */
5326 				ASSERT(pml);
5327 				sfmmu_mlist_exit(pml);
5328 				continue;
5329 			}
5330 
5331 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5332 
5333 			ttemod = tte;
5334 			TTE_SET_LOFLAGS(&ttemod, tteflags, pprot);
5335 #if defined(SF_ERRATA_57)
5336 			if (check_exec && addr < errata57_limit)
5337 				ttemod.tte_exec_perm = 0;
5338 #endif
5339 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5340 			    &sfhmep->hme_tte);
5341 
5342 			if (ret < 0) {
5343 				/* tte changed underneath us */
5344 				if (pml) {
5345 					sfmmu_mlist_exit(pml);
5346 				}
5347 				continue;
5348 			}
5349 
5350 			if (tteflags & TTE_HWWR_INT) {
5351 				/*
5352 				 * need to sync if we are clearing modify bit.
5353 				 */
5354 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5355 			}
5356 
5357 			if (pp && PP_ISRO(pp)) {
5358 				if (pprot & TTE_WRPRM_INT) {
5359 					pmtx = sfmmu_page_enter(pp);
5360 					PP_CLRRO(pp);
5361 					sfmmu_page_exit(pmtx);
5362 				}
5363 			}
5364 
5365 			if (ret > 0 && use_demap_range) {
5366 				DEMAP_RANGE_MARKPG(dmrp, addr);
5367 			} else if (ret > 0) {
5368 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5369 			}
5370 
5371 			if (pml) {
5372 				sfmmu_mlist_exit(pml);
5373 			}
5374 		}
5375 next_addr:
5376 		addr += TTEBYTES(ttesz);
5377 		sfhmep++;
5378 		DEMAP_RANGE_NEXTPG(dmrp);
5379 	}
5380 	return (addr);
5381 }
5382 
5383 /*
5384  * This routine is deprecated and should only be used by hat_chgprot.
5385  * The correct routine is sfmmu_vtop_attr.
5386  * This routine converts virtual page protections to physical ones.  It will
5387  * update the tteflags field with the tte mask corresponding to the protections
5388  * affected and it returns the new protections.  It will also clear the modify
5389  * bit if we are taking away write permission.  This is necessary since the
5390  * modify bit is the hardware permission bit and we need to clear it in order
5391  * to detect write faults.
5392  * It accepts the following special protections:
5393  * ~PROT_WRITE = remove write permissions.
5394  * ~PROT_USER = remove user permissions.
5395  */
5396 static uint_t
5397 sfmmu_vtop_prot(uint_t vprot, uint_t *tteflagsp)
5398 {
5399 	if (vprot == (uint_t)~PROT_WRITE) {
5400 		*tteflagsp = TTE_WRPRM_INT | TTE_HWWR_INT;
5401 		return (0);		/* will cause wrprm to be cleared */
5402 	}
5403 	if (vprot == (uint_t)~PROT_USER) {
5404 		*tteflagsp = TTE_PRIV_INT;
5405 		return (0);		/* will cause privprm to be cleared */
5406 	}
5407 	if ((vprot == 0) || (vprot == PROT_USER) ||
5408 	    ((vprot & PROT_ALL) != vprot)) {
5409 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5410 	}
5411 
5412 	switch (vprot) {
5413 	case (PROT_READ):
5414 	case (PROT_EXEC):
5415 	case (PROT_EXEC | PROT_READ):
5416 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5417 		return (TTE_PRIV_INT);		/* set prv and clr wrt */
5418 	case (PROT_WRITE):
5419 	case (PROT_WRITE | PROT_READ):
5420 	case (PROT_EXEC | PROT_WRITE):
5421 	case (PROT_EXEC | PROT_WRITE | PROT_READ):
5422 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5423 		return (TTE_PRIV_INT | TTE_WRPRM_INT);	/* set prv and wrt */
5424 	case (PROT_USER | PROT_READ):
5425 	case (PROT_USER | PROT_EXEC):
5426 	case (PROT_USER | PROT_EXEC | PROT_READ):
5427 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5428 		return (0);			/* clr prv and wrt */
5429 	case (PROT_USER | PROT_WRITE):
5430 	case (PROT_USER | PROT_WRITE | PROT_READ):
5431 	case (PROT_USER | PROT_EXEC | PROT_WRITE):
5432 	case (PROT_USER | PROT_EXEC | PROT_WRITE | PROT_READ):
5433 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5434 		return (TTE_WRPRM_INT);		/* clr prv and set wrt */
5435 	default:
5436 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5437 	}
5438 	return (0);
5439 }
5440 
5441 /*
5442  * Alternate unload for very large virtual ranges. With a true 64 bit VA,
5443  * the normal algorithm would take too long for a very large VA range with
5444  * few real mappings. This routine just walks thru all HMEs in the global
5445  * hash table to find and remove mappings.
5446  */
5447 static void
5448 hat_unload_large_virtual(struct hat *sfmmup, caddr_t startaddr, size_t len,
5449     uint_t flags, hat_callback_t *callback)
5450 {
5451 	struct hmehash_bucket *hmebp;
5452 	struct hme_blk *hmeblkp;
5453 	struct hme_blk *pr_hblk = NULL;
5454 	struct hme_blk *nx_hblk;
5455 	struct hme_blk *list = NULL;
5456 	int i;
5457 	demap_range_t dmr, *dmrp;
5458 	cpuset_t cpuset;
5459 	caddr_t	endaddr = startaddr + len;
5460 	caddr_t	sa;
5461 	caddr_t	ea;
5462 	caddr_t	cb_sa[MAX_CB_ADDR];
5463 	caddr_t	cb_ea[MAX_CB_ADDR];
5464 	int	addr_cnt = 0;
5465 	int	a = 0;
5466 
5467 	if (sfmmup->sfmmu_free) {
5468 		dmrp = NULL;
5469 	} else {
5470 		dmrp = &dmr;
5471 		DEMAP_RANGE_INIT(sfmmup, dmrp);
5472 	}
5473 
5474 	/*
5475 	 * Loop through all the hash buckets of HME blocks looking for matches.
5476 	 */
5477 	for (i = 0; i <= UHMEHASH_SZ; i++) {
5478 		hmebp = &uhme_hash[i];
5479 		SFMMU_HASH_LOCK(hmebp);
5480 		hmeblkp = hmebp->hmeblkp;
5481 		pr_hblk = NULL;
5482 		while (hmeblkp) {
5483 			nx_hblk = hmeblkp->hblk_next;
5484 
5485 			/*
5486 			 * skip if not this context, if a shadow block or
5487 			 * if the mapping is not in the requested range
5488 			 */
5489 			if (hmeblkp->hblk_tag.htag_id != sfmmup ||
5490 			    hmeblkp->hblk_shw_bit ||
5491 			    (sa = (caddr_t)get_hblk_base(hmeblkp)) >= endaddr ||
5492 			    (ea = get_hblk_endaddr(hmeblkp)) <= startaddr) {
5493 				pr_hblk = hmeblkp;
5494 				goto next_block;
5495 			}
5496 
5497 			ASSERT(!hmeblkp->hblk_shared);
5498 			/*
5499 			 * unload if there are any current valid mappings
5500 			 */
5501 			if (hmeblkp->hblk_vcnt != 0 ||
5502 			    hmeblkp->hblk_hmecnt != 0)
5503 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
5504 				    sa, ea, dmrp, flags);
5505 
5506 			/*
5507 			 * on unmap we also release the HME block itself, once
5508 			 * all mappings are gone.
5509 			 */
5510 			if ((flags & HAT_UNLOAD_UNMAP) != 0 &&
5511 			    !hmeblkp->hblk_vcnt &&
5512 			    !hmeblkp->hblk_hmecnt) {
5513 				ASSERT(!hmeblkp->hblk_lckcnt);
5514 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
5515 				    &list, 0);
5516 			} else {
5517 				pr_hblk = hmeblkp;
5518 			}
5519 
5520 			if (callback == NULL)
5521 				goto next_block;
5522 
5523 			/*
5524 			 * HME blocks may span more than one page, but we may be
5525 			 * unmapping only one page, so check for a smaller range
5526 			 * for the callback
5527 			 */
5528 			if (sa < startaddr)
5529 				sa = startaddr;
5530 			if (--ea > endaddr)
5531 				ea = endaddr - 1;
5532 
5533 			cb_sa[addr_cnt] = sa;
5534 			cb_ea[addr_cnt] = ea;
5535 			if (++addr_cnt == MAX_CB_ADDR) {
5536 				if (dmrp != NULL) {
5537 					DEMAP_RANGE_FLUSH(dmrp);
5538 					cpuset = sfmmup->sfmmu_cpusran;
5539 					xt_sync(cpuset);
5540 				}
5541 
5542 				for (a = 0; a < MAX_CB_ADDR; ++a) {
5543 					callback->hcb_start_addr = cb_sa[a];
5544 					callback->hcb_end_addr = cb_ea[a];
5545 					callback->hcb_function(callback);
5546 				}
5547 				addr_cnt = 0;
5548 			}
5549 
5550 next_block:
5551 			hmeblkp = nx_hblk;
5552 		}
5553 		SFMMU_HASH_UNLOCK(hmebp);
5554 	}
5555 
5556 	sfmmu_hblks_list_purge(&list, 0);
5557 	if (dmrp != NULL) {
5558 		DEMAP_RANGE_FLUSH(dmrp);
5559 		cpuset = sfmmup->sfmmu_cpusran;
5560 		xt_sync(cpuset);
5561 	}
5562 
5563 	for (a = 0; a < addr_cnt; ++a) {
5564 		callback->hcb_start_addr = cb_sa[a];
5565 		callback->hcb_end_addr = cb_ea[a];
5566 		callback->hcb_function(callback);
5567 	}
5568 
5569 	/*
5570 	 * Check TSB and TLB page sizes if the process isn't exiting.
5571 	 */
5572 	if (!sfmmup->sfmmu_free)
5573 		sfmmu_check_page_sizes(sfmmup, 0);
5574 }
5575 
5576 /*
5577  * Unload all the mappings in the range [addr..addr+len). addr and len must
5578  * be MMU_PAGESIZE aligned.
5579  */
5580 
5581 extern struct seg *segkmap;
5582 #define	ISSEGKMAP(sfmmup, addr) (sfmmup == ksfmmup && \
5583 segkmap->s_base <= (addr) && (addr) < (segkmap->s_base + segkmap->s_size))
5584 
5585 
5586 void
5587 hat_unload_callback(struct hat *sfmmup, caddr_t addr, size_t len, uint_t flags,
5588     hat_callback_t *callback)
5589 {
5590 	struct hmehash_bucket *hmebp;
5591 	hmeblk_tag hblktag;
5592 	int hmeshift, hashno, iskernel;
5593 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
5594 	caddr_t endaddr;
5595 	cpuset_t cpuset;
5596 	int addr_count = 0;
5597 	int a;
5598 	caddr_t cb_start_addr[MAX_CB_ADDR];
5599 	caddr_t cb_end_addr[MAX_CB_ADDR];
5600 	int issegkmap = ISSEGKMAP(sfmmup, addr);
5601 	demap_range_t dmr, *dmrp;
5602 
5603 	ASSERT(sfmmup->sfmmu_as != NULL);
5604 
5605 	ASSERT((sfmmup == ksfmmup) || (flags & HAT_UNLOAD_OTHER) || \
5606 	    AS_LOCK_HELD(sfmmup->sfmmu_as));
5607 
5608 	ASSERT(sfmmup != NULL);
5609 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5610 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
5611 
5612 	/*
5613 	 * Probing through a large VA range (say 63 bits) will be slow, even
5614 	 * at 4 Meg steps between the probes. So, when the virtual address range
5615 	 * is very large, search the HME entries for what to unload.
5616 	 *
5617 	 *	len >> TTE_PAGE_SHIFT(TTE4M) is the # of 4Meg probes we'd need
5618 	 *
5619 	 *	UHMEHASH_SZ is number of hash buckets to examine
5620 	 *
5621 	 */
5622 	if (sfmmup != KHATID && (len >> TTE_PAGE_SHIFT(TTE4M)) > UHMEHASH_SZ) {
5623 		hat_unload_large_virtual(sfmmup, addr, len, flags, callback);
5624 		return;
5625 	}
5626 
5627 	CPUSET_ZERO(cpuset);
5628 
5629 	/*
5630 	 * If the process is exiting, we can save a lot of fuss since
5631 	 * we'll flush the TLB when we free the ctx anyway.
5632 	 */
5633 	if (sfmmup->sfmmu_free) {
5634 		dmrp = NULL;
5635 	} else {
5636 		dmrp = &dmr;
5637 		DEMAP_RANGE_INIT(sfmmup, dmrp);
5638 	}
5639 
5640 	endaddr = addr + len;
5641 	hblktag.htag_id = sfmmup;
5642 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5643 
5644 	/*
5645 	 * It is likely for the vm to call unload over a wide range of
5646 	 * addresses that are actually very sparsely populated by
5647 	 * translations.  In order to speed this up the sfmmu hat supports
5648 	 * the concept of shadow hmeblks. Dummy large page hmeblks that
5649 	 * correspond to actual small translations are allocated at tteload
5650 	 * time and are referred to as shadow hmeblks.  Now, during unload
5651 	 * time, we first check if we have a shadow hmeblk for that
5652 	 * translation.  The absence of one means the corresponding address
5653 	 * range is empty and can be skipped.
5654 	 *
5655 	 * The kernel is an exception to above statement and that is why
5656 	 * we don't use shadow hmeblks and hash starting from the smallest
5657 	 * page size.
5658 	 */
5659 	if (sfmmup == KHATID) {
5660 		iskernel = 1;
5661 		hashno = TTE64K;
5662 	} else {
5663 		iskernel = 0;
5664 		if (mmu_page_sizes == max_mmu_page_sizes) {
5665 			hashno = TTE256M;
5666 		} else {
5667 			hashno = TTE4M;
5668 		}
5669 	}
5670 	while (addr < endaddr) {
5671 		hmeshift = HME_HASH_SHIFT(hashno);
5672 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5673 		hblktag.htag_rehash = hashno;
5674 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5675 
5676 		SFMMU_HASH_LOCK(hmebp);
5677 
5678 		HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, pr_hblk, &list);
5679 		if (hmeblkp == NULL) {
5680 			/*
5681 			 * didn't find an hmeblk. skip the appropiate
5682 			 * address range.
5683 			 */
5684 			SFMMU_HASH_UNLOCK(hmebp);
5685 			if (iskernel) {
5686 				if (hashno < mmu_hashcnt) {
5687 					hashno++;
5688 					continue;
5689 				} else {
5690 					hashno = TTE64K;
5691 					addr = (caddr_t)roundup((uintptr_t)addr
5692 					    + 1, MMU_PAGESIZE64K);
5693 					continue;
5694 				}
5695 			}
5696 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5697 			    (1 << hmeshift));
5698 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5699 				ASSERT(hashno == TTE64K);
5700 				continue;
5701 			}
5702 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5703 				hashno = TTE512K;
5704 				continue;
5705 			}
5706 			if (mmu_page_sizes == max_mmu_page_sizes) {
5707 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5708 					hashno = TTE4M;
5709 					continue;
5710 				}
5711 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5712 					hashno = TTE32M;
5713 					continue;
5714 				}
5715 				hashno = TTE256M;
5716 				continue;
5717 			} else {
5718 				hashno = TTE4M;
5719 				continue;
5720 			}
5721 		}
5722 		ASSERT(hmeblkp);
5723 		ASSERT(!hmeblkp->hblk_shared);
5724 		if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5725 			/*
5726 			 * If the valid count is zero we can skip the range
5727 			 * mapped by this hmeblk.
5728 			 * We free hblks in the case of HAT_UNMAP.  HAT_UNMAP
5729 			 * is used by segment drivers as a hint
5730 			 * that the mapping resource won't be used any longer.
5731 			 * The best example of this is during exit().
5732 			 */
5733 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5734 			    get_hblk_span(hmeblkp));
5735 			if ((flags & HAT_UNLOAD_UNMAP) ||
5736 			    (iskernel && !issegkmap)) {
5737 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk,
5738 				    &list, 0);
5739 			}
5740 			SFMMU_HASH_UNLOCK(hmebp);
5741 
5742 			if (iskernel) {
5743 				hashno = TTE64K;
5744 				continue;
5745 			}
5746 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5747 				ASSERT(hashno == TTE64K);
5748 				continue;
5749 			}
5750 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5751 				hashno = TTE512K;
5752 				continue;
5753 			}
5754 			if (mmu_page_sizes == max_mmu_page_sizes) {
5755 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5756 					hashno = TTE4M;
5757 					continue;
5758 				}
5759 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5760 					hashno = TTE32M;
5761 					continue;
5762 				}
5763 				hashno = TTE256M;
5764 				continue;
5765 			} else {
5766 				hashno = TTE4M;
5767 				continue;
5768 			}
5769 		}
5770 		if (hmeblkp->hblk_shw_bit) {
5771 			/*
5772 			 * If we encounter a shadow hmeblk we know there is
5773 			 * smaller sized hmeblks mapping the same address space.
5774 			 * Decrement the hash size and rehash.
5775 			 */
5776 			ASSERT(sfmmup != KHATID);
5777 			hashno--;
5778 			SFMMU_HASH_UNLOCK(hmebp);
5779 			continue;
5780 		}
5781 
5782 		/*
5783 		 * track callback address ranges.
5784 		 * only start a new range when it's not contiguous
5785 		 */
5786 		if (callback != NULL) {
5787 			if (addr_count > 0 &&
5788 			    addr == cb_end_addr[addr_count - 1])
5789 				--addr_count;
5790 			else
5791 				cb_start_addr[addr_count] = addr;
5792 		}
5793 
5794 		addr = sfmmu_hblk_unload(sfmmup, hmeblkp, addr, endaddr,
5795 		    dmrp, flags);
5796 
5797 		if (callback != NULL)
5798 			cb_end_addr[addr_count++] = addr;
5799 
5800 		if (((flags & HAT_UNLOAD_UNMAP) || (iskernel && !issegkmap)) &&
5801 		    !hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5802 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 0);
5803 		}
5804 		SFMMU_HASH_UNLOCK(hmebp);
5805 
5806 		/*
5807 		 * Notify our caller as to exactly which pages
5808 		 * have been unloaded. We do these in clumps,
5809 		 * to minimize the number of xt_sync()s that need to occur.
5810 		 */
5811 		if (callback != NULL && addr_count == MAX_CB_ADDR) {
5812 			if (dmrp != NULL) {
5813 				DEMAP_RANGE_FLUSH(dmrp);
5814 				cpuset = sfmmup->sfmmu_cpusran;
5815 				xt_sync(cpuset);
5816 			}
5817 
5818 			for (a = 0; a < MAX_CB_ADDR; ++a) {
5819 				callback->hcb_start_addr = cb_start_addr[a];
5820 				callback->hcb_end_addr = cb_end_addr[a];
5821 				callback->hcb_function(callback);
5822 			}
5823 			addr_count = 0;
5824 		}
5825 		if (iskernel) {
5826 			hashno = TTE64K;
5827 			continue;
5828 		}
5829 		if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5830 			ASSERT(hashno == TTE64K);
5831 			continue;
5832 		}
5833 		if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5834 			hashno = TTE512K;
5835 			continue;
5836 		}
5837 		if (mmu_page_sizes == max_mmu_page_sizes) {
5838 			if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5839 				hashno = TTE4M;
5840 				continue;
5841 			}
5842 			if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5843 				hashno = TTE32M;
5844 				continue;
5845 			}
5846 			hashno = TTE256M;
5847 		} else {
5848 			hashno = TTE4M;
5849 		}
5850 	}
5851 
5852 	sfmmu_hblks_list_purge(&list, 0);
5853 	if (dmrp != NULL) {
5854 		DEMAP_RANGE_FLUSH(dmrp);
5855 		cpuset = sfmmup->sfmmu_cpusran;
5856 		xt_sync(cpuset);
5857 	}
5858 	if (callback && addr_count != 0) {
5859 		for (a = 0; a < addr_count; ++a) {
5860 			callback->hcb_start_addr = cb_start_addr[a];
5861 			callback->hcb_end_addr = cb_end_addr[a];
5862 			callback->hcb_function(callback);
5863 		}
5864 	}
5865 
5866 	/*
5867 	 * Check TSB and TLB page sizes if the process isn't exiting.
5868 	 */
5869 	if (!sfmmup->sfmmu_free)
5870 		sfmmu_check_page_sizes(sfmmup, 0);
5871 }
5872 
5873 /*
5874  * Unload all the mappings in the range [addr..addr+len). addr and len must
5875  * be MMU_PAGESIZE aligned.
5876  */
5877 void
5878 hat_unload(struct hat *sfmmup, caddr_t addr, size_t len, uint_t flags)
5879 {
5880 	hat_unload_callback(sfmmup, addr, len, flags, NULL);
5881 }
5882 
5883 
5884 /*
5885  * Find the largest mapping size for this page.
5886  */
5887 int
5888 fnd_mapping_sz(page_t *pp)
5889 {
5890 	int sz;
5891 	int p_index;
5892 
5893 	p_index = PP_MAPINDEX(pp);
5894 
5895 	sz = 0;
5896 	p_index >>= 1;	/* don't care about 8K bit */
5897 	for (; p_index; p_index >>= 1) {
5898 		sz++;
5899 	}
5900 
5901 	return (sz);
5902 }
5903 
5904 /*
5905  * This function unloads a range of addresses for an hmeblk.
5906  * It returns the next address to be unloaded.
5907  * It should be called with the hash lock held.
5908  */
5909 static caddr_t
5910 sfmmu_hblk_unload(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5911     caddr_t endaddr, demap_range_t *dmrp, uint_t flags)
5912 {
5913 	tte_t	tte, ttemod;
5914 	struct	sf_hment *sfhmep;
5915 	int	ttesz;
5916 	long	ttecnt;
5917 	page_t *pp;
5918 	kmutex_t *pml;
5919 	int ret;
5920 	int use_demap_range;
5921 
5922 	ASSERT(in_hblk_range(hmeblkp, addr));
5923 	ASSERT(!hmeblkp->hblk_shw_bit);
5924 	ASSERT(sfmmup != NULL || hmeblkp->hblk_shared);
5925 	ASSERT(sfmmup == NULL || !hmeblkp->hblk_shared);
5926 	ASSERT(dmrp == NULL || !hmeblkp->hblk_shared);
5927 
5928 #ifdef DEBUG
5929 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5930 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5931 		panic("sfmmu_hblk_unload: partial unload of large page");
5932 	}
5933 #endif /* DEBUG */
5934 
5935 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5936 	ttesz = get_hblk_ttesz(hmeblkp);
5937 
5938 	use_demap_range = ((dmrp == NULL) ||
5939 	    (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp)));
5940 
5941 	if (use_demap_range) {
5942 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5943 	} else if (dmrp != NULL) {
5944 		DEMAP_RANGE_FLUSH(dmrp);
5945 	}
5946 	ttecnt = 0;
5947 	HBLKTOHME(sfhmep, hmeblkp, addr);
5948 
5949 	while (addr < endaddr) {
5950 		pml = NULL;
5951 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5952 		if (TTE_IS_VALID(&tte)) {
5953 			pp = sfhmep->hme_page;
5954 			if (pp != NULL) {
5955 				pml = sfmmu_mlist_enter(pp);
5956 			}
5957 
5958 			/*
5959 			 * Verify if hme still points to 'pp' now that
5960 			 * we have p_mapping lock.
5961 			 */
5962 			if (sfhmep->hme_page != pp) {
5963 				if (pp != NULL && sfhmep->hme_page != NULL) {
5964 					ASSERT(pml != NULL);
5965 					sfmmu_mlist_exit(pml);
5966 					/* Re-start this iteration. */
5967 					continue;
5968 				}
5969 				ASSERT((pp != NULL) &&
5970 				    (sfhmep->hme_page == NULL));
5971 				goto tte_unloaded;
5972 			}
5973 
5974 			/*
5975 			 * This point on we have both HASH and p_mapping
5976 			 * lock.
5977 			 */
5978 			ASSERT(pp == sfhmep->hme_page);
5979 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5980 
5981 			/*
5982 			 * We need to loop on modify tte because it is
5983 			 * possible for pagesync to come along and
5984 			 * change the software bits beneath us.
5985 			 *
5986 			 * Page_unload can also invalidate the tte after
5987 			 * we read tte outside of p_mapping lock.
5988 			 */
5989 again:
5990 			ttemod = tte;
5991 
5992 			TTE_SET_INVALID(&ttemod);
5993 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5994 			    &sfhmep->hme_tte);
5995 
5996 			if (ret <= 0) {
5997 				if (TTE_IS_VALID(&tte)) {
5998 					ASSERT(ret < 0);
5999 					goto again;
6000 				}
6001 				if (pp != NULL) {
6002 					panic("sfmmu_hblk_unload: pp = 0x%p "
6003 					    "tte became invalid under mlist"
6004 					    " lock = 0x%p", (void *)pp,
6005 					    (void *)pml);
6006 				}
6007 				continue;
6008 			}
6009 
6010 			if (!(flags & HAT_UNLOAD_NOSYNC)) {
6011 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
6012 			}
6013 
6014 			/*
6015 			 * Ok- we invalidated the tte. Do the rest of the job.
6016 			 */
6017 			ttecnt++;
6018 
6019 			if (flags & HAT_UNLOAD_UNLOCK) {
6020 				ASSERT(hmeblkp->hblk_lckcnt > 0);
6021 				atomic_dec_32(&hmeblkp->hblk_lckcnt);
6022 				HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
6023 			}
6024 
6025 			/*
6026 			 * Normally we would need to flush the page
6027 			 * from the virtual cache at this point in
6028 			 * order to prevent a potential cache alias
6029 			 * inconsistency.
6030 			 * The particular scenario we need to worry
6031 			 * about is:
6032 			 * Given:  va1 and va2 are two virtual address
6033 			 * that alias and map the same physical
6034 			 * address.
6035 			 * 1.   mapping exists from va1 to pa and data
6036 			 * has been read into the cache.
6037 			 * 2.   unload va1.
6038 			 * 3.   load va2 and modify data using va2.
6039 			 * 4    unload va2.
6040 			 * 5.   load va1 and reference data.  Unless we
6041 			 * flush the data cache when we unload we will
6042 			 * get stale data.
6043 			 * Fortunately, page coloring eliminates the
6044 			 * above scenario by remembering the color a
6045 			 * physical page was last or is currently
6046 			 * mapped to.  Now, we delay the flush until
6047 			 * the loading of translations.  Only when the
6048 			 * new translation is of a different color
6049 			 * are we forced to flush.
6050 			 */
6051 			if (use_demap_range) {
6052 				/*
6053 				 * Mark this page as needing a demap.
6054 				 */
6055 				DEMAP_RANGE_MARKPG(dmrp, addr);
6056 			} else {
6057 				ASSERT(sfmmup != NULL);
6058 				ASSERT(!hmeblkp->hblk_shared);
6059 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
6060 				    sfmmup->sfmmu_free, 0);
6061 			}
6062 
6063 			if (pp) {
6064 				/*
6065 				 * Remove the hment from the mapping list
6066 				 */
6067 				ASSERT(hmeblkp->hblk_hmecnt > 0);
6068 
6069 				/*
6070 				 * Again, we cannot
6071 				 * ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS);
6072 				 */
6073 				HME_SUB(sfhmep, pp);
6074 				membar_stst();
6075 				atomic_dec_16(&hmeblkp->hblk_hmecnt);
6076 			}
6077 
6078 			ASSERT(hmeblkp->hblk_vcnt > 0);
6079 			atomic_dec_16(&hmeblkp->hblk_vcnt);
6080 
6081 			ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
6082 			    !hmeblkp->hblk_lckcnt);
6083 
6084 #ifdef VAC
6085 			if (pp && (pp->p_nrm & (P_KPMC | P_KPMS | P_TNC))) {
6086 				if (PP_ISTNC(pp)) {
6087 					/*
6088 					 * If page was temporary
6089 					 * uncached, try to recache
6090 					 * it. Note that HME_SUB() was
6091 					 * called above so p_index and
6092 					 * mlist had been updated.
6093 					 */
6094 					conv_tnc(pp, ttesz);
6095 				} else if (pp->p_mapping == NULL) {
6096 					ASSERT(kpm_enable);
6097 					/*
6098 					 * Page is marked to be in VAC conflict
6099 					 * to an existing kpm mapping and/or is
6100 					 * kpm mapped using only the regular
6101 					 * pagesize.
6102 					 */
6103 					sfmmu_kpm_hme_unload(pp);
6104 				}
6105 			}
6106 #endif	/* VAC */
6107 		} else if ((pp = sfhmep->hme_page) != NULL) {
6108 				/*
6109 				 * TTE is invalid but the hme
6110 				 * still exists. let pageunload
6111 				 * complete its job.
6112 				 */
6113 				ASSERT(pml == NULL);
6114 				pml = sfmmu_mlist_enter(pp);
6115 				if (sfhmep->hme_page != NULL) {
6116 					sfmmu_mlist_exit(pml);
6117 					continue;
6118 				}
6119 				ASSERT(sfhmep->hme_page == NULL);
6120 		} else if (hmeblkp->hblk_hmecnt != 0) {
6121 			/*
6122 			 * pageunload may have not finished decrementing
6123 			 * hblk_vcnt and hblk_hmecnt. Find page_t if any and
6124 			 * wait for pageunload to finish. Rely on pageunload
6125 			 * to decrement hblk_hmecnt after hblk_vcnt.
6126 			 */
6127 			pfn_t pfn = TTE_TO_TTEPFN(&tte);
6128 			ASSERT(pml == NULL);
6129 			if (pf_is_memory(pfn)) {
6130 				pp = page_numtopp_nolock(pfn);
6131 				if (pp != NULL) {
6132 					pml = sfmmu_mlist_enter(pp);
6133 					sfmmu_mlist_exit(pml);
6134 					pml = NULL;
6135 				}
6136 			}
6137 		}
6138 
6139 tte_unloaded:
6140 		/*
6141 		 * At this point, the tte we are looking at
6142 		 * should be unloaded, and hme has been unlinked
6143 		 * from page too. This is important because in
6144 		 * pageunload, it does ttesync() then HME_SUB.
6145 		 * We need to make sure HME_SUB has been completed
6146 		 * so we know ttesync() has been completed. Otherwise,
6147 		 * at exit time, after return from hat layer, VM will
6148 		 * release as structure which hat_setstat() (called
6149 		 * by ttesync()) needs.
6150 		 */
6151 #ifdef DEBUG
6152 		{
6153 			tte_t	dtte;
6154 
6155 			ASSERT(sfhmep->hme_page == NULL);
6156 
6157 			sfmmu_copytte(&sfhmep->hme_tte, &dtte);
6158 			ASSERT(!TTE_IS_VALID(&dtte));
6159 		}
6160 #endif
6161 
6162 		if (pml) {
6163 			sfmmu_mlist_exit(pml);
6164 		}
6165 
6166 		addr += TTEBYTES(ttesz);
6167 		sfhmep++;
6168 		DEMAP_RANGE_NEXTPG(dmrp);
6169 	}
6170 	/*
6171 	 * For shared hmeblks this routine is only called when region is freed
6172 	 * and no longer referenced.  So no need to decrement ttecnt
6173 	 * in the region structure here.
6174 	 */
6175 	if (ttecnt > 0 && sfmmup != NULL) {
6176 		atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -ttecnt);
6177 	}
6178 	return (addr);
6179 }
6180 
6181 /*
6182  * Invalidate a virtual address range for the local CPU.
6183  * For best performance ensure that the va range is completely
6184  * mapped, otherwise the entire TLB will be flushed.
6185  */
6186 void
6187 hat_flush_range(struct hat *sfmmup, caddr_t va, size_t size)
6188 {
6189 	ssize_t sz;
6190 	caddr_t endva = va + size;
6191 
6192 	while (va < endva) {
6193 		sz = hat_getpagesize(sfmmup, va);
6194 		if (sz < 0) {
6195 			vtag_flushall();
6196 			break;
6197 		}
6198 		vtag_flushpage(va, (uint64_t)sfmmup);
6199 		va += sz;
6200 	}
6201 }
6202 
6203 /*
6204  * Synchronize all the mappings in the range [addr..addr+len).
6205  * Can be called with clearflag having two states:
6206  * HAT_SYNC_DONTZERO means just return the rm stats
6207  * HAT_SYNC_ZERORM means zero rm bits in the tte and return the stats
6208  */
6209 void
6210 hat_sync(struct hat *sfmmup, caddr_t addr, size_t len, uint_t clearflag)
6211 {
6212 	struct hmehash_bucket *hmebp;
6213 	hmeblk_tag hblktag;
6214 	int hmeshift, hashno = 1;
6215 	struct hme_blk *hmeblkp, *list = NULL;
6216 	caddr_t endaddr;
6217 	cpuset_t cpuset;
6218 
6219 	ASSERT((sfmmup == ksfmmup) || AS_LOCK_HELD(sfmmup->sfmmu_as));
6220 	ASSERT((len & MMU_PAGEOFFSET) == 0);
6221 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
6222 	    (clearflag == HAT_SYNC_ZERORM));
6223 
6224 	CPUSET_ZERO(cpuset);
6225 
6226 	endaddr = addr + len;
6227 	hblktag.htag_id = sfmmup;
6228 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
6229 
6230 	/*
6231 	 * Spitfire supports 4 page sizes.
6232 	 * Most pages are expected to be of the smallest page
6233 	 * size (8K) and these will not need to be rehashed. 64K
6234 	 * pages also don't need to be rehashed because the an hmeblk
6235 	 * spans 64K of address space. 512K pages might need 1 rehash and
6236 	 * and 4M pages 2 rehashes.
6237 	 */
6238 	while (addr < endaddr) {
6239 		hmeshift = HME_HASH_SHIFT(hashno);
6240 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
6241 		hblktag.htag_rehash = hashno;
6242 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
6243 
6244 		SFMMU_HASH_LOCK(hmebp);
6245 
6246 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
6247 		if (hmeblkp != NULL) {
6248 			ASSERT(!hmeblkp->hblk_shared);
6249 			/*
6250 			 * We've encountered a shadow hmeblk so skip the range
6251 			 * of the next smaller mapping size.
6252 			 */
6253 			if (hmeblkp->hblk_shw_bit) {
6254 				ASSERT(sfmmup != ksfmmup);
6255 				ASSERT(hashno > 1);
6256 				addr = (caddr_t)P2END((uintptr_t)addr,
6257 				    TTEBYTES(hashno - 1));
6258 			} else {
6259 				addr = sfmmu_hblk_sync(sfmmup, hmeblkp,
6260 				    addr, endaddr, clearflag);
6261 			}
6262 			SFMMU_HASH_UNLOCK(hmebp);
6263 			hashno = 1;
6264 			continue;
6265 		}
6266 		SFMMU_HASH_UNLOCK(hmebp);
6267 
6268 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
6269 			/*
6270 			 * We have traversed the whole list and rehashed
6271 			 * if necessary without finding the address to sync.
6272 			 * This is ok so we increment the address by the
6273 			 * smallest hmeblk range for kernel mappings and the
6274 			 * largest hmeblk range, to account for shadow hmeblks,
6275 			 * for user mappings and continue.
6276 			 */
6277 			if (sfmmup == ksfmmup)
6278 				addr = (caddr_t)P2END((uintptr_t)addr,
6279 				    TTEBYTES(1));
6280 			else
6281 				addr = (caddr_t)P2END((uintptr_t)addr,
6282 				    TTEBYTES(hashno));
6283 			hashno = 1;
6284 		} else {
6285 			hashno++;
6286 		}
6287 	}
6288 	sfmmu_hblks_list_purge(&list, 0);
6289 	cpuset = sfmmup->sfmmu_cpusran;
6290 	xt_sync(cpuset);
6291 }
6292 
6293 static caddr_t
6294 sfmmu_hblk_sync(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
6295     caddr_t endaddr, int clearflag)
6296 {
6297 	tte_t	tte, ttemod;
6298 	struct sf_hment *sfhmep;
6299 	int ttesz;
6300 	struct page *pp;
6301 	kmutex_t *pml;
6302 	int ret;
6303 
6304 	ASSERT(hmeblkp->hblk_shw_bit == 0);
6305 	ASSERT(!hmeblkp->hblk_shared);
6306 
6307 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
6308 
6309 	ttesz = get_hblk_ttesz(hmeblkp);
6310 	HBLKTOHME(sfhmep, hmeblkp, addr);
6311 
6312 	while (addr < endaddr) {
6313 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6314 		if (TTE_IS_VALID(&tte)) {
6315 			pml = NULL;
6316 			pp = sfhmep->hme_page;
6317 			if (pp) {
6318 				pml = sfmmu_mlist_enter(pp);
6319 			}
6320 			if (pp != sfhmep->hme_page) {
6321 				/*
6322 				 * tte most have been unloaded
6323 				 * underneath us.  Recheck
6324 				 */
6325 				ASSERT(pml);
6326 				sfmmu_mlist_exit(pml);
6327 				continue;
6328 			}
6329 
6330 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
6331 
6332 			if (clearflag == HAT_SYNC_ZERORM) {
6333 				ttemod = tte;
6334 				TTE_CLR_RM(&ttemod);
6335 				ret = sfmmu_modifytte_try(&tte, &ttemod,
6336 				    &sfhmep->hme_tte);
6337 				if (ret < 0) {
6338 					if (pml) {
6339 						sfmmu_mlist_exit(pml);
6340 					}
6341 					continue;
6342 				}
6343 
6344 				if (ret > 0) {
6345 					sfmmu_tlb_demap(addr, sfmmup,
6346 					    hmeblkp, 0, 0);
6347 				}
6348 			}
6349 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
6350 			if (pml) {
6351 				sfmmu_mlist_exit(pml);
6352 			}
6353 		}
6354 		addr += TTEBYTES(ttesz);
6355 		sfhmep++;
6356 	}
6357 	return (addr);
6358 }
6359 
6360 /*
6361  * This function will sync a tte to the page struct and it will
6362  * update the hat stats. Currently it allows us to pass a NULL pp
6363  * and we will simply update the stats.  We may want to change this
6364  * so we only keep stats for pages backed by pp's.
6365  */
6366 static void
6367 sfmmu_ttesync(struct hat *sfmmup, caddr_t addr, tte_t *ttep, page_t *pp)
6368 {
6369 	uint_t rm = 0;
6370 	int	sz;
6371 	pgcnt_t	npgs;
6372 
6373 	ASSERT(TTE_IS_VALID(ttep));
6374 
6375 	if (TTE_IS_NOSYNC(ttep)) {
6376 		return;
6377 	}
6378 
6379 	if (TTE_IS_REF(ttep))  {
6380 		rm = P_REF;
6381 	}
6382 	if (TTE_IS_MOD(ttep))  {
6383 		rm |= P_MOD;
6384 	}
6385 
6386 	if (rm == 0) {
6387 		return;
6388 	}
6389 
6390 	sz = TTE_CSZ(ttep);
6391 	if (sfmmup != NULL && sfmmup->sfmmu_rmstat) {
6392 		int i;
6393 		caddr_t	vaddr = addr;
6394 
6395 		for (i = 0; i < TTEPAGES(sz); i++, vaddr += MMU_PAGESIZE) {
6396 			hat_setstat(sfmmup->sfmmu_as, vaddr, MMU_PAGESIZE, rm);
6397 		}
6398 
6399 	}
6400 
6401 	/*
6402 	 * XXX I want to use cas to update nrm bits but they
6403 	 * currently belong in common/vm and not in hat where
6404 	 * they should be.
6405 	 * The nrm bits are protected by the same mutex as
6406 	 * the one that protects the page's mapping list.
6407 	 */
6408 	if (!pp)
6409 		return;
6410 	ASSERT(sfmmu_mlist_held(pp));
6411 	/*
6412 	 * If the tte is for a large page, we need to sync all the
6413 	 * pages covered by the tte.
6414 	 */
6415 	if (sz != TTE8K) {
6416 		ASSERT(pp->p_szc != 0);
6417 		pp = PP_GROUPLEADER(pp, sz);
6418 		ASSERT(sfmmu_mlist_held(pp));
6419 	}
6420 
6421 	/* Get number of pages from tte size. */
6422 	npgs = TTEPAGES(sz);
6423 
6424 	do {
6425 		ASSERT(pp);
6426 		ASSERT(sfmmu_mlist_held(pp));
6427 		if (((rm & P_REF) != 0 && !PP_ISREF(pp)) ||
6428 		    ((rm & P_MOD) != 0 && !PP_ISMOD(pp)))
6429 			hat_page_setattr(pp, rm);
6430 
6431 		/*
6432 		 * Are we done? If not, we must have a large mapping.
6433 		 * For large mappings we need to sync the rest of the pages
6434 		 * covered by this tte; goto the next page.
6435 		 */
6436 	} while (--npgs > 0 && (pp = PP_PAGENEXT(pp)));
6437 }
6438 
6439 /*
6440  * Execute pre-callback handler of each pa_hment linked to pp
6441  *
6442  * Inputs:
6443  *   flag: either HAT_PRESUSPEND or HAT_SUSPEND.
6444  *   capture_cpus: pointer to return value (below)
6445  *
6446  * Returns:
6447  *   Propagates the subsystem callback return values back to the caller;
6448  *   returns 0 on success.  If capture_cpus is non-NULL, the value returned
6449  *   is zero if all of the pa_hments are of a type that do not require
6450  *   capturing CPUs prior to suspending the mapping, else it is 1.
6451  */
6452 static int
6453 hat_pageprocess_precallbacks(struct page *pp, uint_t flag, int *capture_cpus)
6454 {
6455 	struct sf_hment	*sfhmep;
6456 	struct pa_hment *pahmep;
6457 	int (*f)(caddr_t, uint_t, uint_t, void *);
6458 	int		ret;
6459 	id_t		id;
6460 	int		locked = 0;
6461 	kmutex_t	*pml;
6462 
6463 	ASSERT(PAGE_EXCL(pp));
6464 	if (!sfmmu_mlist_held(pp)) {
6465 		pml = sfmmu_mlist_enter(pp);
6466 		locked = 1;
6467 	}
6468 
6469 	if (capture_cpus)
6470 		*capture_cpus = 0;
6471 
6472 top:
6473 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6474 		/*
6475 		 * skip sf_hments corresponding to VA<->PA mappings;
6476 		 * for pa_hment's, hme_tte.ll is zero
6477 		 */
6478 		if (!IS_PAHME(sfhmep))
6479 			continue;
6480 
6481 		pahmep = sfhmep->hme_data;
6482 		ASSERT(pahmep != NULL);
6483 
6484 		/*
6485 		 * skip if pre-handler has been called earlier in this loop
6486 		 */
6487 		if (pahmep->flags & flag)
6488 			continue;
6489 
6490 		id = pahmep->cb_id;
6491 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6492 		if (capture_cpus && sfmmu_cb_table[id].capture_cpus != 0)
6493 			*capture_cpus = 1;
6494 		if ((f = sfmmu_cb_table[id].prehandler) == NULL) {
6495 			pahmep->flags |= flag;
6496 			continue;
6497 		}
6498 
6499 		/*
6500 		 * Drop the mapping list lock to avoid locking order issues.
6501 		 */
6502 		if (locked)
6503 			sfmmu_mlist_exit(pml);
6504 
6505 		ret = f(pahmep->addr, pahmep->len, flag, pahmep->pvt);
6506 		if (ret != 0)
6507 			return (ret);	/* caller must do the cleanup */
6508 
6509 		if (locked) {
6510 			pml = sfmmu_mlist_enter(pp);
6511 			pahmep->flags |= flag;
6512 			goto top;
6513 		}
6514 
6515 		pahmep->flags |= flag;
6516 	}
6517 
6518 	if (locked)
6519 		sfmmu_mlist_exit(pml);
6520 
6521 	return (0);
6522 }
6523 
6524 /*
6525  * Execute post-callback handler of each pa_hment linked to pp
6526  *
6527  * Same overall assumptions and restrictions apply as for
6528  * hat_pageprocess_precallbacks().
6529  */
6530 static void
6531 hat_pageprocess_postcallbacks(struct page *pp, uint_t flag)
6532 {
6533 	pfn_t pgpfn = pp->p_pagenum;
6534 	pfn_t pgmask = btop(page_get_pagesize(pp->p_szc)) - 1;
6535 	pfn_t newpfn;
6536 	struct sf_hment *sfhmep;
6537 	struct pa_hment *pahmep;
6538 	int (*f)(caddr_t, uint_t, uint_t, void *, pfn_t);
6539 	id_t	id;
6540 	int	locked = 0;
6541 	kmutex_t *pml;
6542 
6543 	ASSERT(PAGE_EXCL(pp));
6544 	if (!sfmmu_mlist_held(pp)) {
6545 		pml = sfmmu_mlist_enter(pp);
6546 		locked = 1;
6547 	}
6548 
6549 top:
6550 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6551 		/*
6552 		 * skip sf_hments corresponding to VA<->PA mappings;
6553 		 * for pa_hment's, hme_tte.ll is zero
6554 		 */
6555 		if (!IS_PAHME(sfhmep))
6556 			continue;
6557 
6558 		pahmep = sfhmep->hme_data;
6559 		ASSERT(pahmep != NULL);
6560 
6561 		if ((pahmep->flags & flag) == 0)
6562 			continue;
6563 
6564 		pahmep->flags &= ~flag;
6565 
6566 		id = pahmep->cb_id;
6567 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6568 		if ((f = sfmmu_cb_table[id].posthandler) == NULL)
6569 			continue;
6570 
6571 		/*
6572 		 * Convert the base page PFN into the constituent PFN
6573 		 * which is needed by the callback handler.
6574 		 */
6575 		newpfn = pgpfn | (btop((uintptr_t)pahmep->addr) & pgmask);
6576 
6577 		/*
6578 		 * Drop the mapping list lock to avoid locking order issues.
6579 		 */
6580 		if (locked)
6581 			sfmmu_mlist_exit(pml);
6582 
6583 		if (f(pahmep->addr, pahmep->len, flag, pahmep->pvt, newpfn)
6584 		    != 0)
6585 			panic("sfmmu: posthandler failed");
6586 
6587 		if (locked) {
6588 			pml = sfmmu_mlist_enter(pp);
6589 			goto top;
6590 		}
6591 	}
6592 
6593 	if (locked)
6594 		sfmmu_mlist_exit(pml);
6595 }
6596 
6597 /*
6598  * Suspend locked kernel mapping
6599  */
6600 void
6601 hat_pagesuspend(struct page *pp)
6602 {
6603 	struct sf_hment *sfhmep;
6604 	sfmmu_t *sfmmup;
6605 	tte_t tte, ttemod;
6606 	struct hme_blk *hmeblkp;
6607 	caddr_t addr;
6608 	int index, cons;
6609 	cpuset_t cpuset;
6610 
6611 	ASSERT(PAGE_EXCL(pp));
6612 	ASSERT(sfmmu_mlist_held(pp));
6613 
6614 	mutex_enter(&kpr_suspendlock);
6615 
6616 	/*
6617 	 * We're about to suspend a kernel mapping so mark this thread as
6618 	 * non-traceable by DTrace. This prevents us from running into issues
6619 	 * with probe context trying to touch a suspended page
6620 	 * in the relocation codepath itself.
6621 	 */
6622 	curthread->t_flag |= T_DONTDTRACE;
6623 
6624 	index = PP_MAPINDEX(pp);
6625 	cons = TTE8K;
6626 
6627 retry:
6628 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6629 
6630 		if (IS_PAHME(sfhmep))
6631 			continue;
6632 
6633 		if (get_hblk_ttesz(sfmmu_hmetohblk(sfhmep)) != cons)
6634 			continue;
6635 
6636 		/*
6637 		 * Loop until we successfully set the suspend bit in
6638 		 * the TTE.
6639 		 */
6640 again:
6641 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6642 		ASSERT(TTE_IS_VALID(&tte));
6643 
6644 		ttemod = tte;
6645 		TTE_SET_SUSPEND(&ttemod);
6646 		if (sfmmu_modifytte_try(&tte, &ttemod,
6647 		    &sfhmep->hme_tte) < 0)
6648 			goto again;
6649 
6650 		/*
6651 		 * Invalidate TSB entry
6652 		 */
6653 		hmeblkp = sfmmu_hmetohblk(sfhmep);
6654 
6655 		sfmmup = hblktosfmmu(hmeblkp);
6656 		ASSERT(sfmmup == ksfmmup);
6657 		ASSERT(!hmeblkp->hblk_shared);
6658 
6659 		addr = tte_to_vaddr(hmeblkp, tte);
6660 
6661 		/*
6662 		 * No need to make sure that the TSB for this sfmmu is
6663 		 * not being relocated since it is ksfmmup and thus it
6664 		 * will never be relocated.
6665 		 */
6666 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
6667 
6668 		/*
6669 		 * Update xcall stats
6670 		 */
6671 		cpuset = cpu_ready_set;
6672 		CPUSET_DEL(cpuset, CPU->cpu_id);
6673 
6674 		/* LINTED: constant in conditional context */
6675 		SFMMU_XCALL_STATS(ksfmmup);
6676 
6677 		/*
6678 		 * Flush TLB entry on remote CPU's
6679 		 */
6680 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
6681 		    (uint64_t)ksfmmup);
6682 		xt_sync(cpuset);
6683 
6684 		/*
6685 		 * Flush TLB entry on local CPU
6686 		 */
6687 		vtag_flushpage(addr, (uint64_t)ksfmmup);
6688 	}
6689 
6690 	while (index != 0) {
6691 		index = index >> 1;
6692 		if (index != 0)
6693 			cons++;
6694 		if (index & 0x1) {
6695 			pp = PP_GROUPLEADER(pp, cons);
6696 			goto retry;
6697 		}
6698 	}
6699 }
6700 
6701 #ifdef	DEBUG
6702 
6703 #define	N_PRLE	1024
6704 struct prle {
6705 	page_t *targ;
6706 	page_t *repl;
6707 	int status;
6708 	int pausecpus;
6709 	hrtime_t whence;
6710 };
6711 
6712 static struct prle page_relocate_log[N_PRLE];
6713 static int prl_entry;
6714 static kmutex_t prl_mutex;
6715 
6716 #define	PAGE_RELOCATE_LOG(t, r, s, p)					\
6717 	mutex_enter(&prl_mutex);					\
6718 	page_relocate_log[prl_entry].targ = *(t);			\
6719 	page_relocate_log[prl_entry].repl = *(r);			\
6720 	page_relocate_log[prl_entry].status = (s);			\
6721 	page_relocate_log[prl_entry].pausecpus = (p);			\
6722 	page_relocate_log[prl_entry].whence = gethrtime();		\
6723 	prl_entry = (prl_entry == (N_PRLE - 1))? 0 : prl_entry + 1;	\
6724 	mutex_exit(&prl_mutex);
6725 
6726 #else	/* !DEBUG */
6727 #define	PAGE_RELOCATE_LOG(t, r, s, p)
6728 #endif
6729 
6730 /*
6731  * Core Kernel Page Relocation Algorithm
6732  *
6733  * Input:
6734  *
6735  * target :	constituent pages are SE_EXCL locked.
6736  * replacement:	constituent pages are SE_EXCL locked.
6737  *
6738  * Output:
6739  *
6740  * nrelocp:	number of pages relocated
6741  */
6742 int
6743 hat_page_relocate(page_t **target, page_t **replacement, spgcnt_t *nrelocp)
6744 {
6745 	page_t		*targ, *repl;
6746 	page_t		*tpp, *rpp;
6747 	kmutex_t	*low, *high;
6748 	spgcnt_t	npages, i;
6749 	page_t		*pl = NULL;
6750 	int		old_pil;
6751 	cpuset_t	cpuset;
6752 	int		cap_cpus;
6753 	int		ret;
6754 #ifdef VAC
6755 	int		cflags = 0;
6756 #endif
6757 
6758 	if (!kcage_on || PP_ISNORELOC(*target)) {
6759 		PAGE_RELOCATE_LOG(target, replacement, EAGAIN, -1);
6760 		return (EAGAIN);
6761 	}
6762 
6763 	mutex_enter(&kpr_mutex);
6764 	kreloc_thread = curthread;
6765 
6766 	targ = *target;
6767 	repl = *replacement;
6768 	ASSERT(repl != NULL);
6769 	ASSERT(targ->p_szc == repl->p_szc);
6770 
6771 	npages = page_get_pagecnt(targ->p_szc);
6772 
6773 	/*
6774 	 * unload VA<->PA mappings that are not locked
6775 	 */
6776 	tpp = targ;
6777 	for (i = 0; i < npages; i++) {
6778 		(void) hat_pageunload(tpp, SFMMU_KERNEL_RELOC);
6779 		tpp++;
6780 	}
6781 
6782 	/*
6783 	 * Do "presuspend" callbacks, in a context from which we can still
6784 	 * block as needed. Note that we don't hold the mapping list lock
6785 	 * of "targ" at this point due to potential locking order issues;
6786 	 * we assume that between the hat_pageunload() above and holding
6787 	 * the SE_EXCL lock that the mapping list *cannot* change at this
6788 	 * point.
6789 	 */
6790 	ret = hat_pageprocess_precallbacks(targ, HAT_PRESUSPEND, &cap_cpus);
6791 	if (ret != 0) {
6792 		/*
6793 		 * EIO translates to fatal error, for all others cleanup
6794 		 * and return EAGAIN.
6795 		 */
6796 		ASSERT(ret != EIO);
6797 		hat_pageprocess_postcallbacks(targ, HAT_POSTUNSUSPEND);
6798 		PAGE_RELOCATE_LOG(target, replacement, ret, -1);
6799 		kreloc_thread = NULL;
6800 		mutex_exit(&kpr_mutex);
6801 		return (EAGAIN);
6802 	}
6803 
6804 	/*
6805 	 * acquire p_mapping list lock for both the target and replacement
6806 	 * root pages.
6807 	 *
6808 	 * low and high refer to the need to grab the mlist locks in a
6809 	 * specific order in order to prevent race conditions.  Thus the
6810 	 * lower lock must be grabbed before the higher lock.
6811 	 *
6812 	 * This will block hat_unload's accessing p_mapping list.  Since
6813 	 * we have SE_EXCL lock, hat_memload and hat_pageunload will be
6814 	 * blocked.  Thus, no one else will be accessing the p_mapping list
6815 	 * while we suspend and reload the locked mapping below.
6816 	 */
6817 	tpp = targ;
6818 	rpp = repl;
6819 	sfmmu_mlist_reloc_enter(tpp, rpp, &low, &high);
6820 
6821 	kpreempt_disable();
6822 
6823 	/*
6824 	 * We raise our PIL to 13 so that we don't get captured by
6825 	 * another CPU or pinned by an interrupt thread.  We can't go to
6826 	 * PIL 14 since the nexus driver(s) may need to interrupt at
6827 	 * that level in the case of IOMMU pseudo mappings.
6828 	 */
6829 	cpuset = cpu_ready_set;
6830 	CPUSET_DEL(cpuset, CPU->cpu_id);
6831 	if (!cap_cpus || CPUSET_ISNULL(cpuset)) {
6832 		old_pil = splr(XCALL_PIL);
6833 	} else {
6834 		old_pil = -1;
6835 		xc_attention(cpuset);
6836 	}
6837 	ASSERT(getpil() == XCALL_PIL);
6838 
6839 	/*
6840 	 * Now do suspend callbacks. In the case of an IOMMU mapping
6841 	 * this will suspend all DMA activity to the page while it is
6842 	 * being relocated. Since we are well above LOCK_LEVEL and CPUs
6843 	 * may be captured at this point we should have acquired any needed
6844 	 * locks in the presuspend callback.
6845 	 */
6846 	ret = hat_pageprocess_precallbacks(targ, HAT_SUSPEND, NULL);
6847 	if (ret != 0) {
6848 		repl = targ;
6849 		goto suspend_fail;
6850 	}
6851 
6852 	/*
6853 	 * Raise the PIL yet again, this time to block all high-level
6854 	 * interrupts on this CPU. This is necessary to prevent an
6855 	 * interrupt routine from pinning the thread which holds the
6856 	 * mapping suspended and then touching the suspended page.
6857 	 *
6858 	 * Once the page is suspended we also need to be careful to
6859 	 * avoid calling any functions which touch any seg_kmem memory
6860 	 * since that memory may be backed by the very page we are
6861 	 * relocating in here!
6862 	 */
6863 	hat_pagesuspend(targ);
6864 
6865 	/*
6866 	 * Now that we are confident everybody has stopped using this page,
6867 	 * copy the page contents.  Note we use a physical copy to prevent
6868 	 * locking issues and to avoid fpRAS because we can't handle it in
6869 	 * this context.
6870 	 */
6871 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6872 #ifdef VAC
6873 		/*
6874 		 * If the replacement has a different vcolor than
6875 		 * the one being replacd, we need to handle VAC
6876 		 * consistency for it just as we were setting up
6877 		 * a new mapping to it.
6878 		 */
6879 		if ((PP_GET_VCOLOR(rpp) != NO_VCOLOR) &&
6880 		    (tpp->p_vcolor != rpp->p_vcolor) &&
6881 		    !CacheColor_IsFlushed(cflags, PP_GET_VCOLOR(rpp))) {
6882 			CacheColor_SetFlushed(cflags, PP_GET_VCOLOR(rpp));
6883 			sfmmu_cache_flushcolor(PP_GET_VCOLOR(rpp),
6884 			    rpp->p_pagenum);
6885 		}
6886 #endif
6887 		/*
6888 		 * Copy the contents of the page.
6889 		 */
6890 		ppcopy_kernel(tpp, rpp);
6891 	}
6892 
6893 	tpp = targ;
6894 	rpp = repl;
6895 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6896 		/*
6897 		 * Copy attributes.  VAC consistency was handled above,
6898 		 * if required.
6899 		 */
6900 		rpp->p_nrm = tpp->p_nrm;
6901 		tpp->p_nrm = 0;
6902 		rpp->p_index = tpp->p_index;
6903 		tpp->p_index = 0;
6904 #ifdef VAC
6905 		rpp->p_vcolor = tpp->p_vcolor;
6906 #endif
6907 	}
6908 
6909 	/*
6910 	 * First, unsuspend the page, if we set the suspend bit, and transfer
6911 	 * the mapping list from the target page to the replacement page.
6912 	 * Next process postcallbacks; since pa_hment's are linked only to the
6913 	 * p_mapping list of root page, we don't iterate over the constituent
6914 	 * pages.
6915 	 */
6916 	hat_pagereload(targ, repl);
6917 
6918 suspend_fail:
6919 	hat_pageprocess_postcallbacks(repl, HAT_UNSUSPEND);
6920 
6921 	/*
6922 	 * Now lower our PIL and release any captured CPUs since we
6923 	 * are out of the "danger zone".  After this it will again be
6924 	 * safe to acquire adaptive mutex locks, or to drop them...
6925 	 */
6926 	if (old_pil != -1) {
6927 		splx(old_pil);
6928 	} else {
6929 		xc_dismissed(cpuset);
6930 	}
6931 
6932 	kpreempt_enable();
6933 
6934 	sfmmu_mlist_reloc_exit(low, high);
6935 
6936 	/*
6937 	 * Postsuspend callbacks should drop any locks held across
6938 	 * the suspend callbacks.  As before, we don't hold the mapping
6939 	 * list lock at this point.. our assumption is that the mapping
6940 	 * list still can't change due to our holding SE_EXCL lock and
6941 	 * there being no unlocked mappings left. Hence the restriction
6942 	 * on calling context to hat_delete_callback()
6943 	 */
6944 	hat_pageprocess_postcallbacks(repl, HAT_POSTUNSUSPEND);
6945 	if (ret != 0) {
6946 		/*
6947 		 * The second presuspend call failed: we got here through
6948 		 * the suspend_fail label above.
6949 		 */
6950 		ASSERT(ret != EIO);
6951 		PAGE_RELOCATE_LOG(target, replacement, ret, cap_cpus);
6952 		kreloc_thread = NULL;
6953 		mutex_exit(&kpr_mutex);
6954 		return (EAGAIN);
6955 	}
6956 
6957 	/*
6958 	 * Now that we're out of the performance critical section we can
6959 	 * take care of updating the hash table, since we still
6960 	 * hold all the pages locked SE_EXCL at this point we
6961 	 * needn't worry about things changing out from under us.
6962 	 */
6963 	tpp = targ;
6964 	rpp = repl;
6965 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6966 
6967 		/*
6968 		 * replace targ with replacement in page_hash table
6969 		 */
6970 		targ = tpp;
6971 		page_relocate_hash(rpp, targ);
6972 
6973 		/*
6974 		 * concatenate target; caller of platform_page_relocate()
6975 		 * expects target to be concatenated after returning.
6976 		 */
6977 		ASSERT(targ->p_next == targ);
6978 		ASSERT(targ->p_prev == targ);
6979 		page_list_concat(&pl, &targ);
6980 	}
6981 
6982 	ASSERT(*target == pl);
6983 	*nrelocp = npages;
6984 	PAGE_RELOCATE_LOG(target, replacement, 0, cap_cpus);
6985 	kreloc_thread = NULL;
6986 	mutex_exit(&kpr_mutex);
6987 	return (0);
6988 }
6989 
6990 /*
6991  * Called when stray pa_hments are found attached to a page which is
6992  * being freed.  Notify the subsystem which attached the pa_hment of
6993  * the error if it registered a suitable handler, else panic.
6994  */
6995 static void
6996 sfmmu_pahment_leaked(struct pa_hment *pahmep)
6997 {
6998 	id_t cb_id = pahmep->cb_id;
6999 
7000 	ASSERT(cb_id >= (id_t)0 && cb_id < sfmmu_cb_nextid);
7001 	if (sfmmu_cb_table[cb_id].errhandler != NULL) {
7002 		if (sfmmu_cb_table[cb_id].errhandler(pahmep->addr, pahmep->len,
7003 		    HAT_CB_ERR_LEAKED, pahmep->pvt) == 0)
7004 			return;		/* non-fatal */
7005 	}
7006 	panic("pa_hment leaked: 0x%p", (void *)pahmep);
7007 }
7008 
7009 /*
7010  * Remove all mappings to page 'pp'.
7011  */
7012 int
7013 hat_pageunload(struct page *pp, uint_t forceflag)
7014 {
7015 	struct page *origpp = pp;
7016 	struct sf_hment *sfhme, *tmphme;
7017 	struct hme_blk *hmeblkp;
7018 	kmutex_t *pml;
7019 #ifdef VAC
7020 	kmutex_t *pmtx;
7021 #endif
7022 	cpuset_t cpuset, tset;
7023 	int index, cons;
7024 	int pa_hments;
7025 
7026 	ASSERT(PAGE_EXCL(pp));
7027 
7028 	tmphme = NULL;
7029 	pa_hments = 0;
7030 	CPUSET_ZERO(cpuset);
7031 
7032 	pml = sfmmu_mlist_enter(pp);
7033 
7034 #ifdef VAC
7035 	if (pp->p_kpmref)
7036 		sfmmu_kpm_pageunload(pp);
7037 	ASSERT(!PP_ISMAPPED_KPM(pp));
7038 #endif
7039 	/*
7040 	 * Clear vpm reference. Since the page is exclusively locked
7041 	 * vpm cannot be referencing it.
7042 	 */
7043 	if (vpm_enable) {
7044 		pp->p_vpmref = 0;
7045 	}
7046 
7047 	index = PP_MAPINDEX(pp);
7048 	cons = TTE8K;
7049 retry:
7050 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7051 		tmphme = sfhme->hme_next;
7052 
7053 		if (IS_PAHME(sfhme)) {
7054 			ASSERT(sfhme->hme_data != NULL);
7055 			pa_hments++;
7056 			continue;
7057 		}
7058 
7059 		hmeblkp = sfmmu_hmetohblk(sfhme);
7060 
7061 		/*
7062 		 * If there are kernel mappings don't unload them, they will
7063 		 * be suspended.
7064 		 */
7065 		if (forceflag == SFMMU_KERNEL_RELOC && hmeblkp->hblk_lckcnt &&
7066 		    hmeblkp->hblk_tag.htag_id == ksfmmup)
7067 			continue;
7068 
7069 		tset = sfmmu_pageunload(pp, sfhme, cons);
7070 		CPUSET_OR(cpuset, tset);
7071 	}
7072 
7073 	while (index != 0) {
7074 		index = index >> 1;
7075 		if (index != 0)
7076 			cons++;
7077 		if (index & 0x1) {
7078 			/* Go to leading page */
7079 			pp = PP_GROUPLEADER(pp, cons);
7080 			ASSERT(sfmmu_mlist_held(pp));
7081 			goto retry;
7082 		}
7083 	}
7084 
7085 	/*
7086 	 * cpuset may be empty if the page was only mapped by segkpm,
7087 	 * in which case we won't actually cross-trap.
7088 	 */
7089 	xt_sync(cpuset);
7090 
7091 	/*
7092 	 * The page should have no mappings at this point, unless
7093 	 * we were called from hat_page_relocate() in which case we
7094 	 * leave the locked mappings which will be suspended later.
7095 	 */
7096 	ASSERT(!PP_ISMAPPED(origpp) || pa_hments ||
7097 	    (forceflag == SFMMU_KERNEL_RELOC));
7098 
7099 #ifdef VAC
7100 	if (PP_ISTNC(pp)) {
7101 		if (cons == TTE8K) {
7102 			pmtx = sfmmu_page_enter(pp);
7103 			PP_CLRTNC(pp);
7104 			sfmmu_page_exit(pmtx);
7105 		} else {
7106 			conv_tnc(pp, cons);
7107 		}
7108 	}
7109 #endif	/* VAC */
7110 
7111 	if (pa_hments && forceflag != SFMMU_KERNEL_RELOC) {
7112 		/*
7113 		 * Unlink any pa_hments and free them, calling back
7114 		 * the responsible subsystem to notify it of the error.
7115 		 * This can occur in situations such as drivers leaking
7116 		 * DMA handles: naughty, but common enough that we'd like
7117 		 * to keep the system running rather than bringing it
7118 		 * down with an obscure error like "pa_hment leaked"
7119 		 * which doesn't aid the user in debugging their driver.
7120 		 */
7121 		for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7122 			tmphme = sfhme->hme_next;
7123 			if (IS_PAHME(sfhme)) {
7124 				struct pa_hment *pahmep = sfhme->hme_data;
7125 				sfmmu_pahment_leaked(pahmep);
7126 				HME_SUB(sfhme, pp);
7127 				kmem_cache_free(pa_hment_cache, pahmep);
7128 			}
7129 		}
7130 
7131 		ASSERT(!PP_ISMAPPED(origpp));
7132 	}
7133 
7134 	sfmmu_mlist_exit(pml);
7135 
7136 	return (0);
7137 }
7138 
7139 cpuset_t
7140 sfmmu_pageunload(page_t *pp, struct sf_hment *sfhme, int cons)
7141 {
7142 	struct hme_blk *hmeblkp;
7143 	sfmmu_t *sfmmup;
7144 	tte_t tte, ttemod;
7145 #ifdef DEBUG
7146 	tte_t orig_old;
7147 #endif /* DEBUG */
7148 	caddr_t addr;
7149 	int ttesz;
7150 	int ret;
7151 	cpuset_t cpuset;
7152 
7153 	ASSERT(pp != NULL);
7154 	ASSERT(sfmmu_mlist_held(pp));
7155 	ASSERT(!PP_ISKAS(pp));
7156 
7157 	CPUSET_ZERO(cpuset);
7158 
7159 	hmeblkp = sfmmu_hmetohblk(sfhme);
7160 
7161 readtte:
7162 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7163 	if (TTE_IS_VALID(&tte)) {
7164 		sfmmup = hblktosfmmu(hmeblkp);
7165 		ttesz = get_hblk_ttesz(hmeblkp);
7166 		/*
7167 		 * Only unload mappings of 'cons' size.
7168 		 */
7169 		if (ttesz != cons)
7170 			return (cpuset);
7171 
7172 		/*
7173 		 * Note that we have p_mapping lock, but no hash lock here.
7174 		 * hblk_unload() has to have both hash lock AND p_mapping
7175 		 * lock before it tries to modify tte. So, the tte could
7176 		 * not become invalid in the sfmmu_modifytte_try() below.
7177 		 */
7178 		ttemod = tte;
7179 #ifdef DEBUG
7180 		orig_old = tte;
7181 #endif /* DEBUG */
7182 
7183 		TTE_SET_INVALID(&ttemod);
7184 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7185 		if (ret < 0) {
7186 #ifdef DEBUG
7187 			/* only R/M bits can change. */
7188 			chk_tte(&orig_old, &tte, &ttemod, hmeblkp);
7189 #endif /* DEBUG */
7190 			goto readtte;
7191 		}
7192 
7193 		if (ret == 0) {
7194 			panic("pageunload: cas failed?");
7195 		}
7196 
7197 		addr = tte_to_vaddr(hmeblkp, tte);
7198 
7199 		if (hmeblkp->hblk_shared) {
7200 			sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7201 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
7202 			sf_region_t *rgnp;
7203 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7204 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7205 			ASSERT(srdp != NULL);
7206 			rgnp = srdp->srd_hmergnp[rid];
7207 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
7208 			cpuset = sfmmu_rgntlb_demap(addr, rgnp, hmeblkp, 1);
7209 			sfmmu_ttesync(NULL, addr, &tte, pp);
7210 			ASSERT(rgnp->rgn_ttecnt[ttesz] > 0);
7211 			atomic_dec_ulong(&rgnp->rgn_ttecnt[ttesz]);
7212 		} else {
7213 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
7214 			atomic_dec_ulong(&sfmmup->sfmmu_ttecnt[ttesz]);
7215 
7216 			/*
7217 			 * We need to flush the page from the virtual cache
7218 			 * in order to prevent a virtual cache alias
7219 			 * inconsistency. The particular scenario we need
7220 			 * to worry about is:
7221 			 * Given:  va1 and va2 are two virtual address that
7222 			 * alias and will map the same physical address.
7223 			 * 1.   mapping exists from va1 to pa and data has
7224 			 *	been read into the cache.
7225 			 * 2.   unload va1.
7226 			 * 3.   load va2 and modify data using va2.
7227 			 * 4    unload va2.
7228 			 * 5.   load va1 and reference data.  Unless we flush
7229 			 *	the data cache when we unload we will get
7230 			 *	stale data.
7231 			 * This scenario is taken care of by using virtual
7232 			 * page coloring.
7233 			 */
7234 			if (sfmmup->sfmmu_ismhat) {
7235 				/*
7236 				 * Flush TSBs, TLBs and caches
7237 				 * of every process
7238 				 * sharing this ism segment.
7239 				 */
7240 				sfmmu_hat_lock_all();
7241 				mutex_enter(&ism_mlist_lock);
7242 				kpreempt_disable();
7243 				sfmmu_ismtlbcache_demap(addr, sfmmup, hmeblkp,
7244 				    pp->p_pagenum, CACHE_NO_FLUSH);
7245 				kpreempt_enable();
7246 				mutex_exit(&ism_mlist_lock);
7247 				sfmmu_hat_unlock_all();
7248 				cpuset = cpu_ready_set;
7249 			} else {
7250 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7251 				cpuset = sfmmup->sfmmu_cpusran;
7252 			}
7253 		}
7254 
7255 		/*
7256 		 * Hme_sub has to run after ttesync() and a_rss update.
7257 		 * See hblk_unload().
7258 		 */
7259 		HME_SUB(sfhme, pp);
7260 		membar_stst();
7261 
7262 		/*
7263 		 * We can not make ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
7264 		 * since pteload may have done a HME_ADD() right after
7265 		 * we did the HME_SUB() above. Hmecnt is now maintained
7266 		 * by cas only. no lock guranteed its value. The only
7267 		 * gurantee we have is the hmecnt should not be less than
7268 		 * what it should be so the hblk will not be taken away.
7269 		 * It's also important that we decremented the hmecnt after
7270 		 * we are done with hmeblkp so that this hmeblk won't be
7271 		 * stolen.
7272 		 */
7273 		ASSERT(hmeblkp->hblk_hmecnt > 0);
7274 		ASSERT(hmeblkp->hblk_vcnt > 0);
7275 		atomic_dec_16(&hmeblkp->hblk_vcnt);
7276 		atomic_dec_16(&hmeblkp->hblk_hmecnt);
7277 		/*
7278 		 * This is bug 4063182.
7279 		 * XXX: fixme
7280 		 * ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
7281 		 *	!hmeblkp->hblk_lckcnt);
7282 		 */
7283 	} else {
7284 		panic("invalid tte? pp %p &tte %p",
7285 		    (void *)pp, (void *)&tte);
7286 	}
7287 
7288 	return (cpuset);
7289 }
7290 
7291 /*
7292  * While relocating a kernel page, this function will move the mappings
7293  * from tpp to dpp and modify any associated data with these mappings.
7294  * It also unsuspends the suspended kernel mapping.
7295  */
7296 static void
7297 hat_pagereload(struct page *tpp, struct page *dpp)
7298 {
7299 	struct sf_hment *sfhme;
7300 	tte_t tte, ttemod;
7301 	int index, cons;
7302 
7303 	ASSERT(getpil() == PIL_MAX);
7304 	ASSERT(sfmmu_mlist_held(tpp));
7305 	ASSERT(sfmmu_mlist_held(dpp));
7306 
7307 	index = PP_MAPINDEX(tpp);
7308 	cons = TTE8K;
7309 
7310 	/* Update real mappings to the page */
7311 retry:
7312 	for (sfhme = tpp->p_mapping; sfhme != NULL; sfhme = sfhme->hme_next) {
7313 		if (IS_PAHME(sfhme))
7314 			continue;
7315 		sfmmu_copytte(&sfhme->hme_tte, &tte);
7316 		ttemod = tte;
7317 
7318 		/*
7319 		 * replace old pfn with new pfn in TTE
7320 		 */
7321 		PFN_TO_TTE(ttemod, dpp->p_pagenum);
7322 
7323 		/*
7324 		 * clear suspend bit
7325 		 */
7326 		ASSERT(TTE_IS_SUSPEND(&ttemod));
7327 		TTE_CLR_SUSPEND(&ttemod);
7328 
7329 		if (sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte) < 0)
7330 			panic("hat_pagereload(): sfmmu_modifytte_try() failed");
7331 
7332 		/*
7333 		 * set hme_page point to new page
7334 		 */
7335 		sfhme->hme_page = dpp;
7336 	}
7337 
7338 	/*
7339 	 * move p_mapping list from old page to new page
7340 	 */
7341 	dpp->p_mapping = tpp->p_mapping;
7342 	tpp->p_mapping = NULL;
7343 	dpp->p_share = tpp->p_share;
7344 	tpp->p_share = 0;
7345 
7346 	while (index != 0) {
7347 		index = index >> 1;
7348 		if (index != 0)
7349 			cons++;
7350 		if (index & 0x1) {
7351 			tpp = PP_GROUPLEADER(tpp, cons);
7352 			dpp = PP_GROUPLEADER(dpp, cons);
7353 			goto retry;
7354 		}
7355 	}
7356 
7357 	curthread->t_flag &= ~T_DONTDTRACE;
7358 	mutex_exit(&kpr_suspendlock);
7359 }
7360 
7361 uint_t
7362 hat_pagesync(struct page *pp, uint_t clearflag)
7363 {
7364 	struct sf_hment *sfhme, *tmphme = NULL;
7365 	struct hme_blk *hmeblkp;
7366 	kmutex_t *pml;
7367 	cpuset_t cpuset, tset;
7368 	int	index, cons;
7369 	extern	ulong_t po_share;
7370 	page_t	*save_pp = pp;
7371 	int	stop_on_sh = 0;
7372 	uint_t	shcnt;
7373 
7374 	CPUSET_ZERO(cpuset);
7375 
7376 	if (PP_ISRO(pp) && (clearflag & HAT_SYNC_STOPON_MOD)) {
7377 		return (PP_GENERIC_ATTR(pp));
7378 	}
7379 
7380 	if ((clearflag & HAT_SYNC_ZERORM) == 0) {
7381 		if ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(pp)) {
7382 			return (PP_GENERIC_ATTR(pp));
7383 		}
7384 		if ((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(pp)) {
7385 			return (PP_GENERIC_ATTR(pp));
7386 		}
7387 		if (clearflag & HAT_SYNC_STOPON_SHARED) {
7388 			if (pp->p_share > po_share) {
7389 				hat_page_setattr(pp, P_REF);
7390 				return (PP_GENERIC_ATTR(pp));
7391 			}
7392 			stop_on_sh = 1;
7393 			shcnt = 0;
7394 		}
7395 	}
7396 
7397 	clearflag &= ~HAT_SYNC_STOPON_SHARED;
7398 	pml = sfmmu_mlist_enter(pp);
7399 	index = PP_MAPINDEX(pp);
7400 	cons = TTE8K;
7401 retry:
7402 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7403 		/*
7404 		 * We need to save the next hment on the list since
7405 		 * it is possible for pagesync to remove an invalid hment
7406 		 * from the list.
7407 		 */
7408 		tmphme = sfhme->hme_next;
7409 		if (IS_PAHME(sfhme))
7410 			continue;
7411 		/*
7412 		 * If we are looking for large mappings and this hme doesn't
7413 		 * reach the range we are seeking, just ignore it.
7414 		 */
7415 		hmeblkp = sfmmu_hmetohblk(sfhme);
7416 
7417 		if (hme_size(sfhme) < cons)
7418 			continue;
7419 
7420 		if (stop_on_sh) {
7421 			if (hmeblkp->hblk_shared) {
7422 				sf_srd_t *srdp = hblktosrd(hmeblkp);
7423 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7424 				sf_region_t *rgnp;
7425 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7426 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7427 				ASSERT(srdp != NULL);
7428 				rgnp = srdp->srd_hmergnp[rid];
7429 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
7430 				    rgnp, rid);
7431 				shcnt += rgnp->rgn_refcnt;
7432 			} else {
7433 				shcnt++;
7434 			}
7435 			if (shcnt > po_share) {
7436 				/*
7437 				 * tell the pager to spare the page this time
7438 				 * around.
7439 				 */
7440 				hat_page_setattr(save_pp, P_REF);
7441 				index = 0;
7442 				break;
7443 			}
7444 		}
7445 		tset = sfmmu_pagesync(pp, sfhme,
7446 		    clearflag & ~HAT_SYNC_STOPON_RM);
7447 		CPUSET_OR(cpuset, tset);
7448 
7449 		/*
7450 		 * If clearflag is HAT_SYNC_DONTZERO, break out as soon
7451 		 * as the "ref" or "mod" is set or share cnt exceeds po_share.
7452 		 */
7453 		if ((clearflag & ~HAT_SYNC_STOPON_RM) == HAT_SYNC_DONTZERO &&
7454 		    (((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(save_pp)) ||
7455 		    ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(save_pp)))) {
7456 			index = 0;
7457 			break;
7458 		}
7459 	}
7460 
7461 	while (index) {
7462 		index = index >> 1;
7463 		cons++;
7464 		if (index & 0x1) {
7465 			/* Go to leading page */
7466 			pp = PP_GROUPLEADER(pp, cons);
7467 			goto retry;
7468 		}
7469 	}
7470 
7471 	xt_sync(cpuset);
7472 	sfmmu_mlist_exit(pml);
7473 	return (PP_GENERIC_ATTR(save_pp));
7474 }
7475 
7476 /*
7477  * Get all the hardware dependent attributes for a page struct
7478  */
7479 static cpuset_t
7480 sfmmu_pagesync(struct page *pp, struct sf_hment *sfhme,
7481     uint_t clearflag)
7482 {
7483 	caddr_t addr;
7484 	tte_t tte, ttemod;
7485 	struct hme_blk *hmeblkp;
7486 	int ret;
7487 	sfmmu_t *sfmmup;
7488 	cpuset_t cpuset;
7489 
7490 	ASSERT(pp != NULL);
7491 	ASSERT(sfmmu_mlist_held(pp));
7492 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
7493 	    (clearflag == HAT_SYNC_ZERORM));
7494 
7495 	SFMMU_STAT(sf_pagesync);
7496 
7497 	CPUSET_ZERO(cpuset);
7498 
7499 sfmmu_pagesync_retry:
7500 
7501 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7502 	if (TTE_IS_VALID(&tte)) {
7503 		hmeblkp = sfmmu_hmetohblk(sfhme);
7504 		sfmmup = hblktosfmmu(hmeblkp);
7505 		addr = tte_to_vaddr(hmeblkp, tte);
7506 		if (clearflag == HAT_SYNC_ZERORM) {
7507 			ttemod = tte;
7508 			TTE_CLR_RM(&ttemod);
7509 			ret = sfmmu_modifytte_try(&tte, &ttemod,
7510 			    &sfhme->hme_tte);
7511 			if (ret < 0) {
7512 				/*
7513 				 * cas failed and the new value is not what
7514 				 * we want.
7515 				 */
7516 				goto sfmmu_pagesync_retry;
7517 			}
7518 
7519 			if (ret > 0) {
7520 				/* we win the cas */
7521 				if (hmeblkp->hblk_shared) {
7522 					sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7523 					uint_t rid =
7524 					    hmeblkp->hblk_tag.htag_rid;
7525 					sf_region_t *rgnp;
7526 					ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7527 					ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7528 					ASSERT(srdp != NULL);
7529 					rgnp = srdp->srd_hmergnp[rid];
7530 					SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7531 					    srdp, rgnp, rid);
7532 					cpuset = sfmmu_rgntlb_demap(addr,
7533 					    rgnp, hmeblkp, 1);
7534 				} else {
7535 					sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
7536 					    0, 0);
7537 					cpuset = sfmmup->sfmmu_cpusran;
7538 				}
7539 			}
7540 		}
7541 		sfmmu_ttesync(hmeblkp->hblk_shared ? NULL : sfmmup, addr,
7542 		    &tte, pp);
7543 	}
7544 	return (cpuset);
7545 }
7546 
7547 /*
7548  * Remove write permission from a mappings to a page, so that
7549  * we can detect the next modification of it. This requires modifying
7550  * the TTE then invalidating (demap) any TLB entry using that TTE.
7551  * This code is similar to sfmmu_pagesync().
7552  */
7553 static cpuset_t
7554 sfmmu_pageclrwrt(struct page *pp, struct sf_hment *sfhme)
7555 {
7556 	caddr_t addr;
7557 	tte_t tte;
7558 	tte_t ttemod;
7559 	struct hme_blk *hmeblkp;
7560 	int ret;
7561 	sfmmu_t *sfmmup;
7562 	cpuset_t cpuset;
7563 
7564 	ASSERT(pp != NULL);
7565 	ASSERT(sfmmu_mlist_held(pp));
7566 
7567 	CPUSET_ZERO(cpuset);
7568 	SFMMU_STAT(sf_clrwrt);
7569 
7570 retry:
7571 
7572 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7573 	if (TTE_IS_VALID(&tte) && TTE_IS_WRITABLE(&tte)) {
7574 		hmeblkp = sfmmu_hmetohblk(sfhme);
7575 		sfmmup = hblktosfmmu(hmeblkp);
7576 		addr = tte_to_vaddr(hmeblkp, tte);
7577 
7578 		ttemod = tte;
7579 		TTE_CLR_WRT(&ttemod);
7580 		TTE_CLR_MOD(&ttemod);
7581 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7582 
7583 		/*
7584 		 * if cas failed and the new value is not what
7585 		 * we want retry
7586 		 */
7587 		if (ret < 0)
7588 			goto retry;
7589 
7590 		/* we win the cas */
7591 		if (ret > 0) {
7592 			if (hmeblkp->hblk_shared) {
7593 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7594 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7595 				sf_region_t *rgnp;
7596 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7597 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7598 				ASSERT(srdp != NULL);
7599 				rgnp = srdp->srd_hmergnp[rid];
7600 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7601 				    srdp, rgnp, rid);
7602 				cpuset = sfmmu_rgntlb_demap(addr,
7603 				    rgnp, hmeblkp, 1);
7604 			} else {
7605 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7606 				cpuset = sfmmup->sfmmu_cpusran;
7607 			}
7608 		}
7609 	}
7610 
7611 	return (cpuset);
7612 }
7613 
7614 /*
7615  * Walk all mappings of a page, removing write permission and clearing the
7616  * ref/mod bits. This code is similar to hat_pagesync()
7617  */
7618 static void
7619 hat_page_clrwrt(page_t *pp)
7620 {
7621 	struct sf_hment *sfhme;
7622 	struct sf_hment *tmphme = NULL;
7623 	kmutex_t *pml;
7624 	cpuset_t cpuset;
7625 	cpuset_t tset;
7626 	int	index;
7627 	int	 cons;
7628 
7629 	CPUSET_ZERO(cpuset);
7630 
7631 	pml = sfmmu_mlist_enter(pp);
7632 	index = PP_MAPINDEX(pp);
7633 	cons = TTE8K;
7634 retry:
7635 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7636 		tmphme = sfhme->hme_next;
7637 
7638 		/*
7639 		 * If we are looking for large mappings and this hme doesn't
7640 		 * reach the range we are seeking, just ignore its.
7641 		 */
7642 
7643 		if (hme_size(sfhme) < cons)
7644 			continue;
7645 
7646 		tset = sfmmu_pageclrwrt(pp, sfhme);
7647 		CPUSET_OR(cpuset, tset);
7648 	}
7649 
7650 	while (index) {
7651 		index = index >> 1;
7652 		cons++;
7653 		if (index & 0x1) {
7654 			/* Go to leading page */
7655 			pp = PP_GROUPLEADER(pp, cons);
7656 			goto retry;
7657 		}
7658 	}
7659 
7660 	xt_sync(cpuset);
7661 	sfmmu_mlist_exit(pml);
7662 }
7663 
7664 /*
7665  * Set the given REF/MOD/RO bits for the given page.
7666  * For a vnode with a sorted v_pages list, we need to change
7667  * the attributes and the v_pages list together under page_vnode_mutex.
7668  */
7669 void
7670 hat_page_setattr(page_t *pp, uint_t flag)
7671 {
7672 	vnode_t		*vp = pp->p_vnode;
7673 	page_t		**listp;
7674 	kmutex_t	*pmtx;
7675 	kmutex_t	*vphm = NULL;
7676 	int		noshuffle;
7677 
7678 	noshuffle = flag & P_NSH;
7679 	flag &= ~P_NSH;
7680 
7681 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7682 
7683 	/*
7684 	 * nothing to do if attribute already set
7685 	 */
7686 	if ((pp->p_nrm & flag) == flag)
7687 		return;
7688 
7689 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp) &&
7690 	    !noshuffle) {
7691 		vphm = page_vnode_mutex(vp);
7692 		mutex_enter(vphm);
7693 	}
7694 
7695 	pmtx = sfmmu_page_enter(pp);
7696 	pp->p_nrm |= flag;
7697 	sfmmu_page_exit(pmtx);
7698 
7699 	if (vphm != NULL) {
7700 		/*
7701 		 * Some File Systems examine v_pages for NULL w/o
7702 		 * grabbing the vphm mutex. Must not let it become NULL when
7703 		 * pp is the only page on the list.
7704 		 */
7705 		if (pp->p_vpnext != pp) {
7706 			page_vpsub(&vp->v_pages, pp);
7707 			if (vp->v_pages != NULL)
7708 				listp = &vp->v_pages->p_vpprev->p_vpnext;
7709 			else
7710 				listp = &vp->v_pages;
7711 			page_vpadd(listp, pp);
7712 		}
7713 		mutex_exit(vphm);
7714 	}
7715 }
7716 
7717 void
7718 hat_page_clrattr(page_t *pp, uint_t flag)
7719 {
7720 	vnode_t		*vp = pp->p_vnode;
7721 	kmutex_t	*pmtx;
7722 
7723 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7724 
7725 	pmtx = sfmmu_page_enter(pp);
7726 
7727 	/*
7728 	 * Caller is expected to hold page's io lock for VMODSORT to work
7729 	 * correctly with pvn_vplist_dirty() and pvn_getdirty() when mod
7730 	 * bit is cleared.
7731 	 * We don't have assert to avoid tripping some existing third party
7732 	 * code. The dirty page is moved back to top of the v_page list
7733 	 * after IO is done in pvn_write_done().
7734 	 */
7735 	pp->p_nrm &= ~flag;
7736 	sfmmu_page_exit(pmtx);
7737 
7738 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) {
7739 
7740 		/*
7741 		 * VMODSORT works by removing write permissions and getting
7742 		 * a fault when a page is made dirty. At this point
7743 		 * we need to remove write permission from all mappings
7744 		 * to this page.
7745 		 */
7746 		hat_page_clrwrt(pp);
7747 	}
7748 }
7749 
7750 uint_t
7751 hat_page_getattr(page_t *pp, uint_t flag)
7752 {
7753 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7754 	return ((uint_t)(pp->p_nrm & flag));
7755 }
7756 
7757 /*
7758  * DEBUG kernels: verify that a kernel va<->pa translation
7759  * is safe by checking the underlying page_t is in a page
7760  * relocation-safe state.
7761  */
7762 #ifdef	DEBUG
7763 void
7764 sfmmu_check_kpfn(pfn_t pfn)
7765 {
7766 	page_t *pp;
7767 	int index, cons;
7768 
7769 	if (hat_check_vtop == 0)
7770 		return;
7771 
7772 	if (kvseg.s_base == NULL || panicstr)
7773 		return;
7774 
7775 	pp = page_numtopp_nolock(pfn);
7776 	if (!pp)
7777 		return;
7778 
7779 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7780 		return;
7781 
7782 	/*
7783 	 * Handed a large kernel page, we dig up the root page since we
7784 	 * know the root page might have the lock also.
7785 	 */
7786 	if (pp->p_szc != 0) {
7787 		index = PP_MAPINDEX(pp);
7788 		cons = TTE8K;
7789 again:
7790 		while (index != 0) {
7791 			index >>= 1;
7792 			if (index != 0)
7793 				cons++;
7794 			if (index & 0x1) {
7795 				pp = PP_GROUPLEADER(pp, cons);
7796 				goto again;
7797 			}
7798 		}
7799 	}
7800 
7801 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7802 		return;
7803 
7804 	/*
7805 	 * Pages need to be locked or allocated "permanent" (either from
7806 	 * static_arena arena or explicitly setting PG_NORELOC when calling
7807 	 * page_create_va()) for VA->PA translations to be valid.
7808 	 */
7809 	if (!PP_ISNORELOC(pp))
7810 		panic("Illegal VA->PA translation, pp 0x%p not permanent",
7811 		    (void *)pp);
7812 	else
7813 		panic("Illegal VA->PA translation, pp 0x%p not locked",
7814 		    (void *)pp);
7815 }
7816 #endif	/* DEBUG */
7817 
7818 /*
7819  * Returns a page frame number for a given virtual address.
7820  * Returns PFN_INVALID to indicate an invalid mapping
7821  */
7822 pfn_t
7823 hat_getpfnum(struct hat *hat, caddr_t addr)
7824 {
7825 	pfn_t pfn;
7826 	tte_t tte;
7827 
7828 	/*
7829 	 * We would like to
7830 	 * ASSERT(AS_LOCK_HELD(as));
7831 	 * but we can't because the iommu driver will call this
7832 	 * routine at interrupt time and it can't grab the as lock
7833 	 * or it will deadlock: A thread could have the as lock
7834 	 * and be waiting for io.  The io can't complete
7835 	 * because the interrupt thread is blocked trying to grab
7836 	 * the as lock.
7837 	 */
7838 
7839 	if (hat == ksfmmup) {
7840 		if (IS_KMEM_VA_LARGEPAGE(addr)) {
7841 			ASSERT(segkmem_lpszc > 0);
7842 			pfn = sfmmu_kvaszc2pfn(addr, segkmem_lpszc);
7843 			if (pfn != PFN_INVALID) {
7844 				sfmmu_check_kpfn(pfn);
7845 				return (pfn);
7846 			}
7847 		} else if (segkpm && IS_KPM_ADDR(addr)) {
7848 			return (sfmmu_kpm_vatopfn(addr));
7849 		}
7850 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
7851 		    == PFN_SUSPENDED) {
7852 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
7853 		}
7854 		sfmmu_check_kpfn(pfn);
7855 		return (pfn);
7856 	} else {
7857 		return (sfmmu_uvatopfn(addr, hat, NULL));
7858 	}
7859 }
7860 
7861 /*
7862  * This routine will return both pfn and tte for the vaddr.
7863  */
7864 static pfn_t
7865 sfmmu_uvatopfn(caddr_t vaddr, struct hat *sfmmup, tte_t *ttep)
7866 {
7867 	struct hmehash_bucket *hmebp;
7868 	hmeblk_tag hblktag;
7869 	int hmeshift, hashno = 1;
7870 	struct hme_blk *hmeblkp = NULL;
7871 	tte_t tte;
7872 
7873 	struct sf_hment *sfhmep;
7874 	pfn_t pfn;
7875 
7876 	/* support for ISM */
7877 	ism_map_t	*ism_map;
7878 	ism_blk_t	*ism_blkp;
7879 	int		i;
7880 	sfmmu_t *ism_hatid = NULL;
7881 	sfmmu_t *locked_hatid = NULL;
7882 	sfmmu_t	*sv_sfmmup = sfmmup;
7883 	caddr_t	sv_vaddr = vaddr;
7884 	sf_srd_t *srdp;
7885 
7886 	if (ttep == NULL) {
7887 		ttep = &tte;
7888 	} else {
7889 		ttep->ll = 0;
7890 	}
7891 
7892 	ASSERT(sfmmup != ksfmmup);
7893 	SFMMU_STAT(sf_user_vtop);
7894 	/*
7895 	 * Set ism_hatid if vaddr falls in a ISM segment.
7896 	 */
7897 	ism_blkp = sfmmup->sfmmu_iblk;
7898 	if (ism_blkp != NULL) {
7899 		sfmmu_ismhat_enter(sfmmup, 0);
7900 		locked_hatid = sfmmup;
7901 	}
7902 	while (ism_blkp != NULL && ism_hatid == NULL) {
7903 		ism_map = ism_blkp->iblk_maps;
7904 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
7905 			if (vaddr >= ism_start(ism_map[i]) &&
7906 			    vaddr < ism_end(ism_map[i])) {
7907 				sfmmup = ism_hatid = ism_map[i].imap_ismhat;
7908 				vaddr = (caddr_t)(vaddr -
7909 				    ism_start(ism_map[i]));
7910 				break;
7911 			}
7912 		}
7913 		ism_blkp = ism_blkp->iblk_next;
7914 	}
7915 	if (locked_hatid) {
7916 		sfmmu_ismhat_exit(locked_hatid, 0);
7917 	}
7918 
7919 	hblktag.htag_id = sfmmup;
7920 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
7921 	do {
7922 		hmeshift = HME_HASH_SHIFT(hashno);
7923 		hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
7924 		hblktag.htag_rehash = hashno;
7925 		hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
7926 
7927 		SFMMU_HASH_LOCK(hmebp);
7928 
7929 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
7930 		if (hmeblkp != NULL) {
7931 			ASSERT(!hmeblkp->hblk_shared);
7932 			HBLKTOHME(sfhmep, hmeblkp, vaddr);
7933 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
7934 			SFMMU_HASH_UNLOCK(hmebp);
7935 			if (TTE_IS_VALID(ttep)) {
7936 				pfn = TTE_TO_PFN(vaddr, ttep);
7937 				return (pfn);
7938 			}
7939 			break;
7940 		}
7941 		SFMMU_HASH_UNLOCK(hmebp);
7942 		hashno++;
7943 	} while (HME_REHASH(sfmmup) && (hashno <= mmu_hashcnt));
7944 
7945 	if (SF_HMERGNMAP_ISNULL(sv_sfmmup)) {
7946 		return (PFN_INVALID);
7947 	}
7948 	srdp = sv_sfmmup->sfmmu_srdp;
7949 	ASSERT(srdp != NULL);
7950 	ASSERT(srdp->srd_refcnt != 0);
7951 	hblktag.htag_id = srdp;
7952 	hashno = 1;
7953 	do {
7954 		hmeshift = HME_HASH_SHIFT(hashno);
7955 		hblktag.htag_bspage = HME_HASH_BSPAGE(sv_vaddr, hmeshift);
7956 		hblktag.htag_rehash = hashno;
7957 		hmebp = HME_HASH_FUNCTION(srdp, sv_vaddr, hmeshift);
7958 
7959 		SFMMU_HASH_LOCK(hmebp);
7960 		for (hmeblkp = hmebp->hmeblkp; hmeblkp != NULL;
7961 		    hmeblkp = hmeblkp->hblk_next) {
7962 			uint_t rid;
7963 			sf_region_t *rgnp;
7964 			caddr_t rsaddr;
7965 			caddr_t readdr;
7966 
7967 			if (!HTAGS_EQ_SHME(hmeblkp->hblk_tag, hblktag,
7968 			    sv_sfmmup->sfmmu_hmeregion_map)) {
7969 				continue;
7970 			}
7971 			ASSERT(hmeblkp->hblk_shared);
7972 			rid = hmeblkp->hblk_tag.htag_rid;
7973 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7974 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7975 			rgnp = srdp->srd_hmergnp[rid];
7976 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
7977 			HBLKTOHME(sfhmep, hmeblkp, sv_vaddr);
7978 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
7979 			rsaddr = rgnp->rgn_saddr;
7980 			readdr = rsaddr + rgnp->rgn_size;
7981 #ifdef DEBUG
7982 			if (TTE_IS_VALID(ttep) ||
7983 			    get_hblk_ttesz(hmeblkp) > TTE8K) {
7984 				caddr_t eva = tte_to_evaddr(hmeblkp, ttep);
7985 				ASSERT(eva > sv_vaddr);
7986 				ASSERT(sv_vaddr >= rsaddr);
7987 				ASSERT(sv_vaddr < readdr);
7988 				ASSERT(eva <= readdr);
7989 			}
7990 #endif /* DEBUG */
7991 			/*
7992 			 * Continue the search if we
7993 			 * found an invalid 8K tte outside of the area
7994 			 * covered by this hmeblk's region.
7995 			 */
7996 			if (TTE_IS_VALID(ttep)) {
7997 				SFMMU_HASH_UNLOCK(hmebp);
7998 				pfn = TTE_TO_PFN(sv_vaddr, ttep);
7999 				return (pfn);
8000 			} else if (get_hblk_ttesz(hmeblkp) > TTE8K ||
8001 			    (sv_vaddr >= rsaddr && sv_vaddr < readdr)) {
8002 				SFMMU_HASH_UNLOCK(hmebp);
8003 				pfn = PFN_INVALID;
8004 				return (pfn);
8005 			}
8006 		}
8007 		SFMMU_HASH_UNLOCK(hmebp);
8008 		hashno++;
8009 	} while (hashno <= mmu_hashcnt);
8010 	return (PFN_INVALID);
8011 }
8012 
8013 
8014 /*
8015  * For compatability with AT&T and later optimizations
8016  */
8017 /* ARGSUSED */
8018 void
8019 hat_map(struct hat *hat, caddr_t addr, size_t len, uint_t flags)
8020 {
8021 	ASSERT(hat != NULL);
8022 }
8023 
8024 /*
8025  * Return the number of mappings to a particular page.  This number is an
8026  * approximation of the number of people sharing the page.
8027  *
8028  * shared hmeblks or ism hmeblks are counted as 1 mapping here.
8029  * hat_page_checkshare() can be used to compare threshold to share
8030  * count that reflects the number of region sharers albeit at higher cost.
8031  */
8032 ulong_t
8033 hat_page_getshare(page_t *pp)
8034 {
8035 	page_t *spp = pp;	/* start page */
8036 	kmutex_t *pml;
8037 	ulong_t	cnt;
8038 	int index, sz = TTE64K;
8039 
8040 	/*
8041 	 * We need to grab the mlist lock to make sure any outstanding
8042 	 * load/unloads complete.  Otherwise we could return zero
8043 	 * even though the unload(s) hasn't finished yet.
8044 	 */
8045 	pml = sfmmu_mlist_enter(spp);
8046 	cnt = spp->p_share;
8047 
8048 #ifdef VAC
8049 	if (kpm_enable)
8050 		cnt += spp->p_kpmref;
8051 #endif
8052 	if (vpm_enable && pp->p_vpmref) {
8053 		cnt += 1;
8054 	}
8055 
8056 	/*
8057 	 * If we have any large mappings, we count the number of
8058 	 * mappings that this large page is part of.
8059 	 */
8060 	index = PP_MAPINDEX(spp);
8061 	index >>= 1;
8062 	while (index) {
8063 		pp = PP_GROUPLEADER(spp, sz);
8064 		if ((index & 0x1) && pp != spp) {
8065 			cnt += pp->p_share;
8066 			spp = pp;
8067 		}
8068 		index >>= 1;
8069 		sz++;
8070 	}
8071 	sfmmu_mlist_exit(pml);
8072 	return (cnt);
8073 }
8074 
8075 /*
8076  * Return 1 if the number of mappings exceeds sh_thresh. Return 0
8077  * otherwise. Count shared hmeblks by region's refcnt.
8078  */
8079 int
8080 hat_page_checkshare(page_t *pp, ulong_t sh_thresh)
8081 {
8082 	kmutex_t *pml;
8083 	ulong_t	cnt = 0;
8084 	int index, sz = TTE8K;
8085 	struct sf_hment *sfhme, *tmphme = NULL;
8086 	struct hme_blk *hmeblkp;
8087 
8088 	pml = sfmmu_mlist_enter(pp);
8089 
8090 #ifdef VAC
8091 	if (kpm_enable)
8092 		cnt = pp->p_kpmref;
8093 #endif
8094 
8095 	if (vpm_enable && pp->p_vpmref) {
8096 		cnt += 1;
8097 	}
8098 
8099 	if (pp->p_share + cnt > sh_thresh) {
8100 		sfmmu_mlist_exit(pml);
8101 		return (1);
8102 	}
8103 
8104 	index = PP_MAPINDEX(pp);
8105 
8106 again:
8107 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
8108 		tmphme = sfhme->hme_next;
8109 		if (IS_PAHME(sfhme)) {
8110 			continue;
8111 		}
8112 
8113 		hmeblkp = sfmmu_hmetohblk(sfhme);
8114 		if (hme_size(sfhme) != sz) {
8115 			continue;
8116 		}
8117 
8118 		if (hmeblkp->hblk_shared) {
8119 			sf_srd_t *srdp = hblktosrd(hmeblkp);
8120 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
8121 			sf_region_t *rgnp;
8122 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8123 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8124 			ASSERT(srdp != NULL);
8125 			rgnp = srdp->srd_hmergnp[rid];
8126 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
8127 			    rgnp, rid);
8128 			cnt += rgnp->rgn_refcnt;
8129 		} else {
8130 			cnt++;
8131 		}
8132 		if (cnt > sh_thresh) {
8133 			sfmmu_mlist_exit(pml);
8134 			return (1);
8135 		}
8136 	}
8137 
8138 	index >>= 1;
8139 	sz++;
8140 	while (index) {
8141 		pp = PP_GROUPLEADER(pp, sz);
8142 		ASSERT(sfmmu_mlist_held(pp));
8143 		if (index & 0x1) {
8144 			goto again;
8145 		}
8146 		index >>= 1;
8147 		sz++;
8148 	}
8149 	sfmmu_mlist_exit(pml);
8150 	return (0);
8151 }
8152 
8153 /*
8154  * Unload all large mappings to the pp and reset the p_szc field of every
8155  * constituent page according to the remaining mappings.
8156  *
8157  * pp must be locked SE_EXCL. Even though no other constituent pages are
8158  * locked it's legal to unload the large mappings to the pp because all
8159  * constituent pages of large locked mappings have to be locked SE_SHARED.
8160  * This means if we have SE_EXCL lock on one of constituent pages none of the
8161  * large mappings to pp are locked.
8162  *
8163  * Decrease p_szc field starting from the last constituent page and ending
8164  * with the root page. This method is used because other threads rely on the
8165  * root's p_szc to find the lock to syncronize on. After a root page_t's p_szc
8166  * is demoted then other threads will succeed in sfmmu_mlspl_enter(). This
8167  * ensures that p_szc changes of the constituent pages appears atomic for all
8168  * threads that use sfmmu_mlspl_enter() to examine p_szc field.
8169  *
8170  * This mechanism is only used for file system pages where it's not always
8171  * possible to get SE_EXCL locks on all constituent pages to demote the size
8172  * code (as is done for anonymous or kernel large pages).
8173  *
8174  * See more comments in front of sfmmu_mlspl_enter().
8175  */
8176 void
8177 hat_page_demote(page_t *pp)
8178 {
8179 	int index;
8180 	int sz;
8181 	cpuset_t cpuset;
8182 	int sync = 0;
8183 	page_t *rootpp;
8184 	struct sf_hment *sfhme;
8185 	struct sf_hment *tmphme = NULL;
8186 	struct hme_blk *hmeblkp;
8187 	uint_t pszc;
8188 	page_t *lastpp;
8189 	cpuset_t tset;
8190 	pgcnt_t npgs;
8191 	kmutex_t *pml;
8192 	kmutex_t *pmtx = NULL;
8193 
8194 	ASSERT(PAGE_EXCL(pp));
8195 	ASSERT(!PP_ISFREE(pp));
8196 	ASSERT(!PP_ISKAS(pp));
8197 	ASSERT(page_szc_lock_assert(pp));
8198 	pml = sfmmu_mlist_enter(pp);
8199 
8200 	pszc = pp->p_szc;
8201 	if (pszc == 0) {
8202 		goto out;
8203 	}
8204 
8205 	index = PP_MAPINDEX(pp) >> 1;
8206 
8207 	if (index) {
8208 		CPUSET_ZERO(cpuset);
8209 		sz = TTE64K;
8210 		sync = 1;
8211 	}
8212 
8213 	while (index) {
8214 		if (!(index & 0x1)) {
8215 			index >>= 1;
8216 			sz++;
8217 			continue;
8218 		}
8219 		ASSERT(sz <= pszc);
8220 		rootpp = PP_GROUPLEADER(pp, sz);
8221 		for (sfhme = rootpp->p_mapping; sfhme; sfhme = tmphme) {
8222 			tmphme = sfhme->hme_next;
8223 			ASSERT(!IS_PAHME(sfhme));
8224 			hmeblkp = sfmmu_hmetohblk(sfhme);
8225 			if (hme_size(sfhme) != sz) {
8226 				continue;
8227 			}
8228 			tset = sfmmu_pageunload(rootpp, sfhme, sz);
8229 			CPUSET_OR(cpuset, tset);
8230 		}
8231 		if (index >>= 1) {
8232 			sz++;
8233 		}
8234 	}
8235 
8236 	ASSERT(!PP_ISMAPPED_LARGE(pp));
8237 
8238 	if (sync) {
8239 		xt_sync(cpuset);
8240 #ifdef VAC
8241 		if (PP_ISTNC(pp)) {
8242 			conv_tnc(rootpp, sz);
8243 		}
8244 #endif	/* VAC */
8245 	}
8246 
8247 	pmtx = sfmmu_page_enter(pp);
8248 
8249 	ASSERT(pp->p_szc == pszc);
8250 	rootpp = PP_PAGEROOT(pp);
8251 	ASSERT(rootpp->p_szc == pszc);
8252 	lastpp = PP_PAGENEXT_N(rootpp, TTEPAGES(pszc) - 1);
8253 
8254 	while (lastpp != rootpp) {
8255 		sz = PP_MAPINDEX(lastpp) ? fnd_mapping_sz(lastpp) : 0;
8256 		ASSERT(sz < pszc);
8257 		npgs = (sz == 0) ? 1 : TTEPAGES(sz);
8258 		ASSERT(P2PHASE(lastpp->p_pagenum, npgs) == npgs - 1);
8259 		while (--npgs > 0) {
8260 			lastpp->p_szc = (uchar_t)sz;
8261 			lastpp = PP_PAGEPREV(lastpp);
8262 		}
8263 		if (sz) {
8264 			/*
8265 			 * make sure before current root's pszc
8266 			 * is updated all updates to constituent pages pszc
8267 			 * fields are globally visible.
8268 			 */
8269 			membar_producer();
8270 		}
8271 		lastpp->p_szc = sz;
8272 		ASSERT(IS_P2ALIGNED(lastpp->p_pagenum, TTEPAGES(sz)));
8273 		if (lastpp != rootpp) {
8274 			lastpp = PP_PAGEPREV(lastpp);
8275 		}
8276 	}
8277 	if (sz == 0) {
8278 		/* the loop above doesn't cover this case */
8279 		rootpp->p_szc = 0;
8280 	}
8281 out:
8282 	ASSERT(pp->p_szc == 0);
8283 	if (pmtx != NULL) {
8284 		sfmmu_page_exit(pmtx);
8285 	}
8286 	sfmmu_mlist_exit(pml);
8287 }
8288 
8289 /*
8290  * Refresh the HAT ismttecnt[] element for size szc.
8291  * Caller must have set ISM busy flag to prevent mapping
8292  * lists from changing while we're traversing them.
8293  */
8294 pgcnt_t
8295 ism_tsb_entries(sfmmu_t *sfmmup, int szc)
8296 {
8297 	ism_blk_t	*ism_blkp = sfmmup->sfmmu_iblk;
8298 	ism_map_t	*ism_map;
8299 	pgcnt_t		npgs = 0;
8300 	pgcnt_t		npgs_scd = 0;
8301 	int		j;
8302 	sf_scd_t	*scdp;
8303 	uchar_t		rid;
8304 
8305 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
8306 	scdp = sfmmup->sfmmu_scdp;
8307 
8308 	for (; ism_blkp != NULL; ism_blkp = ism_blkp->iblk_next) {
8309 		ism_map = ism_blkp->iblk_maps;
8310 		for (j = 0; ism_map[j].imap_ismhat && j < ISM_MAP_SLOTS; j++) {
8311 			rid = ism_map[j].imap_rid;
8312 			ASSERT(rid == SFMMU_INVALID_ISMRID ||
8313 			    rid < sfmmup->sfmmu_srdp->srd_next_ismrid);
8314 
8315 			if (scdp != NULL && rid != SFMMU_INVALID_ISMRID &&
8316 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
8317 				/* ISM is in sfmmup's SCD */
8318 				npgs_scd +=
8319 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8320 			} else {
8321 				/* ISMs is not in SCD */
8322 				npgs +=
8323 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8324 			}
8325 		}
8326 	}
8327 	sfmmup->sfmmu_ismttecnt[szc] = npgs;
8328 	sfmmup->sfmmu_scdismttecnt[szc] = npgs_scd;
8329 	return (npgs);
8330 }
8331 
8332 /*
8333  * Yield the memory claim requirement for an address space.
8334  *
8335  * This is currently implemented as the number of bytes that have active
8336  * hardware translations that have page structures.  Therefore, it can
8337  * underestimate the traditional resident set size, eg, if the
8338  * physical page is present and the hardware translation is missing;
8339  * and it can overestimate the rss, eg, if there are active
8340  * translations to a frame buffer with page structs.
8341  * Also, it does not take sharing into account.
8342  *
8343  * Note that we don't acquire locks here since this function is most often
8344  * called from the clock thread.
8345  */
8346 size_t
8347 hat_get_mapped_size(struct hat *hat)
8348 {
8349 	size_t		assize = 0;
8350 	int		i;
8351 
8352 	if (hat == NULL)
8353 		return (0);
8354 
8355 	for (i = 0; i < mmu_page_sizes; i++)
8356 		assize += ((pgcnt_t)hat->sfmmu_ttecnt[i] +
8357 		    (pgcnt_t)hat->sfmmu_scdrttecnt[i]) * TTEBYTES(i);
8358 
8359 	if (hat->sfmmu_iblk == NULL)
8360 		return (assize);
8361 
8362 	for (i = 0; i < mmu_page_sizes; i++)
8363 		assize += ((pgcnt_t)hat->sfmmu_ismttecnt[i] +
8364 		    (pgcnt_t)hat->sfmmu_scdismttecnt[i]) * TTEBYTES(i);
8365 
8366 	return (assize);
8367 }
8368 
8369 int
8370 hat_stats_enable(struct hat *hat)
8371 {
8372 	hatlock_t	*hatlockp;
8373 
8374 	hatlockp = sfmmu_hat_enter(hat);
8375 	hat->sfmmu_rmstat++;
8376 	sfmmu_hat_exit(hatlockp);
8377 	return (1);
8378 }
8379 
8380 void
8381 hat_stats_disable(struct hat *hat)
8382 {
8383 	hatlock_t	*hatlockp;
8384 
8385 	hatlockp = sfmmu_hat_enter(hat);
8386 	hat->sfmmu_rmstat--;
8387 	sfmmu_hat_exit(hatlockp);
8388 }
8389 
8390 /*
8391  * Routines for entering or removing  ourselves from the
8392  * ism_hat's mapping list. This is used for both private and
8393  * SCD hats.
8394  */
8395 static void
8396 iment_add(struct ism_ment *iment,  struct hat *ism_hat)
8397 {
8398 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8399 
8400 	iment->iment_prev = NULL;
8401 	iment->iment_next = ism_hat->sfmmu_iment;
8402 	if (ism_hat->sfmmu_iment) {
8403 		ism_hat->sfmmu_iment->iment_prev = iment;
8404 	}
8405 	ism_hat->sfmmu_iment = iment;
8406 }
8407 
8408 static void
8409 iment_sub(struct ism_ment *iment, struct hat *ism_hat)
8410 {
8411 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8412 
8413 	if (ism_hat->sfmmu_iment == NULL) {
8414 		panic("ism map entry remove - no entries");
8415 	}
8416 
8417 	if (iment->iment_prev) {
8418 		ASSERT(ism_hat->sfmmu_iment != iment);
8419 		iment->iment_prev->iment_next = iment->iment_next;
8420 	} else {
8421 		ASSERT(ism_hat->sfmmu_iment == iment);
8422 		ism_hat->sfmmu_iment = iment->iment_next;
8423 	}
8424 
8425 	if (iment->iment_next) {
8426 		iment->iment_next->iment_prev = iment->iment_prev;
8427 	}
8428 
8429 	/*
8430 	 * zero out the entry
8431 	 */
8432 	iment->iment_next = NULL;
8433 	iment->iment_prev = NULL;
8434 	iment->iment_hat =  NULL;
8435 	iment->iment_base_va = 0;
8436 }
8437 
8438 /*
8439  * Hat_share()/unshare() return an (non-zero) error
8440  * when saddr and daddr are not properly aligned.
8441  *
8442  * The top level mapping element determines the alignment
8443  * requirement for saddr and daddr, depending on different
8444  * architectures.
8445  *
8446  * When hat_share()/unshare() are not supported,
8447  * HATOP_SHARE()/UNSHARE() return 0
8448  */
8449 int
8450 hat_share(struct hat *sfmmup, caddr_t addr, struct hat *ism_hatid,
8451     caddr_t sptaddr, size_t len, uint_t ismszc)
8452 {
8453 	ism_blk_t	*ism_blkp;
8454 	ism_blk_t	*new_iblk;
8455 	ism_map_t	*ism_map;
8456 	ism_ment_t	*ism_ment;
8457 	int		i, added;
8458 	hatlock_t	*hatlockp;
8459 	int		reload_mmu = 0;
8460 	uint_t		ismshift = page_get_shift(ismszc);
8461 	size_t		ismpgsz = page_get_pagesize(ismszc);
8462 	uint_t		ismmask = (uint_t)ismpgsz - 1;
8463 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8464 	ushort_t	ismhatflag;
8465 	hat_region_cookie_t rcookie;
8466 	sf_scd_t	*old_scdp;
8467 
8468 #ifdef DEBUG
8469 	caddr_t		eaddr = addr + len;
8470 #endif /* DEBUG */
8471 
8472 	ASSERT(ism_hatid != NULL && sfmmup != NULL);
8473 	ASSERT(sptaddr == ISMID_STARTADDR);
8474 	/*
8475 	 * Check the alignment.
8476 	 */
8477 	if (!ISM_ALIGNED(ismshift, addr) || !ISM_ALIGNED(ismshift, sptaddr))
8478 		return (EINVAL);
8479 
8480 	/*
8481 	 * Check size alignment.
8482 	 */
8483 	if (!ISM_ALIGNED(ismshift, len))
8484 		return (EINVAL);
8485 
8486 	/*
8487 	 * Allocate ism_ment for the ism_hat's mapping list, and an
8488 	 * ism map blk in case we need one.  We must do our
8489 	 * allocations before acquiring locks to prevent a deadlock
8490 	 * in the kmem allocator on the mapping list lock.
8491 	 */
8492 	new_iblk = kmem_cache_alloc(ism_blk_cache, KM_SLEEP);
8493 	ism_ment = kmem_cache_alloc(ism_ment_cache, KM_SLEEP);
8494 
8495 	/*
8496 	 * Serialize ISM mappings with the ISM busy flag, and also the
8497 	 * trap handlers.
8498 	 */
8499 	sfmmu_ismhat_enter(sfmmup, 0);
8500 
8501 	/*
8502 	 * Allocate an ism map blk if necessary.
8503 	 */
8504 	if (sfmmup->sfmmu_iblk == NULL) {
8505 		sfmmup->sfmmu_iblk = new_iblk;
8506 		bzero(new_iblk, sizeof (*new_iblk));
8507 		new_iblk->iblk_nextpa = (uint64_t)-1;
8508 		membar_stst();	/* make sure next ptr visible to all CPUs */
8509 		sfmmup->sfmmu_ismblkpa = va_to_pa((caddr_t)new_iblk);
8510 		reload_mmu = 1;
8511 		new_iblk = NULL;
8512 	}
8513 
8514 #ifdef DEBUG
8515 	/*
8516 	 * Make sure mapping does not already exist.
8517 	 */
8518 	ism_blkp = sfmmup->sfmmu_iblk;
8519 	while (ism_blkp != NULL) {
8520 		ism_map = ism_blkp->iblk_maps;
8521 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
8522 			if ((addr >= ism_start(ism_map[i]) &&
8523 			    addr < ism_end(ism_map[i])) ||
8524 			    eaddr > ism_start(ism_map[i]) &&
8525 			    eaddr <= ism_end(ism_map[i])) {
8526 				panic("sfmmu_share: Already mapped!");
8527 			}
8528 		}
8529 		ism_blkp = ism_blkp->iblk_next;
8530 	}
8531 #endif /* DEBUG */
8532 
8533 	ASSERT(ismszc >= TTE4M);
8534 	if (ismszc == TTE4M) {
8535 		ismhatflag = HAT_4M_FLAG;
8536 	} else if (ismszc == TTE32M) {
8537 		ismhatflag = HAT_32M_FLAG;
8538 	} else if (ismszc == TTE256M) {
8539 		ismhatflag = HAT_256M_FLAG;
8540 	}
8541 	/*
8542 	 * Add mapping to first available mapping slot.
8543 	 */
8544 	ism_blkp = sfmmup->sfmmu_iblk;
8545 	added = 0;
8546 	while (!added) {
8547 		ism_map = ism_blkp->iblk_maps;
8548 		for (i = 0; i < ISM_MAP_SLOTS; i++)  {
8549 			if (ism_map[i].imap_ismhat == NULL) {
8550 
8551 				ism_map[i].imap_ismhat = ism_hatid;
8552 				ism_map[i].imap_vb_shift = (uchar_t)ismshift;
8553 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8554 				ism_map[i].imap_hatflags = ismhatflag;
8555 				ism_map[i].imap_sz_mask = ismmask;
8556 				/*
8557 				 * imap_seg is checked in ISM_CHECK to see if
8558 				 * non-NULL, then other info assumed valid.
8559 				 */
8560 				membar_stst();
8561 				ism_map[i].imap_seg = (uintptr_t)addr | sh_size;
8562 				ism_map[i].imap_ment = ism_ment;
8563 
8564 				/*
8565 				 * Now add ourselves to the ism_hat's
8566 				 * mapping list.
8567 				 */
8568 				ism_ment->iment_hat = sfmmup;
8569 				ism_ment->iment_base_va = addr;
8570 				ism_hatid->sfmmu_ismhat = 1;
8571 				mutex_enter(&ism_mlist_lock);
8572 				iment_add(ism_ment, ism_hatid);
8573 				mutex_exit(&ism_mlist_lock);
8574 				added = 1;
8575 				break;
8576 			}
8577 		}
8578 		if (!added && ism_blkp->iblk_next == NULL) {
8579 			ism_blkp->iblk_next = new_iblk;
8580 			new_iblk = NULL;
8581 			bzero(ism_blkp->iblk_next,
8582 			    sizeof (*ism_blkp->iblk_next));
8583 			ism_blkp->iblk_next->iblk_nextpa = (uint64_t)-1;
8584 			membar_stst();
8585 			ism_blkp->iblk_nextpa =
8586 			    va_to_pa((caddr_t)ism_blkp->iblk_next);
8587 		}
8588 		ism_blkp = ism_blkp->iblk_next;
8589 	}
8590 
8591 	/*
8592 	 * After calling hat_join_region, sfmmup may join a new SCD or
8593 	 * move from the old scd to a new scd, in which case, we want to
8594 	 * shrink the sfmmup's private tsb size, i.e., pass shrink to
8595 	 * sfmmu_check_page_sizes at the end of this routine.
8596 	 */
8597 	old_scdp = sfmmup->sfmmu_scdp;
8598 
8599 	rcookie = hat_join_region(sfmmup, addr, len, (void *)ism_hatid, 0,
8600 	    PROT_ALL, ismszc, NULL, HAT_REGION_ISM);
8601 	if (rcookie != HAT_INVALID_REGION_COOKIE) {
8602 		ism_map[i].imap_rid = (uchar_t)((uint64_t)rcookie);
8603 	}
8604 	/*
8605 	 * Update our counters for this sfmmup's ism mappings.
8606 	 */
8607 	for (i = 0; i <= ismszc; i++) {
8608 		if (!(disable_ism_large_pages & (1 << i)))
8609 			(void) ism_tsb_entries(sfmmup, i);
8610 	}
8611 
8612 	/*
8613 	 * For ISM and DISM we do not support 512K pages, so we only only
8614 	 * search the 4M and 8K/64K hashes for 4 pagesize cpus, and search the
8615 	 * 256M or 32M, and 4M and 8K/64K hashes for 6 pagesize cpus.
8616 	 *
8617 	 * Need to set 32M/256M ISM flags to make sure
8618 	 * sfmmu_check_page_sizes() enables them on Panther.
8619 	 */
8620 	ASSERT((disable_ism_large_pages & (1 << TTE512K)) != 0);
8621 
8622 	switch (ismszc) {
8623 	case TTE256M:
8624 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_256M_ISM)) {
8625 			hatlockp = sfmmu_hat_enter(sfmmup);
8626 			SFMMU_FLAGS_SET(sfmmup, HAT_256M_ISM);
8627 			sfmmu_hat_exit(hatlockp);
8628 		}
8629 		break;
8630 	case TTE32M:
8631 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_32M_ISM)) {
8632 			hatlockp = sfmmu_hat_enter(sfmmup);
8633 			SFMMU_FLAGS_SET(sfmmup, HAT_32M_ISM);
8634 			sfmmu_hat_exit(hatlockp);
8635 		}
8636 		break;
8637 	default:
8638 		break;
8639 	}
8640 
8641 	/*
8642 	 * If we updated the ismblkpa for this HAT we must make
8643 	 * sure all CPUs running this process reload their tsbmiss area.
8644 	 * Otherwise they will fail to load the mappings in the tsbmiss
8645 	 * handler and will loop calling pagefault().
8646 	 */
8647 	if (reload_mmu) {
8648 		hatlockp = sfmmu_hat_enter(sfmmup);
8649 		sfmmu_sync_mmustate(sfmmup);
8650 		sfmmu_hat_exit(hatlockp);
8651 	}
8652 
8653 	sfmmu_ismhat_exit(sfmmup, 0);
8654 
8655 	/*
8656 	 * Free up ismblk if we didn't use it.
8657 	 */
8658 	if (new_iblk != NULL)
8659 		kmem_cache_free(ism_blk_cache, new_iblk);
8660 
8661 	/*
8662 	 * Check TSB and TLB page sizes.
8663 	 */
8664 	if (sfmmup->sfmmu_scdp != NULL && old_scdp != sfmmup->sfmmu_scdp) {
8665 		sfmmu_check_page_sizes(sfmmup, 0);
8666 	} else {
8667 		sfmmu_check_page_sizes(sfmmup, 1);
8668 	}
8669 	return (0);
8670 }
8671 
8672 /*
8673  * hat_unshare removes exactly one ism_map from
8674  * this process's as.  It expects multiple calls
8675  * to hat_unshare for multiple shm segments.
8676  */
8677 void
8678 hat_unshare(struct hat *sfmmup, caddr_t addr, size_t len, uint_t ismszc)
8679 {
8680 	ism_map_t	*ism_map;
8681 	ism_ment_t	*free_ment = NULL;
8682 	ism_blk_t	*ism_blkp;
8683 	struct hat	*ism_hatid;
8684 	int		found, i;
8685 	hatlock_t	*hatlockp;
8686 	struct tsb_info	*tsbinfo;
8687 	uint_t		ismshift = page_get_shift(ismszc);
8688 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8689 	uchar_t		ism_rid;
8690 	sf_scd_t	*old_scdp;
8691 
8692 	ASSERT(ISM_ALIGNED(ismshift, addr));
8693 	ASSERT(ISM_ALIGNED(ismshift, len));
8694 	ASSERT(sfmmup != NULL);
8695 	ASSERT(sfmmup != ksfmmup);
8696 
8697 	ASSERT(sfmmup->sfmmu_as != NULL);
8698 
8699 	/*
8700 	 * Make sure that during the entire time ISM mappings are removed,
8701 	 * the trap handlers serialize behind us, and that no one else
8702 	 * can be mucking with ISM mappings.  This also lets us get away
8703 	 * with not doing expensive cross calls to flush the TLB -- we
8704 	 * just discard the context, flush the entire TSB, and call it
8705 	 * a day.
8706 	 */
8707 	sfmmu_ismhat_enter(sfmmup, 0);
8708 
8709 	/*
8710 	 * Remove the mapping.
8711 	 *
8712 	 * We can't have any holes in the ism map.
8713 	 * The tsb miss code while searching the ism map will
8714 	 * stop on an empty map slot.  So we must move
8715 	 * everyone past the hole up 1 if any.
8716 	 *
8717 	 * Also empty ism map blks are not freed until the
8718 	 * process exits. This is to prevent a MT race condition
8719 	 * between sfmmu_unshare() and sfmmu_tsbmiss_exception().
8720 	 */
8721 	found = 0;
8722 	ism_blkp = sfmmup->sfmmu_iblk;
8723 	while (!found && ism_blkp != NULL) {
8724 		ism_map = ism_blkp->iblk_maps;
8725 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
8726 			if (addr == ism_start(ism_map[i]) &&
8727 			    sh_size == (size_t)(ism_size(ism_map[i]))) {
8728 				found = 1;
8729 				break;
8730 			}
8731 		}
8732 		if (!found)
8733 			ism_blkp = ism_blkp->iblk_next;
8734 	}
8735 
8736 	if (found) {
8737 		ism_hatid = ism_map[i].imap_ismhat;
8738 		ism_rid = ism_map[i].imap_rid;
8739 		ASSERT(ism_hatid != NULL);
8740 		ASSERT(ism_hatid->sfmmu_ismhat == 1);
8741 
8742 		/*
8743 		 * After hat_leave_region, the sfmmup may leave SCD,
8744 		 * in which case, we want to grow the private tsb size when
8745 		 * calling sfmmu_check_page_sizes at the end of the routine.
8746 		 */
8747 		old_scdp = sfmmup->sfmmu_scdp;
8748 		/*
8749 		 * Then remove ourselves from the region.
8750 		 */
8751 		if (ism_rid != SFMMU_INVALID_ISMRID) {
8752 			hat_leave_region(sfmmup, (void *)((uint64_t)ism_rid),
8753 			    HAT_REGION_ISM);
8754 		}
8755 
8756 		/*
8757 		 * And now guarantee that any other cpu
8758 		 * that tries to process an ISM miss
8759 		 * will go to tl=0.
8760 		 */
8761 		hatlockp = sfmmu_hat_enter(sfmmup);
8762 		sfmmu_invalidate_ctx(sfmmup);
8763 		sfmmu_hat_exit(hatlockp);
8764 
8765 		/*
8766 		 * Remove ourselves from the ism mapping list.
8767 		 */
8768 		mutex_enter(&ism_mlist_lock);
8769 		iment_sub(ism_map[i].imap_ment, ism_hatid);
8770 		mutex_exit(&ism_mlist_lock);
8771 		free_ment = ism_map[i].imap_ment;
8772 
8773 		/*
8774 		 * We delete the ism map by copying
8775 		 * the next map over the current one.
8776 		 * We will take the next one in the maps
8777 		 * array or from the next ism_blk.
8778 		 */
8779 		while (ism_blkp != NULL) {
8780 			ism_map = ism_blkp->iblk_maps;
8781 			while (i < (ISM_MAP_SLOTS - 1)) {
8782 				ism_map[i] = ism_map[i + 1];
8783 				i++;
8784 			}
8785 			/* i == (ISM_MAP_SLOTS - 1) */
8786 			ism_blkp = ism_blkp->iblk_next;
8787 			if (ism_blkp != NULL) {
8788 				ism_map[i] = ism_blkp->iblk_maps[0];
8789 				i = 0;
8790 			} else {
8791 				ism_map[i].imap_seg = 0;
8792 				ism_map[i].imap_vb_shift = 0;
8793 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8794 				ism_map[i].imap_hatflags = 0;
8795 				ism_map[i].imap_sz_mask = 0;
8796 				ism_map[i].imap_ismhat = NULL;
8797 				ism_map[i].imap_ment = NULL;
8798 			}
8799 		}
8800 
8801 		/*
8802 		 * Now flush entire TSB for the process, since
8803 		 * demapping page by page can be too expensive.
8804 		 * We don't have to flush the TLB here anymore
8805 		 * since we switch to a new TLB ctx instead.
8806 		 * Also, there is no need to flush if the process
8807 		 * is exiting since the TSB will be freed later.
8808 		 */
8809 		if (!sfmmup->sfmmu_free) {
8810 			hatlockp = sfmmu_hat_enter(sfmmup);
8811 			for (tsbinfo = sfmmup->sfmmu_tsb; tsbinfo != NULL;
8812 			    tsbinfo = tsbinfo->tsb_next) {
8813 				if (tsbinfo->tsb_flags & TSB_SWAPPED)
8814 					continue;
8815 				if (tsbinfo->tsb_flags & TSB_RELOC_FLAG) {
8816 					tsbinfo->tsb_flags |=
8817 					    TSB_FLUSH_NEEDED;
8818 					continue;
8819 				}
8820 
8821 				sfmmu_inv_tsb(tsbinfo->tsb_va,
8822 				    TSB_BYTES(tsbinfo->tsb_szc));
8823 			}
8824 			sfmmu_hat_exit(hatlockp);
8825 		}
8826 	}
8827 
8828 	/*
8829 	 * Update our counters for this sfmmup's ism mappings.
8830 	 */
8831 	for (i = 0; i <= ismszc; i++) {
8832 		if (!(disable_ism_large_pages & (1 << i)))
8833 			(void) ism_tsb_entries(sfmmup, i);
8834 	}
8835 
8836 	sfmmu_ismhat_exit(sfmmup, 0);
8837 
8838 	/*
8839 	 * We must do our freeing here after dropping locks
8840 	 * to prevent a deadlock in the kmem allocator on the
8841 	 * mapping list lock.
8842 	 */
8843 	if (free_ment != NULL)
8844 		kmem_cache_free(ism_ment_cache, free_ment);
8845 
8846 	/*
8847 	 * Check TSB and TLB page sizes if the process isn't exiting.
8848 	 */
8849 	if (!sfmmup->sfmmu_free) {
8850 		if (found && old_scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
8851 			sfmmu_check_page_sizes(sfmmup, 1);
8852 		} else {
8853 			sfmmu_check_page_sizes(sfmmup, 0);
8854 		}
8855 	}
8856 }
8857 
8858 /* ARGSUSED */
8859 static int
8860 sfmmu_idcache_constructor(void *buf, void *cdrarg, int kmflags)
8861 {
8862 	/* void *buf is sfmmu_t pointer */
8863 	bzero(buf, sizeof (sfmmu_t));
8864 
8865 	return (0);
8866 }
8867 
8868 /* ARGSUSED */
8869 static void
8870 sfmmu_idcache_destructor(void *buf, void *cdrarg)
8871 {
8872 	/* void *buf is sfmmu_t pointer */
8873 }
8874 
8875 /*
8876  * setup kmem hmeblks by bzeroing all members and initializing the nextpa
8877  * field to be the pa of this hmeblk
8878  */
8879 /* ARGSUSED */
8880 static int
8881 sfmmu_hblkcache_constructor(void *buf, void *cdrarg, int kmflags)
8882 {
8883 	struct hme_blk *hmeblkp;
8884 
8885 	bzero(buf, (size_t)cdrarg);
8886 	hmeblkp = (struct hme_blk *)buf;
8887 	hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
8888 
8889 #ifdef	HBLK_TRACE
8890 	mutex_init(&hmeblkp->hblk_audit_lock, NULL, MUTEX_DEFAULT, NULL);
8891 #endif	/* HBLK_TRACE */
8892 
8893 	return (0);
8894 }
8895 
8896 /* ARGSUSED */
8897 static void
8898 sfmmu_hblkcache_destructor(void *buf, void *cdrarg)
8899 {
8900 
8901 #ifdef	HBLK_TRACE
8902 
8903 	struct hme_blk *hmeblkp;
8904 
8905 	hmeblkp = (struct hme_blk *)buf;
8906 	mutex_destroy(&hmeblkp->hblk_audit_lock);
8907 
8908 #endif	/* HBLK_TRACE */
8909 }
8910 
8911 #define	SFMMU_CACHE_RECLAIM_SCAN_RATIO 8
8912 static int sfmmu_cache_reclaim_scan_ratio = SFMMU_CACHE_RECLAIM_SCAN_RATIO;
8913 /*
8914  * The kmem allocator will callback into our reclaim routine when the system
8915  * is running low in memory.  We traverse the hash and free up all unused but
8916  * still cached hme_blks.  We also traverse the free list and free them up
8917  * as well.
8918  */
8919 /*ARGSUSED*/
8920 static void
8921 sfmmu_hblkcache_reclaim(void *cdrarg)
8922 {
8923 	int i;
8924 	struct hmehash_bucket *hmebp;
8925 	struct hme_blk *hmeblkp, *nx_hblk, *pr_hblk = NULL;
8926 	static struct hmehash_bucket *uhmehash_reclaim_hand;
8927 	static struct hmehash_bucket *khmehash_reclaim_hand;
8928 	struct hme_blk *list = NULL, *last_hmeblkp;
8929 	cpuset_t cpuset = cpu_ready_set;
8930 	cpu_hme_pend_t *cpuhp;
8931 
8932 	/* Free up hmeblks on the cpu pending lists */
8933 	for (i = 0; i < NCPU; i++) {
8934 		cpuhp = &cpu_hme_pend[i];
8935 		if (cpuhp->chp_listp != NULL)  {
8936 			mutex_enter(&cpuhp->chp_mutex);
8937 			if (cpuhp->chp_listp == NULL) {
8938 				mutex_exit(&cpuhp->chp_mutex);
8939 				continue;
8940 			}
8941 			for (last_hmeblkp = cpuhp->chp_listp;
8942 			    last_hmeblkp->hblk_next != NULL;
8943 			    last_hmeblkp = last_hmeblkp->hblk_next)
8944 				;
8945 			last_hmeblkp->hblk_next = list;
8946 			list = cpuhp->chp_listp;
8947 			cpuhp->chp_listp = NULL;
8948 			cpuhp->chp_count = 0;
8949 			mutex_exit(&cpuhp->chp_mutex);
8950 		}
8951 
8952 	}
8953 
8954 	if (list != NULL) {
8955 		kpreempt_disable();
8956 		CPUSET_DEL(cpuset, CPU->cpu_id);
8957 		xt_sync(cpuset);
8958 		xt_sync(cpuset);
8959 		kpreempt_enable();
8960 		sfmmu_hblk_free(&list);
8961 		list = NULL;
8962 	}
8963 
8964 	hmebp = uhmehash_reclaim_hand;
8965 	if (hmebp == NULL || hmebp > &uhme_hash[UHMEHASH_SZ])
8966 		uhmehash_reclaim_hand = hmebp = uhme_hash;
8967 	uhmehash_reclaim_hand += UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
8968 
8969 	for (i = UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
8970 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
8971 			hmeblkp = hmebp->hmeblkp;
8972 			pr_hblk = NULL;
8973 			while (hmeblkp) {
8974 				nx_hblk = hmeblkp->hblk_next;
8975 				if (!hmeblkp->hblk_vcnt &&
8976 				    !hmeblkp->hblk_hmecnt) {
8977 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
8978 					    pr_hblk, &list, 0);
8979 				} else {
8980 					pr_hblk = hmeblkp;
8981 				}
8982 				hmeblkp = nx_hblk;
8983 			}
8984 			SFMMU_HASH_UNLOCK(hmebp);
8985 		}
8986 		if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
8987 			hmebp = uhme_hash;
8988 	}
8989 
8990 	hmebp = khmehash_reclaim_hand;
8991 	if (hmebp == NULL || hmebp > &khme_hash[KHMEHASH_SZ])
8992 		khmehash_reclaim_hand = hmebp = khme_hash;
8993 	khmehash_reclaim_hand += KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
8994 
8995 	for (i = KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
8996 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
8997 			hmeblkp = hmebp->hmeblkp;
8998 			pr_hblk = NULL;
8999 			while (hmeblkp) {
9000 				nx_hblk = hmeblkp->hblk_next;
9001 				if (!hmeblkp->hblk_vcnt &&
9002 				    !hmeblkp->hblk_hmecnt) {
9003 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9004 					    pr_hblk, &list, 0);
9005 				} else {
9006 					pr_hblk = hmeblkp;
9007 				}
9008 				hmeblkp = nx_hblk;
9009 			}
9010 			SFMMU_HASH_UNLOCK(hmebp);
9011 		}
9012 		if (hmebp++ == &khme_hash[KHMEHASH_SZ])
9013 			hmebp = khme_hash;
9014 	}
9015 	sfmmu_hblks_list_purge(&list, 0);
9016 }
9017 
9018 /*
9019  * sfmmu_get_ppvcolor should become a vm_machdep or hatop interface.
9020  * same goes for sfmmu_get_addrvcolor().
9021  *
9022  * This function will return the virtual color for the specified page. The
9023  * virtual color corresponds to this page current mapping or its last mapping.
9024  * It is used by memory allocators to choose addresses with the correct
9025  * alignment so vac consistency is automatically maintained.  If the page
9026  * has no color it returns -1.
9027  */
9028 /*ARGSUSED*/
9029 int
9030 sfmmu_get_ppvcolor(struct page *pp)
9031 {
9032 #ifdef VAC
9033 	int color;
9034 
9035 	if (!(cache & CACHE_VAC) || PP_NEWPAGE(pp)) {
9036 		return (-1);
9037 	}
9038 	color = PP_GET_VCOLOR(pp);
9039 	ASSERT(color < mmu_btop(shm_alignment));
9040 	return (color);
9041 #else
9042 	return (-1);
9043 #endif	/* VAC */
9044 }
9045 
9046 /*
9047  * This function will return the desired alignment for vac consistency
9048  * (vac color) given a virtual address.  If no vac is present it returns -1.
9049  */
9050 /*ARGSUSED*/
9051 int
9052 sfmmu_get_addrvcolor(caddr_t vaddr)
9053 {
9054 #ifdef VAC
9055 	if (cache & CACHE_VAC) {
9056 		return (addr_to_vcolor(vaddr));
9057 	} else {
9058 		return (-1);
9059 	}
9060 #else
9061 	return (-1);
9062 #endif	/* VAC */
9063 }
9064 
9065 #ifdef VAC
9066 /*
9067  * Check for conflicts.
9068  * A conflict exists if the new and existent mappings do not match in
9069  * their "shm_alignment fields. If conflicts exist, the existant mappings
9070  * are flushed unless one of them is locked. If one of them is locked, then
9071  * the mappings are flushed and converted to non-cacheable mappings.
9072  */
9073 static void
9074 sfmmu_vac_conflict(struct hat *hat, caddr_t addr, page_t *pp)
9075 {
9076 	struct hat *tmphat;
9077 	struct sf_hment *sfhmep, *tmphme = NULL;
9078 	struct hme_blk *hmeblkp;
9079 	int vcolor;
9080 	tte_t tte;
9081 
9082 	ASSERT(sfmmu_mlist_held(pp));
9083 	ASSERT(!PP_ISNC(pp));		/* page better be cacheable */
9084 
9085 	vcolor = addr_to_vcolor(addr);
9086 	if (PP_NEWPAGE(pp)) {
9087 		PP_SET_VCOLOR(pp, vcolor);
9088 		return;
9089 	}
9090 
9091 	if (PP_GET_VCOLOR(pp) == vcolor) {
9092 		return;
9093 	}
9094 
9095 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
9096 		/*
9097 		 * Previous user of page had a different color
9098 		 * but since there are no current users
9099 		 * we just flush the cache and change the color.
9100 		 */
9101 		SFMMU_STAT(sf_pgcolor_conflict);
9102 		sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9103 		PP_SET_VCOLOR(pp, vcolor);
9104 		return;
9105 	}
9106 
9107 	/*
9108 	 * If we get here we have a vac conflict with a current
9109 	 * mapping.  VAC conflict policy is as follows.
9110 	 * - The default is to unload the other mappings unless:
9111 	 * - If we have a large mapping we uncache the page.
9112 	 * We need to uncache the rest of the large page too.
9113 	 * - If any of the mappings are locked we uncache the page.
9114 	 * - If the requested mapping is inconsistent
9115 	 * with another mapping and that mapping
9116 	 * is in the same address space we have to
9117 	 * make it non-cached.  The default thing
9118 	 * to do is unload the inconsistent mapping
9119 	 * but if they are in the same address space
9120 	 * we run the risk of unmapping the pc or the
9121 	 * stack which we will use as we return to the user,
9122 	 * in which case we can then fault on the thing
9123 	 * we just unloaded and get into an infinite loop.
9124 	 */
9125 	if (PP_ISMAPPED_LARGE(pp)) {
9126 		int sz;
9127 
9128 		/*
9129 		 * Existing mapping is for big pages. We don't unload
9130 		 * existing big mappings to satisfy new mappings.
9131 		 * Always convert all mappings to TNC.
9132 		 */
9133 		sz = fnd_mapping_sz(pp);
9134 		pp = PP_GROUPLEADER(pp, sz);
9135 		SFMMU_STAT_ADD(sf_uncache_conflict, TTEPAGES(sz));
9136 		sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH,
9137 		    TTEPAGES(sz));
9138 
9139 		return;
9140 	}
9141 
9142 	/*
9143 	 * check if any mapping is in same as or if it is locked
9144 	 * since in that case we need to uncache.
9145 	 */
9146 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9147 		tmphme = sfhmep->hme_next;
9148 		if (IS_PAHME(sfhmep))
9149 			continue;
9150 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9151 		tmphat = hblktosfmmu(hmeblkp);
9152 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
9153 		ASSERT(TTE_IS_VALID(&tte));
9154 		if (hmeblkp->hblk_shared || tmphat == hat ||
9155 		    hmeblkp->hblk_lckcnt) {
9156 			/*
9157 			 * We have an uncache conflict
9158 			 */
9159 			SFMMU_STAT(sf_uncache_conflict);
9160 			sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH, 1);
9161 			return;
9162 		}
9163 	}
9164 
9165 	/*
9166 	 * We have an unload conflict
9167 	 * We have already checked for LARGE mappings, therefore
9168 	 * the remaining mapping(s) must be TTE8K.
9169 	 */
9170 	SFMMU_STAT(sf_unload_conflict);
9171 
9172 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9173 		tmphme = sfhmep->hme_next;
9174 		if (IS_PAHME(sfhmep))
9175 			continue;
9176 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9177 		ASSERT(!hmeblkp->hblk_shared);
9178 		(void) sfmmu_pageunload(pp, sfhmep, TTE8K);
9179 	}
9180 
9181 	if (PP_ISMAPPED_KPM(pp))
9182 		sfmmu_kpm_vac_unload(pp, addr);
9183 
9184 	/*
9185 	 * Unloads only do TLB flushes so we need to flush the
9186 	 * cache here.
9187 	 */
9188 	sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9189 	PP_SET_VCOLOR(pp, vcolor);
9190 }
9191 
9192 /*
9193  * Whenever a mapping is unloaded and the page is in TNC state,
9194  * we see if the page can be made cacheable again. 'pp' is
9195  * the page that we just unloaded a mapping from, the size
9196  * of mapping that was unloaded is 'ottesz'.
9197  * Remark:
9198  * The recache policy for mpss pages can leave a performance problem
9199  * under the following circumstances:
9200  * . A large page in uncached mode has just been unmapped.
9201  * . All constituent pages are TNC due to a conflicting small mapping.
9202  * . There are many other, non conflicting, small mappings around for
9203  *   a lot of the constituent pages.
9204  * . We're called w/ the "old" groupleader page and the old ottesz,
9205  *   but this is irrelevant, since we're no more "PP_ISMAPPED_LARGE", so
9206  *   we end up w/ TTE8K or npages == 1.
9207  * . We call tst_tnc w/ the old groupleader only, and if there is no
9208  *   conflict, we re-cache only this page.
9209  * . All other small mappings are not checked and will be left in TNC mode.
9210  * The problem is not very serious because:
9211  * . mpss is actually only defined for heap and stack, so the probability
9212  *   is not very high that a large page mapping exists in parallel to a small
9213  *   one (this is possible, but seems to be bad programming style in the
9214  *   appl).
9215  * . The problem gets a little bit more serious, when those TNC pages
9216  *   have to be mapped into kernel space, e.g. for networking.
9217  * . When VAC alias conflicts occur in applications, this is regarded
9218  *   as an application bug. So if kstat's show them, the appl should
9219  *   be changed anyway.
9220  */
9221 void
9222 conv_tnc(page_t *pp, int ottesz)
9223 {
9224 	int cursz, dosz;
9225 	pgcnt_t curnpgs, dopgs;
9226 	pgcnt_t pg64k;
9227 	page_t *pp2;
9228 
9229 	/*
9230 	 * Determine how big a range we check for TNC and find
9231 	 * leader page. cursz is the size of the biggest
9232 	 * mapping that still exist on 'pp'.
9233 	 */
9234 	if (PP_ISMAPPED_LARGE(pp)) {
9235 		cursz = fnd_mapping_sz(pp);
9236 	} else {
9237 		cursz = TTE8K;
9238 	}
9239 
9240 	if (ottesz >= cursz) {
9241 		dosz = ottesz;
9242 		pp2 = pp;
9243 	} else {
9244 		dosz = cursz;
9245 		pp2 = PP_GROUPLEADER(pp, dosz);
9246 	}
9247 
9248 	pg64k = TTEPAGES(TTE64K);
9249 	dopgs = TTEPAGES(dosz);
9250 
9251 	ASSERT(dopgs == 1 || ((dopgs & (pg64k - 1)) == 0));
9252 
9253 	while (dopgs != 0) {
9254 		curnpgs = TTEPAGES(cursz);
9255 		if (tst_tnc(pp2, curnpgs)) {
9256 			SFMMU_STAT_ADD(sf_recache, curnpgs);
9257 			sfmmu_page_cache_array(pp2, HAT_CACHE, CACHE_NO_FLUSH,
9258 			    curnpgs);
9259 		}
9260 
9261 		ASSERT(dopgs >= curnpgs);
9262 		dopgs -= curnpgs;
9263 
9264 		if (dopgs == 0) {
9265 			break;
9266 		}
9267 
9268 		pp2 = PP_PAGENEXT_N(pp2, curnpgs);
9269 		if (((dopgs & (pg64k - 1)) == 0) && PP_ISMAPPED_LARGE(pp2)) {
9270 			cursz = fnd_mapping_sz(pp2);
9271 		} else {
9272 			cursz = TTE8K;
9273 		}
9274 	}
9275 }
9276 
9277 /*
9278  * Returns 1 if page(s) can be converted from TNC to cacheable setting,
9279  * returns 0 otherwise. Note that oaddr argument is valid for only
9280  * 8k pages.
9281  */
9282 int
9283 tst_tnc(page_t *pp, pgcnt_t npages)
9284 {
9285 	struct	sf_hment *sfhme;
9286 	struct	hme_blk *hmeblkp;
9287 	tte_t	tte;
9288 	caddr_t	vaddr;
9289 	int	clr_valid = 0;
9290 	int	color, color1, bcolor;
9291 	int	i, ncolors;
9292 
9293 	ASSERT(pp != NULL);
9294 	ASSERT(!(cache & CACHE_WRITEBACK));
9295 
9296 	if (npages > 1) {
9297 		ncolors = CACHE_NUM_COLOR;
9298 	}
9299 
9300 	for (i = 0; i < npages; i++) {
9301 		ASSERT(sfmmu_mlist_held(pp));
9302 		ASSERT(PP_ISTNC(pp));
9303 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
9304 
9305 		if (PP_ISPNC(pp)) {
9306 			return (0);
9307 		}
9308 
9309 		clr_valid = 0;
9310 		if (PP_ISMAPPED_KPM(pp)) {
9311 			caddr_t kpmvaddr;
9312 
9313 			ASSERT(kpm_enable);
9314 			kpmvaddr = hat_kpm_page2va(pp, 1);
9315 			ASSERT(!(npages > 1 && IS_KPM_ALIAS_RANGE(kpmvaddr)));
9316 			color1 = addr_to_vcolor(kpmvaddr);
9317 			clr_valid = 1;
9318 		}
9319 
9320 		for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9321 			if (IS_PAHME(sfhme))
9322 				continue;
9323 			hmeblkp = sfmmu_hmetohblk(sfhme);
9324 
9325 			sfmmu_copytte(&sfhme->hme_tte, &tte);
9326 			ASSERT(TTE_IS_VALID(&tte));
9327 
9328 			vaddr = tte_to_vaddr(hmeblkp, tte);
9329 			color = addr_to_vcolor(vaddr);
9330 
9331 			if (npages > 1) {
9332 				/*
9333 				 * If there is a big mapping, make sure
9334 				 * 8K mapping is consistent with the big
9335 				 * mapping.
9336 				 */
9337 				bcolor = i % ncolors;
9338 				if (color != bcolor) {
9339 					return (0);
9340 				}
9341 			}
9342 			if (!clr_valid) {
9343 				clr_valid = 1;
9344 				color1 = color;
9345 			}
9346 
9347 			if (color1 != color) {
9348 				return (0);
9349 			}
9350 		}
9351 
9352 		pp = PP_PAGENEXT(pp);
9353 	}
9354 
9355 	return (1);
9356 }
9357 
9358 void
9359 sfmmu_page_cache_array(page_t *pp, int flags, int cache_flush_flag,
9360     pgcnt_t npages)
9361 {
9362 	kmutex_t *pmtx;
9363 	int i, ncolors, bcolor;
9364 	kpm_hlk_t *kpmp;
9365 	cpuset_t cpuset;
9366 
9367 	ASSERT(pp != NULL);
9368 	ASSERT(!(cache & CACHE_WRITEBACK));
9369 
9370 	kpmp = sfmmu_kpm_kpmp_enter(pp, npages);
9371 	pmtx = sfmmu_page_enter(pp);
9372 
9373 	/*
9374 	 * Fast path caching single unmapped page
9375 	 */
9376 	if (npages == 1 && !PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp) &&
9377 	    flags == HAT_CACHE) {
9378 		PP_CLRTNC(pp);
9379 		PP_CLRPNC(pp);
9380 		sfmmu_page_exit(pmtx);
9381 		sfmmu_kpm_kpmp_exit(kpmp);
9382 		return;
9383 	}
9384 
9385 	/*
9386 	 * We need to capture all cpus in order to change cacheability
9387 	 * because we can't allow one cpu to access the same physical
9388 	 * page using a cacheable and a non-cachebale mapping at the same
9389 	 * time. Since we may end up walking the ism mapping list
9390 	 * have to grab it's lock now since we can't after all the
9391 	 * cpus have been captured.
9392 	 */
9393 	sfmmu_hat_lock_all();
9394 	mutex_enter(&ism_mlist_lock);
9395 	kpreempt_disable();
9396 	cpuset = cpu_ready_set;
9397 	xc_attention(cpuset);
9398 
9399 	if (npages > 1) {
9400 		/*
9401 		 * Make sure all colors are flushed since the
9402 		 * sfmmu_page_cache() only flushes one color-
9403 		 * it does not know big pages.
9404 		 */
9405 		ncolors = CACHE_NUM_COLOR;
9406 		if (flags & HAT_TMPNC) {
9407 			for (i = 0; i < ncolors; i++) {
9408 				sfmmu_cache_flushcolor(i, pp->p_pagenum);
9409 			}
9410 			cache_flush_flag = CACHE_NO_FLUSH;
9411 		}
9412 	}
9413 
9414 	for (i = 0; i < npages; i++) {
9415 
9416 		ASSERT(sfmmu_mlist_held(pp));
9417 
9418 		if (!(flags == HAT_TMPNC && PP_ISTNC(pp))) {
9419 
9420 			if (npages > 1) {
9421 				bcolor = i % ncolors;
9422 			} else {
9423 				bcolor = NO_VCOLOR;
9424 			}
9425 
9426 			sfmmu_page_cache(pp, flags, cache_flush_flag,
9427 			    bcolor);
9428 		}
9429 
9430 		pp = PP_PAGENEXT(pp);
9431 	}
9432 
9433 	xt_sync(cpuset);
9434 	xc_dismissed(cpuset);
9435 	mutex_exit(&ism_mlist_lock);
9436 	sfmmu_hat_unlock_all();
9437 	sfmmu_page_exit(pmtx);
9438 	sfmmu_kpm_kpmp_exit(kpmp);
9439 	kpreempt_enable();
9440 }
9441 
9442 /*
9443  * This function changes the virtual cacheability of all mappings to a
9444  * particular page.  When changing from uncache to cacheable the mappings will
9445  * only be changed if all of them have the same virtual color.
9446  * We need to flush the cache in all cpus.  It is possible that
9447  * a process referenced a page as cacheable but has sinced exited
9448  * and cleared the mapping list.  We still to flush it but have no
9449  * state so all cpus is the only alternative.
9450  */
9451 static void
9452 sfmmu_page_cache(page_t *pp, int flags, int cache_flush_flag, int bcolor)
9453 {
9454 	struct	sf_hment *sfhme;
9455 	struct	hme_blk *hmeblkp;
9456 	sfmmu_t *sfmmup;
9457 	tte_t	tte, ttemod;
9458 	caddr_t	vaddr;
9459 	int	ret, color;
9460 	pfn_t	pfn;
9461 
9462 	color = bcolor;
9463 	pfn = pp->p_pagenum;
9464 
9465 	for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9466 
9467 		if (IS_PAHME(sfhme))
9468 			continue;
9469 		hmeblkp = sfmmu_hmetohblk(sfhme);
9470 
9471 		sfmmu_copytte(&sfhme->hme_tte, &tte);
9472 		ASSERT(TTE_IS_VALID(&tte));
9473 		vaddr = tte_to_vaddr(hmeblkp, tte);
9474 		color = addr_to_vcolor(vaddr);
9475 
9476 #ifdef DEBUG
9477 		if ((flags & HAT_CACHE) && bcolor != NO_VCOLOR) {
9478 			ASSERT(color == bcolor);
9479 		}
9480 #endif
9481 
9482 		ASSERT(flags != HAT_TMPNC || color == PP_GET_VCOLOR(pp));
9483 
9484 		ttemod = tte;
9485 		if (flags & (HAT_UNCACHE | HAT_TMPNC)) {
9486 			TTE_CLR_VCACHEABLE(&ttemod);
9487 		} else {	/* flags & HAT_CACHE */
9488 			TTE_SET_VCACHEABLE(&ttemod);
9489 		}
9490 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
9491 		if (ret < 0) {
9492 			/*
9493 			 * Since all cpus are captured modifytte should not
9494 			 * fail.
9495 			 */
9496 			panic("sfmmu_page_cache: write to tte failed");
9497 		}
9498 
9499 		sfmmup = hblktosfmmu(hmeblkp);
9500 		if (cache_flush_flag == CACHE_FLUSH) {
9501 			/*
9502 			 * Flush TSBs, TLBs and caches
9503 			 */
9504 			if (hmeblkp->hblk_shared) {
9505 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9506 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9507 				sf_region_t *rgnp;
9508 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9509 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9510 				ASSERT(srdp != NULL);
9511 				rgnp = srdp->srd_hmergnp[rid];
9512 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9513 				    srdp, rgnp, rid);
9514 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9515 				    hmeblkp, 0);
9516 				sfmmu_cache_flush(pfn, addr_to_vcolor(vaddr));
9517 			} else if (sfmmup->sfmmu_ismhat) {
9518 				if (flags & HAT_CACHE) {
9519 					SFMMU_STAT(sf_ism_recache);
9520 				} else {
9521 					SFMMU_STAT(sf_ism_uncache);
9522 				}
9523 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9524 				    pfn, CACHE_FLUSH);
9525 			} else {
9526 				sfmmu_tlbcache_demap(vaddr, sfmmup, hmeblkp,
9527 				    pfn, 0, FLUSH_ALL_CPUS, CACHE_FLUSH, 1);
9528 			}
9529 
9530 			/*
9531 			 * all cache entries belonging to this pfn are
9532 			 * now flushed.
9533 			 */
9534 			cache_flush_flag = CACHE_NO_FLUSH;
9535 		} else {
9536 			/*
9537 			 * Flush only TSBs and TLBs.
9538 			 */
9539 			if (hmeblkp->hblk_shared) {
9540 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9541 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9542 				sf_region_t *rgnp;
9543 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9544 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9545 				ASSERT(srdp != NULL);
9546 				rgnp = srdp->srd_hmergnp[rid];
9547 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9548 				    srdp, rgnp, rid);
9549 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9550 				    hmeblkp, 0);
9551 			} else if (sfmmup->sfmmu_ismhat) {
9552 				if (flags & HAT_CACHE) {
9553 					SFMMU_STAT(sf_ism_recache);
9554 				} else {
9555 					SFMMU_STAT(sf_ism_uncache);
9556 				}
9557 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9558 				    pfn, CACHE_NO_FLUSH);
9559 			} else {
9560 				sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 1);
9561 			}
9562 		}
9563 	}
9564 
9565 	if (PP_ISMAPPED_KPM(pp))
9566 		sfmmu_kpm_page_cache(pp, flags, cache_flush_flag);
9567 
9568 	switch (flags) {
9569 
9570 		default:
9571 			panic("sfmmu_pagecache: unknown flags");
9572 			break;
9573 
9574 		case HAT_CACHE:
9575 			PP_CLRTNC(pp);
9576 			PP_CLRPNC(pp);
9577 			PP_SET_VCOLOR(pp, color);
9578 			break;
9579 
9580 		case HAT_TMPNC:
9581 			PP_SETTNC(pp);
9582 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9583 			break;
9584 
9585 		case HAT_UNCACHE:
9586 			PP_SETPNC(pp);
9587 			PP_CLRTNC(pp);
9588 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9589 			break;
9590 	}
9591 }
9592 #endif	/* VAC */
9593 
9594 
9595 /*
9596  * Wrapper routine used to return a context.
9597  *
9598  * It's the responsibility of the caller to guarantee that the
9599  * process serializes on calls here by taking the HAT lock for
9600  * the hat.
9601  *
9602  */
9603 static void
9604 sfmmu_get_ctx(sfmmu_t *sfmmup)
9605 {
9606 	mmu_ctx_t *mmu_ctxp;
9607 	uint_t pstate_save;
9608 	int ret;
9609 
9610 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9611 	ASSERT(sfmmup != ksfmmup);
9612 
9613 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID)) {
9614 		sfmmu_setup_tsbinfo(sfmmup);
9615 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ALLCTX_INVALID);
9616 	}
9617 
9618 	kpreempt_disable();
9619 
9620 	mmu_ctxp = CPU_MMU_CTXP(CPU);
9621 	ASSERT(mmu_ctxp);
9622 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
9623 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
9624 
9625 	/*
9626 	 * Do a wrap-around if cnum reaches the max # cnum supported by a MMU.
9627 	 */
9628 	if (mmu_ctxp->mmu_cnum == mmu_ctxp->mmu_nctxs)
9629 		sfmmu_ctx_wrap_around(mmu_ctxp, B_TRUE);
9630 
9631 	/*
9632 	 * Let the MMU set up the page sizes to use for
9633 	 * this context in the TLB. Don't program 2nd dtlb for ism hat.
9634 	 */
9635 	if ((&mmu_set_ctx_page_sizes) && (sfmmup->sfmmu_ismhat == 0)) {
9636 		mmu_set_ctx_page_sizes(sfmmup);
9637 	}
9638 
9639 	/*
9640 	 * sfmmu_alloc_ctx and sfmmu_load_mmustate will be performed with
9641 	 * interrupts disabled to prevent race condition with wrap-around
9642 	 * ctx invalidatation. In sun4v, ctx invalidation also involves
9643 	 * a HV call to set the number of TSBs to 0. If interrupts are not
9644 	 * disabled until after sfmmu_load_mmustate is complete TSBs may
9645 	 * become assigned to INVALID_CONTEXT. This is not allowed.
9646 	 */
9647 	pstate_save = sfmmu_disable_intrs();
9648 
9649 	if (sfmmu_alloc_ctx(sfmmup, 1, CPU, SFMMU_PRIVATE) &&
9650 	    sfmmup->sfmmu_scdp != NULL) {
9651 		sf_scd_t *scdp = sfmmup->sfmmu_scdp;
9652 		sfmmu_t *scsfmmup = scdp->scd_sfmmup;
9653 		ret = sfmmu_alloc_ctx(scsfmmup, 1, CPU, SFMMU_SHARED);
9654 		/* debug purpose only */
9655 		ASSERT(!ret || scsfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
9656 		    != INVALID_CONTEXT);
9657 	}
9658 	sfmmu_load_mmustate(sfmmup);
9659 
9660 	sfmmu_enable_intrs(pstate_save);
9661 
9662 	kpreempt_enable();
9663 }
9664 
9665 /*
9666  * When all cnums are used up in a MMU, cnum will wrap around to the
9667  * next generation and start from 2.
9668  */
9669 static void
9670 sfmmu_ctx_wrap_around(mmu_ctx_t *mmu_ctxp, boolean_t reset_cnum)
9671 {
9672 
9673 	/* caller must have disabled the preemption */
9674 	ASSERT(curthread->t_preempt >= 1);
9675 	ASSERT(mmu_ctxp != NULL);
9676 
9677 	/* acquire Per-MMU (PM) spin lock */
9678 	mutex_enter(&mmu_ctxp->mmu_lock);
9679 
9680 	/* re-check to see if wrap-around is needed */
9681 	if (mmu_ctxp->mmu_cnum < mmu_ctxp->mmu_nctxs)
9682 		goto done;
9683 
9684 	SFMMU_MMU_STAT(mmu_wrap_around);
9685 
9686 	/* update gnum */
9687 	ASSERT(mmu_ctxp->mmu_gnum != 0);
9688 	mmu_ctxp->mmu_gnum++;
9689 	if (mmu_ctxp->mmu_gnum == 0 ||
9690 	    mmu_ctxp->mmu_gnum > MAX_SFMMU_GNUM_VAL) {
9691 		cmn_err(CE_PANIC, "mmu_gnum of mmu_ctx 0x%p is out of bound.",
9692 		    (void *)mmu_ctxp);
9693 	}
9694 
9695 	if (mmu_ctxp->mmu_ncpus > 1) {
9696 		cpuset_t cpuset;
9697 
9698 		membar_enter(); /* make sure updated gnum visible */
9699 
9700 		SFMMU_XCALL_STATS(NULL);
9701 
9702 		/* xcall to others on the same MMU to invalidate ctx */
9703 		cpuset = mmu_ctxp->mmu_cpuset;
9704 		ASSERT(CPU_IN_SET(cpuset, CPU->cpu_id) || !reset_cnum);
9705 		CPUSET_DEL(cpuset, CPU->cpu_id);
9706 		CPUSET_AND(cpuset, cpu_ready_set);
9707 
9708 		/*
9709 		 * Pass in INVALID_CONTEXT as the first parameter to
9710 		 * sfmmu_raise_tsb_exception, which invalidates the context
9711 		 * of any process running on the CPUs in the MMU.
9712 		 */
9713 		xt_some(cpuset, sfmmu_raise_tsb_exception,
9714 		    INVALID_CONTEXT, INVALID_CONTEXT);
9715 		xt_sync(cpuset);
9716 
9717 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
9718 	}
9719 
9720 	if (sfmmu_getctx_sec() != INVALID_CONTEXT) {
9721 		sfmmu_setctx_sec(INVALID_CONTEXT);
9722 		sfmmu_clear_utsbinfo();
9723 	}
9724 
9725 	/*
9726 	 * No xcall is needed here. For sun4u systems all CPUs in context
9727 	 * domain share a single physical MMU therefore it's enough to flush
9728 	 * TLB on local CPU. On sun4v systems we use 1 global context
9729 	 * domain and flush all remote TLBs in sfmmu_raise_tsb_exception
9730 	 * handler. Note that vtag_flushall_uctxs() is called
9731 	 * for Ultra II machine, where the equivalent flushall functionality
9732 	 * is implemented in SW, and only user ctx TLB entries are flushed.
9733 	 */
9734 	if (&vtag_flushall_uctxs != NULL) {
9735 		vtag_flushall_uctxs();
9736 	} else {
9737 		vtag_flushall();
9738 	}
9739 
9740 	/* reset mmu cnum, skips cnum 0 and 1 */
9741 	if (reset_cnum == B_TRUE)
9742 		mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
9743 
9744 done:
9745 	mutex_exit(&mmu_ctxp->mmu_lock);
9746 }
9747 
9748 
9749 /*
9750  * For multi-threaded process, set the process context to INVALID_CONTEXT
9751  * so that it faults and reloads the MMU state from TL=0. For single-threaded
9752  * process, we can just load the MMU state directly without having to
9753  * set context invalid. Caller must hold the hat lock since we don't
9754  * acquire it here.
9755  */
9756 static void
9757 sfmmu_sync_mmustate(sfmmu_t *sfmmup)
9758 {
9759 	uint_t cnum;
9760 	uint_t pstate_save;
9761 
9762 	ASSERT(sfmmup != ksfmmup);
9763 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9764 
9765 	kpreempt_disable();
9766 
9767 	/*
9768 	 * We check whether the pass'ed-in sfmmup is the same as the
9769 	 * current running proc. This is to makes sure the current proc
9770 	 * stays single-threaded if it already is.
9771 	 */
9772 	if ((sfmmup == curthread->t_procp->p_as->a_hat) &&
9773 	    (curthread->t_procp->p_lwpcnt == 1)) {
9774 		/* single-thread */
9775 		cnum = sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum;
9776 		if (cnum != INVALID_CONTEXT) {
9777 			uint_t curcnum;
9778 			/*
9779 			 * Disable interrupts to prevent race condition
9780 			 * with sfmmu_ctx_wrap_around ctx invalidation.
9781 			 * In sun4v, ctx invalidation involves setting
9782 			 * TSB to NULL, hence, interrupts should be disabled
9783 			 * untill after sfmmu_load_mmustate is completed.
9784 			 */
9785 			pstate_save = sfmmu_disable_intrs();
9786 			curcnum = sfmmu_getctx_sec();
9787 			if (curcnum == cnum)
9788 				sfmmu_load_mmustate(sfmmup);
9789 			sfmmu_enable_intrs(pstate_save);
9790 			ASSERT(curcnum == cnum || curcnum == INVALID_CONTEXT);
9791 		}
9792 	} else {
9793 		/*
9794 		 * multi-thread
9795 		 * or when sfmmup is not the same as the curproc.
9796 		 */
9797 		sfmmu_invalidate_ctx(sfmmup);
9798 	}
9799 
9800 	kpreempt_enable();
9801 }
9802 
9803 
9804 /*
9805  * Replace the specified TSB with a new TSB.  This function gets called when
9806  * we grow, shrink or swapin a TSB.  When swapping in a TSB (TSB_SWAPIN), the
9807  * TSB_FORCEALLOC flag may be used to force allocation of a minimum-sized TSB
9808  * (8K).
9809  *
9810  * Caller must hold the HAT lock, but should assume any tsb_info
9811  * pointers it has are no longer valid after calling this function.
9812  *
9813  * Return values:
9814  *	TSB_ALLOCFAIL	Failed to allocate a TSB, due to memory constraints
9815  *	TSB_LOSTRACE	HAT is busy, i.e. another thread is already doing
9816  *			something to this tsbinfo/TSB
9817  *	TSB_SUCCESS	Operation succeeded
9818  */
9819 static tsb_replace_rc_t
9820 sfmmu_replace_tsb(sfmmu_t *sfmmup, struct tsb_info *old_tsbinfo, uint_t szc,
9821     hatlock_t *hatlockp, uint_t flags)
9822 {
9823 	struct tsb_info *new_tsbinfo = NULL;
9824 	struct tsb_info *curtsb, *prevtsb;
9825 	uint_t tte_sz_mask;
9826 	int i;
9827 
9828 	ASSERT(sfmmup != ksfmmup);
9829 	ASSERT(sfmmup->sfmmu_ismhat == 0);
9830 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9831 	ASSERT(szc <= tsb_max_growsize);
9832 
9833 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_BUSY))
9834 		return (TSB_LOSTRACE);
9835 
9836 	/*
9837 	 * Find the tsb_info ahead of this one in the list, and
9838 	 * also make sure that the tsb_info passed in really
9839 	 * exists!
9840 	 */
9841 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
9842 	    curtsb != old_tsbinfo && curtsb != NULL;
9843 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
9844 		;
9845 	ASSERT(curtsb != NULL);
9846 
9847 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
9848 		/*
9849 		 * The process is swapped out, so just set the new size
9850 		 * code.  When it swaps back in, we'll allocate a new one
9851 		 * of the new chosen size.
9852 		 */
9853 		curtsb->tsb_szc = szc;
9854 		return (TSB_SUCCESS);
9855 	}
9856 	SFMMU_FLAGS_SET(sfmmup, HAT_BUSY);
9857 
9858 	tte_sz_mask = old_tsbinfo->tsb_ttesz_mask;
9859 
9860 	/*
9861 	 * All initialization is done inside of sfmmu_tsbinfo_alloc().
9862 	 * If we fail to allocate a TSB, exit.
9863 	 *
9864 	 * If tsb grows with new tsb size > 4M and old tsb size < 4M,
9865 	 * then try 4M slab after the initial alloc fails.
9866 	 *
9867 	 * If tsb swapin with tsb size > 4M, then try 4M after the
9868 	 * initial alloc fails.
9869 	 */
9870 	sfmmu_hat_exit(hatlockp);
9871 	if (sfmmu_tsbinfo_alloc(&new_tsbinfo, szc,
9872 	    tte_sz_mask, flags, sfmmup) &&
9873 	    (!(flags & (TSB_GROW | TSB_SWAPIN)) || (szc <= TSB_4M_SZCODE) ||
9874 	    (!(flags & TSB_SWAPIN) &&
9875 	    (old_tsbinfo->tsb_szc >= TSB_4M_SZCODE)) ||
9876 	    sfmmu_tsbinfo_alloc(&new_tsbinfo, TSB_4M_SZCODE,
9877 	    tte_sz_mask, flags, sfmmup))) {
9878 		(void) sfmmu_hat_enter(sfmmup);
9879 		if (!(flags & TSB_SWAPIN))
9880 			SFMMU_STAT(sf_tsb_resize_failures);
9881 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
9882 		return (TSB_ALLOCFAIL);
9883 	}
9884 	(void) sfmmu_hat_enter(sfmmup);
9885 
9886 	/*
9887 	 * Re-check to make sure somebody else didn't muck with us while we
9888 	 * didn't hold the HAT lock.  If the process swapped out, fine, just
9889 	 * exit; this can happen if we try to shrink the TSB from the context
9890 	 * of another process (such as on an ISM unmap), though it is rare.
9891 	 */
9892 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
9893 		SFMMU_STAT(sf_tsb_resize_failures);
9894 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
9895 		sfmmu_hat_exit(hatlockp);
9896 		sfmmu_tsbinfo_free(new_tsbinfo);
9897 		(void) sfmmu_hat_enter(sfmmup);
9898 		return (TSB_LOSTRACE);
9899 	}
9900 
9901 #ifdef	DEBUG
9902 	/* Reverify that the tsb_info still exists.. for debugging only */
9903 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
9904 	    curtsb != old_tsbinfo && curtsb != NULL;
9905 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
9906 		;
9907 	ASSERT(curtsb != NULL);
9908 #endif	/* DEBUG */
9909 
9910 	/*
9911 	 * Quiesce any CPUs running this process on their next TLB miss
9912 	 * so they atomically see the new tsb_info.  We temporarily set the
9913 	 * context to invalid context so new threads that come on processor
9914 	 * after we do the xcall to cpusran will also serialize behind the
9915 	 * HAT lock on TLB miss and will see the new TSB.  Since this short
9916 	 * race with a new thread coming on processor is relatively rare,
9917 	 * this synchronization mechanism should be cheaper than always
9918 	 * pausing all CPUs for the duration of the setup, which is what
9919 	 * the old implementation did.  This is particuarly true if we are
9920 	 * copying a huge chunk of memory around during that window.
9921 	 *
9922 	 * The memory barriers are to make sure things stay consistent
9923 	 * with resume() since it does not hold the HAT lock while
9924 	 * walking the list of tsb_info structures.
9925 	 */
9926 	if ((flags & TSB_SWAPIN) != TSB_SWAPIN) {
9927 		/* The TSB is either growing or shrinking. */
9928 		sfmmu_invalidate_ctx(sfmmup);
9929 	} else {
9930 		/*
9931 		 * It is illegal to swap in TSBs from a process other
9932 		 * than a process being swapped in.  This in turn
9933 		 * implies we do not have a valid MMU context here
9934 		 * since a process needs one to resolve translation
9935 		 * misses.
9936 		 */
9937 		ASSERT(curthread->t_procp->p_as->a_hat == sfmmup);
9938 	}
9939 
9940 #ifdef DEBUG
9941 	ASSERT(max_mmu_ctxdoms > 0);
9942 
9943 	/*
9944 	 * Process should have INVALID_CONTEXT on all MMUs
9945 	 */
9946 	for (i = 0; i < max_mmu_ctxdoms; i++) {
9947 
9948 		ASSERT(sfmmup->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
9949 	}
9950 #endif
9951 
9952 	new_tsbinfo->tsb_next = old_tsbinfo->tsb_next;
9953 	membar_stst();	/* strict ordering required */
9954 	if (prevtsb)
9955 		prevtsb->tsb_next = new_tsbinfo;
9956 	else
9957 		sfmmup->sfmmu_tsb = new_tsbinfo;
9958 	membar_enter();	/* make sure new TSB globally visible */
9959 
9960 	/*
9961 	 * We need to migrate TSB entries from the old TSB to the new TSB
9962 	 * if tsb_remap_ttes is set and the TSB is growing.
9963 	 */
9964 	if (tsb_remap_ttes && ((flags & TSB_GROW) == TSB_GROW))
9965 		sfmmu_copy_tsb(old_tsbinfo, new_tsbinfo);
9966 
9967 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
9968 
9969 	/*
9970 	 * Drop the HAT lock to free our old tsb_info.
9971 	 */
9972 	sfmmu_hat_exit(hatlockp);
9973 
9974 	if ((flags & TSB_GROW) == TSB_GROW) {
9975 		SFMMU_STAT(sf_tsb_grow);
9976 	} else if ((flags & TSB_SHRINK) == TSB_SHRINK) {
9977 		SFMMU_STAT(sf_tsb_shrink);
9978 	}
9979 
9980 	sfmmu_tsbinfo_free(old_tsbinfo);
9981 
9982 	(void) sfmmu_hat_enter(sfmmup);
9983 	return (TSB_SUCCESS);
9984 }
9985 
9986 /*
9987  * This function will re-program hat pgsz array, and invalidate the
9988  * process' context, forcing the process to switch to another
9989  * context on the next TLB miss, and therefore start using the
9990  * TLB that is reprogrammed for the new page sizes.
9991  */
9992 void
9993 sfmmu_reprog_pgsz_arr(sfmmu_t *sfmmup, uint8_t *tmp_pgsz)
9994 {
9995 	int i;
9996 	hatlock_t *hatlockp = NULL;
9997 
9998 	hatlockp = sfmmu_hat_enter(sfmmup);
9999 	/* USIII+-IV+ optimization, requires hat lock */
10000 	if (tmp_pgsz) {
10001 		for (i = 0; i < mmu_page_sizes; i++)
10002 			sfmmup->sfmmu_pgsz[i] = tmp_pgsz[i];
10003 	}
10004 	SFMMU_STAT(sf_tlb_reprog_pgsz);
10005 
10006 	sfmmu_invalidate_ctx(sfmmup);
10007 
10008 	sfmmu_hat_exit(hatlockp);
10009 }
10010 
10011 /*
10012  * The scd_rttecnt field in the SCD must be updated to take account of the
10013  * regions which it contains.
10014  */
10015 static void
10016 sfmmu_set_scd_rttecnt(sf_srd_t *srdp, sf_scd_t *scdp)
10017 {
10018 	uint_t rid;
10019 	uint_t i, j;
10020 	ulong_t w;
10021 	sf_region_t *rgnp;
10022 
10023 	ASSERT(srdp != NULL);
10024 
10025 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
10026 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
10027 			continue;
10028 		}
10029 
10030 		j = 0;
10031 		while (w) {
10032 			if (!(w & 0x1)) {
10033 				j++;
10034 				w >>= 1;
10035 				continue;
10036 			}
10037 			rid = (i << BT_ULSHIFT) | j;
10038 			j++;
10039 			w >>= 1;
10040 
10041 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
10042 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
10043 			rgnp = srdp->srd_hmergnp[rid];
10044 			ASSERT(rgnp->rgn_refcnt > 0);
10045 			ASSERT(rgnp->rgn_id == rid);
10046 
10047 			scdp->scd_rttecnt[rgnp->rgn_pgszc] +=
10048 			    rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
10049 
10050 			/*
10051 			 * Maintain the tsb0 inflation cnt for the regions
10052 			 * in the SCD.
10053 			 */
10054 			if (rgnp->rgn_pgszc >= TTE4M) {
10055 				scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt +=
10056 				    rgnp->rgn_size >>
10057 				    (TTE_PAGE_SHIFT(TTE8K) + 2);
10058 			}
10059 		}
10060 	}
10061 }
10062 
10063 /*
10064  * This function assumes that there are either four or six supported page
10065  * sizes and at most two programmable TLBs, so we need to decide which
10066  * page sizes are most important and then tell the MMU layer so it
10067  * can adjust the TLB page sizes accordingly (if supported).
10068  *
10069  * If these assumptions change, this function will need to be
10070  * updated to support whatever the new limits are.
10071  *
10072  * The growing flag is nonzero if we are growing the address space,
10073  * and zero if it is shrinking.  This allows us to decide whether
10074  * to grow or shrink our TSB, depending upon available memory
10075  * conditions.
10076  */
10077 static void
10078 sfmmu_check_page_sizes(sfmmu_t *sfmmup, int growing)
10079 {
10080 	uint64_t ttecnt[MMU_PAGE_SIZES];
10081 	uint64_t tte8k_cnt, tte4m_cnt;
10082 	uint8_t i;
10083 	int sectsb_thresh;
10084 
10085 	/*
10086 	 * Kernel threads, processes with small address spaces not using
10087 	 * large pages, and dummy ISM HATs need not apply.
10088 	 */
10089 	if (sfmmup == ksfmmup || sfmmup->sfmmu_ismhat != NULL)
10090 		return;
10091 
10092 	if (!SFMMU_LGPGS_INUSE(sfmmup) &&
10093 	    sfmmup->sfmmu_ttecnt[TTE8K] <= tsb_rss_factor)
10094 		return;
10095 
10096 	for (i = 0; i < mmu_page_sizes; i++) {
10097 		ttecnt[i] = sfmmup->sfmmu_ttecnt[i] +
10098 		    sfmmup->sfmmu_ismttecnt[i];
10099 	}
10100 
10101 	/* Check pagesizes in use, and possibly reprogram DTLB. */
10102 	if (&mmu_check_page_sizes)
10103 		mmu_check_page_sizes(sfmmup, ttecnt);
10104 
10105 	/*
10106 	 * Calculate the number of 8k ttes to represent the span of these
10107 	 * pages.
10108 	 */
10109 	tte8k_cnt = ttecnt[TTE8K] +
10110 	    (ttecnt[TTE64K] << (MMU_PAGESHIFT64K - MMU_PAGESHIFT)) +
10111 	    (ttecnt[TTE512K] << (MMU_PAGESHIFT512K - MMU_PAGESHIFT));
10112 	if (mmu_page_sizes == max_mmu_page_sizes) {
10113 		tte4m_cnt = ttecnt[TTE4M] +
10114 		    (ttecnt[TTE32M] << (MMU_PAGESHIFT32M - MMU_PAGESHIFT4M)) +
10115 		    (ttecnt[TTE256M] << (MMU_PAGESHIFT256M - MMU_PAGESHIFT4M));
10116 	} else {
10117 		tte4m_cnt = ttecnt[TTE4M];
10118 	}
10119 
10120 	/*
10121 	 * Inflate tte8k_cnt to allow for region large page allocation failure.
10122 	 */
10123 	tte8k_cnt += sfmmup->sfmmu_tsb0_4minflcnt;
10124 
10125 	/*
10126 	 * Inflate TSB sizes by a factor of 2 if this process
10127 	 * uses 4M text pages to minimize extra conflict misses
10128 	 * in the first TSB since without counting text pages
10129 	 * 8K TSB may become too small.
10130 	 *
10131 	 * Also double the size of the second TSB to minimize
10132 	 * extra conflict misses due to competition between 4M text pages
10133 	 * and data pages.
10134 	 *
10135 	 * We need to adjust the second TSB allocation threshold by the
10136 	 * inflation factor, since there is no point in creating a second
10137 	 * TSB when we know all the mappings can fit in the I/D TLBs.
10138 	 */
10139 	sectsb_thresh = tsb_sectsb_threshold;
10140 	if (sfmmup->sfmmu_flags & HAT_4MTEXT_FLAG) {
10141 		tte8k_cnt <<= 1;
10142 		tte4m_cnt <<= 1;
10143 		sectsb_thresh <<= 1;
10144 	}
10145 
10146 	/*
10147 	 * Check to see if our TSB is the right size; we may need to
10148 	 * grow or shrink it.  If the process is small, our work is
10149 	 * finished at this point.
10150 	 */
10151 	if (tte8k_cnt <= tsb_rss_factor && tte4m_cnt <= sectsb_thresh) {
10152 		return;
10153 	}
10154 	sfmmu_size_tsb(sfmmup, growing, tte8k_cnt, tte4m_cnt, sectsb_thresh);
10155 }
10156 
10157 static void
10158 sfmmu_size_tsb(sfmmu_t *sfmmup, int growing, uint64_t tte8k_cnt,
10159     uint64_t tte4m_cnt, int sectsb_thresh)
10160 {
10161 	int tsb_bits;
10162 	uint_t tsb_szc;
10163 	struct tsb_info *tsbinfop;
10164 	hatlock_t *hatlockp = NULL;
10165 
10166 	hatlockp = sfmmu_hat_enter(sfmmup);
10167 	ASSERT(hatlockp != NULL);
10168 	tsbinfop = sfmmup->sfmmu_tsb;
10169 	ASSERT(tsbinfop != NULL);
10170 
10171 	/*
10172 	 * If we're growing, select the size based on RSS.  If we're
10173 	 * shrinking, leave some room so we don't have to turn around and
10174 	 * grow again immediately.
10175 	 */
10176 	if (growing)
10177 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
10178 	else
10179 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt << 1);
10180 
10181 	if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10182 	    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10183 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10184 		    hatlockp, TSB_SHRINK);
10185 	} else if (growing && tsb_szc > tsbinfop->tsb_szc && TSB_OK_GROW()) {
10186 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10187 		    hatlockp, TSB_GROW);
10188 	}
10189 	tsbinfop = sfmmup->sfmmu_tsb;
10190 
10191 	/*
10192 	 * With the TLB and first TSB out of the way, we need to see if
10193 	 * we need a second TSB for 4M pages.  If we managed to reprogram
10194 	 * the TLB page sizes above, the process will start using this new
10195 	 * TSB right away; otherwise, it will start using it on the next
10196 	 * context switch.  Either way, it's no big deal so there's no
10197 	 * synchronization with the trap handlers here unless we grow the
10198 	 * TSB (in which case it's required to prevent using the old one
10199 	 * after it's freed). Note: second tsb is required for 32M/256M
10200 	 * page sizes.
10201 	 */
10202 	if (tte4m_cnt > sectsb_thresh) {
10203 		/*
10204 		 * If we're growing, select the size based on RSS.  If we're
10205 		 * shrinking, leave some room so we don't have to turn
10206 		 * around and grow again immediately.
10207 		 */
10208 		if (growing)
10209 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
10210 		else
10211 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt << 1);
10212 		if (tsbinfop->tsb_next == NULL) {
10213 			struct tsb_info *newtsb;
10214 			int allocflags = SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)?
10215 			    0 : TSB_ALLOC;
10216 
10217 			sfmmu_hat_exit(hatlockp);
10218 
10219 			/*
10220 			 * Try to allocate a TSB for 4[32|256]M pages.  If we
10221 			 * can't get the size we want, retry w/a minimum sized
10222 			 * TSB.  If that still didn't work, give up; we can
10223 			 * still run without one.
10224 			 */
10225 			tsb_bits = (mmu_page_sizes == max_mmu_page_sizes)?
10226 			    TSB4M|TSB32M|TSB256M:TSB4M;
10227 			if ((sfmmu_tsbinfo_alloc(&newtsb, tsb_szc, tsb_bits,
10228 			    allocflags, sfmmup)) &&
10229 			    (tsb_szc <= TSB_4M_SZCODE ||
10230 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
10231 			    tsb_bits, allocflags, sfmmup)) &&
10232 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_MIN_SZCODE,
10233 			    tsb_bits, allocflags, sfmmup)) {
10234 				return;
10235 			}
10236 
10237 			hatlockp = sfmmu_hat_enter(sfmmup);
10238 
10239 			sfmmu_invalidate_ctx(sfmmup);
10240 
10241 			if (sfmmup->sfmmu_tsb->tsb_next == NULL) {
10242 				sfmmup->sfmmu_tsb->tsb_next = newtsb;
10243 				SFMMU_STAT(sf_tsb_sectsb_create);
10244 				sfmmu_hat_exit(hatlockp);
10245 				return;
10246 			} else {
10247 				/*
10248 				 * It's annoying, but possible for us
10249 				 * to get here.. we dropped the HAT lock
10250 				 * because of locking order in the kmem
10251 				 * allocator, and while we were off getting
10252 				 * our memory, some other thread decided to
10253 				 * do us a favor and won the race to get a
10254 				 * second TSB for this process.  Sigh.
10255 				 */
10256 				sfmmu_hat_exit(hatlockp);
10257 				sfmmu_tsbinfo_free(newtsb);
10258 				return;
10259 			}
10260 		}
10261 
10262 		/*
10263 		 * We have a second TSB, see if it's big enough.
10264 		 */
10265 		tsbinfop = tsbinfop->tsb_next;
10266 
10267 		/*
10268 		 * Check to see if our second TSB is the right size;
10269 		 * we may need to grow or shrink it.
10270 		 * To prevent thrashing (e.g. growing the TSB on a
10271 		 * subsequent map operation), only try to shrink if
10272 		 * the TSB reach exceeds twice the virtual address
10273 		 * space size.
10274 		 */
10275 		if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10276 		    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10277 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10278 			    tsb_szc, hatlockp, TSB_SHRINK);
10279 		} else if (growing && tsb_szc > tsbinfop->tsb_szc &&
10280 		    TSB_OK_GROW()) {
10281 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10282 			    tsb_szc, hatlockp, TSB_GROW);
10283 		}
10284 	}
10285 
10286 	sfmmu_hat_exit(hatlockp);
10287 }
10288 
10289 /*
10290  * Free up a sfmmu
10291  * Since the sfmmu is currently embedded in the hat struct we simply zero
10292  * out our fields and free up the ism map blk list if any.
10293  */
10294 static void
10295 sfmmu_free_sfmmu(sfmmu_t *sfmmup)
10296 {
10297 	ism_blk_t	*blkp, *nx_blkp;
10298 #ifdef	DEBUG
10299 	ism_map_t	*map;
10300 	int		i;
10301 #endif
10302 
10303 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
10304 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
10305 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
10306 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
10307 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
10308 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
10309 	ASSERT(SF_RGNMAP_ISNULL(sfmmup));
10310 
10311 	sfmmup->sfmmu_free = 0;
10312 	sfmmup->sfmmu_ismhat = 0;
10313 
10314 	blkp = sfmmup->sfmmu_iblk;
10315 	sfmmup->sfmmu_iblk = NULL;
10316 
10317 	while (blkp) {
10318 #ifdef	DEBUG
10319 		map = blkp->iblk_maps;
10320 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
10321 			ASSERT(map[i].imap_seg == 0);
10322 			ASSERT(map[i].imap_ismhat == NULL);
10323 			ASSERT(map[i].imap_ment == NULL);
10324 		}
10325 #endif
10326 		nx_blkp = blkp->iblk_next;
10327 		blkp->iblk_next = NULL;
10328 		blkp->iblk_nextpa = (uint64_t)-1;
10329 		kmem_cache_free(ism_blk_cache, blkp);
10330 		blkp = nx_blkp;
10331 	}
10332 }
10333 
10334 /*
10335  * Locking primitves accessed by HATLOCK macros
10336  */
10337 
10338 #define	SFMMU_SPL_MTX	(0x0)
10339 #define	SFMMU_ML_MTX	(0x1)
10340 
10341 #define	SFMMU_MLSPL_MTX(type, pg)	(((type) == SFMMU_SPL_MTX) ? \
10342 					    SPL_HASH(pg) : MLIST_HASH(pg))
10343 
10344 kmutex_t *
10345 sfmmu_page_enter(struct page *pp)
10346 {
10347 	return (sfmmu_mlspl_enter(pp, SFMMU_SPL_MTX));
10348 }
10349 
10350 void
10351 sfmmu_page_exit(kmutex_t *spl)
10352 {
10353 	mutex_exit(spl);
10354 }
10355 
10356 int
10357 sfmmu_page_spl_held(struct page *pp)
10358 {
10359 	return (sfmmu_mlspl_held(pp, SFMMU_SPL_MTX));
10360 }
10361 
10362 kmutex_t *
10363 sfmmu_mlist_enter(struct page *pp)
10364 {
10365 	return (sfmmu_mlspl_enter(pp, SFMMU_ML_MTX));
10366 }
10367 
10368 void
10369 sfmmu_mlist_exit(kmutex_t *mml)
10370 {
10371 	mutex_exit(mml);
10372 }
10373 
10374 int
10375 sfmmu_mlist_held(struct page *pp)
10376 {
10377 
10378 	return (sfmmu_mlspl_held(pp, SFMMU_ML_MTX));
10379 }
10380 
10381 /*
10382  * Common code for sfmmu_mlist_enter() and sfmmu_page_enter().  For
10383  * sfmmu_mlist_enter() case mml_table lock array is used and for
10384  * sfmmu_page_enter() sfmmu_page_lock lock array is used.
10385  *
10386  * The lock is taken on a root page so that it protects an operation on all
10387  * constituent pages of a large page pp belongs to.
10388  *
10389  * The routine takes a lock from the appropriate array. The lock is determined
10390  * by hashing the root page. After taking the lock this routine checks if the
10391  * root page has the same size code that was used to determine the root (i.e
10392  * that root hasn't changed).  If root page has the expected p_szc field we
10393  * have the right lock and it's returned to the caller. If root's p_szc
10394  * decreased we release the lock and retry from the beginning.  This case can
10395  * happen due to hat_page_demote() decreasing p_szc between our load of p_szc
10396  * value and taking the lock. The number of retries due to p_szc decrease is
10397  * limited by the maximum p_szc value. If p_szc is 0 we return the lock
10398  * determined by hashing pp itself.
10399  *
10400  * If our caller doesn't hold a SE_SHARED or SE_EXCL lock on pp it's also
10401  * possible that p_szc can increase. To increase p_szc a thread has to lock
10402  * all constituent pages EXCL and do hat_pageunload() on all of them. All the
10403  * callers that don't hold a page locked recheck if hmeblk through which pp
10404  * was found still maps this pp.  If it doesn't map it anymore returned lock
10405  * is immediately dropped. Therefore if sfmmu_mlspl_enter() hits the case of
10406  * p_szc increase after taking the lock it returns this lock without further
10407  * retries because in this case the caller doesn't care about which lock was
10408  * taken. The caller will drop it right away.
10409  *
10410  * After the routine returns it's guaranteed that hat_page_demote() can't
10411  * change p_szc field of any of constituent pages of a large page pp belongs
10412  * to as long as pp was either locked at least SHARED prior to this call or
10413  * the caller finds that hment that pointed to this pp still references this
10414  * pp (this also assumes that the caller holds hme hash bucket lock so that
10415  * the same pp can't be remapped into the same hmeblk after it was unmapped by
10416  * hat_pageunload()).
10417  */
10418 static kmutex_t *
10419 sfmmu_mlspl_enter(struct page *pp, int type)
10420 {
10421 	kmutex_t	*mtx;
10422 	uint_t		prev_rszc = UINT_MAX;
10423 	page_t		*rootpp;
10424 	uint_t		szc;
10425 	uint_t		rszc;
10426 	uint_t		pszc = pp->p_szc;
10427 
10428 	ASSERT(pp != NULL);
10429 
10430 again:
10431 	if (pszc == 0) {
10432 		mtx = SFMMU_MLSPL_MTX(type, pp);
10433 		mutex_enter(mtx);
10434 		return (mtx);
10435 	}
10436 
10437 	/* The lock lives in the root page */
10438 	rootpp = PP_GROUPLEADER(pp, pszc);
10439 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10440 	mutex_enter(mtx);
10441 
10442 	/*
10443 	 * Return mml in the following 3 cases:
10444 	 *
10445 	 * 1) If pp itself is root since if its p_szc decreased before we took
10446 	 * the lock pp is still the root of smaller szc page. And if its p_szc
10447 	 * increased it doesn't matter what lock we return (see comment in
10448 	 * front of this routine).
10449 	 *
10450 	 * 2) If pp's not root but rootpp is the root of a rootpp->p_szc size
10451 	 * large page we have the right lock since any previous potential
10452 	 * hat_page_demote() is done demoting from greater than current root's
10453 	 * p_szc because hat_page_demote() changes root's p_szc last. No
10454 	 * further hat_page_demote() can start or be in progress since it
10455 	 * would need the same lock we currently hold.
10456 	 *
10457 	 * 3) If rootpp's p_szc increased since previous iteration it doesn't
10458 	 * matter what lock we return (see comment in front of this routine).
10459 	 */
10460 	if (pp == rootpp || (rszc = rootpp->p_szc) == pszc ||
10461 	    rszc >= prev_rszc) {
10462 		return (mtx);
10463 	}
10464 
10465 	/*
10466 	 * hat_page_demote() could have decreased root's p_szc.
10467 	 * In this case pp's p_szc must also be smaller than pszc.
10468 	 * Retry.
10469 	 */
10470 	if (rszc < pszc) {
10471 		szc = pp->p_szc;
10472 		if (szc < pszc) {
10473 			mutex_exit(mtx);
10474 			pszc = szc;
10475 			goto again;
10476 		}
10477 		/*
10478 		 * pp's p_szc increased after it was decreased.
10479 		 * page cannot be mapped. Return current lock. The caller
10480 		 * will drop it right away.
10481 		 */
10482 		return (mtx);
10483 	}
10484 
10485 	/*
10486 	 * root's p_szc is greater than pp's p_szc.
10487 	 * hat_page_demote() is not done with all pages
10488 	 * yet. Wait for it to complete.
10489 	 */
10490 	mutex_exit(mtx);
10491 	rootpp = PP_GROUPLEADER(rootpp, rszc);
10492 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10493 	mutex_enter(mtx);
10494 	mutex_exit(mtx);
10495 	prev_rszc = rszc;
10496 	goto again;
10497 }
10498 
10499 static int
10500 sfmmu_mlspl_held(struct page *pp, int type)
10501 {
10502 	kmutex_t	*mtx;
10503 
10504 	ASSERT(pp != NULL);
10505 	/* The lock lives in the root page */
10506 	pp = PP_PAGEROOT(pp);
10507 	ASSERT(pp != NULL);
10508 
10509 	mtx = SFMMU_MLSPL_MTX(type, pp);
10510 	return (MUTEX_HELD(mtx));
10511 }
10512 
10513 static uint_t
10514 sfmmu_get_free_hblk(struct hme_blk **hmeblkpp, uint_t critical)
10515 {
10516 	struct  hme_blk *hblkp;
10517 
10518 
10519 	if (freehblkp != NULL) {
10520 		mutex_enter(&freehblkp_lock);
10521 		if (freehblkp != NULL) {
10522 			/*
10523 			 * If the current thread is owning hblk_reserve OR
10524 			 * critical request from sfmmu_hblk_steal()
10525 			 * let it succeed even if freehblkcnt is really low.
10526 			 */
10527 			if (freehblkcnt <= HBLK_RESERVE_MIN && !critical) {
10528 				SFMMU_STAT(sf_get_free_throttle);
10529 				mutex_exit(&freehblkp_lock);
10530 				return (0);
10531 			}
10532 			freehblkcnt--;
10533 			*hmeblkpp = freehblkp;
10534 			hblkp = *hmeblkpp;
10535 			freehblkp = hblkp->hblk_next;
10536 			mutex_exit(&freehblkp_lock);
10537 			hblkp->hblk_next = NULL;
10538 			SFMMU_STAT(sf_get_free_success);
10539 
10540 			ASSERT(hblkp->hblk_hmecnt == 0);
10541 			ASSERT(hblkp->hblk_vcnt == 0);
10542 			ASSERT(hblkp->hblk_nextpa == va_to_pa((caddr_t)hblkp));
10543 
10544 			return (1);
10545 		}
10546 		mutex_exit(&freehblkp_lock);
10547 	}
10548 
10549 	/* Check cpu hblk pending queues */
10550 	if ((*hmeblkpp = sfmmu_check_pending_hblks(TTE8K)) != NULL) {
10551 		hblkp = *hmeblkpp;
10552 		hblkp->hblk_next = NULL;
10553 		hblkp->hblk_nextpa = va_to_pa((caddr_t)hblkp);
10554 
10555 		ASSERT(hblkp->hblk_hmecnt == 0);
10556 		ASSERT(hblkp->hblk_vcnt == 0);
10557 
10558 		return (1);
10559 	}
10560 
10561 	SFMMU_STAT(sf_get_free_fail);
10562 	return (0);
10563 }
10564 
10565 static uint_t
10566 sfmmu_put_free_hblk(struct hme_blk *hmeblkp, uint_t critical)
10567 {
10568 	struct  hme_blk *hblkp;
10569 
10570 	ASSERT(hmeblkp->hblk_hmecnt == 0);
10571 	ASSERT(hmeblkp->hblk_vcnt == 0);
10572 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
10573 
10574 	/*
10575 	 * If the current thread is mapping into kernel space,
10576 	 * let it succede even if freehblkcnt is max
10577 	 * so that it will avoid freeing it to kmem.
10578 	 * This will prevent stack overflow due to
10579 	 * possible recursion since kmem_cache_free()
10580 	 * might require creation of a slab which
10581 	 * in turn needs an hmeblk to map that slab;
10582 	 * let's break this vicious chain at the first
10583 	 * opportunity.
10584 	 */
10585 	if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10586 		mutex_enter(&freehblkp_lock);
10587 		if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10588 			SFMMU_STAT(sf_put_free_success);
10589 			freehblkcnt++;
10590 			hmeblkp->hblk_next = freehblkp;
10591 			freehblkp = hmeblkp;
10592 			mutex_exit(&freehblkp_lock);
10593 			return (1);
10594 		}
10595 		mutex_exit(&freehblkp_lock);
10596 	}
10597 
10598 	/*
10599 	 * Bring down freehblkcnt to HBLK_RESERVE_CNT. We are here
10600 	 * only if freehblkcnt is at least HBLK_RESERVE_CNT *and*
10601 	 * we are not in the process of mapping into kernel space.
10602 	 */
10603 	ASSERT(!critical);
10604 	while (freehblkcnt > HBLK_RESERVE_CNT) {
10605 		mutex_enter(&freehblkp_lock);
10606 		if (freehblkcnt > HBLK_RESERVE_CNT) {
10607 			freehblkcnt--;
10608 			hblkp = freehblkp;
10609 			freehblkp = hblkp->hblk_next;
10610 			mutex_exit(&freehblkp_lock);
10611 			ASSERT(get_hblk_cache(hblkp) == sfmmu8_cache);
10612 			kmem_cache_free(sfmmu8_cache, hblkp);
10613 			continue;
10614 		}
10615 		mutex_exit(&freehblkp_lock);
10616 	}
10617 	SFMMU_STAT(sf_put_free_fail);
10618 	return (0);
10619 }
10620 
10621 static void
10622 sfmmu_hblk_swap(struct hme_blk *new)
10623 {
10624 	struct hme_blk *old, *hblkp, *prev;
10625 	uint64_t newpa;
10626 	caddr_t	base, vaddr, endaddr;
10627 	struct hmehash_bucket *hmebp;
10628 	struct sf_hment *osfhme, *nsfhme;
10629 	page_t *pp;
10630 	kmutex_t *pml;
10631 	tte_t tte;
10632 	struct hme_blk *list = NULL;
10633 
10634 #ifdef	DEBUG
10635 	hmeblk_tag		hblktag;
10636 	struct hme_blk		*found;
10637 #endif
10638 	old = HBLK_RESERVE;
10639 	ASSERT(!old->hblk_shared);
10640 
10641 	/*
10642 	 * save pa before bcopy clobbers it
10643 	 */
10644 	newpa = new->hblk_nextpa;
10645 
10646 	base = (caddr_t)get_hblk_base(old);
10647 	endaddr = base + get_hblk_span(old);
10648 
10649 	/*
10650 	 * acquire hash bucket lock.
10651 	 */
10652 	hmebp = sfmmu_tteload_acquire_hashbucket(ksfmmup, base, TTE8K,
10653 	    SFMMU_INVALID_SHMERID);
10654 
10655 	/*
10656 	 * copy contents from old to new
10657 	 */
10658 	bcopy((void *)old, (void *)new, HME8BLK_SZ);
10659 
10660 	/*
10661 	 * add new to hash chain
10662 	 */
10663 	sfmmu_hblk_hash_add(hmebp, new, newpa);
10664 
10665 	/*
10666 	 * search hash chain for hblk_reserve; this needs to be performed
10667 	 * after adding new, otherwise prev won't correspond to the hblk which
10668 	 * is prior to old in hash chain when we call sfmmu_hblk_hash_rm to
10669 	 * remove old later.
10670 	 */
10671 	for (prev = NULL,
10672 	    hblkp = hmebp->hmeblkp; hblkp != NULL && hblkp != old;
10673 	    prev = hblkp, hblkp = hblkp->hblk_next)
10674 		;
10675 
10676 	if (hblkp != old)
10677 		panic("sfmmu_hblk_swap: hblk_reserve not found");
10678 
10679 	/*
10680 	 * p_mapping list is still pointing to hments in hblk_reserve;
10681 	 * fix up p_mapping list so that they point to hments in new.
10682 	 *
10683 	 * Since all these mappings are created by hblk_reserve_thread
10684 	 * on the way and it's using at least one of the buffers from each of
10685 	 * the newly minted slabs, there is no danger of any of these
10686 	 * mappings getting unloaded by another thread.
10687 	 *
10688 	 * tsbmiss could only modify ref/mod bits of hments in old/new.
10689 	 * Since all of these hments hold mappings established by segkmem
10690 	 * and mappings in segkmem are setup with HAT_NOSYNC, ref/mod bits
10691 	 * have no meaning for the mappings in hblk_reserve.  hments in
10692 	 * old and new are identical except for ref/mod bits.
10693 	 */
10694 	for (vaddr = base; vaddr < endaddr; vaddr += TTEBYTES(TTE8K)) {
10695 
10696 		HBLKTOHME(osfhme, old, vaddr);
10697 		sfmmu_copytte(&osfhme->hme_tte, &tte);
10698 
10699 		if (TTE_IS_VALID(&tte)) {
10700 			if ((pp = osfhme->hme_page) == NULL)
10701 				panic("sfmmu_hblk_swap: page not mapped");
10702 
10703 			pml = sfmmu_mlist_enter(pp);
10704 
10705 			if (pp != osfhme->hme_page)
10706 				panic("sfmmu_hblk_swap: mapping changed");
10707 
10708 			HBLKTOHME(nsfhme, new, vaddr);
10709 
10710 			HME_ADD(nsfhme, pp);
10711 			HME_SUB(osfhme, pp);
10712 
10713 			sfmmu_mlist_exit(pml);
10714 		}
10715 	}
10716 
10717 	/*
10718 	 * remove old from hash chain
10719 	 */
10720 	sfmmu_hblk_hash_rm(hmebp, old, prev, &list, 1);
10721 
10722 #ifdef	DEBUG
10723 
10724 	hblktag.htag_id = ksfmmup;
10725 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
10726 	hblktag.htag_bspage = HME_HASH_BSPAGE(base, HME_HASH_SHIFT(TTE8K));
10727 	hblktag.htag_rehash = HME_HASH_REHASH(TTE8K);
10728 	HME_HASH_FAST_SEARCH(hmebp, hblktag, found);
10729 
10730 	if (found != new)
10731 		panic("sfmmu_hblk_swap: new hblk not found");
10732 #endif
10733 
10734 	SFMMU_HASH_UNLOCK(hmebp);
10735 
10736 	/*
10737 	 * Reset hblk_reserve
10738 	 */
10739 	bzero((void *)old, HME8BLK_SZ);
10740 	old->hblk_nextpa = va_to_pa((caddr_t)old);
10741 }
10742 
10743 /*
10744  * Grab the mlist mutex for both pages passed in.
10745  *
10746  * low and high will be returned as pointers to the mutexes for these pages.
10747  * low refers to the mutex residing in the lower bin of the mlist hash, while
10748  * high refers to the mutex residing in the higher bin of the mlist hash.  This
10749  * is due to the locking order restrictions on the same thread grabbing
10750  * multiple mlist mutexes.  The low lock must be acquired before the high lock.
10751  *
10752  * If both pages hash to the same mutex, only grab that single mutex, and
10753  * high will be returned as NULL
10754  * If the pages hash to different bins in the hash, grab the lower addressed
10755  * lock first and then the higher addressed lock in order to follow the locking
10756  * rules involved with the same thread grabbing multiple mlist mutexes.
10757  * low and high will both have non-NULL values.
10758  */
10759 static void
10760 sfmmu_mlist_reloc_enter(struct page *targ, struct page *repl,
10761     kmutex_t **low, kmutex_t **high)
10762 {
10763 	kmutex_t	*mml_targ, *mml_repl;
10764 
10765 	/*
10766 	 * no need to do the dance around szc as in sfmmu_mlist_enter()
10767 	 * because this routine is only called by hat_page_relocate() and all
10768 	 * targ and repl pages are already locked EXCL so szc can't change.
10769 	 */
10770 
10771 	mml_targ = MLIST_HASH(PP_PAGEROOT(targ));
10772 	mml_repl = MLIST_HASH(PP_PAGEROOT(repl));
10773 
10774 	if (mml_targ == mml_repl) {
10775 		*low = mml_targ;
10776 		*high = NULL;
10777 	} else {
10778 		if (mml_targ < mml_repl) {
10779 			*low = mml_targ;
10780 			*high = mml_repl;
10781 		} else {
10782 			*low = mml_repl;
10783 			*high = mml_targ;
10784 		}
10785 	}
10786 
10787 	mutex_enter(*low);
10788 	if (*high)
10789 		mutex_enter(*high);
10790 }
10791 
10792 static void
10793 sfmmu_mlist_reloc_exit(kmutex_t *low, kmutex_t *high)
10794 {
10795 	if (high)
10796 		mutex_exit(high);
10797 	mutex_exit(low);
10798 }
10799 
10800 static hatlock_t *
10801 sfmmu_hat_enter(sfmmu_t *sfmmup)
10802 {
10803 	hatlock_t	*hatlockp;
10804 
10805 	if (sfmmup != ksfmmup) {
10806 		hatlockp = TSB_HASH(sfmmup);
10807 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
10808 		return (hatlockp);
10809 	}
10810 	return (NULL);
10811 }
10812 
10813 static hatlock_t *
10814 sfmmu_hat_tryenter(sfmmu_t *sfmmup)
10815 {
10816 	hatlock_t	*hatlockp;
10817 
10818 	if (sfmmup != ksfmmup) {
10819 		hatlockp = TSB_HASH(sfmmup);
10820 		if (mutex_tryenter(HATLOCK_MUTEXP(hatlockp)) == 0)
10821 			return (NULL);
10822 		return (hatlockp);
10823 	}
10824 	return (NULL);
10825 }
10826 
10827 static void
10828 sfmmu_hat_exit(hatlock_t *hatlockp)
10829 {
10830 	if (hatlockp != NULL)
10831 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
10832 }
10833 
10834 static void
10835 sfmmu_hat_lock_all(void)
10836 {
10837 	int i;
10838 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
10839 		mutex_enter(HATLOCK_MUTEXP(&hat_lock[i]));
10840 }
10841 
10842 static void
10843 sfmmu_hat_unlock_all(void)
10844 {
10845 	int i;
10846 	for (i = SFMMU_NUM_LOCK - 1; i >= 0; i--)
10847 		mutex_exit(HATLOCK_MUTEXP(&hat_lock[i]));
10848 }
10849 
10850 int
10851 sfmmu_hat_lock_held(sfmmu_t *sfmmup)
10852 {
10853 	ASSERT(sfmmup != ksfmmup);
10854 	return (MUTEX_HELD(HATLOCK_MUTEXP(TSB_HASH(sfmmup))));
10855 }
10856 
10857 /*
10858  * Locking primitives to provide consistency between ISM unmap
10859  * and other operations.  Since ISM unmap can take a long time, we
10860  * use HAT_ISMBUSY flag (protected by the hatlock) to avoid creating
10861  * contention on the hatlock buckets while ISM segments are being
10862  * unmapped.  The tradeoff is that the flags don't prevent priority
10863  * inversion from occurring, so we must request kernel priority in
10864  * case we have to sleep to keep from getting buried while holding
10865  * the HAT_ISMBUSY flag set, which in turn could block other kernel
10866  * threads from running (for example, in sfmmu_uvatopfn()).
10867  */
10868 static void
10869 sfmmu_ismhat_enter(sfmmu_t *sfmmup, int hatlock_held)
10870 {
10871 	hatlock_t *hatlockp;
10872 
10873 	if (!hatlock_held)
10874 		hatlockp = sfmmu_hat_enter(sfmmup);
10875 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY))
10876 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
10877 	SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
10878 	if (!hatlock_held)
10879 		sfmmu_hat_exit(hatlockp);
10880 }
10881 
10882 static void
10883 sfmmu_ismhat_exit(sfmmu_t *sfmmup, int hatlock_held)
10884 {
10885 	hatlock_t *hatlockp;
10886 
10887 	if (!hatlock_held)
10888 		hatlockp = sfmmu_hat_enter(sfmmup);
10889 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
10890 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
10891 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
10892 	if (!hatlock_held)
10893 		sfmmu_hat_exit(hatlockp);
10894 }
10895 
10896 /*
10897  *
10898  * Algorithm:
10899  *
10900  * (1) if segkmem is not ready, allocate hblk from an array of pre-alloc'ed
10901  *	hblks.
10902  *
10903  * (2) if we are allocating an hblk for mapping a slab in sfmmu_cache,
10904  *
10905  *		(a) try to return an hblk from reserve pool of free hblks;
10906  *		(b) if the reserve pool is empty, acquire hblk_reserve_lock
10907  *		    and return hblk_reserve.
10908  *
10909  * (3) call kmem_cache_alloc() to allocate hblk;
10910  *
10911  *		(a) if hblk_reserve_lock is held by the current thread,
10912  *		    atomically replace hblk_reserve by the hblk that is
10913  *		    returned by kmem_cache_alloc; release hblk_reserve_lock
10914  *		    and call kmem_cache_alloc() again.
10915  *		(b) if reserve pool is not full, add the hblk that is
10916  *		    returned by kmem_cache_alloc to reserve pool and
10917  *		    call kmem_cache_alloc again.
10918  *
10919  */
10920 static struct hme_blk *
10921 sfmmu_hblk_alloc(sfmmu_t *sfmmup, caddr_t vaddr,
10922     struct hmehash_bucket *hmebp, uint_t size, hmeblk_tag hblktag,
10923     uint_t flags, uint_t rid)
10924 {
10925 	struct hme_blk *hmeblkp = NULL;
10926 	struct hme_blk *newhblkp;
10927 	struct hme_blk *shw_hblkp = NULL;
10928 	struct kmem_cache *sfmmu_cache = NULL;
10929 	uint64_t hblkpa;
10930 	ulong_t index;
10931 	uint_t owner;		/* set to 1 if using hblk_reserve */
10932 	uint_t forcefree;
10933 	int sleep;
10934 	sf_srd_t *srdp;
10935 	sf_region_t *rgnp;
10936 
10937 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
10938 	ASSERT(hblktag.htag_rid == rid);
10939 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
10940 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
10941 	    IS_P2ALIGNED(vaddr, TTEBYTES(size)));
10942 
10943 	/*
10944 	 * If segkmem is not created yet, allocate from static hmeblks
10945 	 * created at the end of startup_modules().  See the block comment
10946 	 * in startup_modules() describing how we estimate the number of
10947 	 * static hmeblks that will be needed during re-map.
10948 	 */
10949 	if (!hblk_alloc_dynamic) {
10950 
10951 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
10952 
10953 		if (size == TTE8K) {
10954 			index = nucleus_hblk8.index;
10955 			if (index >= nucleus_hblk8.len) {
10956 				/*
10957 				 * If we panic here, see startup_modules() to
10958 				 * make sure that we are calculating the
10959 				 * number of hblk8's that we need correctly.
10960 				 */
10961 				prom_panic("no nucleus hblk8 to allocate");
10962 			}
10963 			hmeblkp =
10964 			    (struct hme_blk *)&nucleus_hblk8.list[index];
10965 			nucleus_hblk8.index++;
10966 			SFMMU_STAT(sf_hblk8_nalloc);
10967 		} else {
10968 			index = nucleus_hblk1.index;
10969 			if (nucleus_hblk1.index >= nucleus_hblk1.len) {
10970 				/*
10971 				 * If we panic here, see startup_modules().
10972 				 * Most likely you need to update the
10973 				 * calculation of the number of hblk1 elements
10974 				 * that the kernel needs to boot.
10975 				 */
10976 				prom_panic("no nucleus hblk1 to allocate");
10977 			}
10978 			hmeblkp =
10979 			    (struct hme_blk *)&nucleus_hblk1.list[index];
10980 			nucleus_hblk1.index++;
10981 			SFMMU_STAT(sf_hblk1_nalloc);
10982 		}
10983 
10984 		goto hblk_init;
10985 	}
10986 
10987 	SFMMU_HASH_UNLOCK(hmebp);
10988 
10989 	if (sfmmup != KHATID && !SFMMU_IS_SHMERID_VALID(rid)) {
10990 		if (mmu_page_sizes == max_mmu_page_sizes) {
10991 			if (size < TTE256M)
10992 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
10993 				    size, flags);
10994 		} else {
10995 			if (size < TTE4M)
10996 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
10997 				    size, flags);
10998 		}
10999 	} else if (SFMMU_IS_SHMERID_VALID(rid)) {
11000 		/*
11001 		 * Shared hmes use per region bitmaps in rgn_hmeflag
11002 		 * rather than shadow hmeblks to keep track of the
11003 		 * mapping sizes which have been allocated for the region.
11004 		 * Here we cleanup old invalid hmeblks with this rid,
11005 		 * which may be left around by pageunload().
11006 		 */
11007 		int ttesz;
11008 		caddr_t va;
11009 		caddr_t	eva = vaddr + TTEBYTES(size);
11010 
11011 		ASSERT(sfmmup != KHATID);
11012 
11013 		srdp = sfmmup->sfmmu_srdp;
11014 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11015 		rgnp = srdp->srd_hmergnp[rid];
11016 		ASSERT(rgnp != NULL && rgnp->rgn_id == rid);
11017 		ASSERT(rgnp->rgn_refcnt != 0);
11018 		ASSERT(size <= rgnp->rgn_pgszc);
11019 
11020 		ttesz = HBLK_MIN_TTESZ;
11021 		do {
11022 			if (!(rgnp->rgn_hmeflags & (0x1 << ttesz))) {
11023 				continue;
11024 			}
11025 
11026 			if (ttesz > size && ttesz != HBLK_MIN_TTESZ) {
11027 				sfmmu_cleanup_rhblk(srdp, vaddr, rid, ttesz);
11028 			} else if (ttesz < size) {
11029 				for (va = vaddr; va < eva;
11030 				    va += TTEBYTES(ttesz)) {
11031 					sfmmu_cleanup_rhblk(srdp, va, rid,
11032 					    ttesz);
11033 				}
11034 			}
11035 		} while (++ttesz <= rgnp->rgn_pgszc);
11036 	}
11037 
11038 fill_hblk:
11039 	owner = (hblk_reserve_thread == curthread) ? 1 : 0;
11040 
11041 	if (owner && size == TTE8K) {
11042 
11043 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11044 		/*
11045 		 * We are really in a tight spot. We already own
11046 		 * hblk_reserve and we need another hblk.  In anticipation
11047 		 * of this kind of scenario, we specifically set aside
11048 		 * HBLK_RESERVE_MIN number of hblks to be used exclusively
11049 		 * by owner of hblk_reserve.
11050 		 */
11051 		SFMMU_STAT(sf_hblk_recurse_cnt);
11052 
11053 		if (!sfmmu_get_free_hblk(&hmeblkp, 1))
11054 			panic("sfmmu_hblk_alloc: reserve list is empty");
11055 
11056 		goto hblk_verify;
11057 	}
11058 
11059 	ASSERT(!owner);
11060 
11061 	if ((flags & HAT_NO_KALLOC) == 0) {
11062 
11063 		sfmmu_cache = ((size == TTE8K) ? sfmmu8_cache : sfmmu1_cache);
11064 		sleep = ((sfmmup == KHATID) ? KM_NOSLEEP : KM_SLEEP);
11065 
11066 		if ((hmeblkp = kmem_cache_alloc(sfmmu_cache, sleep)) == NULL) {
11067 			hmeblkp = sfmmu_hblk_steal(size);
11068 		} else {
11069 			/*
11070 			 * if we are the owner of hblk_reserve,
11071 			 * swap hblk_reserve with hmeblkp and
11072 			 * start a fresh life.  Hope things go
11073 			 * better this time.
11074 			 */
11075 			if (hblk_reserve_thread == curthread) {
11076 				ASSERT(sfmmu_cache == sfmmu8_cache);
11077 				sfmmu_hblk_swap(hmeblkp);
11078 				hblk_reserve_thread = NULL;
11079 				mutex_exit(&hblk_reserve_lock);
11080 				goto fill_hblk;
11081 			}
11082 			/*
11083 			 * let's donate this hblk to our reserve list if
11084 			 * we are not mapping kernel range
11085 			 */
11086 			if (size == TTE8K && sfmmup != KHATID) {
11087 				if (sfmmu_put_free_hblk(hmeblkp, 0))
11088 					goto fill_hblk;
11089 			}
11090 		}
11091 	} else {
11092 		/*
11093 		 * We are here to map the slab in sfmmu8_cache; let's
11094 		 * check if we could tap our reserve list; if successful,
11095 		 * this will avoid the pain of going thru sfmmu_hblk_swap
11096 		 */
11097 		SFMMU_STAT(sf_hblk_slab_cnt);
11098 		if (!sfmmu_get_free_hblk(&hmeblkp, 0)) {
11099 			/*
11100 			 * let's start hblk_reserve dance
11101 			 */
11102 			SFMMU_STAT(sf_hblk_reserve_cnt);
11103 			owner = 1;
11104 			mutex_enter(&hblk_reserve_lock);
11105 			hmeblkp = HBLK_RESERVE;
11106 			hblk_reserve_thread = curthread;
11107 		}
11108 	}
11109 
11110 hblk_verify:
11111 	ASSERT(hmeblkp != NULL);
11112 	set_hblk_sz(hmeblkp, size);
11113 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
11114 	SFMMU_HASH_LOCK(hmebp);
11115 	HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11116 	if (newhblkp != NULL) {
11117 		SFMMU_HASH_UNLOCK(hmebp);
11118 		if (hmeblkp != HBLK_RESERVE) {
11119 			/*
11120 			 * This is really tricky!
11121 			 *
11122 			 * vmem_alloc(vmem_seg_arena)
11123 			 *  vmem_alloc(vmem_internal_arena)
11124 			 *   segkmem_alloc(heap_arena)
11125 			 *    vmem_alloc(heap_arena)
11126 			 *    page_create()
11127 			 *    hat_memload()
11128 			 *	kmem_cache_free()
11129 			 *	 kmem_cache_alloc()
11130 			 *	  kmem_slab_create()
11131 			 *	   vmem_alloc(kmem_internal_arena)
11132 			 *	    segkmem_alloc(heap_arena)
11133 			 *		vmem_alloc(heap_arena)
11134 			 *		page_create()
11135 			 *		hat_memload()
11136 			 *		  kmem_cache_free()
11137 			 *		...
11138 			 *
11139 			 * Thus, hat_memload() could call kmem_cache_free
11140 			 * for enough number of times that we could easily
11141 			 * hit the bottom of the stack or run out of reserve
11142 			 * list of vmem_seg structs.  So, we must donate
11143 			 * this hblk to reserve list if it's allocated
11144 			 * from sfmmu8_cache *and* mapping kernel range.
11145 			 * We don't need to worry about freeing hmeblk1's
11146 			 * to kmem since they don't map any kmem slabs.
11147 			 *
11148 			 * Note: When segkmem supports largepages, we must
11149 			 * free hmeblk1's to reserve list as well.
11150 			 */
11151 			forcefree = (sfmmup == KHATID) ? 1 : 0;
11152 			if (size == TTE8K &&
11153 			    sfmmu_put_free_hblk(hmeblkp, forcefree)) {
11154 				goto re_verify;
11155 			}
11156 			ASSERT(sfmmup != KHATID);
11157 			kmem_cache_free(get_hblk_cache(hmeblkp), hmeblkp);
11158 		} else {
11159 			/*
11160 			 * Hey! we don't need hblk_reserve any more.
11161 			 */
11162 			ASSERT(owner);
11163 			hblk_reserve_thread = NULL;
11164 			mutex_exit(&hblk_reserve_lock);
11165 			owner = 0;
11166 		}
11167 re_verify:
11168 		/*
11169 		 * let's check if the goodies are still present
11170 		 */
11171 		SFMMU_HASH_LOCK(hmebp);
11172 		HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11173 		if (newhblkp != NULL) {
11174 			/*
11175 			 * return newhblkp if it's not hblk_reserve;
11176 			 * if newhblkp is hblk_reserve, return it
11177 			 * _only if_ we are the owner of hblk_reserve.
11178 			 */
11179 			if (newhblkp != HBLK_RESERVE || owner) {
11180 				ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
11181 				    newhblkp->hblk_shared);
11182 				ASSERT(SFMMU_IS_SHMERID_VALID(rid) ||
11183 				    !newhblkp->hblk_shared);
11184 				return (newhblkp);
11185 			} else {
11186 				/*
11187 				 * we just hit hblk_reserve in the hash and
11188 				 * we are not the owner of that;
11189 				 *
11190 				 * block until hblk_reserve_thread completes
11191 				 * swapping hblk_reserve and try the dance
11192 				 * once again.
11193 				 */
11194 				SFMMU_HASH_UNLOCK(hmebp);
11195 				mutex_enter(&hblk_reserve_lock);
11196 				mutex_exit(&hblk_reserve_lock);
11197 				SFMMU_STAT(sf_hblk_reserve_hit);
11198 				goto fill_hblk;
11199 			}
11200 		} else {
11201 			/*
11202 			 * it's no more! try the dance once again.
11203 			 */
11204 			SFMMU_HASH_UNLOCK(hmebp);
11205 			goto fill_hblk;
11206 		}
11207 	}
11208 
11209 hblk_init:
11210 	if (SFMMU_IS_SHMERID_VALID(rid)) {
11211 		uint16_t tteflag = 0x1 <<
11212 		    ((size < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ : size);
11213 
11214 		if (!(rgnp->rgn_hmeflags & tteflag)) {
11215 			atomic_or_16(&rgnp->rgn_hmeflags, tteflag);
11216 		}
11217 		hmeblkp->hblk_shared = 1;
11218 	} else {
11219 		hmeblkp->hblk_shared = 0;
11220 	}
11221 	set_hblk_sz(hmeblkp, size);
11222 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11223 	hmeblkp->hblk_next = (struct hme_blk *)NULL;
11224 	hmeblkp->hblk_tag = hblktag;
11225 	hmeblkp->hblk_shadow = shw_hblkp;
11226 	hblkpa = hmeblkp->hblk_nextpa;
11227 	hmeblkp->hblk_nextpa = HMEBLK_ENDPA;
11228 
11229 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
11230 	ASSERT(get_hblk_span(hmeblkp) == HMEBLK_SPAN(size));
11231 	ASSERT(hmeblkp->hblk_hmecnt == 0);
11232 	ASSERT(hmeblkp->hblk_vcnt == 0);
11233 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11234 	ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
11235 	sfmmu_hblk_hash_add(hmebp, hmeblkp, hblkpa);
11236 	return (hmeblkp);
11237 }
11238 
11239 /*
11240  * This function cleans up the hme_blk and returns it to the free list.
11241  */
11242 /* ARGSUSED */
11243 static void
11244 sfmmu_hblk_free(struct hme_blk **listp)
11245 {
11246 	struct hme_blk *hmeblkp, *next_hmeblkp;
11247 	int		size;
11248 	uint_t		critical;
11249 	uint64_t	hblkpa;
11250 
11251 	ASSERT(*listp != NULL);
11252 
11253 	hmeblkp = *listp;
11254 	while (hmeblkp != NULL) {
11255 		next_hmeblkp = hmeblkp->hblk_next;
11256 		ASSERT(!hmeblkp->hblk_hmecnt);
11257 		ASSERT(!hmeblkp->hblk_vcnt);
11258 		ASSERT(!hmeblkp->hblk_lckcnt);
11259 		ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
11260 		ASSERT(hmeblkp->hblk_shared == 0);
11261 		ASSERT(hmeblkp->hblk_shw_bit == 0);
11262 		ASSERT(hmeblkp->hblk_shadow == NULL);
11263 
11264 		hblkpa = va_to_pa((caddr_t)hmeblkp);
11265 		ASSERT(hblkpa != (uint64_t)-1);
11266 		critical = (hblktosfmmu(hmeblkp) == KHATID) ? 1 : 0;
11267 
11268 		size = get_hblk_ttesz(hmeblkp);
11269 		hmeblkp->hblk_next = NULL;
11270 		hmeblkp->hblk_nextpa = hblkpa;
11271 
11272 		if (hmeblkp->hblk_nuc_bit == 0) {
11273 
11274 			if (size != TTE8K ||
11275 			    !sfmmu_put_free_hblk(hmeblkp, critical))
11276 				kmem_cache_free(get_hblk_cache(hmeblkp),
11277 				    hmeblkp);
11278 		}
11279 		hmeblkp = next_hmeblkp;
11280 	}
11281 }
11282 
11283 #define	BUCKETS_TO_SEARCH_BEFORE_UNLOAD	30
11284 #define	SFMMU_HBLK_STEAL_THRESHOLD 5
11285 
11286 static uint_t sfmmu_hblk_steal_twice;
11287 static uint_t sfmmu_hblk_steal_count, sfmmu_hblk_steal_unload_count;
11288 
11289 /*
11290  * Steal a hmeblk from user or kernel hme hash lists.
11291  * For 8K tte grab one from reserve pool (freehblkp) before proceeding to
11292  * steal and if we fail to steal after SFMMU_HBLK_STEAL_THRESHOLD attempts
11293  * tap into critical reserve of freehblkp.
11294  * Note: We remain looping in this routine until we find one.
11295  */
11296 static struct hme_blk *
11297 sfmmu_hblk_steal(int size)
11298 {
11299 	static struct hmehash_bucket *uhmehash_steal_hand = NULL;
11300 	struct hmehash_bucket *hmebp;
11301 	struct hme_blk *hmeblkp = NULL, *pr_hblk;
11302 	uint64_t hblkpa;
11303 	int i;
11304 	uint_t loop_cnt = 0, critical;
11305 
11306 	for (;;) {
11307 		/* Check cpu hblk pending queues */
11308 		if ((hmeblkp = sfmmu_check_pending_hblks(size)) != NULL) {
11309 			hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
11310 			ASSERT(hmeblkp->hblk_hmecnt == 0);
11311 			ASSERT(hmeblkp->hblk_vcnt == 0);
11312 			return (hmeblkp);
11313 		}
11314 
11315 		if (size == TTE8K) {
11316 			critical =
11317 			    (++loop_cnt > SFMMU_HBLK_STEAL_THRESHOLD) ? 1 : 0;
11318 			if (sfmmu_get_free_hblk(&hmeblkp, critical))
11319 				return (hmeblkp);
11320 		}
11321 
11322 		hmebp = (uhmehash_steal_hand == NULL) ? uhme_hash :
11323 		    uhmehash_steal_hand;
11324 		ASSERT(hmebp >= uhme_hash && hmebp <= &uhme_hash[UHMEHASH_SZ]);
11325 
11326 		for (i = 0; hmeblkp == NULL && i <= UHMEHASH_SZ +
11327 		    BUCKETS_TO_SEARCH_BEFORE_UNLOAD; i++) {
11328 			SFMMU_HASH_LOCK(hmebp);
11329 			hmeblkp = hmebp->hmeblkp;
11330 			hblkpa = hmebp->hmeh_nextpa;
11331 			pr_hblk = NULL;
11332 			while (hmeblkp) {
11333 				/*
11334 				 * check if it is a hmeblk that is not locked
11335 				 * and not shared. skip shadow hmeblks with
11336 				 * shadow_mask set i.e valid count non zero.
11337 				 */
11338 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11339 				    (hmeblkp->hblk_shw_bit == 0 ||
11340 				    hmeblkp->hblk_vcnt == 0) &&
11341 				    (hmeblkp->hblk_lckcnt == 0)) {
11342 					/*
11343 					 * there is a high probability that we
11344 					 * will find a free one. search some
11345 					 * buckets for a free hmeblk initially
11346 					 * before unloading a valid hmeblk.
11347 					 */
11348 					if ((hmeblkp->hblk_vcnt == 0 &&
11349 					    hmeblkp->hblk_hmecnt == 0) || (i >=
11350 					    BUCKETS_TO_SEARCH_BEFORE_UNLOAD)) {
11351 						if (sfmmu_steal_this_hblk(hmebp,
11352 						    hmeblkp, hblkpa, pr_hblk)) {
11353 							/*
11354 							 * Hblk is unloaded
11355 							 * successfully
11356 							 */
11357 							break;
11358 						}
11359 					}
11360 				}
11361 				pr_hblk = hmeblkp;
11362 				hblkpa = hmeblkp->hblk_nextpa;
11363 				hmeblkp = hmeblkp->hblk_next;
11364 			}
11365 
11366 			SFMMU_HASH_UNLOCK(hmebp);
11367 			if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
11368 				hmebp = uhme_hash;
11369 		}
11370 		uhmehash_steal_hand = hmebp;
11371 
11372 		if (hmeblkp != NULL)
11373 			break;
11374 
11375 		/*
11376 		 * in the worst case, look for a free one in the kernel
11377 		 * hash table.
11378 		 */
11379 		for (i = 0, hmebp = khme_hash; i <= KHMEHASH_SZ; i++) {
11380 			SFMMU_HASH_LOCK(hmebp);
11381 			hmeblkp = hmebp->hmeblkp;
11382 			hblkpa = hmebp->hmeh_nextpa;
11383 			pr_hblk = NULL;
11384 			while (hmeblkp) {
11385 				/*
11386 				 * check if it is free hmeblk
11387 				 */
11388 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11389 				    (hmeblkp->hblk_lckcnt == 0) &&
11390 				    (hmeblkp->hblk_vcnt == 0) &&
11391 				    (hmeblkp->hblk_hmecnt == 0)) {
11392 					if (sfmmu_steal_this_hblk(hmebp,
11393 					    hmeblkp, hblkpa, pr_hblk)) {
11394 						break;
11395 					} else {
11396 						/*
11397 						 * Cannot fail since we have
11398 						 * hash lock.
11399 						 */
11400 						panic("fail to steal?");
11401 					}
11402 				}
11403 
11404 				pr_hblk = hmeblkp;
11405 				hblkpa = hmeblkp->hblk_nextpa;
11406 				hmeblkp = hmeblkp->hblk_next;
11407 			}
11408 
11409 			SFMMU_HASH_UNLOCK(hmebp);
11410 			if (hmebp++ == &khme_hash[KHMEHASH_SZ])
11411 				hmebp = khme_hash;
11412 		}
11413 
11414 		if (hmeblkp != NULL)
11415 			break;
11416 		sfmmu_hblk_steal_twice++;
11417 	}
11418 	return (hmeblkp);
11419 }
11420 
11421 /*
11422  * This routine does real work to prepare a hblk to be "stolen" by
11423  * unloading the mappings, updating shadow counts ....
11424  * It returns 1 if the block is ready to be reused (stolen), or 0
11425  * means the block cannot be stolen yet- pageunload is still working
11426  * on this hblk.
11427  */
11428 static int
11429 sfmmu_steal_this_hblk(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
11430     uint64_t hblkpa, struct hme_blk *pr_hblk)
11431 {
11432 	int shw_size, vshift;
11433 	struct hme_blk *shw_hblkp;
11434 	caddr_t vaddr;
11435 	uint_t shw_mask, newshw_mask;
11436 	struct hme_blk *list = NULL;
11437 
11438 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11439 
11440 	/*
11441 	 * check if the hmeblk is free, unload if necessary
11442 	 */
11443 	if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11444 		sfmmu_t *sfmmup;
11445 		demap_range_t dmr;
11446 
11447 		sfmmup = hblktosfmmu(hmeblkp);
11448 		if (hmeblkp->hblk_shared || sfmmup->sfmmu_ismhat) {
11449 			return (0);
11450 		}
11451 		DEMAP_RANGE_INIT(sfmmup, &dmr);
11452 		(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
11453 		    (caddr_t)get_hblk_base(hmeblkp),
11454 		    get_hblk_endaddr(hmeblkp), &dmr, HAT_UNLOAD);
11455 		DEMAP_RANGE_FLUSH(&dmr);
11456 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11457 			/*
11458 			 * Pageunload is working on the same hblk.
11459 			 */
11460 			return (0);
11461 		}
11462 
11463 		sfmmu_hblk_steal_unload_count++;
11464 	}
11465 
11466 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11467 	ASSERT(hmeblkp->hblk_vcnt == 0 && hmeblkp->hblk_hmecnt == 0);
11468 
11469 	sfmmu_hblk_hash_rm(hmebp, hmeblkp, pr_hblk, &list, 1);
11470 	hmeblkp->hblk_nextpa = hblkpa;
11471 
11472 	shw_hblkp = hmeblkp->hblk_shadow;
11473 	if (shw_hblkp) {
11474 		ASSERT(!hmeblkp->hblk_shared);
11475 		shw_size = get_hblk_ttesz(shw_hblkp);
11476 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
11477 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
11478 		ASSERT(vshift < 8);
11479 		/*
11480 		 * Atomically clear shadow mask bit
11481 		 */
11482 		do {
11483 			shw_mask = shw_hblkp->hblk_shw_mask;
11484 			ASSERT(shw_mask & (1 << vshift));
11485 			newshw_mask = shw_mask & ~(1 << vshift);
11486 			newshw_mask = atomic_cas_32(&shw_hblkp->hblk_shw_mask,
11487 			    shw_mask, newshw_mask);
11488 		} while (newshw_mask != shw_mask);
11489 		hmeblkp->hblk_shadow = NULL;
11490 	}
11491 
11492 	/*
11493 	 * remove shadow bit if we are stealing an unused shadow hmeblk.
11494 	 * sfmmu_hblk_alloc needs it that way, will set shadow bit later if
11495 	 * we are indeed allocating a shadow hmeblk.
11496 	 */
11497 	hmeblkp->hblk_shw_bit = 0;
11498 
11499 	if (hmeblkp->hblk_shared) {
11500 		sf_srd_t	*srdp;
11501 		sf_region_t	*rgnp;
11502 		uint_t		rid;
11503 
11504 		srdp = hblktosrd(hmeblkp);
11505 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11506 		rid = hmeblkp->hblk_tag.htag_rid;
11507 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
11508 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
11509 		rgnp = srdp->srd_hmergnp[rid];
11510 		ASSERT(rgnp != NULL);
11511 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
11512 		hmeblkp->hblk_shared = 0;
11513 	}
11514 
11515 	sfmmu_hblk_steal_count++;
11516 	SFMMU_STAT(sf_steal_count);
11517 
11518 	return (1);
11519 }
11520 
11521 struct hme_blk *
11522 sfmmu_hmetohblk(struct sf_hment *sfhme)
11523 {
11524 	struct hme_blk *hmeblkp;
11525 	struct sf_hment *sfhme0;
11526 	struct hme_blk *hblk_dummy = 0;
11527 
11528 	/*
11529 	 * No dummy sf_hments, please.
11530 	 */
11531 	ASSERT(sfhme->hme_tte.ll != 0);
11532 
11533 	sfhme0 = sfhme - sfhme->hme_tte.tte_hmenum;
11534 	hmeblkp = (struct hme_blk *)((uintptr_t)sfhme0 -
11535 	    (uintptr_t)&hblk_dummy->hblk_hme[0]);
11536 
11537 	return (hmeblkp);
11538 }
11539 
11540 /*
11541  * On swapin, get appropriately sized TSB(s) and clear the HAT_SWAPPED flag.
11542  * If we can't get appropriately sized TSB(s), try for 8K TSB(s) using
11543  * KM_SLEEP allocation.
11544  *
11545  * Return 0 on success, -1 otherwise.
11546  */
11547 static void
11548 sfmmu_tsb_swapin(sfmmu_t *sfmmup, hatlock_t *hatlockp)
11549 {
11550 	struct tsb_info *tsbinfop, *next;
11551 	tsb_replace_rc_t rc;
11552 	boolean_t gotfirst = B_FALSE;
11553 
11554 	ASSERT(sfmmup != ksfmmup);
11555 	ASSERT(sfmmu_hat_lock_held(sfmmup));
11556 
11557 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPIN)) {
11558 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
11559 	}
11560 
11561 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
11562 		SFMMU_FLAGS_SET(sfmmup, HAT_SWAPIN);
11563 	} else {
11564 		return;
11565 	}
11566 
11567 	ASSERT(sfmmup->sfmmu_tsb != NULL);
11568 
11569 	/*
11570 	 * Loop over all tsbinfo's replacing them with ones that actually have
11571 	 * a TSB.  If any of the replacements ever fail, bail out of the loop.
11572 	 */
11573 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; tsbinfop = next) {
11574 		ASSERT(tsbinfop->tsb_flags & TSB_SWAPPED);
11575 		next = tsbinfop->tsb_next;
11576 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, tsbinfop->tsb_szc,
11577 		    hatlockp, TSB_SWAPIN);
11578 		if (rc != TSB_SUCCESS) {
11579 			break;
11580 		}
11581 		gotfirst = B_TRUE;
11582 	}
11583 
11584 	switch (rc) {
11585 	case TSB_SUCCESS:
11586 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11587 		cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11588 		return;
11589 	case TSB_LOSTRACE:
11590 		break;
11591 	case TSB_ALLOCFAIL:
11592 		break;
11593 	default:
11594 		panic("sfmmu_replace_tsb returned unrecognized failure code "
11595 		    "%d", rc);
11596 	}
11597 
11598 	/*
11599 	 * In this case, we failed to get one of our TSBs.  If we failed to
11600 	 * get the first TSB, get one of minimum size (8KB).  Walk the list
11601 	 * and throw away the tsbinfos, starting where the allocation failed;
11602 	 * we can get by with just one TSB as long as we don't leave the
11603 	 * SWAPPED tsbinfo structures lying around.
11604 	 */
11605 	tsbinfop = sfmmup->sfmmu_tsb;
11606 	next = tsbinfop->tsb_next;
11607 	tsbinfop->tsb_next = NULL;
11608 
11609 	sfmmu_hat_exit(hatlockp);
11610 	for (tsbinfop = next; tsbinfop != NULL; tsbinfop = next) {
11611 		next = tsbinfop->tsb_next;
11612 		sfmmu_tsbinfo_free(tsbinfop);
11613 	}
11614 	hatlockp = sfmmu_hat_enter(sfmmup);
11615 
11616 	/*
11617 	 * If we don't have any TSBs, get a single 8K TSB for 8K, 64K and 512K
11618 	 * pages.
11619 	 */
11620 	if (!gotfirst) {
11621 		tsbinfop = sfmmup->sfmmu_tsb;
11622 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, TSB_MIN_SZCODE,
11623 		    hatlockp, TSB_SWAPIN | TSB_FORCEALLOC);
11624 		ASSERT(rc == TSB_SUCCESS);
11625 	}
11626 
11627 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11628 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11629 }
11630 
11631 static int
11632 sfmmu_is_rgnva(sf_srd_t *srdp, caddr_t addr, ulong_t w, ulong_t bmw)
11633 {
11634 	ulong_t bix = 0;
11635 	uint_t rid;
11636 	sf_region_t *rgnp;
11637 
11638 	ASSERT(srdp != NULL);
11639 	ASSERT(srdp->srd_refcnt != 0);
11640 
11641 	w <<= BT_ULSHIFT;
11642 	while (bmw) {
11643 		if (!(bmw & 0x1)) {
11644 			bix++;
11645 			bmw >>= 1;
11646 			continue;
11647 		}
11648 		rid = w | bix;
11649 		rgnp = srdp->srd_hmergnp[rid];
11650 		ASSERT(rgnp->rgn_refcnt > 0);
11651 		ASSERT(rgnp->rgn_id == rid);
11652 		if (addr < rgnp->rgn_saddr ||
11653 		    addr >= (rgnp->rgn_saddr + rgnp->rgn_size)) {
11654 			bix++;
11655 			bmw >>= 1;
11656 		} else {
11657 			return (1);
11658 		}
11659 	}
11660 	return (0);
11661 }
11662 
11663 /*
11664  * Handle exceptions for low level tsb_handler.
11665  *
11666  * There are many scenarios that could land us here:
11667  *
11668  * If the context is invalid we land here. The context can be invalid
11669  * for 3 reasons: 1) we couldn't allocate a new context and now need to
11670  * perform a wrap around operation in order to allocate a new context.
11671  * 2) Context was invalidated to change pagesize programming 3) ISMs or
11672  * TSBs configuration is changeing for this process and we are forced into
11673  * here to do a syncronization operation. If the context is valid we can
11674  * be here from window trap hanlder. In this case just call trap to handle
11675  * the fault.
11676  *
11677  * Note that the process will run in INVALID_CONTEXT before
11678  * faulting into here and subsequently loading the MMU registers
11679  * (including the TSB base register) associated with this process.
11680  * For this reason, the trap handlers must all test for
11681  * INVALID_CONTEXT before attempting to access any registers other
11682  * than the context registers.
11683  */
11684 void
11685 sfmmu_tsbmiss_exception(struct regs *rp, uintptr_t tagaccess, uint_t traptype)
11686 {
11687 	sfmmu_t *sfmmup, *shsfmmup;
11688 	uint_t ctxtype;
11689 	klwp_id_t lwp;
11690 	char lwp_save_state;
11691 	hatlock_t *hatlockp, *shatlockp;
11692 	struct tsb_info *tsbinfop;
11693 	struct tsbmiss *tsbmp;
11694 	sf_scd_t *scdp;
11695 
11696 	SFMMU_STAT(sf_tsb_exceptions);
11697 	SFMMU_MMU_STAT(mmu_tsb_exceptions);
11698 	sfmmup = astosfmmu(curthread->t_procp->p_as);
11699 	/*
11700 	 * note that in sun4u, tagacces register contains ctxnum
11701 	 * while sun4v passes ctxtype in the tagaccess register.
11702 	 */
11703 	ctxtype = tagaccess & TAGACC_CTX_MASK;
11704 
11705 	ASSERT(sfmmup != ksfmmup && ctxtype != KCONTEXT);
11706 	ASSERT(sfmmup->sfmmu_ismhat == 0);
11707 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED) ||
11708 	    ctxtype == INVALID_CONTEXT);
11709 
11710 	if (ctxtype != INVALID_CONTEXT && traptype != T_DATA_PROT) {
11711 		/*
11712 		 * We may land here because shme bitmap and pagesize
11713 		 * flags are updated lazily in tsbmiss area on other cpus.
11714 		 * If we detect here that tsbmiss area is out of sync with
11715 		 * sfmmu update it and retry the trapped instruction.
11716 		 * Otherwise call trap().
11717 		 */
11718 		int ret = 0;
11719 		uchar_t tteflag_mask = (1 << TTE64K) | (1 << TTE8K);
11720 		caddr_t addr = (caddr_t)(tagaccess & TAGACC_VADDR_MASK);
11721 
11722 		/*
11723 		 * Must set lwp state to LWP_SYS before
11724 		 * trying to acquire any adaptive lock
11725 		 */
11726 		lwp = ttolwp(curthread);
11727 		ASSERT(lwp);
11728 		lwp_save_state = lwp->lwp_state;
11729 		lwp->lwp_state = LWP_SYS;
11730 
11731 		hatlockp = sfmmu_hat_enter(sfmmup);
11732 		kpreempt_disable();
11733 		tsbmp = &tsbmiss_area[CPU->cpu_id];
11734 		ASSERT(sfmmup == tsbmp->usfmmup);
11735 		if (((tsbmp->uhat_tteflags ^ sfmmup->sfmmu_tteflags) &
11736 		    ~tteflag_mask) ||
11737 		    ((tsbmp->uhat_rtteflags ^  sfmmup->sfmmu_rtteflags) &
11738 		    ~tteflag_mask)) {
11739 			tsbmp->uhat_tteflags = sfmmup->sfmmu_tteflags;
11740 			tsbmp->uhat_rtteflags = sfmmup->sfmmu_rtteflags;
11741 			ret = 1;
11742 		}
11743 		if (sfmmup->sfmmu_srdp != NULL) {
11744 			ulong_t *sm = sfmmup->sfmmu_hmeregion_map.bitmap;
11745 			ulong_t *tm = tsbmp->shmermap;
11746 			ulong_t i;
11747 			for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
11748 				ulong_t d = tm[i] ^ sm[i];
11749 				if (d) {
11750 					if (d & sm[i]) {
11751 						if (!ret && sfmmu_is_rgnva(
11752 						    sfmmup->sfmmu_srdp,
11753 						    addr, i, d & sm[i])) {
11754 							ret = 1;
11755 						}
11756 					}
11757 					tm[i] = sm[i];
11758 				}
11759 			}
11760 		}
11761 		kpreempt_enable();
11762 		sfmmu_hat_exit(hatlockp);
11763 		lwp->lwp_state = lwp_save_state;
11764 		if (ret) {
11765 			return;
11766 		}
11767 	} else if (ctxtype == INVALID_CONTEXT) {
11768 		/*
11769 		 * First, make sure we come out of here with a valid ctx,
11770 		 * since if we don't get one we'll simply loop on the
11771 		 * faulting instruction.
11772 		 *
11773 		 * If the ISM mappings are changing, the TSB is relocated,
11774 		 * the process is swapped, the process is joining SCD or
11775 		 * leaving SCD or shared regions we serialize behind the
11776 		 * controlling thread with hat lock, sfmmu_flags and
11777 		 * sfmmu_tsb_cv condition variable.
11778 		 */
11779 
11780 		/*
11781 		 * Must set lwp state to LWP_SYS before
11782 		 * trying to acquire any adaptive lock
11783 		 */
11784 		lwp = ttolwp(curthread);
11785 		ASSERT(lwp);
11786 		lwp_save_state = lwp->lwp_state;
11787 		lwp->lwp_state = LWP_SYS;
11788 
11789 		hatlockp = sfmmu_hat_enter(sfmmup);
11790 retry:
11791 		if ((scdp = sfmmup->sfmmu_scdp) != NULL) {
11792 			shsfmmup = scdp->scd_sfmmup;
11793 			ASSERT(shsfmmup != NULL);
11794 
11795 			for (tsbinfop = shsfmmup->sfmmu_tsb; tsbinfop != NULL;
11796 			    tsbinfop = tsbinfop->tsb_next) {
11797 				if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
11798 					/* drop the private hat lock */
11799 					sfmmu_hat_exit(hatlockp);
11800 					/* acquire the shared hat lock */
11801 					shatlockp = sfmmu_hat_enter(shsfmmup);
11802 					/*
11803 					 * recheck to see if anything changed
11804 					 * after we drop the private hat lock.
11805 					 */
11806 					if (sfmmup->sfmmu_scdp == scdp &&
11807 					    shsfmmup == scdp->scd_sfmmup) {
11808 						sfmmu_tsb_chk_reloc(shsfmmup,
11809 						    shatlockp);
11810 					}
11811 					sfmmu_hat_exit(shatlockp);
11812 					hatlockp = sfmmu_hat_enter(sfmmup);
11813 					goto retry;
11814 				}
11815 			}
11816 		}
11817 
11818 		for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
11819 		    tsbinfop = tsbinfop->tsb_next) {
11820 			if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
11821 				cv_wait(&sfmmup->sfmmu_tsb_cv,
11822 				    HATLOCK_MUTEXP(hatlockp));
11823 				goto retry;
11824 			}
11825 		}
11826 
11827 		/*
11828 		 * Wait for ISM maps to be updated.
11829 		 */
11830 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
11831 			cv_wait(&sfmmup->sfmmu_tsb_cv,
11832 			    HATLOCK_MUTEXP(hatlockp));
11833 			goto retry;
11834 		}
11835 
11836 		/* Is this process joining an SCD? */
11837 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
11838 			/*
11839 			 * Flush private TSB and setup shared TSB.
11840 			 * sfmmu_finish_join_scd() does not drop the
11841 			 * hat lock.
11842 			 */
11843 			sfmmu_finish_join_scd(sfmmup);
11844 			SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
11845 		}
11846 
11847 		/*
11848 		 * If we're swapping in, get TSB(s).  Note that we must do
11849 		 * this before we get a ctx or load the MMU state.  Once
11850 		 * we swap in we have to recheck to make sure the TSB(s) and
11851 		 * ISM mappings didn't change while we slept.
11852 		 */
11853 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
11854 			sfmmu_tsb_swapin(sfmmup, hatlockp);
11855 			goto retry;
11856 		}
11857 
11858 		sfmmu_get_ctx(sfmmup);
11859 
11860 		sfmmu_hat_exit(hatlockp);
11861 		/*
11862 		 * Must restore lwp_state if not calling
11863 		 * trap() for further processing. Restore
11864 		 * it anyway.
11865 		 */
11866 		lwp->lwp_state = lwp_save_state;
11867 		return;
11868 	}
11869 	trap(rp, (caddr_t)tagaccess, traptype, 0);
11870 }
11871 
11872 static void
11873 sfmmu_tsb_chk_reloc(sfmmu_t *sfmmup, hatlock_t *hatlockp)
11874 {
11875 	struct tsb_info *tp;
11876 
11877 	ASSERT(sfmmu_hat_lock_held(sfmmup));
11878 
11879 	for (tp = sfmmup->sfmmu_tsb; tp != NULL; tp = tp->tsb_next) {
11880 		if (tp->tsb_flags & TSB_RELOC_FLAG) {
11881 			cv_wait(&sfmmup->sfmmu_tsb_cv,
11882 			    HATLOCK_MUTEXP(hatlockp));
11883 			break;
11884 		}
11885 	}
11886 }
11887 
11888 /*
11889  * sfmmu_vatopfn_suspended is called from GET_TTE when TL=0 and
11890  * TTE_SUSPENDED bit set in tte we block on aquiring a page lock
11891  * rather than spinning to avoid send mondo timeouts with
11892  * interrupts enabled. When the lock is acquired it is immediately
11893  * released and we return back to sfmmu_vatopfn just after
11894  * the GET_TTE call.
11895  */
11896 void
11897 sfmmu_vatopfn_suspended(caddr_t vaddr, sfmmu_t *sfmmu, tte_t *ttep)
11898 {
11899 	struct page	**pp;
11900 
11901 	(void) as_pagelock(sfmmu->sfmmu_as, &pp, vaddr, TTE_CSZ(ttep), S_WRITE);
11902 	as_pageunlock(sfmmu->sfmmu_as, pp, vaddr, TTE_CSZ(ttep), S_WRITE);
11903 }
11904 
11905 /*
11906  * sfmmu_tsbmiss_suspended is called from GET_TTE when TL>0 and
11907  * TTE_SUSPENDED bit set in tte. We do this so that we can handle
11908  * cross traps which cannot be handled while spinning in the
11909  * trap handlers. Simply enter and exit the kpr_suspendlock spin
11910  * mutex, which is held by the holder of the suspend bit, and then
11911  * retry the trapped instruction after unwinding.
11912  */
11913 /*ARGSUSED*/
11914 void
11915 sfmmu_tsbmiss_suspended(struct regs *rp, uintptr_t tagacc, uint_t traptype)
11916 {
11917 	ASSERT(curthread != kreloc_thread);
11918 	mutex_enter(&kpr_suspendlock);
11919 	mutex_exit(&kpr_suspendlock);
11920 }
11921 
11922 /*
11923  * This routine could be optimized to reduce the number of xcalls by flushing
11924  * the entire TLBs if region reference count is above some threshold but the
11925  * tradeoff will depend on the size of the TLB. So for now flush the specific
11926  * page a context at a time.
11927  *
11928  * If uselocks is 0 then it's called after all cpus were captured and all the
11929  * hat locks were taken. In this case don't take the region lock by relying on
11930  * the order of list region update operations in hat_join_region(),
11931  * hat_leave_region() and hat_dup_region(). The ordering in those routines
11932  * guarantees that list is always forward walkable and reaches active sfmmus
11933  * regardless of where xc_attention() captures a cpu.
11934  */
11935 cpuset_t
11936 sfmmu_rgntlb_demap(caddr_t addr, sf_region_t *rgnp,
11937     struct hme_blk *hmeblkp, int uselocks)
11938 {
11939 	sfmmu_t	*sfmmup;
11940 	cpuset_t cpuset;
11941 	cpuset_t rcpuset;
11942 	hatlock_t *hatlockp;
11943 	uint_t rid = rgnp->rgn_id;
11944 	sf_rgn_link_t *rlink;
11945 	sf_scd_t *scdp;
11946 
11947 	ASSERT(hmeblkp->hblk_shared);
11948 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
11949 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
11950 
11951 	CPUSET_ZERO(rcpuset);
11952 	if (uselocks) {
11953 		mutex_enter(&rgnp->rgn_mutex);
11954 	}
11955 	sfmmup = rgnp->rgn_sfmmu_head;
11956 	while (sfmmup != NULL) {
11957 		if (uselocks) {
11958 			hatlockp = sfmmu_hat_enter(sfmmup);
11959 		}
11960 
11961 		/*
11962 		 * When an SCD is created the SCD hat is linked on the sfmmu
11963 		 * region lists for each hme region which is part of the
11964 		 * SCD. If we find an SCD hat, when walking these lists,
11965 		 * then we flush the shared TSBs, if we find a private hat,
11966 		 * which is part of an SCD, but where the region
11967 		 * is not part of the SCD then we flush the private TSBs.
11968 		 */
11969 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
11970 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
11971 			scdp = sfmmup->sfmmu_scdp;
11972 			if (SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
11973 				if (uselocks) {
11974 					sfmmu_hat_exit(hatlockp);
11975 				}
11976 				goto next;
11977 			}
11978 		}
11979 
11980 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
11981 
11982 		kpreempt_disable();
11983 		cpuset = sfmmup->sfmmu_cpusran;
11984 		CPUSET_AND(cpuset, cpu_ready_set);
11985 		CPUSET_DEL(cpuset, CPU->cpu_id);
11986 		SFMMU_XCALL_STATS(sfmmup);
11987 		xt_some(cpuset, vtag_flushpage_tl1,
11988 		    (uint64_t)addr, (uint64_t)sfmmup);
11989 		vtag_flushpage(addr, (uint64_t)sfmmup);
11990 		if (uselocks) {
11991 			sfmmu_hat_exit(hatlockp);
11992 		}
11993 		kpreempt_enable();
11994 		CPUSET_OR(rcpuset, cpuset);
11995 
11996 next:
11997 		/* LINTED: constant in conditional context */
11998 		SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
11999 		ASSERT(rlink != NULL);
12000 		sfmmup = rlink->next;
12001 	}
12002 	if (uselocks) {
12003 		mutex_exit(&rgnp->rgn_mutex);
12004 	}
12005 	return (rcpuset);
12006 }
12007 
12008 /*
12009  * This routine takes an sfmmu pointer and the va for an adddress in an
12010  * ISM region as input and returns the corresponding region id in ism_rid.
12011  * The return value of 1 indicates that a region has been found and ism_rid
12012  * is valid, otherwise 0 is returned.
12013  */
12014 static int
12015 find_ism_rid(sfmmu_t *sfmmup, sfmmu_t *ism_sfmmup, caddr_t va, uint_t *ism_rid)
12016 {
12017 	ism_blk_t	*ism_blkp;
12018 	int		i;
12019 	ism_map_t	*ism_map;
12020 #ifdef DEBUG
12021 	struct hat	*ism_hatid;
12022 #endif
12023 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12024 
12025 	ism_blkp = sfmmup->sfmmu_iblk;
12026 	while (ism_blkp != NULL) {
12027 		ism_map = ism_blkp->iblk_maps;
12028 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
12029 			if ((va >= ism_start(ism_map[i])) &&
12030 			    (va < ism_end(ism_map[i]))) {
12031 
12032 				*ism_rid = ism_map[i].imap_rid;
12033 #ifdef DEBUG
12034 				ism_hatid = ism_map[i].imap_ismhat;
12035 				ASSERT(ism_hatid == ism_sfmmup);
12036 				ASSERT(ism_hatid->sfmmu_ismhat);
12037 #endif
12038 				return (1);
12039 			}
12040 		}
12041 		ism_blkp = ism_blkp->iblk_next;
12042 	}
12043 	return (0);
12044 }
12045 
12046 /*
12047  * Special routine to flush out ism mappings- TSBs, TLBs and D-caches.
12048  * This routine may be called with all cpu's captured. Therefore, the
12049  * caller is responsible for holding all locks and disabling kernel
12050  * preemption.
12051  */
12052 /* ARGSUSED */
12053 static void
12054 sfmmu_ismtlbcache_demap(caddr_t addr, sfmmu_t *ism_sfmmup,
12055     struct hme_blk *hmeblkp, pfn_t pfnum, int cache_flush_flag)
12056 {
12057 	cpuset_t	cpuset;
12058 	caddr_t		va;
12059 	ism_ment_t	*ment;
12060 	sfmmu_t		*sfmmup;
12061 #ifdef VAC
12062 	int		vcolor;
12063 #endif
12064 
12065 	sf_scd_t	*scdp;
12066 	uint_t		ism_rid;
12067 
12068 	ASSERT(!hmeblkp->hblk_shared);
12069 	/*
12070 	 * Walk the ism_hat's mapping list and flush the page
12071 	 * from every hat sharing this ism_hat. This routine
12072 	 * may be called while all cpu's have been captured.
12073 	 * Therefore we can't attempt to grab any locks. For now
12074 	 * this means we will protect the ism mapping list under
12075 	 * a single lock which will be grabbed by the caller.
12076 	 * If hat_share/unshare scalibility becomes a performance
12077 	 * problem then we may need to re-think ism mapping list locking.
12078 	 */
12079 	ASSERT(ism_sfmmup->sfmmu_ismhat);
12080 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
12081 	addr = addr - ISMID_STARTADDR;
12082 
12083 	for (ment = ism_sfmmup->sfmmu_iment; ment; ment = ment->iment_next) {
12084 
12085 		sfmmup = ment->iment_hat;
12086 
12087 		va = ment->iment_base_va;
12088 		va = (caddr_t)((uintptr_t)va  + (uintptr_t)addr);
12089 
12090 		/*
12091 		 * When an SCD is created the SCD hat is linked on the ism
12092 		 * mapping lists for each ISM segment which is part of the
12093 		 * SCD. If we find an SCD hat, when walking these lists,
12094 		 * then we flush the shared TSBs, if we find a private hat,
12095 		 * which is part of an SCD, but where the region
12096 		 * corresponding to this va is not part of the SCD then we
12097 		 * flush the private TSBs.
12098 		 */
12099 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12100 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD) &&
12101 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
12102 			if (!find_ism_rid(sfmmup, ism_sfmmup, va,
12103 			    &ism_rid)) {
12104 				cmn_err(CE_PANIC,
12105 				    "can't find matching ISM rid!");
12106 			}
12107 
12108 			scdp = sfmmup->sfmmu_scdp;
12109 			if (SFMMU_IS_ISMRID_VALID(ism_rid) &&
12110 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map,
12111 			    ism_rid)) {
12112 				continue;
12113 			}
12114 		}
12115 		SFMMU_UNLOAD_TSB(va, sfmmup, hmeblkp, 1);
12116 
12117 		cpuset = sfmmup->sfmmu_cpusran;
12118 		CPUSET_AND(cpuset, cpu_ready_set);
12119 		CPUSET_DEL(cpuset, CPU->cpu_id);
12120 		SFMMU_XCALL_STATS(sfmmup);
12121 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)va,
12122 		    (uint64_t)sfmmup);
12123 		vtag_flushpage(va, (uint64_t)sfmmup);
12124 
12125 #ifdef VAC
12126 		/*
12127 		 * Flush D$
12128 		 * When flushing D$ we must flush all
12129 		 * cpu's. See sfmmu_cache_flush().
12130 		 */
12131 		if (cache_flush_flag == CACHE_FLUSH) {
12132 			cpuset = cpu_ready_set;
12133 			CPUSET_DEL(cpuset, CPU->cpu_id);
12134 
12135 			SFMMU_XCALL_STATS(sfmmup);
12136 			vcolor = addr_to_vcolor(va);
12137 			xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12138 			vac_flushpage(pfnum, vcolor);
12139 		}
12140 #endif	/* VAC */
12141 	}
12142 }
12143 
12144 /*
12145  * Demaps the TSB, CPU caches, and flushes all TLBs on all CPUs of
12146  * a particular virtual address and ctx.  If noflush is set we do not
12147  * flush the TLB/TSB.  This function may or may not be called with the
12148  * HAT lock held.
12149  */
12150 static void
12151 sfmmu_tlbcache_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12152     pfn_t pfnum, int tlb_noflush, int cpu_flag, int cache_flush_flag,
12153     int hat_lock_held)
12154 {
12155 #ifdef VAC
12156 	int vcolor;
12157 #endif
12158 	cpuset_t cpuset;
12159 	hatlock_t *hatlockp;
12160 
12161 	ASSERT(!hmeblkp->hblk_shared);
12162 
12163 #if defined(lint) && !defined(VAC)
12164 	pfnum = pfnum;
12165 	cpu_flag = cpu_flag;
12166 	cache_flush_flag = cache_flush_flag;
12167 #endif
12168 
12169 	/*
12170 	 * There is no longer a need to protect against ctx being
12171 	 * stolen here since we don't store the ctx in the TSB anymore.
12172 	 */
12173 #ifdef VAC
12174 	vcolor = addr_to_vcolor(addr);
12175 #endif
12176 
12177 	/*
12178 	 * We must hold the hat lock during the flush of TLB,
12179 	 * to avoid a race with sfmmu_invalidate_ctx(), where
12180 	 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12181 	 * causing TLB demap routine to skip flush on that MMU.
12182 	 * If the context on a MMU has already been set to
12183 	 * INVALID_CONTEXT, we just get an extra flush on
12184 	 * that MMU.
12185 	 */
12186 	if (!hat_lock_held && !tlb_noflush)
12187 		hatlockp = sfmmu_hat_enter(sfmmup);
12188 
12189 	kpreempt_disable();
12190 	if (!tlb_noflush) {
12191 		/*
12192 		 * Flush the TSB and TLB.
12193 		 */
12194 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12195 
12196 		cpuset = sfmmup->sfmmu_cpusran;
12197 		CPUSET_AND(cpuset, cpu_ready_set);
12198 		CPUSET_DEL(cpuset, CPU->cpu_id);
12199 
12200 		SFMMU_XCALL_STATS(sfmmup);
12201 
12202 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
12203 		    (uint64_t)sfmmup);
12204 
12205 		vtag_flushpage(addr, (uint64_t)sfmmup);
12206 	}
12207 
12208 	if (!hat_lock_held && !tlb_noflush)
12209 		sfmmu_hat_exit(hatlockp);
12210 
12211 #ifdef VAC
12212 	/*
12213 	 * Flush the D$
12214 	 *
12215 	 * Even if the ctx is stolen, we need to flush the
12216 	 * cache. Our ctx stealer only flushes the TLBs.
12217 	 */
12218 	if (cache_flush_flag == CACHE_FLUSH) {
12219 		if (cpu_flag & FLUSH_ALL_CPUS) {
12220 			cpuset = cpu_ready_set;
12221 		} else {
12222 			cpuset = sfmmup->sfmmu_cpusran;
12223 			CPUSET_AND(cpuset, cpu_ready_set);
12224 		}
12225 		CPUSET_DEL(cpuset, CPU->cpu_id);
12226 		SFMMU_XCALL_STATS(sfmmup);
12227 		xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12228 		vac_flushpage(pfnum, vcolor);
12229 	}
12230 #endif	/* VAC */
12231 	kpreempt_enable();
12232 }
12233 
12234 /*
12235  * Demaps the TSB and flushes all TLBs on all cpus for a particular virtual
12236  * address and ctx.  If noflush is set we do not currently do anything.
12237  * This function may or may not be called with the HAT lock held.
12238  */
12239 static void
12240 sfmmu_tlb_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12241     int tlb_noflush, int hat_lock_held)
12242 {
12243 	cpuset_t cpuset;
12244 	hatlock_t *hatlockp;
12245 
12246 	ASSERT(!hmeblkp->hblk_shared);
12247 
12248 	/*
12249 	 * If the process is exiting we have nothing to do.
12250 	 */
12251 	if (tlb_noflush)
12252 		return;
12253 
12254 	/*
12255 	 * Flush TSB.
12256 	 */
12257 	if (!hat_lock_held)
12258 		hatlockp = sfmmu_hat_enter(sfmmup);
12259 	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12260 
12261 	kpreempt_disable();
12262 
12263 	cpuset = sfmmup->sfmmu_cpusran;
12264 	CPUSET_AND(cpuset, cpu_ready_set);
12265 	CPUSET_DEL(cpuset, CPU->cpu_id);
12266 
12267 	SFMMU_XCALL_STATS(sfmmup);
12268 	xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, (uint64_t)sfmmup);
12269 
12270 	vtag_flushpage(addr, (uint64_t)sfmmup);
12271 
12272 	if (!hat_lock_held)
12273 		sfmmu_hat_exit(hatlockp);
12274 
12275 	kpreempt_enable();
12276 
12277 }
12278 
12279 /*
12280  * Special case of sfmmu_tlb_demap for MMU_PAGESIZE hblks. Use the xcall
12281  * call handler that can flush a range of pages to save on xcalls.
12282  */
12283 static int sfmmu_xcall_save;
12284 
12285 /*
12286  * this routine is never used for demaping addresses backed by SRD hmeblks.
12287  */
12288 static void
12289 sfmmu_tlb_range_demap(demap_range_t *dmrp)
12290 {
12291 	sfmmu_t *sfmmup = dmrp->dmr_sfmmup;
12292 	hatlock_t *hatlockp;
12293 	cpuset_t cpuset;
12294 	uint64_t sfmmu_pgcnt;
12295 	pgcnt_t pgcnt = 0;
12296 	int pgunload = 0;
12297 	int dirtypg = 0;
12298 	caddr_t addr = dmrp->dmr_addr;
12299 	caddr_t eaddr;
12300 	uint64_t bitvec = dmrp->dmr_bitvec;
12301 
12302 	ASSERT(bitvec & 1);
12303 
12304 	/*
12305 	 * Flush TSB and calculate number of pages to flush.
12306 	 */
12307 	while (bitvec != 0) {
12308 		dirtypg = 0;
12309 		/*
12310 		 * Find the first page to flush and then count how many
12311 		 * pages there are after it that also need to be flushed.
12312 		 * This way the number of TSB flushes is minimized.
12313 		 */
12314 		while ((bitvec & 1) == 0) {
12315 			pgcnt++;
12316 			addr += MMU_PAGESIZE;
12317 			bitvec >>= 1;
12318 		}
12319 		while (bitvec & 1) {
12320 			dirtypg++;
12321 			bitvec >>= 1;
12322 		}
12323 		eaddr = addr + ptob(dirtypg);
12324 		hatlockp = sfmmu_hat_enter(sfmmup);
12325 		sfmmu_unload_tsb_range(sfmmup, addr, eaddr, TTE8K);
12326 		sfmmu_hat_exit(hatlockp);
12327 		pgunload += dirtypg;
12328 		addr = eaddr;
12329 		pgcnt += dirtypg;
12330 	}
12331 
12332 	ASSERT((pgcnt<<MMU_PAGESHIFT) <= dmrp->dmr_endaddr - dmrp->dmr_addr);
12333 	if (sfmmup->sfmmu_free == 0) {
12334 		addr = dmrp->dmr_addr;
12335 		bitvec = dmrp->dmr_bitvec;
12336 
12337 		/*
12338 		 * make sure it has SFMMU_PGCNT_SHIFT bits only,
12339 		 * as it will be used to pack argument for xt_some
12340 		 */
12341 		ASSERT((pgcnt > 0) &&
12342 		    (pgcnt <= (1 << SFMMU_PGCNT_SHIFT)));
12343 
12344 		/*
12345 		 * Encode pgcnt as (pgcnt -1 ), and pass (pgcnt - 1) in
12346 		 * the low 6 bits of sfmmup. This is doable since pgcnt
12347 		 * always >= 1.
12348 		 */
12349 		ASSERT(!((uint64_t)sfmmup & SFMMU_PGCNT_MASK));
12350 		sfmmu_pgcnt = (uint64_t)sfmmup |
12351 		    ((pgcnt - 1) & SFMMU_PGCNT_MASK);
12352 
12353 		/*
12354 		 * We must hold the hat lock during the flush of TLB,
12355 		 * to avoid a race with sfmmu_invalidate_ctx(), where
12356 		 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12357 		 * causing TLB demap routine to skip flush on that MMU.
12358 		 * If the context on a MMU has already been set to
12359 		 * INVALID_CONTEXT, we just get an extra flush on
12360 		 * that MMU.
12361 		 */
12362 		hatlockp = sfmmu_hat_enter(sfmmup);
12363 		kpreempt_disable();
12364 
12365 		cpuset = sfmmup->sfmmu_cpusran;
12366 		CPUSET_AND(cpuset, cpu_ready_set);
12367 		CPUSET_DEL(cpuset, CPU->cpu_id);
12368 
12369 		SFMMU_XCALL_STATS(sfmmup);
12370 		xt_some(cpuset, vtag_flush_pgcnt_tl1, (uint64_t)addr,
12371 		    sfmmu_pgcnt);
12372 
12373 		for (; bitvec != 0; bitvec >>= 1) {
12374 			if (bitvec & 1)
12375 				vtag_flushpage(addr, (uint64_t)sfmmup);
12376 			addr += MMU_PAGESIZE;
12377 		}
12378 		kpreempt_enable();
12379 		sfmmu_hat_exit(hatlockp);
12380 
12381 		sfmmu_xcall_save += (pgunload-1);
12382 	}
12383 	dmrp->dmr_bitvec = 0;
12384 }
12385 
12386 /*
12387  * In cases where we need to synchronize with TLB/TSB miss trap
12388  * handlers, _and_ need to flush the TLB, it's a lot easier to
12389  * throw away the context from the process than to do a
12390  * special song and dance to keep things consistent for the
12391  * handlers.
12392  *
12393  * Since the process suddenly ends up without a context and our caller
12394  * holds the hat lock, threads that fault after this function is called
12395  * will pile up on the lock.  We can then do whatever we need to
12396  * atomically from the context of the caller.  The first blocked thread
12397  * to resume executing will get the process a new context, and the
12398  * process will resume executing.
12399  *
12400  * One added advantage of this approach is that on MMUs that
12401  * support a "flush all" operation, we will delay the flush until
12402  * cnum wrap-around, and then flush the TLB one time.  This
12403  * is rather rare, so it's a lot less expensive than making 8000
12404  * x-calls to flush the TLB 8000 times.
12405  *
12406  * A per-process (PP) lock is used to synchronize ctx allocations in
12407  * resume() and ctx invalidations here.
12408  */
12409 static void
12410 sfmmu_invalidate_ctx(sfmmu_t *sfmmup)
12411 {
12412 	cpuset_t cpuset;
12413 	int cnum, currcnum;
12414 	mmu_ctx_t *mmu_ctxp;
12415 	int i;
12416 	uint_t pstate_save;
12417 
12418 	SFMMU_STAT(sf_ctx_inv);
12419 
12420 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12421 	ASSERT(sfmmup != ksfmmup);
12422 
12423 	kpreempt_disable();
12424 
12425 	mmu_ctxp = CPU_MMU_CTXP(CPU);
12426 	ASSERT(mmu_ctxp);
12427 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
12428 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
12429 
12430 	currcnum = sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum;
12431 
12432 	pstate_save = sfmmu_disable_intrs();
12433 
12434 	lock_set(&sfmmup->sfmmu_ctx_lock);	/* acquire PP lock */
12435 	/* set HAT cnum invalid across all context domains. */
12436 	for (i = 0; i < max_mmu_ctxdoms; i++) {
12437 
12438 		cnum = sfmmup->sfmmu_ctxs[i].cnum;
12439 		if (cnum == INVALID_CONTEXT) {
12440 			continue;
12441 		}
12442 
12443 		sfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
12444 	}
12445 	membar_enter();	/* make sure globally visible to all CPUs */
12446 	lock_clear(&sfmmup->sfmmu_ctx_lock);	/* release PP lock */
12447 
12448 	sfmmu_enable_intrs(pstate_save);
12449 
12450 	cpuset = sfmmup->sfmmu_cpusran;
12451 	CPUSET_DEL(cpuset, CPU->cpu_id);
12452 	CPUSET_AND(cpuset, cpu_ready_set);
12453 	if (!CPUSET_ISNULL(cpuset)) {
12454 		SFMMU_XCALL_STATS(sfmmup);
12455 		xt_some(cpuset, sfmmu_raise_tsb_exception,
12456 		    (uint64_t)sfmmup, INVALID_CONTEXT);
12457 		xt_sync(cpuset);
12458 		SFMMU_STAT(sf_tsb_raise_exception);
12459 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
12460 	}
12461 
12462 	/*
12463 	 * If the hat to-be-invalidated is the same as the current
12464 	 * process on local CPU we need to invalidate
12465 	 * this CPU context as well.
12466 	 */
12467 	if ((sfmmu_getctx_sec() == currcnum) &&
12468 	    (currcnum != INVALID_CONTEXT)) {
12469 		/* sets shared context to INVALID too */
12470 		sfmmu_setctx_sec(INVALID_CONTEXT);
12471 		sfmmu_clear_utsbinfo();
12472 	}
12473 
12474 	SFMMU_FLAGS_SET(sfmmup, HAT_ALLCTX_INVALID);
12475 
12476 	kpreempt_enable();
12477 
12478 	/*
12479 	 * we hold the hat lock, so nobody should allocate a context
12480 	 * for us yet
12481 	 */
12482 	ASSERT(sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum == INVALID_CONTEXT);
12483 }
12484 
12485 #ifdef VAC
12486 /*
12487  * We need to flush the cache in all cpus.  It is possible that
12488  * a process referenced a page as cacheable but has sinced exited
12489  * and cleared the mapping list.  We still to flush it but have no
12490  * state so all cpus is the only alternative.
12491  */
12492 void
12493 sfmmu_cache_flush(pfn_t pfnum, int vcolor)
12494 {
12495 	cpuset_t cpuset;
12496 
12497 	kpreempt_disable();
12498 	cpuset = cpu_ready_set;
12499 	CPUSET_DEL(cpuset, CPU->cpu_id);
12500 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12501 	xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12502 	xt_sync(cpuset);
12503 	vac_flushpage(pfnum, vcolor);
12504 	kpreempt_enable();
12505 }
12506 
12507 void
12508 sfmmu_cache_flushcolor(int vcolor, pfn_t pfnum)
12509 {
12510 	cpuset_t cpuset;
12511 
12512 	ASSERT(vcolor >= 0);
12513 
12514 	kpreempt_disable();
12515 	cpuset = cpu_ready_set;
12516 	CPUSET_DEL(cpuset, CPU->cpu_id);
12517 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12518 	xt_some(cpuset, vac_flushcolor_tl1, vcolor, pfnum);
12519 	xt_sync(cpuset);
12520 	vac_flushcolor(vcolor, pfnum);
12521 	kpreempt_enable();
12522 }
12523 #endif	/* VAC */
12524 
12525 /*
12526  * We need to prevent processes from accessing the TSB using a cached physical
12527  * address.  It's alright if they try to access the TSB via virtual address
12528  * since they will just fault on that virtual address once the mapping has
12529  * been suspended.
12530  */
12531 #pragma weak sendmondo_in_recover
12532 
12533 /* ARGSUSED */
12534 static int
12535 sfmmu_tsb_pre_relocator(caddr_t va, uint_t tsbsz, uint_t flags, void *tsbinfo)
12536 {
12537 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12538 	sfmmu_t *sfmmup = tsbinfop->tsb_sfmmu;
12539 	hatlock_t *hatlockp;
12540 	sf_scd_t *scdp;
12541 
12542 	if (flags != HAT_PRESUSPEND)
12543 		return (0);
12544 
12545 	/*
12546 	 * If tsb is a shared TSB with TSB_SHAREDCTX set, sfmmup must
12547 	 * be a shared hat, then set SCD's tsbinfo's flag.
12548 	 * If tsb is not shared, sfmmup is a private hat, then set
12549 	 * its private tsbinfo's flag.
12550 	 */
12551 	hatlockp = sfmmu_hat_enter(sfmmup);
12552 	tsbinfop->tsb_flags |= TSB_RELOC_FLAG;
12553 
12554 	if (!(tsbinfop->tsb_flags & TSB_SHAREDCTX)) {
12555 		sfmmu_tsb_inv_ctx(sfmmup);
12556 		sfmmu_hat_exit(hatlockp);
12557 	} else {
12558 		/* release lock on the shared hat */
12559 		sfmmu_hat_exit(hatlockp);
12560 		/* sfmmup is a shared hat */
12561 		ASSERT(sfmmup->sfmmu_scdhat);
12562 		scdp = sfmmup->sfmmu_scdp;
12563 		ASSERT(scdp != NULL);
12564 		/* get private hat from the scd list */
12565 		mutex_enter(&scdp->scd_mutex);
12566 		sfmmup = scdp->scd_sf_list;
12567 		while (sfmmup != NULL) {
12568 			hatlockp = sfmmu_hat_enter(sfmmup);
12569 			/*
12570 			 * We do not call sfmmu_tsb_inv_ctx here because
12571 			 * sendmondo_in_recover check is only needed for
12572 			 * sun4u.
12573 			 */
12574 			sfmmu_invalidate_ctx(sfmmup);
12575 			sfmmu_hat_exit(hatlockp);
12576 			sfmmup = sfmmup->sfmmu_scd_link.next;
12577 
12578 		}
12579 		mutex_exit(&scdp->scd_mutex);
12580 	}
12581 	return (0);
12582 }
12583 
12584 static void
12585 sfmmu_tsb_inv_ctx(sfmmu_t *sfmmup)
12586 {
12587 	extern uint32_t sendmondo_in_recover;
12588 
12589 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12590 
12591 	/*
12592 	 * For Cheetah+ Erratum 25:
12593 	 * Wait for any active recovery to finish.  We can't risk
12594 	 * relocating the TSB of the thread running mondo_recover_proc()
12595 	 * since, if we did that, we would deadlock.  The scenario we are
12596 	 * trying to avoid is as follows:
12597 	 *
12598 	 * THIS CPU			RECOVER CPU
12599 	 * --------			-----------
12600 	 *				Begins recovery, walking through TSB
12601 	 * hat_pagesuspend() TSB TTE
12602 	 *				TLB miss on TSB TTE, spins at TL1
12603 	 * xt_sync()
12604 	 *	send_mondo_timeout()
12605 	 *	mondo_recover_proc()
12606 	 *	((deadlocked))
12607 	 *
12608 	 * The second half of the workaround is that mondo_recover_proc()
12609 	 * checks to see if the tsb_info has the RELOC flag set, and if it
12610 	 * does, it skips over that TSB without ever touching tsbinfop->tsb_va
12611 	 * and hence avoiding the TLB miss that could result in a deadlock.
12612 	 */
12613 	if (&sendmondo_in_recover) {
12614 		membar_enter();	/* make sure RELOC flag visible */
12615 		while (sendmondo_in_recover) {
12616 			drv_usecwait(1);
12617 			membar_consumer();
12618 		}
12619 	}
12620 
12621 	sfmmu_invalidate_ctx(sfmmup);
12622 }
12623 
12624 /* ARGSUSED */
12625 static int
12626 sfmmu_tsb_post_relocator(caddr_t va, uint_t tsbsz, uint_t flags,
12627     void *tsbinfo, pfn_t newpfn)
12628 {
12629 	hatlock_t *hatlockp;
12630 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12631 	sfmmu_t	*sfmmup = tsbinfop->tsb_sfmmu;
12632 
12633 	if (flags != HAT_POSTUNSUSPEND)
12634 		return (0);
12635 
12636 	hatlockp = sfmmu_hat_enter(sfmmup);
12637 
12638 	SFMMU_STAT(sf_tsb_reloc);
12639 
12640 	/*
12641 	 * The process may have swapped out while we were relocating one
12642 	 * of its TSBs.  If so, don't bother doing the setup since the
12643 	 * process can't be using the memory anymore.
12644 	 */
12645 	if ((tsbinfop->tsb_flags & TSB_SWAPPED) == 0) {
12646 		ASSERT(va == tsbinfop->tsb_va);
12647 		sfmmu_tsbinfo_setup_phys(tsbinfop, newpfn);
12648 
12649 		if (tsbinfop->tsb_flags & TSB_FLUSH_NEEDED) {
12650 			sfmmu_inv_tsb(tsbinfop->tsb_va,
12651 			    TSB_BYTES(tsbinfop->tsb_szc));
12652 			tsbinfop->tsb_flags &= ~TSB_FLUSH_NEEDED;
12653 		}
12654 	}
12655 
12656 	membar_exit();
12657 	tsbinfop->tsb_flags &= ~TSB_RELOC_FLAG;
12658 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
12659 
12660 	sfmmu_hat_exit(hatlockp);
12661 
12662 	return (0);
12663 }
12664 
12665 /*
12666  * Allocate and initialize a tsb_info structure.  Note that we may or may not
12667  * allocate a TSB here, depending on the flags passed in.
12668  */
12669 static int
12670 sfmmu_tsbinfo_alloc(struct tsb_info **tsbinfopp, int tsb_szc, int tte_sz_mask,
12671     uint_t flags, sfmmu_t *sfmmup)
12672 {
12673 	int err;
12674 
12675 	*tsbinfopp = (struct tsb_info *)kmem_cache_alloc(
12676 	    sfmmu_tsbinfo_cache, KM_SLEEP);
12677 
12678 	if ((err = sfmmu_init_tsbinfo(*tsbinfopp, tte_sz_mask,
12679 	    tsb_szc, flags, sfmmup)) != 0) {
12680 		kmem_cache_free(sfmmu_tsbinfo_cache, *tsbinfopp);
12681 		SFMMU_STAT(sf_tsb_allocfail);
12682 		*tsbinfopp = NULL;
12683 		return (err);
12684 	}
12685 	SFMMU_STAT(sf_tsb_alloc);
12686 
12687 	/*
12688 	 * Bump the TSB size counters for this TSB size.
12689 	 */
12690 	(*(((int *)&sfmmu_tsbsize_stat) + tsb_szc))++;
12691 	return (0);
12692 }
12693 
12694 static void
12695 sfmmu_tsb_free(struct tsb_info *tsbinfo)
12696 {
12697 	caddr_t tsbva = tsbinfo->tsb_va;
12698 	uint_t tsb_size = TSB_BYTES(tsbinfo->tsb_szc);
12699 	struct kmem_cache *kmem_cachep = tsbinfo->tsb_cache;
12700 	vmem_t	*vmp = tsbinfo->tsb_vmp;
12701 
12702 	/*
12703 	 * If we allocated this TSB from relocatable kernel memory, then we
12704 	 * need to uninstall the callback handler.
12705 	 */
12706 	if (tsbinfo->tsb_cache != sfmmu_tsb8k_cache) {
12707 		uintptr_t slab_mask;
12708 		caddr_t slab_vaddr;
12709 		page_t **ppl;
12710 		int ret;
12711 
12712 		ASSERT(tsb_size <= MMU_PAGESIZE4M || use_bigtsb_arena);
12713 		if (tsb_size > MMU_PAGESIZE4M)
12714 			slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
12715 		else
12716 			slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
12717 		slab_vaddr = (caddr_t)((uintptr_t)tsbva & slab_mask);
12718 
12719 		ret = as_pagelock(&kas, &ppl, slab_vaddr, PAGESIZE, S_WRITE);
12720 		ASSERT(ret == 0);
12721 		hat_delete_callback(tsbva, (uint_t)tsb_size, (void *)tsbinfo,
12722 		    0, NULL);
12723 		as_pageunlock(&kas, ppl, slab_vaddr, PAGESIZE, S_WRITE);
12724 	}
12725 
12726 	if (kmem_cachep != NULL) {
12727 		kmem_cache_free(kmem_cachep, tsbva);
12728 	} else {
12729 		vmem_xfree(vmp, (void *)tsbva, tsb_size);
12730 	}
12731 	tsbinfo->tsb_va = (caddr_t)0xbad00bad;
12732 	atomic_add_64(&tsb_alloc_bytes, -(int64_t)tsb_size);
12733 }
12734 
12735 static void
12736 sfmmu_tsbinfo_free(struct tsb_info *tsbinfo)
12737 {
12738 	if ((tsbinfo->tsb_flags & TSB_SWAPPED) == 0) {
12739 		sfmmu_tsb_free(tsbinfo);
12740 	}
12741 	kmem_cache_free(sfmmu_tsbinfo_cache, tsbinfo);
12742 
12743 }
12744 
12745 /*
12746  * Setup all the references to physical memory for this tsbinfo.
12747  * The underlying page(s) must be locked.
12748  */
12749 static void
12750 sfmmu_tsbinfo_setup_phys(struct tsb_info *tsbinfo, pfn_t pfn)
12751 {
12752 	ASSERT(pfn != PFN_INVALID);
12753 	ASSERT(pfn == va_to_pfn(tsbinfo->tsb_va));
12754 
12755 #ifndef sun4v
12756 	if (tsbinfo->tsb_szc == 0) {
12757 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn,
12758 		    PROT_WRITE|PROT_READ, TTE8K);
12759 	} else {
12760 		/*
12761 		 * Round down PA and use a large mapping; the handlers will
12762 		 * compute the TSB pointer at the correct offset into the
12763 		 * big virtual page.  NOTE: this assumes all TSBs larger
12764 		 * than 8K must come from physically contiguous slabs of
12765 		 * size tsb_slab_size.
12766 		 */
12767 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn & ~tsb_slab_mask,
12768 		    PROT_WRITE|PROT_READ, tsb_slab_ttesz);
12769 	}
12770 	tsbinfo->tsb_pa = ptob(pfn);
12771 
12772 	TTE_SET_LOCKED(&tsbinfo->tsb_tte); /* lock the tte into dtlb */
12773 	TTE_SET_MOD(&tsbinfo->tsb_tte);    /* enable writes */
12774 
12775 	ASSERT(TTE_IS_PRIVILEGED(&tsbinfo->tsb_tte));
12776 	ASSERT(TTE_IS_LOCKED(&tsbinfo->tsb_tte));
12777 #else /* sun4v */
12778 	tsbinfo->tsb_pa = ptob(pfn);
12779 #endif /* sun4v */
12780 }
12781 
12782 
12783 /*
12784  * Returns zero on success, ENOMEM if over the high water mark,
12785  * or EAGAIN if the caller needs to retry with a smaller TSB
12786  * size (or specify TSB_FORCEALLOC if the allocation can't fail).
12787  *
12788  * This call cannot fail to allocate a TSB if TSB_FORCEALLOC
12789  * is specified and the TSB requested is PAGESIZE, though it
12790  * may sleep waiting for memory if sufficient memory is not
12791  * available.
12792  */
12793 static int
12794 sfmmu_init_tsbinfo(struct tsb_info *tsbinfo, int tteszmask,
12795     int tsbcode, uint_t flags, sfmmu_t *sfmmup)
12796 {
12797 	caddr_t vaddr = NULL;
12798 	caddr_t slab_vaddr;
12799 	uintptr_t slab_mask;
12800 	int tsbbytes = TSB_BYTES(tsbcode);
12801 	int lowmem = 0;
12802 	struct kmem_cache *kmem_cachep = NULL;
12803 	vmem_t *vmp = NULL;
12804 	lgrp_id_t lgrpid = LGRP_NONE;
12805 	pfn_t pfn;
12806 	uint_t cbflags = HAC_SLEEP;
12807 	page_t **pplist;
12808 	int ret;
12809 
12810 	ASSERT(tsbbytes <= MMU_PAGESIZE4M || use_bigtsb_arena);
12811 	if (tsbbytes > MMU_PAGESIZE4M)
12812 		slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
12813 	else
12814 		slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
12815 
12816 	if (flags & (TSB_FORCEALLOC | TSB_SWAPIN | TSB_GROW | TSB_SHRINK))
12817 		flags |= TSB_ALLOC;
12818 
12819 	ASSERT((flags & TSB_FORCEALLOC) == 0 || tsbcode == TSB_MIN_SZCODE);
12820 
12821 	tsbinfo->tsb_sfmmu = sfmmup;
12822 
12823 	/*
12824 	 * If not allocating a TSB, set up the tsbinfo, set TSB_SWAPPED, and
12825 	 * return.
12826 	 */
12827 	if ((flags & TSB_ALLOC) == 0) {
12828 		tsbinfo->tsb_szc = tsbcode;
12829 		tsbinfo->tsb_ttesz_mask = tteszmask;
12830 		tsbinfo->tsb_va = (caddr_t)0xbadbadbeef;
12831 		tsbinfo->tsb_pa = -1;
12832 		tsbinfo->tsb_tte.ll = 0;
12833 		tsbinfo->tsb_next = NULL;
12834 		tsbinfo->tsb_flags = TSB_SWAPPED;
12835 		tsbinfo->tsb_cache = NULL;
12836 		tsbinfo->tsb_vmp = NULL;
12837 		return (0);
12838 	}
12839 
12840 #ifdef DEBUG
12841 	/*
12842 	 * For debugging:
12843 	 * Randomly force allocation failures every tsb_alloc_mtbf
12844 	 * tries if TSB_FORCEALLOC is not specified.  This will
12845 	 * return ENOMEM if tsb_alloc_mtbf is odd, or EAGAIN if
12846 	 * it is even, to allow testing of both failure paths...
12847 	 */
12848 	if (tsb_alloc_mtbf && ((flags & TSB_FORCEALLOC) == 0) &&
12849 	    (tsb_alloc_count++ == tsb_alloc_mtbf)) {
12850 		tsb_alloc_count = 0;
12851 		tsb_alloc_fail_mtbf++;
12852 		return ((tsb_alloc_mtbf & 1)? ENOMEM : EAGAIN);
12853 	}
12854 #endif	/* DEBUG */
12855 
12856 	/*
12857 	 * Enforce high water mark if we are not doing a forced allocation
12858 	 * and are not shrinking a process' TSB.
12859 	 */
12860 	if ((flags & TSB_SHRINK) == 0 &&
12861 	    (tsbbytes + tsb_alloc_bytes) > tsb_alloc_hiwater) {
12862 		if ((flags & TSB_FORCEALLOC) == 0)
12863 			return (ENOMEM);
12864 		lowmem = 1;
12865 	}
12866 
12867 	/*
12868 	 * Allocate from the correct location based upon the size of the TSB
12869 	 * compared to the base page size, and what memory conditions dictate.
12870 	 * Note we always do nonblocking allocations from the TSB arena since
12871 	 * we don't want memory fragmentation to cause processes to block
12872 	 * indefinitely waiting for memory; until the kernel algorithms that
12873 	 * coalesce large pages are improved this is our best option.
12874 	 *
12875 	 * Algorithm:
12876 	 *	If allocating a "large" TSB (>8K), allocate from the
12877 	 *		appropriate kmem_tsb_default_arena vmem arena
12878 	 *	else if low on memory or the TSB_FORCEALLOC flag is set or
12879 	 *	tsb_forceheap is set
12880 	 *		Allocate from kernel heap via sfmmu_tsb8k_cache with
12881 	 *		KM_SLEEP (never fails)
12882 	 *	else
12883 	 *		Allocate from appropriate sfmmu_tsb_cache with
12884 	 *		KM_NOSLEEP
12885 	 *	endif
12886 	 */
12887 	if (tsb_lgrp_affinity)
12888 		lgrpid = lgrp_home_id(curthread);
12889 	if (lgrpid == LGRP_NONE)
12890 		lgrpid = 0;	/* use lgrp of boot CPU */
12891 
12892 	if (tsbbytes > MMU_PAGESIZE) {
12893 		if (tsbbytes > MMU_PAGESIZE4M) {
12894 			vmp = kmem_bigtsb_default_arena[lgrpid];
12895 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
12896 			    0, 0, NULL, NULL, VM_NOSLEEP);
12897 		} else {
12898 			vmp = kmem_tsb_default_arena[lgrpid];
12899 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
12900 			    0, 0, NULL, NULL, VM_NOSLEEP);
12901 		}
12902 #ifdef	DEBUG
12903 	} else if (lowmem || (flags & TSB_FORCEALLOC) || tsb_forceheap) {
12904 #else	/* !DEBUG */
12905 	} else if (lowmem || (flags & TSB_FORCEALLOC)) {
12906 #endif	/* DEBUG */
12907 		kmem_cachep = sfmmu_tsb8k_cache;
12908 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_SLEEP);
12909 		ASSERT(vaddr != NULL);
12910 	} else {
12911 		kmem_cachep = sfmmu_tsb_cache[lgrpid];
12912 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_NOSLEEP);
12913 	}
12914 
12915 	tsbinfo->tsb_cache = kmem_cachep;
12916 	tsbinfo->tsb_vmp = vmp;
12917 
12918 	if (vaddr == NULL) {
12919 		return (EAGAIN);
12920 	}
12921 
12922 	atomic_add_64(&tsb_alloc_bytes, (int64_t)tsbbytes);
12923 	kmem_cachep = tsbinfo->tsb_cache;
12924 
12925 	/*
12926 	 * If we are allocating from outside the cage, then we need to
12927 	 * register a relocation callback handler.  Note that for now
12928 	 * since pseudo mappings always hang off of the slab's root page,
12929 	 * we need only lock the first 8K of the TSB slab.  This is a bit
12930 	 * hacky but it is good for performance.
12931 	 */
12932 	if (kmem_cachep != sfmmu_tsb8k_cache) {
12933 		slab_vaddr = (caddr_t)((uintptr_t)vaddr & slab_mask);
12934 		ret = as_pagelock(&kas, &pplist, slab_vaddr, PAGESIZE, S_WRITE);
12935 		ASSERT(ret == 0);
12936 		ret = hat_add_callback(sfmmu_tsb_cb_id, vaddr, (uint_t)tsbbytes,
12937 		    cbflags, (void *)tsbinfo, &pfn, NULL);
12938 
12939 		/*
12940 		 * Need to free up resources if we could not successfully
12941 		 * add the callback function and return an error condition.
12942 		 */
12943 		if (ret != 0) {
12944 			if (kmem_cachep) {
12945 				kmem_cache_free(kmem_cachep, vaddr);
12946 			} else {
12947 				vmem_xfree(vmp, (void *)vaddr, tsbbytes);
12948 			}
12949 			as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE,
12950 			    S_WRITE);
12951 			return (EAGAIN);
12952 		}
12953 	} else {
12954 		/*
12955 		 * Since allocation of 8K TSBs from heap is rare and occurs
12956 		 * during memory pressure we allocate them from permanent
12957 		 * memory rather than using callbacks to get the PFN.
12958 		 */
12959 		pfn = hat_getpfnum(kas.a_hat, vaddr);
12960 	}
12961 
12962 	tsbinfo->tsb_va = vaddr;
12963 	tsbinfo->tsb_szc = tsbcode;
12964 	tsbinfo->tsb_ttesz_mask = tteszmask;
12965 	tsbinfo->tsb_next = NULL;
12966 	tsbinfo->tsb_flags = 0;
12967 
12968 	sfmmu_tsbinfo_setup_phys(tsbinfo, pfn);
12969 
12970 	sfmmu_inv_tsb(vaddr, tsbbytes);
12971 
12972 	if (kmem_cachep != sfmmu_tsb8k_cache) {
12973 		as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE, S_WRITE);
12974 	}
12975 
12976 	return (0);
12977 }
12978 
12979 /*
12980  * Initialize per cpu tsb and per cpu tsbmiss_area
12981  */
12982 void
12983 sfmmu_init_tsbs(void)
12984 {
12985 	int i;
12986 	struct tsbmiss	*tsbmissp;
12987 	struct kpmtsbm	*kpmtsbmp;
12988 #ifndef sun4v
12989 	extern int	dcache_line_mask;
12990 #endif /* sun4v */
12991 	extern uint_t	vac_colors;
12992 
12993 	/*
12994 	 * Init. tsb miss area.
12995 	 */
12996 	tsbmissp = tsbmiss_area;
12997 
12998 	for (i = 0; i < NCPU; tsbmissp++, i++) {
12999 		/*
13000 		 * initialize the tsbmiss area.
13001 		 * Do this for all possible CPUs as some may be added
13002 		 * while the system is running. There is no cost to this.
13003 		 */
13004 		tsbmissp->ksfmmup = ksfmmup;
13005 #ifndef sun4v
13006 		tsbmissp->dcache_line_mask = (uint16_t)dcache_line_mask;
13007 #endif /* sun4v */
13008 		tsbmissp->khashstart =
13009 		    (struct hmehash_bucket *)va_to_pa((caddr_t)khme_hash);
13010 		tsbmissp->uhashstart =
13011 		    (struct hmehash_bucket *)va_to_pa((caddr_t)uhme_hash);
13012 		tsbmissp->khashsz = khmehash_num;
13013 		tsbmissp->uhashsz = uhmehash_num;
13014 	}
13015 
13016 	sfmmu_tsb_cb_id = hat_register_callback('T'<<16 | 'S' << 8 | 'B',
13017 	    sfmmu_tsb_pre_relocator, sfmmu_tsb_post_relocator, NULL, 0);
13018 
13019 	if (kpm_enable == 0)
13020 		return;
13021 
13022 	/* -- Begin KPM specific init -- */
13023 
13024 	if (kpm_smallpages) {
13025 		/*
13026 		 * If we're using base pagesize pages for seg_kpm
13027 		 * mappings, we use the kernel TSB since we can't afford
13028 		 * to allocate a second huge TSB for these mappings.
13029 		 */
13030 		kpm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13031 		kpm_tsbsz = ktsb_szcode;
13032 		kpmsm_tsbbase = kpm_tsbbase;
13033 		kpmsm_tsbsz = kpm_tsbsz;
13034 	} else {
13035 		/*
13036 		 * In VAC conflict case, just put the entries in the
13037 		 * kernel 8K indexed TSB for now so we can find them.
13038 		 * This could really be changed in the future if we feel
13039 		 * the need...
13040 		 */
13041 		kpmsm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13042 		kpmsm_tsbsz = ktsb_szcode;
13043 		kpm_tsbbase = ktsb_phys? ktsb4m_pbase : (uint64_t)ktsb4m_base;
13044 		kpm_tsbsz = ktsb4m_szcode;
13045 	}
13046 
13047 	kpmtsbmp = kpmtsbm_area;
13048 	for (i = 0; i < NCPU; kpmtsbmp++, i++) {
13049 		/*
13050 		 * Initialize the kpmtsbm area.
13051 		 * Do this for all possible CPUs as some may be added
13052 		 * while the system is running. There is no cost to this.
13053 		 */
13054 		kpmtsbmp->vbase = kpm_vbase;
13055 		kpmtsbmp->vend = kpm_vbase + kpm_size * vac_colors;
13056 		kpmtsbmp->sz_shift = kpm_size_shift;
13057 		kpmtsbmp->kpmp_shift = kpmp_shift;
13058 		kpmtsbmp->kpmp2pshft = (uchar_t)kpmp2pshft;
13059 		if (kpm_smallpages == 0) {
13060 			kpmtsbmp->kpmp_table_sz = kpmp_table_sz;
13061 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_table);
13062 		} else {
13063 			kpmtsbmp->kpmp_table_sz = kpmp_stable_sz;
13064 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_stable);
13065 		}
13066 		kpmtsbmp->msegphashpa = va_to_pa(memseg_phash);
13067 		kpmtsbmp->flags = KPMTSBM_ENABLE_FLAG;
13068 #ifdef	DEBUG
13069 		kpmtsbmp->flags |= (kpm_tsbmtl) ?  KPMTSBM_TLTSBM_FLAG : 0;
13070 #endif	/* DEBUG */
13071 		if (ktsb_phys)
13072 			kpmtsbmp->flags |= KPMTSBM_TSBPHYS_FLAG;
13073 	}
13074 
13075 	/* -- End KPM specific init -- */
13076 }
13077 
13078 /* Avoid using sfmmu_tsbinfo_alloc() to avoid kmem_alloc - no real reason */
13079 struct tsb_info ktsb_info[2];
13080 
13081 /*
13082  * Called from hat_kern_setup() to setup the tsb_info for ksfmmup.
13083  */
13084 void
13085 sfmmu_init_ktsbinfo()
13086 {
13087 	ASSERT(ksfmmup != NULL);
13088 	ASSERT(ksfmmup->sfmmu_tsb == NULL);
13089 	/*
13090 	 * Allocate tsbinfos for kernel and copy in data
13091 	 * to make debug easier and sun4v setup easier.
13092 	 */
13093 	ktsb_info[0].tsb_sfmmu = ksfmmup;
13094 	ktsb_info[0].tsb_szc = ktsb_szcode;
13095 	ktsb_info[0].tsb_ttesz_mask = TSB8K|TSB64K|TSB512K;
13096 	ktsb_info[0].tsb_va = ktsb_base;
13097 	ktsb_info[0].tsb_pa = ktsb_pbase;
13098 	ktsb_info[0].tsb_flags = 0;
13099 	ktsb_info[0].tsb_tte.ll = 0;
13100 	ktsb_info[0].tsb_cache = NULL;
13101 
13102 	ktsb_info[1].tsb_sfmmu = ksfmmup;
13103 	ktsb_info[1].tsb_szc = ktsb4m_szcode;
13104 	ktsb_info[1].tsb_ttesz_mask = TSB4M;
13105 	ktsb_info[1].tsb_va = ktsb4m_base;
13106 	ktsb_info[1].tsb_pa = ktsb4m_pbase;
13107 	ktsb_info[1].tsb_flags = 0;
13108 	ktsb_info[1].tsb_tte.ll = 0;
13109 	ktsb_info[1].tsb_cache = NULL;
13110 
13111 	/* Link them into ksfmmup. */
13112 	ktsb_info[0].tsb_next = &ktsb_info[1];
13113 	ktsb_info[1].tsb_next = NULL;
13114 	ksfmmup->sfmmu_tsb = &ktsb_info[0];
13115 
13116 	sfmmu_setup_tsbinfo(ksfmmup);
13117 }
13118 
13119 /*
13120  * Cache the last value returned from va_to_pa().  If the VA specified
13121  * in the current call to cached_va_to_pa() maps to the same Page (as the
13122  * previous call to cached_va_to_pa()), then compute the PA using
13123  * cached info, else call va_to_pa().
13124  *
13125  * Note: this function is neither MT-safe nor consistent in the presence
13126  * of multiple, interleaved threads.  This function was created to enable
13127  * an optimization used during boot (at a point when there's only one thread
13128  * executing on the "boot CPU", and before startup_vm() has been called).
13129  */
13130 static uint64_t
13131 cached_va_to_pa(void *vaddr)
13132 {
13133 	static uint64_t prev_vaddr_base = 0;
13134 	static uint64_t prev_pfn = 0;
13135 
13136 	if ((((uint64_t)vaddr) & MMU_PAGEMASK) == prev_vaddr_base) {
13137 		return (prev_pfn | ((uint64_t)vaddr & MMU_PAGEOFFSET));
13138 	} else {
13139 		uint64_t pa = va_to_pa(vaddr);
13140 
13141 		if (pa != ((uint64_t)-1)) {
13142 			/*
13143 			 * Computed physical address is valid.  Cache its
13144 			 * related info for the next cached_va_to_pa() call.
13145 			 */
13146 			prev_pfn = pa & MMU_PAGEMASK;
13147 			prev_vaddr_base = ((uint64_t)vaddr) & MMU_PAGEMASK;
13148 		}
13149 
13150 		return (pa);
13151 	}
13152 }
13153 
13154 /*
13155  * Carve up our nucleus hblk region.  We may allocate more hblks than
13156  * asked due to rounding errors but we are guaranteed to have at least
13157  * enough space to allocate the requested number of hblk8's and hblk1's.
13158  */
13159 void
13160 sfmmu_init_nucleus_hblks(caddr_t addr, size_t size, int nhblk8, int nhblk1)
13161 {
13162 	struct hme_blk *hmeblkp;
13163 	size_t hme8blk_sz, hme1blk_sz;
13164 	size_t i;
13165 	size_t hblk8_bound;
13166 	ulong_t j = 0, k = 0;
13167 
13168 	ASSERT(addr != NULL && size != 0);
13169 
13170 	/* Need to use proper structure alignment */
13171 	hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t));
13172 	hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t));
13173 
13174 	nucleus_hblk8.list = (void *)addr;
13175 	nucleus_hblk8.index = 0;
13176 
13177 	/*
13178 	 * Use as much memory as possible for hblk8's since we
13179 	 * expect all bop_alloc'ed memory to be allocated in 8k chunks.
13180 	 * We need to hold back enough space for the hblk1's which
13181 	 * we'll allocate next.
13182 	 */
13183 	hblk8_bound = size - (nhblk1 * hme1blk_sz) - hme8blk_sz;
13184 	for (i = 0; i <= hblk8_bound; i += hme8blk_sz, j++) {
13185 		hmeblkp = (struct hme_blk *)addr;
13186 		addr += hme8blk_sz;
13187 		hmeblkp->hblk_nuc_bit = 1;
13188 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13189 	}
13190 	nucleus_hblk8.len = j;
13191 	ASSERT(j >= nhblk8);
13192 	SFMMU_STAT_ADD(sf_hblk8_ncreate, j);
13193 
13194 	nucleus_hblk1.list = (void *)addr;
13195 	nucleus_hblk1.index = 0;
13196 	for (; i <= (size - hme1blk_sz); i += hme1blk_sz, k++) {
13197 		hmeblkp = (struct hme_blk *)addr;
13198 		addr += hme1blk_sz;
13199 		hmeblkp->hblk_nuc_bit = 1;
13200 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13201 	}
13202 	ASSERT(k >= nhblk1);
13203 	nucleus_hblk1.len = k;
13204 	SFMMU_STAT_ADD(sf_hblk1_ncreate, k);
13205 }
13206 
13207 /*
13208  * This function is currently not supported on this platform. For what
13209  * it's supposed to do, see hat.c and hat_srmmu.c
13210  */
13211 /* ARGSUSED */
13212 faultcode_t
13213 hat_softlock(struct hat *hat, caddr_t addr, size_t *lenp, page_t **ppp,
13214     uint_t flags)
13215 {
13216 	return (FC_NOSUPPORT);
13217 }
13218 
13219 /*
13220  * Searchs the mapping list of the page for a mapping of the same size. If not
13221  * found the corresponding bit is cleared in the p_index field. When large
13222  * pages are more prevalent in the system, we can maintain the mapping list
13223  * in order and we don't have to traverse the list each time. Just check the
13224  * next and prev entries, and if both are of different size, we clear the bit.
13225  */
13226 static void
13227 sfmmu_rm_large_mappings(page_t *pp, int ttesz)
13228 {
13229 	struct sf_hment *sfhmep;
13230 	struct hme_blk *hmeblkp;
13231 	int	index;
13232 	pgcnt_t	npgs;
13233 
13234 	ASSERT(ttesz > TTE8K);
13235 
13236 	ASSERT(sfmmu_mlist_held(pp));
13237 
13238 	ASSERT(PP_ISMAPPED_LARGE(pp));
13239 
13240 	/*
13241 	 * Traverse mapping list looking for another mapping of same size.
13242 	 * since we only want to clear index field if all mappings of
13243 	 * that size are gone.
13244 	 */
13245 
13246 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
13247 		if (IS_PAHME(sfhmep))
13248 			continue;
13249 		hmeblkp = sfmmu_hmetohblk(sfhmep);
13250 		if (hme_size(sfhmep) == ttesz) {
13251 			/*
13252 			 * another mapping of the same size. don't clear index.
13253 			 */
13254 			return;
13255 		}
13256 	}
13257 
13258 	/*
13259 	 * Clear the p_index bit for large page.
13260 	 */
13261 	index = PAGESZ_TO_INDEX(ttesz);
13262 	npgs = TTEPAGES(ttesz);
13263 	while (npgs-- > 0) {
13264 		ASSERT(pp->p_index & index);
13265 		pp->p_index &= ~index;
13266 		pp = PP_PAGENEXT(pp);
13267 	}
13268 }
13269 
13270 /*
13271  * return supported features
13272  */
13273 /* ARGSUSED */
13274 int
13275 hat_supported(enum hat_features feature, void *arg)
13276 {
13277 	switch (feature) {
13278 	case    HAT_SHARED_PT:
13279 	case	HAT_DYNAMIC_ISM_UNMAP:
13280 	case	HAT_VMODSORT:
13281 		return (1);
13282 	case	HAT_SHARED_REGIONS:
13283 		if (shctx_on)
13284 			return (1);
13285 		else
13286 			return (0);
13287 	default:
13288 		return (0);
13289 	}
13290 }
13291 
13292 void
13293 hat_enter(struct hat *hat)
13294 {
13295 	hatlock_t	*hatlockp;
13296 
13297 	if (hat != ksfmmup) {
13298 		hatlockp = TSB_HASH(hat);
13299 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
13300 	}
13301 }
13302 
13303 void
13304 hat_exit(struct hat *hat)
13305 {
13306 	hatlock_t	*hatlockp;
13307 
13308 	if (hat != ksfmmup) {
13309 		hatlockp = TSB_HASH(hat);
13310 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
13311 	}
13312 }
13313 
13314 /*ARGSUSED*/
13315 void
13316 hat_reserve(struct as *as, caddr_t addr, size_t len)
13317 {
13318 }
13319 
13320 static void
13321 hat_kstat_init(void)
13322 {
13323 	kstat_t *ksp;
13324 
13325 	ksp = kstat_create("unix", 0, "sfmmu_global_stat", "hat",
13326 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_global_stat),
13327 	    KSTAT_FLAG_VIRTUAL);
13328 	if (ksp) {
13329 		ksp->ks_data = (void *) &sfmmu_global_stat;
13330 		kstat_install(ksp);
13331 	}
13332 	ksp = kstat_create("unix", 0, "sfmmu_tsbsize_stat", "hat",
13333 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_tsbsize_stat),
13334 	    KSTAT_FLAG_VIRTUAL);
13335 	if (ksp) {
13336 		ksp->ks_data = (void *) &sfmmu_tsbsize_stat;
13337 		kstat_install(ksp);
13338 	}
13339 	ksp = kstat_create("unix", 0, "sfmmu_percpu_stat", "hat",
13340 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_percpu_stat) * NCPU,
13341 	    KSTAT_FLAG_WRITABLE);
13342 	if (ksp) {
13343 		ksp->ks_update = sfmmu_kstat_percpu_update;
13344 		kstat_install(ksp);
13345 	}
13346 }
13347 
13348 /* ARGSUSED */
13349 static int
13350 sfmmu_kstat_percpu_update(kstat_t *ksp, int rw)
13351 {
13352 	struct sfmmu_percpu_stat *cpu_kstat = ksp->ks_data;
13353 	struct tsbmiss *tsbm = tsbmiss_area;
13354 	struct kpmtsbm *kpmtsbm = kpmtsbm_area;
13355 	int i;
13356 
13357 	ASSERT(cpu_kstat);
13358 	if (rw == KSTAT_READ) {
13359 		for (i = 0; i < NCPU; cpu_kstat++, tsbm++, kpmtsbm++, i++) {
13360 			cpu_kstat->sf_itlb_misses = 0;
13361 			cpu_kstat->sf_dtlb_misses = 0;
13362 			cpu_kstat->sf_utsb_misses = tsbm->utsb_misses -
13363 			    tsbm->uprot_traps;
13364 			cpu_kstat->sf_ktsb_misses = tsbm->ktsb_misses +
13365 			    kpmtsbm->kpm_tsb_misses - tsbm->kprot_traps;
13366 			cpu_kstat->sf_tsb_hits = 0;
13367 			cpu_kstat->sf_umod_faults = tsbm->uprot_traps;
13368 			cpu_kstat->sf_kmod_faults = tsbm->kprot_traps;
13369 		}
13370 	} else {
13371 		/* KSTAT_WRITE is used to clear stats */
13372 		for (i = 0; i < NCPU; tsbm++, kpmtsbm++, i++) {
13373 			tsbm->utsb_misses = 0;
13374 			tsbm->ktsb_misses = 0;
13375 			tsbm->uprot_traps = 0;
13376 			tsbm->kprot_traps = 0;
13377 			kpmtsbm->kpm_dtlb_misses = 0;
13378 			kpmtsbm->kpm_tsb_misses = 0;
13379 		}
13380 	}
13381 	return (0);
13382 }
13383 
13384 #ifdef	DEBUG
13385 
13386 tte_t  *gorig[NCPU], *gcur[NCPU], *gnew[NCPU];
13387 
13388 /*
13389  * A tte checker. *orig_old is the value we read before cas.
13390  *	*cur is the value returned by cas.
13391  *	*new is the desired value when we do the cas.
13392  *
13393  *	*hmeblkp is currently unused.
13394  */
13395 
13396 /* ARGSUSED */
13397 void
13398 chk_tte(tte_t *orig_old, tte_t *cur, tte_t *new, struct hme_blk *hmeblkp)
13399 {
13400 	pfn_t i, j, k;
13401 	int cpuid = CPU->cpu_id;
13402 
13403 	gorig[cpuid] = orig_old;
13404 	gcur[cpuid] = cur;
13405 	gnew[cpuid] = new;
13406 
13407 #ifdef lint
13408 	hmeblkp = hmeblkp;
13409 #endif
13410 
13411 	if (TTE_IS_VALID(orig_old)) {
13412 		if (TTE_IS_VALID(cur)) {
13413 			i = TTE_TO_TTEPFN(orig_old);
13414 			j = TTE_TO_TTEPFN(cur);
13415 			k = TTE_TO_TTEPFN(new);
13416 			if (i != j) {
13417 				/* remap error? */
13418 				panic("chk_tte: bad pfn, 0x%lx, 0x%lx", i, j);
13419 			}
13420 
13421 			if (i != k) {
13422 				/* remap error? */
13423 				panic("chk_tte: bad pfn2, 0x%lx, 0x%lx", i, k);
13424 			}
13425 		} else {
13426 			if (TTE_IS_VALID(new)) {
13427 				panic("chk_tte: invalid cur? ");
13428 			}
13429 
13430 			i = TTE_TO_TTEPFN(orig_old);
13431 			k = TTE_TO_TTEPFN(new);
13432 			if (i != k) {
13433 				panic("chk_tte: bad pfn3, 0x%lx, 0x%lx", i, k);
13434 			}
13435 		}
13436 	} else {
13437 		if (TTE_IS_VALID(cur)) {
13438 			j = TTE_TO_TTEPFN(cur);
13439 			if (TTE_IS_VALID(new)) {
13440 				k = TTE_TO_TTEPFN(new);
13441 				if (j != k) {
13442 					panic("chk_tte: bad pfn4, 0x%lx, 0x%lx",
13443 					    j, k);
13444 				}
13445 			} else {
13446 				panic("chk_tte: why here?");
13447 			}
13448 		} else {
13449 			if (!TTE_IS_VALID(new)) {
13450 				panic("chk_tte: why here2 ?");
13451 			}
13452 		}
13453 	}
13454 }
13455 
13456 #endif /* DEBUG */
13457 
13458 extern void prefetch_tsbe_read(struct tsbe *);
13459 extern void prefetch_tsbe_write(struct tsbe *);
13460 
13461 
13462 /*
13463  * We want to prefetch 7 cache lines ahead for our read prefetch.  This gives
13464  * us optimal performance on Cheetah+.  You can only have 8 outstanding
13465  * prefetches at any one time, so we opted for 7 read prefetches and 1 write
13466  * prefetch to make the most utilization of the prefetch capability.
13467  */
13468 #define	TSBE_PREFETCH_STRIDE (7)
13469 
13470 void
13471 sfmmu_copy_tsb(struct tsb_info *old_tsbinfo, struct tsb_info *new_tsbinfo)
13472 {
13473 	int old_bytes = TSB_BYTES(old_tsbinfo->tsb_szc);
13474 	int new_bytes = TSB_BYTES(new_tsbinfo->tsb_szc);
13475 	int old_entries = TSB_ENTRIES(old_tsbinfo->tsb_szc);
13476 	int new_entries = TSB_ENTRIES(new_tsbinfo->tsb_szc);
13477 	struct tsbe *old;
13478 	struct tsbe *new;
13479 	struct tsbe *new_base = (struct tsbe *)new_tsbinfo->tsb_va;
13480 	uint64_t va;
13481 	int new_offset;
13482 	int i;
13483 	int vpshift;
13484 	int last_prefetch;
13485 
13486 	if (old_bytes == new_bytes) {
13487 		bcopy(old_tsbinfo->tsb_va, new_tsbinfo->tsb_va, new_bytes);
13488 	} else {
13489 
13490 		/*
13491 		 * A TSBE is 16 bytes which means there are four TSBE's per
13492 		 * P$ line (64 bytes), thus every 4 TSBE's we prefetch.
13493 		 */
13494 		old = (struct tsbe *)old_tsbinfo->tsb_va;
13495 		last_prefetch = old_entries - (4*(TSBE_PREFETCH_STRIDE+1));
13496 		for (i = 0; i < old_entries; i++, old++) {
13497 			if (((i & (4-1)) == 0) && (i < last_prefetch))
13498 				prefetch_tsbe_read(old);
13499 			if (!old->tte_tag.tag_invalid) {
13500 				/*
13501 				 * We have a valid TTE to remap.  Check the
13502 				 * size.  We won't remap 64K or 512K TTEs
13503 				 * because they span more than one TSB entry
13504 				 * and are indexed using an 8K virt. page.
13505 				 * Ditto for 32M and 256M TTEs.
13506 				 */
13507 				if (TTE_CSZ(&old->tte_data) == TTE64K ||
13508 				    TTE_CSZ(&old->tte_data) == TTE512K)
13509 					continue;
13510 				if (mmu_page_sizes == max_mmu_page_sizes) {
13511 					if (TTE_CSZ(&old->tte_data) == TTE32M ||
13512 					    TTE_CSZ(&old->tte_data) == TTE256M)
13513 						continue;
13514 				}
13515 
13516 				/* clear the lower 22 bits of the va */
13517 				va = *(uint64_t *)old << 22;
13518 				/* turn va into a virtual pfn */
13519 				va >>= 22 - TSB_START_SIZE;
13520 				/*
13521 				 * or in bits from the offset in the tsb
13522 				 * to get the real virtual pfn. These
13523 				 * correspond to bits [21:13] in the va
13524 				 */
13525 				vpshift =
13526 				    TTE_BSZS_SHIFT(TTE_CSZ(&old->tte_data)) &
13527 				    0x1ff;
13528 				va |= (i << vpshift);
13529 				va >>= vpshift;
13530 				new_offset = va & (new_entries - 1);
13531 				new = new_base + new_offset;
13532 				prefetch_tsbe_write(new);
13533 				*new = *old;
13534 			}
13535 		}
13536 	}
13537 }
13538 
13539 /*
13540  * unused in sfmmu
13541  */
13542 void
13543 hat_dump(void)
13544 {
13545 }
13546 
13547 /*
13548  * Called when a thread is exiting and we have switched to the kernel address
13549  * space.  Perform the same VM initialization resume() uses when switching
13550  * processes.
13551  *
13552  * Note that sfmmu_load_mmustate() is currently a no-op for kernel threads, but
13553  * we call it anyway in case the semantics change in the future.
13554  */
13555 /*ARGSUSED*/
13556 void
13557 hat_thread_exit(kthread_t *thd)
13558 {
13559 	uint_t pgsz_cnum;
13560 	uint_t pstate_save;
13561 
13562 	ASSERT(thd->t_procp->p_as == &kas);
13563 
13564 	pgsz_cnum = KCONTEXT;
13565 #ifdef sun4u
13566 	pgsz_cnum |= (ksfmmup->sfmmu_cext << CTXREG_EXT_SHIFT);
13567 #endif
13568 
13569 	/*
13570 	 * Note that sfmmu_load_mmustate() is currently a no-op for
13571 	 * kernel threads. We need to disable interrupts here,
13572 	 * simply because otherwise sfmmu_load_mmustate() would panic
13573 	 * if the caller does not disable interrupts.
13574 	 */
13575 	pstate_save = sfmmu_disable_intrs();
13576 
13577 	/* Compatibility Note: hw takes care of MMU_SCONTEXT1 */
13578 	sfmmu_setctx_sec(pgsz_cnum);
13579 	sfmmu_load_mmustate(ksfmmup);
13580 	sfmmu_enable_intrs(pstate_save);
13581 }
13582 
13583 
13584 /*
13585  * SRD support
13586  */
13587 #define	SRD_HASH_FUNCTION(vp)	(((((uintptr_t)(vp)) >> 4) ^ \
13588 				    (((uintptr_t)(vp)) >> 11)) & \
13589 				    srd_hashmask)
13590 
13591 /*
13592  * Attach the process to the srd struct associated with the exec vnode
13593  * from which the process is started.
13594  */
13595 void
13596 hat_join_srd(struct hat *sfmmup, vnode_t *evp)
13597 {
13598 	uint_t hash = SRD_HASH_FUNCTION(evp);
13599 	sf_srd_t *srdp;
13600 	sf_srd_t *newsrdp;
13601 
13602 	ASSERT(sfmmup != ksfmmup);
13603 	ASSERT(sfmmup->sfmmu_srdp == NULL);
13604 
13605 	if (!shctx_on) {
13606 		return;
13607 	}
13608 
13609 	VN_HOLD(evp);
13610 
13611 	if (srd_buckets[hash].srdb_srdp != NULL) {
13612 		mutex_enter(&srd_buckets[hash].srdb_lock);
13613 		for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13614 		    srdp = srdp->srd_hash) {
13615 			if (srdp->srd_evp == evp) {
13616 				ASSERT(srdp->srd_refcnt >= 0);
13617 				sfmmup->sfmmu_srdp = srdp;
13618 				atomic_inc_32(
13619 				    (volatile uint_t *)&srdp->srd_refcnt);
13620 				mutex_exit(&srd_buckets[hash].srdb_lock);
13621 				return;
13622 			}
13623 		}
13624 		mutex_exit(&srd_buckets[hash].srdb_lock);
13625 	}
13626 	newsrdp = kmem_cache_alloc(srd_cache, KM_SLEEP);
13627 	ASSERT(newsrdp->srd_next_ismrid == 0 && newsrdp->srd_next_hmerid == 0);
13628 
13629 	newsrdp->srd_evp = evp;
13630 	newsrdp->srd_refcnt = 1;
13631 	newsrdp->srd_hmergnfree = NULL;
13632 	newsrdp->srd_ismrgnfree = NULL;
13633 
13634 	mutex_enter(&srd_buckets[hash].srdb_lock);
13635 	for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13636 	    srdp = srdp->srd_hash) {
13637 		if (srdp->srd_evp == evp) {
13638 			ASSERT(srdp->srd_refcnt >= 0);
13639 			sfmmup->sfmmu_srdp = srdp;
13640 			atomic_inc_32((volatile uint_t *)&srdp->srd_refcnt);
13641 			mutex_exit(&srd_buckets[hash].srdb_lock);
13642 			kmem_cache_free(srd_cache, newsrdp);
13643 			return;
13644 		}
13645 	}
13646 	newsrdp->srd_hash = srd_buckets[hash].srdb_srdp;
13647 	srd_buckets[hash].srdb_srdp = newsrdp;
13648 	sfmmup->sfmmu_srdp = newsrdp;
13649 
13650 	mutex_exit(&srd_buckets[hash].srdb_lock);
13651 
13652 }
13653 
13654 static void
13655 sfmmu_leave_srd(sfmmu_t *sfmmup)
13656 {
13657 	vnode_t *evp;
13658 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
13659 	uint_t hash;
13660 	sf_srd_t **prev_srdpp;
13661 	sf_region_t *rgnp;
13662 	sf_region_t *nrgnp;
13663 #ifdef DEBUG
13664 	int rgns = 0;
13665 #endif
13666 	int i;
13667 
13668 	ASSERT(sfmmup != ksfmmup);
13669 	ASSERT(srdp != NULL);
13670 	ASSERT(srdp->srd_refcnt > 0);
13671 	ASSERT(sfmmup->sfmmu_scdp == NULL);
13672 	ASSERT(sfmmup->sfmmu_free == 1);
13673 
13674 	sfmmup->sfmmu_srdp = NULL;
13675 	evp = srdp->srd_evp;
13676 	ASSERT(evp != NULL);
13677 	if (atomic_dec_32_nv((volatile uint_t *)&srdp->srd_refcnt)) {
13678 		VN_RELE(evp);
13679 		return;
13680 	}
13681 
13682 	hash = SRD_HASH_FUNCTION(evp);
13683 	mutex_enter(&srd_buckets[hash].srdb_lock);
13684 	for (prev_srdpp = &srd_buckets[hash].srdb_srdp;
13685 	    (srdp = *prev_srdpp) != NULL; prev_srdpp = &srdp->srd_hash) {
13686 		if (srdp->srd_evp == evp) {
13687 			break;
13688 		}
13689 	}
13690 	if (srdp == NULL || srdp->srd_refcnt) {
13691 		mutex_exit(&srd_buckets[hash].srdb_lock);
13692 		VN_RELE(evp);
13693 		return;
13694 	}
13695 	*prev_srdpp = srdp->srd_hash;
13696 	mutex_exit(&srd_buckets[hash].srdb_lock);
13697 
13698 	ASSERT(srdp->srd_refcnt == 0);
13699 	VN_RELE(evp);
13700 
13701 #ifdef DEBUG
13702 	for (i = 0; i < SFMMU_MAX_REGION_BUCKETS; i++) {
13703 		ASSERT(srdp->srd_rgnhash[i] == NULL);
13704 	}
13705 #endif /* DEBUG */
13706 
13707 	/* free each hme regions in the srd */
13708 	for (rgnp = srdp->srd_hmergnfree; rgnp != NULL; rgnp = nrgnp) {
13709 		nrgnp = rgnp->rgn_next;
13710 		ASSERT(rgnp->rgn_id < srdp->srd_next_hmerid);
13711 		ASSERT(rgnp->rgn_refcnt == 0);
13712 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13713 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13714 		ASSERT(rgnp->rgn_hmeflags == 0);
13715 		ASSERT(srdp->srd_hmergnp[rgnp->rgn_id] == rgnp);
13716 #ifdef DEBUG
13717 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13718 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13719 		}
13720 		rgns++;
13721 #endif /* DEBUG */
13722 		kmem_cache_free(region_cache, rgnp);
13723 	}
13724 	ASSERT(rgns == srdp->srd_next_hmerid);
13725 
13726 #ifdef DEBUG
13727 	rgns = 0;
13728 #endif
13729 	/* free each ism rgns in the srd */
13730 	for (rgnp = srdp->srd_ismrgnfree; rgnp != NULL; rgnp = nrgnp) {
13731 		nrgnp = rgnp->rgn_next;
13732 		ASSERT(rgnp->rgn_id < srdp->srd_next_ismrid);
13733 		ASSERT(rgnp->rgn_refcnt == 0);
13734 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13735 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13736 		ASSERT(srdp->srd_ismrgnp[rgnp->rgn_id] == rgnp);
13737 #ifdef DEBUG
13738 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13739 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13740 		}
13741 		rgns++;
13742 #endif /* DEBUG */
13743 		kmem_cache_free(region_cache, rgnp);
13744 	}
13745 	ASSERT(rgns == srdp->srd_next_ismrid);
13746 	ASSERT(srdp->srd_ismbusyrgns == 0);
13747 	ASSERT(srdp->srd_hmebusyrgns == 0);
13748 
13749 	srdp->srd_next_ismrid = 0;
13750 	srdp->srd_next_hmerid = 0;
13751 
13752 	bzero((void *)srdp->srd_ismrgnp,
13753 	    sizeof (sf_region_t *) * SFMMU_MAX_ISM_REGIONS);
13754 	bzero((void *)srdp->srd_hmergnp,
13755 	    sizeof (sf_region_t *) * SFMMU_MAX_HME_REGIONS);
13756 
13757 	ASSERT(srdp->srd_scdp == NULL);
13758 	kmem_cache_free(srd_cache, srdp);
13759 }
13760 
13761 /* ARGSUSED */
13762 static int
13763 sfmmu_srdcache_constructor(void *buf, void *cdrarg, int kmflags)
13764 {
13765 	sf_srd_t *srdp = (sf_srd_t *)buf;
13766 	bzero(buf, sizeof (*srdp));
13767 
13768 	mutex_init(&srdp->srd_mutex, NULL, MUTEX_DEFAULT, NULL);
13769 	mutex_init(&srdp->srd_scd_mutex, NULL, MUTEX_DEFAULT, NULL);
13770 	return (0);
13771 }
13772 
13773 /* ARGSUSED */
13774 static void
13775 sfmmu_srdcache_destructor(void *buf, void *cdrarg)
13776 {
13777 	sf_srd_t *srdp = (sf_srd_t *)buf;
13778 
13779 	mutex_destroy(&srdp->srd_mutex);
13780 	mutex_destroy(&srdp->srd_scd_mutex);
13781 }
13782 
13783 /*
13784  * The caller makes sure hat_join_region()/hat_leave_region() can't be called
13785  * at the same time for the same process and address range. This is ensured by
13786  * the fact that address space is locked as writer when a process joins the
13787  * regions. Therefore there's no need to hold an srd lock during the entire
13788  * execution of hat_join_region()/hat_leave_region().
13789  */
13790 
13791 #define	RGN_HASH_FUNCTION(obj)	(((((uintptr_t)(obj)) >> 4) ^ \
13792 				    (((uintptr_t)(obj)) >> 11)) & \
13793 					srd_rgn_hashmask)
13794 /*
13795  * This routine implements the shared context functionality required when
13796  * attaching a segment to an address space. It must be called from
13797  * hat_share() for D(ISM) segments and from segvn_create() for segments
13798  * with the MAP_PRIVATE and MAP_TEXT flags set. It returns a region_cookie
13799  * which is saved in the private segment data for hme segments and
13800  * the ism_map structure for ism segments.
13801  */
13802 hat_region_cookie_t
13803 hat_join_region(struct hat *sfmmup, caddr_t r_saddr, size_t r_size,
13804     void *r_obj, u_offset_t r_objoff, uchar_t r_perm, uchar_t r_pgszc,
13805     hat_rgn_cb_func_t r_cb_function, uint_t flags)
13806 {
13807 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
13808 	uint_t rhash;
13809 	uint_t rid;
13810 	hatlock_t *hatlockp;
13811 	sf_region_t *rgnp;
13812 	sf_region_t *new_rgnp = NULL;
13813 	int i;
13814 	uint16_t *nextidp;
13815 	sf_region_t **freelistp;
13816 	int maxids;
13817 	sf_region_t **rarrp;
13818 	uint16_t *busyrgnsp;
13819 	ulong_t rttecnt;
13820 	uchar_t tteflag;
13821 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
13822 	int text = (r_type == HAT_REGION_TEXT);
13823 
13824 	if (srdp == NULL || r_size == 0) {
13825 		return (HAT_INVALID_REGION_COOKIE);
13826 	}
13827 
13828 	ASSERT(sfmmup != ksfmmup);
13829 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as));
13830 	ASSERT(srdp->srd_refcnt > 0);
13831 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
13832 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
13833 	ASSERT(r_pgszc < mmu_page_sizes);
13834 	if (!IS_P2ALIGNED(r_saddr, TTEBYTES(r_pgszc)) ||
13835 	    !IS_P2ALIGNED(r_size, TTEBYTES(r_pgszc))) {
13836 		panic("hat_join_region: region addr or size is not aligned\n");
13837 	}
13838 
13839 
13840 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
13841 	    SFMMU_REGION_HME;
13842 	/*
13843 	 * Currently only support shared hmes for the read only main text
13844 	 * region.
13845 	 */
13846 	if (r_type == SFMMU_REGION_HME && ((r_obj != srdp->srd_evp) ||
13847 	    (r_perm & PROT_WRITE))) {
13848 		return (HAT_INVALID_REGION_COOKIE);
13849 	}
13850 
13851 	rhash = RGN_HASH_FUNCTION(r_obj);
13852 
13853 	if (r_type == SFMMU_REGION_ISM) {
13854 		nextidp = &srdp->srd_next_ismrid;
13855 		freelistp = &srdp->srd_ismrgnfree;
13856 		maxids = SFMMU_MAX_ISM_REGIONS;
13857 		rarrp = srdp->srd_ismrgnp;
13858 		busyrgnsp = &srdp->srd_ismbusyrgns;
13859 	} else {
13860 		nextidp = &srdp->srd_next_hmerid;
13861 		freelistp = &srdp->srd_hmergnfree;
13862 		maxids = SFMMU_MAX_HME_REGIONS;
13863 		rarrp = srdp->srd_hmergnp;
13864 		busyrgnsp = &srdp->srd_hmebusyrgns;
13865 	}
13866 
13867 	mutex_enter(&srdp->srd_mutex);
13868 
13869 	for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
13870 	    rgnp = rgnp->rgn_hash) {
13871 		if (rgnp->rgn_saddr == r_saddr && rgnp->rgn_size == r_size &&
13872 		    rgnp->rgn_obj == r_obj && rgnp->rgn_objoff == r_objoff &&
13873 		    rgnp->rgn_perm == r_perm && rgnp->rgn_pgszc == r_pgszc) {
13874 			break;
13875 		}
13876 	}
13877 
13878 rfound:
13879 	if (rgnp != NULL) {
13880 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
13881 		ASSERT(rgnp->rgn_cb_function == r_cb_function);
13882 		ASSERT(rgnp->rgn_refcnt >= 0);
13883 		rid = rgnp->rgn_id;
13884 		ASSERT(rid < maxids);
13885 		ASSERT(rarrp[rid] == rgnp);
13886 		ASSERT(rid < *nextidp);
13887 		atomic_inc_32((volatile uint_t *)&rgnp->rgn_refcnt);
13888 		mutex_exit(&srdp->srd_mutex);
13889 		if (new_rgnp != NULL) {
13890 			kmem_cache_free(region_cache, new_rgnp);
13891 		}
13892 		if (r_type == SFMMU_REGION_HME) {
13893 			int myjoin =
13894 			    (sfmmup == astosfmmu(curthread->t_procp->p_as));
13895 
13896 			sfmmu_link_to_hmeregion(sfmmup, rgnp);
13897 			/*
13898 			 * bitmap should be updated after linking sfmmu on
13899 			 * region list so that pageunload() doesn't skip
13900 			 * TSB/TLB flush. As soon as bitmap is updated another
13901 			 * thread in this process can already start accessing
13902 			 * this region.
13903 			 */
13904 			/*
13905 			 * Normally ttecnt accounting is done as part of
13906 			 * pagefault handling. But a process may not take any
13907 			 * pagefaults on shared hmeblks created by some other
13908 			 * process. To compensate for this assume that the
13909 			 * entire region will end up faulted in using
13910 			 * the region's pagesize.
13911 			 *
13912 			 */
13913 			if (r_pgszc > TTE8K) {
13914 				tteflag = 1 << r_pgszc;
13915 				if (disable_large_pages & tteflag) {
13916 					tteflag = 0;
13917 				}
13918 			} else {
13919 				tteflag = 0;
13920 			}
13921 			if (tteflag && !(sfmmup->sfmmu_rtteflags & tteflag)) {
13922 				hatlockp = sfmmu_hat_enter(sfmmup);
13923 				sfmmup->sfmmu_rtteflags |= tteflag;
13924 				sfmmu_hat_exit(hatlockp);
13925 			}
13926 			hatlockp = sfmmu_hat_enter(sfmmup);
13927 
13928 			/*
13929 			 * Preallocate 1/4 of ttecnt's in 8K TSB for >= 4M
13930 			 * region to allow for large page allocation failure.
13931 			 */
13932 			if (r_pgszc >= TTE4M) {
13933 				sfmmup->sfmmu_tsb0_4minflcnt +=
13934 				    r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
13935 			}
13936 
13937 			/* update sfmmu_ttecnt with the shme rgn ttecnt */
13938 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
13939 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
13940 			    rttecnt);
13941 
13942 			if (text && r_pgszc >= TTE4M &&
13943 			    (tteflag || ((disable_large_pages >> TTE4M) &
13944 			    ((1 << (r_pgszc - TTE4M + 1)) - 1))) &&
13945 			    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
13946 				SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
13947 			}
13948 
13949 			sfmmu_hat_exit(hatlockp);
13950 			/*
13951 			 * On Panther we need to make sure TLB is programmed
13952 			 * to accept 32M/256M pages.  Call
13953 			 * sfmmu_check_page_sizes() now to make sure TLB is
13954 			 * setup before making hmeregions visible to other
13955 			 * threads.
13956 			 */
13957 			sfmmu_check_page_sizes(sfmmup, 1);
13958 			hatlockp = sfmmu_hat_enter(sfmmup);
13959 			SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
13960 
13961 			/*
13962 			 * if context is invalid tsb miss exception code will
13963 			 * call sfmmu_check_page_sizes() and update tsbmiss
13964 			 * area later.
13965 			 */
13966 			kpreempt_disable();
13967 			if (myjoin &&
13968 			    (sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
13969 			    != INVALID_CONTEXT)) {
13970 				struct tsbmiss *tsbmp;
13971 
13972 				tsbmp = &tsbmiss_area[CPU->cpu_id];
13973 				ASSERT(sfmmup == tsbmp->usfmmup);
13974 				BT_SET(tsbmp->shmermap, rid);
13975 				if (r_pgszc > TTE64K) {
13976 					tsbmp->uhat_rtteflags |= tteflag;
13977 				}
13978 
13979 			}
13980 			kpreempt_enable();
13981 
13982 			sfmmu_hat_exit(hatlockp);
13983 			ASSERT((hat_region_cookie_t)((uint64_t)rid) !=
13984 			    HAT_INVALID_REGION_COOKIE);
13985 		} else {
13986 			hatlockp = sfmmu_hat_enter(sfmmup);
13987 			SF_RGNMAP_ADD(sfmmup->sfmmu_ismregion_map, rid);
13988 			sfmmu_hat_exit(hatlockp);
13989 		}
13990 		ASSERT(rid < maxids);
13991 
13992 		if (r_type == SFMMU_REGION_ISM) {
13993 			sfmmu_find_scd(sfmmup);
13994 		}
13995 		return ((hat_region_cookie_t)((uint64_t)rid));
13996 	}
13997 
13998 	ASSERT(new_rgnp == NULL);
13999 
14000 	if (*busyrgnsp >= maxids) {
14001 		mutex_exit(&srdp->srd_mutex);
14002 		return (HAT_INVALID_REGION_COOKIE);
14003 	}
14004 
14005 	ASSERT(MUTEX_HELD(&srdp->srd_mutex));
14006 	if (*freelistp != NULL) {
14007 		rgnp = *freelistp;
14008 		*freelistp = rgnp->rgn_next;
14009 		ASSERT(rgnp->rgn_id < *nextidp);
14010 		ASSERT(rgnp->rgn_id < maxids);
14011 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
14012 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK)
14013 		    == r_type);
14014 		ASSERT(rarrp[rgnp->rgn_id] == rgnp);
14015 		ASSERT(rgnp->rgn_hmeflags == 0);
14016 	} else {
14017 		/*
14018 		 * release local locks before memory allocation.
14019 		 */
14020 		mutex_exit(&srdp->srd_mutex);
14021 
14022 		new_rgnp = kmem_cache_alloc(region_cache, KM_SLEEP);
14023 
14024 		mutex_enter(&srdp->srd_mutex);
14025 		for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
14026 		    rgnp = rgnp->rgn_hash) {
14027 			if (rgnp->rgn_saddr == r_saddr &&
14028 			    rgnp->rgn_size == r_size &&
14029 			    rgnp->rgn_obj == r_obj &&
14030 			    rgnp->rgn_objoff == r_objoff &&
14031 			    rgnp->rgn_perm == r_perm &&
14032 			    rgnp->rgn_pgszc == r_pgszc) {
14033 				break;
14034 			}
14035 		}
14036 		if (rgnp != NULL) {
14037 			goto rfound;
14038 		}
14039 
14040 		if (*nextidp >= maxids) {
14041 			mutex_exit(&srdp->srd_mutex);
14042 			goto fail;
14043 		}
14044 		rgnp = new_rgnp;
14045 		new_rgnp = NULL;
14046 		rgnp->rgn_id = (*nextidp)++;
14047 		ASSERT(rgnp->rgn_id < maxids);
14048 		ASSERT(rarrp[rgnp->rgn_id] == NULL);
14049 		rarrp[rgnp->rgn_id] = rgnp;
14050 	}
14051 
14052 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14053 	ASSERT(rgnp->rgn_hmeflags == 0);
14054 #ifdef DEBUG
14055 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14056 		ASSERT(rgnp->rgn_ttecnt[i] == 0);
14057 	}
14058 #endif
14059 	rgnp->rgn_saddr = r_saddr;
14060 	rgnp->rgn_size = r_size;
14061 	rgnp->rgn_obj = r_obj;
14062 	rgnp->rgn_objoff = r_objoff;
14063 	rgnp->rgn_perm = r_perm;
14064 	rgnp->rgn_pgszc = r_pgszc;
14065 	rgnp->rgn_flags = r_type;
14066 	rgnp->rgn_refcnt = 0;
14067 	rgnp->rgn_cb_function = r_cb_function;
14068 	rgnp->rgn_hash = srdp->srd_rgnhash[rhash];
14069 	srdp->srd_rgnhash[rhash] = rgnp;
14070 	(*busyrgnsp)++;
14071 	ASSERT(*busyrgnsp <= maxids);
14072 	goto rfound;
14073 
14074 fail:
14075 	ASSERT(new_rgnp != NULL);
14076 	kmem_cache_free(region_cache, new_rgnp);
14077 	return (HAT_INVALID_REGION_COOKIE);
14078 }
14079 
14080 /*
14081  * This function implements the shared context functionality required
14082  * when detaching a segment from an address space. It must be called
14083  * from hat_unshare() for all D(ISM) segments and from segvn_unmap(),
14084  * for segments with a valid region_cookie.
14085  * It will also be called from all seg_vn routines which change a
14086  * segment's attributes such as segvn_setprot(), segvn_setpagesize(),
14087  * segvn_clrszc() & segvn_advise(), as well as in the case of COW fault
14088  * from segvn_fault().
14089  */
14090 void
14091 hat_leave_region(struct hat *sfmmup, hat_region_cookie_t rcookie, uint_t flags)
14092 {
14093 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14094 	sf_scd_t *scdp;
14095 	uint_t rhash;
14096 	uint_t rid = (uint_t)((uint64_t)rcookie);
14097 	hatlock_t *hatlockp = NULL;
14098 	sf_region_t *rgnp;
14099 	sf_region_t **prev_rgnpp;
14100 	sf_region_t *cur_rgnp;
14101 	void *r_obj;
14102 	int i;
14103 	caddr_t	r_saddr;
14104 	caddr_t r_eaddr;
14105 	size_t	r_size;
14106 	uchar_t	r_pgszc;
14107 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
14108 
14109 	ASSERT(sfmmup != ksfmmup);
14110 	ASSERT(srdp != NULL);
14111 	ASSERT(srdp->srd_refcnt > 0);
14112 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
14113 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
14114 	ASSERT(!sfmmup->sfmmu_free || sfmmup->sfmmu_scdp == NULL);
14115 
14116 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
14117 	    SFMMU_REGION_HME;
14118 
14119 	if (r_type == SFMMU_REGION_ISM) {
14120 		ASSERT(SFMMU_IS_ISMRID_VALID(rid));
14121 		ASSERT(rid < SFMMU_MAX_ISM_REGIONS);
14122 		rgnp = srdp->srd_ismrgnp[rid];
14123 	} else {
14124 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14125 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14126 		rgnp = srdp->srd_hmergnp[rid];
14127 	}
14128 	ASSERT(rgnp != NULL);
14129 	ASSERT(rgnp->rgn_id == rid);
14130 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14131 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14132 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as));
14133 
14134 	if (sfmmup->sfmmu_free) {
14135 		ulong_t rttecnt;
14136 		r_pgszc = rgnp->rgn_pgszc;
14137 		r_size = rgnp->rgn_size;
14138 
14139 		ASSERT(sfmmup->sfmmu_scdp == NULL);
14140 		if (r_type == SFMMU_REGION_ISM) {
14141 			SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14142 		} else {
14143 			/* update shme rgns ttecnt in sfmmu_ttecnt */
14144 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14145 			ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14146 
14147 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14148 			    -rttecnt);
14149 
14150 			SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14151 		}
14152 	} else if (r_type == SFMMU_REGION_ISM) {
14153 		hatlockp = sfmmu_hat_enter(sfmmup);
14154 		ASSERT(rid < srdp->srd_next_ismrid);
14155 		SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14156 		scdp = sfmmup->sfmmu_scdp;
14157 		if (scdp != NULL &&
14158 		    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
14159 			sfmmu_leave_scd(sfmmup, r_type);
14160 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14161 		}
14162 		sfmmu_hat_exit(hatlockp);
14163 	} else {
14164 		ulong_t rttecnt;
14165 		r_pgszc = rgnp->rgn_pgszc;
14166 		r_saddr = rgnp->rgn_saddr;
14167 		r_size = rgnp->rgn_size;
14168 		r_eaddr = r_saddr + r_size;
14169 
14170 		ASSERT(r_type == SFMMU_REGION_HME);
14171 		hatlockp = sfmmu_hat_enter(sfmmup);
14172 		ASSERT(rid < srdp->srd_next_hmerid);
14173 		SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14174 
14175 		/*
14176 		 * If region is part of an SCD call sfmmu_leave_scd().
14177 		 * Otherwise if process is not exiting and has valid context
14178 		 * just drop the context on the floor to lose stale TLB
14179 		 * entries and force the update of tsb miss area to reflect
14180 		 * the new region map. After that clean our TSB entries.
14181 		 */
14182 		scdp = sfmmup->sfmmu_scdp;
14183 		if (scdp != NULL &&
14184 		    SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
14185 			sfmmu_leave_scd(sfmmup, r_type);
14186 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14187 		}
14188 		sfmmu_invalidate_ctx(sfmmup);
14189 
14190 		i = TTE8K;
14191 		while (i < mmu_page_sizes) {
14192 			if (rgnp->rgn_ttecnt[i] != 0) {
14193 				sfmmu_unload_tsb_range(sfmmup, r_saddr,
14194 				    r_eaddr, i);
14195 				if (i < TTE4M) {
14196 					i = TTE4M;
14197 					continue;
14198 				} else {
14199 					break;
14200 				}
14201 			}
14202 			i++;
14203 		}
14204 		/* Remove the preallocated 1/4 8k ttecnt for 4M regions. */
14205 		if (r_pgszc >= TTE4M) {
14206 			rttecnt = r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14207 			ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
14208 			    rttecnt);
14209 			sfmmup->sfmmu_tsb0_4minflcnt -= rttecnt;
14210 		}
14211 
14212 		/* update shme rgns ttecnt in sfmmu_ttecnt */
14213 		rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14214 		ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14215 		atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc], -rttecnt);
14216 
14217 		sfmmu_hat_exit(hatlockp);
14218 		if (scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
14219 			/* sfmmup left the scd, grow private tsb */
14220 			sfmmu_check_page_sizes(sfmmup, 1);
14221 		} else {
14222 			sfmmu_check_page_sizes(sfmmup, 0);
14223 		}
14224 	}
14225 
14226 	if (r_type == SFMMU_REGION_HME) {
14227 		sfmmu_unlink_from_hmeregion(sfmmup, rgnp);
14228 	}
14229 
14230 	r_obj = rgnp->rgn_obj;
14231 	if (atomic_dec_32_nv((volatile uint_t *)&rgnp->rgn_refcnt)) {
14232 		return;
14233 	}
14234 
14235 	/*
14236 	 * looks like nobody uses this region anymore. Free it.
14237 	 */
14238 	rhash = RGN_HASH_FUNCTION(r_obj);
14239 	mutex_enter(&srdp->srd_mutex);
14240 	for (prev_rgnpp = &srdp->srd_rgnhash[rhash];
14241 	    (cur_rgnp = *prev_rgnpp) != NULL;
14242 	    prev_rgnpp = &cur_rgnp->rgn_hash) {
14243 		if (cur_rgnp == rgnp && cur_rgnp->rgn_refcnt == 0) {
14244 			break;
14245 		}
14246 	}
14247 
14248 	if (cur_rgnp == NULL) {
14249 		mutex_exit(&srdp->srd_mutex);
14250 		return;
14251 	}
14252 
14253 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14254 	*prev_rgnpp = rgnp->rgn_hash;
14255 	if (r_type == SFMMU_REGION_ISM) {
14256 		rgnp->rgn_flags |= SFMMU_REGION_FREE;
14257 		ASSERT(rid < srdp->srd_next_ismrid);
14258 		rgnp->rgn_next = srdp->srd_ismrgnfree;
14259 		srdp->srd_ismrgnfree = rgnp;
14260 		ASSERT(srdp->srd_ismbusyrgns > 0);
14261 		srdp->srd_ismbusyrgns--;
14262 		mutex_exit(&srdp->srd_mutex);
14263 		return;
14264 	}
14265 	mutex_exit(&srdp->srd_mutex);
14266 
14267 	/*
14268 	 * Destroy region's hmeblks.
14269 	 */
14270 	sfmmu_unload_hmeregion(srdp, rgnp);
14271 
14272 	rgnp->rgn_hmeflags = 0;
14273 
14274 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14275 	ASSERT(rgnp->rgn_id == rid);
14276 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14277 		rgnp->rgn_ttecnt[i] = 0;
14278 	}
14279 	rgnp->rgn_flags |= SFMMU_REGION_FREE;
14280 	mutex_enter(&srdp->srd_mutex);
14281 	ASSERT(rid < srdp->srd_next_hmerid);
14282 	rgnp->rgn_next = srdp->srd_hmergnfree;
14283 	srdp->srd_hmergnfree = rgnp;
14284 	ASSERT(srdp->srd_hmebusyrgns > 0);
14285 	srdp->srd_hmebusyrgns--;
14286 	mutex_exit(&srdp->srd_mutex);
14287 }
14288 
14289 /*
14290  * For now only called for hmeblk regions and not for ISM regions.
14291  */
14292 void
14293 hat_dup_region(struct hat *sfmmup, hat_region_cookie_t rcookie)
14294 {
14295 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14296 	uint_t rid = (uint_t)((uint64_t)rcookie);
14297 	sf_region_t *rgnp;
14298 	sf_rgn_link_t *rlink;
14299 	sf_rgn_link_t *hrlink;
14300 	ulong_t	rttecnt;
14301 
14302 	ASSERT(sfmmup != ksfmmup);
14303 	ASSERT(srdp != NULL);
14304 	ASSERT(srdp->srd_refcnt > 0);
14305 
14306 	ASSERT(rid < srdp->srd_next_hmerid);
14307 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14308 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14309 
14310 	rgnp = srdp->srd_hmergnp[rid];
14311 	ASSERT(rgnp->rgn_refcnt > 0);
14312 	ASSERT(rgnp->rgn_id == rid);
14313 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == SFMMU_REGION_HME);
14314 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14315 
14316 	atomic_inc_32((volatile uint_t *)&rgnp->rgn_refcnt);
14317 
14318 	/* LINTED: constant in conditional context */
14319 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 0);
14320 	ASSERT(rlink != NULL);
14321 	mutex_enter(&rgnp->rgn_mutex);
14322 	ASSERT(rgnp->rgn_sfmmu_head != NULL);
14323 	/* LINTED: constant in conditional context */
14324 	SFMMU_HMERID2RLINKP(rgnp->rgn_sfmmu_head, rid, hrlink, 0, 0);
14325 	ASSERT(hrlink != NULL);
14326 	ASSERT(hrlink->prev == NULL);
14327 	rlink->next = rgnp->rgn_sfmmu_head;
14328 	rlink->prev = NULL;
14329 	hrlink->prev = sfmmup;
14330 	/*
14331 	 * make sure rlink's next field is correct
14332 	 * before making this link visible.
14333 	 */
14334 	membar_stst();
14335 	rgnp->rgn_sfmmu_head = sfmmup;
14336 	mutex_exit(&rgnp->rgn_mutex);
14337 
14338 	/* update sfmmu_ttecnt with the shme rgn ttecnt */
14339 	rttecnt = rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
14340 	atomic_add_long(&sfmmup->sfmmu_ttecnt[rgnp->rgn_pgszc], rttecnt);
14341 	/* update tsb0 inflation count */
14342 	if (rgnp->rgn_pgszc >= TTE4M) {
14343 		sfmmup->sfmmu_tsb0_4minflcnt +=
14344 		    rgnp->rgn_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14345 	}
14346 	/*
14347 	 * Update regionid bitmask without hat lock since no other thread
14348 	 * can update this region bitmask right now.
14349 	 */
14350 	SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14351 }
14352 
14353 /* ARGSUSED */
14354 static int
14355 sfmmu_rgncache_constructor(void *buf, void *cdrarg, int kmflags)
14356 {
14357 	sf_region_t *rgnp = (sf_region_t *)buf;
14358 	bzero(buf, sizeof (*rgnp));
14359 
14360 	mutex_init(&rgnp->rgn_mutex, NULL, MUTEX_DEFAULT, NULL);
14361 
14362 	return (0);
14363 }
14364 
14365 /* ARGSUSED */
14366 static void
14367 sfmmu_rgncache_destructor(void *buf, void *cdrarg)
14368 {
14369 	sf_region_t *rgnp = (sf_region_t *)buf;
14370 	mutex_destroy(&rgnp->rgn_mutex);
14371 }
14372 
14373 static int
14374 sfrgnmap_isnull(sf_region_map_t *map)
14375 {
14376 	int i;
14377 
14378 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14379 		if (map->bitmap[i] != 0) {
14380 			return (0);
14381 		}
14382 	}
14383 	return (1);
14384 }
14385 
14386 static int
14387 sfhmergnmap_isnull(sf_hmeregion_map_t *map)
14388 {
14389 	int i;
14390 
14391 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
14392 		if (map->bitmap[i] != 0) {
14393 			return (0);
14394 		}
14395 	}
14396 	return (1);
14397 }
14398 
14399 #ifdef DEBUG
14400 static void
14401 check_scd_sfmmu_list(sfmmu_t **headp, sfmmu_t *sfmmup, int onlist)
14402 {
14403 	sfmmu_t *sp;
14404 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14405 
14406 	for (sp = *headp; sp != NULL; sp = sp->sfmmu_scd_link.next) {
14407 		ASSERT(srdp == sp->sfmmu_srdp);
14408 		if (sp == sfmmup) {
14409 			if (onlist) {
14410 				return;
14411 			} else {
14412 				panic("shctx: sfmmu 0x%p found on scd"
14413 				    "list 0x%p", (void *)sfmmup,
14414 				    (void *)*headp);
14415 			}
14416 		}
14417 	}
14418 	if (onlist) {
14419 		panic("shctx: sfmmu 0x%p not found on scd list 0x%p",
14420 		    (void *)sfmmup, (void *)*headp);
14421 	} else {
14422 		return;
14423 	}
14424 }
14425 #else /* DEBUG */
14426 #define	check_scd_sfmmu_list(headp, sfmmup, onlist)
14427 #endif /* DEBUG */
14428 
14429 /*
14430  * Removes an sfmmu from the SCD sfmmu list.
14431  */
14432 static void
14433 sfmmu_from_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14434 {
14435 	ASSERT(sfmmup->sfmmu_srdp != NULL);
14436 	check_scd_sfmmu_list(headp, sfmmup, 1);
14437 	if (sfmmup->sfmmu_scd_link.prev != NULL) {
14438 		ASSERT(*headp != sfmmup);
14439 		sfmmup->sfmmu_scd_link.prev->sfmmu_scd_link.next =
14440 		    sfmmup->sfmmu_scd_link.next;
14441 	} else {
14442 		ASSERT(*headp == sfmmup);
14443 		*headp = sfmmup->sfmmu_scd_link.next;
14444 	}
14445 	if (sfmmup->sfmmu_scd_link.next != NULL) {
14446 		sfmmup->sfmmu_scd_link.next->sfmmu_scd_link.prev =
14447 		    sfmmup->sfmmu_scd_link.prev;
14448 	}
14449 }
14450 
14451 
14452 /*
14453  * Adds an sfmmu to the start of the queue.
14454  */
14455 static void
14456 sfmmu_to_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14457 {
14458 	check_scd_sfmmu_list(headp, sfmmup, 0);
14459 	sfmmup->sfmmu_scd_link.prev = NULL;
14460 	sfmmup->sfmmu_scd_link.next = *headp;
14461 	if (*headp != NULL)
14462 		(*headp)->sfmmu_scd_link.prev = sfmmup;
14463 	*headp = sfmmup;
14464 }
14465 
14466 /*
14467  * Remove an scd from the start of the queue.
14468  */
14469 static void
14470 sfmmu_remove_scd(sf_scd_t **headp, sf_scd_t *scdp)
14471 {
14472 	if (scdp->scd_prev != NULL) {
14473 		ASSERT(*headp != scdp);
14474 		scdp->scd_prev->scd_next = scdp->scd_next;
14475 	} else {
14476 		ASSERT(*headp == scdp);
14477 		*headp = scdp->scd_next;
14478 	}
14479 
14480 	if (scdp->scd_next != NULL) {
14481 		scdp->scd_next->scd_prev = scdp->scd_prev;
14482 	}
14483 }
14484 
14485 /*
14486  * Add an scd to the start of the queue.
14487  */
14488 static void
14489 sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *scdp)
14490 {
14491 	scdp->scd_prev = NULL;
14492 	scdp->scd_next = *headp;
14493 	if (*headp != NULL) {
14494 		(*headp)->scd_prev = scdp;
14495 	}
14496 	*headp = scdp;
14497 }
14498 
14499 static int
14500 sfmmu_alloc_scd_tsbs(sf_srd_t *srdp, sf_scd_t *scdp)
14501 {
14502 	uint_t rid;
14503 	uint_t i;
14504 	uint_t j;
14505 	ulong_t w;
14506 	sf_region_t *rgnp;
14507 	ulong_t tte8k_cnt = 0;
14508 	ulong_t tte4m_cnt = 0;
14509 	uint_t tsb_szc;
14510 	sfmmu_t *scsfmmup = scdp->scd_sfmmup;
14511 	sfmmu_t	*ism_hatid;
14512 	struct tsb_info *newtsb;
14513 	int szc;
14514 
14515 	ASSERT(srdp != NULL);
14516 
14517 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14518 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14519 			continue;
14520 		}
14521 		j = 0;
14522 		while (w) {
14523 			if (!(w & 0x1)) {
14524 				j++;
14525 				w >>= 1;
14526 				continue;
14527 			}
14528 			rid = (i << BT_ULSHIFT) | j;
14529 			j++;
14530 			w >>= 1;
14531 
14532 			if (rid < SFMMU_MAX_HME_REGIONS) {
14533 				rgnp = srdp->srd_hmergnp[rid];
14534 				ASSERT(rgnp->rgn_id == rid);
14535 				ASSERT(rgnp->rgn_refcnt > 0);
14536 
14537 				if (rgnp->rgn_pgszc < TTE4M) {
14538 					tte8k_cnt += rgnp->rgn_size >>
14539 					    TTE_PAGE_SHIFT(TTE8K);
14540 				} else {
14541 					ASSERT(rgnp->rgn_pgszc >= TTE4M);
14542 					tte4m_cnt += rgnp->rgn_size >>
14543 					    TTE_PAGE_SHIFT(TTE4M);
14544 					/*
14545 					 * Inflate SCD tsb0 by preallocating
14546 					 * 1/4 8k ttecnt for 4M regions to
14547 					 * allow for lgpg alloc failure.
14548 					 */
14549 					tte8k_cnt += rgnp->rgn_size >>
14550 					    (TTE_PAGE_SHIFT(TTE8K) + 2);
14551 				}
14552 			} else {
14553 				rid -= SFMMU_MAX_HME_REGIONS;
14554 				rgnp = srdp->srd_ismrgnp[rid];
14555 				ASSERT(rgnp->rgn_id == rid);
14556 				ASSERT(rgnp->rgn_refcnt > 0);
14557 
14558 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14559 				ASSERT(ism_hatid->sfmmu_ismhat);
14560 
14561 				for (szc = 0; szc < TTE4M; szc++) {
14562 					tte8k_cnt +=
14563 					    ism_hatid->sfmmu_ttecnt[szc] <<
14564 					    TTE_BSZS_SHIFT(szc);
14565 				}
14566 
14567 				ASSERT(rgnp->rgn_pgszc >= TTE4M);
14568 				if (rgnp->rgn_pgszc >= TTE4M) {
14569 					tte4m_cnt += rgnp->rgn_size >>
14570 					    TTE_PAGE_SHIFT(TTE4M);
14571 				}
14572 			}
14573 		}
14574 	}
14575 
14576 	tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
14577 
14578 	/* Allocate both the SCD TSBs here. */
14579 	if (sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14580 	    tsb_szc, TSB8K|TSB64K|TSB512K, TSB_ALLOC, scsfmmup) &&
14581 	    (tsb_szc <= TSB_4M_SZCODE ||
14582 	    sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14583 	    TSB_4M_SZCODE, TSB8K|TSB64K|TSB512K,
14584 	    TSB_ALLOC, scsfmmup))) {
14585 
14586 		SFMMU_STAT(sf_scd_1sttsb_allocfail);
14587 		return (TSB_ALLOCFAIL);
14588 	} else {
14589 		scsfmmup->sfmmu_tsb->tsb_flags |= TSB_SHAREDCTX;
14590 
14591 		if (tte4m_cnt) {
14592 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
14593 			if (sfmmu_tsbinfo_alloc(&newtsb, tsb_szc,
14594 			    TSB4M|TSB32M|TSB256M, TSB_ALLOC, scsfmmup) &&
14595 			    (tsb_szc <= TSB_4M_SZCODE ||
14596 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
14597 			    TSB4M|TSB32M|TSB256M,
14598 			    TSB_ALLOC, scsfmmup))) {
14599 				/*
14600 				 * If we fail to allocate the 2nd shared tsb,
14601 				 * just free the 1st tsb, return failure.
14602 				 */
14603 				sfmmu_tsbinfo_free(scsfmmup->sfmmu_tsb);
14604 				SFMMU_STAT(sf_scd_2ndtsb_allocfail);
14605 				return (TSB_ALLOCFAIL);
14606 			} else {
14607 				ASSERT(scsfmmup->sfmmu_tsb->tsb_next == NULL);
14608 				newtsb->tsb_flags |= TSB_SHAREDCTX;
14609 				scsfmmup->sfmmu_tsb->tsb_next = newtsb;
14610 				SFMMU_STAT(sf_scd_2ndtsb_alloc);
14611 			}
14612 		}
14613 		SFMMU_STAT(sf_scd_1sttsb_alloc);
14614 	}
14615 	return (TSB_SUCCESS);
14616 }
14617 
14618 static void
14619 sfmmu_free_scd_tsbs(sfmmu_t *scd_sfmmu)
14620 {
14621 	while (scd_sfmmu->sfmmu_tsb != NULL) {
14622 		struct tsb_info *next = scd_sfmmu->sfmmu_tsb->tsb_next;
14623 		sfmmu_tsbinfo_free(scd_sfmmu->sfmmu_tsb);
14624 		scd_sfmmu->sfmmu_tsb = next;
14625 	}
14626 }
14627 
14628 /*
14629  * Link the sfmmu onto the hme region list.
14630  */
14631 void
14632 sfmmu_link_to_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14633 {
14634 	uint_t rid;
14635 	sf_rgn_link_t *rlink;
14636 	sfmmu_t *head;
14637 	sf_rgn_link_t *hrlink;
14638 
14639 	rid = rgnp->rgn_id;
14640 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14641 
14642 	/* LINTED: constant in conditional context */
14643 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 1);
14644 	ASSERT(rlink != NULL);
14645 	mutex_enter(&rgnp->rgn_mutex);
14646 	if ((head = rgnp->rgn_sfmmu_head) == NULL) {
14647 		rlink->next = NULL;
14648 		rlink->prev = NULL;
14649 		/*
14650 		 * make sure rlink's next field is NULL
14651 		 * before making this link visible.
14652 		 */
14653 		membar_stst();
14654 		rgnp->rgn_sfmmu_head = sfmmup;
14655 	} else {
14656 		/* LINTED: constant in conditional context */
14657 		SFMMU_HMERID2RLINKP(head, rid, hrlink, 0, 0);
14658 		ASSERT(hrlink != NULL);
14659 		ASSERT(hrlink->prev == NULL);
14660 		rlink->next = head;
14661 		rlink->prev = NULL;
14662 		hrlink->prev = sfmmup;
14663 		/*
14664 		 * make sure rlink's next field is correct
14665 		 * before making this link visible.
14666 		 */
14667 		membar_stst();
14668 		rgnp->rgn_sfmmu_head = sfmmup;
14669 	}
14670 	mutex_exit(&rgnp->rgn_mutex);
14671 }
14672 
14673 /*
14674  * Unlink the sfmmu from the hme region list.
14675  */
14676 void
14677 sfmmu_unlink_from_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14678 {
14679 	uint_t rid;
14680 	sf_rgn_link_t *rlink;
14681 
14682 	rid = rgnp->rgn_id;
14683 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14684 
14685 	/* LINTED: constant in conditional context */
14686 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
14687 	ASSERT(rlink != NULL);
14688 	mutex_enter(&rgnp->rgn_mutex);
14689 	if (rgnp->rgn_sfmmu_head == sfmmup) {
14690 		sfmmu_t *next = rlink->next;
14691 		rgnp->rgn_sfmmu_head = next;
14692 		/*
14693 		 * if we are stopped by xc_attention() after this
14694 		 * point the forward link walking in
14695 		 * sfmmu_rgntlb_demap() will work correctly since the
14696 		 * head correctly points to the next element.
14697 		 */
14698 		membar_stst();
14699 		rlink->next = NULL;
14700 		ASSERT(rlink->prev == NULL);
14701 		if (next != NULL) {
14702 			sf_rgn_link_t *nrlink;
14703 			/* LINTED: constant in conditional context */
14704 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14705 			ASSERT(nrlink != NULL);
14706 			ASSERT(nrlink->prev == sfmmup);
14707 			nrlink->prev = NULL;
14708 		}
14709 	} else {
14710 		sfmmu_t *next = rlink->next;
14711 		sfmmu_t *prev = rlink->prev;
14712 		sf_rgn_link_t *prlink;
14713 
14714 		ASSERT(prev != NULL);
14715 		/* LINTED: constant in conditional context */
14716 		SFMMU_HMERID2RLINKP(prev, rid, prlink, 0, 0);
14717 		ASSERT(prlink != NULL);
14718 		ASSERT(prlink->next == sfmmup);
14719 		prlink->next = next;
14720 		/*
14721 		 * if we are stopped by xc_attention()
14722 		 * after this point the forward link walking
14723 		 * will work correctly since the prev element
14724 		 * correctly points to the next element.
14725 		 */
14726 		membar_stst();
14727 		rlink->next = NULL;
14728 		rlink->prev = NULL;
14729 		if (next != NULL) {
14730 			sf_rgn_link_t *nrlink;
14731 			/* LINTED: constant in conditional context */
14732 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14733 			ASSERT(nrlink != NULL);
14734 			ASSERT(nrlink->prev == sfmmup);
14735 			nrlink->prev = prev;
14736 		}
14737 	}
14738 	mutex_exit(&rgnp->rgn_mutex);
14739 }
14740 
14741 /*
14742  * Link scd sfmmu onto ism or hme region list for each region in the
14743  * scd region map.
14744  */
14745 void
14746 sfmmu_link_scd_to_regions(sf_srd_t *srdp, sf_scd_t *scdp)
14747 {
14748 	uint_t rid;
14749 	uint_t i;
14750 	uint_t j;
14751 	ulong_t w;
14752 	sf_region_t *rgnp;
14753 	sfmmu_t *scsfmmup;
14754 
14755 	scsfmmup = scdp->scd_sfmmup;
14756 	ASSERT(scsfmmup->sfmmu_scdhat);
14757 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14758 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14759 			continue;
14760 		}
14761 		j = 0;
14762 		while (w) {
14763 			if (!(w & 0x1)) {
14764 				j++;
14765 				w >>= 1;
14766 				continue;
14767 			}
14768 			rid = (i << BT_ULSHIFT) | j;
14769 			j++;
14770 			w >>= 1;
14771 
14772 			if (rid < SFMMU_MAX_HME_REGIONS) {
14773 				rgnp = srdp->srd_hmergnp[rid];
14774 				ASSERT(rgnp->rgn_id == rid);
14775 				ASSERT(rgnp->rgn_refcnt > 0);
14776 				sfmmu_link_to_hmeregion(scsfmmup, rgnp);
14777 			} else {
14778 				sfmmu_t *ism_hatid = NULL;
14779 				ism_ment_t *ism_ment;
14780 				rid -= SFMMU_MAX_HME_REGIONS;
14781 				rgnp = srdp->srd_ismrgnp[rid];
14782 				ASSERT(rgnp->rgn_id == rid);
14783 				ASSERT(rgnp->rgn_refcnt > 0);
14784 
14785 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14786 				ASSERT(ism_hatid->sfmmu_ismhat);
14787 				ism_ment = &scdp->scd_ism_links[rid];
14788 				ism_ment->iment_hat = scsfmmup;
14789 				ism_ment->iment_base_va = rgnp->rgn_saddr;
14790 				mutex_enter(&ism_mlist_lock);
14791 				iment_add(ism_ment, ism_hatid);
14792 				mutex_exit(&ism_mlist_lock);
14793 
14794 			}
14795 		}
14796 	}
14797 }
14798 /*
14799  * Unlink scd sfmmu from ism or hme region list for each region in the
14800  * scd region map.
14801  */
14802 void
14803 sfmmu_unlink_scd_from_regions(sf_srd_t *srdp, sf_scd_t *scdp)
14804 {
14805 	uint_t rid;
14806 	uint_t i;
14807 	uint_t j;
14808 	ulong_t w;
14809 	sf_region_t *rgnp;
14810 	sfmmu_t *scsfmmup;
14811 
14812 	scsfmmup = scdp->scd_sfmmup;
14813 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14814 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14815 			continue;
14816 		}
14817 		j = 0;
14818 		while (w) {
14819 			if (!(w & 0x1)) {
14820 				j++;
14821 				w >>= 1;
14822 				continue;
14823 			}
14824 			rid = (i << BT_ULSHIFT) | j;
14825 			j++;
14826 			w >>= 1;
14827 
14828 			if (rid < SFMMU_MAX_HME_REGIONS) {
14829 				rgnp = srdp->srd_hmergnp[rid];
14830 				ASSERT(rgnp->rgn_id == rid);
14831 				ASSERT(rgnp->rgn_refcnt > 0);
14832 				sfmmu_unlink_from_hmeregion(scsfmmup,
14833 				    rgnp);
14834 
14835 			} else {
14836 				sfmmu_t *ism_hatid = NULL;
14837 				ism_ment_t *ism_ment;
14838 				rid -= SFMMU_MAX_HME_REGIONS;
14839 				rgnp = srdp->srd_ismrgnp[rid];
14840 				ASSERT(rgnp->rgn_id == rid);
14841 				ASSERT(rgnp->rgn_refcnt > 0);
14842 
14843 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14844 				ASSERT(ism_hatid->sfmmu_ismhat);
14845 				ism_ment = &scdp->scd_ism_links[rid];
14846 				ASSERT(ism_ment->iment_hat == scdp->scd_sfmmup);
14847 				ASSERT(ism_ment->iment_base_va ==
14848 				    rgnp->rgn_saddr);
14849 				mutex_enter(&ism_mlist_lock);
14850 				iment_sub(ism_ment, ism_hatid);
14851 				mutex_exit(&ism_mlist_lock);
14852 
14853 			}
14854 		}
14855 	}
14856 }
14857 /*
14858  * Allocates and initialises a new SCD structure, this is called with
14859  * the srd_scd_mutex held and returns with the reference count
14860  * initialised to 1.
14861  */
14862 static sf_scd_t *
14863 sfmmu_alloc_scd(sf_srd_t *srdp, sf_region_map_t *new_map)
14864 {
14865 	sf_scd_t *new_scdp;
14866 	sfmmu_t *scsfmmup;
14867 	int i;
14868 
14869 	ASSERT(MUTEX_HELD(&srdp->srd_scd_mutex));
14870 	new_scdp = kmem_cache_alloc(scd_cache, KM_SLEEP);
14871 
14872 	scsfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
14873 	new_scdp->scd_sfmmup = scsfmmup;
14874 	scsfmmup->sfmmu_srdp = srdp;
14875 	scsfmmup->sfmmu_scdp = new_scdp;
14876 	scsfmmup->sfmmu_tsb0_4minflcnt = 0;
14877 	scsfmmup->sfmmu_scdhat = 1;
14878 	CPUSET_ALL(scsfmmup->sfmmu_cpusran);
14879 	bzero(scsfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
14880 
14881 	ASSERT(max_mmu_ctxdoms > 0);
14882 	for (i = 0; i < max_mmu_ctxdoms; i++) {
14883 		scsfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
14884 		scsfmmup->sfmmu_ctxs[i].gnum = 0;
14885 	}
14886 
14887 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14888 		new_scdp->scd_rttecnt[i] = 0;
14889 	}
14890 
14891 	new_scdp->scd_region_map = *new_map;
14892 	new_scdp->scd_refcnt = 1;
14893 	if (sfmmu_alloc_scd_tsbs(srdp, new_scdp) != TSB_SUCCESS) {
14894 		kmem_cache_free(scd_cache, new_scdp);
14895 		kmem_cache_free(sfmmuid_cache, scsfmmup);
14896 		return (NULL);
14897 	}
14898 	if (&mmu_init_scd) {
14899 		mmu_init_scd(new_scdp);
14900 	}
14901 	return (new_scdp);
14902 }
14903 
14904 /*
14905  * The first phase of a process joining an SCD. The hat structure is
14906  * linked to the SCD queue and then the HAT_JOIN_SCD sfmmu flag is set
14907  * and a cross-call with context invalidation is used to cause the
14908  * remaining work to be carried out in the sfmmu_tsbmiss_exception()
14909  * routine.
14910  */
14911 static void
14912 sfmmu_join_scd(sf_scd_t *scdp, sfmmu_t *sfmmup)
14913 {
14914 	hatlock_t *hatlockp;
14915 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14916 	int i;
14917 	sf_scd_t *old_scdp;
14918 
14919 	ASSERT(srdp != NULL);
14920 	ASSERT(scdp != NULL);
14921 	ASSERT(scdp->scd_refcnt > 0);
14922 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as));
14923 
14924 	if ((old_scdp = sfmmup->sfmmu_scdp) != NULL) {
14925 		ASSERT(old_scdp != scdp);
14926 
14927 		mutex_enter(&old_scdp->scd_mutex);
14928 		sfmmu_from_scd_list(&old_scdp->scd_sf_list, sfmmup);
14929 		mutex_exit(&old_scdp->scd_mutex);
14930 		/*
14931 		 * sfmmup leaves the old scd. Update sfmmu_ttecnt to
14932 		 * include the shme rgn ttecnt for rgns that
14933 		 * were in the old SCD
14934 		 */
14935 		for (i = 0; i < mmu_page_sizes; i++) {
14936 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
14937 			    old_scdp->scd_rttecnt[i]);
14938 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
14939 			    sfmmup->sfmmu_scdrttecnt[i]);
14940 		}
14941 	}
14942 
14943 	/*
14944 	 * Move sfmmu to the scd lists.
14945 	 */
14946 	mutex_enter(&scdp->scd_mutex);
14947 	sfmmu_to_scd_list(&scdp->scd_sf_list, sfmmup);
14948 	mutex_exit(&scdp->scd_mutex);
14949 	SF_SCD_INCR_REF(scdp);
14950 
14951 	hatlockp = sfmmu_hat_enter(sfmmup);
14952 	/*
14953 	 * For a multi-thread process, we must stop
14954 	 * all the other threads before joining the scd.
14955 	 */
14956 
14957 	SFMMU_FLAGS_SET(sfmmup, HAT_JOIN_SCD);
14958 
14959 	sfmmu_invalidate_ctx(sfmmup);
14960 	sfmmup->sfmmu_scdp = scdp;
14961 
14962 	/*
14963 	 * Copy scd_rttecnt into sfmmup's sfmmu_scdrttecnt, and update
14964 	 * sfmmu_ttecnt to not include the rgn ttecnt just joined in SCD.
14965 	 */
14966 	for (i = 0; i < mmu_page_sizes; i++) {
14967 		sfmmup->sfmmu_scdrttecnt[i] = scdp->scd_rttecnt[i];
14968 		ASSERT(sfmmup->sfmmu_ttecnt[i] >= scdp->scd_rttecnt[i]);
14969 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
14970 		    -sfmmup->sfmmu_scdrttecnt[i]);
14971 	}
14972 	/* update tsb0 inflation count */
14973 	if (old_scdp != NULL) {
14974 		sfmmup->sfmmu_tsb0_4minflcnt +=
14975 		    old_scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
14976 	}
14977 	ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
14978 	    scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt);
14979 	sfmmup->sfmmu_tsb0_4minflcnt -= scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
14980 
14981 	sfmmu_hat_exit(hatlockp);
14982 
14983 	if (old_scdp != NULL) {
14984 		SF_SCD_DECR_REF(srdp, old_scdp);
14985 	}
14986 
14987 }
14988 
14989 /*
14990  * This routine is called by a process to become part of an SCD. It is called
14991  * from sfmmu_tsbmiss_exception() once most of the initial work has been
14992  * done by sfmmu_join_scd(). This routine must not drop the hat lock.
14993  */
14994 static void
14995 sfmmu_finish_join_scd(sfmmu_t *sfmmup)
14996 {
14997 	struct tsb_info	*tsbinfop;
14998 
14999 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15000 	ASSERT(sfmmup->sfmmu_scdp != NULL);
15001 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD));
15002 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15003 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID));
15004 
15005 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
15006 	    tsbinfop = tsbinfop->tsb_next) {
15007 		if (tsbinfop->tsb_flags & TSB_SWAPPED) {
15008 			continue;
15009 		}
15010 		ASSERT(!(tsbinfop->tsb_flags & TSB_RELOC_FLAG));
15011 
15012 		sfmmu_inv_tsb(tsbinfop->tsb_va,
15013 		    TSB_BYTES(tsbinfop->tsb_szc));
15014 	}
15015 
15016 	/* Set HAT_CTX1_FLAG for all SCD ISMs */
15017 	sfmmu_ism_hatflags(sfmmup, 1);
15018 
15019 	SFMMU_STAT(sf_join_scd);
15020 }
15021 
15022 /*
15023  * This routine is called in order to check if there is an SCD which matches
15024  * the process's region map if not then a new SCD may be created.
15025  */
15026 static void
15027 sfmmu_find_scd(sfmmu_t *sfmmup)
15028 {
15029 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15030 	sf_scd_t *scdp, *new_scdp;
15031 	int ret;
15032 
15033 	ASSERT(srdp != NULL);
15034 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as));
15035 
15036 	mutex_enter(&srdp->srd_scd_mutex);
15037 	for (scdp = srdp->srd_scdp; scdp != NULL;
15038 	    scdp = scdp->scd_next) {
15039 		SF_RGNMAP_EQUAL(&scdp->scd_region_map,
15040 		    &sfmmup->sfmmu_region_map, ret);
15041 		if (ret == 1) {
15042 			SF_SCD_INCR_REF(scdp);
15043 			mutex_exit(&srdp->srd_scd_mutex);
15044 			sfmmu_join_scd(scdp, sfmmup);
15045 			ASSERT(scdp->scd_refcnt >= 2);
15046 			atomic_dec_32((volatile uint32_t *)&scdp->scd_refcnt);
15047 			return;
15048 		} else {
15049 			/*
15050 			 * If the sfmmu region map is a subset of the scd
15051 			 * region map, then the assumption is that this process
15052 			 * will continue attaching to ISM segments until the
15053 			 * region maps are equal.
15054 			 */
15055 			SF_RGNMAP_IS_SUBSET(&scdp->scd_region_map,
15056 			    &sfmmup->sfmmu_region_map, ret);
15057 			if (ret == 1) {
15058 				mutex_exit(&srdp->srd_scd_mutex);
15059 				return;
15060 			}
15061 		}
15062 	}
15063 
15064 	ASSERT(scdp == NULL);
15065 	/*
15066 	 * No matching SCD has been found, create a new one.
15067 	 */
15068 	if ((new_scdp = sfmmu_alloc_scd(srdp, &sfmmup->sfmmu_region_map)) ==
15069 	    NULL) {
15070 		mutex_exit(&srdp->srd_scd_mutex);
15071 		return;
15072 	}
15073 
15074 	/*
15075 	 * sfmmu_alloc_scd() returns with a ref count of 1 on the scd.
15076 	 */
15077 
15078 	/* Set scd_rttecnt for shme rgns in SCD */
15079 	sfmmu_set_scd_rttecnt(srdp, new_scdp);
15080 
15081 	/*
15082 	 * Link scd onto srd_scdp list and scd sfmmu onto region/iment lists.
15083 	 */
15084 	sfmmu_link_scd_to_regions(srdp, new_scdp);
15085 	sfmmu_add_scd(&srdp->srd_scdp, new_scdp);
15086 	SFMMU_STAT_ADD(sf_create_scd, 1);
15087 
15088 	mutex_exit(&srdp->srd_scd_mutex);
15089 	sfmmu_join_scd(new_scdp, sfmmup);
15090 	ASSERT(new_scdp->scd_refcnt >= 2);
15091 	atomic_dec_32((volatile uint32_t *)&new_scdp->scd_refcnt);
15092 }
15093 
15094 /*
15095  * This routine is called by a process to remove itself from an SCD. It is
15096  * either called when the processes has detached from a segment or from
15097  * hat_free_start() as a result of calling exit.
15098  */
15099 static void
15100 sfmmu_leave_scd(sfmmu_t *sfmmup, uchar_t r_type)
15101 {
15102 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15103 	sf_srd_t *srdp =  sfmmup->sfmmu_srdp;
15104 	hatlock_t *hatlockp = TSB_HASH(sfmmup);
15105 	int i;
15106 
15107 	ASSERT(scdp != NULL);
15108 	ASSERT(srdp != NULL);
15109 
15110 	if (sfmmup->sfmmu_free) {
15111 		/*
15112 		 * If the process is part of an SCD the sfmmu is unlinked
15113 		 * from scd_sf_list.
15114 		 */
15115 		mutex_enter(&scdp->scd_mutex);
15116 		sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15117 		mutex_exit(&scdp->scd_mutex);
15118 		/*
15119 		 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15120 		 * are about to leave the SCD
15121 		 */
15122 		for (i = 0; i < mmu_page_sizes; i++) {
15123 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15124 			    scdp->scd_rttecnt[i]);
15125 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15126 			    sfmmup->sfmmu_scdrttecnt[i]);
15127 			sfmmup->sfmmu_scdrttecnt[i] = 0;
15128 		}
15129 		sfmmup->sfmmu_scdp = NULL;
15130 
15131 		SF_SCD_DECR_REF(srdp, scdp);
15132 		return;
15133 	}
15134 
15135 	ASSERT(r_type != SFMMU_REGION_ISM ||
15136 	    SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15137 	ASSERT(scdp->scd_refcnt);
15138 	ASSERT(!sfmmup->sfmmu_free);
15139 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15140 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as));
15141 
15142 	/*
15143 	 * Wait for ISM maps to be updated.
15144 	 */
15145 	if (r_type != SFMMU_REGION_ISM) {
15146 		while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY) &&
15147 		    sfmmup->sfmmu_scdp != NULL) {
15148 			cv_wait(&sfmmup->sfmmu_tsb_cv,
15149 			    HATLOCK_MUTEXP(hatlockp));
15150 		}
15151 
15152 		if (sfmmup->sfmmu_scdp == NULL) {
15153 			sfmmu_hat_exit(hatlockp);
15154 			return;
15155 		}
15156 		SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
15157 	}
15158 
15159 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
15160 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
15161 		/*
15162 		 * Since HAT_JOIN_SCD was set our context
15163 		 * is still invalid.
15164 		 */
15165 	} else {
15166 		/*
15167 		 * For a multi-thread process, we must stop
15168 		 * all the other threads before leaving the scd.
15169 		 */
15170 
15171 		sfmmu_invalidate_ctx(sfmmup);
15172 	}
15173 
15174 	/* Clear all the rid's for ISM, delete flags, etc */
15175 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15176 	sfmmu_ism_hatflags(sfmmup, 0);
15177 
15178 	/*
15179 	 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15180 	 * are in SCD before this sfmmup leaves the SCD.
15181 	 */
15182 	for (i = 0; i < mmu_page_sizes; i++) {
15183 		ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15184 		    scdp->scd_rttecnt[i]);
15185 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15186 		    sfmmup->sfmmu_scdrttecnt[i]);
15187 		sfmmup->sfmmu_scdrttecnt[i] = 0;
15188 		/* update ismttecnt to include SCD ism before hat leaves SCD */
15189 		sfmmup->sfmmu_ismttecnt[i] += sfmmup->sfmmu_scdismttecnt[i];
15190 		sfmmup->sfmmu_scdismttecnt[i] = 0;
15191 	}
15192 	/* update tsb0 inflation count */
15193 	sfmmup->sfmmu_tsb0_4minflcnt += scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15194 
15195 	if (r_type != SFMMU_REGION_ISM) {
15196 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
15197 	}
15198 	sfmmup->sfmmu_scdp = NULL;
15199 
15200 	sfmmu_hat_exit(hatlockp);
15201 
15202 	/*
15203 	 * Unlink sfmmu from scd_sf_list this can be done without holding
15204 	 * the hat lock as we hold the sfmmu_as lock which prevents
15205 	 * hat_join_region from adding this thread to the scd again. Other
15206 	 * threads check if sfmmu_scdp is NULL under hat lock and if it's NULL
15207 	 * they won't get here, since sfmmu_leave_scd() clears sfmmu_scdp
15208 	 * while holding the hat lock.
15209 	 */
15210 	mutex_enter(&scdp->scd_mutex);
15211 	sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15212 	mutex_exit(&scdp->scd_mutex);
15213 	SFMMU_STAT(sf_leave_scd);
15214 
15215 	SF_SCD_DECR_REF(srdp, scdp);
15216 	hatlockp = sfmmu_hat_enter(sfmmup);
15217 
15218 }
15219 
15220 /*
15221  * Unlink and free up an SCD structure with a reference count of 0.
15222  */
15223 static void
15224 sfmmu_destroy_scd(sf_srd_t *srdp, sf_scd_t *scdp, sf_region_map_t *scd_rmap)
15225 {
15226 	sfmmu_t *scsfmmup;
15227 	sf_scd_t *sp;
15228 	hatlock_t *shatlockp;
15229 	int i, ret;
15230 
15231 	mutex_enter(&srdp->srd_scd_mutex);
15232 	for (sp = srdp->srd_scdp; sp != NULL; sp = sp->scd_next) {
15233 		if (sp == scdp)
15234 			break;
15235 	}
15236 	if (sp == NULL || sp->scd_refcnt) {
15237 		mutex_exit(&srdp->srd_scd_mutex);
15238 		return;
15239 	}
15240 
15241 	/*
15242 	 * It is possible that the scd has been freed and reallocated with a
15243 	 * different region map while we've been waiting for the srd_scd_mutex.
15244 	 */
15245 	SF_RGNMAP_EQUAL(scd_rmap, &sp->scd_region_map, ret);
15246 	if (ret != 1) {
15247 		mutex_exit(&srdp->srd_scd_mutex);
15248 		return;
15249 	}
15250 
15251 	ASSERT(scdp->scd_sf_list == NULL);
15252 	/*
15253 	 * Unlink scd from srd_scdp list.
15254 	 */
15255 	sfmmu_remove_scd(&srdp->srd_scdp, scdp);
15256 	mutex_exit(&srdp->srd_scd_mutex);
15257 
15258 	sfmmu_unlink_scd_from_regions(srdp, scdp);
15259 
15260 	/* Clear shared context tsb and release ctx */
15261 	scsfmmup = scdp->scd_sfmmup;
15262 
15263 	/*
15264 	 * create a barrier so that scd will not be destroyed
15265 	 * if other thread still holds the same shared hat lock.
15266 	 * E.g., sfmmu_tsbmiss_exception() needs to acquire the
15267 	 * shared hat lock before checking the shared tsb reloc flag.
15268 	 */
15269 	shatlockp = sfmmu_hat_enter(scsfmmup);
15270 	sfmmu_hat_exit(shatlockp);
15271 
15272 	sfmmu_free_scd_tsbs(scsfmmup);
15273 
15274 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
15275 		if (scsfmmup->sfmmu_hmeregion_links[i] != NULL) {
15276 			kmem_free(scsfmmup->sfmmu_hmeregion_links[i],
15277 			    SFMMU_L2_HMERLINKS_SIZE);
15278 			scsfmmup->sfmmu_hmeregion_links[i] = NULL;
15279 		}
15280 	}
15281 	kmem_cache_free(sfmmuid_cache, scsfmmup);
15282 	kmem_cache_free(scd_cache, scdp);
15283 	SFMMU_STAT(sf_destroy_scd);
15284 }
15285 
15286 /*
15287  * Modifies the HAT_CTX1_FLAG for each of the ISM segments which correspond to
15288  * bits which are set in the ism_region_map parameter. This flag indicates to
15289  * the tsbmiss handler that mapping for these segments should be loaded using
15290  * the shared context.
15291  */
15292 static void
15293 sfmmu_ism_hatflags(sfmmu_t *sfmmup, int addflag)
15294 {
15295 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15296 	ism_blk_t *ism_blkp;
15297 	ism_map_t *ism_map;
15298 	int i, rid;
15299 
15300 	ASSERT(sfmmup->sfmmu_iblk != NULL);
15301 	ASSERT(scdp != NULL);
15302 	/*
15303 	 * Note that the caller either set HAT_ISMBUSY flag or checked
15304 	 * under hat lock that HAT_ISMBUSY was not set by another thread.
15305 	 */
15306 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15307 
15308 	ism_blkp = sfmmup->sfmmu_iblk;
15309 	while (ism_blkp != NULL) {
15310 		ism_map = ism_blkp->iblk_maps;
15311 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
15312 			rid = ism_map[i].imap_rid;
15313 			if (rid == SFMMU_INVALID_ISMRID) {
15314 				continue;
15315 			}
15316 			ASSERT(rid >= 0 && rid < SFMMU_MAX_ISM_REGIONS);
15317 			if (SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid) &&
15318 			    addflag) {
15319 				ism_map[i].imap_hatflags |=
15320 				    HAT_CTX1_FLAG;
15321 			} else {
15322 				ism_map[i].imap_hatflags &=
15323 				    ~HAT_CTX1_FLAG;
15324 			}
15325 		}
15326 		ism_blkp = ism_blkp->iblk_next;
15327 	}
15328 }
15329 
15330 static int
15331 sfmmu_srd_lock_held(sf_srd_t *srdp)
15332 {
15333 	return (MUTEX_HELD(&srdp->srd_mutex));
15334 }
15335 
15336 /* ARGSUSED */
15337 static int
15338 sfmmu_scdcache_constructor(void *buf, void *cdrarg, int kmflags)
15339 {
15340 	sf_scd_t *scdp = (sf_scd_t *)buf;
15341 
15342 	bzero(buf, sizeof (sf_scd_t));
15343 	mutex_init(&scdp->scd_mutex, NULL, MUTEX_DEFAULT, NULL);
15344 	return (0);
15345 }
15346 
15347 /* ARGSUSED */
15348 static void
15349 sfmmu_scdcache_destructor(void *buf, void *cdrarg)
15350 {
15351 	sf_scd_t *scdp = (sf_scd_t *)buf;
15352 
15353 	mutex_destroy(&scdp->scd_mutex);
15354 }
15355 
15356 /*
15357  * The listp parameter is a pointer to a list of hmeblks which are partially
15358  * freed as result of calling sfmmu_hblk_hash_rm(), the last phase of the
15359  * freeing process is to cross-call all cpus to ensure that there are no
15360  * remaining cached references.
15361  *
15362  * If the local generation number is less than the global then we can free
15363  * hmeblks which are already on the pending queue as another cpu has completed
15364  * the cross-call.
15365  *
15366  * We cross-call to make sure that there are no threads on other cpus accessing
15367  * these hmblks and then complete the process of freeing them under the
15368  * following conditions:
15369  *	The total number of pending hmeblks is greater than the threshold
15370  *	The reserve list has fewer than HBLK_RESERVE_CNT hmeblks
15371  *	It is at least 1 second since the last time we cross-called
15372  *
15373  * Otherwise, we add the hmeblks to the per-cpu pending queue.
15374  */
15375 static void
15376 sfmmu_hblks_list_purge(struct hme_blk **listp, int dontfree)
15377 {
15378 	struct hme_blk *hblkp, *pr_hblkp = NULL;
15379 	int		count = 0;
15380 	cpuset_t	cpuset = cpu_ready_set;
15381 	cpu_hme_pend_t	*cpuhp;
15382 	timestruc_t	now;
15383 	int		one_second_expired = 0;
15384 
15385 	gethrestime_lasttick(&now);
15386 
15387 	for (hblkp = *listp; hblkp != NULL; hblkp = hblkp->hblk_next) {
15388 		ASSERT(hblkp->hblk_shw_bit == 0);
15389 		ASSERT(hblkp->hblk_shared == 0);
15390 		count++;
15391 		pr_hblkp = hblkp;
15392 	}
15393 
15394 	cpuhp = &cpu_hme_pend[CPU->cpu_seqid];
15395 	mutex_enter(&cpuhp->chp_mutex);
15396 
15397 	if ((cpuhp->chp_count + count) == 0) {
15398 		mutex_exit(&cpuhp->chp_mutex);
15399 		return;
15400 	}
15401 
15402 	if ((now.tv_sec - cpuhp->chp_timestamp) > 1) {
15403 		one_second_expired  = 1;
15404 	}
15405 
15406 	if (!dontfree && (freehblkcnt < HBLK_RESERVE_CNT ||
15407 	    (cpuhp->chp_count + count) > cpu_hme_pend_thresh ||
15408 	    one_second_expired)) {
15409 		/* Append global list to local */
15410 		if (pr_hblkp == NULL) {
15411 			*listp = cpuhp->chp_listp;
15412 		} else {
15413 			pr_hblkp->hblk_next = cpuhp->chp_listp;
15414 		}
15415 		cpuhp->chp_listp = NULL;
15416 		cpuhp->chp_count = 0;
15417 		cpuhp->chp_timestamp = now.tv_sec;
15418 		mutex_exit(&cpuhp->chp_mutex);
15419 
15420 		kpreempt_disable();
15421 		CPUSET_DEL(cpuset, CPU->cpu_id);
15422 		xt_sync(cpuset);
15423 		xt_sync(cpuset);
15424 		kpreempt_enable();
15425 
15426 		/*
15427 		 * At this stage we know that no trap handlers on other
15428 		 * cpus can have references to hmeblks on the list.
15429 		 */
15430 		sfmmu_hblk_free(listp);
15431 	} else if (*listp != NULL) {
15432 		pr_hblkp->hblk_next = cpuhp->chp_listp;
15433 		cpuhp->chp_listp = *listp;
15434 		cpuhp->chp_count += count;
15435 		*listp = NULL;
15436 		mutex_exit(&cpuhp->chp_mutex);
15437 	} else {
15438 		mutex_exit(&cpuhp->chp_mutex);
15439 	}
15440 }
15441 
15442 /*
15443  * Add an hmeblk to the the hash list.
15444  */
15445 void
15446 sfmmu_hblk_hash_add(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
15447     uint64_t hblkpa)
15448 {
15449 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
15450 #ifdef	DEBUG
15451 	if (hmebp->hmeblkp == NULL) {
15452 		ASSERT(hmebp->hmeh_nextpa == HMEBLK_ENDPA);
15453 	}
15454 #endif /* DEBUG */
15455 
15456 	hmeblkp->hblk_nextpa = hmebp->hmeh_nextpa;
15457 	/*
15458 	 * Since the TSB miss handler now does not lock the hash chain before
15459 	 * walking it, make sure that the hmeblks nextpa is globally visible
15460 	 * before we make the hmeblk globally visible by updating the chain root
15461 	 * pointer in the hash bucket.
15462 	 */
15463 	membar_producer();
15464 	hmebp->hmeh_nextpa = hblkpa;
15465 	hmeblkp->hblk_next = hmebp->hmeblkp;
15466 	hmebp->hmeblkp = hmeblkp;
15467 
15468 }
15469 
15470 /*
15471  * This function is the first part of a 2 part process to remove an hmeblk
15472  * from the hash chain. In this phase we unlink the hmeblk from the hash chain
15473  * but leave the next physical pointer unchanged. The hmeblk is then linked onto
15474  * a per-cpu pending list using the virtual address pointer.
15475  *
15476  * TSB miss trap handlers that start after this phase will no longer see
15477  * this hmeblk. TSB miss handlers that still cache this hmeblk in a register
15478  * can still use it for further chain traversal because we haven't yet modifed
15479  * the next physical pointer or freed it.
15480  *
15481  * In the second phase of hmeblk removal we'll issue a barrier xcall before
15482  * we reuse or free this hmeblk. This will make sure all lingering references to
15483  * the hmeblk after first phase disappear before we finally reclaim it.
15484  * This scheme eliminates the need for TSB miss handlers to lock hmeblk chains
15485  * during their traversal.
15486  *
15487  * The hmehash_mutex must be held when calling this function.
15488  *
15489  * Input:
15490  *	 hmebp - hme hash bucket pointer
15491  *	 hmeblkp - address of hmeblk to be removed
15492  *	 pr_hblk - virtual address of previous hmeblkp
15493  *	 listp - pointer to list of hmeblks linked by virtual address
15494  *	 free_now flag - indicates that a complete removal from the hash chains
15495  *			 is necessary.
15496  *
15497  * It is inefficient to use the free_now flag as a cross-call is required to
15498  * remove a single hmeblk from the hash chain but is necessary when hmeblks are
15499  * in short supply.
15500  */
15501 void
15502 sfmmu_hblk_hash_rm(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
15503     struct hme_blk *pr_hblk, struct hme_blk **listp, int free_now)
15504 {
15505 	int shw_size, vshift;
15506 	struct hme_blk *shw_hblkp;
15507 	uint_t		shw_mask, newshw_mask;
15508 	caddr_t		vaddr;
15509 	int		size;
15510 	cpuset_t cpuset = cpu_ready_set;
15511 
15512 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
15513 
15514 	if (hmebp->hmeblkp == hmeblkp) {
15515 		hmebp->hmeh_nextpa = hmeblkp->hblk_nextpa;
15516 		hmebp->hmeblkp = hmeblkp->hblk_next;
15517 	} else {
15518 		pr_hblk->hblk_nextpa = hmeblkp->hblk_nextpa;
15519 		pr_hblk->hblk_next = hmeblkp->hblk_next;
15520 	}
15521 
15522 	size = get_hblk_ttesz(hmeblkp);
15523 	shw_hblkp = hmeblkp->hblk_shadow;
15524 	if (shw_hblkp) {
15525 		ASSERT(hblktosfmmu(hmeblkp) != KHATID);
15526 		ASSERT(!hmeblkp->hblk_shared);
15527 #ifdef	DEBUG
15528 		if (mmu_page_sizes == max_mmu_page_sizes) {
15529 			ASSERT(size < TTE256M);
15530 		} else {
15531 			ASSERT(size < TTE4M);
15532 		}
15533 #endif /* DEBUG */
15534 
15535 		shw_size = get_hblk_ttesz(shw_hblkp);
15536 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
15537 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
15538 		ASSERT(vshift < 8);
15539 		/*
15540 		 * Atomically clear shadow mask bit
15541 		 */
15542 		do {
15543 			shw_mask = shw_hblkp->hblk_shw_mask;
15544 			ASSERT(shw_mask & (1 << vshift));
15545 			newshw_mask = shw_mask & ~(1 << vshift);
15546 			newshw_mask = atomic_cas_32(&shw_hblkp->hblk_shw_mask,
15547 			    shw_mask, newshw_mask);
15548 		} while (newshw_mask != shw_mask);
15549 		hmeblkp->hblk_shadow = NULL;
15550 	}
15551 	hmeblkp->hblk_shw_bit = 0;
15552 
15553 	if (hmeblkp->hblk_shared) {
15554 #ifdef	DEBUG
15555 		sf_srd_t	*srdp;
15556 		sf_region_t	*rgnp;
15557 		uint_t		rid;
15558 
15559 		srdp = hblktosrd(hmeblkp);
15560 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
15561 		rid = hmeblkp->hblk_tag.htag_rid;
15562 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
15563 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
15564 		rgnp = srdp->srd_hmergnp[rid];
15565 		ASSERT(rgnp != NULL);
15566 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
15567 #endif /* DEBUG */
15568 		hmeblkp->hblk_shared = 0;
15569 	}
15570 	if (free_now) {
15571 		kpreempt_disable();
15572 		CPUSET_DEL(cpuset, CPU->cpu_id);
15573 		xt_sync(cpuset);
15574 		xt_sync(cpuset);
15575 		kpreempt_enable();
15576 
15577 		hmeblkp->hblk_nextpa = HMEBLK_ENDPA;
15578 		hmeblkp->hblk_next = NULL;
15579 	} else {
15580 		/* Append hmeblkp to listp for processing later. */
15581 		hmeblkp->hblk_next = *listp;
15582 		*listp = hmeblkp;
15583 	}
15584 }
15585 
15586 /*
15587  * This routine is called when memory is in short supply and returns a free
15588  * hmeblk of the requested size from the cpu pending lists.
15589  */
15590 static struct hme_blk *
15591 sfmmu_check_pending_hblks(int size)
15592 {
15593 	int i;
15594 	struct hme_blk *hmeblkp = NULL, *last_hmeblkp;
15595 	int found_hmeblk;
15596 	cpuset_t cpuset = cpu_ready_set;
15597 	cpu_hme_pend_t *cpuhp;
15598 
15599 	/* Flush cpu hblk pending queues */
15600 	for (i = 0; i < NCPU; i++) {
15601 		cpuhp = &cpu_hme_pend[i];
15602 		if (cpuhp->chp_listp != NULL)  {
15603 			mutex_enter(&cpuhp->chp_mutex);
15604 			if (cpuhp->chp_listp == NULL)  {
15605 				mutex_exit(&cpuhp->chp_mutex);
15606 				continue;
15607 			}
15608 			found_hmeblk = 0;
15609 			last_hmeblkp = NULL;
15610 			for (hmeblkp = cpuhp->chp_listp; hmeblkp != NULL;
15611 			    hmeblkp = hmeblkp->hblk_next) {
15612 				if (get_hblk_ttesz(hmeblkp) == size) {
15613 					if (last_hmeblkp == NULL) {
15614 						cpuhp->chp_listp =
15615 						    hmeblkp->hblk_next;
15616 					} else {
15617 						last_hmeblkp->hblk_next =
15618 						    hmeblkp->hblk_next;
15619 					}
15620 					ASSERT(cpuhp->chp_count > 0);
15621 					cpuhp->chp_count--;
15622 					found_hmeblk = 1;
15623 					break;
15624 				} else {
15625 					last_hmeblkp = hmeblkp;
15626 				}
15627 			}
15628 			mutex_exit(&cpuhp->chp_mutex);
15629 
15630 			if (found_hmeblk) {
15631 				kpreempt_disable();
15632 				CPUSET_DEL(cpuset, CPU->cpu_id);
15633 				xt_sync(cpuset);
15634 				xt_sync(cpuset);
15635 				kpreempt_enable();
15636 				return (hmeblkp);
15637 			}
15638 		}
15639 	}
15640 	return (NULL);
15641 }
15642