xref: /titanic_51/usr/src/uts/sfmmu/vm/hat_sfmmu.c (revision 72b4c2cd5756cbce91ec4b298168685490456787)
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 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * VM - Hardware Address Translation management for Spitfire MMU.
30  *
31  * This file implements the machine specific hardware translation
32  * needed by the VM system.  The machine independent interface is
33  * described in <vm/hat.h> while the machine dependent interface
34  * and data structures are described in <vm/hat_sfmmu.h>.
35  *
36  * The hat layer manages the address translation hardware as a cache
37  * driven by calls from the higher levels in the VM system.
38  */
39 
40 #include <sys/types.h>
41 #include <sys/kstat.h>
42 #include <vm/hat.h>
43 #include <vm/hat_sfmmu.h>
44 #include <vm/page.h>
45 #include <sys/pte.h>
46 #include <sys/systm.h>
47 #include <sys/mman.h>
48 #include <sys/sysmacros.h>
49 #include <sys/machparam.h>
50 #include <sys/vtrace.h>
51 #include <sys/kmem.h>
52 #include <sys/mmu.h>
53 #include <sys/cmn_err.h>
54 #include <sys/cpu.h>
55 #include <sys/cpuvar.h>
56 #include <sys/debug.h>
57 #include <sys/lgrp.h>
58 #include <sys/archsystm.h>
59 #include <sys/machsystm.h>
60 #include <sys/vmsystm.h>
61 #include <vm/as.h>
62 #include <vm/seg.h>
63 #include <vm/seg_kp.h>
64 #include <vm/seg_kmem.h>
65 #include <vm/seg_kpm.h>
66 #include <vm/rm.h>
67 #include <sys/t_lock.h>
68 #include <sys/obpdefs.h>
69 #include <sys/vm_machparam.h>
70 #include <sys/var.h>
71 #include <sys/trap.h>
72 #include <sys/machtrap.h>
73 #include <sys/scb.h>
74 #include <sys/bitmap.h>
75 #include <sys/machlock.h>
76 #include <sys/membar.h>
77 #include <sys/atomic.h>
78 #include <sys/cpu_module.h>
79 #include <sys/prom_debug.h>
80 #include <sys/ksynch.h>
81 #include <sys/mem_config.h>
82 #include <sys/mem_cage.h>
83 #include <vm/vm_dep.h>
84 #include <vm/xhat_sfmmu.h>
85 #include <sys/fpu/fpusystm.h>
86 #include <vm/mach_kpm.h>
87 #include <sys/callb.h>
88 
89 #ifdef	DEBUG
90 #define	SFMMU_VALIDATE_HMERID(hat, rid, saddr, len)			\
91 	if (SFMMU_IS_SHMERID_VALID(rid)) {				\
92 		caddr_t _eaddr = (saddr) + (len);			\
93 		sf_srd_t *_srdp;					\
94 		sf_region_t *_rgnp;					\
95 		ASSERT((rid) < SFMMU_MAX_HME_REGIONS);			\
96 		ASSERT(SF_RGNMAP_TEST(hat->sfmmu_hmeregion_map, rid));	\
97 		ASSERT((hat) != ksfmmup);				\
98 		_srdp = (hat)->sfmmu_srdp;				\
99 		ASSERT(_srdp != NULL);					\
100 		ASSERT(_srdp->srd_refcnt != 0);				\
101 		_rgnp = _srdp->srd_hmergnp[(rid)];			\
102 		ASSERT(_rgnp != NULL && _rgnp->rgn_id == rid);		\
103 		ASSERT(_rgnp->rgn_refcnt != 0);				\
104 		ASSERT(!(_rgnp->rgn_flags & SFMMU_REGION_FREE));	\
105 		ASSERT((_rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) ==	\
106 		    SFMMU_REGION_HME);					\
107 		ASSERT((saddr) >= _rgnp->rgn_saddr);			\
108 		ASSERT((saddr) < _rgnp->rgn_saddr + _rgnp->rgn_size);	\
109 		ASSERT(_eaddr > _rgnp->rgn_saddr);			\
110 		ASSERT(_eaddr <= _rgnp->rgn_saddr + _rgnp->rgn_size);	\
111 	}
112 
113 #define	SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid) 	 	 \
114 {						 			 \
115 		caddr_t _hsva;						 \
116 		caddr_t _heva;						 \
117 		caddr_t _rsva;					 	 \
118 		caddr_t _reva;					 	 \
119 		int	_ttesz = get_hblk_ttesz(hmeblkp);		 \
120 		int	_flagtte;					 \
121 		ASSERT((srdp)->srd_refcnt != 0);			 \
122 		ASSERT((rid) < SFMMU_MAX_HME_REGIONS);			 \
123 		ASSERT((rgnp)->rgn_id == rid);				 \
124 		ASSERT(!((rgnp)->rgn_flags & SFMMU_REGION_FREE));	 \
125 		ASSERT(((rgnp)->rgn_flags & SFMMU_REGION_TYPE_MASK) ==	 \
126 		    SFMMU_REGION_HME);					 \
127 		ASSERT(_ttesz <= (rgnp)->rgn_pgszc);			 \
128 		_hsva = (caddr_t)get_hblk_base(hmeblkp);		 \
129 		_heva = get_hblk_endaddr(hmeblkp);			 \
130 		_rsva = (caddr_t)P2ALIGN(				 \
131 		    (uintptr_t)(rgnp)->rgn_saddr, HBLK_MIN_BYTES);	 \
132 		_reva = (caddr_t)P2ROUNDUP(				 \
133 		    (uintptr_t)((rgnp)->rgn_saddr + (rgnp)->rgn_size),	 \
134 		    HBLK_MIN_BYTES);					 \
135 		ASSERT(_hsva >= _rsva);				 	 \
136 		ASSERT(_hsva < _reva);				 	 \
137 		ASSERT(_heva > _rsva);				 	 \
138 		ASSERT(_heva <= _reva);				 	 \
139 		_flagtte = (_ttesz < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ :  \
140 			_ttesz;						 \
141 		ASSERT(rgnp->rgn_hmeflags & (0x1 << _flagtte));		 \
142 }
143 
144 #else /* DEBUG */
145 #define	SFMMU_VALIDATE_HMERID(hat, rid, addr, len)
146 #define	SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid)
147 #endif /* DEBUG */
148 
149 #if defined(SF_ERRATA_57)
150 extern caddr_t errata57_limit;
151 #endif
152 
153 #define	HME8BLK_SZ_RND		((roundup(HME8BLK_SZ, sizeof (int64_t))) /  \
154 				(sizeof (int64_t)))
155 #define	HBLK_RESERVE		((struct hme_blk *)hblk_reserve)
156 
157 #define	HBLK_RESERVE_CNT	128
158 #define	HBLK_RESERVE_MIN	20
159 
160 static struct hme_blk		*freehblkp;
161 static kmutex_t			freehblkp_lock;
162 static int			freehblkcnt;
163 
164 static int64_t			hblk_reserve[HME8BLK_SZ_RND];
165 static kmutex_t			hblk_reserve_lock;
166 static kthread_t		*hblk_reserve_thread;
167 
168 static nucleus_hblk8_info_t	nucleus_hblk8;
169 static nucleus_hblk1_info_t	nucleus_hblk1;
170 
171 /*
172  * SFMMU specific hat functions
173  */
174 void	hat_pagecachectl(struct page *, int);
175 
176 /* flags for hat_pagecachectl */
177 #define	HAT_CACHE	0x1
178 #define	HAT_UNCACHE	0x2
179 #define	HAT_TMPNC	0x4
180 
181 /*
182  * Flag to allow the creation of non-cacheable translations
183  * to system memory. It is off by default. At the moment this
184  * flag is used by the ecache error injector. The error injector
185  * will turn it on when creating such a translation then shut it
186  * off when it's finished.
187  */
188 
189 int	sfmmu_allow_nc_trans = 0;
190 
191 /*
192  * Flag to disable large page support.
193  * 	value of 1 => disable all large pages.
194  *	bits 1, 2, and 3 are to disable 64K, 512K and 4M pages respectively.
195  *
196  * For example, use the value 0x4 to disable 512K pages.
197  *
198  */
199 #define	LARGE_PAGES_OFF		0x1
200 
201 /*
202  * The disable_large_pages and disable_ism_large_pages variables control
203  * hat_memload_array and the page sizes to be used by ISM and the kernel.
204  *
205  * The disable_auto_data_large_pages and disable_auto_text_large_pages variables
206  * are only used to control which OOB pages to use at upper VM segment creation
207  * time, and are set in hat_init_pagesizes and used in the map_pgsz* routines.
208  * Their values may come from platform or CPU specific code to disable page
209  * sizes that should not be used.
210  *
211  * WARNING: 512K pages are currently not supported for ISM/DISM.
212  */
213 uint_t	disable_large_pages = 0;
214 uint_t	disable_ism_large_pages = (1 << TTE512K);
215 uint_t	disable_auto_data_large_pages = 0;
216 uint_t	disable_auto_text_large_pages = 0;
217 
218 /*
219  * Private sfmmu data structures for hat management
220  */
221 static struct kmem_cache *sfmmuid_cache;
222 static struct kmem_cache *mmuctxdom_cache;
223 
224 /*
225  * Private sfmmu data structures for tsb management
226  */
227 static struct kmem_cache *sfmmu_tsbinfo_cache;
228 static struct kmem_cache *sfmmu_tsb8k_cache;
229 static struct kmem_cache *sfmmu_tsb_cache[NLGRPS_MAX];
230 static vmem_t *kmem_bigtsb_arena;
231 static vmem_t *kmem_tsb_arena;
232 
233 /*
234  * sfmmu static variables for hmeblk resource management.
235  */
236 static vmem_t *hat_memload1_arena; /* HAT translation arena for sfmmu1_cache */
237 static struct kmem_cache *sfmmu8_cache;
238 static struct kmem_cache *sfmmu1_cache;
239 static struct kmem_cache *pa_hment_cache;
240 
241 static kmutex_t 	ism_mlist_lock;	/* mutex for ism mapping list */
242 /*
243  * private data for ism
244  */
245 static struct kmem_cache *ism_blk_cache;
246 static struct kmem_cache *ism_ment_cache;
247 #define	ISMID_STARTADDR	NULL
248 
249 /*
250  * Region management data structures and function declarations.
251  */
252 
253 static void	sfmmu_leave_srd(sfmmu_t *);
254 static int	sfmmu_srdcache_constructor(void *, void *, int);
255 static void	sfmmu_srdcache_destructor(void *, void *);
256 static int	sfmmu_rgncache_constructor(void *, void *, int);
257 static void	sfmmu_rgncache_destructor(void *, void *);
258 static int	sfrgnmap_isnull(sf_region_map_t *);
259 static int	sfhmergnmap_isnull(sf_hmeregion_map_t *);
260 static int	sfmmu_scdcache_constructor(void *, void *, int);
261 static void	sfmmu_scdcache_destructor(void *, void *);
262 static void	sfmmu_rgn_cb_noop(caddr_t, caddr_t, caddr_t,
263     size_t, void *, u_offset_t);
264 
265 static uint_t srd_hashmask = SFMMU_MAX_SRD_BUCKETS - 1;
266 static sf_srd_bucket_t *srd_buckets;
267 static struct kmem_cache *srd_cache;
268 static uint_t srd_rgn_hashmask = SFMMU_MAX_REGION_BUCKETS - 1;
269 static struct kmem_cache *region_cache;
270 static struct kmem_cache *scd_cache;
271 
272 #ifdef sun4v
273 int use_bigtsb_arena = 1;
274 #else
275 int use_bigtsb_arena = 0;
276 #endif
277 
278 /* External /etc/system tunable, for turning on&off the shctx support */
279 int disable_shctx = 0;
280 /* Internal variable, set by MD if the HW supports shctx feature */
281 int shctx_on = 0;
282 
283 #ifdef DEBUG
284 static void check_scd_sfmmu_list(sfmmu_t **, sfmmu_t *, int);
285 #endif
286 static void sfmmu_to_scd_list(sfmmu_t **, sfmmu_t *);
287 static void sfmmu_from_scd_list(sfmmu_t **, sfmmu_t *);
288 
289 static sf_scd_t *sfmmu_alloc_scd(sf_srd_t *, sf_region_map_t *);
290 static void sfmmu_find_scd(sfmmu_t *);
291 static void sfmmu_join_scd(sf_scd_t *, sfmmu_t *);
292 static void sfmmu_finish_join_scd(sfmmu_t *);
293 static void sfmmu_leave_scd(sfmmu_t *, uchar_t);
294 static void sfmmu_destroy_scd(sf_srd_t *, sf_scd_t *, sf_region_map_t *);
295 static int sfmmu_alloc_scd_tsbs(sf_srd_t *, sf_scd_t *);
296 static void sfmmu_free_scd_tsbs(sfmmu_t *);
297 static void sfmmu_tsb_inv_ctx(sfmmu_t *);
298 static int find_ism_rid(sfmmu_t *, sfmmu_t *, caddr_t, uint_t *);
299 static void sfmmu_ism_hatflags(sfmmu_t *, int);
300 static int sfmmu_srd_lock_held(sf_srd_t *);
301 static void sfmmu_remove_scd(sf_scd_t **, sf_scd_t *);
302 static void sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *);
303 static void sfmmu_link_scd_to_regions(sf_srd_t *, sf_scd_t *);
304 static void sfmmu_unlink_scd_from_regions(sf_srd_t *, sf_scd_t *);
305 static void sfmmu_link_to_hmeregion(sfmmu_t *, sf_region_t *);
306 static void sfmmu_unlink_from_hmeregion(sfmmu_t *, sf_region_t *);
307 
308 /*
309  * ``hat_lock'' is a hashed mutex lock for protecting sfmmu TSB lists,
310  * HAT flags, synchronizing TLB/TSB coherency, and context management.
311  * The lock is hashed on the sfmmup since the case where we need to lock
312  * all processes is rare but does occur (e.g. we need to unload a shared
313  * mapping from all processes using the mapping).  We have a lot of buckets,
314  * and each slab of sfmmu_t's can use about a quarter of them, giving us
315  * a fairly good distribution without wasting too much space and overhead
316  * when we have to grab them all.
317  */
318 #define	SFMMU_NUM_LOCK	128		/* must be power of two */
319 hatlock_t	hat_lock[SFMMU_NUM_LOCK];
320 
321 /*
322  * Hash algorithm optimized for a small number of slabs.
323  *  7 is (highbit((sizeof sfmmu_t)) - 1)
324  * This hash algorithm is based upon the knowledge that sfmmu_t's come from a
325  * kmem_cache, and thus they will be sequential within that cache.  In
326  * addition, each new slab will have a different "color" up to cache_maxcolor
327  * which will skew the hashing for each successive slab which is allocated.
328  * If the size of sfmmu_t changed to a larger size, this algorithm may need
329  * to be revisited.
330  */
331 #define	TSB_HASH_SHIFT_BITS (7)
332 #define	PTR_HASH(x) ((uintptr_t)x >> TSB_HASH_SHIFT_BITS)
333 
334 #ifdef DEBUG
335 int tsb_hash_debug = 0;
336 #define	TSB_HASH(sfmmup)	\
337 	(tsb_hash_debug ? &hat_lock[0] : \
338 	&hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)])
339 #else	/* DEBUG */
340 #define	TSB_HASH(sfmmup)	&hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)]
341 #endif	/* DEBUG */
342 
343 
344 /* sfmmu_replace_tsb() return codes. */
345 typedef enum tsb_replace_rc {
346 	TSB_SUCCESS,
347 	TSB_ALLOCFAIL,
348 	TSB_LOSTRACE,
349 	TSB_ALREADY_SWAPPED,
350 	TSB_CANTGROW
351 } tsb_replace_rc_t;
352 
353 /*
354  * Flags for TSB allocation routines.
355  */
356 #define	TSB_ALLOC	0x01
357 #define	TSB_FORCEALLOC	0x02
358 #define	TSB_GROW	0x04
359 #define	TSB_SHRINK	0x08
360 #define	TSB_SWAPIN	0x10
361 
362 /*
363  * Support for HAT callbacks.
364  */
365 #define	SFMMU_MAX_RELOC_CALLBACKS	10
366 int sfmmu_max_cb_id = SFMMU_MAX_RELOC_CALLBACKS;
367 static id_t sfmmu_cb_nextid = 0;
368 static id_t sfmmu_tsb_cb_id;
369 struct sfmmu_callback *sfmmu_cb_table;
370 
371 /*
372  * Kernel page relocation is enabled by default for non-caged
373  * kernel pages.  This has little effect unless segkmem_reloc is
374  * set, since by default kernel memory comes from inside the
375  * kernel cage.
376  */
377 int hat_kpr_enabled = 1;
378 
379 kmutex_t	kpr_mutex;
380 kmutex_t	kpr_suspendlock;
381 kthread_t	*kreloc_thread;
382 
383 /*
384  * Enable VA->PA translation sanity checking on DEBUG kernels.
385  * Disabled by default.  This is incompatible with some
386  * drivers (error injector, RSM) so if it breaks you get
387  * to keep both pieces.
388  */
389 int hat_check_vtop = 0;
390 
391 /*
392  * Private sfmmu routines (prototypes)
393  */
394 static struct hme_blk *sfmmu_shadow_hcreate(sfmmu_t *, caddr_t, int, uint_t);
395 static struct 	hme_blk *sfmmu_hblk_alloc(sfmmu_t *, caddr_t,
396 			struct hmehash_bucket *, uint_t, hmeblk_tag, uint_t,
397 			uint_t);
398 static caddr_t	sfmmu_hblk_unload(struct hat *, struct hme_blk *, caddr_t,
399 			caddr_t, demap_range_t *, uint_t);
400 static caddr_t	sfmmu_hblk_sync(struct hat *, struct hme_blk *, caddr_t,
401 			caddr_t, int);
402 static void	sfmmu_hblk_free(struct hmehash_bucket *, struct hme_blk *,
403 			uint64_t, struct hme_blk **);
404 static void	sfmmu_hblks_list_purge(struct hme_blk **);
405 static uint_t	sfmmu_get_free_hblk(struct hme_blk **, uint_t);
406 static uint_t	sfmmu_put_free_hblk(struct hme_blk *, uint_t);
407 static struct hme_blk *sfmmu_hblk_steal(int);
408 static int	sfmmu_steal_this_hblk(struct hmehash_bucket *,
409 			struct hme_blk *, uint64_t, uint64_t,
410 			struct hme_blk *);
411 static caddr_t	sfmmu_hblk_unlock(struct hme_blk *, caddr_t, caddr_t);
412 
413 static void	hat_do_memload_array(struct hat *, caddr_t, size_t,
414 		    struct page **, uint_t, uint_t, uint_t);
415 static void	hat_do_memload(struct hat *, caddr_t, struct page *,
416 		    uint_t, uint_t, uint_t);
417 static void	sfmmu_memload_batchsmall(struct hat *, caddr_t, page_t **,
418 		    uint_t, uint_t, pgcnt_t, uint_t);
419 void		sfmmu_tteload(struct hat *, tte_t *, caddr_t, page_t *,
420 			uint_t);
421 static int	sfmmu_tteload_array(sfmmu_t *, tte_t *, caddr_t, page_t **,
422 			uint_t, uint_t);
423 static struct hmehash_bucket *sfmmu_tteload_acquire_hashbucket(sfmmu_t *,
424 					caddr_t, int, uint_t);
425 static struct hme_blk *sfmmu_tteload_find_hmeblk(sfmmu_t *,
426 			struct hmehash_bucket *, caddr_t, uint_t, uint_t,
427 			uint_t);
428 static int	sfmmu_tteload_addentry(sfmmu_t *, struct hme_blk *, tte_t *,
429 			caddr_t, page_t **, uint_t, uint_t);
430 static void	sfmmu_tteload_release_hashbucket(struct hmehash_bucket *);
431 
432 static int	sfmmu_pagearray_setup(caddr_t, page_t **, tte_t *, int);
433 static pfn_t	sfmmu_uvatopfn(caddr_t, sfmmu_t *, tte_t *);
434 void		sfmmu_memtte(tte_t *, pfn_t, uint_t, int);
435 #ifdef VAC
436 static void	sfmmu_vac_conflict(struct hat *, caddr_t, page_t *);
437 static int	sfmmu_vacconflict_array(caddr_t, page_t *, int *);
438 int	tst_tnc(page_t *pp, pgcnt_t);
439 void	conv_tnc(page_t *pp, int);
440 #endif
441 
442 static void	sfmmu_get_ctx(sfmmu_t *);
443 static void	sfmmu_free_sfmmu(sfmmu_t *);
444 
445 static void	sfmmu_ttesync(struct hat *, caddr_t, tte_t *, page_t *);
446 static void	sfmmu_chgattr(struct hat *, caddr_t, size_t, uint_t, int);
447 
448 cpuset_t	sfmmu_pageunload(page_t *, struct sf_hment *, int);
449 static void	hat_pagereload(struct page *, struct page *);
450 static cpuset_t	sfmmu_pagesync(page_t *, struct sf_hment *, uint_t);
451 #ifdef VAC
452 void	sfmmu_page_cache_array(page_t *, int, int, pgcnt_t);
453 static void	sfmmu_page_cache(page_t *, int, int, int);
454 #endif
455 
456 cpuset_t	sfmmu_rgntlb_demap(caddr_t, sf_region_t *,
457     struct hme_blk *, int);
458 static void	sfmmu_tlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *,
459 			pfn_t, int, int, int, int);
460 static void	sfmmu_ismtlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *,
461 			pfn_t, int);
462 static void	sfmmu_tlb_demap(caddr_t, sfmmu_t *, struct hme_blk *, int, int);
463 static void	sfmmu_tlb_range_demap(demap_range_t *);
464 static void	sfmmu_invalidate_ctx(sfmmu_t *);
465 static void	sfmmu_sync_mmustate(sfmmu_t *);
466 
467 static void 	sfmmu_tsbinfo_setup_phys(struct tsb_info *, pfn_t);
468 static int	sfmmu_tsbinfo_alloc(struct tsb_info **, int, int, uint_t,
469 			sfmmu_t *);
470 static void	sfmmu_tsb_free(struct tsb_info *);
471 static void	sfmmu_tsbinfo_free(struct tsb_info *);
472 static int	sfmmu_init_tsbinfo(struct tsb_info *, int, int, uint_t,
473 			sfmmu_t *);
474 static void	sfmmu_tsb_chk_reloc(sfmmu_t *, hatlock_t *);
475 static void	sfmmu_tsb_swapin(sfmmu_t *, hatlock_t *);
476 static int	sfmmu_select_tsb_szc(pgcnt_t);
477 static void	sfmmu_mod_tsb(sfmmu_t *, caddr_t, tte_t *, int);
478 #define		sfmmu_load_tsb(sfmmup, vaddr, tte, szc) \
479 	sfmmu_mod_tsb(sfmmup, vaddr, tte, szc)
480 #define		sfmmu_unload_tsb(sfmmup, vaddr, szc)    \
481 	sfmmu_mod_tsb(sfmmup, vaddr, NULL, szc)
482 static void	sfmmu_copy_tsb(struct tsb_info *, struct tsb_info *);
483 static tsb_replace_rc_t sfmmu_replace_tsb(sfmmu_t *, struct tsb_info *, uint_t,
484     hatlock_t *, uint_t);
485 static void	sfmmu_size_tsb(sfmmu_t *, int, uint64_t, uint64_t, int);
486 
487 #ifdef VAC
488 void	sfmmu_cache_flush(pfn_t, int);
489 void	sfmmu_cache_flushcolor(int, pfn_t);
490 #endif
491 static caddr_t	sfmmu_hblk_chgattr(sfmmu_t *, struct hme_blk *, caddr_t,
492 			caddr_t, demap_range_t *, uint_t, int);
493 
494 static uint64_t	sfmmu_vtop_attr(uint_t, int mode, tte_t *);
495 static uint_t	sfmmu_ptov_attr(tte_t *);
496 static caddr_t	sfmmu_hblk_chgprot(sfmmu_t *, struct hme_blk *, caddr_t,
497 			caddr_t, demap_range_t *, uint_t);
498 static uint_t	sfmmu_vtop_prot(uint_t, uint_t *);
499 static int	sfmmu_idcache_constructor(void *, void *, int);
500 static void	sfmmu_idcache_destructor(void *, void *);
501 static int	sfmmu_hblkcache_constructor(void *, void *, int);
502 static void	sfmmu_hblkcache_destructor(void *, void *);
503 static void	sfmmu_hblkcache_reclaim(void *);
504 static void	sfmmu_shadow_hcleanup(sfmmu_t *, struct hme_blk *,
505 			struct hmehash_bucket *);
506 static void	sfmmu_free_hblks(sfmmu_t *, caddr_t, caddr_t, int);
507 static void	sfmmu_cleanup_rhblk(sf_srd_t *, caddr_t, uint_t, int);
508 static void	sfmmu_unload_hmeregion_va(sf_srd_t *, uint_t, caddr_t, caddr_t,
509 			int, caddr_t *);
510 static void	sfmmu_unload_hmeregion(sf_srd_t *, sf_region_t *);
511 
512 static void	sfmmu_rm_large_mappings(page_t *, int);
513 
514 static void	hat_lock_init(void);
515 static void	hat_kstat_init(void);
516 static int	sfmmu_kstat_percpu_update(kstat_t *ksp, int rw);
517 static void	sfmmu_set_scd_rttecnt(sf_srd_t *, sf_scd_t *);
518 static	int	sfmmu_is_rgnva(sf_srd_t *, caddr_t, ulong_t, ulong_t);
519 static void	sfmmu_check_page_sizes(sfmmu_t *, int);
520 int	fnd_mapping_sz(page_t *);
521 static void	iment_add(struct ism_ment *,  struct hat *);
522 static void	iment_sub(struct ism_ment *, struct hat *);
523 static pgcnt_t	ism_tsb_entries(sfmmu_t *, int szc);
524 extern void	sfmmu_setup_tsbinfo(sfmmu_t *);
525 extern void	sfmmu_clear_utsbinfo(void);
526 
527 static void	sfmmu_ctx_wrap_around(mmu_ctx_t *);
528 
529 /* kpm globals */
530 #ifdef	DEBUG
531 /*
532  * Enable trap level tsbmiss handling
533  */
534 int	kpm_tsbmtl = 1;
535 
536 /*
537  * Flush the TLB on kpm mapout. Note: Xcalls are used (again) for the
538  * required TLB shootdowns in this case, so handle w/ care. Off by default.
539  */
540 int	kpm_tlb_flush;
541 #endif	/* DEBUG */
542 
543 static void	*sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *, size_t, int);
544 
545 #ifdef DEBUG
546 static void	sfmmu_check_hblk_flist();
547 #endif
548 
549 /*
550  * Semi-private sfmmu data structures.  Some of them are initialize in
551  * startup or in hat_init. Some of them are private but accessed by
552  * assembly code or mach_sfmmu.c
553  */
554 struct hmehash_bucket *uhme_hash;	/* user hmeblk hash table */
555 struct hmehash_bucket *khme_hash;	/* kernel hmeblk hash table */
556 uint64_t	uhme_hash_pa;		/* PA of uhme_hash */
557 uint64_t	khme_hash_pa;		/* PA of khme_hash */
558 int 		uhmehash_num;		/* # of buckets in user hash table */
559 int 		khmehash_num;		/* # of buckets in kernel hash table */
560 
561 uint_t		max_mmu_ctxdoms = 0;	/* max context domains in the system */
562 mmu_ctx_t	**mmu_ctxs_tbl;		/* global array of context domains */
563 uint64_t	mmu_saved_gnum = 0;	/* to init incoming MMUs' gnums */
564 
565 #define	DEFAULT_NUM_CTXS_PER_MMU 8192
566 static uint_t	nctxs = DEFAULT_NUM_CTXS_PER_MMU;
567 
568 int		cache;			/* describes system cache */
569 
570 caddr_t		ktsb_base;		/* kernel 8k-indexed tsb base address */
571 uint64_t	ktsb_pbase;		/* kernel 8k-indexed tsb phys address */
572 int		ktsb_szcode;		/* kernel 8k-indexed tsb size code */
573 int		ktsb_sz;		/* kernel 8k-indexed tsb size */
574 
575 caddr_t		ktsb4m_base;		/* kernel 4m-indexed tsb base address */
576 uint64_t	ktsb4m_pbase;		/* kernel 4m-indexed tsb phys address */
577 int		ktsb4m_szcode;		/* kernel 4m-indexed tsb size code */
578 int		ktsb4m_sz;		/* kernel 4m-indexed tsb size */
579 
580 uint64_t	kpm_tsbbase;		/* kernel seg_kpm 4M TSB base address */
581 int		kpm_tsbsz;		/* kernel seg_kpm 4M TSB size code */
582 uint64_t	kpmsm_tsbbase;		/* kernel seg_kpm 8K TSB base address */
583 int		kpmsm_tsbsz;		/* kernel seg_kpm 8K TSB size code */
584 
585 #ifndef sun4v
586 int		utsb_dtlb_ttenum = -1;	/* index in TLB for utsb locked TTE */
587 int		utsb4m_dtlb_ttenum = -1; /* index in TLB for 4M TSB TTE */
588 int		dtlb_resv_ttenum;	/* index in TLB of first reserved TTE */
589 caddr_t		utsb_vabase;		/* reserved kernel virtual memory */
590 caddr_t		utsb4m_vabase;		/* for trap handler TSB accesses */
591 #endif /* sun4v */
592 uint64_t	tsb_alloc_bytes = 0;	/* bytes allocated to TSBs */
593 vmem_t		*kmem_tsb_default_arena[NLGRPS_MAX];	/* For dynamic TSBs */
594 vmem_t		*kmem_bigtsb_default_arena[NLGRPS_MAX]; /* dynamic 256M TSBs */
595 
596 /*
597  * Size to use for TSB slabs.  Future platforms that support page sizes
598  * larger than 4M may wish to change these values, and provide their own
599  * assembly macros for building and decoding the TSB base register contents.
600  * Note disable_large_pages will override the value set here.
601  */
602 static	uint_t tsb_slab_ttesz = TTE4M;
603 size_t	tsb_slab_size = MMU_PAGESIZE4M;
604 uint_t	tsb_slab_shift = MMU_PAGESHIFT4M;
605 /* PFN mask for TTE */
606 size_t	tsb_slab_mask = MMU_PAGEOFFSET4M >> MMU_PAGESHIFT;
607 
608 /*
609  * Size to use for TSB slabs.  These are used only when 256M tsb arenas
610  * exist.
611  */
612 static uint_t	bigtsb_slab_ttesz = TTE256M;
613 static size_t	bigtsb_slab_size = MMU_PAGESIZE256M;
614 static uint_t	bigtsb_slab_shift = MMU_PAGESHIFT256M;
615 /* 256M page alignment for 8K pfn */
616 static size_t	bigtsb_slab_mask = MMU_PAGEOFFSET256M >> MMU_PAGESHIFT;
617 
618 /* largest TSB size to grow to, will be smaller on smaller memory systems */
619 static int	tsb_max_growsize = 0;
620 
621 /*
622  * Tunable parameters dealing with TSB policies.
623  */
624 
625 /*
626  * This undocumented tunable forces all 8K TSBs to be allocated from
627  * the kernel heap rather than from the kmem_tsb_default_arena arenas.
628  */
629 #ifdef	DEBUG
630 int	tsb_forceheap = 0;
631 #endif	/* DEBUG */
632 
633 /*
634  * Decide whether to use per-lgroup arenas, or one global set of
635  * TSB arenas.  The default is not to break up per-lgroup, since
636  * most platforms don't recognize any tangible benefit from it.
637  */
638 int	tsb_lgrp_affinity = 0;
639 
640 /*
641  * Used for growing the TSB based on the process RSS.
642  * tsb_rss_factor is based on the smallest TSB, and is
643  * shifted by the TSB size to determine if we need to grow.
644  * The default will grow the TSB if the number of TTEs for
645  * this page size exceeds 75% of the number of TSB entries,
646  * which should _almost_ eliminate all conflict misses
647  * (at the expense of using up lots and lots of memory).
648  */
649 #define	TSB_RSS_FACTOR		(TSB_ENTRIES(TSB_MIN_SZCODE) * 0.75)
650 #define	SFMMU_RSS_TSBSIZE(tsbszc)	(tsb_rss_factor << tsbszc)
651 #define	SELECT_TSB_SIZECODE(pgcnt) ( \
652 	(enable_tsb_rss_sizing)? sfmmu_select_tsb_szc(pgcnt) : \
653 	default_tsb_size)
654 #define	TSB_OK_SHRINK()	\
655 	(tsb_alloc_bytes > tsb_alloc_hiwater || freemem < desfree)
656 #define	TSB_OK_GROW()	\
657 	(tsb_alloc_bytes < tsb_alloc_hiwater && freemem > desfree)
658 
659 int	enable_tsb_rss_sizing = 1;
660 int	tsb_rss_factor	= (int)TSB_RSS_FACTOR;
661 
662 /* which TSB size code to use for new address spaces or if rss sizing off */
663 int default_tsb_size = TSB_8K_SZCODE;
664 
665 static uint64_t tsb_alloc_hiwater; /* limit TSB reserved memory */
666 uint64_t tsb_alloc_hiwater_factor; /* tsb_alloc_hiwater = physmem / this */
667 #define	TSB_ALLOC_HIWATER_FACTOR_DEFAULT	32
668 
669 #ifdef DEBUG
670 static int tsb_random_size = 0;	/* set to 1 to test random tsb sizes on alloc */
671 static int tsb_grow_stress = 0;	/* if set to 1, keep replacing TSB w/ random */
672 static int tsb_alloc_mtbf = 0;	/* fail allocation every n attempts */
673 static int tsb_alloc_fail_mtbf = 0;
674 static int tsb_alloc_count = 0;
675 #endif /* DEBUG */
676 
677 /* if set to 1, will remap valid TTEs when growing TSB. */
678 int tsb_remap_ttes = 1;
679 
680 /*
681  * If we have more than this many mappings, allocate a second TSB.
682  * This default is chosen because the I/D fully associative TLBs are
683  * assumed to have at least 8 available entries. Platforms with a
684  * larger fully-associative TLB could probably override the default.
685  */
686 
687 #ifdef sun4v
688 int tsb_sectsb_threshold = 0;
689 #else
690 int tsb_sectsb_threshold = 8;
691 #endif
692 
693 /*
694  * kstat data
695  */
696 struct sfmmu_global_stat sfmmu_global_stat;
697 struct sfmmu_tsbsize_stat sfmmu_tsbsize_stat;
698 
699 /*
700  * Global data
701  */
702 sfmmu_t 	*ksfmmup;		/* kernel's hat id */
703 
704 #ifdef DEBUG
705 static void	chk_tte(tte_t *, tte_t *, tte_t *, struct hme_blk *);
706 #endif
707 
708 /* sfmmu locking operations */
709 static kmutex_t *sfmmu_mlspl_enter(struct page *, int);
710 static int	sfmmu_mlspl_held(struct page *, int);
711 
712 kmutex_t *sfmmu_page_enter(page_t *);
713 void	sfmmu_page_exit(kmutex_t *);
714 int	sfmmu_page_spl_held(struct page *);
715 
716 /* sfmmu internal locking operations - accessed directly */
717 static void	sfmmu_mlist_reloc_enter(page_t *, page_t *,
718 				kmutex_t **, kmutex_t **);
719 static void	sfmmu_mlist_reloc_exit(kmutex_t *, kmutex_t *);
720 static hatlock_t *
721 		sfmmu_hat_enter(sfmmu_t *);
722 static hatlock_t *
723 		sfmmu_hat_tryenter(sfmmu_t *);
724 static void	sfmmu_hat_exit(hatlock_t *);
725 static void	sfmmu_hat_lock_all(void);
726 static void	sfmmu_hat_unlock_all(void);
727 static void	sfmmu_ismhat_enter(sfmmu_t *, int);
728 static void	sfmmu_ismhat_exit(sfmmu_t *, int);
729 
730 /*
731  * Array of mutexes protecting a page's mapping list and p_nrm field.
732  *
733  * The hash function looks complicated, but is made up so that:
734  *
735  * "pp" not shifted, so adjacent pp values will hash to different cache lines
736  *  (8 byte alignment * 8 bytes/mutes == 64 byte coherency subblock)
737  *
738  * "pp" >> mml_shift, incorporates more source bits into the hash result
739  *
740  *  "& (mml_table_size - 1), should be faster than using remainder "%"
741  *
742  * Hopefully, mml_table, mml_table_size and mml_shift are all in the same
743  * cacheline, since they get declared next to each other below. We'll trust
744  * ld not to do something random.
745  */
746 #ifdef	DEBUG
747 int mlist_hash_debug = 0;
748 #define	MLIST_HASH(pp)	(mlist_hash_debug ? &mml_table[0] : \
749 	&mml_table[((uintptr_t)(pp) + \
750 	((uintptr_t)(pp) >> mml_shift)) & (mml_table_sz - 1)])
751 #else	/* !DEBUG */
752 #define	MLIST_HASH(pp)   &mml_table[ \
753 	((uintptr_t)(pp) + ((uintptr_t)(pp) >> mml_shift)) & (mml_table_sz - 1)]
754 #endif	/* !DEBUG */
755 
756 kmutex_t		*mml_table;
757 uint_t			mml_table_sz;	/* must be a power of 2 */
758 uint_t			mml_shift;	/* log2(mml_table_sz) + 3 for align */
759 
760 kpm_hlk_t	*kpmp_table;
761 uint_t		kpmp_table_sz;	/* must be a power of 2 */
762 uchar_t		kpmp_shift;
763 
764 kpm_shlk_t	*kpmp_stable;
765 uint_t		kpmp_stable_sz;	/* must be a power of 2 */
766 
767 /*
768  * SPL_HASH was improved to avoid false cache line sharing
769  */
770 #define	SPL_TABLE_SIZE	128
771 #define	SPL_MASK	(SPL_TABLE_SIZE - 1)
772 #define	SPL_SHIFT	7		/* log2(SPL_TABLE_SIZE) */
773 
774 #define	SPL_INDEX(pp) \
775 	((((uintptr_t)(pp) >> SPL_SHIFT) ^ \
776 	((uintptr_t)(pp) >> (SPL_SHIFT << 1))) & \
777 	(SPL_TABLE_SIZE - 1))
778 
779 #define	SPL_HASH(pp)    \
780 	(&sfmmu_page_lock[SPL_INDEX(pp) & SPL_MASK].pad_mutex)
781 
782 static	pad_mutex_t	sfmmu_page_lock[SPL_TABLE_SIZE];
783 
784 
785 /*
786  * hat_unload_callback() will group together callbacks in order
787  * to avoid xt_sync() calls.  This is the maximum size of the group.
788  */
789 #define	MAX_CB_ADDR	32
790 
791 tte_t	hw_tte;
792 static ulong_t sfmmu_dmr_maxbit = DMR_MAXBIT;
793 
794 static char	*mmu_ctx_kstat_names[] = {
795 	"mmu_ctx_tsb_exceptions",
796 	"mmu_ctx_tsb_raise_exception",
797 	"mmu_ctx_wrap_around",
798 };
799 
800 /*
801  * Wrapper for vmem_xalloc since vmem_create only allows limited
802  * parameters for vm_source_alloc functions.  This function allows us
803  * to specify alignment consistent with the size of the object being
804  * allocated.
805  */
806 static void *
807 sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag)
808 {
809 	return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag));
810 }
811 
812 /* Common code for setting tsb_alloc_hiwater. */
813 #define	SFMMU_SET_TSB_ALLOC_HIWATER(pages)	tsb_alloc_hiwater = \
814 		ptob(pages) / tsb_alloc_hiwater_factor
815 
816 /*
817  * Set tsb_max_growsize to allow at most all of physical memory to be mapped by
818  * a single TSB.  physmem is the number of physical pages so we need physmem 8K
819  * TTEs to represent all those physical pages.  We round this up by using
820  * 1<<highbit().  To figure out which size code to use, remember that the size
821  * code is just an amount to shift the smallest TSB size to get the size of
822  * this TSB.  So we subtract that size, TSB_START_SIZE, from highbit() (or
823  * highbit() - 1) to get the size code for the smallest TSB that can represent
824  * all of physical memory, while erring on the side of too much.
825  *
826  * Restrict tsb_max_growsize to make sure that:
827  *	1) TSBs can't grow larger than the TSB slab size
828  *	2) TSBs can't grow larger than UTSB_MAX_SZCODE.
829  */
830 #define	SFMMU_SET_TSB_MAX_GROWSIZE(pages) {				\
831 	int	_i, _szc, _slabszc, _tsbszc;				\
832 									\
833 	_i = highbit(pages);						\
834 	if ((1 << (_i - 1)) == (pages))					\
835 		_i--;		/* 2^n case, round down */              \
836 	_szc = _i - TSB_START_SIZE;					\
837 	_slabszc = bigtsb_slab_shift - (TSB_START_SIZE + TSB_ENTRY_SHIFT); \
838 	_tsbszc = MIN(_szc, _slabszc);                                  \
839 	tsb_max_growsize = MIN(_tsbszc, UTSB_MAX_SZCODE);               \
840 }
841 
842 /*
843  * Given a pointer to an sfmmu and a TTE size code, return a pointer to the
844  * tsb_info which handles that TTE size.
845  */
846 #define	SFMMU_GET_TSBINFO(tsbinfop, sfmmup, tte_szc) {			\
847 	(tsbinfop) = (sfmmup)->sfmmu_tsb;				\
848 	ASSERT(((tsbinfop)->tsb_flags & TSB_SHAREDCTX) ||		\
849 	    sfmmu_hat_lock_held(sfmmup));				\
850 	if ((tte_szc) >= TTE4M)	{					\
851 		ASSERT((tsbinfop) != NULL);				\
852 		(tsbinfop) = (tsbinfop)->tsb_next;			\
853 	}								\
854 }
855 
856 /*
857  * Macro to use to unload entries from the TSB.
858  * It has knowledge of which page sizes get replicated in the TSB
859  * and will call the appropriate unload routine for the appropriate size.
860  */
861 #define	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, ismhat)		\
862 {									\
863 	int ttesz = get_hblk_ttesz(hmeblkp);				\
864 	if (ttesz == TTE8K || ttesz == TTE4M) {				\
865 		sfmmu_unload_tsb(sfmmup, addr, ttesz);			\
866 	} else {							\
867 		caddr_t sva = ismhat ? addr : 				\
868 		    (caddr_t)get_hblk_base(hmeblkp);			\
869 		caddr_t eva = sva + get_hblk_span(hmeblkp);		\
870 		ASSERT(addr >= sva && addr < eva);			\
871 		sfmmu_unload_tsb_range(sfmmup, sva, eva, ttesz);	\
872 	}								\
873 }
874 
875 
876 /* Update tsb_alloc_hiwater after memory is configured. */
877 /*ARGSUSED*/
878 static void
879 sfmmu_update_post_add(void *arg, pgcnt_t delta_pages)
880 {
881 	/* Assumes physmem has already been updated. */
882 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
883 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
884 }
885 
886 /*
887  * Update tsb_alloc_hiwater before memory is deleted.  We'll do nothing here
888  * and update tsb_alloc_hiwater and tsb_max_growsize after the memory is
889  * deleted.
890  */
891 /*ARGSUSED*/
892 static int
893 sfmmu_update_pre_del(void *arg, pgcnt_t delta_pages)
894 {
895 	return (0);
896 }
897 
898 /* Update tsb_alloc_hiwater after memory fails to be unconfigured. */
899 /*ARGSUSED*/
900 static void
901 sfmmu_update_post_del(void *arg, pgcnt_t delta_pages, int cancelled)
902 {
903 	/*
904 	 * Whether the delete was cancelled or not, just go ahead and update
905 	 * tsb_alloc_hiwater and tsb_max_growsize.
906 	 */
907 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
908 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
909 }
910 
911 static kphysm_setup_vector_t sfmmu_update_vec = {
912 	KPHYSM_SETUP_VECTOR_VERSION,	/* version */
913 	sfmmu_update_post_add,		/* post_add */
914 	sfmmu_update_pre_del,		/* pre_del */
915 	sfmmu_update_post_del		/* post_del */
916 };
917 
918 
919 /*
920  * HME_BLK HASH PRIMITIVES
921  */
922 
923 /*
924  * Enter a hme on the mapping list for page pp.
925  * When large pages are more prevalent in the system we might want to
926  * keep the mapping list in ascending order by the hment size. For now,
927  * small pages are more frequent, so don't slow it down.
928  */
929 #define	HME_ADD(hme, pp)					\
930 {								\
931 	ASSERT(sfmmu_mlist_held(pp));				\
932 								\
933 	hme->hme_prev = NULL;					\
934 	hme->hme_next = pp->p_mapping;				\
935 	hme->hme_page = pp;					\
936 	if (pp->p_mapping) {					\
937 		((struct sf_hment *)(pp->p_mapping))->hme_prev = hme;\
938 		ASSERT(pp->p_share > 0);			\
939 	} else  {						\
940 		/* EMPTY */					\
941 		ASSERT(pp->p_share == 0);			\
942 	}							\
943 	pp->p_mapping = hme;					\
944 	pp->p_share++;						\
945 }
946 
947 /*
948  * Enter a hme on the mapping list for page pp.
949  * If we are unmapping a large translation, we need to make sure that the
950  * change is reflect in the corresponding bit of the p_index field.
951  */
952 #define	HME_SUB(hme, pp)					\
953 {								\
954 	ASSERT(sfmmu_mlist_held(pp));				\
955 	ASSERT(hme->hme_page == pp || IS_PAHME(hme));		\
956 								\
957 	if (pp->p_mapping == NULL) {				\
958 		panic("hme_remove - no mappings");		\
959 	}							\
960 								\
961 	membar_stst();	/* ensure previous stores finish */	\
962 								\
963 	ASSERT(pp->p_share > 0);				\
964 	pp->p_share--;						\
965 								\
966 	if (hme->hme_prev) {					\
967 		ASSERT(pp->p_mapping != hme);			\
968 		ASSERT(hme->hme_prev->hme_page == pp ||		\
969 			IS_PAHME(hme->hme_prev));		\
970 		hme->hme_prev->hme_next = hme->hme_next;	\
971 	} else {						\
972 		ASSERT(pp->p_mapping == hme);			\
973 		pp->p_mapping = hme->hme_next;			\
974 		ASSERT((pp->p_mapping == NULL) ?		\
975 			(pp->p_share == 0) : 1);		\
976 	}							\
977 								\
978 	if (hme->hme_next) {					\
979 		ASSERT(hme->hme_next->hme_page == pp ||		\
980 			IS_PAHME(hme->hme_next));		\
981 		hme->hme_next->hme_prev = hme->hme_prev;	\
982 	}							\
983 								\
984 	/* zero out the entry */				\
985 	hme->hme_next = NULL;					\
986 	hme->hme_prev = NULL;					\
987 	hme->hme_page = NULL;					\
988 								\
989 	if (hme_size(hme) > TTE8K) {				\
990 		/* remove mappings for remainder of large pg */	\
991 		sfmmu_rm_large_mappings(pp, hme_size(hme));	\
992 	}							\
993 }
994 
995 /*
996  * This function returns the hment given the hme_blk and a vaddr.
997  * It assumes addr has already been checked to belong to hme_blk's
998  * range.
999  */
1000 #define	HBLKTOHME(hment, hmeblkp, addr)					\
1001 {									\
1002 	int index;							\
1003 	HBLKTOHME_IDX(hment, hmeblkp, addr, index)			\
1004 }
1005 
1006 /*
1007  * Version of HBLKTOHME that also returns the index in hmeblkp
1008  * of the hment.
1009  */
1010 #define	HBLKTOHME_IDX(hment, hmeblkp, addr, idx)			\
1011 {									\
1012 	ASSERT(in_hblk_range((hmeblkp), (addr)));			\
1013 									\
1014 	if (get_hblk_ttesz(hmeblkp) == TTE8K) {				\
1015 		idx = (((uintptr_t)(addr) >> MMU_PAGESHIFT) & (NHMENTS-1)); \
1016 	} else								\
1017 		idx = 0;						\
1018 									\
1019 	(hment) = &(hmeblkp)->hblk_hme[idx];				\
1020 }
1021 
1022 /*
1023  * Disable any page sizes not supported by the CPU
1024  */
1025 void
1026 hat_init_pagesizes()
1027 {
1028 	int 		i;
1029 
1030 	mmu_exported_page_sizes = 0;
1031 	for (i = TTE8K; i < max_mmu_page_sizes; i++) {
1032 
1033 		szc_2_userszc[i] = (uint_t)-1;
1034 		userszc_2_szc[i] = (uint_t)-1;
1035 
1036 		if ((mmu_exported_pagesize_mask & (1 << i)) == 0) {
1037 			disable_large_pages |= (1 << i);
1038 		} else {
1039 			szc_2_userszc[i] = mmu_exported_page_sizes;
1040 			userszc_2_szc[mmu_exported_page_sizes] = i;
1041 			mmu_exported_page_sizes++;
1042 		}
1043 	}
1044 
1045 	disable_ism_large_pages |= disable_large_pages;
1046 	disable_auto_data_large_pages = disable_large_pages;
1047 	disable_auto_text_large_pages = disable_large_pages;
1048 
1049 	/*
1050 	 * Initialize mmu-specific large page sizes.
1051 	 */
1052 	if (&mmu_large_pages_disabled) {
1053 		disable_large_pages |= mmu_large_pages_disabled(HAT_LOAD);
1054 		disable_ism_large_pages |=
1055 		    mmu_large_pages_disabled(HAT_LOAD_SHARE);
1056 		disable_auto_data_large_pages |=
1057 		    mmu_large_pages_disabled(HAT_AUTO_DATA);
1058 		disable_auto_text_large_pages |=
1059 		    mmu_large_pages_disabled(HAT_AUTO_TEXT);
1060 	}
1061 }
1062 
1063 /*
1064  * Initialize the hardware address translation structures.
1065  */
1066 void
1067 hat_init(void)
1068 {
1069 	int 		i;
1070 	uint_t		sz;
1071 	size_t		size;
1072 
1073 	hat_lock_init();
1074 	hat_kstat_init();
1075 
1076 	/*
1077 	 * Hardware-only bits in a TTE
1078 	 */
1079 	MAKE_TTE_MASK(&hw_tte);
1080 
1081 	hat_init_pagesizes();
1082 
1083 	/* Initialize the hash locks */
1084 	for (i = 0; i < khmehash_num; i++) {
1085 		mutex_init(&khme_hash[i].hmehash_mutex, NULL,
1086 		    MUTEX_DEFAULT, NULL);
1087 	}
1088 	for (i = 0; i < uhmehash_num; i++) {
1089 		mutex_init(&uhme_hash[i].hmehash_mutex, NULL,
1090 		    MUTEX_DEFAULT, NULL);
1091 	}
1092 	khmehash_num--;		/* make sure counter starts from 0 */
1093 	uhmehash_num--;		/* make sure counter starts from 0 */
1094 
1095 	/*
1096 	 * Allocate context domain structures.
1097 	 *
1098 	 * A platform may choose to modify max_mmu_ctxdoms in
1099 	 * set_platform_defaults(). If a platform does not define
1100 	 * a set_platform_defaults() or does not choose to modify
1101 	 * max_mmu_ctxdoms, it gets one MMU context domain for every CPU.
1102 	 *
1103 	 * For sun4v, there will be one global context domain, this is to
1104 	 * avoid the ldom cpu substitution problem.
1105 	 *
1106 	 * For all platforms that have CPUs sharing MMUs, this
1107 	 * value must be defined.
1108 	 */
1109 	if (max_mmu_ctxdoms == 0) {
1110 #ifndef sun4v
1111 		max_mmu_ctxdoms = max_ncpus;
1112 #else /* sun4v */
1113 		max_mmu_ctxdoms = 1;
1114 #endif /* sun4v */
1115 	}
1116 
1117 	size = max_mmu_ctxdoms * sizeof (mmu_ctx_t *);
1118 	mmu_ctxs_tbl = kmem_zalloc(size, KM_SLEEP);
1119 
1120 	/* mmu_ctx_t is 64 bytes aligned */
1121 	mmuctxdom_cache = kmem_cache_create("mmuctxdom_cache",
1122 	    sizeof (mmu_ctx_t), 64, NULL, NULL, NULL, NULL, NULL, 0);
1123 	/*
1124 	 * MMU context domain initialization for the Boot CPU.
1125 	 * This needs the context domains array allocated above.
1126 	 */
1127 	mutex_enter(&cpu_lock);
1128 	sfmmu_cpu_init(CPU);
1129 	mutex_exit(&cpu_lock);
1130 
1131 	/*
1132 	 * Intialize ism mapping list lock.
1133 	 */
1134 
1135 	mutex_init(&ism_mlist_lock, NULL, MUTEX_DEFAULT, NULL);
1136 
1137 	/*
1138 	 * Each sfmmu structure carries an array of MMU context info
1139 	 * structures, one per context domain. The size of this array depends
1140 	 * on the maximum number of context domains. So, the size of the
1141 	 * sfmmu structure varies per platform.
1142 	 *
1143 	 * sfmmu is allocated from static arena, because trap
1144 	 * handler at TL > 0 is not allowed to touch kernel relocatable
1145 	 * memory. sfmmu's alignment is changed to 64 bytes from
1146 	 * default 8 bytes, as the lower 6 bits will be used to pass
1147 	 * pgcnt to vtag_flush_pgcnt_tl1.
1148 	 */
1149 	size = sizeof (sfmmu_t) + sizeof (sfmmu_ctx_t) * (max_mmu_ctxdoms - 1);
1150 
1151 	sfmmuid_cache = kmem_cache_create("sfmmuid_cache", size,
1152 	    64, sfmmu_idcache_constructor, sfmmu_idcache_destructor,
1153 	    NULL, NULL, static_arena, 0);
1154 
1155 	sfmmu_tsbinfo_cache = kmem_cache_create("sfmmu_tsbinfo_cache",
1156 	    sizeof (struct tsb_info), 0, NULL, NULL, NULL, NULL, NULL, 0);
1157 
1158 	/*
1159 	 * Since we only use the tsb8k cache to "borrow" pages for TSBs
1160 	 * from the heap when low on memory or when TSB_FORCEALLOC is
1161 	 * specified, don't use magazines to cache them--we want to return
1162 	 * them to the system as quickly as possible.
1163 	 */
1164 	sfmmu_tsb8k_cache = kmem_cache_create("sfmmu_tsb8k_cache",
1165 	    MMU_PAGESIZE, MMU_PAGESIZE, NULL, NULL, NULL, NULL,
1166 	    static_arena, KMC_NOMAGAZINE);
1167 
1168 	/*
1169 	 * Set tsb_alloc_hiwater to 1/tsb_alloc_hiwater_factor of physical
1170 	 * memory, which corresponds to the old static reserve for TSBs.
1171 	 * tsb_alloc_hiwater_factor defaults to 32.  This caps the amount of
1172 	 * memory we'll allocate for TSB slabs; beyond this point TSB
1173 	 * allocations will be taken from the kernel heap (via
1174 	 * sfmmu_tsb8k_cache) and will be throttled as would any other kmem
1175 	 * consumer.
1176 	 */
1177 	if (tsb_alloc_hiwater_factor == 0) {
1178 		tsb_alloc_hiwater_factor = TSB_ALLOC_HIWATER_FACTOR_DEFAULT;
1179 	}
1180 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
1181 
1182 	for (sz = tsb_slab_ttesz; sz > 0; sz--) {
1183 		if (!(disable_large_pages & (1 << sz)))
1184 			break;
1185 	}
1186 
1187 	if (sz < tsb_slab_ttesz) {
1188 		tsb_slab_ttesz = sz;
1189 		tsb_slab_shift = MMU_PAGESHIFT + (sz << 1) + sz;
1190 		tsb_slab_size = 1 << tsb_slab_shift;
1191 		tsb_slab_mask = (1 << (tsb_slab_shift - MMU_PAGESHIFT)) - 1;
1192 		use_bigtsb_arena = 0;
1193 	} else if (use_bigtsb_arena &&
1194 	    (disable_large_pages & (1 << bigtsb_slab_ttesz))) {
1195 		use_bigtsb_arena = 0;
1196 	}
1197 
1198 	if (!use_bigtsb_arena) {
1199 		bigtsb_slab_shift = tsb_slab_shift;
1200 	}
1201 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
1202 
1203 	/*
1204 	 * On smaller memory systems, allocate TSB memory in smaller chunks
1205 	 * than the default 4M slab size. We also honor disable_large_pages
1206 	 * here.
1207 	 *
1208 	 * The trap handlers need to be patched with the final slab shift,
1209 	 * since they need to be able to construct the TSB pointer at runtime.
1210 	 */
1211 	if ((tsb_max_growsize <= TSB_512K_SZCODE) &&
1212 	    !(disable_large_pages & (1 << TTE512K))) {
1213 		tsb_slab_ttesz = TTE512K;
1214 		tsb_slab_shift = MMU_PAGESHIFT512K;
1215 		tsb_slab_size = MMU_PAGESIZE512K;
1216 		tsb_slab_mask = MMU_PAGEOFFSET512K >> MMU_PAGESHIFT;
1217 		use_bigtsb_arena = 0;
1218 	}
1219 
1220 	if (!use_bigtsb_arena) {
1221 		bigtsb_slab_ttesz = tsb_slab_ttesz;
1222 		bigtsb_slab_shift = tsb_slab_shift;
1223 		bigtsb_slab_size = tsb_slab_size;
1224 		bigtsb_slab_mask = tsb_slab_mask;
1225 	}
1226 
1227 
1228 	/*
1229 	 * Set up memory callback to update tsb_alloc_hiwater and
1230 	 * tsb_max_growsize.
1231 	 */
1232 	i = kphysm_setup_func_register(&sfmmu_update_vec, (void *) 0);
1233 	ASSERT(i == 0);
1234 
1235 	/*
1236 	 * kmem_tsb_arena is the source from which large TSB slabs are
1237 	 * drawn.  The quantum of this arena corresponds to the largest
1238 	 * TSB size we can dynamically allocate for user processes.
1239 	 * Currently it must also be a supported page size since we
1240 	 * use exactly one translation entry to map each slab page.
1241 	 *
1242 	 * The per-lgroup kmem_tsb_default_arena arenas are the arenas from
1243 	 * which most TSBs are allocated.  Since most TSB allocations are
1244 	 * typically 8K we have a kmem cache we stack on top of each
1245 	 * kmem_tsb_default_arena to speed up those allocations.
1246 	 *
1247 	 * Note the two-level scheme of arenas is required only
1248 	 * because vmem_create doesn't allow us to specify alignment
1249 	 * requirements.  If this ever changes the code could be
1250 	 * simplified to use only one level of arenas.
1251 	 *
1252 	 * If 256M page support exists on sun4v, 256MB kmem_bigtsb_arena
1253 	 * will be provided in addition to the 4M kmem_tsb_arena.
1254 	 */
1255 	if (use_bigtsb_arena) {
1256 		kmem_bigtsb_arena = vmem_create("kmem_bigtsb", NULL, 0,
1257 		    bigtsb_slab_size, sfmmu_vmem_xalloc_aligned_wrapper,
1258 		    vmem_xfree, heap_arena, 0, VM_SLEEP);
1259 	}
1260 
1261 	kmem_tsb_arena = vmem_create("kmem_tsb", NULL, 0, tsb_slab_size,
1262 	    sfmmu_vmem_xalloc_aligned_wrapper,
1263 	    vmem_xfree, heap_arena, 0, VM_SLEEP);
1264 
1265 	if (tsb_lgrp_affinity) {
1266 		char s[50];
1267 		for (i = 0; i < NLGRPS_MAX; i++) {
1268 			if (use_bigtsb_arena) {
1269 				(void) sprintf(s, "kmem_bigtsb_lgrp%d", i);
1270 				kmem_bigtsb_default_arena[i] = vmem_create(s,
1271 				    NULL, 0, 2 * tsb_slab_size,
1272 				    sfmmu_tsb_segkmem_alloc,
1273 				    sfmmu_tsb_segkmem_free, kmem_bigtsb_arena,
1274 				    0, VM_SLEEP | VM_BESTFIT);
1275 			}
1276 
1277 			(void) sprintf(s, "kmem_tsb_lgrp%d", i);
1278 			kmem_tsb_default_arena[i] = vmem_create(s,
1279 			    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1280 			    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1281 			    VM_SLEEP | VM_BESTFIT);
1282 
1283 			(void) sprintf(s, "sfmmu_tsb_lgrp%d_cache", i);
1284 			sfmmu_tsb_cache[i] = kmem_cache_create(s,
1285 			    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1286 			    kmem_tsb_default_arena[i], 0);
1287 		}
1288 	} else {
1289 		if (use_bigtsb_arena) {
1290 			kmem_bigtsb_default_arena[0] =
1291 			    vmem_create("kmem_bigtsb_default", NULL, 0,
1292 			    2 * tsb_slab_size, sfmmu_tsb_segkmem_alloc,
1293 			    sfmmu_tsb_segkmem_free, kmem_bigtsb_arena, 0,
1294 			    VM_SLEEP | VM_BESTFIT);
1295 		}
1296 
1297 		kmem_tsb_default_arena[0] = vmem_create("kmem_tsb_default",
1298 		    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1299 		    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1300 		    VM_SLEEP | VM_BESTFIT);
1301 		sfmmu_tsb_cache[0] = kmem_cache_create("sfmmu_tsb_cache",
1302 		    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1303 		    kmem_tsb_default_arena[0], 0);
1304 	}
1305 
1306 	sfmmu8_cache = kmem_cache_create("sfmmu8_cache", HME8BLK_SZ,
1307 	    HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1308 	    sfmmu_hblkcache_destructor,
1309 	    sfmmu_hblkcache_reclaim, (void *)HME8BLK_SZ,
1310 	    hat_memload_arena, KMC_NOHASH);
1311 
1312 	hat_memload1_arena = vmem_create("hat_memload1", NULL, 0, PAGESIZE,
1313 	    segkmem_alloc_permanent, segkmem_free, heap_arena, 0, VM_SLEEP);
1314 
1315 	sfmmu1_cache = kmem_cache_create("sfmmu1_cache", HME1BLK_SZ,
1316 	    HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1317 	    sfmmu_hblkcache_destructor,
1318 	    NULL, (void *)HME1BLK_SZ,
1319 	    hat_memload1_arena, KMC_NOHASH);
1320 
1321 	pa_hment_cache = kmem_cache_create("pa_hment_cache", PAHME_SZ,
1322 	    0, NULL, NULL, NULL, NULL, static_arena, KMC_NOHASH);
1323 
1324 	ism_blk_cache = kmem_cache_create("ism_blk_cache",
1325 	    sizeof (ism_blk_t), ecache_alignsize, NULL, NULL,
1326 	    NULL, NULL, static_arena, KMC_NOHASH);
1327 
1328 	ism_ment_cache = kmem_cache_create("ism_ment_cache",
1329 	    sizeof (ism_ment_t), 0, NULL, NULL,
1330 	    NULL, NULL, NULL, 0);
1331 
1332 	/*
1333 	 * We grab the first hat for the kernel,
1334 	 */
1335 	AS_LOCK_ENTER(&kas, &kas.a_lock, RW_WRITER);
1336 	kas.a_hat = hat_alloc(&kas);
1337 	AS_LOCK_EXIT(&kas, &kas.a_lock);
1338 
1339 	/*
1340 	 * Initialize hblk_reserve.
1341 	 */
1342 	((struct hme_blk *)hblk_reserve)->hblk_nextpa =
1343 	    va_to_pa((caddr_t)hblk_reserve);
1344 
1345 #ifndef UTSB_PHYS
1346 	/*
1347 	 * Reserve some kernel virtual address space for the locked TTEs
1348 	 * that allow us to probe the TSB from TL>0.
1349 	 */
1350 	utsb_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1351 	    0, 0, NULL, NULL, VM_SLEEP);
1352 	utsb4m_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1353 	    0, 0, NULL, NULL, VM_SLEEP);
1354 #endif
1355 
1356 #ifdef VAC
1357 	/*
1358 	 * The big page VAC handling code assumes VAC
1359 	 * will not be bigger than the smallest big
1360 	 * page- which is 64K.
1361 	 */
1362 	if (TTEPAGES(TTE64K) < CACHE_NUM_COLOR) {
1363 		cmn_err(CE_PANIC, "VAC too big!");
1364 	}
1365 #endif
1366 
1367 	(void) xhat_init();
1368 
1369 	uhme_hash_pa = va_to_pa(uhme_hash);
1370 	khme_hash_pa = va_to_pa(khme_hash);
1371 
1372 	/*
1373 	 * Initialize relocation locks. kpr_suspendlock is held
1374 	 * at PIL_MAX to prevent interrupts from pinning the holder
1375 	 * of a suspended TTE which may access it leading to a
1376 	 * deadlock condition.
1377 	 */
1378 	mutex_init(&kpr_mutex, NULL, MUTEX_DEFAULT, NULL);
1379 	mutex_init(&kpr_suspendlock, NULL, MUTEX_SPIN, (void *)PIL_MAX);
1380 
1381 	/*
1382 	 * If Shared context support is disabled via /etc/system
1383 	 * set shctx_on to 0 here if it was set to 1 earlier in boot
1384 	 * sequence by cpu module initialization code.
1385 	 */
1386 	if (shctx_on && disable_shctx) {
1387 		shctx_on = 0;
1388 	}
1389 
1390 	if (shctx_on) {
1391 		srd_buckets = kmem_zalloc(SFMMU_MAX_SRD_BUCKETS *
1392 		    sizeof (srd_buckets[0]), KM_SLEEP);
1393 		for (i = 0; i < SFMMU_MAX_SRD_BUCKETS; i++) {
1394 			mutex_init(&srd_buckets[i].srdb_lock, NULL,
1395 			    MUTEX_DEFAULT, NULL);
1396 		}
1397 
1398 		srd_cache = kmem_cache_create("srd_cache", sizeof (sf_srd_t),
1399 		    0, sfmmu_srdcache_constructor, sfmmu_srdcache_destructor,
1400 		    NULL, NULL, NULL, 0);
1401 		region_cache = kmem_cache_create("region_cache",
1402 		    sizeof (sf_region_t), 0, sfmmu_rgncache_constructor,
1403 		    sfmmu_rgncache_destructor, NULL, NULL, NULL, 0);
1404 		scd_cache = kmem_cache_create("scd_cache", sizeof (sf_scd_t),
1405 		    0, sfmmu_scdcache_constructor,  sfmmu_scdcache_destructor,
1406 		    NULL, NULL, NULL, 0);
1407 	}
1408 
1409 	/*
1410 	 * Pre-allocate hrm_hashtab before enabling the collection of
1411 	 * refmod statistics.  Allocating on the fly would mean us
1412 	 * running the risk of suffering recursive mutex enters or
1413 	 * deadlocks.
1414 	 */
1415 	hrm_hashtab = kmem_zalloc(HRM_HASHSIZE * sizeof (struct hrmstat *),
1416 	    KM_SLEEP);
1417 }
1418 
1419 /*
1420  * Initialize locking for the hat layer, called early during boot.
1421  */
1422 static void
1423 hat_lock_init()
1424 {
1425 	int i;
1426 
1427 	/*
1428 	 * initialize the array of mutexes protecting a page's mapping
1429 	 * list and p_nrm field.
1430 	 */
1431 	for (i = 0; i < mml_table_sz; i++)
1432 		mutex_init(&mml_table[i], NULL, MUTEX_DEFAULT, NULL);
1433 
1434 	if (kpm_enable) {
1435 		for (i = 0; i < kpmp_table_sz; i++) {
1436 			mutex_init(&kpmp_table[i].khl_mutex, NULL,
1437 			    MUTEX_DEFAULT, NULL);
1438 		}
1439 	}
1440 
1441 	/*
1442 	 * Initialize array of mutex locks that protects sfmmu fields and
1443 	 * TSB lists.
1444 	 */
1445 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
1446 		mutex_init(HATLOCK_MUTEXP(&hat_lock[i]), NULL, MUTEX_DEFAULT,
1447 		    NULL);
1448 }
1449 
1450 #define	SFMMU_KERNEL_MAXVA \
1451 	(kmem64_base ? (uintptr_t)kmem64_end : (SYSLIMIT))
1452 
1453 /*
1454  * Allocate a hat structure.
1455  * Called when an address space first uses a hat.
1456  */
1457 struct hat *
1458 hat_alloc(struct as *as)
1459 {
1460 	sfmmu_t *sfmmup;
1461 	int i;
1462 	uint64_t cnum;
1463 	extern uint_t get_color_start(struct as *);
1464 
1465 	ASSERT(AS_WRITE_HELD(as, &as->a_lock));
1466 	sfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
1467 	sfmmup->sfmmu_as = as;
1468 	sfmmup->sfmmu_flags = 0;
1469 	sfmmup->sfmmu_tteflags = 0;
1470 	sfmmup->sfmmu_rtteflags = 0;
1471 	LOCK_INIT_CLEAR(&sfmmup->sfmmu_ctx_lock);
1472 
1473 	if (as == &kas) {
1474 		ksfmmup = sfmmup;
1475 		sfmmup->sfmmu_cext = 0;
1476 		cnum = KCONTEXT;
1477 
1478 		sfmmup->sfmmu_clrstart = 0;
1479 		sfmmup->sfmmu_tsb = NULL;
1480 		/*
1481 		 * hat_kern_setup() will call sfmmu_init_ktsbinfo()
1482 		 * to setup tsb_info for ksfmmup.
1483 		 */
1484 	} else {
1485 
1486 		/*
1487 		 * Just set to invalid ctx. When it faults, it will
1488 		 * get a valid ctx. This would avoid the situation
1489 		 * where we get a ctx, but it gets stolen and then
1490 		 * we fault when we try to run and so have to get
1491 		 * another ctx.
1492 		 */
1493 		sfmmup->sfmmu_cext = 0;
1494 		cnum = INVALID_CONTEXT;
1495 
1496 		/* initialize original physical page coloring bin */
1497 		sfmmup->sfmmu_clrstart = get_color_start(as);
1498 #ifdef DEBUG
1499 		if (tsb_random_size) {
1500 			uint32_t randval = (uint32_t)gettick() >> 4;
1501 			int size = randval % (tsb_max_growsize + 1);
1502 
1503 			/* chose a random tsb size for stress testing */
1504 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb, size,
1505 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1506 		} else
1507 #endif /* DEBUG */
1508 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb,
1509 			    default_tsb_size,
1510 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1511 		sfmmup->sfmmu_flags = HAT_SWAPPED | HAT_ALLCTX_INVALID;
1512 		ASSERT(sfmmup->sfmmu_tsb != NULL);
1513 	}
1514 
1515 	ASSERT(max_mmu_ctxdoms > 0);
1516 	for (i = 0; i < max_mmu_ctxdoms; i++) {
1517 		sfmmup->sfmmu_ctxs[i].cnum = cnum;
1518 		sfmmup->sfmmu_ctxs[i].gnum = 0;
1519 	}
1520 
1521 	for (i = 0; i < max_mmu_page_sizes; i++) {
1522 		sfmmup->sfmmu_ttecnt[i] = 0;
1523 		sfmmup->sfmmu_scdrttecnt[i] = 0;
1524 		sfmmup->sfmmu_ismttecnt[i] = 0;
1525 		sfmmup->sfmmu_scdismttecnt[i] = 0;
1526 		sfmmup->sfmmu_pgsz[i] = TTE8K;
1527 	}
1528 	sfmmup->sfmmu_tsb0_4minflcnt = 0;
1529 	sfmmup->sfmmu_iblk = NULL;
1530 	sfmmup->sfmmu_ismhat = 0;
1531 	sfmmup->sfmmu_scdhat = 0;
1532 	sfmmup->sfmmu_ismblkpa = (uint64_t)-1;
1533 	if (sfmmup == ksfmmup) {
1534 		CPUSET_ALL(sfmmup->sfmmu_cpusran);
1535 	} else {
1536 		CPUSET_ZERO(sfmmup->sfmmu_cpusran);
1537 	}
1538 	sfmmup->sfmmu_free = 0;
1539 	sfmmup->sfmmu_rmstat = 0;
1540 	sfmmup->sfmmu_clrbin = sfmmup->sfmmu_clrstart;
1541 	sfmmup->sfmmu_xhat_provider = NULL;
1542 	cv_init(&sfmmup->sfmmu_tsb_cv, NULL, CV_DEFAULT, NULL);
1543 	sfmmup->sfmmu_srdp = NULL;
1544 	SF_RGNMAP_ZERO(sfmmup->sfmmu_region_map);
1545 	bzero(sfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
1546 	sfmmup->sfmmu_scdp = NULL;
1547 	sfmmup->sfmmu_scd_link.next = NULL;
1548 	sfmmup->sfmmu_scd_link.prev = NULL;
1549 	return (sfmmup);
1550 }
1551 
1552 /*
1553  * Create per-MMU context domain kstats for a given MMU ctx.
1554  */
1555 static void
1556 sfmmu_mmu_kstat_create(mmu_ctx_t *mmu_ctxp)
1557 {
1558 	mmu_ctx_stat_t	stat;
1559 	kstat_t		*mmu_kstat;
1560 
1561 	ASSERT(MUTEX_HELD(&cpu_lock));
1562 	ASSERT(mmu_ctxp->mmu_kstat == NULL);
1563 
1564 	mmu_kstat = kstat_create("unix", mmu_ctxp->mmu_idx, "mmu_ctx",
1565 	    "hat", KSTAT_TYPE_NAMED, MMU_CTX_NUM_STATS, KSTAT_FLAG_VIRTUAL);
1566 
1567 	if (mmu_kstat == NULL) {
1568 		cmn_err(CE_WARN, "kstat_create for MMU %d failed",
1569 		    mmu_ctxp->mmu_idx);
1570 	} else {
1571 		mmu_kstat->ks_data = mmu_ctxp->mmu_kstat_data;
1572 		for (stat = 0; stat < MMU_CTX_NUM_STATS; stat++)
1573 			kstat_named_init(&mmu_ctxp->mmu_kstat_data[stat],
1574 			    mmu_ctx_kstat_names[stat], KSTAT_DATA_INT64);
1575 		mmu_ctxp->mmu_kstat = mmu_kstat;
1576 		kstat_install(mmu_kstat);
1577 	}
1578 }
1579 
1580 /*
1581  * plat_cpuid_to_mmu_ctx_info() is a platform interface that returns MMU
1582  * context domain information for a given CPU. If a platform does not
1583  * specify that interface, then the function below is used instead to return
1584  * default information. The defaults are as follows:
1585  *
1586  *	- For sun4u systems there's one MMU context domain per CPU.
1587  *	  This default is used by all sun4u systems except OPL. OPL systems
1588  *	  provide platform specific interface to map CPU ids to MMU ids
1589  *	  because on OPL more than 1 CPU shares a single MMU.
1590  *        Note that on sun4v, there is one global context domain for
1591  *	  the entire system. This is to avoid running into potential problem
1592  *	  with ldom physical cpu substitution feature.
1593  *	- The number of MMU context IDs supported on any CPU in the
1594  *	  system is 8K.
1595  */
1596 /*ARGSUSED*/
1597 static void
1598 sfmmu_cpuid_to_mmu_ctx_info(processorid_t cpuid, mmu_ctx_info_t *infop)
1599 {
1600 	infop->mmu_nctxs = nctxs;
1601 #ifndef sun4v
1602 	infop->mmu_idx = cpu[cpuid]->cpu_seqid;
1603 #else /* sun4v */
1604 	infop->mmu_idx = 0;
1605 #endif /* sun4v */
1606 }
1607 
1608 /*
1609  * Called during CPU initialization to set the MMU context-related information
1610  * for a CPU.
1611  *
1612  * cpu_lock serializes accesses to mmu_ctxs and mmu_saved_gnum.
1613  */
1614 void
1615 sfmmu_cpu_init(cpu_t *cp)
1616 {
1617 	mmu_ctx_info_t	info;
1618 	mmu_ctx_t	*mmu_ctxp;
1619 
1620 	ASSERT(MUTEX_HELD(&cpu_lock));
1621 
1622 	if (&plat_cpuid_to_mmu_ctx_info == NULL)
1623 		sfmmu_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1624 	else
1625 		plat_cpuid_to_mmu_ctx_info(cp->cpu_id, &info);
1626 
1627 	ASSERT(info.mmu_idx < max_mmu_ctxdoms);
1628 
1629 	if ((mmu_ctxp = mmu_ctxs_tbl[info.mmu_idx]) == NULL) {
1630 		/* Each mmu_ctx is cacheline aligned. */
1631 		mmu_ctxp = kmem_cache_alloc(mmuctxdom_cache, KM_SLEEP);
1632 		bzero(mmu_ctxp, sizeof (mmu_ctx_t));
1633 
1634 		mutex_init(&mmu_ctxp->mmu_lock, NULL, MUTEX_SPIN,
1635 		    (void *)ipltospl(DISP_LEVEL));
1636 		mmu_ctxp->mmu_idx = info.mmu_idx;
1637 		mmu_ctxp->mmu_nctxs = info.mmu_nctxs;
1638 		/*
1639 		 * Globally for lifetime of a system,
1640 		 * gnum must always increase.
1641 		 * mmu_saved_gnum is protected by the cpu_lock.
1642 		 */
1643 		mmu_ctxp->mmu_gnum = mmu_saved_gnum + 1;
1644 		mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
1645 
1646 		sfmmu_mmu_kstat_create(mmu_ctxp);
1647 
1648 		mmu_ctxs_tbl[info.mmu_idx] = mmu_ctxp;
1649 	} else {
1650 		ASSERT(mmu_ctxp->mmu_idx == info.mmu_idx);
1651 	}
1652 
1653 	/*
1654 	 * The mmu_lock is acquired here to prevent races with
1655 	 * the wrap-around code.
1656 	 */
1657 	mutex_enter(&mmu_ctxp->mmu_lock);
1658 
1659 
1660 	mmu_ctxp->mmu_ncpus++;
1661 	CPUSET_ADD(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1662 	CPU_MMU_IDX(cp) = info.mmu_idx;
1663 	CPU_MMU_CTXP(cp) = mmu_ctxp;
1664 
1665 	mutex_exit(&mmu_ctxp->mmu_lock);
1666 }
1667 
1668 /*
1669  * Called to perform MMU context-related cleanup for a CPU.
1670  */
1671 void
1672 sfmmu_cpu_cleanup(cpu_t *cp)
1673 {
1674 	mmu_ctx_t	*mmu_ctxp;
1675 
1676 	ASSERT(MUTEX_HELD(&cpu_lock));
1677 
1678 	mmu_ctxp = CPU_MMU_CTXP(cp);
1679 	ASSERT(mmu_ctxp != NULL);
1680 
1681 	/*
1682 	 * The mmu_lock is acquired here to prevent races with
1683 	 * the wrap-around code.
1684 	 */
1685 	mutex_enter(&mmu_ctxp->mmu_lock);
1686 
1687 	CPU_MMU_CTXP(cp) = NULL;
1688 
1689 	CPUSET_DEL(mmu_ctxp->mmu_cpuset, cp->cpu_id);
1690 	if (--mmu_ctxp->mmu_ncpus == 0) {
1691 		mmu_ctxs_tbl[mmu_ctxp->mmu_idx] = NULL;
1692 		mutex_exit(&mmu_ctxp->mmu_lock);
1693 		mutex_destroy(&mmu_ctxp->mmu_lock);
1694 
1695 		if (mmu_ctxp->mmu_kstat)
1696 			kstat_delete(mmu_ctxp->mmu_kstat);
1697 
1698 		/* mmu_saved_gnum is protected by the cpu_lock. */
1699 		if (mmu_saved_gnum < mmu_ctxp->mmu_gnum)
1700 			mmu_saved_gnum = mmu_ctxp->mmu_gnum;
1701 
1702 		kmem_cache_free(mmuctxdom_cache, mmu_ctxp);
1703 
1704 		return;
1705 	}
1706 
1707 	mutex_exit(&mmu_ctxp->mmu_lock);
1708 }
1709 
1710 /*
1711  * Hat_setup, makes an address space context the current active one.
1712  * In sfmmu this translates to setting the secondary context with the
1713  * corresponding context.
1714  */
1715 void
1716 hat_setup(struct hat *sfmmup, int allocflag)
1717 {
1718 	hatlock_t *hatlockp;
1719 
1720 	/* Init needs some special treatment. */
1721 	if (allocflag == HAT_INIT) {
1722 		/*
1723 		 * Make sure that we have
1724 		 * 1. a TSB
1725 		 * 2. a valid ctx that doesn't get stolen after this point.
1726 		 */
1727 		hatlockp = sfmmu_hat_enter(sfmmup);
1728 
1729 		/*
1730 		 * Swap in the TSB.  hat_init() allocates tsbinfos without
1731 		 * TSBs, but we need one for init, since the kernel does some
1732 		 * special things to set up its stack and needs the TSB to
1733 		 * resolve page faults.
1734 		 */
1735 		sfmmu_tsb_swapin(sfmmup, hatlockp);
1736 
1737 		sfmmu_get_ctx(sfmmup);
1738 
1739 		sfmmu_hat_exit(hatlockp);
1740 	} else {
1741 		ASSERT(allocflag == HAT_ALLOC);
1742 
1743 		hatlockp = sfmmu_hat_enter(sfmmup);
1744 		kpreempt_disable();
1745 
1746 		CPUSET_ADD(sfmmup->sfmmu_cpusran, CPU->cpu_id);
1747 		/*
1748 		 * sfmmu_setctx_sec takes <pgsz|cnum> as a parameter,
1749 		 * pagesize bits don't matter in this case since we are passing
1750 		 * INVALID_CONTEXT to it.
1751 		 * Compatibility Note: hw takes care of MMU_SCONTEXT1
1752 		 */
1753 		sfmmu_setctx_sec(INVALID_CONTEXT);
1754 		sfmmu_clear_utsbinfo();
1755 
1756 		kpreempt_enable();
1757 		sfmmu_hat_exit(hatlockp);
1758 	}
1759 }
1760 
1761 /*
1762  * Free all the translation resources for the specified address space.
1763  * Called from as_free when an address space is being destroyed.
1764  */
1765 void
1766 hat_free_start(struct hat *sfmmup)
1767 {
1768 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
1769 	ASSERT(sfmmup != ksfmmup);
1770 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1771 
1772 	sfmmup->sfmmu_free = 1;
1773 	if (sfmmup->sfmmu_scdp != NULL) {
1774 		sfmmu_leave_scd(sfmmup, 0);
1775 	}
1776 
1777 	ASSERT(sfmmup->sfmmu_scdp == NULL);
1778 }
1779 
1780 void
1781 hat_free_end(struct hat *sfmmup)
1782 {
1783 	int i;
1784 
1785 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1786 	ASSERT(sfmmup->sfmmu_free == 1);
1787 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
1788 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
1789 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
1790 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
1791 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
1792 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
1793 
1794 	if (sfmmup->sfmmu_rmstat) {
1795 		hat_freestat(sfmmup->sfmmu_as, NULL);
1796 	}
1797 
1798 	while (sfmmup->sfmmu_tsb != NULL) {
1799 		struct tsb_info *next = sfmmup->sfmmu_tsb->tsb_next;
1800 		sfmmu_tsbinfo_free(sfmmup->sfmmu_tsb);
1801 		sfmmup->sfmmu_tsb = next;
1802 	}
1803 
1804 	if (sfmmup->sfmmu_srdp != NULL) {
1805 		sfmmu_leave_srd(sfmmup);
1806 		ASSERT(sfmmup->sfmmu_srdp == NULL);
1807 		for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1808 			if (sfmmup->sfmmu_hmeregion_links[i] != NULL) {
1809 				kmem_free(sfmmup->sfmmu_hmeregion_links[i],
1810 				    SFMMU_L2_HMERLINKS_SIZE);
1811 				sfmmup->sfmmu_hmeregion_links[i] = NULL;
1812 			}
1813 		}
1814 	}
1815 	sfmmu_free_sfmmu(sfmmup);
1816 
1817 #ifdef DEBUG
1818 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
1819 		ASSERT(sfmmup->sfmmu_hmeregion_links[i] == NULL);
1820 	}
1821 #endif
1822 
1823 	kmem_cache_free(sfmmuid_cache, sfmmup);
1824 }
1825 
1826 /*
1827  * Set up any translation structures, for the specified address space,
1828  * that are needed or preferred when the process is being swapped in.
1829  */
1830 /* ARGSUSED */
1831 void
1832 hat_swapin(struct hat *hat)
1833 {
1834 	ASSERT(hat->sfmmu_xhat_provider == NULL);
1835 }
1836 
1837 /*
1838  * Free all of the translation resources, for the specified address space,
1839  * that can be freed while the process is swapped out. Called from as_swapout.
1840  * Also, free up the ctx that this process was using.
1841  */
1842 void
1843 hat_swapout(struct hat *sfmmup)
1844 {
1845 	struct hmehash_bucket *hmebp;
1846 	struct hme_blk *hmeblkp;
1847 	struct hme_blk *pr_hblk = NULL;
1848 	struct hme_blk *nx_hblk;
1849 	int i;
1850 	uint64_t hblkpa, prevpa, nx_pa;
1851 	struct hme_blk *list = NULL;
1852 	hatlock_t *hatlockp;
1853 	struct tsb_info *tsbinfop;
1854 	struct free_tsb {
1855 		struct free_tsb *next;
1856 		struct tsb_info *tsbinfop;
1857 	};			/* free list of TSBs */
1858 	struct free_tsb *freelist, *last, *next;
1859 
1860 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1861 	SFMMU_STAT(sf_swapout);
1862 
1863 	/*
1864 	 * There is no way to go from an as to all its translations in sfmmu.
1865 	 * Here is one of the times when we take the big hit and traverse
1866 	 * the hash looking for hme_blks to free up.  Not only do we free up
1867 	 * this as hme_blks but all those that are free.  We are obviously
1868 	 * swapping because we need memory so let's free up as much
1869 	 * as we can.
1870 	 *
1871 	 * Note that we don't flush TLB/TSB here -- it's not necessary
1872 	 * because:
1873 	 *  1) we free the ctx we're using and throw away the TSB(s);
1874 	 *  2) processes aren't runnable while being swapped out.
1875 	 */
1876 	ASSERT(sfmmup != KHATID);
1877 	for (i = 0; i <= UHMEHASH_SZ; i++) {
1878 		hmebp = &uhme_hash[i];
1879 		SFMMU_HASH_LOCK(hmebp);
1880 		hmeblkp = hmebp->hmeblkp;
1881 		hblkpa = hmebp->hmeh_nextpa;
1882 		prevpa = 0;
1883 		pr_hblk = NULL;
1884 		while (hmeblkp) {
1885 
1886 			ASSERT(!hmeblkp->hblk_xhat_bit);
1887 
1888 			if ((hmeblkp->hblk_tag.htag_id == sfmmup) &&
1889 			    !hmeblkp->hblk_shw_bit && !hmeblkp->hblk_lckcnt) {
1890 				ASSERT(!hmeblkp->hblk_shared);
1891 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
1892 				    (caddr_t)get_hblk_base(hmeblkp),
1893 				    get_hblk_endaddr(hmeblkp),
1894 				    NULL, HAT_UNLOAD);
1895 			}
1896 			nx_hblk = hmeblkp->hblk_next;
1897 			nx_pa = hmeblkp->hblk_nextpa;
1898 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
1899 				ASSERT(!hmeblkp->hblk_lckcnt);
1900 				sfmmu_hblk_hash_rm(hmebp, hmeblkp,
1901 				    prevpa, pr_hblk);
1902 				sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
1903 			} else {
1904 				pr_hblk = hmeblkp;
1905 				prevpa = hblkpa;
1906 			}
1907 			hmeblkp = nx_hblk;
1908 			hblkpa = nx_pa;
1909 		}
1910 		SFMMU_HASH_UNLOCK(hmebp);
1911 	}
1912 
1913 	sfmmu_hblks_list_purge(&list);
1914 
1915 	/*
1916 	 * Now free up the ctx so that others can reuse it.
1917 	 */
1918 	hatlockp = sfmmu_hat_enter(sfmmup);
1919 
1920 	sfmmu_invalidate_ctx(sfmmup);
1921 
1922 	/*
1923 	 * Free TSBs, but not tsbinfos, and set SWAPPED flag.
1924 	 * If TSBs were never swapped in, just return.
1925 	 * This implies that we don't support partial swapping
1926 	 * of TSBs -- either all are swapped out, or none are.
1927 	 *
1928 	 * We must hold the HAT lock here to prevent racing with another
1929 	 * thread trying to unmap TTEs from the TSB or running the post-
1930 	 * relocator after relocating the TSB's memory.  Unfortunately, we
1931 	 * can't free memory while holding the HAT lock or we could
1932 	 * deadlock, so we build a list of TSBs to be freed after marking
1933 	 * the tsbinfos as swapped out and free them after dropping the
1934 	 * lock.
1935 	 */
1936 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
1937 		sfmmu_hat_exit(hatlockp);
1938 		return;
1939 	}
1940 
1941 	SFMMU_FLAGS_SET(sfmmup, HAT_SWAPPED);
1942 	last = freelist = NULL;
1943 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
1944 	    tsbinfop = tsbinfop->tsb_next) {
1945 		ASSERT((tsbinfop->tsb_flags & TSB_SWAPPED) == 0);
1946 
1947 		/*
1948 		 * Cast the TSB into a struct free_tsb and put it on the free
1949 		 * list.
1950 		 */
1951 		if (freelist == NULL) {
1952 			last = freelist = (struct free_tsb *)tsbinfop->tsb_va;
1953 		} else {
1954 			last->next = (struct free_tsb *)tsbinfop->tsb_va;
1955 			last = last->next;
1956 		}
1957 		last->next = NULL;
1958 		last->tsbinfop = tsbinfop;
1959 		tsbinfop->tsb_flags |= TSB_SWAPPED;
1960 		/*
1961 		 * Zero out the TTE to clear the valid bit.
1962 		 * Note we can't use a value like 0xbad because we want to
1963 		 * ensure diagnostic bits are NEVER set on TTEs that might
1964 		 * be loaded.  The intent is to catch any invalid access
1965 		 * to the swapped TSB, such as a thread running with a valid
1966 		 * context without first calling sfmmu_tsb_swapin() to
1967 		 * allocate TSB memory.
1968 		 */
1969 		tsbinfop->tsb_tte.ll = 0;
1970 	}
1971 
1972 	/* Now we can drop the lock and free the TSB memory. */
1973 	sfmmu_hat_exit(hatlockp);
1974 	for (; freelist != NULL; freelist = next) {
1975 		next = freelist->next;
1976 		sfmmu_tsb_free(freelist->tsbinfop);
1977 	}
1978 }
1979 
1980 /*
1981  * Duplicate the translations of an as into another newas
1982  */
1983 /* ARGSUSED */
1984 int
1985 hat_dup(struct hat *hat, struct hat *newhat, caddr_t addr, size_t len,
1986 	uint_t flag)
1987 {
1988 	sf_srd_t *srdp;
1989 	sf_scd_t *scdp;
1990 	int i;
1991 	extern uint_t get_color_start(struct as *);
1992 
1993 	ASSERT(hat->sfmmu_xhat_provider == NULL);
1994 	ASSERT((flag == 0) || (flag == HAT_DUP_ALL) || (flag == HAT_DUP_COW) ||
1995 	    (flag == HAT_DUP_SRD));
1996 	ASSERT(hat != ksfmmup);
1997 	ASSERT(newhat != ksfmmup);
1998 	ASSERT(flag != HAT_DUP_ALL || hat->sfmmu_srdp == newhat->sfmmu_srdp);
1999 
2000 	if (flag == HAT_DUP_COW) {
2001 		panic("hat_dup: HAT_DUP_COW not supported");
2002 	}
2003 
2004 	if (flag == HAT_DUP_SRD && ((srdp = hat->sfmmu_srdp) != NULL)) {
2005 		ASSERT(srdp->srd_evp != NULL);
2006 		VN_HOLD(srdp->srd_evp);
2007 		ASSERT(srdp->srd_refcnt > 0);
2008 		newhat->sfmmu_srdp = srdp;
2009 		atomic_add_32((volatile uint_t *)&srdp->srd_refcnt, 1);
2010 	}
2011 
2012 	/*
2013 	 * HAT_DUP_ALL flag is used after as duplication is done.
2014 	 */
2015 	if (flag == HAT_DUP_ALL && ((srdp = newhat->sfmmu_srdp) != NULL)) {
2016 		ASSERT(newhat->sfmmu_srdp->srd_refcnt >= 2);
2017 		newhat->sfmmu_rtteflags = hat->sfmmu_rtteflags;
2018 		if (hat->sfmmu_flags & HAT_4MTEXT_FLAG) {
2019 			newhat->sfmmu_flags |= HAT_4MTEXT_FLAG;
2020 		}
2021 
2022 		/* check if need to join scd */
2023 		if ((scdp = hat->sfmmu_scdp) != NULL &&
2024 		    newhat->sfmmu_scdp != scdp) {
2025 			int ret;
2026 			SF_RGNMAP_IS_SUBSET(&newhat->sfmmu_region_map,
2027 			    &scdp->scd_region_map, ret);
2028 			ASSERT(ret);
2029 			sfmmu_join_scd(scdp, newhat);
2030 			ASSERT(newhat->sfmmu_scdp == scdp &&
2031 			    scdp->scd_refcnt >= 2);
2032 			for (i = 0; i < max_mmu_page_sizes; i++) {
2033 				newhat->sfmmu_ismttecnt[i] =
2034 				    hat->sfmmu_ismttecnt[i];
2035 				newhat->sfmmu_scdismttecnt[i] =
2036 				    hat->sfmmu_scdismttecnt[i];
2037 			}
2038 		}
2039 
2040 		sfmmu_check_page_sizes(newhat, 1);
2041 	}
2042 
2043 	if (flag == HAT_DUP_ALL && consistent_coloring == 0 &&
2044 	    update_proc_pgcolorbase_after_fork != 0) {
2045 		hat->sfmmu_clrbin = get_color_start(hat->sfmmu_as);
2046 	}
2047 	return (0);
2048 }
2049 
2050 void
2051 hat_memload(struct hat *hat, caddr_t addr, struct page *pp,
2052 	uint_t attr, uint_t flags)
2053 {
2054 	hat_do_memload(hat, addr, pp, attr, flags,
2055 	    SFMMU_INVALID_SHMERID);
2056 }
2057 
2058 void
2059 hat_memload_region(struct hat *hat, caddr_t addr, struct page *pp,
2060 	uint_t attr, uint_t flags, hat_region_cookie_t rcookie)
2061 {
2062 	uint_t rid;
2063 	if (rcookie == HAT_INVALID_REGION_COOKIE ||
2064 	    hat->sfmmu_xhat_provider != NULL) {
2065 		hat_do_memload(hat, addr, pp, attr, flags,
2066 		    SFMMU_INVALID_SHMERID);
2067 		return;
2068 	}
2069 	rid = (uint_t)((uint64_t)rcookie);
2070 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2071 	hat_do_memload(hat, addr, pp, attr, flags, rid);
2072 }
2073 
2074 /*
2075  * Set up addr to map to page pp with protection prot.
2076  * As an optimization we also load the TSB with the
2077  * corresponding tte but it is no big deal if  the tte gets kicked out.
2078  */
2079 static void
2080 hat_do_memload(struct hat *hat, caddr_t addr, struct page *pp,
2081 	uint_t attr, uint_t flags, uint_t rid)
2082 {
2083 	tte_t tte;
2084 
2085 
2086 	ASSERT(hat != NULL);
2087 	ASSERT(PAGE_LOCKED(pp));
2088 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2089 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2090 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2091 	SFMMU_VALIDATE_HMERID(hat, rid, addr, MMU_PAGESIZE);
2092 
2093 	if (PP_ISFREE(pp)) {
2094 		panic("hat_memload: loading a mapping to free page %p",
2095 		    (void *)pp);
2096 	}
2097 
2098 	if (hat->sfmmu_xhat_provider) {
2099 		/* no regions for xhats */
2100 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
2101 		XHAT_MEMLOAD(hat, addr, pp, attr, flags);
2102 		return;
2103 	}
2104 
2105 	ASSERT((hat == ksfmmup) ||
2106 	    AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock));
2107 
2108 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2109 		cmn_err(CE_NOTE, "hat_memload: unsupported flags %d",
2110 		    flags & ~SFMMU_LOAD_ALLFLAG);
2111 
2112 	if (hat->sfmmu_rmstat)
2113 		hat_resvstat(MMU_PAGESIZE, hat->sfmmu_as, addr);
2114 
2115 #if defined(SF_ERRATA_57)
2116 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2117 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2118 	    !(flags & HAT_LOAD_SHARE)) {
2119 		cmn_err(CE_WARN, "hat_memload: illegal attempt to make user "
2120 		    " page executable");
2121 		attr &= ~PROT_EXEC;
2122 	}
2123 #endif
2124 
2125 	sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2126 	(void) sfmmu_tteload_array(hat, &tte, addr, &pp, flags, rid);
2127 
2128 	/*
2129 	 * Check TSB and TLB page sizes.
2130 	 */
2131 	if ((flags & HAT_LOAD_SHARE) == 0) {
2132 		sfmmu_check_page_sizes(hat, 1);
2133 	}
2134 }
2135 
2136 /*
2137  * hat_devload can be called to map real memory (e.g.
2138  * /dev/kmem) and even though hat_devload will determine pf is
2139  * for memory, it will be unable to get a shared lock on the
2140  * page (because someone else has it exclusively) and will
2141  * pass dp = NULL.  If tteload doesn't get a non-NULL
2142  * page pointer it can't cache memory.
2143  */
2144 void
2145 hat_devload(struct hat *hat, caddr_t addr, size_t len, pfn_t pfn,
2146 	uint_t attr, int flags)
2147 {
2148 	tte_t tte;
2149 	struct page *pp = NULL;
2150 	int use_lgpg = 0;
2151 
2152 	ASSERT(hat != NULL);
2153 
2154 	if (hat->sfmmu_xhat_provider) {
2155 		XHAT_DEVLOAD(hat, addr, len, pfn, attr, flags);
2156 		return;
2157 	}
2158 
2159 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
2160 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2161 	ASSERT((hat == ksfmmup) ||
2162 	    AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock));
2163 	if (len == 0)
2164 		panic("hat_devload: zero len");
2165 	if (flags & ~SFMMU_LOAD_ALLFLAG)
2166 		cmn_err(CE_NOTE, "hat_devload: unsupported flags %d",
2167 		    flags & ~SFMMU_LOAD_ALLFLAG);
2168 
2169 #if defined(SF_ERRATA_57)
2170 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2171 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2172 	    !(flags & HAT_LOAD_SHARE)) {
2173 		cmn_err(CE_WARN, "hat_devload: illegal attempt to make user "
2174 		    " page executable");
2175 		attr &= ~PROT_EXEC;
2176 	}
2177 #endif
2178 
2179 	/*
2180 	 * If it's a memory page find its pp
2181 	 */
2182 	if (!(flags & HAT_LOAD_NOCONSIST) && pf_is_memory(pfn)) {
2183 		pp = page_numtopp_nolock(pfn);
2184 		if (pp == NULL) {
2185 			flags |= HAT_LOAD_NOCONSIST;
2186 		} else {
2187 			if (PP_ISFREE(pp)) {
2188 				panic("hat_memload: loading "
2189 				    "a mapping to free page %p",
2190 				    (void *)pp);
2191 			}
2192 			if (!PAGE_LOCKED(pp) && !PP_ISNORELOC(pp)) {
2193 				panic("hat_memload: loading a mapping "
2194 				    "to unlocked relocatable page %p",
2195 				    (void *)pp);
2196 			}
2197 			ASSERT(len == MMU_PAGESIZE);
2198 		}
2199 	}
2200 
2201 	if (hat->sfmmu_rmstat)
2202 		hat_resvstat(len, hat->sfmmu_as, addr);
2203 
2204 	if (flags & HAT_LOAD_NOCONSIST) {
2205 		attr |= SFMMU_UNCACHEVTTE;
2206 		use_lgpg = 1;
2207 	}
2208 	if (!pf_is_memory(pfn)) {
2209 		attr |= SFMMU_UNCACHEPTTE | HAT_NOSYNC;
2210 		use_lgpg = 1;
2211 		switch (attr & HAT_ORDER_MASK) {
2212 			case HAT_STRICTORDER:
2213 			case HAT_UNORDERED_OK:
2214 				/*
2215 				 * we set the side effect bit for all non
2216 				 * memory mappings unless merging is ok
2217 				 */
2218 				attr |= SFMMU_SIDEFFECT;
2219 				break;
2220 			case HAT_MERGING_OK:
2221 			case HAT_LOADCACHING_OK:
2222 			case HAT_STORECACHING_OK:
2223 				break;
2224 			default:
2225 				panic("hat_devload: bad attr");
2226 				break;
2227 		}
2228 	}
2229 	while (len) {
2230 		if (!use_lgpg) {
2231 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2232 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2233 			    flags, SFMMU_INVALID_SHMERID);
2234 			len -= MMU_PAGESIZE;
2235 			addr += MMU_PAGESIZE;
2236 			pfn++;
2237 			continue;
2238 		}
2239 		/*
2240 		 *  try to use large pages, check va/pa alignments
2241 		 *  Note that 32M/256M page sizes are not (yet) supported.
2242 		 */
2243 		if ((len >= MMU_PAGESIZE4M) &&
2244 		    !((uintptr_t)addr & MMU_PAGEOFFSET4M) &&
2245 		    !(disable_large_pages & (1 << TTE4M)) &&
2246 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET4M)) {
2247 			sfmmu_memtte(&tte, pfn, attr, TTE4M);
2248 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2249 			    flags, SFMMU_INVALID_SHMERID);
2250 			len -= MMU_PAGESIZE4M;
2251 			addr += MMU_PAGESIZE4M;
2252 			pfn += MMU_PAGESIZE4M / MMU_PAGESIZE;
2253 		} else if ((len >= MMU_PAGESIZE512K) &&
2254 		    !((uintptr_t)addr & MMU_PAGEOFFSET512K) &&
2255 		    !(disable_large_pages & (1 << TTE512K)) &&
2256 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET512K)) {
2257 			sfmmu_memtte(&tte, pfn, attr, TTE512K);
2258 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2259 			    flags, SFMMU_INVALID_SHMERID);
2260 			len -= MMU_PAGESIZE512K;
2261 			addr += MMU_PAGESIZE512K;
2262 			pfn += MMU_PAGESIZE512K / MMU_PAGESIZE;
2263 		} else if ((len >= MMU_PAGESIZE64K) &&
2264 		    !((uintptr_t)addr & MMU_PAGEOFFSET64K) &&
2265 		    !(disable_large_pages & (1 << TTE64K)) &&
2266 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET64K)) {
2267 			sfmmu_memtte(&tte, pfn, attr, TTE64K);
2268 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2269 			    flags, SFMMU_INVALID_SHMERID);
2270 			len -= MMU_PAGESIZE64K;
2271 			addr += MMU_PAGESIZE64K;
2272 			pfn += MMU_PAGESIZE64K / MMU_PAGESIZE;
2273 		} else {
2274 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
2275 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
2276 			    flags, SFMMU_INVALID_SHMERID);
2277 			len -= MMU_PAGESIZE;
2278 			addr += MMU_PAGESIZE;
2279 			pfn++;
2280 		}
2281 	}
2282 
2283 	/*
2284 	 * Check TSB and TLB page sizes.
2285 	 */
2286 	if ((flags & HAT_LOAD_SHARE) == 0) {
2287 		sfmmu_check_page_sizes(hat, 1);
2288 	}
2289 }
2290 
2291 void
2292 hat_memload_array(struct hat *hat, caddr_t addr, size_t len,
2293 	struct page **pps, uint_t attr, uint_t flags)
2294 {
2295 	hat_do_memload_array(hat, addr, len, pps, attr, flags,
2296 	    SFMMU_INVALID_SHMERID);
2297 }
2298 
2299 void
2300 hat_memload_array_region(struct hat *hat, caddr_t addr, size_t len,
2301 	struct page **pps, uint_t attr, uint_t flags,
2302 	hat_region_cookie_t rcookie)
2303 {
2304 	uint_t rid;
2305 	if (rcookie == HAT_INVALID_REGION_COOKIE ||
2306 	    hat->sfmmu_xhat_provider != NULL) {
2307 		hat_do_memload_array(hat, addr, len, pps, attr, flags,
2308 		    SFMMU_INVALID_SHMERID);
2309 		return;
2310 	}
2311 	rid = (uint_t)((uint64_t)rcookie);
2312 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
2313 	hat_do_memload_array(hat, addr, len, pps, attr, flags, rid);
2314 }
2315 
2316 /*
2317  * Map the largest extend possible out of the page array. The array may NOT
2318  * be in order.  The largest possible mapping a page can have
2319  * is specified in the p_szc field.  The p_szc field
2320  * cannot change as long as there any mappings (large or small)
2321  * to any of the pages that make up the large page. (ie. any
2322  * promotion/demotion of page size is not up to the hat but up to
2323  * the page free list manager).  The array
2324  * should consist of properly aligned contigous pages that are
2325  * part of a big page for a large mapping to be created.
2326  */
2327 static void
2328 hat_do_memload_array(struct hat *hat, caddr_t addr, size_t len,
2329 	struct page **pps, uint_t attr, uint_t flags, uint_t rid)
2330 {
2331 	int  ttesz;
2332 	size_t mapsz;
2333 	pgcnt_t	numpg, npgs;
2334 	tte_t tte;
2335 	page_t *pp;
2336 	uint_t large_pages_disable;
2337 
2338 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
2339 	SFMMU_VALIDATE_HMERID(hat, rid, addr, len);
2340 
2341 	if (hat->sfmmu_xhat_provider) {
2342 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
2343 		XHAT_MEMLOAD_ARRAY(hat, addr, len, pps, attr, flags);
2344 		return;
2345 	}
2346 
2347 	if (hat->sfmmu_rmstat)
2348 		hat_resvstat(len, hat->sfmmu_as, addr);
2349 
2350 #if defined(SF_ERRATA_57)
2351 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
2352 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
2353 	    !(flags & HAT_LOAD_SHARE)) {
2354 		cmn_err(CE_WARN, "hat_memload_array: illegal attempt to make "
2355 		    "user page executable");
2356 		attr &= ~PROT_EXEC;
2357 	}
2358 #endif
2359 
2360 	/* Get number of pages */
2361 	npgs = len >> MMU_PAGESHIFT;
2362 
2363 	if (flags & HAT_LOAD_SHARE) {
2364 		large_pages_disable = disable_ism_large_pages;
2365 	} else {
2366 		large_pages_disable = disable_large_pages;
2367 	}
2368 
2369 	if (npgs < NHMENTS || large_pages_disable == LARGE_PAGES_OFF) {
2370 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2371 		    rid);
2372 		return;
2373 	}
2374 
2375 	while (npgs >= NHMENTS) {
2376 		pp = *pps;
2377 		for (ttesz = pp->p_szc; ttesz != TTE8K; ttesz--) {
2378 			/*
2379 			 * Check if this page size is disabled.
2380 			 */
2381 			if (large_pages_disable & (1 << ttesz))
2382 				continue;
2383 
2384 			numpg = TTEPAGES(ttesz);
2385 			mapsz = numpg << MMU_PAGESHIFT;
2386 			if ((npgs >= numpg) &&
2387 			    IS_P2ALIGNED(addr, mapsz) &&
2388 			    IS_P2ALIGNED(pp->p_pagenum, numpg)) {
2389 				/*
2390 				 * At this point we have enough pages and
2391 				 * we know the virtual address and the pfn
2392 				 * are properly aligned.  We still need
2393 				 * to check for physical contiguity but since
2394 				 * it is very likely that this is the case
2395 				 * we will assume they are so and undo
2396 				 * the request if necessary.  It would
2397 				 * be great if we could get a hint flag
2398 				 * like HAT_CONTIG which would tell us
2399 				 * the pages are contigous for sure.
2400 				 */
2401 				sfmmu_memtte(&tte, (*pps)->p_pagenum,
2402 				    attr, ttesz);
2403 				if (!sfmmu_tteload_array(hat, &tte, addr,
2404 				    pps, flags, rid)) {
2405 					break;
2406 				}
2407 			}
2408 		}
2409 		if (ttesz == TTE8K) {
2410 			/*
2411 			 * We were not able to map array using a large page
2412 			 * batch a hmeblk or fraction at a time.
2413 			 */
2414 			numpg = ((uintptr_t)addr >> MMU_PAGESHIFT)
2415 			    & (NHMENTS-1);
2416 			numpg = NHMENTS - numpg;
2417 			ASSERT(numpg <= npgs);
2418 			mapsz = numpg * MMU_PAGESIZE;
2419 			sfmmu_memload_batchsmall(hat, addr, pps, attr, flags,
2420 			    numpg, rid);
2421 		}
2422 		addr += mapsz;
2423 		npgs -= numpg;
2424 		pps += numpg;
2425 	}
2426 
2427 	if (npgs) {
2428 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs,
2429 		    rid);
2430 	}
2431 
2432 	/*
2433 	 * Check TSB and TLB page sizes.
2434 	 */
2435 	if ((flags & HAT_LOAD_SHARE) == 0) {
2436 		sfmmu_check_page_sizes(hat, 1);
2437 	}
2438 }
2439 
2440 /*
2441  * Function tries to batch 8K pages into the same hme blk.
2442  */
2443 static void
2444 sfmmu_memload_batchsmall(struct hat *hat, caddr_t vaddr, page_t **pps,
2445 		    uint_t attr, uint_t flags, pgcnt_t npgs, uint_t rid)
2446 {
2447 	tte_t	tte;
2448 	page_t *pp;
2449 	struct hmehash_bucket *hmebp;
2450 	struct hme_blk *hmeblkp;
2451 	int	index;
2452 
2453 	while (npgs) {
2454 		/*
2455 		 * Acquire the hash bucket.
2456 		 */
2457 		hmebp = sfmmu_tteload_acquire_hashbucket(hat, vaddr, TTE8K,
2458 		    rid);
2459 		ASSERT(hmebp);
2460 
2461 		/*
2462 		 * Find the hment block.
2463 		 */
2464 		hmeblkp = sfmmu_tteload_find_hmeblk(hat, hmebp, vaddr,
2465 		    TTE8K, flags, rid);
2466 		ASSERT(hmeblkp);
2467 
2468 		do {
2469 			/*
2470 			 * Make the tte.
2471 			 */
2472 			pp = *pps;
2473 			sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2474 
2475 			/*
2476 			 * Add the translation.
2477 			 */
2478 			(void) sfmmu_tteload_addentry(hat, hmeblkp, &tte,
2479 			    vaddr, pps, flags, rid);
2480 
2481 			/*
2482 			 * Goto next page.
2483 			 */
2484 			pps++;
2485 			npgs--;
2486 
2487 			/*
2488 			 * Goto next address.
2489 			 */
2490 			vaddr += MMU_PAGESIZE;
2491 
2492 			/*
2493 			 * Don't crossover into a different hmentblk.
2494 			 */
2495 			index = (int)(((uintptr_t)vaddr >> MMU_PAGESHIFT) &
2496 			    (NHMENTS-1));
2497 
2498 		} while (index != 0 && npgs != 0);
2499 
2500 		/*
2501 		 * Release the hash bucket.
2502 		 */
2503 
2504 		sfmmu_tteload_release_hashbucket(hmebp);
2505 	}
2506 }
2507 
2508 /*
2509  * Construct a tte for a page:
2510  *
2511  * tte_valid = 1
2512  * tte_size2 = size & TTE_SZ2_BITS (Panther and Olympus-C only)
2513  * tte_size = size
2514  * tte_nfo = attr & HAT_NOFAULT
2515  * tte_ie = attr & HAT_STRUCTURE_LE
2516  * tte_hmenum = hmenum
2517  * tte_pahi = pp->p_pagenum >> TTE_PASHIFT;
2518  * tte_palo = pp->p_pagenum & TTE_PALOMASK;
2519  * tte_ref = 1 (optimization)
2520  * tte_wr_perm = attr & PROT_WRITE;
2521  * tte_no_sync = attr & HAT_NOSYNC
2522  * tte_lock = attr & SFMMU_LOCKTTE
2523  * tte_cp = !(attr & SFMMU_UNCACHEPTTE)
2524  * tte_cv = !(attr & SFMMU_UNCACHEVTTE)
2525  * tte_e = attr & SFMMU_SIDEFFECT
2526  * tte_priv = !(attr & PROT_USER)
2527  * tte_hwwr = if nosync is set and it is writable we set the mod bit (opt)
2528  * tte_glb = 0
2529  */
2530 void
2531 sfmmu_memtte(tte_t *ttep, pfn_t pfn, uint_t attr, int tte_sz)
2532 {
2533 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2534 
2535 	ttep->tte_inthi = MAKE_TTE_INTHI(pfn, attr, tte_sz, 0 /* hmenum */);
2536 	ttep->tte_intlo = MAKE_TTE_INTLO(pfn, attr, tte_sz, 0 /* hmenum */);
2537 
2538 	if (TTE_IS_NOSYNC(ttep)) {
2539 		TTE_SET_REF(ttep);
2540 		if (TTE_IS_WRITABLE(ttep)) {
2541 			TTE_SET_MOD(ttep);
2542 		}
2543 	}
2544 	if (TTE_IS_NFO(ttep) && TTE_IS_EXECUTABLE(ttep)) {
2545 		panic("sfmmu_memtte: can't set both NFO and EXEC bits");
2546 	}
2547 }
2548 
2549 /*
2550  * This function will add a translation to the hme_blk and allocate the
2551  * hme_blk if one does not exist.
2552  * If a page structure is specified then it will add the
2553  * corresponding hment to the mapping list.
2554  * It will also update the hmenum field for the tte.
2555  *
2556  * Currently this function is only used for kernel mappings.
2557  * So pass invalid region to sfmmu_tteload_array().
2558  */
2559 void
2560 sfmmu_tteload(struct hat *sfmmup, tte_t *ttep, caddr_t vaddr, page_t *pp,
2561 	uint_t flags)
2562 {
2563 	ASSERT(sfmmup == ksfmmup);
2564 	(void) sfmmu_tteload_array(sfmmup, ttep, vaddr, &pp, flags,
2565 	    SFMMU_INVALID_SHMERID);
2566 }
2567 
2568 /*
2569  * Load (ttep != NULL) or unload (ttep == NULL) one entry in the TSB.
2570  * Assumes that a particular page size may only be resident in one TSB.
2571  */
2572 static void
2573 sfmmu_mod_tsb(sfmmu_t *sfmmup, caddr_t vaddr, tte_t *ttep, int ttesz)
2574 {
2575 	struct tsb_info *tsbinfop = NULL;
2576 	uint64_t tag;
2577 	struct tsbe *tsbe_addr;
2578 	uint64_t tsb_base;
2579 	uint_t tsb_size;
2580 	int vpshift = MMU_PAGESHIFT;
2581 	int phys = 0;
2582 
2583 	if (sfmmup == ksfmmup) { /* No support for 32/256M ksfmmu pages */
2584 		phys = ktsb_phys;
2585 		if (ttesz >= TTE4M) {
2586 #ifndef sun4v
2587 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2588 #endif
2589 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2590 			tsb_size = ktsb4m_szcode;
2591 		} else {
2592 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2593 			tsb_size = ktsb_szcode;
2594 		}
2595 	} else {
2596 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2597 
2598 		/*
2599 		 * If there isn't a TSB for this page size, or the TSB is
2600 		 * swapped out, there is nothing to do.  Note that the latter
2601 		 * case seems impossible but can occur if hat_pageunload()
2602 		 * is called on an ISM mapping while the process is swapped
2603 		 * out.
2604 		 */
2605 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2606 			return;
2607 
2608 		/*
2609 		 * If another thread is in the middle of relocating a TSB
2610 		 * we can't unload the entry so set a flag so that the
2611 		 * TSB will be flushed before it can be accessed by the
2612 		 * process.
2613 		 */
2614 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2615 			if (ttep == NULL)
2616 				tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2617 			return;
2618 		}
2619 #if defined(UTSB_PHYS)
2620 		phys = 1;
2621 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2622 #else
2623 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2624 #endif
2625 		tsb_size = tsbinfop->tsb_szc;
2626 	}
2627 	if (ttesz >= TTE4M)
2628 		vpshift = MMU_PAGESHIFT4M;
2629 
2630 	tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2631 	tag = sfmmu_make_tsbtag(vaddr);
2632 
2633 	if (ttep == NULL) {
2634 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2635 	} else {
2636 		if (ttesz >= TTE4M) {
2637 			SFMMU_STAT(sf_tsb_load4m);
2638 		} else {
2639 			SFMMU_STAT(sf_tsb_load8k);
2640 		}
2641 
2642 		sfmmu_load_tsbe(tsbe_addr, tag, ttep, phys);
2643 	}
2644 }
2645 
2646 /*
2647  * Unmap all entries from [start, end) matching the given page size.
2648  *
2649  * This function is used primarily to unmap replicated 64K or 512K entries
2650  * from the TSB that are inserted using the base page size TSB pointer, but
2651  * it may also be called to unmap a range of addresses from the TSB.
2652  */
2653 void
2654 sfmmu_unload_tsb_range(sfmmu_t *sfmmup, caddr_t start, caddr_t end, int ttesz)
2655 {
2656 	struct tsb_info *tsbinfop;
2657 	uint64_t tag;
2658 	struct tsbe *tsbe_addr;
2659 	caddr_t vaddr;
2660 	uint64_t tsb_base;
2661 	int vpshift, vpgsz;
2662 	uint_t tsb_size;
2663 	int phys = 0;
2664 
2665 	/*
2666 	 * Assumptions:
2667 	 *  If ttesz == 8K, 64K or 512K, we walk through the range 8K
2668 	 *  at a time shooting down any valid entries we encounter.
2669 	 *
2670 	 *  If ttesz >= 4M we walk the range 4M at a time shooting
2671 	 *  down any valid mappings we find.
2672 	 */
2673 	if (sfmmup == ksfmmup) {
2674 		phys = ktsb_phys;
2675 		if (ttesz >= TTE4M) {
2676 #ifndef sun4v
2677 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2678 #endif
2679 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2680 			tsb_size = ktsb4m_szcode;
2681 		} else {
2682 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2683 			tsb_size = ktsb_szcode;
2684 		}
2685 	} else {
2686 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2687 
2688 		/*
2689 		 * If there isn't a TSB for this page size, or the TSB is
2690 		 * swapped out, there is nothing to do.  Note that the latter
2691 		 * case seems impossible but can occur if hat_pageunload()
2692 		 * is called on an ISM mapping while the process is swapped
2693 		 * out.
2694 		 */
2695 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2696 			return;
2697 
2698 		/*
2699 		 * If another thread is in the middle of relocating a TSB
2700 		 * we can't unload the entry so set a flag so that the
2701 		 * TSB will be flushed before it can be accessed by the
2702 		 * process.
2703 		 */
2704 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2705 			tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2706 			return;
2707 		}
2708 #if defined(UTSB_PHYS)
2709 		phys = 1;
2710 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2711 #else
2712 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2713 #endif
2714 		tsb_size = tsbinfop->tsb_szc;
2715 	}
2716 	if (ttesz >= TTE4M) {
2717 		vpshift = MMU_PAGESHIFT4M;
2718 		vpgsz = MMU_PAGESIZE4M;
2719 	} else {
2720 		vpshift = MMU_PAGESHIFT;
2721 		vpgsz = MMU_PAGESIZE;
2722 	}
2723 
2724 	for (vaddr = start; vaddr < end; vaddr += vpgsz) {
2725 		tag = sfmmu_make_tsbtag(vaddr);
2726 		tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2727 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2728 	}
2729 }
2730 
2731 /*
2732  * Select the optimum TSB size given the number of mappings
2733  * that need to be cached.
2734  */
2735 static int
2736 sfmmu_select_tsb_szc(pgcnt_t pgcnt)
2737 {
2738 	int szc = 0;
2739 
2740 #ifdef DEBUG
2741 	if (tsb_grow_stress) {
2742 		uint32_t randval = (uint32_t)gettick() >> 4;
2743 		return (randval % (tsb_max_growsize + 1));
2744 	}
2745 #endif	/* DEBUG */
2746 
2747 	while ((szc < tsb_max_growsize) && (pgcnt > SFMMU_RSS_TSBSIZE(szc)))
2748 		szc++;
2749 	return (szc);
2750 }
2751 
2752 /*
2753  * This function will add a translation to the hme_blk and allocate the
2754  * hme_blk if one does not exist.
2755  * If a page structure is specified then it will add the
2756  * corresponding hment to the mapping list.
2757  * It will also update the hmenum field for the tte.
2758  * Furthermore, it attempts to create a large page translation
2759  * for <addr,hat> at page array pps.  It assumes addr and first
2760  * pp is correctly aligned.  It returns 0 if successful and 1 otherwise.
2761  */
2762 static int
2763 sfmmu_tteload_array(sfmmu_t *sfmmup, tte_t *ttep, caddr_t vaddr,
2764 	page_t **pps, uint_t flags, uint_t rid)
2765 {
2766 	struct hmehash_bucket *hmebp;
2767 	struct hme_blk *hmeblkp;
2768 	int 	ret;
2769 	uint_t	size;
2770 
2771 	/*
2772 	 * Get mapping size.
2773 	 */
2774 	size = TTE_CSZ(ttep);
2775 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
2776 
2777 	/*
2778 	 * Acquire the hash bucket.
2779 	 */
2780 	hmebp = sfmmu_tteload_acquire_hashbucket(sfmmup, vaddr, size, rid);
2781 	ASSERT(hmebp);
2782 
2783 	/*
2784 	 * Find the hment block.
2785 	 */
2786 	hmeblkp = sfmmu_tteload_find_hmeblk(sfmmup, hmebp, vaddr, size, flags,
2787 	    rid);
2788 	ASSERT(hmeblkp);
2789 
2790 	/*
2791 	 * Add the translation.
2792 	 */
2793 	ret = sfmmu_tteload_addentry(sfmmup, hmeblkp, ttep, vaddr, pps, flags,
2794 	    rid);
2795 
2796 	/*
2797 	 * Release the hash bucket.
2798 	 */
2799 	sfmmu_tteload_release_hashbucket(hmebp);
2800 
2801 	return (ret);
2802 }
2803 
2804 /*
2805  * Function locks and returns a pointer to the hash bucket for vaddr and size.
2806  */
2807 static struct hmehash_bucket *
2808 sfmmu_tteload_acquire_hashbucket(sfmmu_t *sfmmup, caddr_t vaddr, int size,
2809     uint_t rid)
2810 {
2811 	struct hmehash_bucket *hmebp;
2812 	int hmeshift;
2813 	void *htagid = sfmmutohtagid(sfmmup, rid);
2814 
2815 	ASSERT(htagid != NULL);
2816 
2817 	hmeshift = HME_HASH_SHIFT(size);
2818 
2819 	hmebp = HME_HASH_FUNCTION(htagid, vaddr, hmeshift);
2820 
2821 	SFMMU_HASH_LOCK(hmebp);
2822 
2823 	return (hmebp);
2824 }
2825 
2826 /*
2827  * Function returns a pointer to an hmeblk in the hash bucket, hmebp. If the
2828  * hmeblk doesn't exists for the [sfmmup, vaddr & size] signature, a hmeblk is
2829  * allocated.
2830  */
2831 static struct hme_blk *
2832 sfmmu_tteload_find_hmeblk(sfmmu_t *sfmmup, struct hmehash_bucket *hmebp,
2833 	caddr_t vaddr, uint_t size, uint_t flags, uint_t rid)
2834 {
2835 	hmeblk_tag hblktag;
2836 	int hmeshift;
2837 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
2838 	uint64_t hblkpa, prevpa;
2839 	struct kmem_cache *sfmmu_cache;
2840 	uint_t forcefree;
2841 
2842 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
2843 
2844 	hblktag.htag_id = sfmmutohtagid(sfmmup, rid);
2845 	ASSERT(hblktag.htag_id != NULL);
2846 	hmeshift = HME_HASH_SHIFT(size);
2847 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
2848 	hblktag.htag_rehash = HME_HASH_REHASH(size);
2849 	hblktag.htag_rid = rid;
2850 
2851 ttearray_realloc:
2852 
2853 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, hblkpa,
2854 	    pr_hblk, prevpa, &list);
2855 
2856 	/*
2857 	 * We block until hblk_reserve_lock is released; it's held by
2858 	 * the thread, temporarily using hblk_reserve, until hblk_reserve is
2859 	 * replaced by a hblk from sfmmu8_cache.
2860 	 */
2861 	if (hmeblkp == (struct hme_blk *)hblk_reserve &&
2862 	    hblk_reserve_thread != curthread) {
2863 		SFMMU_HASH_UNLOCK(hmebp);
2864 		mutex_enter(&hblk_reserve_lock);
2865 		mutex_exit(&hblk_reserve_lock);
2866 		SFMMU_STAT(sf_hblk_reserve_hit);
2867 		SFMMU_HASH_LOCK(hmebp);
2868 		goto ttearray_realloc;
2869 	}
2870 
2871 	if (hmeblkp == NULL) {
2872 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
2873 		    hblktag, flags, rid);
2874 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
2875 		ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
2876 	} else {
2877 		/*
2878 		 * It is possible for 8k and 64k hblks to collide since they
2879 		 * have the same rehash value. This is because we
2880 		 * lazily free hblks and 8K/64K blks could be lingering.
2881 		 * If we find size mismatch we free the block and & try again.
2882 		 */
2883 		if (get_hblk_ttesz(hmeblkp) != size) {
2884 			ASSERT(!hmeblkp->hblk_vcnt);
2885 			ASSERT(!hmeblkp->hblk_hmecnt);
2886 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa, pr_hblk);
2887 			sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
2888 			goto ttearray_realloc;
2889 		}
2890 		if (hmeblkp->hblk_shw_bit) {
2891 			/*
2892 			 * if the hblk was previously used as a shadow hblk then
2893 			 * we will change it to a normal hblk
2894 			 */
2895 			ASSERT(!hmeblkp->hblk_shared);
2896 			if (hmeblkp->hblk_shw_mask) {
2897 				sfmmu_shadow_hcleanup(sfmmup, hmeblkp, hmebp);
2898 				ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
2899 				goto ttearray_realloc;
2900 			} else {
2901 				hmeblkp->hblk_shw_bit = 0;
2902 			}
2903 		}
2904 		SFMMU_STAT(sf_hblk_hit);
2905 	}
2906 
2907 	/*
2908 	 * hat_memload() should never call kmem_cache_free(); see block
2909 	 * comment showing the stacktrace in sfmmu_hblk_alloc();
2910 	 * enqueue each hblk in the list to reserve list if it's created
2911 	 * from sfmmu8_cache *and* sfmmup == KHATID.
2912 	 */
2913 	forcefree = (sfmmup == KHATID) ? 1 : 0;
2914 	while ((pr_hblk = list) != NULL) {
2915 		list = pr_hblk->hblk_next;
2916 		sfmmu_cache = get_hblk_cache(pr_hblk);
2917 		if ((sfmmu_cache == sfmmu8_cache) &&
2918 		    sfmmu_put_free_hblk(pr_hblk, forcefree))
2919 			continue;
2920 
2921 		ASSERT(sfmmup != KHATID);
2922 		kmem_cache_free(sfmmu_cache, pr_hblk);
2923 	}
2924 
2925 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
2926 	ASSERT(!hmeblkp->hblk_shw_bit);
2927 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
2928 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
2929 	ASSERT(hmeblkp->hblk_tag.htag_rid == rid);
2930 
2931 	return (hmeblkp);
2932 }
2933 
2934 /*
2935  * Function adds a tte entry into the hmeblk. It returns 0 if successful and 1
2936  * otherwise.
2937  */
2938 static int
2939 sfmmu_tteload_addentry(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, tte_t *ttep,
2940 	caddr_t vaddr, page_t **pps, uint_t flags, uint_t rid)
2941 {
2942 	page_t *pp = *pps;
2943 	int hmenum, size, remap;
2944 	tte_t tteold, flush_tte;
2945 #ifdef DEBUG
2946 	tte_t orig_old;
2947 #endif /* DEBUG */
2948 	struct sf_hment *sfhme;
2949 	kmutex_t *pml, *pmtx;
2950 	hatlock_t *hatlockp;
2951 	int myflt;
2952 
2953 	/*
2954 	 * remove this panic when we decide to let user virtual address
2955 	 * space be >= USERLIMIT.
2956 	 */
2957 	if (!TTE_IS_PRIVILEGED(ttep) && vaddr >= (caddr_t)USERLIMIT)
2958 		panic("user addr %p in kernel space", vaddr);
2959 #if defined(TTE_IS_GLOBAL)
2960 	if (TTE_IS_GLOBAL(ttep))
2961 		panic("sfmmu_tteload: creating global tte");
2962 #endif
2963 
2964 #ifdef DEBUG
2965 	if (pf_is_memory(sfmmu_ttetopfn(ttep, vaddr)) &&
2966 	    !TTE_IS_PCACHEABLE(ttep) && !sfmmu_allow_nc_trans)
2967 		panic("sfmmu_tteload: non cacheable memory tte");
2968 #endif /* DEBUG */
2969 
2970 	/* don't simulate dirty bit for writeable ISM/DISM mappings */
2971 	if ((flags & HAT_LOAD_SHARE) && TTE_IS_WRITABLE(ttep)) {
2972 		TTE_SET_REF(ttep);
2973 		TTE_SET_MOD(ttep);
2974 	}
2975 
2976 	if ((flags & HAT_LOAD_SHARE) || !TTE_IS_REF(ttep) ||
2977 	    !TTE_IS_MOD(ttep)) {
2978 		/*
2979 		 * Don't load TSB for dummy as in ISM.  Also don't preload
2980 		 * the TSB if the TTE isn't writable since we're likely to
2981 		 * fault on it again -- preloading can be fairly expensive.
2982 		 */
2983 		flags |= SFMMU_NO_TSBLOAD;
2984 	}
2985 
2986 	size = TTE_CSZ(ttep);
2987 	switch (size) {
2988 	case TTE8K:
2989 		SFMMU_STAT(sf_tteload8k);
2990 		break;
2991 	case TTE64K:
2992 		SFMMU_STAT(sf_tteload64k);
2993 		break;
2994 	case TTE512K:
2995 		SFMMU_STAT(sf_tteload512k);
2996 		break;
2997 	case TTE4M:
2998 		SFMMU_STAT(sf_tteload4m);
2999 		break;
3000 	case (TTE32M):
3001 		SFMMU_STAT(sf_tteload32m);
3002 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3003 		break;
3004 	case (TTE256M):
3005 		SFMMU_STAT(sf_tteload256m);
3006 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
3007 		break;
3008 	}
3009 
3010 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
3011 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
3012 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) || hmeblkp->hblk_shared);
3013 	ASSERT(SFMMU_IS_SHMERID_VALID(rid) || !hmeblkp->hblk_shared);
3014 
3015 	HBLKTOHME_IDX(sfhme, hmeblkp, vaddr, hmenum);
3016 
3017 	/*
3018 	 * Need to grab mlist lock here so that pageunload
3019 	 * will not change tte behind us.
3020 	 */
3021 	if (pp) {
3022 		pml = sfmmu_mlist_enter(pp);
3023 	}
3024 
3025 	sfmmu_copytte(&sfhme->hme_tte, &tteold);
3026 	/*
3027 	 * Look for corresponding hment and if valid verify
3028 	 * pfns are equal.
3029 	 */
3030 	remap = TTE_IS_VALID(&tteold);
3031 	if (remap) {
3032 		pfn_t	new_pfn, old_pfn;
3033 
3034 		old_pfn = TTE_TO_PFN(vaddr, &tteold);
3035 		new_pfn = TTE_TO_PFN(vaddr, ttep);
3036 
3037 		if (flags & HAT_LOAD_REMAP) {
3038 			/* make sure we are remapping same type of pages */
3039 			if (pf_is_memory(old_pfn) != pf_is_memory(new_pfn)) {
3040 				panic("sfmmu_tteload - tte remap io<->memory");
3041 			}
3042 			if (old_pfn != new_pfn &&
3043 			    (pp != NULL || sfhme->hme_page != NULL)) {
3044 				panic("sfmmu_tteload - tte remap pp != NULL");
3045 			}
3046 		} else if (old_pfn != new_pfn) {
3047 			panic("sfmmu_tteload - tte remap, hmeblkp 0x%p",
3048 			    (void *)hmeblkp);
3049 		}
3050 		ASSERT(TTE_CSZ(&tteold) == TTE_CSZ(ttep));
3051 	}
3052 
3053 	if (pp) {
3054 		if (size == TTE8K) {
3055 #ifdef VAC
3056 			/*
3057 			 * Handle VAC consistency
3058 			 */
3059 			if (!remap && (cache & CACHE_VAC) && !PP_ISNC(pp)) {
3060 				sfmmu_vac_conflict(sfmmup, vaddr, pp);
3061 			}
3062 #endif
3063 
3064 			if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3065 				pmtx = sfmmu_page_enter(pp);
3066 				PP_CLRRO(pp);
3067 				sfmmu_page_exit(pmtx);
3068 			} else if (!PP_ISMAPPED(pp) &&
3069 			    (!TTE_IS_WRITABLE(ttep)) && !(PP_ISMOD(pp))) {
3070 				pmtx = sfmmu_page_enter(pp);
3071 				if (!(PP_ISMOD(pp))) {
3072 					PP_SETRO(pp);
3073 				}
3074 				sfmmu_page_exit(pmtx);
3075 			}
3076 
3077 		} else if (sfmmu_pagearray_setup(vaddr, pps, ttep, remap)) {
3078 			/*
3079 			 * sfmmu_pagearray_setup failed so return
3080 			 */
3081 			sfmmu_mlist_exit(pml);
3082 			return (1);
3083 		}
3084 	}
3085 
3086 	/*
3087 	 * Make sure hment is not on a mapping list.
3088 	 */
3089 	ASSERT(remap || (sfhme->hme_page == NULL));
3090 
3091 	/* if it is not a remap then hme->next better be NULL */
3092 	ASSERT((!remap) ? sfhme->hme_next == NULL : 1);
3093 
3094 	if (flags & HAT_LOAD_LOCK) {
3095 		if ((hmeblkp->hblk_lckcnt + 1) >= MAX_HBLK_LCKCNT) {
3096 			panic("too high lckcnt-hmeblk %p",
3097 			    (void *)hmeblkp);
3098 		}
3099 		atomic_add_32(&hmeblkp->hblk_lckcnt, 1);
3100 
3101 		HBLK_STACK_TRACE(hmeblkp, HBLK_LOCK);
3102 	}
3103 
3104 #ifdef VAC
3105 	if (pp && PP_ISNC(pp)) {
3106 		/*
3107 		 * If the physical page is marked to be uncacheable, like
3108 		 * by a vac conflict, make sure the new mapping is also
3109 		 * uncacheable.
3110 		 */
3111 		TTE_CLR_VCACHEABLE(ttep);
3112 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
3113 	}
3114 #endif
3115 	ttep->tte_hmenum = hmenum;
3116 
3117 #ifdef DEBUG
3118 	orig_old = tteold;
3119 #endif /* DEBUG */
3120 
3121 	while (sfmmu_modifytte_try(&tteold, ttep, &sfhme->hme_tte) < 0) {
3122 		if ((sfmmup == KHATID) &&
3123 		    (flags & (HAT_LOAD_LOCK | HAT_LOAD_REMAP))) {
3124 			sfmmu_copytte(&sfhme->hme_tte, &tteold);
3125 		}
3126 #ifdef DEBUG
3127 		chk_tte(&orig_old, &tteold, ttep, hmeblkp);
3128 #endif /* DEBUG */
3129 	}
3130 	ASSERT(TTE_IS_VALID(&sfhme->hme_tte));
3131 
3132 	if (!TTE_IS_VALID(&tteold)) {
3133 
3134 		atomic_add_16(&hmeblkp->hblk_vcnt, 1);
3135 		if (rid == SFMMU_INVALID_SHMERID) {
3136 			atomic_add_long(&sfmmup->sfmmu_ttecnt[size], 1);
3137 		} else {
3138 			sf_srd_t *srdp = sfmmup->sfmmu_srdp;
3139 			sf_region_t *rgnp = srdp->srd_hmergnp[rid];
3140 			/*
3141 			 * We already accounted for region ttecnt's in sfmmu
3142 			 * during hat_join_region() processing. Here we
3143 			 * only update ttecnt's in region struture.
3144 			 */
3145 			atomic_add_long(&rgnp->rgn_ttecnt[size], 1);
3146 		}
3147 	}
3148 
3149 	myflt = (astosfmmu(curthread->t_procp->p_as) == sfmmup);
3150 	if (size > TTE8K && (flags & HAT_LOAD_SHARE) == 0 &&
3151 	    sfmmup != ksfmmup) {
3152 		uchar_t tteflag = 1 << size;
3153 		if (rid == SFMMU_INVALID_SHMERID) {
3154 			if (!(sfmmup->sfmmu_tteflags & tteflag)) {
3155 				hatlockp = sfmmu_hat_enter(sfmmup);
3156 				sfmmup->sfmmu_tteflags |= tteflag;
3157 				sfmmu_hat_exit(hatlockp);
3158 			}
3159 		} else if (!(sfmmup->sfmmu_rtteflags & tteflag)) {
3160 			hatlockp = sfmmu_hat_enter(sfmmup);
3161 			sfmmup->sfmmu_rtteflags |= tteflag;
3162 			sfmmu_hat_exit(hatlockp);
3163 		}
3164 		/*
3165 		 * Update the current CPU tsbmiss area, so the current thread
3166 		 * won't need to take the tsbmiss for the new pagesize.
3167 		 * The other threads in the process will update their tsb
3168 		 * miss area lazily in sfmmu_tsbmiss_exception() when they
3169 		 * fail to find the translation for a newly added pagesize.
3170 		 */
3171 		if (size > TTE64K && myflt) {
3172 			struct tsbmiss *tsbmp;
3173 			kpreempt_disable();
3174 			tsbmp = &tsbmiss_area[CPU->cpu_id];
3175 			if (rid == SFMMU_INVALID_SHMERID) {
3176 				if (!(tsbmp->uhat_tteflags & tteflag)) {
3177 					tsbmp->uhat_tteflags |= tteflag;
3178 				}
3179 			} else {
3180 				if (!(tsbmp->uhat_rtteflags & tteflag)) {
3181 					tsbmp->uhat_rtteflags |= tteflag;
3182 				}
3183 			}
3184 			kpreempt_enable();
3185 		}
3186 	}
3187 
3188 	if (size >= TTE4M && (flags & HAT_LOAD_TEXT) &&
3189 	    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
3190 		hatlockp = sfmmu_hat_enter(sfmmup);
3191 		SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
3192 		sfmmu_hat_exit(hatlockp);
3193 	}
3194 
3195 	flush_tte.tte_intlo = (tteold.tte_intlo ^ ttep->tte_intlo) &
3196 	    hw_tte.tte_intlo;
3197 	flush_tte.tte_inthi = (tteold.tte_inthi ^ ttep->tte_inthi) &
3198 	    hw_tte.tte_inthi;
3199 
3200 	if (remap && (flush_tte.tte_inthi || flush_tte.tte_intlo)) {
3201 		/*
3202 		 * If remap and new tte differs from old tte we need
3203 		 * to sync the mod bit and flush TLB/TSB.  We don't
3204 		 * need to sync ref bit because we currently always set
3205 		 * ref bit in tteload.
3206 		 */
3207 		ASSERT(TTE_IS_REF(ttep));
3208 		if (TTE_IS_MOD(&tteold)) {
3209 			sfmmu_ttesync(sfmmup, vaddr, &tteold, pp);
3210 		}
3211 		/*
3212 		 * hwtte bits shouldn't change for SRD hmeblks as long as SRD
3213 		 * hmes are only used for read only text. Adding this code for
3214 		 * completeness and future use of shared hmeblks with writable
3215 		 * mappings of VMODSORT vnodes.
3216 		 */
3217 		if (hmeblkp->hblk_shared) {
3218 			cpuset_t cpuset = sfmmu_rgntlb_demap(vaddr,
3219 			    sfmmup->sfmmu_srdp->srd_hmergnp[rid], hmeblkp, 1);
3220 			xt_sync(cpuset);
3221 			SFMMU_STAT_ADD(sf_region_remap_demap, 1);
3222 		} else {
3223 			sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 0);
3224 			xt_sync(sfmmup->sfmmu_cpusran);
3225 		}
3226 	}
3227 
3228 	if ((flags & SFMMU_NO_TSBLOAD) == 0) {
3229 		/*
3230 		 * We only preload 8K and 4M mappings into the TSB, since
3231 		 * 64K and 512K mappings are replicated and hence don't
3232 		 * have a single, unique TSB entry. Ditto for 32M/256M.
3233 		 */
3234 		if (size == TTE8K || size == TTE4M) {
3235 			sf_scd_t *scdp;
3236 			hatlockp = sfmmu_hat_enter(sfmmup);
3237 			/*
3238 			 * Don't preload private TSB if the mapping is used
3239 			 * by the shctx in the SCD.
3240 			 */
3241 			scdp = sfmmup->sfmmu_scdp;
3242 			if (rid == SFMMU_INVALID_SHMERID || scdp == NULL ||
3243 			    !SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
3244 				sfmmu_load_tsb(sfmmup, vaddr, &sfhme->hme_tte,
3245 				    size);
3246 			}
3247 			sfmmu_hat_exit(hatlockp);
3248 		}
3249 	}
3250 	if (pp) {
3251 		if (!remap) {
3252 			HME_ADD(sfhme, pp);
3253 			atomic_add_16(&hmeblkp->hblk_hmecnt, 1);
3254 			ASSERT(hmeblkp->hblk_hmecnt > 0);
3255 
3256 			/*
3257 			 * Cannot ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
3258 			 * see pageunload() for comment.
3259 			 */
3260 		}
3261 		sfmmu_mlist_exit(pml);
3262 	}
3263 
3264 	return (0);
3265 }
3266 /*
3267  * Function unlocks hash bucket.
3268  */
3269 static void
3270 sfmmu_tteload_release_hashbucket(struct hmehash_bucket *hmebp)
3271 {
3272 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3273 	SFMMU_HASH_UNLOCK(hmebp);
3274 }
3275 
3276 /*
3277  * function which checks and sets up page array for a large
3278  * translation.  Will set p_vcolor, p_index, p_ro fields.
3279  * Assumes addr and pfnum of first page are properly aligned.
3280  * Will check for physical contiguity. If check fails it return
3281  * non null.
3282  */
3283 static int
3284 sfmmu_pagearray_setup(caddr_t addr, page_t **pps, tte_t *ttep, int remap)
3285 {
3286 	int 	i, index, ttesz;
3287 	pfn_t	pfnum;
3288 	pgcnt_t	npgs;
3289 	page_t *pp, *pp1;
3290 	kmutex_t *pmtx;
3291 #ifdef VAC
3292 	int osz;
3293 	int cflags = 0;
3294 	int vac_err = 0;
3295 #endif
3296 	int newidx = 0;
3297 
3298 	ttesz = TTE_CSZ(ttep);
3299 
3300 	ASSERT(ttesz > TTE8K);
3301 
3302 	npgs = TTEPAGES(ttesz);
3303 	index = PAGESZ_TO_INDEX(ttesz);
3304 
3305 	pfnum = (*pps)->p_pagenum;
3306 	ASSERT(IS_P2ALIGNED(pfnum, npgs));
3307 
3308 	/*
3309 	 * Save the first pp so we can do HAT_TMPNC at the end.
3310 	 */
3311 	pp1 = *pps;
3312 #ifdef VAC
3313 	osz = fnd_mapping_sz(pp1);
3314 #endif
3315 
3316 	for (i = 0; i < npgs; i++, pps++) {
3317 		pp = *pps;
3318 		ASSERT(PAGE_LOCKED(pp));
3319 		ASSERT(pp->p_szc >= ttesz);
3320 		ASSERT(pp->p_szc == pp1->p_szc);
3321 		ASSERT(sfmmu_mlist_held(pp));
3322 
3323 		/*
3324 		 * XXX is it possible to maintain P_RO on the root only?
3325 		 */
3326 		if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
3327 			pmtx = sfmmu_page_enter(pp);
3328 			PP_CLRRO(pp);
3329 			sfmmu_page_exit(pmtx);
3330 		} else if (!PP_ISMAPPED(pp) && !TTE_IS_WRITABLE(ttep) &&
3331 		    !PP_ISMOD(pp)) {
3332 			pmtx = sfmmu_page_enter(pp);
3333 			if (!(PP_ISMOD(pp))) {
3334 				PP_SETRO(pp);
3335 			}
3336 			sfmmu_page_exit(pmtx);
3337 		}
3338 
3339 		/*
3340 		 * If this is a remap we skip vac & contiguity checks.
3341 		 */
3342 		if (remap)
3343 			continue;
3344 
3345 		/*
3346 		 * set p_vcolor and detect any vac conflicts.
3347 		 */
3348 #ifdef VAC
3349 		if (vac_err == 0) {
3350 			vac_err = sfmmu_vacconflict_array(addr, pp, &cflags);
3351 
3352 		}
3353 #endif
3354 
3355 		/*
3356 		 * Save current index in case we need to undo it.
3357 		 * Note: "PAGESZ_TO_INDEX(sz)	(1 << (sz))"
3358 		 *	"SFMMU_INDEX_SHIFT	6"
3359 		 *	 "SFMMU_INDEX_MASK	((1 << SFMMU_INDEX_SHIFT) - 1)"
3360 		 *	 "PP_MAPINDEX(p_index)	(p_index & SFMMU_INDEX_MASK)"
3361 		 *
3362 		 * So:	index = PAGESZ_TO_INDEX(ttesz);
3363 		 *	if ttesz == 1 then index = 0x2
3364 		 *		    2 then index = 0x4
3365 		 *		    3 then index = 0x8
3366 		 *		    4 then index = 0x10
3367 		 *		    5 then index = 0x20
3368 		 * The code below checks if it's a new pagesize (ie, newidx)
3369 		 * in case we need to take it back out of p_index,
3370 		 * and then or's the new index into the existing index.
3371 		 */
3372 		if ((PP_MAPINDEX(pp) & index) == 0)
3373 			newidx = 1;
3374 		pp->p_index = (PP_MAPINDEX(pp) | index);
3375 
3376 		/*
3377 		 * contiguity check
3378 		 */
3379 		if (pp->p_pagenum != pfnum) {
3380 			/*
3381 			 * If we fail the contiguity test then
3382 			 * the only thing we need to fix is the p_index field.
3383 			 * We might get a few extra flushes but since this
3384 			 * path is rare that is ok.  The p_ro field will
3385 			 * get automatically fixed on the next tteload to
3386 			 * the page.  NO TNC bit is set yet.
3387 			 */
3388 			while (i >= 0) {
3389 				pp = *pps;
3390 				if (newidx)
3391 					pp->p_index = (PP_MAPINDEX(pp) &
3392 					    ~index);
3393 				pps--;
3394 				i--;
3395 			}
3396 			return (1);
3397 		}
3398 		pfnum++;
3399 		addr += MMU_PAGESIZE;
3400 	}
3401 
3402 #ifdef VAC
3403 	if (vac_err) {
3404 		if (ttesz > osz) {
3405 			/*
3406 			 * There are some smaller mappings that causes vac
3407 			 * conflicts. Convert all existing small mappings to
3408 			 * TNC.
3409 			 */
3410 			SFMMU_STAT_ADD(sf_uncache_conflict, npgs);
3411 			sfmmu_page_cache_array(pp1, HAT_TMPNC, CACHE_FLUSH,
3412 			    npgs);
3413 		} else {
3414 			/* EMPTY */
3415 			/*
3416 			 * If there exists an big page mapping,
3417 			 * that means the whole existing big page
3418 			 * has TNC setting already. No need to covert to
3419 			 * TNC again.
3420 			 */
3421 			ASSERT(PP_ISTNC(pp1));
3422 		}
3423 	}
3424 #endif	/* VAC */
3425 
3426 	return (0);
3427 }
3428 
3429 #ifdef VAC
3430 /*
3431  * Routine that detects vac consistency for a large page. It also
3432  * sets virtual color for all pp's for this big mapping.
3433  */
3434 static int
3435 sfmmu_vacconflict_array(caddr_t addr, page_t *pp, int *cflags)
3436 {
3437 	int vcolor, ocolor;
3438 
3439 	ASSERT(sfmmu_mlist_held(pp));
3440 
3441 	if (PP_ISNC(pp)) {
3442 		return (HAT_TMPNC);
3443 	}
3444 
3445 	vcolor = addr_to_vcolor(addr);
3446 	if (PP_NEWPAGE(pp)) {
3447 		PP_SET_VCOLOR(pp, vcolor);
3448 		return (0);
3449 	}
3450 
3451 	ocolor = PP_GET_VCOLOR(pp);
3452 	if (ocolor == vcolor) {
3453 		return (0);
3454 	}
3455 
3456 	if (!PP_ISMAPPED(pp)) {
3457 		/*
3458 		 * Previous user of page had a differnet color
3459 		 * but since there are no current users
3460 		 * we just flush the cache and change the color.
3461 		 * As an optimization for large pages we flush the
3462 		 * entire cache of that color and set a flag.
3463 		 */
3464 		SFMMU_STAT(sf_pgcolor_conflict);
3465 		if (!CacheColor_IsFlushed(*cflags, ocolor)) {
3466 			CacheColor_SetFlushed(*cflags, ocolor);
3467 			sfmmu_cache_flushcolor(ocolor, pp->p_pagenum);
3468 		}
3469 		PP_SET_VCOLOR(pp, vcolor);
3470 		return (0);
3471 	}
3472 
3473 	/*
3474 	 * We got a real conflict with a current mapping.
3475 	 * set flags to start unencaching all mappings
3476 	 * and return failure so we restart looping
3477 	 * the pp array from the beginning.
3478 	 */
3479 	return (HAT_TMPNC);
3480 }
3481 #endif	/* VAC */
3482 
3483 /*
3484  * creates a large page shadow hmeblk for a tte.
3485  * The purpose of this routine is to allow us to do quick unloads because
3486  * the vm layer can easily pass a very large but sparsely populated range.
3487  */
3488 static struct hme_blk *
3489 sfmmu_shadow_hcreate(sfmmu_t *sfmmup, caddr_t vaddr, int ttesz, uint_t flags)
3490 {
3491 	struct hmehash_bucket *hmebp;
3492 	hmeblk_tag hblktag;
3493 	int hmeshift, size, vshift;
3494 	uint_t shw_mask, newshw_mask;
3495 	struct hme_blk *hmeblkp;
3496 
3497 	ASSERT(sfmmup != KHATID);
3498 	if (mmu_page_sizes == max_mmu_page_sizes) {
3499 		ASSERT(ttesz < TTE256M);
3500 	} else {
3501 		ASSERT(ttesz < TTE4M);
3502 		ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
3503 		ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
3504 	}
3505 
3506 	if (ttesz == TTE8K) {
3507 		size = TTE512K;
3508 	} else {
3509 		size = ++ttesz;
3510 	}
3511 
3512 	hblktag.htag_id = sfmmup;
3513 	hmeshift = HME_HASH_SHIFT(size);
3514 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
3515 	hblktag.htag_rehash = HME_HASH_REHASH(size);
3516 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3517 	hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
3518 
3519 	SFMMU_HASH_LOCK(hmebp);
3520 
3521 	HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
3522 	ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
3523 	if (hmeblkp == NULL) {
3524 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
3525 		    hblktag, flags, SFMMU_INVALID_SHMERID);
3526 	}
3527 	ASSERT(hmeblkp);
3528 	if (!hmeblkp->hblk_shw_mask) {
3529 		/*
3530 		 * if this is a unused hblk it was just allocated or could
3531 		 * potentially be a previous large page hblk so we need to
3532 		 * set the shadow bit.
3533 		 */
3534 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3535 		hmeblkp->hblk_shw_bit = 1;
3536 	} else if (hmeblkp->hblk_shw_bit == 0) {
3537 		panic("sfmmu_shadow_hcreate: shw bit not set in hmeblkp 0x%p",
3538 		    (void *)hmeblkp);
3539 	}
3540 	ASSERT(hmeblkp->hblk_shw_bit == 1);
3541 	ASSERT(!hmeblkp->hblk_shared);
3542 	vshift = vaddr_to_vshift(hblktag, vaddr, size);
3543 	ASSERT(vshift < 8);
3544 	/*
3545 	 * Atomically set shw mask bit
3546 	 */
3547 	do {
3548 		shw_mask = hmeblkp->hblk_shw_mask;
3549 		newshw_mask = shw_mask | (1 << vshift);
3550 		newshw_mask = cas32(&hmeblkp->hblk_shw_mask, shw_mask,
3551 		    newshw_mask);
3552 	} while (newshw_mask != shw_mask);
3553 
3554 	SFMMU_HASH_UNLOCK(hmebp);
3555 
3556 	return (hmeblkp);
3557 }
3558 
3559 /*
3560  * This routine cleanup a previous shadow hmeblk and changes it to
3561  * a regular hblk.  This happens rarely but it is possible
3562  * when a process wants to use large pages and there are hblks still
3563  * lying around from the previous as that used these hmeblks.
3564  * The alternative was to cleanup the shadow hblks at unload time
3565  * but since so few user processes actually use large pages, it is
3566  * better to be lazy and cleanup at this time.
3567  */
3568 static void
3569 sfmmu_shadow_hcleanup(sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
3570 	struct hmehash_bucket *hmebp)
3571 {
3572 	caddr_t addr, endaddr;
3573 	int hashno, size;
3574 
3575 	ASSERT(hmeblkp->hblk_shw_bit);
3576 	ASSERT(!hmeblkp->hblk_shared);
3577 
3578 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3579 
3580 	if (!hmeblkp->hblk_shw_mask) {
3581 		hmeblkp->hblk_shw_bit = 0;
3582 		return;
3583 	}
3584 	addr = (caddr_t)get_hblk_base(hmeblkp);
3585 	endaddr = get_hblk_endaddr(hmeblkp);
3586 	size = get_hblk_ttesz(hmeblkp);
3587 	hashno = size - 1;
3588 	ASSERT(hashno > 0);
3589 	SFMMU_HASH_UNLOCK(hmebp);
3590 
3591 	sfmmu_free_hblks(sfmmup, addr, endaddr, hashno);
3592 
3593 	SFMMU_HASH_LOCK(hmebp);
3594 }
3595 
3596 static void
3597 sfmmu_free_hblks(sfmmu_t *sfmmup, caddr_t addr, caddr_t endaddr,
3598 	int hashno)
3599 {
3600 	int hmeshift, shadow = 0;
3601 	hmeblk_tag hblktag;
3602 	struct hmehash_bucket *hmebp;
3603 	struct hme_blk *hmeblkp;
3604 	struct hme_blk *nx_hblk, *pr_hblk, *list = NULL;
3605 	uint64_t hblkpa, prevpa, nx_pa;
3606 
3607 	ASSERT(hashno > 0);
3608 	hblktag.htag_id = sfmmup;
3609 	hblktag.htag_rehash = hashno;
3610 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3611 
3612 	hmeshift = HME_HASH_SHIFT(hashno);
3613 
3614 	while (addr < endaddr) {
3615 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3616 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3617 		SFMMU_HASH_LOCK(hmebp);
3618 		/* inline HME_HASH_SEARCH */
3619 		hmeblkp = hmebp->hmeblkp;
3620 		hblkpa = hmebp->hmeh_nextpa;
3621 		prevpa = 0;
3622 		pr_hblk = NULL;
3623 		while (hmeblkp) {
3624 			ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
3625 			if (HTAGS_EQ(hmeblkp->hblk_tag, hblktag)) {
3626 				/* found hme_blk */
3627 				ASSERT(!hmeblkp->hblk_shared);
3628 				if (hmeblkp->hblk_shw_bit) {
3629 					if (hmeblkp->hblk_shw_mask) {
3630 						shadow = 1;
3631 						sfmmu_shadow_hcleanup(sfmmup,
3632 						    hmeblkp, hmebp);
3633 						break;
3634 					} else {
3635 						hmeblkp->hblk_shw_bit = 0;
3636 					}
3637 				}
3638 
3639 				/*
3640 				 * Hblk_hmecnt and hblk_vcnt could be non zero
3641 				 * since hblk_unload() does not gurantee that.
3642 				 *
3643 				 * XXX - this could cause tteload() to spin
3644 				 * where sfmmu_shadow_hcleanup() is called.
3645 				 */
3646 			}
3647 
3648 			nx_hblk = hmeblkp->hblk_next;
3649 			nx_pa = hmeblkp->hblk_nextpa;
3650 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
3651 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa,
3652 				    pr_hblk);
3653 				sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
3654 			} else {
3655 				pr_hblk = hmeblkp;
3656 				prevpa = hblkpa;
3657 			}
3658 			hmeblkp = nx_hblk;
3659 			hblkpa = nx_pa;
3660 		}
3661 
3662 		SFMMU_HASH_UNLOCK(hmebp);
3663 
3664 		if (shadow) {
3665 			/*
3666 			 * We found another shadow hblk so cleaned its
3667 			 * children.  We need to go back and cleanup
3668 			 * the original hblk so we don't change the
3669 			 * addr.
3670 			 */
3671 			shadow = 0;
3672 		} else {
3673 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
3674 			    (1 << hmeshift));
3675 		}
3676 	}
3677 	sfmmu_hblks_list_purge(&list);
3678 }
3679 
3680 /*
3681  * This routine's job is to delete stale invalid shared hmeregions hmeblks that
3682  * may still linger on after pageunload.
3683  */
3684 static void
3685 sfmmu_cleanup_rhblk(sf_srd_t *srdp, caddr_t addr, uint_t rid, int ttesz)
3686 {
3687 	int hmeshift;
3688 	hmeblk_tag hblktag;
3689 	struct hmehash_bucket *hmebp;
3690 	struct hme_blk *hmeblkp;
3691 	struct hme_blk *pr_hblk;
3692 	struct hme_blk *list = NULL;
3693 	uint64_t hblkpa, prevpa;
3694 
3695 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3696 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3697 
3698 	hmeshift = HME_HASH_SHIFT(ttesz);
3699 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3700 	hblktag.htag_rehash = ttesz;
3701 	hblktag.htag_rid = rid;
3702 	hblktag.htag_id = srdp;
3703 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3704 
3705 	SFMMU_HASH_LOCK(hmebp);
3706 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, hblkpa, pr_hblk,
3707 	    prevpa, &list);
3708 	if (hmeblkp != NULL) {
3709 		ASSERT(hmeblkp->hblk_shared);
3710 		ASSERT(!hmeblkp->hblk_shw_bit);
3711 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3712 			panic("sfmmu_cleanup_rhblk: valid hmeblk");
3713 		}
3714 		ASSERT(!hmeblkp->hblk_lckcnt);
3715 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa, pr_hblk);
3716 		sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
3717 	}
3718 	SFMMU_HASH_UNLOCK(hmebp);
3719 	sfmmu_hblks_list_purge(&list);
3720 }
3721 
3722 /* ARGSUSED */
3723 static void
3724 sfmmu_rgn_cb_noop(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
3725     size_t r_size, void *r_obj, u_offset_t r_objoff)
3726 {
3727 }
3728 
3729 /*
3730  * Searches for an hmeblk which maps addr, then unloads this mapping
3731  * and updates *eaddrp, if the hmeblk is found.
3732  */
3733 static void
3734 sfmmu_unload_hmeregion_va(sf_srd_t *srdp, uint_t rid, caddr_t addr,
3735     caddr_t eaddr, int ttesz, caddr_t *eaddrp)
3736 {
3737 	int hmeshift;
3738 	hmeblk_tag hblktag;
3739 	struct hmehash_bucket *hmebp;
3740 	struct hme_blk *hmeblkp;
3741 	struct hme_blk *pr_hblk;
3742 	struct hme_blk *list = NULL;
3743 	uint64_t hblkpa, prevpa;
3744 
3745 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3746 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3747 	ASSERT(ttesz >= HBLK_MIN_TTESZ);
3748 
3749 	hmeshift = HME_HASH_SHIFT(ttesz);
3750 	hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3751 	hblktag.htag_rehash = ttesz;
3752 	hblktag.htag_rid = rid;
3753 	hblktag.htag_id = srdp;
3754 	hmebp = HME_HASH_FUNCTION(srdp, addr, hmeshift);
3755 
3756 	SFMMU_HASH_LOCK(hmebp);
3757 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, hblkpa, pr_hblk,
3758 	    prevpa, &list);
3759 	if (hmeblkp != NULL) {
3760 		ASSERT(hmeblkp->hblk_shared);
3761 		ASSERT(!hmeblkp->hblk_lckcnt);
3762 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
3763 			*eaddrp = sfmmu_hblk_unload(NULL, hmeblkp, addr,
3764 			    eaddr, NULL, HAT_UNLOAD);
3765 			ASSERT(*eaddrp > addr);
3766 		}
3767 		ASSERT(!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt);
3768 		sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa, pr_hblk);
3769 		sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
3770 	}
3771 	SFMMU_HASH_UNLOCK(hmebp);
3772 	sfmmu_hblks_list_purge(&list);
3773 }
3774 
3775 static void
3776 sfmmu_unload_hmeregion(sf_srd_t *srdp, sf_region_t *rgnp)
3777 {
3778 	int ttesz = rgnp->rgn_pgszc;
3779 	size_t rsz = rgnp->rgn_size;
3780 	caddr_t rsaddr = rgnp->rgn_saddr;
3781 	caddr_t readdr = rsaddr + rsz;
3782 	caddr_t rhsaddr;
3783 	caddr_t va;
3784 	uint_t rid = rgnp->rgn_id;
3785 	caddr_t cbsaddr;
3786 	caddr_t cbeaddr;
3787 	hat_rgn_cb_func_t rcbfunc;
3788 	ulong_t cnt;
3789 
3790 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
3791 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3792 
3793 	ASSERT(IS_P2ALIGNED(rsaddr, TTEBYTES(ttesz)));
3794 	ASSERT(IS_P2ALIGNED(rsz, TTEBYTES(ttesz)));
3795 	if (ttesz < HBLK_MIN_TTESZ) {
3796 		ttesz = HBLK_MIN_TTESZ;
3797 		rhsaddr = (caddr_t)P2ALIGN((uintptr_t)rsaddr, HBLK_MIN_BYTES);
3798 	} else {
3799 		rhsaddr = rsaddr;
3800 	}
3801 
3802 	if ((rcbfunc = rgnp->rgn_cb_function) == NULL) {
3803 		rcbfunc = sfmmu_rgn_cb_noop;
3804 	}
3805 
3806 	while (ttesz >= HBLK_MIN_TTESZ) {
3807 		cbsaddr = rsaddr;
3808 		cbeaddr = rsaddr;
3809 		if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
3810 			ttesz--;
3811 			continue;
3812 		}
3813 		cnt = 0;
3814 		va = rsaddr;
3815 		while (va < readdr) {
3816 			ASSERT(va >= rhsaddr);
3817 			if (va != cbeaddr) {
3818 				if (cbeaddr != cbsaddr) {
3819 					ASSERT(cbeaddr > cbsaddr);
3820 					(*rcbfunc)(cbsaddr, cbeaddr,
3821 					    rsaddr, rsz, rgnp->rgn_obj,
3822 					    rgnp->rgn_objoff);
3823 				}
3824 				cbsaddr = va;
3825 				cbeaddr = va;
3826 			}
3827 			sfmmu_unload_hmeregion_va(srdp, rid, va, readdr,
3828 			    ttesz, &cbeaddr);
3829 			cnt++;
3830 			va = rhsaddr + (cnt << TTE_PAGE_SHIFT(ttesz));
3831 		}
3832 		if (cbeaddr != cbsaddr) {
3833 			ASSERT(cbeaddr > cbsaddr);
3834 			(*rcbfunc)(cbsaddr, cbeaddr, rsaddr,
3835 			    rsz, rgnp->rgn_obj,
3836 			    rgnp->rgn_objoff);
3837 		}
3838 		ttesz--;
3839 	}
3840 }
3841 
3842 /*
3843  * Release one hardware address translation lock on the given address range.
3844  */
3845 void
3846 hat_unlock(struct hat *sfmmup, caddr_t addr, size_t len)
3847 {
3848 	struct hmehash_bucket *hmebp;
3849 	hmeblk_tag hblktag;
3850 	int hmeshift, hashno = 1;
3851 	struct hme_blk *hmeblkp, *list = NULL;
3852 	caddr_t endaddr;
3853 
3854 	ASSERT(sfmmup != NULL);
3855 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3856 
3857 	ASSERT((sfmmup == ksfmmup) ||
3858 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
3859 	ASSERT((len & MMU_PAGEOFFSET) == 0);
3860 	endaddr = addr + len;
3861 	hblktag.htag_id = sfmmup;
3862 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
3863 
3864 	/*
3865 	 * Spitfire supports 4 page sizes.
3866 	 * Most pages are expected to be of the smallest page size (8K) and
3867 	 * these will not need to be rehashed. 64K pages also don't need to be
3868 	 * rehashed because an hmeblk spans 64K of address space. 512K pages
3869 	 * might need 1 rehash and and 4M pages might need 2 rehashes.
3870 	 */
3871 	while (addr < endaddr) {
3872 		hmeshift = HME_HASH_SHIFT(hashno);
3873 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3874 		hblktag.htag_rehash = hashno;
3875 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3876 
3877 		SFMMU_HASH_LOCK(hmebp);
3878 
3879 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
3880 		if (hmeblkp != NULL) {
3881 			ASSERT(!hmeblkp->hblk_shared);
3882 			/*
3883 			 * If we encounter a shadow hmeblk then
3884 			 * we know there are no valid hmeblks mapping
3885 			 * this address at this size or larger.
3886 			 * Just increment address by the smallest
3887 			 * page size.
3888 			 */
3889 			if (hmeblkp->hblk_shw_bit) {
3890 				addr += MMU_PAGESIZE;
3891 			} else {
3892 				addr = sfmmu_hblk_unlock(hmeblkp, addr,
3893 				    endaddr);
3894 			}
3895 			SFMMU_HASH_UNLOCK(hmebp);
3896 			hashno = 1;
3897 			continue;
3898 		}
3899 		SFMMU_HASH_UNLOCK(hmebp);
3900 
3901 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
3902 			/*
3903 			 * We have traversed the whole list and rehashed
3904 			 * if necessary without finding the address to unlock
3905 			 * which should never happen.
3906 			 */
3907 			panic("sfmmu_unlock: addr not found. "
3908 			    "addr %p hat %p", (void *)addr, (void *)sfmmup);
3909 		} else {
3910 			hashno++;
3911 		}
3912 	}
3913 
3914 	sfmmu_hblks_list_purge(&list);
3915 }
3916 
3917 void
3918 hat_unlock_region(struct hat *sfmmup, caddr_t addr, size_t len,
3919     hat_region_cookie_t rcookie)
3920 {
3921 	sf_srd_t *srdp;
3922 	sf_region_t *rgnp;
3923 	int ttesz;
3924 	uint_t rid;
3925 	caddr_t eaddr;
3926 	caddr_t va;
3927 	int hmeshift;
3928 	hmeblk_tag hblktag;
3929 	struct hmehash_bucket *hmebp;
3930 	struct hme_blk *hmeblkp;
3931 	struct hme_blk *pr_hblk;
3932 	struct hme_blk *list;
3933 	uint64_t hblkpa, prevpa;
3934 
3935 	if (rcookie == HAT_INVALID_REGION_COOKIE) {
3936 		hat_unlock(sfmmup, addr, len);
3937 		return;
3938 	}
3939 
3940 	ASSERT(sfmmup != NULL);
3941 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3942 	ASSERT(sfmmup != ksfmmup);
3943 
3944 	srdp = sfmmup->sfmmu_srdp;
3945 	rid = (uint_t)((uint64_t)rcookie);
3946 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
3947 	eaddr = addr + len;
3948 	va = addr;
3949 	list = NULL;
3950 	rgnp = srdp->srd_hmergnp[rid];
3951 	SFMMU_VALIDATE_HMERID(sfmmup, rid, addr, len);
3952 
3953 	ASSERT(IS_P2ALIGNED(addr, TTEBYTES(rgnp->rgn_pgszc)));
3954 	ASSERT(IS_P2ALIGNED(len, TTEBYTES(rgnp->rgn_pgszc)));
3955 	if (rgnp->rgn_pgszc < HBLK_MIN_TTESZ) {
3956 		ttesz = HBLK_MIN_TTESZ;
3957 	} else {
3958 		ttesz = rgnp->rgn_pgszc;
3959 	}
3960 	while (va < eaddr) {
3961 		while (ttesz < rgnp->rgn_pgszc &&
3962 		    IS_P2ALIGNED(va, TTEBYTES(ttesz + 1))) {
3963 			ttesz++;
3964 		}
3965 		while (ttesz >= HBLK_MIN_TTESZ) {
3966 			if (!(rgnp->rgn_hmeflags & (1 << ttesz))) {
3967 				ttesz--;
3968 				continue;
3969 			}
3970 			hmeshift = HME_HASH_SHIFT(ttesz);
3971 			hblktag.htag_bspage = HME_HASH_BSPAGE(va, hmeshift);
3972 			hblktag.htag_rehash = ttesz;
3973 			hblktag.htag_rid = rid;
3974 			hblktag.htag_id = srdp;
3975 			hmebp = HME_HASH_FUNCTION(srdp, va, hmeshift);
3976 			SFMMU_HASH_LOCK(hmebp);
3977 			HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, hblkpa,
3978 			    pr_hblk, prevpa, &list);
3979 			if (hmeblkp == NULL) {
3980 				SFMMU_HASH_UNLOCK(hmebp);
3981 				ttesz--;
3982 				continue;
3983 			}
3984 			ASSERT(hmeblkp->hblk_shared);
3985 			va = sfmmu_hblk_unlock(hmeblkp, va, eaddr);
3986 			ASSERT(va >= eaddr ||
3987 			    IS_P2ALIGNED((uintptr_t)va, TTEBYTES(ttesz)));
3988 			SFMMU_HASH_UNLOCK(hmebp);
3989 			break;
3990 		}
3991 		if (ttesz < HBLK_MIN_TTESZ) {
3992 			panic("hat_unlock_region: addr not found "
3993 			    "addr %p hat %p", va, sfmmup);
3994 		}
3995 	}
3996 	sfmmu_hblks_list_purge(&list);
3997 }
3998 
3999 /*
4000  * Function to unlock a range of addresses in an hmeblk.  It returns the
4001  * next address that needs to be unlocked.
4002  * Should be called with the hash lock held.
4003  */
4004 static caddr_t
4005 sfmmu_hblk_unlock(struct hme_blk *hmeblkp, caddr_t addr, caddr_t endaddr)
4006 {
4007 	struct sf_hment *sfhme;
4008 	tte_t tteold, ttemod;
4009 	int ttesz, ret;
4010 
4011 	ASSERT(in_hblk_range(hmeblkp, addr));
4012 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4013 
4014 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4015 	ttesz = get_hblk_ttesz(hmeblkp);
4016 
4017 	HBLKTOHME(sfhme, hmeblkp, addr);
4018 	while (addr < endaddr) {
4019 readtte:
4020 		sfmmu_copytte(&sfhme->hme_tte, &tteold);
4021 		if (TTE_IS_VALID(&tteold)) {
4022 
4023 			ttemod = tteold;
4024 
4025 			ret = sfmmu_modifytte_try(&tteold, &ttemod,
4026 			    &sfhme->hme_tte);
4027 
4028 			if (ret < 0)
4029 				goto readtte;
4030 
4031 			if (hmeblkp->hblk_lckcnt == 0)
4032 				panic("zero hblk lckcnt");
4033 
4034 			if (((uintptr_t)addr + TTEBYTES(ttesz)) >
4035 			    (uintptr_t)endaddr)
4036 				panic("can't unlock large tte");
4037 
4038 			ASSERT(hmeblkp->hblk_lckcnt > 0);
4039 			atomic_add_32(&hmeblkp->hblk_lckcnt, -1);
4040 			HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
4041 		} else {
4042 			panic("sfmmu_hblk_unlock: invalid tte");
4043 		}
4044 		addr += TTEBYTES(ttesz);
4045 		sfhme++;
4046 	}
4047 	return (addr);
4048 }
4049 
4050 /*
4051  * Physical Address Mapping Framework
4052  *
4053  * General rules:
4054  *
4055  * (1) Applies only to seg_kmem memory pages. To make things easier,
4056  *     seg_kpm addresses are also accepted by the routines, but nothing
4057  *     is done with them since by definition their PA mappings are static.
4058  * (2) hat_add_callback() may only be called while holding the page lock
4059  *     SE_SHARED or SE_EXCL of the underlying page (e.g., as_pagelock()),
4060  *     or passing HAC_PAGELOCK flag.
4061  * (3) prehandler() and posthandler() may not call hat_add_callback() or
4062  *     hat_delete_callback(), nor should they allocate memory. Post quiesce
4063  *     callbacks may not sleep or acquire adaptive mutex locks.
4064  * (4) Either prehandler() or posthandler() (but not both) may be specified
4065  *     as being NULL.  Specifying an errhandler() is optional.
4066  *
4067  * Details of using the framework:
4068  *
4069  * registering a callback (hat_register_callback())
4070  *
4071  *	Pass prehandler, posthandler, errhandler addresses
4072  *	as described below. If capture_cpus argument is nonzero,
4073  *	suspend callback to the prehandler will occur with CPUs
4074  *	captured and executing xc_loop() and CPUs will remain
4075  *	captured until after the posthandler suspend callback
4076  *	occurs.
4077  *
4078  * adding a callback (hat_add_callback())
4079  *
4080  *      as_pagelock();
4081  *	hat_add_callback();
4082  *      save returned pfn in private data structures or program registers;
4083  *      as_pageunlock();
4084  *
4085  * prehandler()
4086  *
4087  *	Stop all accesses by physical address to this memory page.
4088  *	Called twice: the first, PRESUSPEND, is a context safe to acquire
4089  *	adaptive locks. The second, SUSPEND, is called at high PIL with
4090  *	CPUs captured so adaptive locks may NOT be acquired (and all spin
4091  *	locks must be XCALL_PIL or higher locks).
4092  *
4093  *	May return the following errors:
4094  *		EIO:	A fatal error has occurred. This will result in panic.
4095  *		EAGAIN:	The page cannot be suspended. This will fail the
4096  *			relocation.
4097  *		0:	Success.
4098  *
4099  * posthandler()
4100  *
4101  *      Save new pfn in private data structures or program registers;
4102  *	not allowed to fail (non-zero return values will result in panic).
4103  *
4104  * errhandler()
4105  *
4106  *	called when an error occurs related to the callback.  Currently
4107  *	the only such error is HAT_CB_ERR_LEAKED which indicates that
4108  *	a page is being freed, but there are still outstanding callback(s)
4109  *	registered on the page.
4110  *
4111  * removing a callback (hat_delete_callback(); e.g., prior to freeing memory)
4112  *
4113  *	stop using physical address
4114  *	hat_delete_callback();
4115  *
4116  */
4117 
4118 /*
4119  * Register a callback class.  Each subsystem should do this once and
4120  * cache the id_t returned for use in setting up and tearing down callbacks.
4121  *
4122  * There is no facility for removing callback IDs once they are created;
4123  * the "key" should be unique for each module, so in case a module is unloaded
4124  * and subsequently re-loaded, we can recycle the module's previous entry.
4125  */
4126 id_t
4127 hat_register_callback(int key,
4128 	int (*prehandler)(caddr_t, uint_t, uint_t, void *),
4129 	int (*posthandler)(caddr_t, uint_t, uint_t, void *, pfn_t),
4130 	int (*errhandler)(caddr_t, uint_t, uint_t, void *),
4131 	int capture_cpus)
4132 {
4133 	id_t id;
4134 
4135 	/*
4136 	 * Search the table for a pre-existing callback associated with
4137 	 * the identifier "key".  If one exists, we re-use that entry in
4138 	 * the table for this instance, otherwise we assign the next
4139 	 * available table slot.
4140 	 */
4141 	for (id = 0; id < sfmmu_max_cb_id; id++) {
4142 		if (sfmmu_cb_table[id].key == key)
4143 			break;
4144 	}
4145 
4146 	if (id == sfmmu_max_cb_id) {
4147 		id = sfmmu_cb_nextid++;
4148 		if (id >= sfmmu_max_cb_id)
4149 			panic("hat_register_callback: out of callback IDs");
4150 	}
4151 
4152 	ASSERT(prehandler != NULL || posthandler != NULL);
4153 
4154 	sfmmu_cb_table[id].key = key;
4155 	sfmmu_cb_table[id].prehandler = prehandler;
4156 	sfmmu_cb_table[id].posthandler = posthandler;
4157 	sfmmu_cb_table[id].errhandler = errhandler;
4158 	sfmmu_cb_table[id].capture_cpus = capture_cpus;
4159 
4160 	return (id);
4161 }
4162 
4163 #define	HAC_COOKIE_NONE	(void *)-1
4164 
4165 /*
4166  * Add relocation callbacks to the specified addr/len which will be called
4167  * when relocating the associated page. See the description of pre and
4168  * posthandler above for more details.
4169  *
4170  * If HAC_PAGELOCK is included in flags, the underlying memory page is
4171  * locked internally so the caller must be able to deal with the callback
4172  * running even before this function has returned.  If HAC_PAGELOCK is not
4173  * set, it is assumed that the underlying memory pages are locked.
4174  *
4175  * Since the caller must track the individual page boundaries anyway,
4176  * we only allow a callback to be added to a single page (large
4177  * or small).  Thus [addr, addr + len) MUST be contained within a single
4178  * page.
4179  *
4180  * Registering multiple callbacks on the same [addr, addr+len) is supported,
4181  * _provided_that_ a unique parameter is specified for each callback.
4182  * If multiple callbacks are registered on the same range the callback will
4183  * be invoked with each unique parameter. Registering the same callback with
4184  * the same argument more than once will result in corrupted kernel state.
4185  *
4186  * Returns the pfn of the underlying kernel page in *rpfn
4187  * on success, or PFN_INVALID on failure.
4188  *
4189  * cookiep (if passed) provides storage space for an opaque cookie
4190  * to return later to hat_delete_callback(). This cookie makes the callback
4191  * deletion significantly quicker by avoiding a potentially lengthy hash
4192  * search.
4193  *
4194  * Returns values:
4195  *    0:      success
4196  *    ENOMEM: memory allocation failure (e.g. flags was passed as HAC_NOSLEEP)
4197  *    EINVAL: callback ID is not valid
4198  *    ENXIO:  ["vaddr", "vaddr" + len) is not mapped in the kernel's address
4199  *            space
4200  *    ERANGE: ["vaddr", "vaddr" + len) crosses a page boundary
4201  */
4202 int
4203 hat_add_callback(id_t callback_id, caddr_t vaddr, uint_t len, uint_t flags,
4204 	void *pvt, pfn_t *rpfn, void **cookiep)
4205 {
4206 	struct 		hmehash_bucket *hmebp;
4207 	hmeblk_tag 	hblktag;
4208 	struct hme_blk	*hmeblkp;
4209 	int 		hmeshift, hashno;
4210 	caddr_t 	saddr, eaddr, baseaddr;
4211 	struct pa_hment *pahmep;
4212 	struct sf_hment *sfhmep, *osfhmep;
4213 	kmutex_t	*pml;
4214 	tte_t   	tte;
4215 	page_t		*pp;
4216 	vnode_t		*vp;
4217 	u_offset_t	off;
4218 	pfn_t		pfn;
4219 	int		kmflags = (flags & HAC_SLEEP)? KM_SLEEP : KM_NOSLEEP;
4220 	int		locked = 0;
4221 
4222 	/*
4223 	 * For KPM mappings, just return the physical address since we
4224 	 * don't need to register any callbacks.
4225 	 */
4226 	if (IS_KPM_ADDR(vaddr)) {
4227 		uint64_t paddr;
4228 		SFMMU_KPM_VTOP(vaddr, paddr);
4229 		*rpfn = btop(paddr);
4230 		if (cookiep != NULL)
4231 			*cookiep = HAC_COOKIE_NONE;
4232 		return (0);
4233 	}
4234 
4235 	if (callback_id < (id_t)0 || callback_id >= sfmmu_cb_nextid) {
4236 		*rpfn = PFN_INVALID;
4237 		return (EINVAL);
4238 	}
4239 
4240 	if ((pahmep = kmem_cache_alloc(pa_hment_cache, kmflags)) == NULL) {
4241 		*rpfn = PFN_INVALID;
4242 		return (ENOMEM);
4243 	}
4244 
4245 	sfhmep = &pahmep->sfment;
4246 
4247 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4248 	eaddr = saddr + len;
4249 
4250 rehash:
4251 	/* Find the mapping(s) for this page */
4252 	for (hashno = TTE64K, hmeblkp = NULL;
4253 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4254 	    hashno++) {
4255 		hmeshift = HME_HASH_SHIFT(hashno);
4256 		hblktag.htag_id = ksfmmup;
4257 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4258 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4259 		hblktag.htag_rehash = hashno;
4260 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4261 
4262 		SFMMU_HASH_LOCK(hmebp);
4263 
4264 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4265 
4266 		if (hmeblkp == NULL)
4267 			SFMMU_HASH_UNLOCK(hmebp);
4268 	}
4269 
4270 	if (hmeblkp == NULL) {
4271 		kmem_cache_free(pa_hment_cache, pahmep);
4272 		*rpfn = PFN_INVALID;
4273 		return (ENXIO);
4274 	}
4275 
4276 	ASSERT(!hmeblkp->hblk_shared);
4277 
4278 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4279 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4280 
4281 	if (!TTE_IS_VALID(&tte)) {
4282 		SFMMU_HASH_UNLOCK(hmebp);
4283 		kmem_cache_free(pa_hment_cache, pahmep);
4284 		*rpfn = PFN_INVALID;
4285 		return (ENXIO);
4286 	}
4287 
4288 	/*
4289 	 * Make sure the boundaries for the callback fall within this
4290 	 * single mapping.
4291 	 */
4292 	baseaddr = (caddr_t)get_hblk_base(hmeblkp);
4293 	ASSERT(saddr >= baseaddr);
4294 	if (eaddr > saddr + TTEBYTES(TTE_CSZ(&tte))) {
4295 		SFMMU_HASH_UNLOCK(hmebp);
4296 		kmem_cache_free(pa_hment_cache, pahmep);
4297 		*rpfn = PFN_INVALID;
4298 		return (ERANGE);
4299 	}
4300 
4301 	pfn = sfmmu_ttetopfn(&tte, vaddr);
4302 
4303 	/*
4304 	 * The pfn may not have a page_t underneath in which case we
4305 	 * just return it. This can happen if we are doing I/O to a
4306 	 * static portion of the kernel's address space, for instance.
4307 	 */
4308 	pp = osfhmep->hme_page;
4309 	if (pp == NULL) {
4310 		SFMMU_HASH_UNLOCK(hmebp);
4311 		kmem_cache_free(pa_hment_cache, pahmep);
4312 		*rpfn = pfn;
4313 		if (cookiep)
4314 			*cookiep = HAC_COOKIE_NONE;
4315 		return (0);
4316 	}
4317 	ASSERT(pp == PP_PAGEROOT(pp));
4318 
4319 	vp = pp->p_vnode;
4320 	off = pp->p_offset;
4321 
4322 	pml = sfmmu_mlist_enter(pp);
4323 
4324 	if (flags & HAC_PAGELOCK) {
4325 		if (!page_trylock(pp, SE_SHARED)) {
4326 			/*
4327 			 * Somebody is holding SE_EXCL lock. Might
4328 			 * even be hat_page_relocate(). Drop all
4329 			 * our locks, lookup the page in &kvp, and
4330 			 * retry. If it doesn't exist in &kvp and &zvp,
4331 			 * then we must be dealing with a kernel mapped
4332 			 * page which doesn't actually belong to
4333 			 * segkmem so we punt.
4334 			 */
4335 			sfmmu_mlist_exit(pml);
4336 			SFMMU_HASH_UNLOCK(hmebp);
4337 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4338 
4339 			/* check zvp before giving up */
4340 			if (pp == NULL)
4341 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4342 				    SE_SHARED);
4343 
4344 			/* Okay, we didn't find it, give up */
4345 			if (pp == NULL) {
4346 				kmem_cache_free(pa_hment_cache, pahmep);
4347 				*rpfn = pfn;
4348 				if (cookiep)
4349 					*cookiep = HAC_COOKIE_NONE;
4350 				return (0);
4351 			}
4352 			page_unlock(pp);
4353 			goto rehash;
4354 		}
4355 		locked = 1;
4356 	}
4357 
4358 	if (!PAGE_LOCKED(pp) && !panicstr)
4359 		panic("hat_add_callback: page 0x%p not locked", pp);
4360 
4361 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4362 	    pp->p_offset != off) {
4363 		/*
4364 		 * The page moved before we got our hands on it.  Drop
4365 		 * all the locks and try again.
4366 		 */
4367 		ASSERT((flags & HAC_PAGELOCK) != 0);
4368 		sfmmu_mlist_exit(pml);
4369 		SFMMU_HASH_UNLOCK(hmebp);
4370 		page_unlock(pp);
4371 		locked = 0;
4372 		goto rehash;
4373 	}
4374 
4375 	if (!VN_ISKAS(vp)) {
4376 		/*
4377 		 * This is not a segkmem page but another page which
4378 		 * has been kernel mapped. It had better have at least
4379 		 * a share lock on it. Return the pfn.
4380 		 */
4381 		sfmmu_mlist_exit(pml);
4382 		SFMMU_HASH_UNLOCK(hmebp);
4383 		if (locked)
4384 			page_unlock(pp);
4385 		kmem_cache_free(pa_hment_cache, pahmep);
4386 		ASSERT(PAGE_LOCKED(pp));
4387 		*rpfn = pfn;
4388 		if (cookiep)
4389 			*cookiep = HAC_COOKIE_NONE;
4390 		return (0);
4391 	}
4392 
4393 	/*
4394 	 * Setup this pa_hment and link its embedded dummy sf_hment into
4395 	 * the mapping list.
4396 	 */
4397 	pp->p_share++;
4398 	pahmep->cb_id = callback_id;
4399 	pahmep->addr = vaddr;
4400 	pahmep->len = len;
4401 	pahmep->refcnt = 1;
4402 	pahmep->flags = 0;
4403 	pahmep->pvt = pvt;
4404 
4405 	sfhmep->hme_tte.ll = 0;
4406 	sfhmep->hme_data = pahmep;
4407 	sfhmep->hme_prev = osfhmep;
4408 	sfhmep->hme_next = osfhmep->hme_next;
4409 
4410 	if (osfhmep->hme_next)
4411 		osfhmep->hme_next->hme_prev = sfhmep;
4412 
4413 	osfhmep->hme_next = sfhmep;
4414 
4415 	sfmmu_mlist_exit(pml);
4416 	SFMMU_HASH_UNLOCK(hmebp);
4417 
4418 	if (locked)
4419 		page_unlock(pp);
4420 
4421 	*rpfn = pfn;
4422 	if (cookiep)
4423 		*cookiep = (void *)pahmep;
4424 
4425 	return (0);
4426 }
4427 
4428 /*
4429  * Remove the relocation callbacks from the specified addr/len.
4430  */
4431 void
4432 hat_delete_callback(caddr_t vaddr, uint_t len, void *pvt, uint_t flags,
4433 	void *cookie)
4434 {
4435 	struct		hmehash_bucket *hmebp;
4436 	hmeblk_tag	hblktag;
4437 	struct hme_blk	*hmeblkp;
4438 	int		hmeshift, hashno;
4439 	caddr_t		saddr;
4440 	struct pa_hment	*pahmep;
4441 	struct sf_hment	*sfhmep, *osfhmep;
4442 	kmutex_t	*pml;
4443 	tte_t		tte;
4444 	page_t		*pp;
4445 	vnode_t		*vp;
4446 	u_offset_t	off;
4447 	int		locked = 0;
4448 
4449 	/*
4450 	 * If the cookie is HAC_COOKIE_NONE then there is no pa_hment to
4451 	 * remove so just return.
4452 	 */
4453 	if (cookie == HAC_COOKIE_NONE || IS_KPM_ADDR(vaddr))
4454 		return;
4455 
4456 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
4457 
4458 rehash:
4459 	/* Find the mapping(s) for this page */
4460 	for (hashno = TTE64K, hmeblkp = NULL;
4461 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
4462 	    hashno++) {
4463 		hmeshift = HME_HASH_SHIFT(hashno);
4464 		hblktag.htag_id = ksfmmup;
4465 		hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4466 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
4467 		hblktag.htag_rehash = hashno;
4468 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
4469 
4470 		SFMMU_HASH_LOCK(hmebp);
4471 
4472 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
4473 
4474 		if (hmeblkp == NULL)
4475 			SFMMU_HASH_UNLOCK(hmebp);
4476 	}
4477 
4478 	if (hmeblkp == NULL)
4479 		return;
4480 
4481 	ASSERT(!hmeblkp->hblk_shared);
4482 
4483 	HBLKTOHME(osfhmep, hmeblkp, saddr);
4484 
4485 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
4486 	if (!TTE_IS_VALID(&tte)) {
4487 		SFMMU_HASH_UNLOCK(hmebp);
4488 		return;
4489 	}
4490 
4491 	pp = osfhmep->hme_page;
4492 	if (pp == NULL) {
4493 		SFMMU_HASH_UNLOCK(hmebp);
4494 		ASSERT(cookie == NULL);
4495 		return;
4496 	}
4497 
4498 	vp = pp->p_vnode;
4499 	off = pp->p_offset;
4500 
4501 	pml = sfmmu_mlist_enter(pp);
4502 
4503 	if (flags & HAC_PAGELOCK) {
4504 		if (!page_trylock(pp, SE_SHARED)) {
4505 			/*
4506 			 * Somebody is holding SE_EXCL lock. Might
4507 			 * even be hat_page_relocate(). Drop all
4508 			 * our locks, lookup the page in &kvp, and
4509 			 * retry. If it doesn't exist in &kvp and &zvp,
4510 			 * then we must be dealing with a kernel mapped
4511 			 * page which doesn't actually belong to
4512 			 * segkmem so we punt.
4513 			 */
4514 			sfmmu_mlist_exit(pml);
4515 			SFMMU_HASH_UNLOCK(hmebp);
4516 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
4517 			/* check zvp before giving up */
4518 			if (pp == NULL)
4519 				pp = page_lookup(&zvp, (u_offset_t)saddr,
4520 				    SE_SHARED);
4521 
4522 			if (pp == NULL) {
4523 				ASSERT(cookie == NULL);
4524 				return;
4525 			}
4526 			page_unlock(pp);
4527 			goto rehash;
4528 		}
4529 		locked = 1;
4530 	}
4531 
4532 	ASSERT(PAGE_LOCKED(pp));
4533 
4534 	if (osfhmep->hme_page != pp || pp->p_vnode != vp ||
4535 	    pp->p_offset != off) {
4536 		/*
4537 		 * The page moved before we got our hands on it.  Drop
4538 		 * all the locks and try again.
4539 		 */
4540 		ASSERT((flags & HAC_PAGELOCK) != 0);
4541 		sfmmu_mlist_exit(pml);
4542 		SFMMU_HASH_UNLOCK(hmebp);
4543 		page_unlock(pp);
4544 		locked = 0;
4545 		goto rehash;
4546 	}
4547 
4548 	if (!VN_ISKAS(vp)) {
4549 		/*
4550 		 * This is not a segkmem page but another page which
4551 		 * has been kernel mapped.
4552 		 */
4553 		sfmmu_mlist_exit(pml);
4554 		SFMMU_HASH_UNLOCK(hmebp);
4555 		if (locked)
4556 			page_unlock(pp);
4557 		ASSERT(cookie == NULL);
4558 		return;
4559 	}
4560 
4561 	if (cookie != NULL) {
4562 		pahmep = (struct pa_hment *)cookie;
4563 		sfhmep = &pahmep->sfment;
4564 	} else {
4565 		for (sfhmep = pp->p_mapping; sfhmep != NULL;
4566 		    sfhmep = sfhmep->hme_next) {
4567 
4568 			/*
4569 			 * skip va<->pa mappings
4570 			 */
4571 			if (!IS_PAHME(sfhmep))
4572 				continue;
4573 
4574 			pahmep = sfhmep->hme_data;
4575 			ASSERT(pahmep != NULL);
4576 
4577 			/*
4578 			 * if pa_hment matches, remove it
4579 			 */
4580 			if ((pahmep->pvt == pvt) &&
4581 			    (pahmep->addr == vaddr) &&
4582 			    (pahmep->len == len)) {
4583 				break;
4584 			}
4585 		}
4586 	}
4587 
4588 	if (sfhmep == NULL) {
4589 		if (!panicstr) {
4590 			panic("hat_delete_callback: pa_hment not found, pp %p",
4591 			    (void *)pp);
4592 		}
4593 		return;
4594 	}
4595 
4596 	/*
4597 	 * Note: at this point a valid kernel mapping must still be
4598 	 * present on this page.
4599 	 */
4600 	pp->p_share--;
4601 	if (pp->p_share <= 0)
4602 		panic("hat_delete_callback: zero p_share");
4603 
4604 	if (--pahmep->refcnt == 0) {
4605 		if (pahmep->flags != 0)
4606 			panic("hat_delete_callback: pa_hment is busy");
4607 
4608 		/*
4609 		 * Remove sfhmep from the mapping list for the page.
4610 		 */
4611 		if (sfhmep->hme_prev) {
4612 			sfhmep->hme_prev->hme_next = sfhmep->hme_next;
4613 		} else {
4614 			pp->p_mapping = sfhmep->hme_next;
4615 		}
4616 
4617 		if (sfhmep->hme_next)
4618 			sfhmep->hme_next->hme_prev = sfhmep->hme_prev;
4619 
4620 		sfmmu_mlist_exit(pml);
4621 		SFMMU_HASH_UNLOCK(hmebp);
4622 
4623 		if (locked)
4624 			page_unlock(pp);
4625 
4626 		kmem_cache_free(pa_hment_cache, pahmep);
4627 		return;
4628 	}
4629 
4630 	sfmmu_mlist_exit(pml);
4631 	SFMMU_HASH_UNLOCK(hmebp);
4632 	if (locked)
4633 		page_unlock(pp);
4634 }
4635 
4636 /*
4637  * hat_probe returns 1 if the translation for the address 'addr' is
4638  * loaded, zero otherwise.
4639  *
4640  * hat_probe should be used only for advisorary purposes because it may
4641  * occasionally return the wrong value. The implementation must guarantee that
4642  * returning the wrong value is a very rare event. hat_probe is used
4643  * to implement optimizations in the segment drivers.
4644  *
4645  */
4646 int
4647 hat_probe(struct hat *sfmmup, caddr_t addr)
4648 {
4649 	pfn_t pfn;
4650 	tte_t tte;
4651 
4652 	ASSERT(sfmmup != NULL);
4653 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4654 
4655 	ASSERT((sfmmup == ksfmmup) ||
4656 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
4657 
4658 	if (sfmmup == ksfmmup) {
4659 		while ((pfn = sfmmu_vatopfn(addr, sfmmup, &tte))
4660 		    == PFN_SUSPENDED) {
4661 			sfmmu_vatopfn_suspended(addr, sfmmup, &tte);
4662 		}
4663 	} else {
4664 		pfn = sfmmu_uvatopfn(addr, sfmmup, NULL);
4665 	}
4666 
4667 	if (pfn != PFN_INVALID)
4668 		return (1);
4669 	else
4670 		return (0);
4671 }
4672 
4673 ssize_t
4674 hat_getpagesize(struct hat *sfmmup, caddr_t addr)
4675 {
4676 	tte_t tte;
4677 
4678 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4679 
4680 	if (sfmmup == ksfmmup) {
4681 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4682 			return (-1);
4683 		}
4684 	} else {
4685 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4686 			return (-1);
4687 		}
4688 	}
4689 
4690 	ASSERT(TTE_IS_VALID(&tte));
4691 	return (TTEBYTES(TTE_CSZ(&tte)));
4692 }
4693 
4694 uint_t
4695 hat_getattr(struct hat *sfmmup, caddr_t addr, uint_t *attr)
4696 {
4697 	tte_t tte;
4698 
4699 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
4700 
4701 	if (sfmmup == ksfmmup) {
4702 		if (sfmmu_vatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4703 			tte.ll = 0;
4704 		}
4705 	} else {
4706 		if (sfmmu_uvatopfn(addr, sfmmup, &tte) == PFN_INVALID) {
4707 			tte.ll = 0;
4708 		}
4709 	}
4710 	if (TTE_IS_VALID(&tte)) {
4711 		*attr = sfmmu_ptov_attr(&tte);
4712 		return (0);
4713 	}
4714 	*attr = 0;
4715 	return ((uint_t)0xffffffff);
4716 }
4717 
4718 /*
4719  * Enables more attributes on specified address range (ie. logical OR)
4720  */
4721 void
4722 hat_setattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4723 {
4724 	if (hat->sfmmu_xhat_provider) {
4725 		XHAT_SETATTR(hat, addr, len, attr);
4726 		return;
4727 	} else {
4728 		/*
4729 		 * This must be a CPU HAT. If the address space has
4730 		 * XHATs attached, change attributes for all of them,
4731 		 * just in case
4732 		 */
4733 		ASSERT(hat->sfmmu_as != NULL);
4734 		if (hat->sfmmu_as->a_xhat != NULL)
4735 			xhat_setattr_all(hat->sfmmu_as, addr, len, attr);
4736 	}
4737 
4738 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_SETATTR);
4739 }
4740 
4741 /*
4742  * Assigns attributes to the specified address range.  All the attributes
4743  * are specified.
4744  */
4745 void
4746 hat_chgattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4747 {
4748 	if (hat->sfmmu_xhat_provider) {
4749 		XHAT_CHGATTR(hat, addr, len, attr);
4750 		return;
4751 	} else {
4752 		/*
4753 		 * This must be a CPU HAT. If the address space has
4754 		 * XHATs attached, change attributes for all of them,
4755 		 * just in case
4756 		 */
4757 		ASSERT(hat->sfmmu_as != NULL);
4758 		if (hat->sfmmu_as->a_xhat != NULL)
4759 			xhat_chgattr_all(hat->sfmmu_as, addr, len, attr);
4760 	}
4761 
4762 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CHGATTR);
4763 }
4764 
4765 /*
4766  * Remove attributes on the specified address range (ie. loginal NAND)
4767  */
4768 void
4769 hat_clrattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
4770 {
4771 	if (hat->sfmmu_xhat_provider) {
4772 		XHAT_CLRATTR(hat, addr, len, attr);
4773 		return;
4774 	} else {
4775 		/*
4776 		 * This must be a CPU HAT. If the address space has
4777 		 * XHATs attached, change attributes for all of them,
4778 		 * just in case
4779 		 */
4780 		ASSERT(hat->sfmmu_as != NULL);
4781 		if (hat->sfmmu_as->a_xhat != NULL)
4782 			xhat_clrattr_all(hat->sfmmu_as, addr, len, attr);
4783 	}
4784 
4785 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CLRATTR);
4786 }
4787 
4788 /*
4789  * Change attributes on an address range to that specified by attr and mode.
4790  */
4791 static void
4792 sfmmu_chgattr(struct hat *sfmmup, caddr_t addr, size_t len, uint_t attr,
4793 	int mode)
4794 {
4795 	struct hmehash_bucket *hmebp;
4796 	hmeblk_tag hblktag;
4797 	int hmeshift, hashno = 1;
4798 	struct hme_blk *hmeblkp, *list = NULL;
4799 	caddr_t endaddr;
4800 	cpuset_t cpuset;
4801 	demap_range_t dmr;
4802 
4803 	CPUSET_ZERO(cpuset);
4804 
4805 	ASSERT((sfmmup == ksfmmup) ||
4806 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
4807 	ASSERT((len & MMU_PAGEOFFSET) == 0);
4808 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
4809 
4810 	if ((attr & PROT_USER) && (mode != SFMMU_CLRATTR) &&
4811 	    ((addr + len) > (caddr_t)USERLIMIT)) {
4812 		panic("user addr %p in kernel space",
4813 		    (void *)addr);
4814 	}
4815 
4816 	endaddr = addr + len;
4817 	hblktag.htag_id = sfmmup;
4818 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
4819 	DEMAP_RANGE_INIT(sfmmup, &dmr);
4820 
4821 	while (addr < endaddr) {
4822 		hmeshift = HME_HASH_SHIFT(hashno);
4823 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
4824 		hblktag.htag_rehash = hashno;
4825 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
4826 
4827 		SFMMU_HASH_LOCK(hmebp);
4828 
4829 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
4830 		if (hmeblkp != NULL) {
4831 			ASSERT(!hmeblkp->hblk_shared);
4832 			/*
4833 			 * We've encountered a shadow hmeblk so skip the range
4834 			 * of the next smaller mapping size.
4835 			 */
4836 			if (hmeblkp->hblk_shw_bit) {
4837 				ASSERT(sfmmup != ksfmmup);
4838 				ASSERT(hashno > 1);
4839 				addr = (caddr_t)P2END((uintptr_t)addr,
4840 				    TTEBYTES(hashno - 1));
4841 			} else {
4842 				addr = sfmmu_hblk_chgattr(sfmmup,
4843 				    hmeblkp, addr, endaddr, &dmr, attr, mode);
4844 			}
4845 			SFMMU_HASH_UNLOCK(hmebp);
4846 			hashno = 1;
4847 			continue;
4848 		}
4849 		SFMMU_HASH_UNLOCK(hmebp);
4850 
4851 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
4852 			/*
4853 			 * We have traversed the whole list and rehashed
4854 			 * if necessary without finding the address to chgattr.
4855 			 * This is ok, so we increment the address by the
4856 			 * smallest hmeblk range for kernel mappings or for
4857 			 * user mappings with no large pages, and the largest
4858 			 * hmeblk range, to account for shadow hmeblks, for
4859 			 * user mappings with large pages and continue.
4860 			 */
4861 			if (sfmmup == ksfmmup)
4862 				addr = (caddr_t)P2END((uintptr_t)addr,
4863 				    TTEBYTES(1));
4864 			else
4865 				addr = (caddr_t)P2END((uintptr_t)addr,
4866 				    TTEBYTES(hashno));
4867 			hashno = 1;
4868 		} else {
4869 			hashno++;
4870 		}
4871 	}
4872 
4873 	sfmmu_hblks_list_purge(&list);
4874 	DEMAP_RANGE_FLUSH(&dmr);
4875 	cpuset = sfmmup->sfmmu_cpusran;
4876 	xt_sync(cpuset);
4877 }
4878 
4879 /*
4880  * This function chgattr on a range of addresses in an hmeblk.  It returns the
4881  * next addres that needs to be chgattr.
4882  * It should be called with the hash lock held.
4883  * XXX It should be possible to optimize chgattr by not flushing every time but
4884  * on the other hand:
4885  * 1. do one flush crosscall.
4886  * 2. only flush if we are increasing permissions (make sure this will work)
4887  */
4888 static caddr_t
4889 sfmmu_hblk_chgattr(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
4890 	caddr_t endaddr, demap_range_t *dmrp, uint_t attr, int mode)
4891 {
4892 	tte_t tte, tteattr, tteflags, ttemod;
4893 	struct sf_hment *sfhmep;
4894 	int ttesz;
4895 	struct page *pp = NULL;
4896 	kmutex_t *pml, *pmtx;
4897 	int ret;
4898 	int use_demap_range;
4899 #if defined(SF_ERRATA_57)
4900 	int check_exec;
4901 #endif
4902 
4903 	ASSERT(in_hblk_range(hmeblkp, addr));
4904 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4905 	ASSERT(!hmeblkp->hblk_shared);
4906 
4907 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4908 	ttesz = get_hblk_ttesz(hmeblkp);
4909 
4910 	/*
4911 	 * Flush the current demap region if addresses have been
4912 	 * skipped or the page size doesn't match.
4913 	 */
4914 	use_demap_range = (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp));
4915 	if (use_demap_range) {
4916 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
4917 	} else {
4918 		DEMAP_RANGE_FLUSH(dmrp);
4919 	}
4920 
4921 	tteattr.ll = sfmmu_vtop_attr(attr, mode, &tteflags);
4922 #if defined(SF_ERRATA_57)
4923 	check_exec = (sfmmup != ksfmmup) &&
4924 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
4925 	    TTE_IS_EXECUTABLE(&tteattr);
4926 #endif
4927 	HBLKTOHME(sfhmep, hmeblkp, addr);
4928 	while (addr < endaddr) {
4929 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
4930 		if (TTE_IS_VALID(&tte)) {
4931 			if ((tte.ll & tteflags.ll) == tteattr.ll) {
4932 				/*
4933 				 * if the new attr is the same as old
4934 				 * continue
4935 				 */
4936 				goto next_addr;
4937 			}
4938 			if (!TTE_IS_WRITABLE(&tteattr)) {
4939 				/*
4940 				 * make sure we clear hw modify bit if we
4941 				 * removing write protections
4942 				 */
4943 				tteflags.tte_intlo |= TTE_HWWR_INT;
4944 			}
4945 
4946 			pml = NULL;
4947 			pp = sfhmep->hme_page;
4948 			if (pp) {
4949 				pml = sfmmu_mlist_enter(pp);
4950 			}
4951 
4952 			if (pp != sfhmep->hme_page) {
4953 				/*
4954 				 * tte must have been unloaded.
4955 				 */
4956 				ASSERT(pml);
4957 				sfmmu_mlist_exit(pml);
4958 				continue;
4959 			}
4960 
4961 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
4962 
4963 			ttemod = tte;
4964 			ttemod.ll = (ttemod.ll & ~tteflags.ll) | tteattr.ll;
4965 			ASSERT(TTE_TO_TTEPFN(&ttemod) == TTE_TO_TTEPFN(&tte));
4966 
4967 #if defined(SF_ERRATA_57)
4968 			if (check_exec && addr < errata57_limit)
4969 				ttemod.tte_exec_perm = 0;
4970 #endif
4971 			ret = sfmmu_modifytte_try(&tte, &ttemod,
4972 			    &sfhmep->hme_tte);
4973 
4974 			if (ret < 0) {
4975 				/* tte changed underneath us */
4976 				if (pml) {
4977 					sfmmu_mlist_exit(pml);
4978 				}
4979 				continue;
4980 			}
4981 
4982 			if (tteflags.tte_intlo & TTE_HWWR_INT) {
4983 				/*
4984 				 * need to sync if we are clearing modify bit.
4985 				 */
4986 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
4987 			}
4988 
4989 			if (pp && PP_ISRO(pp)) {
4990 				if (tteattr.tte_intlo & TTE_WRPRM_INT) {
4991 					pmtx = sfmmu_page_enter(pp);
4992 					PP_CLRRO(pp);
4993 					sfmmu_page_exit(pmtx);
4994 				}
4995 			}
4996 
4997 			if (ret > 0 && use_demap_range) {
4998 				DEMAP_RANGE_MARKPG(dmrp, addr);
4999 			} else if (ret > 0) {
5000 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5001 			}
5002 
5003 			if (pml) {
5004 				sfmmu_mlist_exit(pml);
5005 			}
5006 		}
5007 next_addr:
5008 		addr += TTEBYTES(ttesz);
5009 		sfhmep++;
5010 		DEMAP_RANGE_NEXTPG(dmrp);
5011 	}
5012 	return (addr);
5013 }
5014 
5015 /*
5016  * This routine converts virtual attributes to physical ones.  It will
5017  * update the tteflags field with the tte mask corresponding to the attributes
5018  * affected and it returns the new attributes.  It will also clear the modify
5019  * bit if we are taking away write permission.  This is necessary since the
5020  * modify bit is the hardware permission bit and we need to clear it in order
5021  * to detect write faults.
5022  */
5023 static uint64_t
5024 sfmmu_vtop_attr(uint_t attr, int mode, tte_t *ttemaskp)
5025 {
5026 	tte_t ttevalue;
5027 
5028 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
5029 
5030 	switch (mode) {
5031 	case SFMMU_CHGATTR:
5032 		/* all attributes specified */
5033 		ttevalue.tte_inthi = MAKE_TTEATTR_INTHI(attr);
5034 		ttevalue.tte_intlo = MAKE_TTEATTR_INTLO(attr);
5035 		ttemaskp->tte_inthi = TTEINTHI_ATTR;
5036 		ttemaskp->tte_intlo = TTEINTLO_ATTR;
5037 		break;
5038 	case SFMMU_SETATTR:
5039 		ASSERT(!(attr & ~HAT_PROT_MASK));
5040 		ttemaskp->ll = 0;
5041 		ttevalue.ll = 0;
5042 		/*
5043 		 * a valid tte implies exec and read for sfmmu
5044 		 * so no need to do anything about them.
5045 		 * since priviledged access implies user access
5046 		 * PROT_USER doesn't make sense either.
5047 		 */
5048 		if (attr & PROT_WRITE) {
5049 			ttemaskp->tte_intlo |= TTE_WRPRM_INT;
5050 			ttevalue.tte_intlo |= TTE_WRPRM_INT;
5051 		}
5052 		break;
5053 	case SFMMU_CLRATTR:
5054 		/* attributes will be nand with current ones */
5055 		if (attr & ~(PROT_WRITE | PROT_USER)) {
5056 			panic("sfmmu: attr %x not supported", attr);
5057 		}
5058 		ttemaskp->ll = 0;
5059 		ttevalue.ll = 0;
5060 		if (attr & PROT_WRITE) {
5061 			/* clear both writable and modify bit */
5062 			ttemaskp->tte_intlo |= TTE_WRPRM_INT | TTE_HWWR_INT;
5063 		}
5064 		if (attr & PROT_USER) {
5065 			ttemaskp->tte_intlo |= TTE_PRIV_INT;
5066 			ttevalue.tte_intlo |= TTE_PRIV_INT;
5067 		}
5068 		break;
5069 	default:
5070 		panic("sfmmu_vtop_attr: bad mode %x", mode);
5071 	}
5072 	ASSERT(TTE_TO_TTEPFN(&ttevalue) == 0);
5073 	return (ttevalue.ll);
5074 }
5075 
5076 static uint_t
5077 sfmmu_ptov_attr(tte_t *ttep)
5078 {
5079 	uint_t attr;
5080 
5081 	ASSERT(TTE_IS_VALID(ttep));
5082 
5083 	attr = PROT_READ;
5084 
5085 	if (TTE_IS_WRITABLE(ttep)) {
5086 		attr |= PROT_WRITE;
5087 	}
5088 	if (TTE_IS_EXECUTABLE(ttep)) {
5089 		attr |= PROT_EXEC;
5090 	}
5091 	if (!TTE_IS_PRIVILEGED(ttep)) {
5092 		attr |= PROT_USER;
5093 	}
5094 	if (TTE_IS_NFO(ttep)) {
5095 		attr |= HAT_NOFAULT;
5096 	}
5097 	if (TTE_IS_NOSYNC(ttep)) {
5098 		attr |= HAT_NOSYNC;
5099 	}
5100 	if (TTE_IS_SIDEFFECT(ttep)) {
5101 		attr |= SFMMU_SIDEFFECT;
5102 	}
5103 	if (!TTE_IS_VCACHEABLE(ttep)) {
5104 		attr |= SFMMU_UNCACHEVTTE;
5105 	}
5106 	if (!TTE_IS_PCACHEABLE(ttep)) {
5107 		attr |= SFMMU_UNCACHEPTTE;
5108 	}
5109 	return (attr);
5110 }
5111 
5112 /*
5113  * hat_chgprot is a deprecated hat call.  New segment drivers
5114  * should store all attributes and use hat_*attr calls.
5115  *
5116  * Change the protections in the virtual address range
5117  * given to the specified virtual protection.  If vprot is ~PROT_WRITE,
5118  * then remove write permission, leaving the other
5119  * permissions unchanged.  If vprot is ~PROT_USER, remove user permissions.
5120  *
5121  */
5122 void
5123 hat_chgprot(struct hat *sfmmup, caddr_t addr, size_t len, uint_t vprot)
5124 {
5125 	struct hmehash_bucket *hmebp;
5126 	hmeblk_tag hblktag;
5127 	int hmeshift, hashno = 1;
5128 	struct hme_blk *hmeblkp, *list = NULL;
5129 	caddr_t endaddr;
5130 	cpuset_t cpuset;
5131 	demap_range_t dmr;
5132 
5133 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5134 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
5135 
5136 	if (sfmmup->sfmmu_xhat_provider) {
5137 		XHAT_CHGPROT(sfmmup, addr, len, vprot);
5138 		return;
5139 	} else {
5140 		/*
5141 		 * This must be a CPU HAT. If the address space has
5142 		 * XHATs attached, change attributes for all of them,
5143 		 * just in case
5144 		 */
5145 		ASSERT(sfmmup->sfmmu_as != NULL);
5146 		if (sfmmup->sfmmu_as->a_xhat != NULL)
5147 			xhat_chgprot_all(sfmmup->sfmmu_as, addr, len, vprot);
5148 	}
5149 
5150 	CPUSET_ZERO(cpuset);
5151 
5152 	if ((vprot != (uint_t)~PROT_WRITE) && (vprot & PROT_USER) &&
5153 	    ((addr + len) > (caddr_t)USERLIMIT)) {
5154 		panic("user addr %p vprot %x in kernel space",
5155 		    (void *)addr, vprot);
5156 	}
5157 	endaddr = addr + len;
5158 	hblktag.htag_id = sfmmup;
5159 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5160 	DEMAP_RANGE_INIT(sfmmup, &dmr);
5161 
5162 	while (addr < endaddr) {
5163 		hmeshift = HME_HASH_SHIFT(hashno);
5164 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5165 		hblktag.htag_rehash = hashno;
5166 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5167 
5168 		SFMMU_HASH_LOCK(hmebp);
5169 
5170 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
5171 		if (hmeblkp != NULL) {
5172 			ASSERT(!hmeblkp->hblk_shared);
5173 			/*
5174 			 * We've encountered a shadow hmeblk so skip the range
5175 			 * of the next smaller mapping size.
5176 			 */
5177 			if (hmeblkp->hblk_shw_bit) {
5178 				ASSERT(sfmmup != ksfmmup);
5179 				ASSERT(hashno > 1);
5180 				addr = (caddr_t)P2END((uintptr_t)addr,
5181 				    TTEBYTES(hashno - 1));
5182 			} else {
5183 				addr = sfmmu_hblk_chgprot(sfmmup, hmeblkp,
5184 				    addr, endaddr, &dmr, vprot);
5185 			}
5186 			SFMMU_HASH_UNLOCK(hmebp);
5187 			hashno = 1;
5188 			continue;
5189 		}
5190 		SFMMU_HASH_UNLOCK(hmebp);
5191 
5192 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
5193 			/*
5194 			 * We have traversed the whole list and rehashed
5195 			 * if necessary without finding the address to chgprot.
5196 			 * This is ok so we increment the address by the
5197 			 * smallest hmeblk range for kernel mappings and the
5198 			 * largest hmeblk range, to account for shadow hmeblks,
5199 			 * for user mappings and continue.
5200 			 */
5201 			if (sfmmup == ksfmmup)
5202 				addr = (caddr_t)P2END((uintptr_t)addr,
5203 				    TTEBYTES(1));
5204 			else
5205 				addr = (caddr_t)P2END((uintptr_t)addr,
5206 				    TTEBYTES(hashno));
5207 			hashno = 1;
5208 		} else {
5209 			hashno++;
5210 		}
5211 	}
5212 
5213 	sfmmu_hblks_list_purge(&list);
5214 	DEMAP_RANGE_FLUSH(&dmr);
5215 	cpuset = sfmmup->sfmmu_cpusran;
5216 	xt_sync(cpuset);
5217 }
5218 
5219 /*
5220  * This function chgprots a range of addresses in an hmeblk.  It returns the
5221  * next addres that needs to be chgprot.
5222  * It should be called with the hash lock held.
5223  * XXX It shold be possible to optimize chgprot by not flushing every time but
5224  * on the other hand:
5225  * 1. do one flush crosscall.
5226  * 2. only flush if we are increasing permissions (make sure this will work)
5227  */
5228 static caddr_t
5229 sfmmu_hblk_chgprot(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5230 	caddr_t endaddr, demap_range_t *dmrp, uint_t vprot)
5231 {
5232 	uint_t pprot;
5233 	tte_t tte, ttemod;
5234 	struct sf_hment *sfhmep;
5235 	uint_t tteflags;
5236 	int ttesz;
5237 	struct page *pp = NULL;
5238 	kmutex_t *pml, *pmtx;
5239 	int ret;
5240 	int use_demap_range;
5241 #if defined(SF_ERRATA_57)
5242 	int check_exec;
5243 #endif
5244 
5245 	ASSERT(in_hblk_range(hmeblkp, addr));
5246 	ASSERT(hmeblkp->hblk_shw_bit == 0);
5247 	ASSERT(!hmeblkp->hblk_shared);
5248 
5249 #ifdef DEBUG
5250 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5251 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5252 		panic("sfmmu_hblk_chgprot: partial chgprot of large page");
5253 	}
5254 #endif /* DEBUG */
5255 
5256 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5257 	ttesz = get_hblk_ttesz(hmeblkp);
5258 
5259 	pprot = sfmmu_vtop_prot(vprot, &tteflags);
5260 #if defined(SF_ERRATA_57)
5261 	check_exec = (sfmmup != ksfmmup) &&
5262 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
5263 	    ((vprot & PROT_EXEC) == PROT_EXEC);
5264 #endif
5265 	HBLKTOHME(sfhmep, hmeblkp, addr);
5266 
5267 	/*
5268 	 * Flush the current demap region if addresses have been
5269 	 * skipped or the page size doesn't match.
5270 	 */
5271 	use_demap_range = (TTEBYTES(ttesz) == MMU_PAGESIZE);
5272 	if (use_demap_range) {
5273 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5274 	} else {
5275 		DEMAP_RANGE_FLUSH(dmrp);
5276 	}
5277 
5278 	while (addr < endaddr) {
5279 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5280 		if (TTE_IS_VALID(&tte)) {
5281 			if (TTE_GET_LOFLAGS(&tte, tteflags) == pprot) {
5282 				/*
5283 				 * if the new protection is the same as old
5284 				 * continue
5285 				 */
5286 				goto next_addr;
5287 			}
5288 			pml = NULL;
5289 			pp = sfhmep->hme_page;
5290 			if (pp) {
5291 				pml = sfmmu_mlist_enter(pp);
5292 			}
5293 			if (pp != sfhmep->hme_page) {
5294 				/*
5295 				 * tte most have been unloaded
5296 				 * underneath us.  Recheck
5297 				 */
5298 				ASSERT(pml);
5299 				sfmmu_mlist_exit(pml);
5300 				continue;
5301 			}
5302 
5303 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5304 
5305 			ttemod = tte;
5306 			TTE_SET_LOFLAGS(&ttemod, tteflags, pprot);
5307 #if defined(SF_ERRATA_57)
5308 			if (check_exec && addr < errata57_limit)
5309 				ttemod.tte_exec_perm = 0;
5310 #endif
5311 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5312 			    &sfhmep->hme_tte);
5313 
5314 			if (ret < 0) {
5315 				/* tte changed underneath us */
5316 				if (pml) {
5317 					sfmmu_mlist_exit(pml);
5318 				}
5319 				continue;
5320 			}
5321 
5322 			if (tteflags & TTE_HWWR_INT) {
5323 				/*
5324 				 * need to sync if we are clearing modify bit.
5325 				 */
5326 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5327 			}
5328 
5329 			if (pp && PP_ISRO(pp)) {
5330 				if (pprot & TTE_WRPRM_INT) {
5331 					pmtx = sfmmu_page_enter(pp);
5332 					PP_CLRRO(pp);
5333 					sfmmu_page_exit(pmtx);
5334 				}
5335 			}
5336 
5337 			if (ret > 0 && use_demap_range) {
5338 				DEMAP_RANGE_MARKPG(dmrp, addr);
5339 			} else if (ret > 0) {
5340 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
5341 			}
5342 
5343 			if (pml) {
5344 				sfmmu_mlist_exit(pml);
5345 			}
5346 		}
5347 next_addr:
5348 		addr += TTEBYTES(ttesz);
5349 		sfhmep++;
5350 		DEMAP_RANGE_NEXTPG(dmrp);
5351 	}
5352 	return (addr);
5353 }
5354 
5355 /*
5356  * This routine is deprecated and should only be used by hat_chgprot.
5357  * The correct routine is sfmmu_vtop_attr.
5358  * This routine converts virtual page protections to physical ones.  It will
5359  * update the tteflags field with the tte mask corresponding to the protections
5360  * affected and it returns the new protections.  It will also clear the modify
5361  * bit if we are taking away write permission.  This is necessary since the
5362  * modify bit is the hardware permission bit and we need to clear it in order
5363  * to detect write faults.
5364  * It accepts the following special protections:
5365  * ~PROT_WRITE = remove write permissions.
5366  * ~PROT_USER = remove user permissions.
5367  */
5368 static uint_t
5369 sfmmu_vtop_prot(uint_t vprot, uint_t *tteflagsp)
5370 {
5371 	if (vprot == (uint_t)~PROT_WRITE) {
5372 		*tteflagsp = TTE_WRPRM_INT | TTE_HWWR_INT;
5373 		return (0);		/* will cause wrprm to be cleared */
5374 	}
5375 	if (vprot == (uint_t)~PROT_USER) {
5376 		*tteflagsp = TTE_PRIV_INT;
5377 		return (0);		/* will cause privprm to be cleared */
5378 	}
5379 	if ((vprot == 0) || (vprot == PROT_USER) ||
5380 	    ((vprot & PROT_ALL) != vprot)) {
5381 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5382 	}
5383 
5384 	switch (vprot) {
5385 	case (PROT_READ):
5386 	case (PROT_EXEC):
5387 	case (PROT_EXEC | PROT_READ):
5388 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5389 		return (TTE_PRIV_INT); 		/* set prv and clr wrt */
5390 	case (PROT_WRITE):
5391 	case (PROT_WRITE | PROT_READ):
5392 	case (PROT_EXEC | PROT_WRITE):
5393 	case (PROT_EXEC | PROT_WRITE | PROT_READ):
5394 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5395 		return (TTE_PRIV_INT | TTE_WRPRM_INT); 	/* set prv and wrt */
5396 	case (PROT_USER | PROT_READ):
5397 	case (PROT_USER | PROT_EXEC):
5398 	case (PROT_USER | PROT_EXEC | PROT_READ):
5399 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
5400 		return (0); 			/* clr prv and wrt */
5401 	case (PROT_USER | PROT_WRITE):
5402 	case (PROT_USER | PROT_WRITE | PROT_READ):
5403 	case (PROT_USER | PROT_EXEC | PROT_WRITE):
5404 	case (PROT_USER | PROT_EXEC | PROT_WRITE | PROT_READ):
5405 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
5406 		return (TTE_WRPRM_INT); 	/* clr prv and set wrt */
5407 	default:
5408 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
5409 	}
5410 	return (0);
5411 }
5412 
5413 /*
5414  * Alternate unload for very large virtual ranges. With a true 64 bit VA,
5415  * the normal algorithm would take too long for a very large VA range with
5416  * few real mappings. This routine just walks thru all HMEs in the global
5417  * hash table to find and remove mappings.
5418  */
5419 static void
5420 hat_unload_large_virtual(
5421 	struct hat		*sfmmup,
5422 	caddr_t			startaddr,
5423 	size_t			len,
5424 	uint_t			flags,
5425 	hat_callback_t		*callback)
5426 {
5427 	struct hmehash_bucket *hmebp;
5428 	struct hme_blk *hmeblkp;
5429 	struct hme_blk *pr_hblk = NULL;
5430 	struct hme_blk *nx_hblk;
5431 	struct hme_blk *list = NULL;
5432 	int i;
5433 	uint64_t hblkpa, prevpa, nx_pa;
5434 	demap_range_t dmr, *dmrp;
5435 	cpuset_t cpuset;
5436 	caddr_t	endaddr = startaddr + len;
5437 	caddr_t	sa;
5438 	caddr_t	ea;
5439 	caddr_t	cb_sa[MAX_CB_ADDR];
5440 	caddr_t	cb_ea[MAX_CB_ADDR];
5441 	int	addr_cnt = 0;
5442 	int	a = 0;
5443 
5444 	if (sfmmup->sfmmu_free) {
5445 		dmrp = NULL;
5446 	} else {
5447 		dmrp = &dmr;
5448 		DEMAP_RANGE_INIT(sfmmup, dmrp);
5449 	}
5450 
5451 	/*
5452 	 * Loop through all the hash buckets of HME blocks looking for matches.
5453 	 */
5454 	for (i = 0; i <= UHMEHASH_SZ; i++) {
5455 		hmebp = &uhme_hash[i];
5456 		SFMMU_HASH_LOCK(hmebp);
5457 		hmeblkp = hmebp->hmeblkp;
5458 		hblkpa = hmebp->hmeh_nextpa;
5459 		prevpa = 0;
5460 		pr_hblk = NULL;
5461 		while (hmeblkp) {
5462 			nx_hblk = hmeblkp->hblk_next;
5463 			nx_pa = hmeblkp->hblk_nextpa;
5464 
5465 			/*
5466 			 * skip if not this context, if a shadow block or
5467 			 * if the mapping is not in the requested range
5468 			 */
5469 			if (hmeblkp->hblk_tag.htag_id != sfmmup ||
5470 			    hmeblkp->hblk_shw_bit ||
5471 			    (sa = (caddr_t)get_hblk_base(hmeblkp)) >= endaddr ||
5472 			    (ea = get_hblk_endaddr(hmeblkp)) <= startaddr) {
5473 				pr_hblk = hmeblkp;
5474 				prevpa = hblkpa;
5475 				goto next_block;
5476 			}
5477 
5478 			ASSERT(!hmeblkp->hblk_shared);
5479 			/*
5480 			 * unload if there are any current valid mappings
5481 			 */
5482 			if (hmeblkp->hblk_vcnt != 0 ||
5483 			    hmeblkp->hblk_hmecnt != 0)
5484 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
5485 				    sa, ea, dmrp, flags);
5486 
5487 			/*
5488 			 * on unmap we also release the HME block itself, once
5489 			 * all mappings are gone.
5490 			 */
5491 			if ((flags & HAT_UNLOAD_UNMAP) != 0 &&
5492 			    !hmeblkp->hblk_vcnt &&
5493 			    !hmeblkp->hblk_hmecnt) {
5494 				ASSERT(!hmeblkp->hblk_lckcnt);
5495 				sfmmu_hblk_hash_rm(hmebp, hmeblkp,
5496 				    prevpa, pr_hblk);
5497 				sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
5498 			} else {
5499 				pr_hblk = hmeblkp;
5500 				prevpa = hblkpa;
5501 			}
5502 
5503 			if (callback == NULL)
5504 				goto next_block;
5505 
5506 			/*
5507 			 * HME blocks may span more than one page, but we may be
5508 			 * unmapping only one page, so check for a smaller range
5509 			 * for the callback
5510 			 */
5511 			if (sa < startaddr)
5512 				sa = startaddr;
5513 			if (--ea > endaddr)
5514 				ea = endaddr - 1;
5515 
5516 			cb_sa[addr_cnt] = sa;
5517 			cb_ea[addr_cnt] = ea;
5518 			if (++addr_cnt == MAX_CB_ADDR) {
5519 				if (dmrp != NULL) {
5520 					DEMAP_RANGE_FLUSH(dmrp);
5521 					cpuset = sfmmup->sfmmu_cpusran;
5522 					xt_sync(cpuset);
5523 				}
5524 
5525 				for (a = 0; a < MAX_CB_ADDR; ++a) {
5526 					callback->hcb_start_addr = cb_sa[a];
5527 					callback->hcb_end_addr = cb_ea[a];
5528 					callback->hcb_function(callback);
5529 				}
5530 				addr_cnt = 0;
5531 			}
5532 
5533 next_block:
5534 			hmeblkp = nx_hblk;
5535 			hblkpa = nx_pa;
5536 		}
5537 		SFMMU_HASH_UNLOCK(hmebp);
5538 	}
5539 
5540 	sfmmu_hblks_list_purge(&list);
5541 	if (dmrp != NULL) {
5542 		DEMAP_RANGE_FLUSH(dmrp);
5543 		cpuset = sfmmup->sfmmu_cpusran;
5544 		xt_sync(cpuset);
5545 	}
5546 
5547 	for (a = 0; a < addr_cnt; ++a) {
5548 		callback->hcb_start_addr = cb_sa[a];
5549 		callback->hcb_end_addr = cb_ea[a];
5550 		callback->hcb_function(callback);
5551 	}
5552 
5553 	/*
5554 	 * Check TSB and TLB page sizes if the process isn't exiting.
5555 	 */
5556 	if (!sfmmup->sfmmu_free)
5557 		sfmmu_check_page_sizes(sfmmup, 0);
5558 }
5559 
5560 /*
5561  * Unload all the mappings in the range [addr..addr+len). addr and len must
5562  * be MMU_PAGESIZE aligned.
5563  */
5564 
5565 extern struct seg *segkmap;
5566 #define	ISSEGKMAP(sfmmup, addr) (sfmmup == ksfmmup && \
5567 segkmap->s_base <= (addr) && (addr) < (segkmap->s_base + segkmap->s_size))
5568 
5569 
5570 void
5571 hat_unload_callback(
5572 	struct hat *sfmmup,
5573 	caddr_t addr,
5574 	size_t len,
5575 	uint_t flags,
5576 	hat_callback_t *callback)
5577 {
5578 	struct hmehash_bucket *hmebp;
5579 	hmeblk_tag hblktag;
5580 	int hmeshift, hashno, iskernel;
5581 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
5582 	caddr_t endaddr;
5583 	cpuset_t cpuset;
5584 	uint64_t hblkpa, prevpa;
5585 	int addr_count = 0;
5586 	int a;
5587 	caddr_t cb_start_addr[MAX_CB_ADDR];
5588 	caddr_t cb_end_addr[MAX_CB_ADDR];
5589 	int issegkmap = ISSEGKMAP(sfmmup, addr);
5590 	demap_range_t dmr, *dmrp;
5591 
5592 	if (sfmmup->sfmmu_xhat_provider) {
5593 		XHAT_UNLOAD_CALLBACK(sfmmup, addr, len, flags, callback);
5594 		return;
5595 	} else {
5596 		/*
5597 		 * This must be a CPU HAT. If the address space has
5598 		 * XHATs attached, unload the mappings for all of them,
5599 		 * just in case
5600 		 */
5601 		ASSERT(sfmmup->sfmmu_as != NULL);
5602 		if (sfmmup->sfmmu_as->a_xhat != NULL)
5603 			xhat_unload_callback_all(sfmmup->sfmmu_as, addr,
5604 			    len, flags, callback);
5605 	}
5606 
5607 	ASSERT((sfmmup == ksfmmup) || (flags & HAT_UNLOAD_OTHER) || \
5608 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
5609 
5610 	ASSERT(sfmmup != NULL);
5611 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5612 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
5613 
5614 	/*
5615 	 * Probing through a large VA range (say 63 bits) will be slow, even
5616 	 * at 4 Meg steps between the probes. So, when the virtual address range
5617 	 * is very large, search the HME entries for what to unload.
5618 	 *
5619 	 *	len >> TTE_PAGE_SHIFT(TTE4M) is the # of 4Meg probes we'd need
5620 	 *
5621 	 *	UHMEHASH_SZ is number of hash buckets to examine
5622 	 *
5623 	 */
5624 	if (sfmmup != KHATID && (len >> TTE_PAGE_SHIFT(TTE4M)) > UHMEHASH_SZ) {
5625 		hat_unload_large_virtual(sfmmup, addr, len, flags, callback);
5626 		return;
5627 	}
5628 
5629 	CPUSET_ZERO(cpuset);
5630 
5631 	/*
5632 	 * If the process is exiting, we can save a lot of fuss since
5633 	 * we'll flush the TLB when we free the ctx anyway.
5634 	 */
5635 	if (sfmmup->sfmmu_free)
5636 		dmrp = NULL;
5637 	else
5638 		dmrp = &dmr;
5639 
5640 	DEMAP_RANGE_INIT(sfmmup, dmrp);
5641 	endaddr = addr + len;
5642 	hblktag.htag_id = sfmmup;
5643 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
5644 
5645 	/*
5646 	 * It is likely for the vm to call unload over a wide range of
5647 	 * addresses that are actually very sparsely populated by
5648 	 * translations.  In order to speed this up the sfmmu hat supports
5649 	 * the concept of shadow hmeblks. Dummy large page hmeblks that
5650 	 * correspond to actual small translations are allocated at tteload
5651 	 * time and are referred to as shadow hmeblks.  Now, during unload
5652 	 * time, we first check if we have a shadow hmeblk for that
5653 	 * translation.  The absence of one means the corresponding address
5654 	 * range is empty and can be skipped.
5655 	 *
5656 	 * The kernel is an exception to above statement and that is why
5657 	 * we don't use shadow hmeblks and hash starting from the smallest
5658 	 * page size.
5659 	 */
5660 	if (sfmmup == KHATID) {
5661 		iskernel = 1;
5662 		hashno = TTE64K;
5663 	} else {
5664 		iskernel = 0;
5665 		if (mmu_page_sizes == max_mmu_page_sizes) {
5666 			hashno = TTE256M;
5667 		} else {
5668 			hashno = TTE4M;
5669 		}
5670 	}
5671 	while (addr < endaddr) {
5672 		hmeshift = HME_HASH_SHIFT(hashno);
5673 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5674 		hblktag.htag_rehash = hashno;
5675 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5676 
5677 		SFMMU_HASH_LOCK(hmebp);
5678 
5679 		HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, hblkpa, pr_hblk,
5680 		    prevpa, &list);
5681 		if (hmeblkp == NULL) {
5682 			/*
5683 			 * didn't find an hmeblk. skip the appropiate
5684 			 * address range.
5685 			 */
5686 			SFMMU_HASH_UNLOCK(hmebp);
5687 			if (iskernel) {
5688 				if (hashno < mmu_hashcnt) {
5689 					hashno++;
5690 					continue;
5691 				} else {
5692 					hashno = TTE64K;
5693 					addr = (caddr_t)roundup((uintptr_t)addr
5694 					    + 1, MMU_PAGESIZE64K);
5695 					continue;
5696 				}
5697 			}
5698 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5699 			    (1 << hmeshift));
5700 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5701 				ASSERT(hashno == TTE64K);
5702 				continue;
5703 			}
5704 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5705 				hashno = TTE512K;
5706 				continue;
5707 			}
5708 			if (mmu_page_sizes == max_mmu_page_sizes) {
5709 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5710 					hashno = TTE4M;
5711 					continue;
5712 				}
5713 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5714 					hashno = TTE32M;
5715 					continue;
5716 				}
5717 				hashno = TTE256M;
5718 				continue;
5719 			} else {
5720 				hashno = TTE4M;
5721 				continue;
5722 			}
5723 		}
5724 		ASSERT(hmeblkp);
5725 		ASSERT(!hmeblkp->hblk_shared);
5726 		if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5727 			/*
5728 			 * If the valid count is zero we can skip the range
5729 			 * mapped by this hmeblk.
5730 			 * We free hblks in the case of HAT_UNMAP.  HAT_UNMAP
5731 			 * is used by segment drivers as a hint
5732 			 * that the mapping resource won't be used any longer.
5733 			 * The best example of this is during exit().
5734 			 */
5735 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
5736 			    get_hblk_span(hmeblkp));
5737 			if ((flags & HAT_UNLOAD_UNMAP) ||
5738 			    (iskernel && !issegkmap)) {
5739 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa,
5740 				    pr_hblk);
5741 				sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
5742 			}
5743 			SFMMU_HASH_UNLOCK(hmebp);
5744 
5745 			if (iskernel) {
5746 				hashno = TTE64K;
5747 				continue;
5748 			}
5749 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5750 				ASSERT(hashno == TTE64K);
5751 				continue;
5752 			}
5753 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5754 				hashno = TTE512K;
5755 				continue;
5756 			}
5757 			if (mmu_page_sizes == max_mmu_page_sizes) {
5758 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5759 					hashno = TTE4M;
5760 					continue;
5761 				}
5762 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5763 					hashno = TTE32M;
5764 					continue;
5765 				}
5766 				hashno = TTE256M;
5767 				continue;
5768 			} else {
5769 				hashno = TTE4M;
5770 				continue;
5771 			}
5772 		}
5773 		if (hmeblkp->hblk_shw_bit) {
5774 			/*
5775 			 * If we encounter a shadow hmeblk we know there is
5776 			 * smaller sized hmeblks mapping the same address space.
5777 			 * Decrement the hash size and rehash.
5778 			 */
5779 			ASSERT(sfmmup != KHATID);
5780 			hashno--;
5781 			SFMMU_HASH_UNLOCK(hmebp);
5782 			continue;
5783 		}
5784 
5785 		/*
5786 		 * track callback address ranges.
5787 		 * only start a new range when it's not contiguous
5788 		 */
5789 		if (callback != NULL) {
5790 			if (addr_count > 0 &&
5791 			    addr == cb_end_addr[addr_count - 1])
5792 				--addr_count;
5793 			else
5794 				cb_start_addr[addr_count] = addr;
5795 		}
5796 
5797 		addr = sfmmu_hblk_unload(sfmmup, hmeblkp, addr, endaddr,
5798 		    dmrp, flags);
5799 
5800 		if (callback != NULL)
5801 			cb_end_addr[addr_count++] = addr;
5802 
5803 		if (((flags & HAT_UNLOAD_UNMAP) || (iskernel && !issegkmap)) &&
5804 		    !hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5805 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa,
5806 			    pr_hblk);
5807 			sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
5808 		}
5809 		SFMMU_HASH_UNLOCK(hmebp);
5810 
5811 		/*
5812 		 * Notify our caller as to exactly which pages
5813 		 * have been unloaded. We do these in clumps,
5814 		 * to minimize the number of xt_sync()s that need to occur.
5815 		 */
5816 		if (callback != NULL && addr_count == MAX_CB_ADDR) {
5817 			DEMAP_RANGE_FLUSH(dmrp);
5818 			if (dmrp != NULL) {
5819 				cpuset = sfmmup->sfmmu_cpusran;
5820 				xt_sync(cpuset);
5821 			}
5822 
5823 			for (a = 0; a < MAX_CB_ADDR; ++a) {
5824 				callback->hcb_start_addr = cb_start_addr[a];
5825 				callback->hcb_end_addr = cb_end_addr[a];
5826 				callback->hcb_function(callback);
5827 			}
5828 			addr_count = 0;
5829 		}
5830 		if (iskernel) {
5831 			hashno = TTE64K;
5832 			continue;
5833 		}
5834 		if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5835 			ASSERT(hashno == TTE64K);
5836 			continue;
5837 		}
5838 		if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5839 			hashno = TTE512K;
5840 			continue;
5841 		}
5842 		if (mmu_page_sizes == max_mmu_page_sizes) {
5843 			if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5844 				hashno = TTE4M;
5845 				continue;
5846 			}
5847 			if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5848 				hashno = TTE32M;
5849 				continue;
5850 			}
5851 			hashno = TTE256M;
5852 		} else {
5853 			hashno = TTE4M;
5854 		}
5855 	}
5856 
5857 	sfmmu_hblks_list_purge(&list);
5858 	DEMAP_RANGE_FLUSH(dmrp);
5859 	if (dmrp != NULL) {
5860 		cpuset = sfmmup->sfmmu_cpusran;
5861 		xt_sync(cpuset);
5862 	}
5863 	if (callback && addr_count != 0) {
5864 		for (a = 0; a < addr_count; ++a) {
5865 			callback->hcb_start_addr = cb_start_addr[a];
5866 			callback->hcb_end_addr = cb_end_addr[a];
5867 			callback->hcb_function(callback);
5868 		}
5869 	}
5870 
5871 	/*
5872 	 * Check TSB and TLB page sizes if the process isn't exiting.
5873 	 */
5874 	if (!sfmmup->sfmmu_free)
5875 		sfmmu_check_page_sizes(sfmmup, 0);
5876 }
5877 
5878 /*
5879  * Unload all the mappings in the range [addr..addr+len). addr and len must
5880  * be MMU_PAGESIZE aligned.
5881  */
5882 void
5883 hat_unload(struct hat *sfmmup, caddr_t addr, size_t len, uint_t flags)
5884 {
5885 	if (sfmmup->sfmmu_xhat_provider) {
5886 		XHAT_UNLOAD(sfmmup, addr, len, flags);
5887 		return;
5888 	}
5889 	hat_unload_callback(sfmmup, addr, len, flags, NULL);
5890 }
5891 
5892 
5893 /*
5894  * Find the largest mapping size for this page.
5895  */
5896 int
5897 fnd_mapping_sz(page_t *pp)
5898 {
5899 	int sz;
5900 	int p_index;
5901 
5902 	p_index = PP_MAPINDEX(pp);
5903 
5904 	sz = 0;
5905 	p_index >>= 1;	/* don't care about 8K bit */
5906 	for (; p_index; p_index >>= 1) {
5907 		sz++;
5908 	}
5909 
5910 	return (sz);
5911 }
5912 
5913 /*
5914  * This function unloads a range of addresses for an hmeblk.
5915  * It returns the next address to be unloaded.
5916  * It should be called with the hash lock held.
5917  */
5918 static caddr_t
5919 sfmmu_hblk_unload(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5920 	caddr_t endaddr, demap_range_t *dmrp, uint_t flags)
5921 {
5922 	tte_t	tte, ttemod;
5923 	struct	sf_hment *sfhmep;
5924 	int	ttesz;
5925 	long	ttecnt;
5926 	page_t *pp;
5927 	kmutex_t *pml;
5928 	int ret;
5929 	int use_demap_range;
5930 
5931 	ASSERT(in_hblk_range(hmeblkp, addr));
5932 	ASSERT(!hmeblkp->hblk_shw_bit);
5933 	ASSERT(sfmmup != NULL || hmeblkp->hblk_shared);
5934 	ASSERT(sfmmup == NULL || !hmeblkp->hblk_shared);
5935 	ASSERT(dmrp == NULL || !hmeblkp->hblk_shared);
5936 
5937 #ifdef DEBUG
5938 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5939 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5940 		panic("sfmmu_hblk_unload: partial unload of large page");
5941 	}
5942 #endif /* DEBUG */
5943 
5944 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5945 	ttesz = get_hblk_ttesz(hmeblkp);
5946 
5947 	use_demap_range = ((dmrp == NULL) ||
5948 	    (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp)));
5949 
5950 	if (use_demap_range) {
5951 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5952 	} else {
5953 		DEMAP_RANGE_FLUSH(dmrp);
5954 	}
5955 	ttecnt = 0;
5956 	HBLKTOHME(sfhmep, hmeblkp, addr);
5957 
5958 	while (addr < endaddr) {
5959 		pml = NULL;
5960 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5961 		if (TTE_IS_VALID(&tte)) {
5962 			pp = sfhmep->hme_page;
5963 			if (pp != NULL) {
5964 				pml = sfmmu_mlist_enter(pp);
5965 			}
5966 
5967 			/*
5968 			 * Verify if hme still points to 'pp' now that
5969 			 * we have p_mapping lock.
5970 			 */
5971 			if (sfhmep->hme_page != pp) {
5972 				if (pp != NULL && sfhmep->hme_page != NULL) {
5973 					ASSERT(pml != NULL);
5974 					sfmmu_mlist_exit(pml);
5975 					/* Re-start this iteration. */
5976 					continue;
5977 				}
5978 				ASSERT((pp != NULL) &&
5979 				    (sfhmep->hme_page == NULL));
5980 				goto tte_unloaded;
5981 			}
5982 
5983 			/*
5984 			 * This point on we have both HASH and p_mapping
5985 			 * lock.
5986 			 */
5987 			ASSERT(pp == sfhmep->hme_page);
5988 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5989 
5990 			/*
5991 			 * We need to loop on modify tte because it is
5992 			 * possible for pagesync to come along and
5993 			 * change the software bits beneath us.
5994 			 *
5995 			 * Page_unload can also invalidate the tte after
5996 			 * we read tte outside of p_mapping lock.
5997 			 */
5998 again:
5999 			ttemod = tte;
6000 
6001 			TTE_SET_INVALID(&ttemod);
6002 			ret = sfmmu_modifytte_try(&tte, &ttemod,
6003 			    &sfhmep->hme_tte);
6004 
6005 			if (ret <= 0) {
6006 				if (TTE_IS_VALID(&tte)) {
6007 					ASSERT(ret < 0);
6008 					goto again;
6009 				}
6010 				if (pp != NULL) {
6011 					panic("sfmmu_hblk_unload: pp = 0x%p "
6012 					    "tte became invalid under mlist"
6013 					    " lock = 0x%p", pp, pml);
6014 				}
6015 				continue;
6016 			}
6017 
6018 			if (!(flags & HAT_UNLOAD_NOSYNC)) {
6019 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
6020 			}
6021 
6022 			/*
6023 			 * Ok- we invalidated the tte. Do the rest of the job.
6024 			 */
6025 			ttecnt++;
6026 
6027 			if (flags & HAT_UNLOAD_UNLOCK) {
6028 				ASSERT(hmeblkp->hblk_lckcnt > 0);
6029 				atomic_add_32(&hmeblkp->hblk_lckcnt, -1);
6030 				HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
6031 			}
6032 
6033 			/*
6034 			 * Normally we would need to flush the page
6035 			 * from the virtual cache at this point in
6036 			 * order to prevent a potential cache alias
6037 			 * inconsistency.
6038 			 * The particular scenario we need to worry
6039 			 * about is:
6040 			 * Given:  va1 and va2 are two virtual address
6041 			 * that alias and map the same physical
6042 			 * address.
6043 			 * 1.   mapping exists from va1 to pa and data
6044 			 * has been read into the cache.
6045 			 * 2.   unload va1.
6046 			 * 3.   load va2 and modify data using va2.
6047 			 * 4    unload va2.
6048 			 * 5.   load va1 and reference data.  Unless we
6049 			 * flush the data cache when we unload we will
6050 			 * get stale data.
6051 			 * Fortunately, page coloring eliminates the
6052 			 * above scenario by remembering the color a
6053 			 * physical page was last or is currently
6054 			 * mapped to.  Now, we delay the flush until
6055 			 * the loading of translations.  Only when the
6056 			 * new translation is of a different color
6057 			 * are we forced to flush.
6058 			 */
6059 			if (use_demap_range) {
6060 				/*
6061 				 * Mark this page as needing a demap.
6062 				 */
6063 				DEMAP_RANGE_MARKPG(dmrp, addr);
6064 			} else {
6065 				ASSERT(sfmmup != NULL);
6066 				ASSERT(!hmeblkp->hblk_shared);
6067 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
6068 				    sfmmup->sfmmu_free, 0);
6069 			}
6070 
6071 			if (pp) {
6072 				/*
6073 				 * Remove the hment from the mapping list
6074 				 */
6075 				ASSERT(hmeblkp->hblk_hmecnt > 0);
6076 
6077 				/*
6078 				 * Again, we cannot
6079 				 * ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS);
6080 				 */
6081 				HME_SUB(sfhmep, pp);
6082 				membar_stst();
6083 				atomic_add_16(&hmeblkp->hblk_hmecnt, -1);
6084 			}
6085 
6086 			ASSERT(hmeblkp->hblk_vcnt > 0);
6087 			atomic_add_16(&hmeblkp->hblk_vcnt, -1);
6088 
6089 			ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
6090 			    !hmeblkp->hblk_lckcnt);
6091 
6092 #ifdef VAC
6093 			if (pp && (pp->p_nrm & (P_KPMC | P_KPMS | P_TNC))) {
6094 				if (PP_ISTNC(pp)) {
6095 					/*
6096 					 * If page was temporary
6097 					 * uncached, try to recache
6098 					 * it. Note that HME_SUB() was
6099 					 * called above so p_index and
6100 					 * mlist had been updated.
6101 					 */
6102 					conv_tnc(pp, ttesz);
6103 				} else if (pp->p_mapping == NULL) {
6104 					ASSERT(kpm_enable);
6105 					/*
6106 					 * Page is marked to be in VAC conflict
6107 					 * to an existing kpm mapping and/or is
6108 					 * kpm mapped using only the regular
6109 					 * pagesize.
6110 					 */
6111 					sfmmu_kpm_hme_unload(pp);
6112 				}
6113 			}
6114 #endif	/* VAC */
6115 		} else if ((pp = sfhmep->hme_page) != NULL) {
6116 				/*
6117 				 * TTE is invalid but the hme
6118 				 * still exists. let pageunload
6119 				 * complete its job.
6120 				 */
6121 				ASSERT(pml == NULL);
6122 				pml = sfmmu_mlist_enter(pp);
6123 				if (sfhmep->hme_page != NULL) {
6124 					sfmmu_mlist_exit(pml);
6125 					continue;
6126 				}
6127 				ASSERT(sfhmep->hme_page == NULL);
6128 		} else if (hmeblkp->hblk_hmecnt != 0) {
6129 			/*
6130 			 * pageunload may have not finished decrementing
6131 			 * hblk_vcnt and hblk_hmecnt. Find page_t if any and
6132 			 * wait for pageunload to finish. Rely on pageunload
6133 			 * to decrement hblk_hmecnt after hblk_vcnt.
6134 			 */
6135 			pfn_t pfn = TTE_TO_TTEPFN(&tte);
6136 			ASSERT(pml == NULL);
6137 			if (pf_is_memory(pfn)) {
6138 				pp = page_numtopp_nolock(pfn);
6139 				if (pp != NULL) {
6140 					pml = sfmmu_mlist_enter(pp);
6141 					sfmmu_mlist_exit(pml);
6142 					pml = NULL;
6143 				}
6144 			}
6145 		}
6146 
6147 tte_unloaded:
6148 		/*
6149 		 * At this point, the tte we are looking at
6150 		 * should be unloaded, and hme has been unlinked
6151 		 * from page too. This is important because in
6152 		 * pageunload, it does ttesync() then HME_SUB.
6153 		 * We need to make sure HME_SUB has been completed
6154 		 * so we know ttesync() has been completed. Otherwise,
6155 		 * at exit time, after return from hat layer, VM will
6156 		 * release as structure which hat_setstat() (called
6157 		 * by ttesync()) needs.
6158 		 */
6159 #ifdef DEBUG
6160 		{
6161 			tte_t	dtte;
6162 
6163 			ASSERT(sfhmep->hme_page == NULL);
6164 
6165 			sfmmu_copytte(&sfhmep->hme_tte, &dtte);
6166 			ASSERT(!TTE_IS_VALID(&dtte));
6167 		}
6168 #endif
6169 
6170 		if (pml) {
6171 			sfmmu_mlist_exit(pml);
6172 		}
6173 
6174 		addr += TTEBYTES(ttesz);
6175 		sfhmep++;
6176 		DEMAP_RANGE_NEXTPG(dmrp);
6177 	}
6178 	/*
6179 	 * For shared hmeblks this routine is only called when region is freed
6180 	 * and no longer referenced.  So no need to decrement ttecnt
6181 	 * in the region structure here.
6182 	 */
6183 	if (ttecnt > 0 && sfmmup != NULL) {
6184 		atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -ttecnt);
6185 	}
6186 	return (addr);
6187 }
6188 
6189 /*
6190  * Synchronize all the mappings in the range [addr..addr+len).
6191  * Can be called with clearflag having two states:
6192  * HAT_SYNC_DONTZERO means just return the rm stats
6193  * HAT_SYNC_ZERORM means zero rm bits in the tte and return the stats
6194  */
6195 void
6196 hat_sync(struct hat *sfmmup, caddr_t addr, size_t len, uint_t clearflag)
6197 {
6198 	struct hmehash_bucket *hmebp;
6199 	hmeblk_tag hblktag;
6200 	int hmeshift, hashno = 1;
6201 	struct hme_blk *hmeblkp, *list = NULL;
6202 	caddr_t endaddr;
6203 	cpuset_t cpuset;
6204 
6205 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
6206 	ASSERT((sfmmup == ksfmmup) ||
6207 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
6208 	ASSERT((len & MMU_PAGEOFFSET) == 0);
6209 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
6210 	    (clearflag == HAT_SYNC_ZERORM));
6211 
6212 	CPUSET_ZERO(cpuset);
6213 
6214 	endaddr = addr + len;
6215 	hblktag.htag_id = sfmmup;
6216 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
6217 
6218 	/*
6219 	 * Spitfire supports 4 page sizes.
6220 	 * Most pages are expected to be of the smallest page
6221 	 * size (8K) and these will not need to be rehashed. 64K
6222 	 * pages also don't need to be rehashed because the an hmeblk
6223 	 * spans 64K of address space. 512K pages might need 1 rehash and
6224 	 * and 4M pages 2 rehashes.
6225 	 */
6226 	while (addr < endaddr) {
6227 		hmeshift = HME_HASH_SHIFT(hashno);
6228 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
6229 		hblktag.htag_rehash = hashno;
6230 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
6231 
6232 		SFMMU_HASH_LOCK(hmebp);
6233 
6234 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
6235 		if (hmeblkp != NULL) {
6236 			ASSERT(!hmeblkp->hblk_shared);
6237 			/*
6238 			 * We've encountered a shadow hmeblk so skip the range
6239 			 * of the next smaller mapping size.
6240 			 */
6241 			if (hmeblkp->hblk_shw_bit) {
6242 				ASSERT(sfmmup != ksfmmup);
6243 				ASSERT(hashno > 1);
6244 				addr = (caddr_t)P2END((uintptr_t)addr,
6245 				    TTEBYTES(hashno - 1));
6246 			} else {
6247 				addr = sfmmu_hblk_sync(sfmmup, hmeblkp,
6248 				    addr, endaddr, clearflag);
6249 			}
6250 			SFMMU_HASH_UNLOCK(hmebp);
6251 			hashno = 1;
6252 			continue;
6253 		}
6254 		SFMMU_HASH_UNLOCK(hmebp);
6255 
6256 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
6257 			/*
6258 			 * We have traversed the whole list and rehashed
6259 			 * if necessary without finding the address to sync.
6260 			 * This is ok so we increment the address by the
6261 			 * smallest hmeblk range for kernel mappings and the
6262 			 * largest hmeblk range, to account for shadow hmeblks,
6263 			 * for user mappings and continue.
6264 			 */
6265 			if (sfmmup == ksfmmup)
6266 				addr = (caddr_t)P2END((uintptr_t)addr,
6267 				    TTEBYTES(1));
6268 			else
6269 				addr = (caddr_t)P2END((uintptr_t)addr,
6270 				    TTEBYTES(hashno));
6271 			hashno = 1;
6272 		} else {
6273 			hashno++;
6274 		}
6275 	}
6276 	sfmmu_hblks_list_purge(&list);
6277 	cpuset = sfmmup->sfmmu_cpusran;
6278 	xt_sync(cpuset);
6279 }
6280 
6281 static caddr_t
6282 sfmmu_hblk_sync(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
6283 	caddr_t endaddr, int clearflag)
6284 {
6285 	tte_t	tte, ttemod;
6286 	struct sf_hment *sfhmep;
6287 	int ttesz;
6288 	struct page *pp;
6289 	kmutex_t *pml;
6290 	int ret;
6291 
6292 	ASSERT(hmeblkp->hblk_shw_bit == 0);
6293 	ASSERT(!hmeblkp->hblk_shared);
6294 
6295 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
6296 
6297 	ttesz = get_hblk_ttesz(hmeblkp);
6298 	HBLKTOHME(sfhmep, hmeblkp, addr);
6299 
6300 	while (addr < endaddr) {
6301 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6302 		if (TTE_IS_VALID(&tte)) {
6303 			pml = NULL;
6304 			pp = sfhmep->hme_page;
6305 			if (pp) {
6306 				pml = sfmmu_mlist_enter(pp);
6307 			}
6308 			if (pp != sfhmep->hme_page) {
6309 				/*
6310 				 * tte most have been unloaded
6311 				 * underneath us.  Recheck
6312 				 */
6313 				ASSERT(pml);
6314 				sfmmu_mlist_exit(pml);
6315 				continue;
6316 			}
6317 
6318 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
6319 
6320 			if (clearflag == HAT_SYNC_ZERORM) {
6321 				ttemod = tte;
6322 				TTE_CLR_RM(&ttemod);
6323 				ret = sfmmu_modifytte_try(&tte, &ttemod,
6324 				    &sfhmep->hme_tte);
6325 				if (ret < 0) {
6326 					if (pml) {
6327 						sfmmu_mlist_exit(pml);
6328 					}
6329 					continue;
6330 				}
6331 
6332 				if (ret > 0) {
6333 					sfmmu_tlb_demap(addr, sfmmup,
6334 					    hmeblkp, 0, 0);
6335 				}
6336 			}
6337 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
6338 			if (pml) {
6339 				sfmmu_mlist_exit(pml);
6340 			}
6341 		}
6342 		addr += TTEBYTES(ttesz);
6343 		sfhmep++;
6344 	}
6345 	return (addr);
6346 }
6347 
6348 /*
6349  * This function will sync a tte to the page struct and it will
6350  * update the hat stats. Currently it allows us to pass a NULL pp
6351  * and we will simply update the stats.  We may want to change this
6352  * so we only keep stats for pages backed by pp's.
6353  */
6354 static void
6355 sfmmu_ttesync(struct hat *sfmmup, caddr_t addr, tte_t *ttep, page_t *pp)
6356 {
6357 	uint_t rm = 0;
6358 	int   	sz;
6359 	pgcnt_t	npgs;
6360 
6361 	ASSERT(TTE_IS_VALID(ttep));
6362 
6363 	if (TTE_IS_NOSYNC(ttep)) {
6364 		return;
6365 	}
6366 
6367 	if (TTE_IS_REF(ttep))  {
6368 		rm = P_REF;
6369 	}
6370 	if (TTE_IS_MOD(ttep))  {
6371 		rm |= P_MOD;
6372 	}
6373 
6374 	if (rm == 0) {
6375 		return;
6376 	}
6377 
6378 	sz = TTE_CSZ(ttep);
6379 	if (sfmmup != NULL && sfmmup->sfmmu_rmstat) {
6380 		int i;
6381 		caddr_t	vaddr = addr;
6382 
6383 		for (i = 0; i < TTEPAGES(sz); i++, vaddr += MMU_PAGESIZE) {
6384 			hat_setstat(sfmmup->sfmmu_as, vaddr, MMU_PAGESIZE, rm);
6385 		}
6386 
6387 	}
6388 
6389 	/*
6390 	 * XXX I want to use cas to update nrm bits but they
6391 	 * currently belong in common/vm and not in hat where
6392 	 * they should be.
6393 	 * The nrm bits are protected by the same mutex as
6394 	 * the one that protects the page's mapping list.
6395 	 */
6396 	if (!pp)
6397 		return;
6398 	ASSERT(sfmmu_mlist_held(pp));
6399 	/*
6400 	 * If the tte is for a large page, we need to sync all the
6401 	 * pages covered by the tte.
6402 	 */
6403 	if (sz != TTE8K) {
6404 		ASSERT(pp->p_szc != 0);
6405 		pp = PP_GROUPLEADER(pp, sz);
6406 		ASSERT(sfmmu_mlist_held(pp));
6407 	}
6408 
6409 	/* Get number of pages from tte size. */
6410 	npgs = TTEPAGES(sz);
6411 
6412 	do {
6413 		ASSERT(pp);
6414 		ASSERT(sfmmu_mlist_held(pp));
6415 		if (((rm & P_REF) != 0 && !PP_ISREF(pp)) ||
6416 		    ((rm & P_MOD) != 0 && !PP_ISMOD(pp)))
6417 			hat_page_setattr(pp, rm);
6418 
6419 		/*
6420 		 * Are we done? If not, we must have a large mapping.
6421 		 * For large mappings we need to sync the rest of the pages
6422 		 * covered by this tte; goto the next page.
6423 		 */
6424 	} while (--npgs > 0 && (pp = PP_PAGENEXT(pp)));
6425 }
6426 
6427 /*
6428  * Execute pre-callback handler of each pa_hment linked to pp
6429  *
6430  * Inputs:
6431  *   flag: either HAT_PRESUSPEND or HAT_SUSPEND.
6432  *   capture_cpus: pointer to return value (below)
6433  *
6434  * Returns:
6435  *   Propagates the subsystem callback return values back to the caller;
6436  *   returns 0 on success.  If capture_cpus is non-NULL, the value returned
6437  *   is zero if all of the pa_hments are of a type that do not require
6438  *   capturing CPUs prior to suspending the mapping, else it is 1.
6439  */
6440 static int
6441 hat_pageprocess_precallbacks(struct page *pp, uint_t flag, int *capture_cpus)
6442 {
6443 	struct sf_hment	*sfhmep;
6444 	struct pa_hment *pahmep;
6445 	int (*f)(caddr_t, uint_t, uint_t, void *);
6446 	int		ret;
6447 	id_t		id;
6448 	int		locked = 0;
6449 	kmutex_t	*pml;
6450 
6451 	ASSERT(PAGE_EXCL(pp));
6452 	if (!sfmmu_mlist_held(pp)) {
6453 		pml = sfmmu_mlist_enter(pp);
6454 		locked = 1;
6455 	}
6456 
6457 	if (capture_cpus)
6458 		*capture_cpus = 0;
6459 
6460 top:
6461 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6462 		/*
6463 		 * skip sf_hments corresponding to VA<->PA mappings;
6464 		 * for pa_hment's, hme_tte.ll is zero
6465 		 */
6466 		if (!IS_PAHME(sfhmep))
6467 			continue;
6468 
6469 		pahmep = sfhmep->hme_data;
6470 		ASSERT(pahmep != NULL);
6471 
6472 		/*
6473 		 * skip if pre-handler has been called earlier in this loop
6474 		 */
6475 		if (pahmep->flags & flag)
6476 			continue;
6477 
6478 		id = pahmep->cb_id;
6479 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6480 		if (capture_cpus && sfmmu_cb_table[id].capture_cpus != 0)
6481 			*capture_cpus = 1;
6482 		if ((f = sfmmu_cb_table[id].prehandler) == NULL) {
6483 			pahmep->flags |= flag;
6484 			continue;
6485 		}
6486 
6487 		/*
6488 		 * Drop the mapping list lock to avoid locking order issues.
6489 		 */
6490 		if (locked)
6491 			sfmmu_mlist_exit(pml);
6492 
6493 		ret = f(pahmep->addr, pahmep->len, flag, pahmep->pvt);
6494 		if (ret != 0)
6495 			return (ret);	/* caller must do the cleanup */
6496 
6497 		if (locked) {
6498 			pml = sfmmu_mlist_enter(pp);
6499 			pahmep->flags |= flag;
6500 			goto top;
6501 		}
6502 
6503 		pahmep->flags |= flag;
6504 	}
6505 
6506 	if (locked)
6507 		sfmmu_mlist_exit(pml);
6508 
6509 	return (0);
6510 }
6511 
6512 /*
6513  * Execute post-callback handler of each pa_hment linked to pp
6514  *
6515  * Same overall assumptions and restrictions apply as for
6516  * hat_pageprocess_precallbacks().
6517  */
6518 static void
6519 hat_pageprocess_postcallbacks(struct page *pp, uint_t flag)
6520 {
6521 	pfn_t pgpfn = pp->p_pagenum;
6522 	pfn_t pgmask = btop(page_get_pagesize(pp->p_szc)) - 1;
6523 	pfn_t newpfn;
6524 	struct sf_hment *sfhmep;
6525 	struct pa_hment *pahmep;
6526 	int (*f)(caddr_t, uint_t, uint_t, void *, pfn_t);
6527 	id_t	id;
6528 	int	locked = 0;
6529 	kmutex_t *pml;
6530 
6531 	ASSERT(PAGE_EXCL(pp));
6532 	if (!sfmmu_mlist_held(pp)) {
6533 		pml = sfmmu_mlist_enter(pp);
6534 		locked = 1;
6535 	}
6536 
6537 top:
6538 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6539 		/*
6540 		 * skip sf_hments corresponding to VA<->PA mappings;
6541 		 * for pa_hment's, hme_tte.ll is zero
6542 		 */
6543 		if (!IS_PAHME(sfhmep))
6544 			continue;
6545 
6546 		pahmep = sfhmep->hme_data;
6547 		ASSERT(pahmep != NULL);
6548 
6549 		if ((pahmep->flags & flag) == 0)
6550 			continue;
6551 
6552 		pahmep->flags &= ~flag;
6553 
6554 		id = pahmep->cb_id;
6555 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
6556 		if ((f = sfmmu_cb_table[id].posthandler) == NULL)
6557 			continue;
6558 
6559 		/*
6560 		 * Convert the base page PFN into the constituent PFN
6561 		 * which is needed by the callback handler.
6562 		 */
6563 		newpfn = pgpfn | (btop((uintptr_t)pahmep->addr) & pgmask);
6564 
6565 		/*
6566 		 * Drop the mapping list lock to avoid locking order issues.
6567 		 */
6568 		if (locked)
6569 			sfmmu_mlist_exit(pml);
6570 
6571 		if (f(pahmep->addr, pahmep->len, flag, pahmep->pvt, newpfn)
6572 		    != 0)
6573 			panic("sfmmu: posthandler failed");
6574 
6575 		if (locked) {
6576 			pml = sfmmu_mlist_enter(pp);
6577 			goto top;
6578 		}
6579 	}
6580 
6581 	if (locked)
6582 		sfmmu_mlist_exit(pml);
6583 }
6584 
6585 /*
6586  * Suspend locked kernel mapping
6587  */
6588 void
6589 hat_pagesuspend(struct page *pp)
6590 {
6591 	struct sf_hment *sfhmep;
6592 	sfmmu_t *sfmmup;
6593 	tte_t tte, ttemod;
6594 	struct hme_blk *hmeblkp;
6595 	caddr_t addr;
6596 	int index, cons;
6597 	cpuset_t cpuset;
6598 
6599 	ASSERT(PAGE_EXCL(pp));
6600 	ASSERT(sfmmu_mlist_held(pp));
6601 
6602 	mutex_enter(&kpr_suspendlock);
6603 
6604 	/*
6605 	 * We're about to suspend a kernel mapping so mark this thread as
6606 	 * non-traceable by DTrace. This prevents us from running into issues
6607 	 * with probe context trying to touch a suspended page
6608 	 * in the relocation codepath itself.
6609 	 */
6610 	curthread->t_flag |= T_DONTDTRACE;
6611 
6612 	index = PP_MAPINDEX(pp);
6613 	cons = TTE8K;
6614 
6615 retry:
6616 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
6617 
6618 		if (IS_PAHME(sfhmep))
6619 			continue;
6620 
6621 		if (get_hblk_ttesz(sfmmu_hmetohblk(sfhmep)) != cons)
6622 			continue;
6623 
6624 		/*
6625 		 * Loop until we successfully set the suspend bit in
6626 		 * the TTE.
6627 		 */
6628 again:
6629 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
6630 		ASSERT(TTE_IS_VALID(&tte));
6631 
6632 		ttemod = tte;
6633 		TTE_SET_SUSPEND(&ttemod);
6634 		if (sfmmu_modifytte_try(&tte, &ttemod,
6635 		    &sfhmep->hme_tte) < 0)
6636 			goto again;
6637 
6638 		/*
6639 		 * Invalidate TSB entry
6640 		 */
6641 		hmeblkp = sfmmu_hmetohblk(sfhmep);
6642 
6643 		sfmmup = hblktosfmmu(hmeblkp);
6644 		ASSERT(sfmmup == ksfmmup);
6645 		ASSERT(!hmeblkp->hblk_shared);
6646 
6647 		addr = tte_to_vaddr(hmeblkp, tte);
6648 
6649 		/*
6650 		 * No need to make sure that the TSB for this sfmmu is
6651 		 * not being relocated since it is ksfmmup and thus it
6652 		 * will never be relocated.
6653 		 */
6654 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
6655 
6656 		/*
6657 		 * Update xcall stats
6658 		 */
6659 		cpuset = cpu_ready_set;
6660 		CPUSET_DEL(cpuset, CPU->cpu_id);
6661 
6662 		/* LINTED: constant in conditional context */
6663 		SFMMU_XCALL_STATS(ksfmmup);
6664 
6665 		/*
6666 		 * Flush TLB entry on remote CPU's
6667 		 */
6668 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
6669 		    (uint64_t)ksfmmup);
6670 		xt_sync(cpuset);
6671 
6672 		/*
6673 		 * Flush TLB entry on local CPU
6674 		 */
6675 		vtag_flushpage(addr, (uint64_t)ksfmmup);
6676 	}
6677 
6678 	while (index != 0) {
6679 		index = index >> 1;
6680 		if (index != 0)
6681 			cons++;
6682 		if (index & 0x1) {
6683 			pp = PP_GROUPLEADER(pp, cons);
6684 			goto retry;
6685 		}
6686 	}
6687 }
6688 
6689 #ifdef	DEBUG
6690 
6691 #define	N_PRLE	1024
6692 struct prle {
6693 	page_t *targ;
6694 	page_t *repl;
6695 	int status;
6696 	int pausecpus;
6697 	hrtime_t whence;
6698 };
6699 
6700 static struct prle page_relocate_log[N_PRLE];
6701 static int prl_entry;
6702 static kmutex_t prl_mutex;
6703 
6704 #define	PAGE_RELOCATE_LOG(t, r, s, p)					\
6705 	mutex_enter(&prl_mutex);					\
6706 	page_relocate_log[prl_entry].targ = *(t);			\
6707 	page_relocate_log[prl_entry].repl = *(r);			\
6708 	page_relocate_log[prl_entry].status = (s);			\
6709 	page_relocate_log[prl_entry].pausecpus = (p);			\
6710 	page_relocate_log[prl_entry].whence = gethrtime();		\
6711 	prl_entry = (prl_entry == (N_PRLE - 1))? 0 : prl_entry + 1;	\
6712 	mutex_exit(&prl_mutex);
6713 
6714 #else	/* !DEBUG */
6715 #define	PAGE_RELOCATE_LOG(t, r, s, p)
6716 #endif
6717 
6718 /*
6719  * Core Kernel Page Relocation Algorithm
6720  *
6721  * Input:
6722  *
6723  * target : 	constituent pages are SE_EXCL locked.
6724  * replacement:	constituent pages are SE_EXCL locked.
6725  *
6726  * Output:
6727  *
6728  * nrelocp:	number of pages relocated
6729  */
6730 int
6731 hat_page_relocate(page_t **target, page_t **replacement, spgcnt_t *nrelocp)
6732 {
6733 	page_t		*targ, *repl;
6734 	page_t		*tpp, *rpp;
6735 	kmutex_t	*low, *high;
6736 	spgcnt_t	npages, i;
6737 	page_t		*pl = NULL;
6738 	int		old_pil;
6739 	cpuset_t	cpuset;
6740 	int		cap_cpus;
6741 	int		ret;
6742 
6743 	if (hat_kpr_enabled == 0 || !kcage_on || PP_ISNORELOC(*target)) {
6744 		PAGE_RELOCATE_LOG(target, replacement, EAGAIN, -1);
6745 		return (EAGAIN);
6746 	}
6747 
6748 	mutex_enter(&kpr_mutex);
6749 	kreloc_thread = curthread;
6750 
6751 	targ = *target;
6752 	repl = *replacement;
6753 	ASSERT(repl != NULL);
6754 	ASSERT(targ->p_szc == repl->p_szc);
6755 
6756 	npages = page_get_pagecnt(targ->p_szc);
6757 
6758 	/*
6759 	 * unload VA<->PA mappings that are not locked
6760 	 */
6761 	tpp = targ;
6762 	for (i = 0; i < npages; i++) {
6763 		(void) hat_pageunload(tpp, SFMMU_KERNEL_RELOC);
6764 		tpp++;
6765 	}
6766 
6767 	/*
6768 	 * Do "presuspend" callbacks, in a context from which we can still
6769 	 * block as needed. Note that we don't hold the mapping list lock
6770 	 * of "targ" at this point due to potential locking order issues;
6771 	 * we assume that between the hat_pageunload() above and holding
6772 	 * the SE_EXCL lock that the mapping list *cannot* change at this
6773 	 * point.
6774 	 */
6775 	ret = hat_pageprocess_precallbacks(targ, HAT_PRESUSPEND, &cap_cpus);
6776 	if (ret != 0) {
6777 		/*
6778 		 * EIO translates to fatal error, for all others cleanup
6779 		 * and return EAGAIN.
6780 		 */
6781 		ASSERT(ret != EIO);
6782 		hat_pageprocess_postcallbacks(targ, HAT_POSTUNSUSPEND);
6783 		PAGE_RELOCATE_LOG(target, replacement, ret, -1);
6784 		kreloc_thread = NULL;
6785 		mutex_exit(&kpr_mutex);
6786 		return (EAGAIN);
6787 	}
6788 
6789 	/*
6790 	 * acquire p_mapping list lock for both the target and replacement
6791 	 * root pages.
6792 	 *
6793 	 * low and high refer to the need to grab the mlist locks in a
6794 	 * specific order in order to prevent race conditions.  Thus the
6795 	 * lower lock must be grabbed before the higher lock.
6796 	 *
6797 	 * This will block hat_unload's accessing p_mapping list.  Since
6798 	 * we have SE_EXCL lock, hat_memload and hat_pageunload will be
6799 	 * blocked.  Thus, no one else will be accessing the p_mapping list
6800 	 * while we suspend and reload the locked mapping below.
6801 	 */
6802 	tpp = targ;
6803 	rpp = repl;
6804 	sfmmu_mlist_reloc_enter(tpp, rpp, &low, &high);
6805 
6806 	kpreempt_disable();
6807 
6808 #ifdef VAC
6809 	/*
6810 	 * If the replacement page is of a different virtual color
6811 	 * than the page it is replacing, we need to handle the VAC
6812 	 * consistency for it just as we would if we were setting up
6813 	 * a new mapping to a page.
6814 	 */
6815 	if ((tpp->p_szc == 0) && (PP_GET_VCOLOR(rpp) != NO_VCOLOR)) {
6816 		if (tpp->p_vcolor != rpp->p_vcolor) {
6817 			sfmmu_cache_flushcolor(PP_GET_VCOLOR(rpp),
6818 			    rpp->p_pagenum);
6819 		}
6820 	}
6821 #endif
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 		/*
6873 		 * Copy the contents of the page.
6874 		 */
6875 		ppcopy_kernel(tpp, rpp);
6876 	}
6877 
6878 	tpp = targ;
6879 	rpp = repl;
6880 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6881 		/*
6882 		 * Copy attributes.  VAC consistency was handled above,
6883 		 * if required.
6884 		 */
6885 		rpp->p_nrm = tpp->p_nrm;
6886 		tpp->p_nrm = 0;
6887 		rpp->p_index = tpp->p_index;
6888 		tpp->p_index = 0;
6889 #ifdef VAC
6890 		rpp->p_vcolor = tpp->p_vcolor;
6891 #endif
6892 	}
6893 
6894 	/*
6895 	 * First, unsuspend the page, if we set the suspend bit, and transfer
6896 	 * the mapping list from the target page to the replacement page.
6897 	 * Next process postcallbacks; since pa_hment's are linked only to the
6898 	 * p_mapping list of root page, we don't iterate over the constituent
6899 	 * pages.
6900 	 */
6901 	hat_pagereload(targ, repl);
6902 
6903 suspend_fail:
6904 	hat_pageprocess_postcallbacks(repl, HAT_UNSUSPEND);
6905 
6906 	/*
6907 	 * Now lower our PIL and release any captured CPUs since we
6908 	 * are out of the "danger zone".  After this it will again be
6909 	 * safe to acquire adaptive mutex locks, or to drop them...
6910 	 */
6911 	if (old_pil != -1) {
6912 		splx(old_pil);
6913 	} else {
6914 		xc_dismissed(cpuset);
6915 	}
6916 
6917 	kpreempt_enable();
6918 
6919 	sfmmu_mlist_reloc_exit(low, high);
6920 
6921 	/*
6922 	 * Postsuspend callbacks should drop any locks held across
6923 	 * the suspend callbacks.  As before, we don't hold the mapping
6924 	 * list lock at this point.. our assumption is that the mapping
6925 	 * list still can't change due to our holding SE_EXCL lock and
6926 	 * there being no unlocked mappings left. Hence the restriction
6927 	 * on calling context to hat_delete_callback()
6928 	 */
6929 	hat_pageprocess_postcallbacks(repl, HAT_POSTUNSUSPEND);
6930 	if (ret != 0) {
6931 		/*
6932 		 * The second presuspend call failed: we got here through
6933 		 * the suspend_fail label above.
6934 		 */
6935 		ASSERT(ret != EIO);
6936 		PAGE_RELOCATE_LOG(target, replacement, ret, cap_cpus);
6937 		kreloc_thread = NULL;
6938 		mutex_exit(&kpr_mutex);
6939 		return (EAGAIN);
6940 	}
6941 
6942 	/*
6943 	 * Now that we're out of the performance critical section we can
6944 	 * take care of updating the hash table, since we still
6945 	 * hold all the pages locked SE_EXCL at this point we
6946 	 * needn't worry about things changing out from under us.
6947 	 */
6948 	tpp = targ;
6949 	rpp = repl;
6950 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6951 
6952 		/*
6953 		 * replace targ with replacement in page_hash table
6954 		 */
6955 		targ = tpp;
6956 		page_relocate_hash(rpp, targ);
6957 
6958 		/*
6959 		 * concatenate target; caller of platform_page_relocate()
6960 		 * expects target to be concatenated after returning.
6961 		 */
6962 		ASSERT(targ->p_next == targ);
6963 		ASSERT(targ->p_prev == targ);
6964 		page_list_concat(&pl, &targ);
6965 	}
6966 
6967 	ASSERT(*target == pl);
6968 	*nrelocp = npages;
6969 	PAGE_RELOCATE_LOG(target, replacement, 0, cap_cpus);
6970 	kreloc_thread = NULL;
6971 	mutex_exit(&kpr_mutex);
6972 	return (0);
6973 }
6974 
6975 /*
6976  * Called when stray pa_hments are found attached to a page which is
6977  * being freed.  Notify the subsystem which attached the pa_hment of
6978  * the error if it registered a suitable handler, else panic.
6979  */
6980 static void
6981 sfmmu_pahment_leaked(struct pa_hment *pahmep)
6982 {
6983 	id_t cb_id = pahmep->cb_id;
6984 
6985 	ASSERT(cb_id >= (id_t)0 && cb_id < sfmmu_cb_nextid);
6986 	if (sfmmu_cb_table[cb_id].errhandler != NULL) {
6987 		if (sfmmu_cb_table[cb_id].errhandler(pahmep->addr, pahmep->len,
6988 		    HAT_CB_ERR_LEAKED, pahmep->pvt) == 0)
6989 			return;		/* non-fatal */
6990 	}
6991 	panic("pa_hment leaked: 0x%p", pahmep);
6992 }
6993 
6994 /*
6995  * Remove all mappings to page 'pp'.
6996  */
6997 int
6998 hat_pageunload(struct page *pp, uint_t forceflag)
6999 {
7000 	struct page *origpp = pp;
7001 	struct sf_hment *sfhme, *tmphme;
7002 	struct hme_blk *hmeblkp;
7003 	kmutex_t *pml;
7004 #ifdef VAC
7005 	kmutex_t *pmtx;
7006 #endif
7007 	cpuset_t cpuset, tset;
7008 	int index, cons;
7009 	int xhme_blks;
7010 	int pa_hments;
7011 
7012 	ASSERT(PAGE_EXCL(pp));
7013 
7014 retry_xhat:
7015 	tmphme = NULL;
7016 	xhme_blks = 0;
7017 	pa_hments = 0;
7018 	CPUSET_ZERO(cpuset);
7019 
7020 	pml = sfmmu_mlist_enter(pp);
7021 
7022 #ifdef VAC
7023 	if (pp->p_kpmref)
7024 		sfmmu_kpm_pageunload(pp);
7025 	ASSERT(!PP_ISMAPPED_KPM(pp));
7026 #endif
7027 
7028 	index = PP_MAPINDEX(pp);
7029 	cons = TTE8K;
7030 retry:
7031 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7032 		tmphme = sfhme->hme_next;
7033 
7034 		if (IS_PAHME(sfhme)) {
7035 			ASSERT(sfhme->hme_data != NULL);
7036 			pa_hments++;
7037 			continue;
7038 		}
7039 
7040 		hmeblkp = sfmmu_hmetohblk(sfhme);
7041 		if (hmeblkp->hblk_xhat_bit) {
7042 			struct xhat_hme_blk *xblk =
7043 			    (struct xhat_hme_blk *)hmeblkp;
7044 
7045 			(void) XHAT_PAGEUNLOAD(xblk->xhat_hme_blk_hat,
7046 			    pp, forceflag, XBLK2PROVBLK(xblk));
7047 
7048 			xhme_blks = 1;
7049 			continue;
7050 		}
7051 
7052 		/*
7053 		 * If there are kernel mappings don't unload them, they will
7054 		 * be suspended.
7055 		 */
7056 		if (forceflag == SFMMU_KERNEL_RELOC && hmeblkp->hblk_lckcnt &&
7057 		    hmeblkp->hblk_tag.htag_id == ksfmmup)
7058 			continue;
7059 
7060 		tset = sfmmu_pageunload(pp, sfhme, cons);
7061 		CPUSET_OR(cpuset, tset);
7062 	}
7063 
7064 	while (index != 0) {
7065 		index = index >> 1;
7066 		if (index != 0)
7067 			cons++;
7068 		if (index & 0x1) {
7069 			/* Go to leading page */
7070 			pp = PP_GROUPLEADER(pp, cons);
7071 			ASSERT(sfmmu_mlist_held(pp));
7072 			goto retry;
7073 		}
7074 	}
7075 
7076 	/*
7077 	 * cpuset may be empty if the page was only mapped by segkpm,
7078 	 * in which case we won't actually cross-trap.
7079 	 */
7080 	xt_sync(cpuset);
7081 
7082 	/*
7083 	 * The page should have no mappings at this point, unless
7084 	 * we were called from hat_page_relocate() in which case we
7085 	 * leave the locked mappings which will be suspended later.
7086 	 */
7087 	ASSERT(!PP_ISMAPPED(origpp) || xhme_blks || pa_hments ||
7088 	    (forceflag == SFMMU_KERNEL_RELOC));
7089 
7090 #ifdef VAC
7091 	if (PP_ISTNC(pp)) {
7092 		if (cons == TTE8K) {
7093 			pmtx = sfmmu_page_enter(pp);
7094 			PP_CLRTNC(pp);
7095 			sfmmu_page_exit(pmtx);
7096 		} else {
7097 			conv_tnc(pp, cons);
7098 		}
7099 	}
7100 #endif	/* VAC */
7101 
7102 	if (pa_hments && forceflag != SFMMU_KERNEL_RELOC) {
7103 		/*
7104 		 * Unlink any pa_hments and free them, calling back
7105 		 * the responsible subsystem to notify it of the error.
7106 		 * This can occur in situations such as drivers leaking
7107 		 * DMA handles: naughty, but common enough that we'd like
7108 		 * to keep the system running rather than bringing it
7109 		 * down with an obscure error like "pa_hment leaked"
7110 		 * which doesn't aid the user in debugging their driver.
7111 		 */
7112 		for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7113 			tmphme = sfhme->hme_next;
7114 			if (IS_PAHME(sfhme)) {
7115 				struct pa_hment *pahmep = sfhme->hme_data;
7116 				sfmmu_pahment_leaked(pahmep);
7117 				HME_SUB(sfhme, pp);
7118 				kmem_cache_free(pa_hment_cache, pahmep);
7119 			}
7120 		}
7121 
7122 		ASSERT(!PP_ISMAPPED(origpp) || xhme_blks);
7123 	}
7124 
7125 	sfmmu_mlist_exit(pml);
7126 
7127 	/*
7128 	 * XHAT may not have finished unloading pages
7129 	 * because some other thread was waiting for
7130 	 * mlist lock and XHAT_PAGEUNLOAD let it do
7131 	 * the job.
7132 	 */
7133 	if (xhme_blks) {
7134 		pp = origpp;
7135 		goto retry_xhat;
7136 	}
7137 
7138 	return (0);
7139 }
7140 
7141 cpuset_t
7142 sfmmu_pageunload(page_t *pp, struct sf_hment *sfhme, int cons)
7143 {
7144 	struct hme_blk *hmeblkp;
7145 	sfmmu_t *sfmmup;
7146 	tte_t tte, ttemod;
7147 #ifdef DEBUG
7148 	tte_t orig_old;
7149 #endif /* DEBUG */
7150 	caddr_t addr;
7151 	int ttesz;
7152 	int ret;
7153 	cpuset_t cpuset;
7154 
7155 	ASSERT(pp != NULL);
7156 	ASSERT(sfmmu_mlist_held(pp));
7157 	ASSERT(!PP_ISKAS(pp));
7158 
7159 	CPUSET_ZERO(cpuset);
7160 
7161 	hmeblkp = sfmmu_hmetohblk(sfhme);
7162 
7163 readtte:
7164 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7165 	if (TTE_IS_VALID(&tte)) {
7166 		sfmmup = hblktosfmmu(hmeblkp);
7167 		ttesz = get_hblk_ttesz(hmeblkp);
7168 		/*
7169 		 * Only unload mappings of 'cons' size.
7170 		 */
7171 		if (ttesz != cons)
7172 			return (cpuset);
7173 
7174 		/*
7175 		 * Note that we have p_mapping lock, but no hash lock here.
7176 		 * hblk_unload() has to have both hash lock AND p_mapping
7177 		 * lock before it tries to modify tte. So, the tte could
7178 		 * not become invalid in the sfmmu_modifytte_try() below.
7179 		 */
7180 		ttemod = tte;
7181 #ifdef DEBUG
7182 		orig_old = tte;
7183 #endif /* DEBUG */
7184 
7185 		TTE_SET_INVALID(&ttemod);
7186 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7187 		if (ret < 0) {
7188 #ifdef DEBUG
7189 			/* only R/M bits can change. */
7190 			chk_tte(&orig_old, &tte, &ttemod, hmeblkp);
7191 #endif /* DEBUG */
7192 			goto readtte;
7193 		}
7194 
7195 		if (ret == 0) {
7196 			panic("pageunload: cas failed?");
7197 		}
7198 
7199 		addr = tte_to_vaddr(hmeblkp, tte);
7200 
7201 		if (hmeblkp->hblk_shared) {
7202 			sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7203 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
7204 			sf_region_t *rgnp;
7205 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7206 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7207 			ASSERT(srdp != NULL);
7208 			rgnp = srdp->srd_hmergnp[rid];
7209 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
7210 			cpuset = sfmmu_rgntlb_demap(addr, rgnp, hmeblkp, 1);
7211 			sfmmu_ttesync(NULL, addr, &tte, pp);
7212 			ASSERT(rgnp->rgn_ttecnt[ttesz] > 0);
7213 			atomic_add_long(&rgnp->rgn_ttecnt[ttesz], -1);
7214 		} else {
7215 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
7216 			atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -1);
7217 
7218 			/*
7219 			 * We need to flush the page from the virtual cache
7220 			 * in order to prevent a virtual cache alias
7221 			 * inconsistency. The particular scenario we need
7222 			 * to worry about is:
7223 			 * Given:  va1 and va2 are two virtual address that
7224 			 * alias and will map the same physical address.
7225 			 * 1.   mapping exists from va1 to pa and data has
7226 			 *	been read into the cache.
7227 			 * 2.   unload va1.
7228 			 * 3.   load va2 and modify data using va2.
7229 			 * 4    unload va2.
7230 			 * 5.   load va1 and reference data.  Unless we flush
7231 			 *	the data cache when we unload we will get
7232 			 *	stale data.
7233 			 * This scenario is taken care of by using virtual
7234 			 * page coloring.
7235 			 */
7236 			if (sfmmup->sfmmu_ismhat) {
7237 				/*
7238 				 * Flush TSBs, TLBs and caches
7239 				 * of every process
7240 				 * sharing this ism segment.
7241 				 */
7242 				sfmmu_hat_lock_all();
7243 				mutex_enter(&ism_mlist_lock);
7244 				kpreempt_disable();
7245 				sfmmu_ismtlbcache_demap(addr, sfmmup, hmeblkp,
7246 				    pp->p_pagenum, CACHE_NO_FLUSH);
7247 				kpreempt_enable();
7248 				mutex_exit(&ism_mlist_lock);
7249 				sfmmu_hat_unlock_all();
7250 				cpuset = cpu_ready_set;
7251 			} else {
7252 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7253 				cpuset = sfmmup->sfmmu_cpusran;
7254 			}
7255 		}
7256 
7257 		/*
7258 		 * Hme_sub has to run after ttesync() and a_rss update.
7259 		 * See hblk_unload().
7260 		 */
7261 		HME_SUB(sfhme, pp);
7262 		membar_stst();
7263 
7264 		/*
7265 		 * We can not make ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
7266 		 * since pteload may have done a HME_ADD() right after
7267 		 * we did the HME_SUB() above. Hmecnt is now maintained
7268 		 * by cas only. no lock guranteed its value. The only
7269 		 * gurantee we have is the hmecnt should not be less than
7270 		 * what it should be so the hblk will not be taken away.
7271 		 * It's also important that we decremented the hmecnt after
7272 		 * we are done with hmeblkp so that this hmeblk won't be
7273 		 * stolen.
7274 		 */
7275 		ASSERT(hmeblkp->hblk_hmecnt > 0);
7276 		ASSERT(hmeblkp->hblk_vcnt > 0);
7277 		atomic_add_16(&hmeblkp->hblk_vcnt, -1);
7278 		atomic_add_16(&hmeblkp->hblk_hmecnt, -1);
7279 		/*
7280 		 * This is bug 4063182.
7281 		 * XXX: fixme
7282 		 * ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
7283 		 *	!hmeblkp->hblk_lckcnt);
7284 		 */
7285 	} else {
7286 		panic("invalid tte? pp %p &tte %p",
7287 		    (void *)pp, (void *)&tte);
7288 	}
7289 
7290 	return (cpuset);
7291 }
7292 
7293 /*
7294  * While relocating a kernel page, this function will move the mappings
7295  * from tpp to dpp and modify any associated data with these mappings.
7296  * It also unsuspends the suspended kernel mapping.
7297  */
7298 static void
7299 hat_pagereload(struct page *tpp, struct page *dpp)
7300 {
7301 	struct sf_hment *sfhme;
7302 	tte_t tte, ttemod;
7303 	int index, cons;
7304 
7305 	ASSERT(getpil() == PIL_MAX);
7306 	ASSERT(sfmmu_mlist_held(tpp));
7307 	ASSERT(sfmmu_mlist_held(dpp));
7308 
7309 	index = PP_MAPINDEX(tpp);
7310 	cons = TTE8K;
7311 
7312 	/* Update real mappings to the page */
7313 retry:
7314 	for (sfhme = tpp->p_mapping; sfhme != NULL; sfhme = sfhme->hme_next) {
7315 		if (IS_PAHME(sfhme))
7316 			continue;
7317 		sfmmu_copytte(&sfhme->hme_tte, &tte);
7318 		ttemod = tte;
7319 
7320 		/*
7321 		 * replace old pfn with new pfn in TTE
7322 		 */
7323 		PFN_TO_TTE(ttemod, dpp->p_pagenum);
7324 
7325 		/*
7326 		 * clear suspend bit
7327 		 */
7328 		ASSERT(TTE_IS_SUSPEND(&ttemod));
7329 		TTE_CLR_SUSPEND(&ttemod);
7330 
7331 		if (sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte) < 0)
7332 			panic("hat_pagereload(): sfmmu_modifytte_try() failed");
7333 
7334 		/*
7335 		 * set hme_page point to new page
7336 		 */
7337 		sfhme->hme_page = dpp;
7338 	}
7339 
7340 	/*
7341 	 * move p_mapping list from old page to new page
7342 	 */
7343 	dpp->p_mapping = tpp->p_mapping;
7344 	tpp->p_mapping = NULL;
7345 	dpp->p_share = tpp->p_share;
7346 	tpp->p_share = 0;
7347 
7348 	while (index != 0) {
7349 		index = index >> 1;
7350 		if (index != 0)
7351 			cons++;
7352 		if (index & 0x1) {
7353 			tpp = PP_GROUPLEADER(tpp, cons);
7354 			dpp = PP_GROUPLEADER(dpp, cons);
7355 			goto retry;
7356 		}
7357 	}
7358 
7359 	curthread->t_flag &= ~T_DONTDTRACE;
7360 	mutex_exit(&kpr_suspendlock);
7361 }
7362 
7363 uint_t
7364 hat_pagesync(struct page *pp, uint_t clearflag)
7365 {
7366 	struct sf_hment *sfhme, *tmphme = NULL;
7367 	struct hme_blk *hmeblkp;
7368 	kmutex_t *pml;
7369 	cpuset_t cpuset, tset;
7370 	int	index, cons;
7371 	extern	ulong_t po_share;
7372 	page_t	*save_pp = pp;
7373 	int	stop_on_sh = 0;
7374 	uint_t	shcnt;
7375 
7376 	CPUSET_ZERO(cpuset);
7377 
7378 	if (PP_ISRO(pp) && (clearflag & HAT_SYNC_STOPON_MOD)) {
7379 		return (PP_GENERIC_ATTR(pp));
7380 	}
7381 
7382 	if ((clearflag & HAT_SYNC_ZERORM) == 0) {
7383 		if ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(pp)) {
7384 			return (PP_GENERIC_ATTR(pp));
7385 		}
7386 		if ((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(pp)) {
7387 			return (PP_GENERIC_ATTR(pp));
7388 		}
7389 		if (clearflag & HAT_SYNC_STOPON_SHARED) {
7390 			if (pp->p_share > po_share) {
7391 				hat_page_setattr(pp, P_REF);
7392 				return (PP_GENERIC_ATTR(pp));
7393 			}
7394 			stop_on_sh = 1;
7395 			shcnt = 0;
7396 		}
7397 	}
7398 
7399 	clearflag &= ~HAT_SYNC_STOPON_SHARED;
7400 	pml = sfmmu_mlist_enter(pp);
7401 	index = PP_MAPINDEX(pp);
7402 	cons = TTE8K;
7403 retry:
7404 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7405 		/*
7406 		 * We need to save the next hment on the list since
7407 		 * it is possible for pagesync to remove an invalid hment
7408 		 * from the list.
7409 		 */
7410 		tmphme = sfhme->hme_next;
7411 		if (IS_PAHME(sfhme))
7412 			continue;
7413 		/*
7414 		 * If we are looking for large mappings and this hme doesn't
7415 		 * reach the range we are seeking, just ignore it.
7416 		 */
7417 		hmeblkp = sfmmu_hmetohblk(sfhme);
7418 		if (hmeblkp->hblk_xhat_bit)
7419 			continue;
7420 
7421 		if (hme_size(sfhme) < cons)
7422 			continue;
7423 
7424 		if (stop_on_sh) {
7425 			if (hmeblkp->hblk_shared) {
7426 				sf_srd_t *srdp = hblktosrd(hmeblkp);
7427 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7428 				sf_region_t *rgnp;
7429 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7430 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7431 				ASSERT(srdp != NULL);
7432 				rgnp = srdp->srd_hmergnp[rid];
7433 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
7434 				    rgnp, rid);
7435 				shcnt += rgnp->rgn_refcnt;
7436 			} else {
7437 				shcnt++;
7438 			}
7439 			if (shcnt > po_share) {
7440 				/*
7441 				 * tell the pager to spare the page this time
7442 				 * around.
7443 				 */
7444 				hat_page_setattr(save_pp, P_REF);
7445 				index = 0;
7446 				break;
7447 			}
7448 		}
7449 		tset = sfmmu_pagesync(pp, sfhme,
7450 		    clearflag & ~HAT_SYNC_STOPON_RM);
7451 		CPUSET_OR(cpuset, tset);
7452 
7453 		/*
7454 		 * If clearflag is HAT_SYNC_DONTZERO, break out as soon
7455 		 * as the "ref" or "mod" is set or share cnt exceeds po_share.
7456 		 */
7457 		if ((clearflag & ~HAT_SYNC_STOPON_RM) == HAT_SYNC_DONTZERO &&
7458 		    (((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(save_pp)) ||
7459 		    ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(save_pp)))) {
7460 			index = 0;
7461 			break;
7462 		}
7463 	}
7464 
7465 	while (index) {
7466 		index = index >> 1;
7467 		cons++;
7468 		if (index & 0x1) {
7469 			/* Go to leading page */
7470 			pp = PP_GROUPLEADER(pp, cons);
7471 			goto retry;
7472 		}
7473 	}
7474 
7475 	xt_sync(cpuset);
7476 	sfmmu_mlist_exit(pml);
7477 	return (PP_GENERIC_ATTR(save_pp));
7478 }
7479 
7480 /*
7481  * Get all the hardware dependent attributes for a page struct
7482  */
7483 static cpuset_t
7484 sfmmu_pagesync(struct page *pp, struct sf_hment *sfhme,
7485 	uint_t clearflag)
7486 {
7487 	caddr_t addr;
7488 	tte_t tte, ttemod;
7489 	struct hme_blk *hmeblkp;
7490 	int ret;
7491 	sfmmu_t *sfmmup;
7492 	cpuset_t cpuset;
7493 
7494 	ASSERT(pp != NULL);
7495 	ASSERT(sfmmu_mlist_held(pp));
7496 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
7497 	    (clearflag == HAT_SYNC_ZERORM));
7498 
7499 	SFMMU_STAT(sf_pagesync);
7500 
7501 	CPUSET_ZERO(cpuset);
7502 
7503 sfmmu_pagesync_retry:
7504 
7505 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7506 	if (TTE_IS_VALID(&tte)) {
7507 		hmeblkp = sfmmu_hmetohblk(sfhme);
7508 		sfmmup = hblktosfmmu(hmeblkp);
7509 		addr = tte_to_vaddr(hmeblkp, tte);
7510 		if (clearflag == HAT_SYNC_ZERORM) {
7511 			ttemod = tte;
7512 			TTE_CLR_RM(&ttemod);
7513 			ret = sfmmu_modifytte_try(&tte, &ttemod,
7514 			    &sfhme->hme_tte);
7515 			if (ret < 0) {
7516 				/*
7517 				 * cas failed and the new value is not what
7518 				 * we want.
7519 				 */
7520 				goto sfmmu_pagesync_retry;
7521 			}
7522 
7523 			if (ret > 0) {
7524 				/* we win the cas */
7525 				if (hmeblkp->hblk_shared) {
7526 					sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7527 					uint_t rid =
7528 					    hmeblkp->hblk_tag.htag_rid;
7529 					sf_region_t *rgnp;
7530 					ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7531 					ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7532 					ASSERT(srdp != NULL);
7533 					rgnp = srdp->srd_hmergnp[rid];
7534 					SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7535 					    srdp, rgnp, rid);
7536 					cpuset = sfmmu_rgntlb_demap(addr,
7537 					    rgnp, hmeblkp, 1);
7538 				} else {
7539 					sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
7540 					    0, 0);
7541 					cpuset = sfmmup->sfmmu_cpusran;
7542 				}
7543 			}
7544 		}
7545 		sfmmu_ttesync(hmeblkp->hblk_shared ? NULL : sfmmup, addr,
7546 		    &tte, pp);
7547 	}
7548 	return (cpuset);
7549 }
7550 
7551 /*
7552  * Remove write permission from a mappings to a page, so that
7553  * we can detect the next modification of it. This requires modifying
7554  * the TTE then invalidating (demap) any TLB entry using that TTE.
7555  * This code is similar to sfmmu_pagesync().
7556  */
7557 static cpuset_t
7558 sfmmu_pageclrwrt(struct page *pp, struct sf_hment *sfhme)
7559 {
7560 	caddr_t addr;
7561 	tte_t tte;
7562 	tte_t ttemod;
7563 	struct hme_blk *hmeblkp;
7564 	int ret;
7565 	sfmmu_t *sfmmup;
7566 	cpuset_t cpuset;
7567 
7568 	ASSERT(pp != NULL);
7569 	ASSERT(sfmmu_mlist_held(pp));
7570 
7571 	CPUSET_ZERO(cpuset);
7572 	SFMMU_STAT(sf_clrwrt);
7573 
7574 retry:
7575 
7576 	sfmmu_copytte(&sfhme->hme_tte, &tte);
7577 	if (TTE_IS_VALID(&tte) && TTE_IS_WRITABLE(&tte)) {
7578 		hmeblkp = sfmmu_hmetohblk(sfhme);
7579 
7580 		/*
7581 		 * xhat mappings should never be to a VMODSORT page.
7582 		 */
7583 		ASSERT(hmeblkp->hblk_xhat_bit == 0);
7584 
7585 		sfmmup = hblktosfmmu(hmeblkp);
7586 		addr = tte_to_vaddr(hmeblkp, tte);
7587 
7588 		ttemod = tte;
7589 		TTE_CLR_WRT(&ttemod);
7590 		TTE_CLR_MOD(&ttemod);
7591 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
7592 
7593 		/*
7594 		 * if cas failed and the new value is not what
7595 		 * we want retry
7596 		 */
7597 		if (ret < 0)
7598 			goto retry;
7599 
7600 		/* we win the cas */
7601 		if (ret > 0) {
7602 			if (hmeblkp->hblk_shared) {
7603 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
7604 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
7605 				sf_region_t *rgnp;
7606 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
7607 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
7608 				ASSERT(srdp != NULL);
7609 				rgnp = srdp->srd_hmergnp[rid];
7610 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
7611 				    srdp, rgnp, rid);
7612 				cpuset = sfmmu_rgntlb_demap(addr,
7613 				    rgnp, hmeblkp, 1);
7614 			} else {
7615 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
7616 				cpuset = sfmmup->sfmmu_cpusran;
7617 			}
7618 		}
7619 	}
7620 
7621 	return (cpuset);
7622 }
7623 
7624 /*
7625  * Walk all mappings of a page, removing write permission and clearing the
7626  * ref/mod bits. This code is similar to hat_pagesync()
7627  */
7628 static void
7629 hat_page_clrwrt(page_t *pp)
7630 {
7631 	struct sf_hment *sfhme;
7632 	struct sf_hment *tmphme = NULL;
7633 	kmutex_t *pml;
7634 	cpuset_t cpuset;
7635 	cpuset_t tset;
7636 	int	index;
7637 	int	 cons;
7638 
7639 	CPUSET_ZERO(cpuset);
7640 
7641 	pml = sfmmu_mlist_enter(pp);
7642 	index = PP_MAPINDEX(pp);
7643 	cons = TTE8K;
7644 retry:
7645 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
7646 		tmphme = sfhme->hme_next;
7647 
7648 		/*
7649 		 * If we are looking for large mappings and this hme doesn't
7650 		 * reach the range we are seeking, just ignore its.
7651 		 */
7652 
7653 		if (hme_size(sfhme) < cons)
7654 			continue;
7655 
7656 		tset = sfmmu_pageclrwrt(pp, sfhme);
7657 		CPUSET_OR(cpuset, tset);
7658 	}
7659 
7660 	while (index) {
7661 		index = index >> 1;
7662 		cons++;
7663 		if (index & 0x1) {
7664 			/* Go to leading page */
7665 			pp = PP_GROUPLEADER(pp, cons);
7666 			goto retry;
7667 		}
7668 	}
7669 
7670 	xt_sync(cpuset);
7671 	sfmmu_mlist_exit(pml);
7672 }
7673 
7674 /*
7675  * Set the given REF/MOD/RO bits for the given page.
7676  * For a vnode with a sorted v_pages list, we need to change
7677  * the attributes and the v_pages list together under page_vnode_mutex.
7678  */
7679 void
7680 hat_page_setattr(page_t *pp, uint_t flag)
7681 {
7682 	vnode_t		*vp = pp->p_vnode;
7683 	page_t		**listp;
7684 	kmutex_t	*pmtx;
7685 	kmutex_t	*vphm = NULL;
7686 	int		noshuffle;
7687 
7688 	noshuffle = flag & P_NSH;
7689 	flag &= ~P_NSH;
7690 
7691 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7692 
7693 	/*
7694 	 * nothing to do if attribute already set
7695 	 */
7696 	if ((pp->p_nrm & flag) == flag)
7697 		return;
7698 
7699 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp) &&
7700 	    !noshuffle) {
7701 		vphm = page_vnode_mutex(vp);
7702 		mutex_enter(vphm);
7703 	}
7704 
7705 	pmtx = sfmmu_page_enter(pp);
7706 	pp->p_nrm |= flag;
7707 	sfmmu_page_exit(pmtx);
7708 
7709 	if (vphm != NULL) {
7710 		/*
7711 		 * Some File Systems examine v_pages for NULL w/o
7712 		 * grabbing the vphm mutex. Must not let it become NULL when
7713 		 * pp is the only page on the list.
7714 		 */
7715 		if (pp->p_vpnext != pp) {
7716 			page_vpsub(&vp->v_pages, pp);
7717 			if (vp->v_pages != NULL)
7718 				listp = &vp->v_pages->p_vpprev->p_vpnext;
7719 			else
7720 				listp = &vp->v_pages;
7721 			page_vpadd(listp, pp);
7722 		}
7723 		mutex_exit(vphm);
7724 	}
7725 }
7726 
7727 void
7728 hat_page_clrattr(page_t *pp, uint_t flag)
7729 {
7730 	vnode_t		*vp = pp->p_vnode;
7731 	kmutex_t	*pmtx;
7732 
7733 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7734 
7735 	pmtx = sfmmu_page_enter(pp);
7736 
7737 	/*
7738 	 * Caller is expected to hold page's io lock for VMODSORT to work
7739 	 * correctly with pvn_vplist_dirty() and pvn_getdirty() when mod
7740 	 * bit is cleared.
7741 	 * We don't have assert to avoid tripping some existing third party
7742 	 * code. The dirty page is moved back to top of the v_page list
7743 	 * after IO is done in pvn_write_done().
7744 	 */
7745 	pp->p_nrm &= ~flag;
7746 	sfmmu_page_exit(pmtx);
7747 
7748 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) {
7749 
7750 		/*
7751 		 * VMODSORT works by removing write permissions and getting
7752 		 * a fault when a page is made dirty. At this point
7753 		 * we need to remove write permission from all mappings
7754 		 * to this page.
7755 		 */
7756 		hat_page_clrwrt(pp);
7757 	}
7758 }
7759 
7760 uint_t
7761 hat_page_getattr(page_t *pp, uint_t flag)
7762 {
7763 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
7764 	return ((uint_t)(pp->p_nrm & flag));
7765 }
7766 
7767 /*
7768  * DEBUG kernels: verify that a kernel va<->pa translation
7769  * is safe by checking the underlying page_t is in a page
7770  * relocation-safe state.
7771  */
7772 #ifdef	DEBUG
7773 void
7774 sfmmu_check_kpfn(pfn_t pfn)
7775 {
7776 	page_t *pp;
7777 	int index, cons;
7778 
7779 	if (hat_check_vtop == 0)
7780 		return;
7781 
7782 	if (hat_kpr_enabled == 0 || kvseg.s_base == NULL || panicstr)
7783 		return;
7784 
7785 	pp = page_numtopp_nolock(pfn);
7786 	if (!pp)
7787 		return;
7788 
7789 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7790 		return;
7791 
7792 	/*
7793 	 * Handed a large kernel page, we dig up the root page since we
7794 	 * know the root page might have the lock also.
7795 	 */
7796 	if (pp->p_szc != 0) {
7797 		index = PP_MAPINDEX(pp);
7798 		cons = TTE8K;
7799 again:
7800 		while (index != 0) {
7801 			index >>= 1;
7802 			if (index != 0)
7803 				cons++;
7804 			if (index & 0x1) {
7805 				pp = PP_GROUPLEADER(pp, cons);
7806 				goto again;
7807 			}
7808 		}
7809 	}
7810 
7811 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
7812 		return;
7813 
7814 	/*
7815 	 * Pages need to be locked or allocated "permanent" (either from
7816 	 * static_arena arena or explicitly setting PG_NORELOC when calling
7817 	 * page_create_va()) for VA->PA translations to be valid.
7818 	 */
7819 	if (!PP_ISNORELOC(pp))
7820 		panic("Illegal VA->PA translation, pp 0x%p not permanent", pp);
7821 	else
7822 		panic("Illegal VA->PA translation, pp 0x%p not locked", pp);
7823 }
7824 #endif	/* DEBUG */
7825 
7826 /*
7827  * Returns a page frame number for a given virtual address.
7828  * Returns PFN_INVALID to indicate an invalid mapping
7829  */
7830 pfn_t
7831 hat_getpfnum(struct hat *hat, caddr_t addr)
7832 {
7833 	pfn_t pfn;
7834 	tte_t tte;
7835 
7836 	/*
7837 	 * We would like to
7838 	 * ASSERT(AS_LOCK_HELD(as, &as->a_lock));
7839 	 * but we can't because the iommu driver will call this
7840 	 * routine at interrupt time and it can't grab the as lock
7841 	 * or it will deadlock: A thread could have the as lock
7842 	 * and be waiting for io.  The io can't complete
7843 	 * because the interrupt thread is blocked trying to grab
7844 	 * the as lock.
7845 	 */
7846 
7847 	ASSERT(hat->sfmmu_xhat_provider == NULL);
7848 
7849 	if (hat == ksfmmup) {
7850 		if (IS_KMEM_VA_LARGEPAGE(addr)) {
7851 			ASSERT(segkmem_lpszc > 0);
7852 			pfn = sfmmu_kvaszc2pfn(addr, segkmem_lpszc);
7853 			if (pfn != PFN_INVALID) {
7854 				sfmmu_check_kpfn(pfn);
7855 				return (pfn);
7856 			}
7857 		} else if (segkpm && IS_KPM_ADDR(addr)) {
7858 			return (sfmmu_kpm_vatopfn(addr));
7859 		}
7860 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
7861 		    == PFN_SUSPENDED) {
7862 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
7863 		}
7864 		sfmmu_check_kpfn(pfn);
7865 		return (pfn);
7866 	} else {
7867 		return (sfmmu_uvatopfn(addr, hat, NULL));
7868 	}
7869 }
7870 
7871 /*
7872  * hat_getkpfnum() is an obsolete DDI routine, and its use is discouraged.
7873  * Use hat_getpfnum(kas.a_hat, ...) instead.
7874  *
7875  * We'd like to return PFN_INVALID if the mappings have underlying page_t's
7876  * but can't right now due to the fact that some software has grown to use
7877  * this interface incorrectly. So for now when the interface is misused,
7878  * return a warning to the user that in the future it won't work in the
7879  * way they're abusing it, and carry on (after disabling page relocation).
7880  */
7881 pfn_t
7882 hat_getkpfnum(caddr_t addr)
7883 {
7884 	pfn_t pfn;
7885 	tte_t tte;
7886 	int badcaller = 0;
7887 	extern int segkmem_reloc;
7888 
7889 	if (segkpm && IS_KPM_ADDR(addr)) {
7890 		badcaller = 1;
7891 		pfn = sfmmu_kpm_vatopfn(addr);
7892 	} else {
7893 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
7894 		    == PFN_SUSPENDED) {
7895 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
7896 		}
7897 		badcaller = pf_is_memory(pfn);
7898 	}
7899 
7900 	if (badcaller) {
7901 		/*
7902 		 * We can't return PFN_INVALID or the caller may panic
7903 		 * or corrupt the system.  The only alternative is to
7904 		 * disable page relocation at this point for all kernel
7905 		 * memory.  This will impact any callers of page_relocate()
7906 		 * such as FMA or DR.
7907 		 *
7908 		 * RFE: Add junk here to spit out an ereport so the sysadmin
7909 		 * can be advised that he should upgrade his device driver
7910 		 * so that this doesn't happen.
7911 		 */
7912 		hat_getkpfnum_badcall(caller());
7913 		if (hat_kpr_enabled && segkmem_reloc) {
7914 			hat_kpr_enabled = 0;
7915 			segkmem_reloc = 0;
7916 			cmn_err(CE_WARN, "Kernel Page Relocation is DISABLED");
7917 		}
7918 	}
7919 	return (pfn);
7920 }
7921 
7922 /*
7923  * This routine will return both pfn and tte for the vaddr.
7924  */
7925 static pfn_t
7926 sfmmu_uvatopfn(caddr_t vaddr, struct hat *sfmmup, tte_t *ttep)
7927 {
7928 	struct hmehash_bucket *hmebp;
7929 	hmeblk_tag hblktag;
7930 	int hmeshift, hashno = 1;
7931 	struct hme_blk *hmeblkp = NULL;
7932 	tte_t tte;
7933 
7934 	struct sf_hment *sfhmep;
7935 	pfn_t pfn;
7936 
7937 	/* support for ISM */
7938 	ism_map_t	*ism_map;
7939 	ism_blk_t	*ism_blkp;
7940 	int		i;
7941 	sfmmu_t *ism_hatid = NULL;
7942 	sfmmu_t *locked_hatid = NULL;
7943 	sfmmu_t	*sv_sfmmup = sfmmup;
7944 	caddr_t	sv_vaddr = vaddr;
7945 	sf_srd_t *srdp;
7946 
7947 	if (ttep == NULL) {
7948 		ttep = &tte;
7949 	} else {
7950 		ttep->ll = 0;
7951 	}
7952 
7953 	ASSERT(sfmmup != ksfmmup);
7954 	SFMMU_STAT(sf_user_vtop);
7955 	/*
7956 	 * Set ism_hatid if vaddr falls in a ISM segment.
7957 	 */
7958 	ism_blkp = sfmmup->sfmmu_iblk;
7959 	if (ism_blkp != NULL) {
7960 		sfmmu_ismhat_enter(sfmmup, 0);
7961 		locked_hatid = sfmmup;
7962 	}
7963 	while (ism_blkp != NULL && ism_hatid == NULL) {
7964 		ism_map = ism_blkp->iblk_maps;
7965 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
7966 			if (vaddr >= ism_start(ism_map[i]) &&
7967 			    vaddr < ism_end(ism_map[i])) {
7968 				sfmmup = ism_hatid = ism_map[i].imap_ismhat;
7969 				vaddr = (caddr_t)(vaddr -
7970 				    ism_start(ism_map[i]));
7971 				break;
7972 			}
7973 		}
7974 		ism_blkp = ism_blkp->iblk_next;
7975 	}
7976 	if (locked_hatid) {
7977 		sfmmu_ismhat_exit(locked_hatid, 0);
7978 	}
7979 
7980 	hblktag.htag_id = sfmmup;
7981 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
7982 	do {
7983 		hmeshift = HME_HASH_SHIFT(hashno);
7984 		hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
7985 		hblktag.htag_rehash = hashno;
7986 		hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
7987 
7988 		SFMMU_HASH_LOCK(hmebp);
7989 
7990 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
7991 		if (hmeblkp != NULL) {
7992 			ASSERT(!hmeblkp->hblk_shared);
7993 			HBLKTOHME(sfhmep, hmeblkp, vaddr);
7994 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
7995 			SFMMU_HASH_UNLOCK(hmebp);
7996 			if (TTE_IS_VALID(ttep)) {
7997 				pfn = TTE_TO_PFN(vaddr, ttep);
7998 				return (pfn);
7999 			}
8000 			break;
8001 		}
8002 		SFMMU_HASH_UNLOCK(hmebp);
8003 		hashno++;
8004 	} while (HME_REHASH(sfmmup) && (hashno <= mmu_hashcnt));
8005 
8006 	if (SF_HMERGNMAP_ISNULL(sv_sfmmup)) {
8007 		return (PFN_INVALID);
8008 	}
8009 	srdp = sv_sfmmup->sfmmu_srdp;
8010 	ASSERT(srdp != NULL);
8011 	ASSERT(srdp->srd_refcnt != 0);
8012 	hblktag.htag_id = srdp;
8013 	hashno = 1;
8014 	do {
8015 		hmeshift = HME_HASH_SHIFT(hashno);
8016 		hblktag.htag_bspage = HME_HASH_BSPAGE(sv_vaddr, hmeshift);
8017 		hblktag.htag_rehash = hashno;
8018 		hmebp = HME_HASH_FUNCTION(srdp, sv_vaddr, hmeshift);
8019 
8020 		SFMMU_HASH_LOCK(hmebp);
8021 		for (hmeblkp = hmebp->hmeblkp; hmeblkp != NULL;
8022 		    hmeblkp = hmeblkp->hblk_next) {
8023 			uint_t rid;
8024 			sf_region_t *rgnp;
8025 			caddr_t rsaddr;
8026 			caddr_t readdr;
8027 
8028 			if (!HTAGS_EQ_SHME(hmeblkp->hblk_tag, hblktag,
8029 			    sv_sfmmup->sfmmu_hmeregion_map)) {
8030 				continue;
8031 			}
8032 			ASSERT(hmeblkp->hblk_shared);
8033 			rid = hmeblkp->hblk_tag.htag_rid;
8034 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8035 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8036 			rgnp = srdp->srd_hmergnp[rid];
8037 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
8038 			HBLKTOHME(sfhmep, hmeblkp, sv_vaddr);
8039 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
8040 			rsaddr = rgnp->rgn_saddr;
8041 			readdr = rsaddr + rgnp->rgn_size;
8042 #ifdef DEBUG
8043 			if (TTE_IS_VALID(ttep) ||
8044 			    get_hblk_ttesz(hmeblkp) > TTE8K) {
8045 				caddr_t eva = tte_to_evaddr(hmeblkp, ttep);
8046 				ASSERT(eva > sv_vaddr);
8047 				ASSERT(sv_vaddr >= rsaddr);
8048 				ASSERT(sv_vaddr < readdr);
8049 				ASSERT(eva <= readdr);
8050 			}
8051 #endif /* DEBUG */
8052 			/*
8053 			 * Continue the search if we
8054 			 * found an invalid 8K tte outside of the area
8055 			 * covered by this hmeblk's region.
8056 			 */
8057 			if (TTE_IS_VALID(ttep)) {
8058 				SFMMU_HASH_UNLOCK(hmebp);
8059 				pfn = TTE_TO_PFN(sv_vaddr, ttep);
8060 				return (pfn);
8061 			} else if (get_hblk_ttesz(hmeblkp) > TTE8K ||
8062 			    (sv_vaddr >= rsaddr && sv_vaddr < readdr)) {
8063 				SFMMU_HASH_UNLOCK(hmebp);
8064 				pfn = PFN_INVALID;
8065 				return (pfn);
8066 			}
8067 		}
8068 		SFMMU_HASH_UNLOCK(hmebp);
8069 		hashno++;
8070 	} while (hashno <= mmu_hashcnt);
8071 	return (PFN_INVALID);
8072 }
8073 
8074 
8075 /*
8076  * For compatability with AT&T and later optimizations
8077  */
8078 /* ARGSUSED */
8079 void
8080 hat_map(struct hat *hat, caddr_t addr, size_t len, uint_t flags)
8081 {
8082 	ASSERT(hat != NULL);
8083 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8084 }
8085 
8086 /*
8087  * Return the number of mappings to a particular page.  This number is an
8088  * approximation of the number of people sharing the page.
8089  *
8090  * shared hmeblks or ism hmeblks are counted as 1 mapping here.
8091  * hat_page_checkshare() can be used to compare threshold to share
8092  * count that reflects the number of region sharers albeit at higher cost.
8093  */
8094 ulong_t
8095 hat_page_getshare(page_t *pp)
8096 {
8097 	page_t *spp = pp;	/* start page */
8098 	kmutex_t *pml;
8099 	ulong_t	cnt;
8100 	int index, sz = TTE64K;
8101 
8102 	/*
8103 	 * We need to grab the mlist lock to make sure any outstanding
8104 	 * load/unloads complete.  Otherwise we could return zero
8105 	 * even though the unload(s) hasn't finished yet.
8106 	 */
8107 	pml = sfmmu_mlist_enter(spp);
8108 	cnt = spp->p_share;
8109 
8110 #ifdef VAC
8111 	if (kpm_enable)
8112 		cnt += spp->p_kpmref;
8113 #endif
8114 
8115 	/*
8116 	 * If we have any large mappings, we count the number of
8117 	 * mappings that this large page is part of.
8118 	 */
8119 	index = PP_MAPINDEX(spp);
8120 	index >>= 1;
8121 	while (index) {
8122 		pp = PP_GROUPLEADER(spp, sz);
8123 		if ((index & 0x1) && pp != spp) {
8124 			cnt += pp->p_share;
8125 			spp = pp;
8126 		}
8127 		index >>= 1;
8128 		sz++;
8129 	}
8130 	sfmmu_mlist_exit(pml);
8131 	return (cnt);
8132 }
8133 
8134 /*
8135  * Return 1 if the number of mappings exceeds sh_thresh. Return 0
8136  * otherwise. Count shared hmeblks by region's refcnt.
8137  */
8138 int
8139 hat_page_checkshare(page_t *pp, ulong_t sh_thresh)
8140 {
8141 	kmutex_t *pml;
8142 	ulong_t	cnt = 0;
8143 	int index, sz = TTE8K;
8144 	struct sf_hment *sfhme, *tmphme = NULL;
8145 	struct hme_blk *hmeblkp;
8146 
8147 	pml = sfmmu_mlist_enter(pp);
8148 
8149 	if (kpm_enable)
8150 		cnt = pp->p_kpmref;
8151 
8152 	if (pp->p_share + cnt > sh_thresh) {
8153 		sfmmu_mlist_exit(pml);
8154 		return (1);
8155 	}
8156 
8157 	index = PP_MAPINDEX(pp);
8158 
8159 again:
8160 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
8161 		tmphme = sfhme->hme_next;
8162 		if (IS_PAHME(sfhme)) {
8163 			continue;
8164 		}
8165 
8166 		hmeblkp = sfmmu_hmetohblk(sfhme);
8167 		if (hmeblkp->hblk_xhat_bit) {
8168 			cnt++;
8169 			if (cnt > sh_thresh) {
8170 				sfmmu_mlist_exit(pml);
8171 				return (1);
8172 			}
8173 			continue;
8174 		}
8175 		if (hme_size(sfhme) != sz) {
8176 			continue;
8177 		}
8178 
8179 		if (hmeblkp->hblk_shared) {
8180 			sf_srd_t *srdp = hblktosrd(hmeblkp);
8181 			uint_t rid = hmeblkp->hblk_tag.htag_rid;
8182 			sf_region_t *rgnp;
8183 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
8184 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
8185 			ASSERT(srdp != NULL);
8186 			rgnp = srdp->srd_hmergnp[rid];
8187 			SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp,
8188 			    rgnp, rid);
8189 			cnt += rgnp->rgn_refcnt;
8190 		} else {
8191 			cnt++;
8192 		}
8193 		if (cnt > sh_thresh) {
8194 			sfmmu_mlist_exit(pml);
8195 			return (1);
8196 		}
8197 	}
8198 
8199 	index >>= 1;
8200 	sz++;
8201 	while (index) {
8202 		pp = PP_GROUPLEADER(pp, sz);
8203 		ASSERT(sfmmu_mlist_held(pp));
8204 		if (index & 0x1) {
8205 			goto again;
8206 		}
8207 		index >>= 1;
8208 		sz++;
8209 	}
8210 	sfmmu_mlist_exit(pml);
8211 	return (0);
8212 }
8213 
8214 /*
8215  * Unload all large mappings to the pp and reset the p_szc field of every
8216  * constituent page according to the remaining mappings.
8217  *
8218  * pp must be locked SE_EXCL. Even though no other constituent pages are
8219  * locked it's legal to unload the large mappings to the pp because all
8220  * constituent pages of large locked mappings have to be locked SE_SHARED.
8221  * This means if we have SE_EXCL lock on one of constituent pages none of the
8222  * large mappings to pp are locked.
8223  *
8224  * Decrease p_szc field starting from the last constituent page and ending
8225  * with the root page. This method is used because other threads rely on the
8226  * root's p_szc to find the lock to syncronize on. After a root page_t's p_szc
8227  * is demoted then other threads will succeed in sfmmu_mlspl_enter(). This
8228  * ensures that p_szc changes of the constituent pages appears atomic for all
8229  * threads that use sfmmu_mlspl_enter() to examine p_szc field.
8230  *
8231  * This mechanism is only used for file system pages where it's not always
8232  * possible to get SE_EXCL locks on all constituent pages to demote the size
8233  * code (as is done for anonymous or kernel large pages).
8234  *
8235  * See more comments in front of sfmmu_mlspl_enter().
8236  */
8237 void
8238 hat_page_demote(page_t *pp)
8239 {
8240 	int index;
8241 	int sz;
8242 	cpuset_t cpuset;
8243 	int sync = 0;
8244 	page_t *rootpp;
8245 	struct sf_hment *sfhme;
8246 	struct sf_hment *tmphme = NULL;
8247 	struct hme_blk *hmeblkp;
8248 	uint_t pszc;
8249 	page_t *lastpp;
8250 	cpuset_t tset;
8251 	pgcnt_t npgs;
8252 	kmutex_t *pml;
8253 	kmutex_t *pmtx = NULL;
8254 
8255 	ASSERT(PAGE_EXCL(pp));
8256 	ASSERT(!PP_ISFREE(pp));
8257 	ASSERT(!PP_ISKAS(pp));
8258 	ASSERT(page_szc_lock_assert(pp));
8259 	pml = sfmmu_mlist_enter(pp);
8260 
8261 	pszc = pp->p_szc;
8262 	if (pszc == 0) {
8263 		goto out;
8264 	}
8265 
8266 	index = PP_MAPINDEX(pp) >> 1;
8267 
8268 	if (index) {
8269 		CPUSET_ZERO(cpuset);
8270 		sz = TTE64K;
8271 		sync = 1;
8272 	}
8273 
8274 	while (index) {
8275 		if (!(index & 0x1)) {
8276 			index >>= 1;
8277 			sz++;
8278 			continue;
8279 		}
8280 		ASSERT(sz <= pszc);
8281 		rootpp = PP_GROUPLEADER(pp, sz);
8282 		for (sfhme = rootpp->p_mapping; sfhme; sfhme = tmphme) {
8283 			tmphme = sfhme->hme_next;
8284 			ASSERT(!IS_PAHME(sfhme));
8285 			hmeblkp = sfmmu_hmetohblk(sfhme);
8286 			if (hme_size(sfhme) != sz) {
8287 				continue;
8288 			}
8289 			if (hmeblkp->hblk_xhat_bit) {
8290 				cmn_err(CE_PANIC,
8291 				    "hat_page_demote: xhat hmeblk");
8292 			}
8293 			tset = sfmmu_pageunload(rootpp, sfhme, sz);
8294 			CPUSET_OR(cpuset, tset);
8295 		}
8296 		if (index >>= 1) {
8297 			sz++;
8298 		}
8299 	}
8300 
8301 	ASSERT(!PP_ISMAPPED_LARGE(pp));
8302 
8303 	if (sync) {
8304 		xt_sync(cpuset);
8305 #ifdef VAC
8306 		if (PP_ISTNC(pp)) {
8307 			conv_tnc(rootpp, sz);
8308 		}
8309 #endif	/* VAC */
8310 	}
8311 
8312 	pmtx = sfmmu_page_enter(pp);
8313 
8314 	ASSERT(pp->p_szc == pszc);
8315 	rootpp = PP_PAGEROOT(pp);
8316 	ASSERT(rootpp->p_szc == pszc);
8317 	lastpp = PP_PAGENEXT_N(rootpp, TTEPAGES(pszc) - 1);
8318 
8319 	while (lastpp != rootpp) {
8320 		sz = PP_MAPINDEX(lastpp) ? fnd_mapping_sz(lastpp) : 0;
8321 		ASSERT(sz < pszc);
8322 		npgs = (sz == 0) ? 1 : TTEPAGES(sz);
8323 		ASSERT(P2PHASE(lastpp->p_pagenum, npgs) == npgs - 1);
8324 		while (--npgs > 0) {
8325 			lastpp->p_szc = (uchar_t)sz;
8326 			lastpp = PP_PAGEPREV(lastpp);
8327 		}
8328 		if (sz) {
8329 			/*
8330 			 * make sure before current root's pszc
8331 			 * is updated all updates to constituent pages pszc
8332 			 * fields are globally visible.
8333 			 */
8334 			membar_producer();
8335 		}
8336 		lastpp->p_szc = sz;
8337 		ASSERT(IS_P2ALIGNED(lastpp->p_pagenum, TTEPAGES(sz)));
8338 		if (lastpp != rootpp) {
8339 			lastpp = PP_PAGEPREV(lastpp);
8340 		}
8341 	}
8342 	if (sz == 0) {
8343 		/* the loop above doesn't cover this case */
8344 		rootpp->p_szc = 0;
8345 	}
8346 out:
8347 	ASSERT(pp->p_szc == 0);
8348 	if (pmtx != NULL) {
8349 		sfmmu_page_exit(pmtx);
8350 	}
8351 	sfmmu_mlist_exit(pml);
8352 }
8353 
8354 /*
8355  * Refresh the HAT ismttecnt[] element for size szc.
8356  * Caller must have set ISM busy flag to prevent mapping
8357  * lists from changing while we're traversing them.
8358  */
8359 pgcnt_t
8360 ism_tsb_entries(sfmmu_t *sfmmup, int szc)
8361 {
8362 	ism_blk_t	*ism_blkp = sfmmup->sfmmu_iblk;
8363 	ism_map_t	*ism_map;
8364 	pgcnt_t		npgs = 0;
8365 	pgcnt_t		npgs_scd = 0;
8366 	int		j;
8367 	sf_scd_t	*scdp;
8368 	uchar_t		rid;
8369 
8370 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
8371 	scdp = sfmmup->sfmmu_scdp;
8372 
8373 	for (; ism_blkp != NULL; ism_blkp = ism_blkp->iblk_next) {
8374 		ism_map = ism_blkp->iblk_maps;
8375 		for (j = 0; ism_map[j].imap_ismhat && j < ISM_MAP_SLOTS; j++) {
8376 			rid = ism_map[j].imap_rid;
8377 			ASSERT(rid == SFMMU_INVALID_ISMRID ||
8378 			    rid < sfmmup->sfmmu_srdp->srd_next_ismrid);
8379 
8380 			if (scdp != NULL && rid != SFMMU_INVALID_ISMRID &&
8381 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
8382 				/* ISM is in sfmmup's SCD */
8383 				npgs_scd +=
8384 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8385 			} else {
8386 				/* ISMs is not in SCD */
8387 				npgs +=
8388 				    ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
8389 			}
8390 		}
8391 	}
8392 	sfmmup->sfmmu_ismttecnt[szc] = npgs;
8393 	sfmmup->sfmmu_scdismttecnt[szc] = npgs_scd;
8394 	return (npgs);
8395 }
8396 
8397 /*
8398  * Yield the memory claim requirement for an address space.
8399  *
8400  * This is currently implemented as the number of bytes that have active
8401  * hardware translations that have page structures.  Therefore, it can
8402  * underestimate the traditional resident set size, eg, if the
8403  * physical page is present and the hardware translation is missing;
8404  * and it can overestimate the rss, eg, if there are active
8405  * translations to a frame buffer with page structs.
8406  * Also, it does not take sharing into account.
8407  *
8408  * Note that we don't acquire locks here since this function is most often
8409  * called from the clock thread.
8410  */
8411 size_t
8412 hat_get_mapped_size(struct hat *hat)
8413 {
8414 	size_t		assize = 0;
8415 	int 		i;
8416 
8417 	if (hat == NULL)
8418 		return (0);
8419 
8420 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8421 
8422 	for (i = 0; i < mmu_page_sizes; i++)
8423 		assize += ((pgcnt_t)hat->sfmmu_ttecnt[i] +
8424 		    (pgcnt_t)hat->sfmmu_scdrttecnt[i]) * TTEBYTES(i);
8425 
8426 	if (hat->sfmmu_iblk == NULL)
8427 		return (assize);
8428 
8429 	for (i = 0; i < mmu_page_sizes; i++)
8430 		assize += ((pgcnt_t)hat->sfmmu_ismttecnt[i] +
8431 		    (pgcnt_t)hat->sfmmu_scdismttecnt[i]) * TTEBYTES(i);
8432 
8433 	return (assize);
8434 }
8435 
8436 int
8437 hat_stats_enable(struct hat *hat)
8438 {
8439 	hatlock_t	*hatlockp;
8440 
8441 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8442 
8443 	hatlockp = sfmmu_hat_enter(hat);
8444 	hat->sfmmu_rmstat++;
8445 	sfmmu_hat_exit(hatlockp);
8446 	return (1);
8447 }
8448 
8449 void
8450 hat_stats_disable(struct hat *hat)
8451 {
8452 	hatlock_t	*hatlockp;
8453 
8454 	ASSERT(hat->sfmmu_xhat_provider == NULL);
8455 
8456 	hatlockp = sfmmu_hat_enter(hat);
8457 	hat->sfmmu_rmstat--;
8458 	sfmmu_hat_exit(hatlockp);
8459 }
8460 
8461 /*
8462  * Routines for entering or removing  ourselves from the
8463  * ism_hat's mapping list. This is used for both private and
8464  * SCD hats.
8465  */
8466 static void
8467 iment_add(struct ism_ment *iment,  struct hat *ism_hat)
8468 {
8469 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8470 
8471 	iment->iment_prev = NULL;
8472 	iment->iment_next = ism_hat->sfmmu_iment;
8473 	if (ism_hat->sfmmu_iment) {
8474 		ism_hat->sfmmu_iment->iment_prev = iment;
8475 	}
8476 	ism_hat->sfmmu_iment = iment;
8477 }
8478 
8479 static void
8480 iment_sub(struct ism_ment *iment, struct hat *ism_hat)
8481 {
8482 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
8483 
8484 	if (ism_hat->sfmmu_iment == NULL) {
8485 		panic("ism map entry remove - no entries");
8486 	}
8487 
8488 	if (iment->iment_prev) {
8489 		ASSERT(ism_hat->sfmmu_iment != iment);
8490 		iment->iment_prev->iment_next = iment->iment_next;
8491 	} else {
8492 		ASSERT(ism_hat->sfmmu_iment == iment);
8493 		ism_hat->sfmmu_iment = iment->iment_next;
8494 	}
8495 
8496 	if (iment->iment_next) {
8497 		iment->iment_next->iment_prev = iment->iment_prev;
8498 	}
8499 
8500 	/*
8501 	 * zero out the entry
8502 	 */
8503 	iment->iment_next = NULL;
8504 	iment->iment_prev = NULL;
8505 	iment->iment_hat =  NULL;
8506 }
8507 
8508 /*
8509  * Hat_share()/unshare() return an (non-zero) error
8510  * when saddr and daddr are not properly aligned.
8511  *
8512  * The top level mapping element determines the alignment
8513  * requirement for saddr and daddr, depending on different
8514  * architectures.
8515  *
8516  * When hat_share()/unshare() are not supported,
8517  * HATOP_SHARE()/UNSHARE() return 0
8518  */
8519 int
8520 hat_share(struct hat *sfmmup, caddr_t addr,
8521 	struct hat *ism_hatid, caddr_t sptaddr, size_t len, uint_t ismszc)
8522 {
8523 	ism_blk_t	*ism_blkp;
8524 	ism_blk_t	*new_iblk;
8525 	ism_map_t 	*ism_map;
8526 	ism_ment_t	*ism_ment;
8527 	int		i, added;
8528 	hatlock_t	*hatlockp;
8529 	int		reload_mmu = 0;
8530 	uint_t		ismshift = page_get_shift(ismszc);
8531 	size_t		ismpgsz = page_get_pagesize(ismszc);
8532 	uint_t		ismmask = (uint_t)ismpgsz - 1;
8533 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8534 	ushort_t	ismhatflag;
8535 	hat_region_cookie_t rcookie;
8536 	sf_scd_t	*old_scdp;
8537 
8538 #ifdef DEBUG
8539 	caddr_t		eaddr = addr + len;
8540 #endif /* DEBUG */
8541 
8542 	ASSERT(ism_hatid != NULL && sfmmup != NULL);
8543 	ASSERT(sptaddr == ISMID_STARTADDR);
8544 	/*
8545 	 * Check the alignment.
8546 	 */
8547 	if (!ISM_ALIGNED(ismshift, addr) || !ISM_ALIGNED(ismshift, sptaddr))
8548 		return (EINVAL);
8549 
8550 	/*
8551 	 * Check size alignment.
8552 	 */
8553 	if (!ISM_ALIGNED(ismshift, len))
8554 		return (EINVAL);
8555 
8556 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
8557 
8558 	/*
8559 	 * Allocate ism_ment for the ism_hat's mapping list, and an
8560 	 * ism map blk in case we need one.  We must do our
8561 	 * allocations before acquiring locks to prevent a deadlock
8562 	 * in the kmem allocator on the mapping list lock.
8563 	 */
8564 	new_iblk = kmem_cache_alloc(ism_blk_cache, KM_SLEEP);
8565 	ism_ment = kmem_cache_alloc(ism_ment_cache, KM_SLEEP);
8566 
8567 	/*
8568 	 * Serialize ISM mappings with the ISM busy flag, and also the
8569 	 * trap handlers.
8570 	 */
8571 	sfmmu_ismhat_enter(sfmmup, 0);
8572 
8573 	/*
8574 	 * Allocate an ism map blk if necessary.
8575 	 */
8576 	if (sfmmup->sfmmu_iblk == NULL) {
8577 		sfmmup->sfmmu_iblk = new_iblk;
8578 		bzero(new_iblk, sizeof (*new_iblk));
8579 		new_iblk->iblk_nextpa = (uint64_t)-1;
8580 		membar_stst();	/* make sure next ptr visible to all CPUs */
8581 		sfmmup->sfmmu_ismblkpa = va_to_pa((caddr_t)new_iblk);
8582 		reload_mmu = 1;
8583 		new_iblk = NULL;
8584 	}
8585 
8586 #ifdef DEBUG
8587 	/*
8588 	 * Make sure mapping does not already exist.
8589 	 */
8590 	ism_blkp = sfmmup->sfmmu_iblk;
8591 	while (ism_blkp != NULL) {
8592 		ism_map = ism_blkp->iblk_maps;
8593 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
8594 			if ((addr >= ism_start(ism_map[i]) &&
8595 			    addr < ism_end(ism_map[i])) ||
8596 			    eaddr > ism_start(ism_map[i]) &&
8597 			    eaddr <= ism_end(ism_map[i])) {
8598 				panic("sfmmu_share: Already mapped!");
8599 			}
8600 		}
8601 		ism_blkp = ism_blkp->iblk_next;
8602 	}
8603 #endif /* DEBUG */
8604 
8605 	ASSERT(ismszc >= TTE4M);
8606 	if (ismszc == TTE4M) {
8607 		ismhatflag = HAT_4M_FLAG;
8608 	} else if (ismszc == TTE32M) {
8609 		ismhatflag = HAT_32M_FLAG;
8610 	} else if (ismszc == TTE256M) {
8611 		ismhatflag = HAT_256M_FLAG;
8612 	}
8613 	/*
8614 	 * Add mapping to first available mapping slot.
8615 	 */
8616 	ism_blkp = sfmmup->sfmmu_iblk;
8617 	added = 0;
8618 	while (!added) {
8619 		ism_map = ism_blkp->iblk_maps;
8620 		for (i = 0; i < ISM_MAP_SLOTS; i++)  {
8621 			if (ism_map[i].imap_ismhat == NULL) {
8622 
8623 				ism_map[i].imap_ismhat = ism_hatid;
8624 				ism_map[i].imap_vb_shift = (uchar_t)ismshift;
8625 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8626 				ism_map[i].imap_hatflags = ismhatflag;
8627 				ism_map[i].imap_sz_mask = ismmask;
8628 				/*
8629 				 * imap_seg is checked in ISM_CHECK to see if
8630 				 * non-NULL, then other info assumed valid.
8631 				 */
8632 				membar_stst();
8633 				ism_map[i].imap_seg = (uintptr_t)addr | sh_size;
8634 				ism_map[i].imap_ment = ism_ment;
8635 
8636 				/*
8637 				 * Now add ourselves to the ism_hat's
8638 				 * mapping list.
8639 				 */
8640 				ism_ment->iment_hat = sfmmup;
8641 				ism_ment->iment_base_va = addr;
8642 				ism_hatid->sfmmu_ismhat = 1;
8643 				mutex_enter(&ism_mlist_lock);
8644 				iment_add(ism_ment, ism_hatid);
8645 				mutex_exit(&ism_mlist_lock);
8646 				added = 1;
8647 				break;
8648 			}
8649 		}
8650 		if (!added && ism_blkp->iblk_next == NULL) {
8651 			ism_blkp->iblk_next = new_iblk;
8652 			new_iblk = NULL;
8653 			bzero(ism_blkp->iblk_next,
8654 			    sizeof (*ism_blkp->iblk_next));
8655 			ism_blkp->iblk_next->iblk_nextpa = (uint64_t)-1;
8656 			membar_stst();
8657 			ism_blkp->iblk_nextpa =
8658 			    va_to_pa((caddr_t)ism_blkp->iblk_next);
8659 		}
8660 		ism_blkp = ism_blkp->iblk_next;
8661 	}
8662 
8663 	/*
8664 	 * After calling hat_join_region, sfmmup may join a new SCD or
8665 	 * move from the old scd to a new scd, in which case, we want to
8666 	 * shrink the sfmmup's private tsb size, i.e., pass shrink to
8667 	 * sfmmu_check_page_sizes at the end of this routine.
8668 	 */
8669 	old_scdp = sfmmup->sfmmu_scdp;
8670 
8671 	rcookie = hat_join_region(sfmmup, addr, len, (void *)ism_hatid, 0,
8672 	    PROT_ALL, ismszc, NULL, HAT_REGION_ISM);
8673 	if (rcookie != HAT_INVALID_REGION_COOKIE) {
8674 		ism_map[i].imap_rid = (uchar_t)((uint64_t)rcookie);
8675 	}
8676 	/*
8677 	 * Update our counters for this sfmmup's ism mappings.
8678 	 */
8679 	for (i = 0; i <= ismszc; i++) {
8680 		if (!(disable_ism_large_pages & (1 << i)))
8681 			(void) ism_tsb_entries(sfmmup, i);
8682 	}
8683 
8684 	/*
8685 	 * For ISM and DISM we do not support 512K pages, so we only only
8686 	 * search the 4M and 8K/64K hashes for 4 pagesize cpus, and search the
8687 	 * 256M or 32M, and 4M and 8K/64K hashes for 6 pagesize cpus.
8688 	 *
8689 	 * Need to set 32M/256M ISM flags to make sure
8690 	 * sfmmu_check_page_sizes() enables them on Panther.
8691 	 */
8692 	ASSERT((disable_ism_large_pages & (1 << TTE512K)) != 0);
8693 
8694 	switch (ismszc) {
8695 	case TTE256M:
8696 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_256M_ISM)) {
8697 			hatlockp = sfmmu_hat_enter(sfmmup);
8698 			SFMMU_FLAGS_SET(sfmmup, HAT_256M_ISM);
8699 			sfmmu_hat_exit(hatlockp);
8700 		}
8701 		break;
8702 	case TTE32M:
8703 		if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_32M_ISM)) {
8704 			hatlockp = sfmmu_hat_enter(sfmmup);
8705 			SFMMU_FLAGS_SET(sfmmup, HAT_32M_ISM);
8706 			sfmmu_hat_exit(hatlockp);
8707 		}
8708 		break;
8709 	default:
8710 		break;
8711 	}
8712 
8713 	/*
8714 	 * If we updated the ismblkpa for this HAT we must make
8715 	 * sure all CPUs running this process reload their tsbmiss area.
8716 	 * Otherwise they will fail to load the mappings in the tsbmiss
8717 	 * handler and will loop calling pagefault().
8718 	 */
8719 	if (reload_mmu) {
8720 		hatlockp = sfmmu_hat_enter(sfmmup);
8721 		sfmmu_sync_mmustate(sfmmup);
8722 		sfmmu_hat_exit(hatlockp);
8723 	}
8724 
8725 	sfmmu_ismhat_exit(sfmmup, 0);
8726 
8727 	/*
8728 	 * Free up ismblk if we didn't use it.
8729 	 */
8730 	if (new_iblk != NULL)
8731 		kmem_cache_free(ism_blk_cache, new_iblk);
8732 
8733 	/*
8734 	 * Check TSB and TLB page sizes.
8735 	 */
8736 	if (sfmmup->sfmmu_scdp != NULL && old_scdp != sfmmup->sfmmu_scdp) {
8737 		sfmmu_check_page_sizes(sfmmup, 0);
8738 	} else {
8739 		sfmmu_check_page_sizes(sfmmup, 1);
8740 	}
8741 	return (0);
8742 }
8743 
8744 /*
8745  * hat_unshare removes exactly one ism_map from
8746  * this process's as.  It expects multiple calls
8747  * to hat_unshare for multiple shm segments.
8748  */
8749 void
8750 hat_unshare(struct hat *sfmmup, caddr_t addr, size_t len, uint_t ismszc)
8751 {
8752 	ism_map_t 	*ism_map;
8753 	ism_ment_t	*free_ment = NULL;
8754 	ism_blk_t	*ism_blkp;
8755 	struct hat	*ism_hatid;
8756 	int 		found, i;
8757 	hatlock_t	*hatlockp;
8758 	struct tsb_info	*tsbinfo;
8759 	uint_t		ismshift = page_get_shift(ismszc);
8760 	size_t		sh_size = ISM_SHIFT(ismshift, len);
8761 	uchar_t		ism_rid;
8762 	sf_scd_t	*old_scdp;
8763 
8764 	ASSERT(ISM_ALIGNED(ismshift, addr));
8765 	ASSERT(ISM_ALIGNED(ismshift, len));
8766 	ASSERT(sfmmup != NULL);
8767 	ASSERT(sfmmup != ksfmmup);
8768 
8769 	if (sfmmup->sfmmu_xhat_provider) {
8770 		XHAT_UNSHARE(sfmmup, addr, len);
8771 		return;
8772 	} else {
8773 		/*
8774 		 * This must be a CPU HAT. If the address space has
8775 		 * XHATs attached, inform all XHATs that ISM segment
8776 		 * is going away
8777 		 */
8778 		ASSERT(sfmmup->sfmmu_as != NULL);
8779 		if (sfmmup->sfmmu_as->a_xhat != NULL)
8780 			xhat_unshare_all(sfmmup->sfmmu_as, addr, len);
8781 	}
8782 
8783 	/*
8784 	 * Make sure that during the entire time ISM mappings are removed,
8785 	 * the trap handlers serialize behind us, and that no one else
8786 	 * can be mucking with ISM mappings.  This also lets us get away
8787 	 * with not doing expensive cross calls to flush the TLB -- we
8788 	 * just discard the context, flush the entire TSB, and call it
8789 	 * a day.
8790 	 */
8791 	sfmmu_ismhat_enter(sfmmup, 0);
8792 
8793 	/*
8794 	 * Remove the mapping.
8795 	 *
8796 	 * We can't have any holes in the ism map.
8797 	 * The tsb miss code while searching the ism map will
8798 	 * stop on an empty map slot.  So we must move
8799 	 * everyone past the hole up 1 if any.
8800 	 *
8801 	 * Also empty ism map blks are not freed until the
8802 	 * process exits. This is to prevent a MT race condition
8803 	 * between sfmmu_unshare() and sfmmu_tsbmiss_exception().
8804 	 */
8805 	found = 0;
8806 	ism_blkp = sfmmup->sfmmu_iblk;
8807 	while (!found && ism_blkp != NULL) {
8808 		ism_map = ism_blkp->iblk_maps;
8809 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
8810 			if (addr == ism_start(ism_map[i]) &&
8811 			    sh_size == (size_t)(ism_size(ism_map[i]))) {
8812 				found = 1;
8813 				break;
8814 			}
8815 		}
8816 		if (!found)
8817 			ism_blkp = ism_blkp->iblk_next;
8818 	}
8819 
8820 	if (found) {
8821 		ism_hatid = ism_map[i].imap_ismhat;
8822 		ism_rid = ism_map[i].imap_rid;
8823 		ASSERT(ism_hatid != NULL);
8824 		ASSERT(ism_hatid->sfmmu_ismhat == 1);
8825 
8826 		/*
8827 		 * After hat_leave_region, the sfmmup may leave SCD,
8828 		 * in which case, we want to grow the private tsb size when
8829 		 * calling sfmmu_check_page_sizes at the end of the routine.
8830 		 */
8831 		old_scdp = sfmmup->sfmmu_scdp;
8832 		/*
8833 		 * Then remove ourselves from the region.
8834 		 */
8835 		if (ism_rid != SFMMU_INVALID_ISMRID) {
8836 			hat_leave_region(sfmmup, (void *)((uint64_t)ism_rid),
8837 			    HAT_REGION_ISM);
8838 		}
8839 
8840 		/*
8841 		 * And now guarantee that any other cpu
8842 		 * that tries to process an ISM miss
8843 		 * will go to tl=0.
8844 		 */
8845 		hatlockp = sfmmu_hat_enter(sfmmup);
8846 		sfmmu_invalidate_ctx(sfmmup);
8847 		sfmmu_hat_exit(hatlockp);
8848 
8849 		/*
8850 		 * Remove ourselves from the ism mapping list.
8851 		 */
8852 		mutex_enter(&ism_mlist_lock);
8853 		iment_sub(ism_map[i].imap_ment, ism_hatid);
8854 		mutex_exit(&ism_mlist_lock);
8855 		free_ment = ism_map[i].imap_ment;
8856 
8857 		/*
8858 		 * We delete the ism map by copying
8859 		 * the next map over the current one.
8860 		 * We will take the next one in the maps
8861 		 * array or from the next ism_blk.
8862 		 */
8863 		while (ism_blkp != NULL) {
8864 			ism_map = ism_blkp->iblk_maps;
8865 			while (i < (ISM_MAP_SLOTS - 1)) {
8866 				ism_map[i] = ism_map[i + 1];
8867 				i++;
8868 			}
8869 			/* i == (ISM_MAP_SLOTS - 1) */
8870 			ism_blkp = ism_blkp->iblk_next;
8871 			if (ism_blkp != NULL) {
8872 				ism_map[i] = ism_blkp->iblk_maps[0];
8873 				i = 0;
8874 			} else {
8875 				ism_map[i].imap_seg = 0;
8876 				ism_map[i].imap_vb_shift = 0;
8877 				ism_map[i].imap_rid = SFMMU_INVALID_ISMRID;
8878 				ism_map[i].imap_hatflags = 0;
8879 				ism_map[i].imap_sz_mask = 0;
8880 				ism_map[i].imap_ismhat = NULL;
8881 				ism_map[i].imap_ment = NULL;
8882 			}
8883 		}
8884 
8885 		/*
8886 		 * Now flush entire TSB for the process, since
8887 		 * demapping page by page can be too expensive.
8888 		 * We don't have to flush the TLB here anymore
8889 		 * since we switch to a new TLB ctx instead.
8890 		 * Also, there is no need to flush if the process
8891 		 * is exiting since the TSB will be freed later.
8892 		 */
8893 		if (!sfmmup->sfmmu_free) {
8894 			hatlockp = sfmmu_hat_enter(sfmmup);
8895 			for (tsbinfo = sfmmup->sfmmu_tsb; tsbinfo != NULL;
8896 			    tsbinfo = tsbinfo->tsb_next) {
8897 				if (tsbinfo->tsb_flags & TSB_SWAPPED)
8898 					continue;
8899 				if (tsbinfo->tsb_flags & TSB_RELOC_FLAG) {
8900 					tsbinfo->tsb_flags |=
8901 					    TSB_FLUSH_NEEDED;
8902 					continue;
8903 				}
8904 
8905 				sfmmu_inv_tsb(tsbinfo->tsb_va,
8906 				    TSB_BYTES(tsbinfo->tsb_szc));
8907 			}
8908 			sfmmu_hat_exit(hatlockp);
8909 		}
8910 	}
8911 
8912 	/*
8913 	 * Update our counters for this sfmmup's ism mappings.
8914 	 */
8915 	for (i = 0; i <= ismszc; i++) {
8916 		if (!(disable_ism_large_pages & (1 << i)))
8917 			(void) ism_tsb_entries(sfmmup, i);
8918 	}
8919 
8920 	sfmmu_ismhat_exit(sfmmup, 0);
8921 
8922 	/*
8923 	 * We must do our freeing here after dropping locks
8924 	 * to prevent a deadlock in the kmem allocator on the
8925 	 * mapping list lock.
8926 	 */
8927 	if (free_ment != NULL)
8928 		kmem_cache_free(ism_ment_cache, free_ment);
8929 
8930 	/*
8931 	 * Check TSB and TLB page sizes if the process isn't exiting.
8932 	 */
8933 	if (!sfmmup->sfmmu_free) {
8934 		if (found && old_scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
8935 			sfmmu_check_page_sizes(sfmmup, 1);
8936 		} else {
8937 			sfmmu_check_page_sizes(sfmmup, 0);
8938 		}
8939 	}
8940 }
8941 
8942 /* ARGSUSED */
8943 static int
8944 sfmmu_idcache_constructor(void *buf, void *cdrarg, int kmflags)
8945 {
8946 	/* void *buf is sfmmu_t pointer */
8947 	bzero(buf, sizeof (sfmmu_t));
8948 
8949 	return (0);
8950 }
8951 
8952 /* ARGSUSED */
8953 static void
8954 sfmmu_idcache_destructor(void *buf, void *cdrarg)
8955 {
8956 	/* void *buf is sfmmu_t pointer */
8957 }
8958 
8959 /*
8960  * setup kmem hmeblks by bzeroing all members and initializing the nextpa
8961  * field to be the pa of this hmeblk
8962  */
8963 /* ARGSUSED */
8964 static int
8965 sfmmu_hblkcache_constructor(void *buf, void *cdrarg, int kmflags)
8966 {
8967 	struct hme_blk *hmeblkp;
8968 
8969 	bzero(buf, (size_t)cdrarg);
8970 	hmeblkp = (struct hme_blk *)buf;
8971 	hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
8972 
8973 #ifdef	HBLK_TRACE
8974 	mutex_init(&hmeblkp->hblk_audit_lock, NULL, MUTEX_DEFAULT, NULL);
8975 #endif	/* HBLK_TRACE */
8976 
8977 	return (0);
8978 }
8979 
8980 /* ARGSUSED */
8981 static void
8982 sfmmu_hblkcache_destructor(void *buf, void *cdrarg)
8983 {
8984 
8985 #ifdef	HBLK_TRACE
8986 
8987 	struct hme_blk *hmeblkp;
8988 
8989 	hmeblkp = (struct hme_blk *)buf;
8990 	mutex_destroy(&hmeblkp->hblk_audit_lock);
8991 
8992 #endif	/* HBLK_TRACE */
8993 }
8994 
8995 #define	SFMMU_CACHE_RECLAIM_SCAN_RATIO 8
8996 static int sfmmu_cache_reclaim_scan_ratio = SFMMU_CACHE_RECLAIM_SCAN_RATIO;
8997 /*
8998  * The kmem allocator will callback into our reclaim routine when the system
8999  * is running low in memory.  We traverse the hash and free up all unused but
9000  * still cached hme_blks.  We also traverse the free list and free them up
9001  * as well.
9002  */
9003 /*ARGSUSED*/
9004 static void
9005 sfmmu_hblkcache_reclaim(void *cdrarg)
9006 {
9007 	int i;
9008 	uint64_t hblkpa, prevpa, nx_pa;
9009 	struct hmehash_bucket *hmebp;
9010 	struct hme_blk *hmeblkp, *nx_hblk, *pr_hblk = NULL;
9011 	static struct hmehash_bucket *uhmehash_reclaim_hand;
9012 	static struct hmehash_bucket *khmehash_reclaim_hand;
9013 	struct hme_blk *list = NULL;
9014 
9015 	hmebp = uhmehash_reclaim_hand;
9016 	if (hmebp == NULL || hmebp > &uhme_hash[UHMEHASH_SZ])
9017 		uhmehash_reclaim_hand = hmebp = uhme_hash;
9018 	uhmehash_reclaim_hand += UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
9019 
9020 	for (i = UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
9021 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
9022 			hmeblkp = hmebp->hmeblkp;
9023 			hblkpa = hmebp->hmeh_nextpa;
9024 			prevpa = 0;
9025 			pr_hblk = NULL;
9026 			while (hmeblkp) {
9027 				nx_hblk = hmeblkp->hblk_next;
9028 				nx_pa = hmeblkp->hblk_nextpa;
9029 				if (!hmeblkp->hblk_vcnt &&
9030 				    !hmeblkp->hblk_hmecnt) {
9031 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9032 					    prevpa, pr_hblk);
9033 					sfmmu_hblk_free(hmebp, hmeblkp,
9034 					    hblkpa, &list);
9035 				} else {
9036 					pr_hblk = hmeblkp;
9037 					prevpa = hblkpa;
9038 				}
9039 				hmeblkp = nx_hblk;
9040 				hblkpa = nx_pa;
9041 			}
9042 			SFMMU_HASH_UNLOCK(hmebp);
9043 		}
9044 		if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
9045 			hmebp = uhme_hash;
9046 	}
9047 
9048 	hmebp = khmehash_reclaim_hand;
9049 	if (hmebp == NULL || hmebp > &khme_hash[KHMEHASH_SZ])
9050 		khmehash_reclaim_hand = hmebp = khme_hash;
9051 	khmehash_reclaim_hand += KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
9052 
9053 	for (i = KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
9054 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
9055 			hmeblkp = hmebp->hmeblkp;
9056 			hblkpa = hmebp->hmeh_nextpa;
9057 			prevpa = 0;
9058 			pr_hblk = NULL;
9059 			while (hmeblkp) {
9060 				nx_hblk = hmeblkp->hblk_next;
9061 				nx_pa = hmeblkp->hblk_nextpa;
9062 				if (!hmeblkp->hblk_vcnt &&
9063 				    !hmeblkp->hblk_hmecnt) {
9064 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
9065 					    prevpa, pr_hblk);
9066 					sfmmu_hblk_free(hmebp, hmeblkp,
9067 					    hblkpa, &list);
9068 				} else {
9069 					pr_hblk = hmeblkp;
9070 					prevpa = hblkpa;
9071 				}
9072 				hmeblkp = nx_hblk;
9073 				hblkpa = nx_pa;
9074 			}
9075 			SFMMU_HASH_UNLOCK(hmebp);
9076 		}
9077 		if (hmebp++ == &khme_hash[KHMEHASH_SZ])
9078 			hmebp = khme_hash;
9079 	}
9080 	sfmmu_hblks_list_purge(&list);
9081 }
9082 
9083 /*
9084  * sfmmu_get_ppvcolor should become a vm_machdep or hatop interface.
9085  * same goes for sfmmu_get_addrvcolor().
9086  *
9087  * This function will return the virtual color for the specified page. The
9088  * virtual color corresponds to this page current mapping or its last mapping.
9089  * It is used by memory allocators to choose addresses with the correct
9090  * alignment so vac consistency is automatically maintained.  If the page
9091  * has no color it returns -1.
9092  */
9093 /*ARGSUSED*/
9094 int
9095 sfmmu_get_ppvcolor(struct page *pp)
9096 {
9097 #ifdef VAC
9098 	int color;
9099 
9100 	if (!(cache & CACHE_VAC) || PP_NEWPAGE(pp)) {
9101 		return (-1);
9102 	}
9103 	color = PP_GET_VCOLOR(pp);
9104 	ASSERT(color < mmu_btop(shm_alignment));
9105 	return (color);
9106 #else
9107 	return (-1);
9108 #endif	/* VAC */
9109 }
9110 
9111 /*
9112  * This function will return the desired alignment for vac consistency
9113  * (vac color) given a virtual address.  If no vac is present it returns -1.
9114  */
9115 /*ARGSUSED*/
9116 int
9117 sfmmu_get_addrvcolor(caddr_t vaddr)
9118 {
9119 #ifdef VAC
9120 	if (cache & CACHE_VAC) {
9121 		return (addr_to_vcolor(vaddr));
9122 	} else {
9123 		return (-1);
9124 	}
9125 #else
9126 	return (-1);
9127 #endif	/* VAC */
9128 }
9129 
9130 #ifdef VAC
9131 /*
9132  * Check for conflicts.
9133  * A conflict exists if the new and existent mappings do not match in
9134  * their "shm_alignment fields. If conflicts exist, the existant mappings
9135  * are flushed unless one of them is locked. If one of them is locked, then
9136  * the mappings are flushed and converted to non-cacheable mappings.
9137  */
9138 static void
9139 sfmmu_vac_conflict(struct hat *hat, caddr_t addr, page_t *pp)
9140 {
9141 	struct hat *tmphat;
9142 	struct sf_hment *sfhmep, *tmphme = NULL;
9143 	struct hme_blk *hmeblkp;
9144 	int vcolor;
9145 	tte_t tte;
9146 
9147 	ASSERT(sfmmu_mlist_held(pp));
9148 	ASSERT(!PP_ISNC(pp));		/* page better be cacheable */
9149 
9150 	vcolor = addr_to_vcolor(addr);
9151 	if (PP_NEWPAGE(pp)) {
9152 		PP_SET_VCOLOR(pp, vcolor);
9153 		return;
9154 	}
9155 
9156 	if (PP_GET_VCOLOR(pp) == vcolor) {
9157 		return;
9158 	}
9159 
9160 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
9161 		/*
9162 		 * Previous user of page had a different color
9163 		 * but since there are no current users
9164 		 * we just flush the cache and change the color.
9165 		 */
9166 		SFMMU_STAT(sf_pgcolor_conflict);
9167 		sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9168 		PP_SET_VCOLOR(pp, vcolor);
9169 		return;
9170 	}
9171 
9172 	/*
9173 	 * If we get here we have a vac conflict with a current
9174 	 * mapping.  VAC conflict policy is as follows.
9175 	 * - The default is to unload the other mappings unless:
9176 	 * - If we have a large mapping we uncache the page.
9177 	 * We need to uncache the rest of the large page too.
9178 	 * - If any of the mappings are locked we uncache the page.
9179 	 * - If the requested mapping is inconsistent
9180 	 * with another mapping and that mapping
9181 	 * is in the same address space we have to
9182 	 * make it non-cached.  The default thing
9183 	 * to do is unload the inconsistent mapping
9184 	 * but if they are in the same address space
9185 	 * we run the risk of unmapping the pc or the
9186 	 * stack which we will use as we return to the user,
9187 	 * in which case we can then fault on the thing
9188 	 * we just unloaded and get into an infinite loop.
9189 	 */
9190 	if (PP_ISMAPPED_LARGE(pp)) {
9191 		int sz;
9192 
9193 		/*
9194 		 * Existing mapping is for big pages. We don't unload
9195 		 * existing big mappings to satisfy new mappings.
9196 		 * Always convert all mappings to TNC.
9197 		 */
9198 		sz = fnd_mapping_sz(pp);
9199 		pp = PP_GROUPLEADER(pp, sz);
9200 		SFMMU_STAT_ADD(sf_uncache_conflict, TTEPAGES(sz));
9201 		sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH,
9202 		    TTEPAGES(sz));
9203 
9204 		return;
9205 	}
9206 
9207 	/*
9208 	 * check if any mapping is in same as or if it is locked
9209 	 * since in that case we need to uncache.
9210 	 */
9211 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9212 		tmphme = sfhmep->hme_next;
9213 		if (IS_PAHME(sfhmep))
9214 			continue;
9215 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9216 		if (hmeblkp->hblk_xhat_bit)
9217 			continue;
9218 		tmphat = hblktosfmmu(hmeblkp);
9219 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
9220 		ASSERT(TTE_IS_VALID(&tte));
9221 		if (hmeblkp->hblk_shared || tmphat == hat ||
9222 		    hmeblkp->hblk_lckcnt) {
9223 			/*
9224 			 * We have an uncache conflict
9225 			 */
9226 			SFMMU_STAT(sf_uncache_conflict);
9227 			sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH, 1);
9228 			return;
9229 		}
9230 	}
9231 
9232 	/*
9233 	 * We have an unload conflict
9234 	 * We have already checked for LARGE mappings, therefore
9235 	 * the remaining mapping(s) must be TTE8K.
9236 	 */
9237 	SFMMU_STAT(sf_unload_conflict);
9238 
9239 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
9240 		tmphme = sfhmep->hme_next;
9241 		if (IS_PAHME(sfhmep))
9242 			continue;
9243 		hmeblkp = sfmmu_hmetohblk(sfhmep);
9244 		if (hmeblkp->hblk_xhat_bit)
9245 			continue;
9246 		ASSERT(!hmeblkp->hblk_shared);
9247 		(void) sfmmu_pageunload(pp, sfhmep, TTE8K);
9248 	}
9249 
9250 	if (PP_ISMAPPED_KPM(pp))
9251 		sfmmu_kpm_vac_unload(pp, addr);
9252 
9253 	/*
9254 	 * Unloads only do TLB flushes so we need to flush the
9255 	 * cache here.
9256 	 */
9257 	sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
9258 	PP_SET_VCOLOR(pp, vcolor);
9259 }
9260 
9261 /*
9262  * Whenever a mapping is unloaded and the page is in TNC state,
9263  * we see if the page can be made cacheable again. 'pp' is
9264  * the page that we just unloaded a mapping from, the size
9265  * of mapping that was unloaded is 'ottesz'.
9266  * Remark:
9267  * The recache policy for mpss pages can leave a performance problem
9268  * under the following circumstances:
9269  * . A large page in uncached mode has just been unmapped.
9270  * . All constituent pages are TNC due to a conflicting small mapping.
9271  * . There are many other, non conflicting, small mappings around for
9272  *   a lot of the constituent pages.
9273  * . We're called w/ the "old" groupleader page and the old ottesz,
9274  *   but this is irrelevant, since we're no more "PP_ISMAPPED_LARGE", so
9275  *   we end up w/ TTE8K or npages == 1.
9276  * . We call tst_tnc w/ the old groupleader only, and if there is no
9277  *   conflict, we re-cache only this page.
9278  * . All other small mappings are not checked and will be left in TNC mode.
9279  * The problem is not very serious because:
9280  * . mpss is actually only defined for heap and stack, so the probability
9281  *   is not very high that a large page mapping exists in parallel to a small
9282  *   one (this is possible, but seems to be bad programming style in the
9283  *   appl).
9284  * . The problem gets a little bit more serious, when those TNC pages
9285  *   have to be mapped into kernel space, e.g. for networking.
9286  * . When VAC alias conflicts occur in applications, this is regarded
9287  *   as an application bug. So if kstat's show them, the appl should
9288  *   be changed anyway.
9289  */
9290 void
9291 conv_tnc(page_t *pp, int ottesz)
9292 {
9293 	int cursz, dosz;
9294 	pgcnt_t curnpgs, dopgs;
9295 	pgcnt_t pg64k;
9296 	page_t *pp2;
9297 
9298 	/*
9299 	 * Determine how big a range we check for TNC and find
9300 	 * leader page. cursz is the size of the biggest
9301 	 * mapping that still exist on 'pp'.
9302 	 */
9303 	if (PP_ISMAPPED_LARGE(pp)) {
9304 		cursz = fnd_mapping_sz(pp);
9305 	} else {
9306 		cursz = TTE8K;
9307 	}
9308 
9309 	if (ottesz >= cursz) {
9310 		dosz = ottesz;
9311 		pp2 = pp;
9312 	} else {
9313 		dosz = cursz;
9314 		pp2 = PP_GROUPLEADER(pp, dosz);
9315 	}
9316 
9317 	pg64k = TTEPAGES(TTE64K);
9318 	dopgs = TTEPAGES(dosz);
9319 
9320 	ASSERT(dopgs == 1 || ((dopgs & (pg64k - 1)) == 0));
9321 
9322 	while (dopgs != 0) {
9323 		curnpgs = TTEPAGES(cursz);
9324 		if (tst_tnc(pp2, curnpgs)) {
9325 			SFMMU_STAT_ADD(sf_recache, curnpgs);
9326 			sfmmu_page_cache_array(pp2, HAT_CACHE, CACHE_NO_FLUSH,
9327 			    curnpgs);
9328 		}
9329 
9330 		ASSERT(dopgs >= curnpgs);
9331 		dopgs -= curnpgs;
9332 
9333 		if (dopgs == 0) {
9334 			break;
9335 		}
9336 
9337 		pp2 = PP_PAGENEXT_N(pp2, curnpgs);
9338 		if (((dopgs & (pg64k - 1)) == 0) && PP_ISMAPPED_LARGE(pp2)) {
9339 			cursz = fnd_mapping_sz(pp2);
9340 		} else {
9341 			cursz = TTE8K;
9342 		}
9343 	}
9344 }
9345 
9346 /*
9347  * Returns 1 if page(s) can be converted from TNC to cacheable setting,
9348  * returns 0 otherwise. Note that oaddr argument is valid for only
9349  * 8k pages.
9350  */
9351 int
9352 tst_tnc(page_t *pp, pgcnt_t npages)
9353 {
9354 	struct	sf_hment *sfhme;
9355 	struct	hme_blk *hmeblkp;
9356 	tte_t	tte;
9357 	caddr_t	vaddr;
9358 	int	clr_valid = 0;
9359 	int 	color, color1, bcolor;
9360 	int	i, ncolors;
9361 
9362 	ASSERT(pp != NULL);
9363 	ASSERT(!(cache & CACHE_WRITEBACK));
9364 
9365 	if (npages > 1) {
9366 		ncolors = CACHE_NUM_COLOR;
9367 	}
9368 
9369 	for (i = 0; i < npages; i++) {
9370 		ASSERT(sfmmu_mlist_held(pp));
9371 		ASSERT(PP_ISTNC(pp));
9372 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
9373 
9374 		if (PP_ISPNC(pp)) {
9375 			return (0);
9376 		}
9377 
9378 		clr_valid = 0;
9379 		if (PP_ISMAPPED_KPM(pp)) {
9380 			caddr_t kpmvaddr;
9381 
9382 			ASSERT(kpm_enable);
9383 			kpmvaddr = hat_kpm_page2va(pp, 1);
9384 			ASSERT(!(npages > 1 && IS_KPM_ALIAS_RANGE(kpmvaddr)));
9385 			color1 = addr_to_vcolor(kpmvaddr);
9386 			clr_valid = 1;
9387 		}
9388 
9389 		for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9390 			if (IS_PAHME(sfhme))
9391 				continue;
9392 			hmeblkp = sfmmu_hmetohblk(sfhme);
9393 			if (hmeblkp->hblk_xhat_bit)
9394 				continue;
9395 
9396 			sfmmu_copytte(&sfhme->hme_tte, &tte);
9397 			ASSERT(TTE_IS_VALID(&tte));
9398 
9399 			vaddr = tte_to_vaddr(hmeblkp, tte);
9400 			color = addr_to_vcolor(vaddr);
9401 
9402 			if (npages > 1) {
9403 				/*
9404 				 * If there is a big mapping, make sure
9405 				 * 8K mapping is consistent with the big
9406 				 * mapping.
9407 				 */
9408 				bcolor = i % ncolors;
9409 				if (color != bcolor) {
9410 					return (0);
9411 				}
9412 			}
9413 			if (!clr_valid) {
9414 				clr_valid = 1;
9415 				color1 = color;
9416 			}
9417 
9418 			if (color1 != color) {
9419 				return (0);
9420 			}
9421 		}
9422 
9423 		pp = PP_PAGENEXT(pp);
9424 	}
9425 
9426 	return (1);
9427 }
9428 
9429 void
9430 sfmmu_page_cache_array(page_t *pp, int flags, int cache_flush_flag,
9431 	pgcnt_t npages)
9432 {
9433 	kmutex_t *pmtx;
9434 	int i, ncolors, bcolor;
9435 	kpm_hlk_t *kpmp;
9436 	cpuset_t cpuset;
9437 
9438 	ASSERT(pp != NULL);
9439 	ASSERT(!(cache & CACHE_WRITEBACK));
9440 
9441 	kpmp = sfmmu_kpm_kpmp_enter(pp, npages);
9442 	pmtx = sfmmu_page_enter(pp);
9443 
9444 	/*
9445 	 * Fast path caching single unmapped page
9446 	 */
9447 	if (npages == 1 && !PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp) &&
9448 	    flags == HAT_CACHE) {
9449 		PP_CLRTNC(pp);
9450 		PP_CLRPNC(pp);
9451 		sfmmu_page_exit(pmtx);
9452 		sfmmu_kpm_kpmp_exit(kpmp);
9453 		return;
9454 	}
9455 
9456 	/*
9457 	 * We need to capture all cpus in order to change cacheability
9458 	 * because we can't allow one cpu to access the same physical
9459 	 * page using a cacheable and a non-cachebale mapping at the same
9460 	 * time. Since we may end up walking the ism mapping list
9461 	 * have to grab it's lock now since we can't after all the
9462 	 * cpus have been captured.
9463 	 */
9464 	sfmmu_hat_lock_all();
9465 	mutex_enter(&ism_mlist_lock);
9466 	kpreempt_disable();
9467 	cpuset = cpu_ready_set;
9468 	xc_attention(cpuset);
9469 
9470 	if (npages > 1) {
9471 		/*
9472 		 * Make sure all colors are flushed since the
9473 		 * sfmmu_page_cache() only flushes one color-
9474 		 * it does not know big pages.
9475 		 */
9476 		ncolors = CACHE_NUM_COLOR;
9477 		if (flags & HAT_TMPNC) {
9478 			for (i = 0; i < ncolors; i++) {
9479 				sfmmu_cache_flushcolor(i, pp->p_pagenum);
9480 			}
9481 			cache_flush_flag = CACHE_NO_FLUSH;
9482 		}
9483 	}
9484 
9485 	for (i = 0; i < npages; i++) {
9486 
9487 		ASSERT(sfmmu_mlist_held(pp));
9488 
9489 		if (!(flags == HAT_TMPNC && PP_ISTNC(pp))) {
9490 
9491 			if (npages > 1) {
9492 				bcolor = i % ncolors;
9493 			} else {
9494 				bcolor = NO_VCOLOR;
9495 			}
9496 
9497 			sfmmu_page_cache(pp, flags, cache_flush_flag,
9498 			    bcolor);
9499 		}
9500 
9501 		pp = PP_PAGENEXT(pp);
9502 	}
9503 
9504 	xt_sync(cpuset);
9505 	xc_dismissed(cpuset);
9506 	mutex_exit(&ism_mlist_lock);
9507 	sfmmu_hat_unlock_all();
9508 	sfmmu_page_exit(pmtx);
9509 	sfmmu_kpm_kpmp_exit(kpmp);
9510 	kpreempt_enable();
9511 }
9512 
9513 /*
9514  * This function changes the virtual cacheability of all mappings to a
9515  * particular page.  When changing from uncache to cacheable the mappings will
9516  * only be changed if all of them have the same virtual color.
9517  * We need to flush the cache in all cpus.  It is possible that
9518  * a process referenced a page as cacheable but has sinced exited
9519  * and cleared the mapping list.  We still to flush it but have no
9520  * state so all cpus is the only alternative.
9521  */
9522 static void
9523 sfmmu_page_cache(page_t *pp, int flags, int cache_flush_flag, int bcolor)
9524 {
9525 	struct	sf_hment *sfhme;
9526 	struct	hme_blk *hmeblkp;
9527 	sfmmu_t *sfmmup;
9528 	tte_t	tte, ttemod;
9529 	caddr_t	vaddr;
9530 	int	ret, color;
9531 	pfn_t	pfn;
9532 
9533 	color = bcolor;
9534 	pfn = pp->p_pagenum;
9535 
9536 	for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
9537 
9538 		if (IS_PAHME(sfhme))
9539 			continue;
9540 		hmeblkp = sfmmu_hmetohblk(sfhme);
9541 
9542 		if (hmeblkp->hblk_xhat_bit)
9543 			continue;
9544 
9545 		sfmmu_copytte(&sfhme->hme_tte, &tte);
9546 		ASSERT(TTE_IS_VALID(&tte));
9547 		vaddr = tte_to_vaddr(hmeblkp, tte);
9548 		color = addr_to_vcolor(vaddr);
9549 
9550 #ifdef DEBUG
9551 		if ((flags & HAT_CACHE) && bcolor != NO_VCOLOR) {
9552 			ASSERT(color == bcolor);
9553 		}
9554 #endif
9555 
9556 		ASSERT(flags != HAT_TMPNC || color == PP_GET_VCOLOR(pp));
9557 
9558 		ttemod = tte;
9559 		if (flags & (HAT_UNCACHE | HAT_TMPNC)) {
9560 			TTE_CLR_VCACHEABLE(&ttemod);
9561 		} else {	/* flags & HAT_CACHE */
9562 			TTE_SET_VCACHEABLE(&ttemod);
9563 		}
9564 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
9565 		if (ret < 0) {
9566 			/*
9567 			 * Since all cpus are captured modifytte should not
9568 			 * fail.
9569 			 */
9570 			panic("sfmmu_page_cache: write to tte failed");
9571 		}
9572 
9573 		sfmmup = hblktosfmmu(hmeblkp);
9574 		if (cache_flush_flag == CACHE_FLUSH) {
9575 			/*
9576 			 * Flush TSBs, TLBs and caches
9577 			 */
9578 			if (hmeblkp->hblk_shared) {
9579 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9580 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9581 				sf_region_t *rgnp;
9582 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9583 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9584 				ASSERT(srdp != NULL);
9585 				rgnp = srdp->srd_hmergnp[rid];
9586 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9587 				    srdp, rgnp, rid);
9588 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9589 				    hmeblkp, 0);
9590 				sfmmu_cache_flush(pfn, addr_to_vcolor(vaddr));
9591 			} else if (sfmmup->sfmmu_ismhat) {
9592 				if (flags & HAT_CACHE) {
9593 					SFMMU_STAT(sf_ism_recache);
9594 				} else {
9595 					SFMMU_STAT(sf_ism_uncache);
9596 				}
9597 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9598 				    pfn, CACHE_FLUSH);
9599 			} else {
9600 				sfmmu_tlbcache_demap(vaddr, sfmmup, hmeblkp,
9601 				    pfn, 0, FLUSH_ALL_CPUS, CACHE_FLUSH, 1);
9602 			}
9603 
9604 			/*
9605 			 * all cache entries belonging to this pfn are
9606 			 * now flushed.
9607 			 */
9608 			cache_flush_flag = CACHE_NO_FLUSH;
9609 		} else {
9610 			/*
9611 			 * Flush only TSBs and TLBs.
9612 			 */
9613 			if (hmeblkp->hblk_shared) {
9614 				sf_srd_t *srdp = (sf_srd_t *)sfmmup;
9615 				uint_t rid = hmeblkp->hblk_tag.htag_rid;
9616 				sf_region_t *rgnp;
9617 				ASSERT(SFMMU_IS_SHMERID_VALID(rid));
9618 				ASSERT(rid < SFMMU_MAX_HME_REGIONS);
9619 				ASSERT(srdp != NULL);
9620 				rgnp = srdp->srd_hmergnp[rid];
9621 				SFMMU_VALIDATE_SHAREDHBLK(hmeblkp,
9622 				    srdp, rgnp, rid);
9623 				(void) sfmmu_rgntlb_demap(vaddr, rgnp,
9624 				    hmeblkp, 0);
9625 			} else if (sfmmup->sfmmu_ismhat) {
9626 				if (flags & HAT_CACHE) {
9627 					SFMMU_STAT(sf_ism_recache);
9628 				} else {
9629 					SFMMU_STAT(sf_ism_uncache);
9630 				}
9631 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
9632 				    pfn, CACHE_NO_FLUSH);
9633 			} else {
9634 				sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 1);
9635 			}
9636 		}
9637 	}
9638 
9639 	if (PP_ISMAPPED_KPM(pp))
9640 		sfmmu_kpm_page_cache(pp, flags, cache_flush_flag);
9641 
9642 	switch (flags) {
9643 
9644 		default:
9645 			panic("sfmmu_pagecache: unknown flags");
9646 			break;
9647 
9648 		case HAT_CACHE:
9649 			PP_CLRTNC(pp);
9650 			PP_CLRPNC(pp);
9651 			PP_SET_VCOLOR(pp, color);
9652 			break;
9653 
9654 		case HAT_TMPNC:
9655 			PP_SETTNC(pp);
9656 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9657 			break;
9658 
9659 		case HAT_UNCACHE:
9660 			PP_SETPNC(pp);
9661 			PP_CLRTNC(pp);
9662 			PP_SET_VCOLOR(pp, NO_VCOLOR);
9663 			break;
9664 	}
9665 }
9666 #endif	/* VAC */
9667 
9668 
9669 /*
9670  * Wrapper routine used to return a context.
9671  *
9672  * It's the responsibility of the caller to guarantee that the
9673  * process serializes on calls here by taking the HAT lock for
9674  * the hat.
9675  *
9676  */
9677 static void
9678 sfmmu_get_ctx(sfmmu_t *sfmmup)
9679 {
9680 	mmu_ctx_t *mmu_ctxp;
9681 	uint_t pstate_save;
9682 	int ret;
9683 
9684 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9685 	ASSERT(sfmmup != ksfmmup);
9686 
9687 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID)) {
9688 		sfmmu_setup_tsbinfo(sfmmup);
9689 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ALLCTX_INVALID);
9690 	}
9691 
9692 	kpreempt_disable();
9693 
9694 	mmu_ctxp = CPU_MMU_CTXP(CPU);
9695 	ASSERT(mmu_ctxp);
9696 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
9697 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
9698 
9699 	/*
9700 	 * Do a wrap-around if cnum reaches the max # cnum supported by a MMU.
9701 	 */
9702 	if (mmu_ctxp->mmu_cnum == mmu_ctxp->mmu_nctxs)
9703 		sfmmu_ctx_wrap_around(mmu_ctxp);
9704 
9705 	/*
9706 	 * Let the MMU set up the page sizes to use for
9707 	 * this context in the TLB. Don't program 2nd dtlb for ism hat.
9708 	 */
9709 	if ((&mmu_set_ctx_page_sizes) && (sfmmup->sfmmu_ismhat == 0)) {
9710 		mmu_set_ctx_page_sizes(sfmmup);
9711 	}
9712 
9713 	/*
9714 	 * sfmmu_alloc_ctx and sfmmu_load_mmustate will be performed with
9715 	 * interrupts disabled to prevent race condition with wrap-around
9716 	 * ctx invalidatation. In sun4v, ctx invalidation also involves
9717 	 * a HV call to set the number of TSBs to 0. If interrupts are not
9718 	 * disabled until after sfmmu_load_mmustate is complete TSBs may
9719 	 * become assigned to INVALID_CONTEXT. This is not allowed.
9720 	 */
9721 	pstate_save = sfmmu_disable_intrs();
9722 
9723 	if (sfmmu_alloc_ctx(sfmmup, 1, CPU, SFMMU_PRIVATE) &&
9724 	    sfmmup->sfmmu_scdp != NULL) {
9725 		sf_scd_t *scdp = sfmmup->sfmmu_scdp;
9726 		sfmmu_t *scsfmmup = scdp->scd_sfmmup;
9727 		ret = sfmmu_alloc_ctx(scsfmmup, 1, CPU, SFMMU_SHARED);
9728 		/* debug purpose only */
9729 		ASSERT(!ret || scsfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
9730 		    != INVALID_CONTEXT);
9731 	}
9732 	sfmmu_load_mmustate(sfmmup);
9733 
9734 	sfmmu_enable_intrs(pstate_save);
9735 
9736 	kpreempt_enable();
9737 }
9738 
9739 /*
9740  * When all cnums are used up in a MMU, cnum will wrap around to the
9741  * next generation and start from 2.
9742  */
9743 static void
9744 sfmmu_ctx_wrap_around(mmu_ctx_t *mmu_ctxp)
9745 {
9746 
9747 	/* caller must have disabled the preemption */
9748 	ASSERT(curthread->t_preempt >= 1);
9749 	ASSERT(mmu_ctxp != NULL);
9750 
9751 	/* acquire Per-MMU (PM) spin lock */
9752 	mutex_enter(&mmu_ctxp->mmu_lock);
9753 
9754 	/* re-check to see if wrap-around is needed */
9755 	if (mmu_ctxp->mmu_cnum < mmu_ctxp->mmu_nctxs)
9756 		goto done;
9757 
9758 	SFMMU_MMU_STAT(mmu_wrap_around);
9759 
9760 	/* update gnum */
9761 	ASSERT(mmu_ctxp->mmu_gnum != 0);
9762 	mmu_ctxp->mmu_gnum++;
9763 	if (mmu_ctxp->mmu_gnum == 0 ||
9764 	    mmu_ctxp->mmu_gnum > MAX_SFMMU_GNUM_VAL) {
9765 		cmn_err(CE_PANIC, "mmu_gnum of mmu_ctx 0x%p is out of bound.",
9766 		    (void *)mmu_ctxp);
9767 	}
9768 
9769 	if (mmu_ctxp->mmu_ncpus > 1) {
9770 		cpuset_t cpuset;
9771 
9772 		membar_enter(); /* make sure updated gnum visible */
9773 
9774 		SFMMU_XCALL_STATS(NULL);
9775 
9776 		/* xcall to others on the same MMU to invalidate ctx */
9777 		cpuset = mmu_ctxp->mmu_cpuset;
9778 		ASSERT(CPU_IN_SET(cpuset, CPU->cpu_id));
9779 		CPUSET_DEL(cpuset, CPU->cpu_id);
9780 		CPUSET_AND(cpuset, cpu_ready_set);
9781 
9782 		/*
9783 		 * Pass in INVALID_CONTEXT as the first parameter to
9784 		 * sfmmu_raise_tsb_exception, which invalidates the context
9785 		 * of any process running on the CPUs in the MMU.
9786 		 */
9787 		xt_some(cpuset, sfmmu_raise_tsb_exception,
9788 		    INVALID_CONTEXT, INVALID_CONTEXT);
9789 		xt_sync(cpuset);
9790 
9791 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
9792 	}
9793 
9794 	if (sfmmu_getctx_sec() != INVALID_CONTEXT) {
9795 		sfmmu_setctx_sec(INVALID_CONTEXT);
9796 		sfmmu_clear_utsbinfo();
9797 	}
9798 
9799 	/*
9800 	 * No xcall is needed here. For sun4u systems all CPUs in context
9801 	 * domain share a single physical MMU therefore it's enough to flush
9802 	 * TLB on local CPU. On sun4v systems we use 1 global context
9803 	 * domain and flush all remote TLBs in sfmmu_raise_tsb_exception
9804 	 * handler. Note that vtag_flushall_uctxs() is called
9805 	 * for Ultra II machine, where the equivalent flushall functionality
9806 	 * is implemented in SW, and only user ctx TLB entries are flushed.
9807 	 */
9808 	if (&vtag_flushall_uctxs != NULL) {
9809 		vtag_flushall_uctxs();
9810 	} else {
9811 		vtag_flushall();
9812 	}
9813 
9814 	/* reset mmu cnum, skips cnum 0 and 1 */
9815 	mmu_ctxp->mmu_cnum = NUM_LOCKED_CTXS;
9816 
9817 done:
9818 	mutex_exit(&mmu_ctxp->mmu_lock);
9819 }
9820 
9821 
9822 /*
9823  * For multi-threaded process, set the process context to INVALID_CONTEXT
9824  * so that it faults and reloads the MMU state from TL=0. For single-threaded
9825  * process, we can just load the MMU state directly without having to
9826  * set context invalid. Caller must hold the hat lock since we don't
9827  * acquire it here.
9828  */
9829 static void
9830 sfmmu_sync_mmustate(sfmmu_t *sfmmup)
9831 {
9832 	uint_t cnum;
9833 	uint_t pstate_save;
9834 
9835 	ASSERT(sfmmup != ksfmmup);
9836 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9837 
9838 	kpreempt_disable();
9839 
9840 	/*
9841 	 * We check whether the pass'ed-in sfmmup is the same as the
9842 	 * current running proc. This is to makes sure the current proc
9843 	 * stays single-threaded if it already is.
9844 	 */
9845 	if ((sfmmup == curthread->t_procp->p_as->a_hat) &&
9846 	    (curthread->t_procp->p_lwpcnt == 1)) {
9847 		/* single-thread */
9848 		cnum = sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum;
9849 		if (cnum != INVALID_CONTEXT) {
9850 			uint_t curcnum;
9851 			/*
9852 			 * Disable interrupts to prevent race condition
9853 			 * with sfmmu_ctx_wrap_around ctx invalidation.
9854 			 * In sun4v, ctx invalidation involves setting
9855 			 * TSB to NULL, hence, interrupts should be disabled
9856 			 * untill after sfmmu_load_mmustate is completed.
9857 			 */
9858 			pstate_save = sfmmu_disable_intrs();
9859 			curcnum = sfmmu_getctx_sec();
9860 			if (curcnum == cnum)
9861 				sfmmu_load_mmustate(sfmmup);
9862 			sfmmu_enable_intrs(pstate_save);
9863 			ASSERT(curcnum == cnum || curcnum == INVALID_CONTEXT);
9864 		}
9865 	} else {
9866 		/*
9867 		 * multi-thread
9868 		 * or when sfmmup is not the same as the curproc.
9869 		 */
9870 		sfmmu_invalidate_ctx(sfmmup);
9871 	}
9872 
9873 	kpreempt_enable();
9874 }
9875 
9876 
9877 /*
9878  * Replace the specified TSB with a new TSB.  This function gets called when
9879  * we grow, shrink or swapin a TSB.  When swapping in a TSB (TSB_SWAPIN), the
9880  * TSB_FORCEALLOC flag may be used to force allocation of a minimum-sized TSB
9881  * (8K).
9882  *
9883  * Caller must hold the HAT lock, but should assume any tsb_info
9884  * pointers it has are no longer valid after calling this function.
9885  *
9886  * Return values:
9887  *	TSB_ALLOCFAIL	Failed to allocate a TSB, due to memory constraints
9888  *	TSB_LOSTRACE	HAT is busy, i.e. another thread is already doing
9889  *			something to this tsbinfo/TSB
9890  *	TSB_SUCCESS	Operation succeeded
9891  */
9892 static tsb_replace_rc_t
9893 sfmmu_replace_tsb(sfmmu_t *sfmmup, struct tsb_info *old_tsbinfo, uint_t szc,
9894     hatlock_t *hatlockp, uint_t flags)
9895 {
9896 	struct tsb_info *new_tsbinfo = NULL;
9897 	struct tsb_info *curtsb, *prevtsb;
9898 	uint_t tte_sz_mask;
9899 	int i;
9900 
9901 	ASSERT(sfmmup != ksfmmup);
9902 	ASSERT(sfmmup->sfmmu_ismhat == 0);
9903 	ASSERT(sfmmu_hat_lock_held(sfmmup));
9904 	ASSERT(szc <= tsb_max_growsize);
9905 
9906 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_BUSY))
9907 		return (TSB_LOSTRACE);
9908 
9909 	/*
9910 	 * Find the tsb_info ahead of this one in the list, and
9911 	 * also make sure that the tsb_info passed in really
9912 	 * exists!
9913 	 */
9914 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
9915 	    curtsb != old_tsbinfo && curtsb != NULL;
9916 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
9917 		;
9918 	ASSERT(curtsb != NULL);
9919 
9920 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
9921 		/*
9922 		 * The process is swapped out, so just set the new size
9923 		 * code.  When it swaps back in, we'll allocate a new one
9924 		 * of the new chosen size.
9925 		 */
9926 		curtsb->tsb_szc = szc;
9927 		return (TSB_SUCCESS);
9928 	}
9929 	SFMMU_FLAGS_SET(sfmmup, HAT_BUSY);
9930 
9931 	tte_sz_mask = old_tsbinfo->tsb_ttesz_mask;
9932 
9933 	/*
9934 	 * All initialization is done inside of sfmmu_tsbinfo_alloc().
9935 	 * If we fail to allocate a TSB, exit.
9936 	 *
9937 	 * If tsb grows with new tsb size > 4M and old tsb size < 4M,
9938 	 * then try 4M slab after the initial alloc fails.
9939 	 *
9940 	 * If tsb swapin with tsb size > 4M, then try 4M after the
9941 	 * initial alloc fails.
9942 	 */
9943 	sfmmu_hat_exit(hatlockp);
9944 	if (sfmmu_tsbinfo_alloc(&new_tsbinfo, szc,
9945 	    tte_sz_mask, flags, sfmmup) &&
9946 	    (!(flags & (TSB_GROW | TSB_SWAPIN)) || (szc <= TSB_4M_SZCODE) ||
9947 	    (!(flags & TSB_SWAPIN) &&
9948 	    (old_tsbinfo->tsb_szc >= TSB_4M_SZCODE)) ||
9949 	    sfmmu_tsbinfo_alloc(&new_tsbinfo, TSB_4M_SZCODE,
9950 	    tte_sz_mask, flags, sfmmup))) {
9951 		(void) sfmmu_hat_enter(sfmmup);
9952 		if (!(flags & TSB_SWAPIN))
9953 			SFMMU_STAT(sf_tsb_resize_failures);
9954 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
9955 		return (TSB_ALLOCFAIL);
9956 	}
9957 	(void) sfmmu_hat_enter(sfmmup);
9958 
9959 	/*
9960 	 * Re-check to make sure somebody else didn't muck with us while we
9961 	 * didn't hold the HAT lock.  If the process swapped out, fine, just
9962 	 * exit; this can happen if we try to shrink the TSB from the context
9963 	 * of another process (such as on an ISM unmap), though it is rare.
9964 	 */
9965 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
9966 		SFMMU_STAT(sf_tsb_resize_failures);
9967 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
9968 		sfmmu_hat_exit(hatlockp);
9969 		sfmmu_tsbinfo_free(new_tsbinfo);
9970 		(void) sfmmu_hat_enter(sfmmup);
9971 		return (TSB_LOSTRACE);
9972 	}
9973 
9974 #ifdef	DEBUG
9975 	/* Reverify that the tsb_info still exists.. for debugging only */
9976 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
9977 	    curtsb != old_tsbinfo && curtsb != NULL;
9978 	    prevtsb = curtsb, curtsb = curtsb->tsb_next)
9979 		;
9980 	ASSERT(curtsb != NULL);
9981 #endif	/* DEBUG */
9982 
9983 	/*
9984 	 * Quiesce any CPUs running this process on their next TLB miss
9985 	 * so they atomically see the new tsb_info.  We temporarily set the
9986 	 * context to invalid context so new threads that come on processor
9987 	 * after we do the xcall to cpusran will also serialize behind the
9988 	 * HAT lock on TLB miss and will see the new TSB.  Since this short
9989 	 * race with a new thread coming on processor is relatively rare,
9990 	 * this synchronization mechanism should be cheaper than always
9991 	 * pausing all CPUs for the duration of the setup, which is what
9992 	 * the old implementation did.  This is particuarly true if we are
9993 	 * copying a huge chunk of memory around during that window.
9994 	 *
9995 	 * The memory barriers are to make sure things stay consistent
9996 	 * with resume() since it does not hold the HAT lock while
9997 	 * walking the list of tsb_info structures.
9998 	 */
9999 	if ((flags & TSB_SWAPIN) != TSB_SWAPIN) {
10000 		/* The TSB is either growing or shrinking. */
10001 		sfmmu_invalidate_ctx(sfmmup);
10002 	} else {
10003 		/*
10004 		 * It is illegal to swap in TSBs from a process other
10005 		 * than a process being swapped in.  This in turn
10006 		 * implies we do not have a valid MMU context here
10007 		 * since a process needs one to resolve translation
10008 		 * misses.
10009 		 */
10010 		ASSERT(curthread->t_procp->p_as->a_hat == sfmmup);
10011 	}
10012 
10013 #ifdef DEBUG
10014 	ASSERT(max_mmu_ctxdoms > 0);
10015 
10016 	/*
10017 	 * Process should have INVALID_CONTEXT on all MMUs
10018 	 */
10019 	for (i = 0; i < max_mmu_ctxdoms; i++) {
10020 
10021 		ASSERT(sfmmup->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
10022 	}
10023 #endif
10024 
10025 	new_tsbinfo->tsb_next = old_tsbinfo->tsb_next;
10026 	membar_stst();	/* strict ordering required */
10027 	if (prevtsb)
10028 		prevtsb->tsb_next = new_tsbinfo;
10029 	else
10030 		sfmmup->sfmmu_tsb = new_tsbinfo;
10031 	membar_enter();	/* make sure new TSB globally visible */
10032 
10033 	/*
10034 	 * We need to migrate TSB entries from the old TSB to the new TSB
10035 	 * if tsb_remap_ttes is set and the TSB is growing.
10036 	 */
10037 	if (tsb_remap_ttes && ((flags & TSB_GROW) == TSB_GROW))
10038 		sfmmu_copy_tsb(old_tsbinfo, new_tsbinfo);
10039 
10040 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
10041 
10042 	/*
10043 	 * Drop the HAT lock to free our old tsb_info.
10044 	 */
10045 	sfmmu_hat_exit(hatlockp);
10046 
10047 	if ((flags & TSB_GROW) == TSB_GROW) {
10048 		SFMMU_STAT(sf_tsb_grow);
10049 	} else if ((flags & TSB_SHRINK) == TSB_SHRINK) {
10050 		SFMMU_STAT(sf_tsb_shrink);
10051 	}
10052 
10053 	sfmmu_tsbinfo_free(old_tsbinfo);
10054 
10055 	(void) sfmmu_hat_enter(sfmmup);
10056 	return (TSB_SUCCESS);
10057 }
10058 
10059 /*
10060  * This function will re-program hat pgsz array, and invalidate the
10061  * process' context, forcing the process to switch to another
10062  * context on the next TLB miss, and therefore start using the
10063  * TLB that is reprogrammed for the new page sizes.
10064  */
10065 void
10066 sfmmu_reprog_pgsz_arr(sfmmu_t *sfmmup, uint8_t *tmp_pgsz)
10067 {
10068 	int i;
10069 	hatlock_t *hatlockp = NULL;
10070 
10071 	hatlockp = sfmmu_hat_enter(sfmmup);
10072 	/* USIII+-IV+ optimization, requires hat lock */
10073 	if (tmp_pgsz) {
10074 		for (i = 0; i < mmu_page_sizes; i++)
10075 			sfmmup->sfmmu_pgsz[i] = tmp_pgsz[i];
10076 	}
10077 	SFMMU_STAT(sf_tlb_reprog_pgsz);
10078 
10079 	sfmmu_invalidate_ctx(sfmmup);
10080 
10081 	sfmmu_hat_exit(hatlockp);
10082 }
10083 
10084 /*
10085  * The scd_rttecnt field in the SCD must be updated to take account of the
10086  * regions which it contains.
10087  */
10088 static void
10089 sfmmu_set_scd_rttecnt(sf_srd_t *srdp, sf_scd_t *scdp)
10090 {
10091 	uint_t rid;
10092 	uint_t i, j;
10093 	ulong_t w;
10094 	sf_region_t *rgnp;
10095 
10096 	ASSERT(srdp != NULL);
10097 
10098 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
10099 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
10100 			continue;
10101 		}
10102 
10103 		j = 0;
10104 		while (w) {
10105 			if (!(w & 0x1)) {
10106 				j++;
10107 				w >>= 1;
10108 				continue;
10109 			}
10110 			rid = (i << BT_ULSHIFT) | j;
10111 			j++;
10112 			w >>= 1;
10113 
10114 			ASSERT(SFMMU_IS_SHMERID_VALID(rid));
10115 			ASSERT(rid < SFMMU_MAX_HME_REGIONS);
10116 			rgnp = srdp->srd_hmergnp[rid];
10117 			ASSERT(rgnp->rgn_refcnt > 0);
10118 			ASSERT(rgnp->rgn_id == rid);
10119 
10120 			scdp->scd_rttecnt[rgnp->rgn_pgszc] +=
10121 			    rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
10122 
10123 			/*
10124 			 * Maintain the tsb0 inflation cnt for the regions
10125 			 * in the SCD.
10126 			 */
10127 			if (rgnp->rgn_pgszc >= TTE4M) {
10128 				scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt +=
10129 				    rgnp->rgn_size >>
10130 				    (TTE_PAGE_SHIFT(TTE8K) + 2);
10131 			}
10132 		}
10133 	}
10134 }
10135 
10136 /*
10137  * This function assumes that there are either four or six supported page
10138  * sizes and at most two programmable TLBs, so we need to decide which
10139  * page sizes are most important and then tell the MMU layer so it
10140  * can adjust the TLB page sizes accordingly (if supported).
10141  *
10142  * If these assumptions change, this function will need to be
10143  * updated to support whatever the new limits are.
10144  *
10145  * The growing flag is nonzero if we are growing the address space,
10146  * and zero if it is shrinking.  This allows us to decide whether
10147  * to grow or shrink our TSB, depending upon available memory
10148  * conditions.
10149  */
10150 static void
10151 sfmmu_check_page_sizes(sfmmu_t *sfmmup, int growing)
10152 {
10153 	uint64_t ttecnt[MMU_PAGE_SIZES];
10154 	uint64_t tte8k_cnt, tte4m_cnt;
10155 	uint8_t i;
10156 	int sectsb_thresh;
10157 
10158 	/*
10159 	 * Kernel threads, processes with small address spaces not using
10160 	 * large pages, and dummy ISM HATs need not apply.
10161 	 */
10162 	if (sfmmup == ksfmmup || sfmmup->sfmmu_ismhat != NULL)
10163 		return;
10164 
10165 	if (!SFMMU_LGPGS_INUSE(sfmmup) &&
10166 	    sfmmup->sfmmu_ttecnt[TTE8K] <= tsb_rss_factor)
10167 		return;
10168 
10169 	for (i = 0; i < mmu_page_sizes; i++) {
10170 		ttecnt[i] = sfmmup->sfmmu_ttecnt[i] +
10171 		    sfmmup->sfmmu_ismttecnt[i];
10172 	}
10173 
10174 	/* Check pagesizes in use, and possibly reprogram DTLB. */
10175 	if (&mmu_check_page_sizes)
10176 		mmu_check_page_sizes(sfmmup, ttecnt);
10177 
10178 	/*
10179 	 * Calculate the number of 8k ttes to represent the span of these
10180 	 * pages.
10181 	 */
10182 	tte8k_cnt = ttecnt[TTE8K] +
10183 	    (ttecnt[TTE64K] << (MMU_PAGESHIFT64K - MMU_PAGESHIFT)) +
10184 	    (ttecnt[TTE512K] << (MMU_PAGESHIFT512K - MMU_PAGESHIFT));
10185 	if (mmu_page_sizes == max_mmu_page_sizes) {
10186 		tte4m_cnt = ttecnt[TTE4M] +
10187 		    (ttecnt[TTE32M] << (MMU_PAGESHIFT32M - MMU_PAGESHIFT4M)) +
10188 		    (ttecnt[TTE256M] << (MMU_PAGESHIFT256M - MMU_PAGESHIFT4M));
10189 	} else {
10190 		tte4m_cnt = ttecnt[TTE4M];
10191 	}
10192 
10193 	/*
10194 	 * Inflate tte8k_cnt to allow for region large page allocation failure.
10195 	 */
10196 	tte8k_cnt += sfmmup->sfmmu_tsb0_4minflcnt;
10197 
10198 	/*
10199 	 * Inflate TSB sizes by a factor of 2 if this process
10200 	 * uses 4M text pages to minimize extra conflict misses
10201 	 * in the first TSB since without counting text pages
10202 	 * 8K TSB may become too small.
10203 	 *
10204 	 * Also double the size of the second TSB to minimize
10205 	 * extra conflict misses due to competition between 4M text pages
10206 	 * and data pages.
10207 	 *
10208 	 * We need to adjust the second TSB allocation threshold by the
10209 	 * inflation factor, since there is no point in creating a second
10210 	 * TSB when we know all the mappings can fit in the I/D TLBs.
10211 	 */
10212 	sectsb_thresh = tsb_sectsb_threshold;
10213 	if (sfmmup->sfmmu_flags & HAT_4MTEXT_FLAG) {
10214 		tte8k_cnt <<= 1;
10215 		tte4m_cnt <<= 1;
10216 		sectsb_thresh <<= 1;
10217 	}
10218 
10219 	/*
10220 	 * Check to see if our TSB is the right size; we may need to
10221 	 * grow or shrink it.  If the process is small, our work is
10222 	 * finished at this point.
10223 	 */
10224 	if (tte8k_cnt <= tsb_rss_factor && tte4m_cnt <= sectsb_thresh) {
10225 		return;
10226 	}
10227 	sfmmu_size_tsb(sfmmup, growing, tte8k_cnt, tte4m_cnt, sectsb_thresh);
10228 }
10229 
10230 static void
10231 sfmmu_size_tsb(sfmmu_t *sfmmup, int growing, uint64_t tte8k_cnt,
10232 	uint64_t tte4m_cnt, int sectsb_thresh)
10233 {
10234 	int tsb_bits;
10235 	uint_t tsb_szc;
10236 	struct tsb_info *tsbinfop;
10237 	hatlock_t *hatlockp = NULL;
10238 
10239 	hatlockp = sfmmu_hat_enter(sfmmup);
10240 	ASSERT(hatlockp != NULL);
10241 	tsbinfop = sfmmup->sfmmu_tsb;
10242 	ASSERT(tsbinfop != NULL);
10243 
10244 	/*
10245 	 * If we're growing, select the size based on RSS.  If we're
10246 	 * shrinking, leave some room so we don't have to turn around and
10247 	 * grow again immediately.
10248 	 */
10249 	if (growing)
10250 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
10251 	else
10252 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt << 1);
10253 
10254 	if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10255 	    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10256 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10257 		    hatlockp, TSB_SHRINK);
10258 	} else if (growing && tsb_szc > tsbinfop->tsb_szc && TSB_OK_GROW()) {
10259 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
10260 		    hatlockp, TSB_GROW);
10261 	}
10262 	tsbinfop = sfmmup->sfmmu_tsb;
10263 
10264 	/*
10265 	 * With the TLB and first TSB out of the way, we need to see if
10266 	 * we need a second TSB for 4M pages.  If we managed to reprogram
10267 	 * the TLB page sizes above, the process will start using this new
10268 	 * TSB right away; otherwise, it will start using it on the next
10269 	 * context switch.  Either way, it's no big deal so there's no
10270 	 * synchronization with the trap handlers here unless we grow the
10271 	 * TSB (in which case it's required to prevent using the old one
10272 	 * after it's freed). Note: second tsb is required for 32M/256M
10273 	 * page sizes.
10274 	 */
10275 	if (tte4m_cnt > sectsb_thresh) {
10276 		/*
10277 		 * If we're growing, select the size based on RSS.  If we're
10278 		 * shrinking, leave some room so we don't have to turn
10279 		 * around and grow again immediately.
10280 		 */
10281 		if (growing)
10282 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
10283 		else
10284 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt << 1);
10285 		if (tsbinfop->tsb_next == NULL) {
10286 			struct tsb_info *newtsb;
10287 			int allocflags = SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)?
10288 			    0 : TSB_ALLOC;
10289 
10290 			sfmmu_hat_exit(hatlockp);
10291 
10292 			/*
10293 			 * Try to allocate a TSB for 4[32|256]M pages.  If we
10294 			 * can't get the size we want, retry w/a minimum sized
10295 			 * TSB.  If that still didn't work, give up; we can
10296 			 * still run without one.
10297 			 */
10298 			tsb_bits = (mmu_page_sizes == max_mmu_page_sizes)?
10299 			    TSB4M|TSB32M|TSB256M:TSB4M;
10300 			if ((sfmmu_tsbinfo_alloc(&newtsb, tsb_szc, tsb_bits,
10301 			    allocflags, sfmmup)) &&
10302 			    (tsb_szc <= TSB_4M_SZCODE ||
10303 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
10304 			    tsb_bits, allocflags, sfmmup)) &&
10305 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_MIN_SZCODE,
10306 			    tsb_bits, allocflags, sfmmup)) {
10307 				return;
10308 			}
10309 
10310 			hatlockp = sfmmu_hat_enter(sfmmup);
10311 
10312 			sfmmu_invalidate_ctx(sfmmup);
10313 
10314 			if (sfmmup->sfmmu_tsb->tsb_next == NULL) {
10315 				sfmmup->sfmmu_tsb->tsb_next = newtsb;
10316 				SFMMU_STAT(sf_tsb_sectsb_create);
10317 				sfmmu_hat_exit(hatlockp);
10318 				return;
10319 			} else {
10320 				/*
10321 				 * It's annoying, but possible for us
10322 				 * to get here.. we dropped the HAT lock
10323 				 * because of locking order in the kmem
10324 				 * allocator, and while we were off getting
10325 				 * our memory, some other thread decided to
10326 				 * do us a favor and won the race to get a
10327 				 * second TSB for this process.  Sigh.
10328 				 */
10329 				sfmmu_hat_exit(hatlockp);
10330 				sfmmu_tsbinfo_free(newtsb);
10331 				return;
10332 			}
10333 		}
10334 
10335 		/*
10336 		 * We have a second TSB, see if it's big enough.
10337 		 */
10338 		tsbinfop = tsbinfop->tsb_next;
10339 
10340 		/*
10341 		 * Check to see if our second TSB is the right size;
10342 		 * we may need to grow or shrink it.
10343 		 * To prevent thrashing (e.g. growing the TSB on a
10344 		 * subsequent map operation), only try to shrink if
10345 		 * the TSB reach exceeds twice the virtual address
10346 		 * space size.
10347 		 */
10348 		if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
10349 		    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
10350 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10351 			    tsb_szc, hatlockp, TSB_SHRINK);
10352 		} else if (growing && tsb_szc > tsbinfop->tsb_szc &&
10353 		    TSB_OK_GROW()) {
10354 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
10355 			    tsb_szc, hatlockp, TSB_GROW);
10356 		}
10357 	}
10358 
10359 	sfmmu_hat_exit(hatlockp);
10360 }
10361 
10362 /*
10363  * Free up a sfmmu
10364  * Since the sfmmu is currently embedded in the hat struct we simply zero
10365  * out our fields and free up the ism map blk list if any.
10366  */
10367 static void
10368 sfmmu_free_sfmmu(sfmmu_t *sfmmup)
10369 {
10370 	ism_blk_t	*blkp, *nx_blkp;
10371 #ifdef	DEBUG
10372 	ism_map_t	*map;
10373 	int 		i;
10374 #endif
10375 
10376 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
10377 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
10378 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
10379 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
10380 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
10381 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
10382 	ASSERT(SF_RGNMAP_ISNULL(sfmmup));
10383 
10384 	sfmmup->sfmmu_free = 0;
10385 	sfmmup->sfmmu_ismhat = 0;
10386 
10387 	blkp = sfmmup->sfmmu_iblk;
10388 	sfmmup->sfmmu_iblk = NULL;
10389 
10390 	while (blkp) {
10391 #ifdef	DEBUG
10392 		map = blkp->iblk_maps;
10393 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
10394 			ASSERT(map[i].imap_seg == 0);
10395 			ASSERT(map[i].imap_ismhat == NULL);
10396 			ASSERT(map[i].imap_ment == NULL);
10397 		}
10398 #endif
10399 		nx_blkp = blkp->iblk_next;
10400 		blkp->iblk_next = NULL;
10401 		blkp->iblk_nextpa = (uint64_t)-1;
10402 		kmem_cache_free(ism_blk_cache, blkp);
10403 		blkp = nx_blkp;
10404 	}
10405 }
10406 
10407 /*
10408  * Locking primitves accessed by HATLOCK macros
10409  */
10410 
10411 #define	SFMMU_SPL_MTX	(0x0)
10412 #define	SFMMU_ML_MTX	(0x1)
10413 
10414 #define	SFMMU_MLSPL_MTX(type, pg)	(((type) == SFMMU_SPL_MTX) ? \
10415 					    SPL_HASH(pg) : MLIST_HASH(pg))
10416 
10417 kmutex_t *
10418 sfmmu_page_enter(struct page *pp)
10419 {
10420 	return (sfmmu_mlspl_enter(pp, SFMMU_SPL_MTX));
10421 }
10422 
10423 void
10424 sfmmu_page_exit(kmutex_t *spl)
10425 {
10426 	mutex_exit(spl);
10427 }
10428 
10429 int
10430 sfmmu_page_spl_held(struct page *pp)
10431 {
10432 	return (sfmmu_mlspl_held(pp, SFMMU_SPL_MTX));
10433 }
10434 
10435 kmutex_t *
10436 sfmmu_mlist_enter(struct page *pp)
10437 {
10438 	return (sfmmu_mlspl_enter(pp, SFMMU_ML_MTX));
10439 }
10440 
10441 void
10442 sfmmu_mlist_exit(kmutex_t *mml)
10443 {
10444 	mutex_exit(mml);
10445 }
10446 
10447 int
10448 sfmmu_mlist_held(struct page *pp)
10449 {
10450 
10451 	return (sfmmu_mlspl_held(pp, SFMMU_ML_MTX));
10452 }
10453 
10454 /*
10455  * Common code for sfmmu_mlist_enter() and sfmmu_page_enter().  For
10456  * sfmmu_mlist_enter() case mml_table lock array is used and for
10457  * sfmmu_page_enter() sfmmu_page_lock lock array is used.
10458  *
10459  * The lock is taken on a root page so that it protects an operation on all
10460  * constituent pages of a large page pp belongs to.
10461  *
10462  * The routine takes a lock from the appropriate array. The lock is determined
10463  * by hashing the root page. After taking the lock this routine checks if the
10464  * root page has the same size code that was used to determine the root (i.e
10465  * that root hasn't changed).  If root page has the expected p_szc field we
10466  * have the right lock and it's returned to the caller. If root's p_szc
10467  * decreased we release the lock and retry from the beginning.  This case can
10468  * happen due to hat_page_demote() decreasing p_szc between our load of p_szc
10469  * value and taking the lock. The number of retries due to p_szc decrease is
10470  * limited by the maximum p_szc value. If p_szc is 0 we return the lock
10471  * determined by hashing pp itself.
10472  *
10473  * If our caller doesn't hold a SE_SHARED or SE_EXCL lock on pp it's also
10474  * possible that p_szc can increase. To increase p_szc a thread has to lock
10475  * all constituent pages EXCL and do hat_pageunload() on all of them. All the
10476  * callers that don't hold a page locked recheck if hmeblk through which pp
10477  * was found still maps this pp.  If it doesn't map it anymore returned lock
10478  * is immediately dropped. Therefore if sfmmu_mlspl_enter() hits the case of
10479  * p_szc increase after taking the lock it returns this lock without further
10480  * retries because in this case the caller doesn't care about which lock was
10481  * taken. The caller will drop it right away.
10482  *
10483  * After the routine returns it's guaranteed that hat_page_demote() can't
10484  * change p_szc field of any of constituent pages of a large page pp belongs
10485  * to as long as pp was either locked at least SHARED prior to this call or
10486  * the caller finds that hment that pointed to this pp still references this
10487  * pp (this also assumes that the caller holds hme hash bucket lock so that
10488  * the same pp can't be remapped into the same hmeblk after it was unmapped by
10489  * hat_pageunload()).
10490  */
10491 static kmutex_t *
10492 sfmmu_mlspl_enter(struct page *pp, int type)
10493 {
10494 	kmutex_t	*mtx;
10495 	uint_t		prev_rszc = UINT_MAX;
10496 	page_t		*rootpp;
10497 	uint_t		szc;
10498 	uint_t		rszc;
10499 	uint_t		pszc = pp->p_szc;
10500 
10501 	ASSERT(pp != NULL);
10502 
10503 again:
10504 	if (pszc == 0) {
10505 		mtx = SFMMU_MLSPL_MTX(type, pp);
10506 		mutex_enter(mtx);
10507 		return (mtx);
10508 	}
10509 
10510 	/* The lock lives in the root page */
10511 	rootpp = PP_GROUPLEADER(pp, pszc);
10512 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10513 	mutex_enter(mtx);
10514 
10515 	/*
10516 	 * Return mml in the following 3 cases:
10517 	 *
10518 	 * 1) If pp itself is root since if its p_szc decreased before we took
10519 	 * the lock pp is still the root of smaller szc page. And if its p_szc
10520 	 * increased it doesn't matter what lock we return (see comment in
10521 	 * front of this routine).
10522 	 *
10523 	 * 2) If pp's not root but rootpp is the root of a rootpp->p_szc size
10524 	 * large page we have the right lock since any previous potential
10525 	 * hat_page_demote() is done demoting from greater than current root's
10526 	 * p_szc because hat_page_demote() changes root's p_szc last. No
10527 	 * further hat_page_demote() can start or be in progress since it
10528 	 * would need the same lock we currently hold.
10529 	 *
10530 	 * 3) If rootpp's p_szc increased since previous iteration it doesn't
10531 	 * matter what lock we return (see comment in front of this routine).
10532 	 */
10533 	if (pp == rootpp || (rszc = rootpp->p_szc) == pszc ||
10534 	    rszc >= prev_rszc) {
10535 		return (mtx);
10536 	}
10537 
10538 	/*
10539 	 * hat_page_demote() could have decreased root's p_szc.
10540 	 * In this case pp's p_szc must also be smaller than pszc.
10541 	 * Retry.
10542 	 */
10543 	if (rszc < pszc) {
10544 		szc = pp->p_szc;
10545 		if (szc < pszc) {
10546 			mutex_exit(mtx);
10547 			pszc = szc;
10548 			goto again;
10549 		}
10550 		/*
10551 		 * pp's p_szc increased after it was decreased.
10552 		 * page cannot be mapped. Return current lock. The caller
10553 		 * will drop it right away.
10554 		 */
10555 		return (mtx);
10556 	}
10557 
10558 	/*
10559 	 * root's p_szc is greater than pp's p_szc.
10560 	 * hat_page_demote() is not done with all pages
10561 	 * yet. Wait for it to complete.
10562 	 */
10563 	mutex_exit(mtx);
10564 	rootpp = PP_GROUPLEADER(rootpp, rszc);
10565 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
10566 	mutex_enter(mtx);
10567 	mutex_exit(mtx);
10568 	prev_rszc = rszc;
10569 	goto again;
10570 }
10571 
10572 static int
10573 sfmmu_mlspl_held(struct page *pp, int type)
10574 {
10575 	kmutex_t	*mtx;
10576 
10577 	ASSERT(pp != NULL);
10578 	/* The lock lives in the root page */
10579 	pp = PP_PAGEROOT(pp);
10580 	ASSERT(pp != NULL);
10581 
10582 	mtx = SFMMU_MLSPL_MTX(type, pp);
10583 	return (MUTEX_HELD(mtx));
10584 }
10585 
10586 static uint_t
10587 sfmmu_get_free_hblk(struct hme_blk **hmeblkpp, uint_t critical)
10588 {
10589 	struct  hme_blk *hblkp;
10590 
10591 	if (freehblkp != NULL) {
10592 		mutex_enter(&freehblkp_lock);
10593 		if (freehblkp != NULL) {
10594 			/*
10595 			 * If the current thread is owning hblk_reserve OR
10596 			 * critical request from sfmmu_hblk_steal()
10597 			 * let it succeed even if freehblkcnt is really low.
10598 			 */
10599 			if (freehblkcnt <= HBLK_RESERVE_MIN && !critical) {
10600 				SFMMU_STAT(sf_get_free_throttle);
10601 				mutex_exit(&freehblkp_lock);
10602 				return (0);
10603 			}
10604 			freehblkcnt--;
10605 			*hmeblkpp = freehblkp;
10606 			hblkp = *hmeblkpp;
10607 			freehblkp = hblkp->hblk_next;
10608 			mutex_exit(&freehblkp_lock);
10609 			hblkp->hblk_next = NULL;
10610 			SFMMU_STAT(sf_get_free_success);
10611 			return (1);
10612 		}
10613 		mutex_exit(&freehblkp_lock);
10614 	}
10615 	SFMMU_STAT(sf_get_free_fail);
10616 	return (0);
10617 }
10618 
10619 static uint_t
10620 sfmmu_put_free_hblk(struct hme_blk *hmeblkp, uint_t critical)
10621 {
10622 	struct  hme_blk *hblkp;
10623 
10624 	/*
10625 	 * If the current thread is mapping into kernel space,
10626 	 * let it succede even if freehblkcnt is max
10627 	 * so that it will avoid freeing it to kmem.
10628 	 * This will prevent stack overflow due to
10629 	 * possible recursion since kmem_cache_free()
10630 	 * might require creation of a slab which
10631 	 * in turn needs an hmeblk to map that slab;
10632 	 * let's break this vicious chain at the first
10633 	 * opportunity.
10634 	 */
10635 	if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10636 		mutex_enter(&freehblkp_lock);
10637 		if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
10638 			SFMMU_STAT(sf_put_free_success);
10639 			freehblkcnt++;
10640 			hmeblkp->hblk_next = freehblkp;
10641 			freehblkp = hmeblkp;
10642 			mutex_exit(&freehblkp_lock);
10643 			return (1);
10644 		}
10645 		mutex_exit(&freehblkp_lock);
10646 	}
10647 
10648 	/*
10649 	 * Bring down freehblkcnt to HBLK_RESERVE_CNT. We are here
10650 	 * only if freehblkcnt is at least HBLK_RESERVE_CNT *and*
10651 	 * we are not in the process of mapping into kernel space.
10652 	 */
10653 	ASSERT(!critical);
10654 	while (freehblkcnt > HBLK_RESERVE_CNT) {
10655 		mutex_enter(&freehblkp_lock);
10656 		if (freehblkcnt > HBLK_RESERVE_CNT) {
10657 			freehblkcnt--;
10658 			hblkp = freehblkp;
10659 			freehblkp = hblkp->hblk_next;
10660 			mutex_exit(&freehblkp_lock);
10661 			ASSERT(get_hblk_cache(hblkp) == sfmmu8_cache);
10662 			kmem_cache_free(sfmmu8_cache, hblkp);
10663 			continue;
10664 		}
10665 		mutex_exit(&freehblkp_lock);
10666 	}
10667 	SFMMU_STAT(sf_put_free_fail);
10668 	return (0);
10669 }
10670 
10671 static void
10672 sfmmu_hblk_swap(struct hme_blk *new)
10673 {
10674 	struct hme_blk *old, *hblkp, *prev;
10675 	uint64_t hblkpa, prevpa, newpa;
10676 	caddr_t	base, vaddr, endaddr;
10677 	struct hmehash_bucket *hmebp;
10678 	struct sf_hment *osfhme, *nsfhme;
10679 	page_t *pp;
10680 	kmutex_t *pml;
10681 	tte_t tte;
10682 
10683 #ifdef	DEBUG
10684 	hmeblk_tag		hblktag;
10685 	struct hme_blk		*found;
10686 #endif
10687 	old = HBLK_RESERVE;
10688 	ASSERT(!old->hblk_shared);
10689 
10690 	/*
10691 	 * save pa before bcopy clobbers it
10692 	 */
10693 	newpa = new->hblk_nextpa;
10694 
10695 	base = (caddr_t)get_hblk_base(old);
10696 	endaddr = base + get_hblk_span(old);
10697 
10698 	/*
10699 	 * acquire hash bucket lock.
10700 	 */
10701 	hmebp = sfmmu_tteload_acquire_hashbucket(ksfmmup, base, TTE8K,
10702 	    SFMMU_INVALID_SHMERID);
10703 
10704 	/*
10705 	 * copy contents from old to new
10706 	 */
10707 	bcopy((void *)old, (void *)new, HME8BLK_SZ);
10708 
10709 	/*
10710 	 * add new to hash chain
10711 	 */
10712 	sfmmu_hblk_hash_add(hmebp, new, newpa);
10713 
10714 	/*
10715 	 * search hash chain for hblk_reserve; this needs to be performed
10716 	 * after adding new, otherwise prevpa and prev won't correspond
10717 	 * to the hblk which is prior to old in hash chain when we call
10718 	 * sfmmu_hblk_hash_rm to remove old later.
10719 	 */
10720 	for (prevpa = 0, prev = NULL,
10721 	    hblkpa = hmebp->hmeh_nextpa, hblkp = hmebp->hmeblkp;
10722 	    hblkp != NULL && hblkp != old;
10723 	    prevpa = hblkpa, prev = hblkp,
10724 	    hblkpa = hblkp->hblk_nextpa, hblkp = hblkp->hblk_next)
10725 		;
10726 
10727 	if (hblkp != old)
10728 		panic("sfmmu_hblk_swap: hblk_reserve not found");
10729 
10730 	/*
10731 	 * p_mapping list is still pointing to hments in hblk_reserve;
10732 	 * fix up p_mapping list so that they point to hments in new.
10733 	 *
10734 	 * Since all these mappings are created by hblk_reserve_thread
10735 	 * on the way and it's using at least one of the buffers from each of
10736 	 * the newly minted slabs, there is no danger of any of these
10737 	 * mappings getting unloaded by another thread.
10738 	 *
10739 	 * tsbmiss could only modify ref/mod bits of hments in old/new.
10740 	 * Since all of these hments hold mappings established by segkmem
10741 	 * and mappings in segkmem are setup with HAT_NOSYNC, ref/mod bits
10742 	 * have no meaning for the mappings in hblk_reserve.  hments in
10743 	 * old and new are identical except for ref/mod bits.
10744 	 */
10745 	for (vaddr = base; vaddr < endaddr; vaddr += TTEBYTES(TTE8K)) {
10746 
10747 		HBLKTOHME(osfhme, old, vaddr);
10748 		sfmmu_copytte(&osfhme->hme_tte, &tte);
10749 
10750 		if (TTE_IS_VALID(&tte)) {
10751 			if ((pp = osfhme->hme_page) == NULL)
10752 				panic("sfmmu_hblk_swap: page not mapped");
10753 
10754 			pml = sfmmu_mlist_enter(pp);
10755 
10756 			if (pp != osfhme->hme_page)
10757 				panic("sfmmu_hblk_swap: mapping changed");
10758 
10759 			HBLKTOHME(nsfhme, new, vaddr);
10760 
10761 			HME_ADD(nsfhme, pp);
10762 			HME_SUB(osfhme, pp);
10763 
10764 			sfmmu_mlist_exit(pml);
10765 		}
10766 	}
10767 
10768 	/*
10769 	 * remove old from hash chain
10770 	 */
10771 	sfmmu_hblk_hash_rm(hmebp, old, prevpa, prev);
10772 
10773 #ifdef	DEBUG
10774 
10775 	hblktag.htag_id = ksfmmup;
10776 	hblktag.htag_rid = SFMMU_INVALID_SHMERID;
10777 	hblktag.htag_bspage = HME_HASH_BSPAGE(base, HME_HASH_SHIFT(TTE8K));
10778 	hblktag.htag_rehash = HME_HASH_REHASH(TTE8K);
10779 	HME_HASH_FAST_SEARCH(hmebp, hblktag, found);
10780 
10781 	if (found != new)
10782 		panic("sfmmu_hblk_swap: new hblk not found");
10783 #endif
10784 
10785 	SFMMU_HASH_UNLOCK(hmebp);
10786 
10787 	/*
10788 	 * Reset hblk_reserve
10789 	 */
10790 	bzero((void *)old, HME8BLK_SZ);
10791 	old->hblk_nextpa = va_to_pa((caddr_t)old);
10792 }
10793 
10794 /*
10795  * Grab the mlist mutex for both pages passed in.
10796  *
10797  * low and high will be returned as pointers to the mutexes for these pages.
10798  * low refers to the mutex residing in the lower bin of the mlist hash, while
10799  * high refers to the mutex residing in the higher bin of the mlist hash.  This
10800  * is due to the locking order restrictions on the same thread grabbing
10801  * multiple mlist mutexes.  The low lock must be acquired before the high lock.
10802  *
10803  * If both pages hash to the same mutex, only grab that single mutex, and
10804  * high will be returned as NULL
10805  * If the pages hash to different bins in the hash, grab the lower addressed
10806  * lock first and then the higher addressed lock in order to follow the locking
10807  * rules involved with the same thread grabbing multiple mlist mutexes.
10808  * low and high will both have non-NULL values.
10809  */
10810 static void
10811 sfmmu_mlist_reloc_enter(struct page *targ, struct page *repl,
10812     kmutex_t **low, kmutex_t **high)
10813 {
10814 	kmutex_t	*mml_targ, *mml_repl;
10815 
10816 	/*
10817 	 * no need to do the dance around szc as in sfmmu_mlist_enter()
10818 	 * because this routine is only called by hat_page_relocate() and all
10819 	 * targ and repl pages are already locked EXCL so szc can't change.
10820 	 */
10821 
10822 	mml_targ = MLIST_HASH(PP_PAGEROOT(targ));
10823 	mml_repl = MLIST_HASH(PP_PAGEROOT(repl));
10824 
10825 	if (mml_targ == mml_repl) {
10826 		*low = mml_targ;
10827 		*high = NULL;
10828 	} else {
10829 		if (mml_targ < mml_repl) {
10830 			*low = mml_targ;
10831 			*high = mml_repl;
10832 		} else {
10833 			*low = mml_repl;
10834 			*high = mml_targ;
10835 		}
10836 	}
10837 
10838 	mutex_enter(*low);
10839 	if (*high)
10840 		mutex_enter(*high);
10841 }
10842 
10843 static void
10844 sfmmu_mlist_reloc_exit(kmutex_t *low, kmutex_t *high)
10845 {
10846 	if (high)
10847 		mutex_exit(high);
10848 	mutex_exit(low);
10849 }
10850 
10851 static hatlock_t *
10852 sfmmu_hat_enter(sfmmu_t *sfmmup)
10853 {
10854 	hatlock_t	*hatlockp;
10855 
10856 	if (sfmmup != ksfmmup) {
10857 		hatlockp = TSB_HASH(sfmmup);
10858 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
10859 		return (hatlockp);
10860 	}
10861 	return (NULL);
10862 }
10863 
10864 static hatlock_t *
10865 sfmmu_hat_tryenter(sfmmu_t *sfmmup)
10866 {
10867 	hatlock_t	*hatlockp;
10868 
10869 	if (sfmmup != ksfmmup) {
10870 		hatlockp = TSB_HASH(sfmmup);
10871 		if (mutex_tryenter(HATLOCK_MUTEXP(hatlockp)) == 0)
10872 			return (NULL);
10873 		return (hatlockp);
10874 	}
10875 	return (NULL);
10876 }
10877 
10878 static void
10879 sfmmu_hat_exit(hatlock_t *hatlockp)
10880 {
10881 	if (hatlockp != NULL)
10882 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
10883 }
10884 
10885 static void
10886 sfmmu_hat_lock_all(void)
10887 {
10888 	int i;
10889 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
10890 		mutex_enter(HATLOCK_MUTEXP(&hat_lock[i]));
10891 }
10892 
10893 static void
10894 sfmmu_hat_unlock_all(void)
10895 {
10896 	int i;
10897 	for (i = SFMMU_NUM_LOCK - 1; i >= 0; i--)
10898 		mutex_exit(HATLOCK_MUTEXP(&hat_lock[i]));
10899 }
10900 
10901 int
10902 sfmmu_hat_lock_held(sfmmu_t *sfmmup)
10903 {
10904 	ASSERT(sfmmup != ksfmmup);
10905 	return (MUTEX_HELD(HATLOCK_MUTEXP(TSB_HASH(sfmmup))));
10906 }
10907 
10908 /*
10909  * Locking primitives to provide consistency between ISM unmap
10910  * and other operations.  Since ISM unmap can take a long time, we
10911  * use HAT_ISMBUSY flag (protected by the hatlock) to avoid creating
10912  * contention on the hatlock buckets while ISM segments are being
10913  * unmapped.  The tradeoff is that the flags don't prevent priority
10914  * inversion from occurring, so we must request kernel priority in
10915  * case we have to sleep to keep from getting buried while holding
10916  * the HAT_ISMBUSY flag set, which in turn could block other kernel
10917  * threads from running (for example, in sfmmu_uvatopfn()).
10918  */
10919 static void
10920 sfmmu_ismhat_enter(sfmmu_t *sfmmup, int hatlock_held)
10921 {
10922 	hatlock_t *hatlockp;
10923 
10924 	THREAD_KPRI_REQUEST();
10925 	if (!hatlock_held)
10926 		hatlockp = sfmmu_hat_enter(sfmmup);
10927 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY))
10928 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
10929 	SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
10930 	if (!hatlock_held)
10931 		sfmmu_hat_exit(hatlockp);
10932 }
10933 
10934 static void
10935 sfmmu_ismhat_exit(sfmmu_t *sfmmup, int hatlock_held)
10936 {
10937 	hatlock_t *hatlockp;
10938 
10939 	if (!hatlock_held)
10940 		hatlockp = sfmmu_hat_enter(sfmmup);
10941 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
10942 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
10943 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
10944 	if (!hatlock_held)
10945 		sfmmu_hat_exit(hatlockp);
10946 	THREAD_KPRI_RELEASE();
10947 }
10948 
10949 /*
10950  *
10951  * Algorithm:
10952  *
10953  * (1) if segkmem is not ready, allocate hblk from an array of pre-alloc'ed
10954  *	hblks.
10955  *
10956  * (2) if we are allocating an hblk for mapping a slab in sfmmu_cache,
10957  *
10958  * 		(a) try to return an hblk from reserve pool of free hblks;
10959  *		(b) if the reserve pool is empty, acquire hblk_reserve_lock
10960  *		    and return hblk_reserve.
10961  *
10962  * (3) call kmem_cache_alloc() to allocate hblk;
10963  *
10964  *		(a) if hblk_reserve_lock is held by the current thread,
10965  *		    atomically replace hblk_reserve by the hblk that is
10966  *		    returned by kmem_cache_alloc; release hblk_reserve_lock
10967  *		    and call kmem_cache_alloc() again.
10968  *		(b) if reserve pool is not full, add the hblk that is
10969  *		    returned by kmem_cache_alloc to reserve pool and
10970  *		    call kmem_cache_alloc again.
10971  *
10972  */
10973 static struct hme_blk *
10974 sfmmu_hblk_alloc(sfmmu_t *sfmmup, caddr_t vaddr,
10975 	struct hmehash_bucket *hmebp, uint_t size, hmeblk_tag hblktag,
10976 	uint_t flags, uint_t rid)
10977 {
10978 	struct hme_blk *hmeblkp = NULL;
10979 	struct hme_blk *newhblkp;
10980 	struct hme_blk *shw_hblkp = NULL;
10981 	struct kmem_cache *sfmmu_cache = NULL;
10982 	uint64_t hblkpa;
10983 	ulong_t index;
10984 	uint_t owner;		/* set to 1 if using hblk_reserve */
10985 	uint_t forcefree;
10986 	int sleep;
10987 	sf_srd_t *srdp;
10988 	sf_region_t *rgnp;
10989 
10990 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
10991 	ASSERT(hblktag.htag_rid == rid);
10992 	SFMMU_VALIDATE_HMERID(sfmmup, rid, vaddr, TTEBYTES(size));
10993 	ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
10994 	    IS_P2ALIGNED(vaddr, TTEBYTES(size)));
10995 
10996 	/*
10997 	 * If segkmem is not created yet, allocate from static hmeblks
10998 	 * created at the end of startup_modules().  See the block comment
10999 	 * in startup_modules() describing how we estimate the number of
11000 	 * static hmeblks that will be needed during re-map.
11001 	 */
11002 	if (!hblk_alloc_dynamic) {
11003 
11004 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11005 
11006 		if (size == TTE8K) {
11007 			index = nucleus_hblk8.index;
11008 			if (index >= nucleus_hblk8.len) {
11009 				/*
11010 				 * If we panic here, see startup_modules() to
11011 				 * make sure that we are calculating the
11012 				 * number of hblk8's that we need correctly.
11013 				 */
11014 				prom_panic("no nucleus hblk8 to allocate");
11015 			}
11016 			hmeblkp =
11017 			    (struct hme_blk *)&nucleus_hblk8.list[index];
11018 			nucleus_hblk8.index++;
11019 			SFMMU_STAT(sf_hblk8_nalloc);
11020 		} else {
11021 			index = nucleus_hblk1.index;
11022 			if (nucleus_hblk1.index >= nucleus_hblk1.len) {
11023 				/*
11024 				 * If we panic here, see startup_modules().
11025 				 * Most likely you need to update the
11026 				 * calculation of the number of hblk1 elements
11027 				 * that the kernel needs to boot.
11028 				 */
11029 				prom_panic("no nucleus hblk1 to allocate");
11030 			}
11031 			hmeblkp =
11032 			    (struct hme_blk *)&nucleus_hblk1.list[index];
11033 			nucleus_hblk1.index++;
11034 			SFMMU_STAT(sf_hblk1_nalloc);
11035 		}
11036 
11037 		goto hblk_init;
11038 	}
11039 
11040 	SFMMU_HASH_UNLOCK(hmebp);
11041 
11042 	if (sfmmup != KHATID && !SFMMU_IS_SHMERID_VALID(rid)) {
11043 		if (mmu_page_sizes == max_mmu_page_sizes) {
11044 			if (size < TTE256M)
11045 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
11046 				    size, flags);
11047 		} else {
11048 			if (size < TTE4M)
11049 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
11050 				    size, flags);
11051 		}
11052 	} else if (SFMMU_IS_SHMERID_VALID(rid)) {
11053 		/*
11054 		 * Shared hmes use per region bitmaps in rgn_hmeflag
11055 		 * rather than shadow hmeblks to keep track of the
11056 		 * mapping sizes which have been allocated for the region.
11057 		 * Here we cleanup old invalid hmeblks with this rid,
11058 		 * which may be left around by pageunload().
11059 		 */
11060 		int ttesz;
11061 		caddr_t va;
11062 		caddr_t	eva = vaddr + TTEBYTES(size);
11063 
11064 		ASSERT(sfmmup != KHATID);
11065 
11066 		srdp = sfmmup->sfmmu_srdp;
11067 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11068 		rgnp = srdp->srd_hmergnp[rid];
11069 		ASSERT(rgnp != NULL && rgnp->rgn_id == rid);
11070 		ASSERT(rgnp->rgn_refcnt != 0);
11071 		ASSERT(size <= rgnp->rgn_pgszc);
11072 
11073 		ttesz = HBLK_MIN_TTESZ;
11074 		do {
11075 			if (!(rgnp->rgn_hmeflags & (0x1 << ttesz))) {
11076 				continue;
11077 			}
11078 
11079 			if (ttesz > size && ttesz != HBLK_MIN_TTESZ) {
11080 				sfmmu_cleanup_rhblk(srdp, vaddr, rid, ttesz);
11081 			} else if (ttesz < size) {
11082 				for (va = vaddr; va < eva;
11083 				    va += TTEBYTES(ttesz)) {
11084 					sfmmu_cleanup_rhblk(srdp, va, rid,
11085 					    ttesz);
11086 				}
11087 			}
11088 		} while (++ttesz <= rgnp->rgn_pgszc);
11089 	}
11090 
11091 fill_hblk:
11092 	owner = (hblk_reserve_thread == curthread) ? 1 : 0;
11093 
11094 	if (owner && size == TTE8K) {
11095 
11096 		ASSERT(!SFMMU_IS_SHMERID_VALID(rid));
11097 		/*
11098 		 * We are really in a tight spot. We already own
11099 		 * hblk_reserve and we need another hblk.  In anticipation
11100 		 * of this kind of scenario, we specifically set aside
11101 		 * HBLK_RESERVE_MIN number of hblks to be used exclusively
11102 		 * by owner of hblk_reserve.
11103 		 */
11104 		SFMMU_STAT(sf_hblk_recurse_cnt);
11105 
11106 		if (!sfmmu_get_free_hblk(&hmeblkp, 1))
11107 			panic("sfmmu_hblk_alloc: reserve list is empty");
11108 
11109 		goto hblk_verify;
11110 	}
11111 
11112 	ASSERT(!owner);
11113 
11114 	if ((flags & HAT_NO_KALLOC) == 0) {
11115 
11116 		sfmmu_cache = ((size == TTE8K) ? sfmmu8_cache : sfmmu1_cache);
11117 		sleep = ((sfmmup == KHATID) ? KM_NOSLEEP : KM_SLEEP);
11118 
11119 		if ((hmeblkp = kmem_cache_alloc(sfmmu_cache, sleep)) == NULL) {
11120 			hmeblkp = sfmmu_hblk_steal(size);
11121 		} else {
11122 			/*
11123 			 * if we are the owner of hblk_reserve,
11124 			 * swap hblk_reserve with hmeblkp and
11125 			 * start a fresh life.  Hope things go
11126 			 * better this time.
11127 			 */
11128 			if (hblk_reserve_thread == curthread) {
11129 				ASSERT(sfmmu_cache == sfmmu8_cache);
11130 				sfmmu_hblk_swap(hmeblkp);
11131 				hblk_reserve_thread = NULL;
11132 				mutex_exit(&hblk_reserve_lock);
11133 				goto fill_hblk;
11134 			}
11135 			/*
11136 			 * let's donate this hblk to our reserve list if
11137 			 * we are not mapping kernel range
11138 			 */
11139 			if (size == TTE8K && sfmmup != KHATID)
11140 				if (sfmmu_put_free_hblk(hmeblkp, 0))
11141 					goto fill_hblk;
11142 		}
11143 	} else {
11144 		/*
11145 		 * We are here to map the slab in sfmmu8_cache; let's
11146 		 * check if we could tap our reserve list; if successful,
11147 		 * this will avoid the pain of going thru sfmmu_hblk_swap
11148 		 */
11149 		SFMMU_STAT(sf_hblk_slab_cnt);
11150 		if (!sfmmu_get_free_hblk(&hmeblkp, 0)) {
11151 			/*
11152 			 * let's start hblk_reserve dance
11153 			 */
11154 			SFMMU_STAT(sf_hblk_reserve_cnt);
11155 			owner = 1;
11156 			mutex_enter(&hblk_reserve_lock);
11157 			hmeblkp = HBLK_RESERVE;
11158 			hblk_reserve_thread = curthread;
11159 		}
11160 	}
11161 
11162 hblk_verify:
11163 	ASSERT(hmeblkp != NULL);
11164 	set_hblk_sz(hmeblkp, size);
11165 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
11166 	SFMMU_HASH_LOCK(hmebp);
11167 	HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11168 	if (newhblkp != NULL) {
11169 		SFMMU_HASH_UNLOCK(hmebp);
11170 		if (hmeblkp != HBLK_RESERVE) {
11171 			/*
11172 			 * This is really tricky!
11173 			 *
11174 			 * vmem_alloc(vmem_seg_arena)
11175 			 *  vmem_alloc(vmem_internal_arena)
11176 			 *   segkmem_alloc(heap_arena)
11177 			 *    vmem_alloc(heap_arena)
11178 			 *    page_create()
11179 			 *    hat_memload()
11180 			 *	kmem_cache_free()
11181 			 *	 kmem_cache_alloc()
11182 			 *	  kmem_slab_create()
11183 			 *	   vmem_alloc(kmem_internal_arena)
11184 			 *	    segkmem_alloc(heap_arena)
11185 			 *		vmem_alloc(heap_arena)
11186 			 *		page_create()
11187 			 *		hat_memload()
11188 			 *		  kmem_cache_free()
11189 			 *		...
11190 			 *
11191 			 * Thus, hat_memload() could call kmem_cache_free
11192 			 * for enough number of times that we could easily
11193 			 * hit the bottom of the stack or run out of reserve
11194 			 * list of vmem_seg structs.  So, we must donate
11195 			 * this hblk to reserve list if it's allocated
11196 			 * from sfmmu8_cache *and* mapping kernel range.
11197 			 * We don't need to worry about freeing hmeblk1's
11198 			 * to kmem since they don't map any kmem slabs.
11199 			 *
11200 			 * Note: When segkmem supports largepages, we must
11201 			 * free hmeblk1's to reserve list as well.
11202 			 */
11203 			forcefree = (sfmmup == KHATID) ? 1 : 0;
11204 			if (size == TTE8K &&
11205 			    sfmmu_put_free_hblk(hmeblkp, forcefree)) {
11206 				goto re_verify;
11207 			}
11208 			ASSERT(sfmmup != KHATID);
11209 			kmem_cache_free(get_hblk_cache(hmeblkp), hmeblkp);
11210 		} else {
11211 			/*
11212 			 * Hey! we don't need hblk_reserve any more.
11213 			 */
11214 			ASSERT(owner);
11215 			hblk_reserve_thread = NULL;
11216 			mutex_exit(&hblk_reserve_lock);
11217 			owner = 0;
11218 		}
11219 re_verify:
11220 		/*
11221 		 * let's check if the goodies are still present
11222 		 */
11223 		SFMMU_HASH_LOCK(hmebp);
11224 		HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
11225 		if (newhblkp != NULL) {
11226 			/*
11227 			 * return newhblkp if it's not hblk_reserve;
11228 			 * if newhblkp is hblk_reserve, return it
11229 			 * _only if_ we are the owner of hblk_reserve.
11230 			 */
11231 			if (newhblkp != HBLK_RESERVE || owner) {
11232 				ASSERT(!SFMMU_IS_SHMERID_VALID(rid) ||
11233 				    newhblkp->hblk_shared);
11234 				ASSERT(SFMMU_IS_SHMERID_VALID(rid) ||
11235 				    !newhblkp->hblk_shared);
11236 				return (newhblkp);
11237 			} else {
11238 				/*
11239 				 * we just hit hblk_reserve in the hash and
11240 				 * we are not the owner of that;
11241 				 *
11242 				 * block until hblk_reserve_thread completes
11243 				 * swapping hblk_reserve and try the dance
11244 				 * once again.
11245 				 */
11246 				SFMMU_HASH_UNLOCK(hmebp);
11247 				mutex_enter(&hblk_reserve_lock);
11248 				mutex_exit(&hblk_reserve_lock);
11249 				SFMMU_STAT(sf_hblk_reserve_hit);
11250 				goto fill_hblk;
11251 			}
11252 		} else {
11253 			/*
11254 			 * it's no more! try the dance once again.
11255 			 */
11256 			SFMMU_HASH_UNLOCK(hmebp);
11257 			goto fill_hblk;
11258 		}
11259 	}
11260 
11261 hblk_init:
11262 	if (SFMMU_IS_SHMERID_VALID(rid)) {
11263 		uint16_t tteflag = 0x1 <<
11264 		    ((size < HBLK_MIN_TTESZ) ? HBLK_MIN_TTESZ : size);
11265 
11266 		if (!(rgnp->rgn_hmeflags & tteflag)) {
11267 			atomic_or_16(&rgnp->rgn_hmeflags, tteflag);
11268 		}
11269 		hmeblkp->hblk_shared = 1;
11270 	} else {
11271 		hmeblkp->hblk_shared = 0;
11272 	}
11273 	set_hblk_sz(hmeblkp, size);
11274 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11275 	hmeblkp->hblk_next = (struct hme_blk *)NULL;
11276 	hmeblkp->hblk_tag = hblktag;
11277 	hmeblkp->hblk_shadow = shw_hblkp;
11278 	hblkpa = hmeblkp->hblk_nextpa;
11279 	hmeblkp->hblk_nextpa = 0;
11280 
11281 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
11282 	ASSERT(get_hblk_span(hmeblkp) == HMEBLK_SPAN(size));
11283 	ASSERT(hmeblkp->hblk_hmecnt == 0);
11284 	ASSERT(hmeblkp->hblk_vcnt == 0);
11285 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11286 	ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
11287 	sfmmu_hblk_hash_add(hmebp, hmeblkp, hblkpa);
11288 	return (hmeblkp);
11289 }
11290 
11291 /*
11292  * This function performs any cleanup required on the hme_blk
11293  * and returns it to the free list.
11294  */
11295 /* ARGSUSED */
11296 static void
11297 sfmmu_hblk_free(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
11298 	uint64_t hblkpa, struct hme_blk **listp)
11299 {
11300 	int shw_size, vshift;
11301 	struct hme_blk *shw_hblkp;
11302 	uint_t		shw_mask, newshw_mask;
11303 	caddr_t		vaddr;
11304 	int		size;
11305 	uint_t		critical;
11306 
11307 	ASSERT(hmeblkp);
11308 	ASSERT(!hmeblkp->hblk_hmecnt);
11309 	ASSERT(!hmeblkp->hblk_vcnt);
11310 	ASSERT(!hmeblkp->hblk_lckcnt);
11311 	ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
11312 	ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
11313 
11314 	critical = (hblktosfmmu(hmeblkp) == KHATID) ? 1 : 0;
11315 
11316 	size = get_hblk_ttesz(hmeblkp);
11317 	shw_hblkp = hmeblkp->hblk_shadow;
11318 	if (shw_hblkp) {
11319 		ASSERT(hblktosfmmu(hmeblkp) != KHATID);
11320 		ASSERT(!hmeblkp->hblk_shared);
11321 		if (mmu_page_sizes == max_mmu_page_sizes) {
11322 			ASSERT(size < TTE256M);
11323 		} else {
11324 			ASSERT(size < TTE4M);
11325 		}
11326 
11327 		shw_size = get_hblk_ttesz(shw_hblkp);
11328 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
11329 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
11330 		ASSERT(vshift < 8);
11331 		/*
11332 		 * Atomically clear shadow mask bit
11333 		 */
11334 		do {
11335 			shw_mask = shw_hblkp->hblk_shw_mask;
11336 			ASSERT(shw_mask & (1 << vshift));
11337 			newshw_mask = shw_mask & ~(1 << vshift);
11338 			newshw_mask = cas32(&shw_hblkp->hblk_shw_mask,
11339 			    shw_mask, newshw_mask);
11340 		} while (newshw_mask != shw_mask);
11341 		hmeblkp->hblk_shadow = NULL;
11342 	}
11343 	hmeblkp->hblk_next = NULL;
11344 	hmeblkp->hblk_nextpa = hblkpa;
11345 	hmeblkp->hblk_shw_bit = 0;
11346 
11347 	if (hmeblkp->hblk_shared) {
11348 		sf_srd_t	*srdp;
11349 		sf_region_t	*rgnp;
11350 		uint_t		rid;
11351 
11352 		srdp = hblktosrd(hmeblkp);
11353 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11354 		rid = hmeblkp->hblk_tag.htag_rid;
11355 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
11356 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
11357 		rgnp = srdp->srd_hmergnp[rid];
11358 		ASSERT(rgnp != NULL);
11359 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
11360 		hmeblkp->hblk_shared = 0;
11361 	}
11362 
11363 	if (hmeblkp->hblk_nuc_bit == 0) {
11364 
11365 		if (size == TTE8K && sfmmu_put_free_hblk(hmeblkp, critical))
11366 			return;
11367 
11368 		hmeblkp->hblk_next = *listp;
11369 		*listp = hmeblkp;
11370 	}
11371 }
11372 
11373 static void
11374 sfmmu_hblks_list_purge(struct hme_blk **listp)
11375 {
11376 	struct hme_blk	*hmeblkp;
11377 
11378 	while ((hmeblkp = *listp) != NULL) {
11379 		*listp = hmeblkp->hblk_next;
11380 		kmem_cache_free(get_hblk_cache(hmeblkp), hmeblkp);
11381 	}
11382 }
11383 
11384 #define	BUCKETS_TO_SEARCH_BEFORE_UNLOAD	30
11385 #define	SFMMU_HBLK_STEAL_THRESHOLD 5
11386 
11387 static uint_t sfmmu_hblk_steal_twice;
11388 static uint_t sfmmu_hblk_steal_count, sfmmu_hblk_steal_unload_count;
11389 
11390 /*
11391  * Steal a hmeblk from user or kernel hme hash lists.
11392  * For 8K tte grab one from reserve pool (freehblkp) before proceeding to
11393  * steal and if we fail to steal after SFMMU_HBLK_STEAL_THRESHOLD attempts
11394  * tap into critical reserve of freehblkp.
11395  * Note: We remain looping in this routine until we find one.
11396  */
11397 static struct hme_blk *
11398 sfmmu_hblk_steal(int size)
11399 {
11400 	static struct hmehash_bucket *uhmehash_steal_hand = NULL;
11401 	struct hmehash_bucket *hmebp;
11402 	struct hme_blk *hmeblkp = NULL, *pr_hblk;
11403 	uint64_t hblkpa, prevpa;
11404 	int i;
11405 	uint_t loop_cnt = 0, critical;
11406 
11407 	for (;;) {
11408 		if (size == TTE8K) {
11409 			critical =
11410 			    (++loop_cnt > SFMMU_HBLK_STEAL_THRESHOLD) ? 1 : 0;
11411 			if (sfmmu_get_free_hblk(&hmeblkp, critical))
11412 				return (hmeblkp);
11413 		}
11414 
11415 		hmebp = (uhmehash_steal_hand == NULL) ? uhme_hash :
11416 		    uhmehash_steal_hand;
11417 		ASSERT(hmebp >= uhme_hash && hmebp <= &uhme_hash[UHMEHASH_SZ]);
11418 
11419 		for (i = 0; hmeblkp == NULL && i <= UHMEHASH_SZ +
11420 		    BUCKETS_TO_SEARCH_BEFORE_UNLOAD; i++) {
11421 			SFMMU_HASH_LOCK(hmebp);
11422 			hmeblkp = hmebp->hmeblkp;
11423 			hblkpa = hmebp->hmeh_nextpa;
11424 			prevpa = 0;
11425 			pr_hblk = NULL;
11426 			while (hmeblkp) {
11427 				/*
11428 				 * check if it is a hmeblk that is not locked
11429 				 * and not shared. skip shadow hmeblks with
11430 				 * shadow_mask set i.e valid count non zero.
11431 				 */
11432 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11433 				    (hmeblkp->hblk_shw_bit == 0 ||
11434 				    hmeblkp->hblk_vcnt == 0) &&
11435 				    (hmeblkp->hblk_lckcnt == 0)) {
11436 					/*
11437 					 * there is a high probability that we
11438 					 * will find a free one. search some
11439 					 * buckets for a free hmeblk initially
11440 					 * before unloading a valid hmeblk.
11441 					 */
11442 					if ((hmeblkp->hblk_vcnt == 0 &&
11443 					    hmeblkp->hblk_hmecnt == 0) || (i >=
11444 					    BUCKETS_TO_SEARCH_BEFORE_UNLOAD)) {
11445 						if (sfmmu_steal_this_hblk(hmebp,
11446 						    hmeblkp, hblkpa, prevpa,
11447 						    pr_hblk)) {
11448 							/*
11449 							 * Hblk is unloaded
11450 							 * successfully
11451 							 */
11452 							break;
11453 						}
11454 					}
11455 				}
11456 				pr_hblk = hmeblkp;
11457 				prevpa = hblkpa;
11458 				hblkpa = hmeblkp->hblk_nextpa;
11459 				hmeblkp = hmeblkp->hblk_next;
11460 			}
11461 
11462 			SFMMU_HASH_UNLOCK(hmebp);
11463 			if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
11464 				hmebp = uhme_hash;
11465 		}
11466 		uhmehash_steal_hand = hmebp;
11467 
11468 		if (hmeblkp != NULL)
11469 			break;
11470 
11471 		/*
11472 		 * in the worst case, look for a free one in the kernel
11473 		 * hash table.
11474 		 */
11475 		for (i = 0, hmebp = khme_hash; i <= KHMEHASH_SZ; i++) {
11476 			SFMMU_HASH_LOCK(hmebp);
11477 			hmeblkp = hmebp->hmeblkp;
11478 			hblkpa = hmebp->hmeh_nextpa;
11479 			prevpa = 0;
11480 			pr_hblk = NULL;
11481 			while (hmeblkp) {
11482 				/*
11483 				 * check if it is free hmeblk
11484 				 */
11485 				if ((get_hblk_ttesz(hmeblkp) == size) &&
11486 				    (hmeblkp->hblk_lckcnt == 0) &&
11487 				    (hmeblkp->hblk_vcnt == 0) &&
11488 				    (hmeblkp->hblk_hmecnt == 0)) {
11489 					if (sfmmu_steal_this_hblk(hmebp,
11490 					    hmeblkp, hblkpa, prevpa, pr_hblk)) {
11491 						break;
11492 					} else {
11493 						/*
11494 						 * Cannot fail since we have
11495 						 * hash lock.
11496 						 */
11497 						panic("fail to steal?");
11498 					}
11499 				}
11500 
11501 				pr_hblk = hmeblkp;
11502 				prevpa = hblkpa;
11503 				hblkpa = hmeblkp->hblk_nextpa;
11504 				hmeblkp = hmeblkp->hblk_next;
11505 			}
11506 
11507 			SFMMU_HASH_UNLOCK(hmebp);
11508 			if (hmebp++ == &khme_hash[KHMEHASH_SZ])
11509 				hmebp = khme_hash;
11510 		}
11511 
11512 		if (hmeblkp != NULL)
11513 			break;
11514 		sfmmu_hblk_steal_twice++;
11515 	}
11516 	return (hmeblkp);
11517 }
11518 
11519 /*
11520  * This routine does real work to prepare a hblk to be "stolen" by
11521  * unloading the mappings, updating shadow counts ....
11522  * It returns 1 if the block is ready to be reused (stolen), or 0
11523  * means the block cannot be stolen yet- pageunload is still working
11524  * on this hblk.
11525  */
11526 static int
11527 sfmmu_steal_this_hblk(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
11528 	uint64_t hblkpa, uint64_t prevpa, struct hme_blk *pr_hblk)
11529 {
11530 	int shw_size, vshift;
11531 	struct hme_blk *shw_hblkp;
11532 	caddr_t vaddr;
11533 	uint_t shw_mask, newshw_mask;
11534 
11535 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
11536 
11537 	/*
11538 	 * check if the hmeblk is free, unload if necessary
11539 	 */
11540 	if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11541 		sfmmu_t *sfmmup;
11542 		demap_range_t dmr;
11543 
11544 		sfmmup = hblktosfmmu(hmeblkp);
11545 		if (hmeblkp->hblk_shared || sfmmup->sfmmu_ismhat) {
11546 			return (0);
11547 		}
11548 		DEMAP_RANGE_INIT(sfmmup, &dmr);
11549 		(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
11550 		    (caddr_t)get_hblk_base(hmeblkp),
11551 		    get_hblk_endaddr(hmeblkp), &dmr, HAT_UNLOAD);
11552 		DEMAP_RANGE_FLUSH(&dmr);
11553 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
11554 			/*
11555 			 * Pageunload is working on the same hblk.
11556 			 */
11557 			return (0);
11558 		}
11559 
11560 		sfmmu_hblk_steal_unload_count++;
11561 	}
11562 
11563 	ASSERT(hmeblkp->hblk_lckcnt == 0);
11564 	ASSERT(hmeblkp->hblk_vcnt == 0 && hmeblkp->hblk_hmecnt == 0);
11565 
11566 	sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa, pr_hblk);
11567 	hmeblkp->hblk_nextpa = hblkpa;
11568 
11569 	shw_hblkp = hmeblkp->hblk_shadow;
11570 	if (shw_hblkp) {
11571 		ASSERT(!hmeblkp->hblk_shared);
11572 		shw_size = get_hblk_ttesz(shw_hblkp);
11573 		vaddr = (caddr_t)get_hblk_base(hmeblkp);
11574 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
11575 		ASSERT(vshift < 8);
11576 		/*
11577 		 * Atomically clear shadow mask bit
11578 		 */
11579 		do {
11580 			shw_mask = shw_hblkp->hblk_shw_mask;
11581 			ASSERT(shw_mask & (1 << vshift));
11582 			newshw_mask = shw_mask & ~(1 << vshift);
11583 			newshw_mask = cas32(&shw_hblkp->hblk_shw_mask,
11584 			    shw_mask, newshw_mask);
11585 		} while (newshw_mask != shw_mask);
11586 		hmeblkp->hblk_shadow = NULL;
11587 	}
11588 
11589 	/*
11590 	 * remove shadow bit if we are stealing an unused shadow hmeblk.
11591 	 * sfmmu_hblk_alloc needs it that way, will set shadow bit later if
11592 	 * we are indeed allocating a shadow hmeblk.
11593 	 */
11594 	hmeblkp->hblk_shw_bit = 0;
11595 
11596 	if (hmeblkp->hblk_shared) {
11597 		sf_srd_t	*srdp;
11598 		sf_region_t	*rgnp;
11599 		uint_t		rid;
11600 
11601 		srdp = hblktosrd(hmeblkp);
11602 		ASSERT(srdp != NULL && srdp->srd_refcnt != 0);
11603 		rid = hmeblkp->hblk_tag.htag_rid;
11604 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
11605 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
11606 		rgnp = srdp->srd_hmergnp[rid];
11607 		ASSERT(rgnp != NULL);
11608 		SFMMU_VALIDATE_SHAREDHBLK(hmeblkp, srdp, rgnp, rid);
11609 		hmeblkp->hblk_shared = 0;
11610 	}
11611 
11612 	sfmmu_hblk_steal_count++;
11613 	SFMMU_STAT(sf_steal_count);
11614 
11615 	return (1);
11616 }
11617 
11618 struct hme_blk *
11619 sfmmu_hmetohblk(struct sf_hment *sfhme)
11620 {
11621 	struct hme_blk *hmeblkp;
11622 	struct sf_hment *sfhme0;
11623 	struct hme_blk *hblk_dummy = 0;
11624 
11625 	/*
11626 	 * No dummy sf_hments, please.
11627 	 */
11628 	ASSERT(sfhme->hme_tte.ll != 0);
11629 
11630 	sfhme0 = sfhme - sfhme->hme_tte.tte_hmenum;
11631 	hmeblkp = (struct hme_blk *)((uintptr_t)sfhme0 -
11632 	    (uintptr_t)&hblk_dummy->hblk_hme[0]);
11633 
11634 	return (hmeblkp);
11635 }
11636 
11637 /*
11638  * On swapin, get appropriately sized TSB(s) and clear the HAT_SWAPPED flag.
11639  * If we can't get appropriately sized TSB(s), try for 8K TSB(s) using
11640  * KM_SLEEP allocation.
11641  *
11642  * Return 0 on success, -1 otherwise.
11643  */
11644 static void
11645 sfmmu_tsb_swapin(sfmmu_t *sfmmup, hatlock_t *hatlockp)
11646 {
11647 	struct tsb_info *tsbinfop, *next;
11648 	tsb_replace_rc_t rc;
11649 	boolean_t gotfirst = B_FALSE;
11650 
11651 	ASSERT(sfmmup != ksfmmup);
11652 	ASSERT(sfmmu_hat_lock_held(sfmmup));
11653 
11654 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPIN)) {
11655 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
11656 	}
11657 
11658 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
11659 		SFMMU_FLAGS_SET(sfmmup, HAT_SWAPIN);
11660 	} else {
11661 		return;
11662 	}
11663 
11664 	ASSERT(sfmmup->sfmmu_tsb != NULL);
11665 
11666 	/*
11667 	 * Loop over all tsbinfo's replacing them with ones that actually have
11668 	 * a TSB.  If any of the replacements ever fail, bail out of the loop.
11669 	 */
11670 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; tsbinfop = next) {
11671 		ASSERT(tsbinfop->tsb_flags & TSB_SWAPPED);
11672 		next = tsbinfop->tsb_next;
11673 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, tsbinfop->tsb_szc,
11674 		    hatlockp, TSB_SWAPIN);
11675 		if (rc != TSB_SUCCESS) {
11676 			break;
11677 		}
11678 		gotfirst = B_TRUE;
11679 	}
11680 
11681 	switch (rc) {
11682 	case TSB_SUCCESS:
11683 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11684 		cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11685 		return;
11686 	case TSB_LOSTRACE:
11687 		break;
11688 	case TSB_ALLOCFAIL:
11689 		break;
11690 	default:
11691 		panic("sfmmu_replace_tsb returned unrecognized failure code "
11692 		    "%d", rc);
11693 	}
11694 
11695 	/*
11696 	 * In this case, we failed to get one of our TSBs.  If we failed to
11697 	 * get the first TSB, get one of minimum size (8KB).  Walk the list
11698 	 * and throw away the tsbinfos, starting where the allocation failed;
11699 	 * we can get by with just one TSB as long as we don't leave the
11700 	 * SWAPPED tsbinfo structures lying around.
11701 	 */
11702 	tsbinfop = sfmmup->sfmmu_tsb;
11703 	next = tsbinfop->tsb_next;
11704 	tsbinfop->tsb_next = NULL;
11705 
11706 	sfmmu_hat_exit(hatlockp);
11707 	for (tsbinfop = next; tsbinfop != NULL; tsbinfop = next) {
11708 		next = tsbinfop->tsb_next;
11709 		sfmmu_tsbinfo_free(tsbinfop);
11710 	}
11711 	hatlockp = sfmmu_hat_enter(sfmmup);
11712 
11713 	/*
11714 	 * If we don't have any TSBs, get a single 8K TSB for 8K, 64K and 512K
11715 	 * pages.
11716 	 */
11717 	if (!gotfirst) {
11718 		tsbinfop = sfmmup->sfmmu_tsb;
11719 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, TSB_MIN_SZCODE,
11720 		    hatlockp, TSB_SWAPIN | TSB_FORCEALLOC);
11721 		ASSERT(rc == TSB_SUCCESS);
11722 	}
11723 
11724 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
11725 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11726 }
11727 
11728 static int
11729 sfmmu_is_rgnva(sf_srd_t *srdp, caddr_t addr, ulong_t w, ulong_t bmw)
11730 {
11731 	ulong_t bix = 0;
11732 	uint_t rid;
11733 	sf_region_t *rgnp;
11734 
11735 	ASSERT(srdp != NULL);
11736 	ASSERT(srdp->srd_refcnt != 0);
11737 
11738 	w <<= BT_ULSHIFT;
11739 	while (bmw) {
11740 		if (!(bmw & 0x1)) {
11741 			bix++;
11742 			bmw >>= 1;
11743 			continue;
11744 		}
11745 		rid = w | bix;
11746 		rgnp = srdp->srd_hmergnp[rid];
11747 		ASSERT(rgnp->rgn_refcnt > 0);
11748 		ASSERT(rgnp->rgn_id == rid);
11749 		if (addr < rgnp->rgn_saddr ||
11750 		    addr >= (rgnp->rgn_saddr + rgnp->rgn_size)) {
11751 			bix++;
11752 			bmw >>= 1;
11753 		} else {
11754 			return (1);
11755 		}
11756 	}
11757 	return (0);
11758 }
11759 
11760 /*
11761  * Handle exceptions for low level tsb_handler.
11762  *
11763  * There are many scenarios that could land us here:
11764  *
11765  * If the context is invalid we land here. The context can be invalid
11766  * for 3 reasons: 1) we couldn't allocate a new context and now need to
11767  * perform a wrap around operation in order to allocate a new context.
11768  * 2) Context was invalidated to change pagesize programming 3) ISMs or
11769  * TSBs configuration is changeing for this process and we are forced into
11770  * here to do a syncronization operation. If the context is valid we can
11771  * be here from window trap hanlder. In this case just call trap to handle
11772  * the fault.
11773  *
11774  * Note that the process will run in INVALID_CONTEXT before
11775  * faulting into here and subsequently loading the MMU registers
11776  * (including the TSB base register) associated with this process.
11777  * For this reason, the trap handlers must all test for
11778  * INVALID_CONTEXT before attempting to access any registers other
11779  * than the context registers.
11780  */
11781 void
11782 sfmmu_tsbmiss_exception(struct regs *rp, uintptr_t tagaccess, uint_t traptype)
11783 {
11784 	sfmmu_t *sfmmup, *shsfmmup;
11785 	uint_t ctxtype;
11786 	klwp_id_t lwp;
11787 	char lwp_save_state;
11788 	hatlock_t *hatlockp, *shatlockp;
11789 	struct tsb_info *tsbinfop;
11790 	struct tsbmiss *tsbmp;
11791 	sf_scd_t *scdp;
11792 
11793 	SFMMU_STAT(sf_tsb_exceptions);
11794 	SFMMU_MMU_STAT(mmu_tsb_exceptions);
11795 	sfmmup = astosfmmu(curthread->t_procp->p_as);
11796 	/*
11797 	 * note that in sun4u, tagacces register contains ctxnum
11798 	 * while sun4v passes ctxtype in the tagaccess register.
11799 	 */
11800 	ctxtype = tagaccess & TAGACC_CTX_MASK;
11801 
11802 	ASSERT(sfmmup != ksfmmup && ctxtype != KCONTEXT);
11803 	ASSERT(sfmmup->sfmmu_ismhat == 0);
11804 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED) ||
11805 	    ctxtype == INVALID_CONTEXT);
11806 
11807 	if (ctxtype != INVALID_CONTEXT && traptype != T_DATA_PROT) {
11808 		/*
11809 		 * We may land here because shme bitmap and pagesize
11810 		 * flags are updated lazily in tsbmiss area on other cpus.
11811 		 * If we detect here that tsbmiss area is out of sync with
11812 		 * sfmmu update it and retry the trapped instruction.
11813 		 * Otherwise call trap().
11814 		 */
11815 		int ret = 0;
11816 		uchar_t tteflag_mask = (1 << TTE64K) | (1 << TTE8K);
11817 		caddr_t addr = (caddr_t)(tagaccess & TAGACC_VADDR_MASK);
11818 
11819 		/*
11820 		 * Must set lwp state to LWP_SYS before
11821 		 * trying to acquire any adaptive lock
11822 		 */
11823 		lwp = ttolwp(curthread);
11824 		ASSERT(lwp);
11825 		lwp_save_state = lwp->lwp_state;
11826 		lwp->lwp_state = LWP_SYS;
11827 
11828 		hatlockp = sfmmu_hat_enter(sfmmup);
11829 		kpreempt_disable();
11830 		tsbmp = &tsbmiss_area[CPU->cpu_id];
11831 		ASSERT(sfmmup == tsbmp->usfmmup);
11832 		if (((tsbmp->uhat_tteflags ^ sfmmup->sfmmu_tteflags) &
11833 		    ~tteflag_mask) ||
11834 		    ((tsbmp->uhat_rtteflags ^  sfmmup->sfmmu_rtteflags) &
11835 		    ~tteflag_mask)) {
11836 			tsbmp->uhat_tteflags = sfmmup->sfmmu_tteflags;
11837 			tsbmp->uhat_rtteflags = sfmmup->sfmmu_rtteflags;
11838 			ret = 1;
11839 		}
11840 		if (sfmmup->sfmmu_srdp != NULL) {
11841 			ulong_t *sm = sfmmup->sfmmu_hmeregion_map.bitmap;
11842 			ulong_t *tm = tsbmp->shmermap;
11843 			ulong_t i;
11844 			for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
11845 				ulong_t d = tm[i] ^ sm[i];
11846 				if (d) {
11847 					if (d & sm[i]) {
11848 						if (!ret && sfmmu_is_rgnva(
11849 						    sfmmup->sfmmu_srdp,
11850 						    addr, i, d & sm[i])) {
11851 							ret = 1;
11852 						}
11853 					}
11854 					tm[i] = sm[i];
11855 				}
11856 			}
11857 		}
11858 		kpreempt_enable();
11859 		sfmmu_hat_exit(hatlockp);
11860 		lwp->lwp_state = lwp_save_state;
11861 		if (ret) {
11862 			return;
11863 		}
11864 	} else if (ctxtype == INVALID_CONTEXT) {
11865 		/*
11866 		 * First, make sure we come out of here with a valid ctx,
11867 		 * since if we don't get one we'll simply loop on the
11868 		 * faulting instruction.
11869 		 *
11870 		 * If the ISM mappings are changing, the TSB is relocated,
11871 		 * the process is swapped, the process is joining SCD or
11872 		 * leaving SCD or shared regions we serialize behind the
11873 		 * controlling thread with hat lock, sfmmu_flags and
11874 		 * sfmmu_tsb_cv condition variable.
11875 		 */
11876 
11877 		/*
11878 		 * Must set lwp state to LWP_SYS before
11879 		 * trying to acquire any adaptive lock
11880 		 */
11881 		lwp = ttolwp(curthread);
11882 		ASSERT(lwp);
11883 		lwp_save_state = lwp->lwp_state;
11884 		lwp->lwp_state = LWP_SYS;
11885 
11886 		hatlockp = sfmmu_hat_enter(sfmmup);
11887 retry:
11888 		if ((scdp = sfmmup->sfmmu_scdp) != NULL) {
11889 			shsfmmup = scdp->scd_sfmmup;
11890 			ASSERT(shsfmmup != NULL);
11891 
11892 			for (tsbinfop = shsfmmup->sfmmu_tsb; tsbinfop != NULL;
11893 			    tsbinfop = tsbinfop->tsb_next) {
11894 				if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
11895 					/* drop the private hat lock */
11896 					sfmmu_hat_exit(hatlockp);
11897 					/* acquire the shared hat lock */
11898 					shatlockp = sfmmu_hat_enter(shsfmmup);
11899 					/*
11900 					 * recheck to see if anything changed
11901 					 * after we drop the private hat lock.
11902 					 */
11903 					if (sfmmup->sfmmu_scdp == scdp &&
11904 					    shsfmmup == scdp->scd_sfmmup) {
11905 						sfmmu_tsb_chk_reloc(shsfmmup,
11906 						    shatlockp);
11907 					}
11908 					sfmmu_hat_exit(shatlockp);
11909 					hatlockp = sfmmu_hat_enter(sfmmup);
11910 					goto retry;
11911 				}
11912 			}
11913 		}
11914 
11915 		for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
11916 		    tsbinfop = tsbinfop->tsb_next) {
11917 			if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
11918 				cv_wait(&sfmmup->sfmmu_tsb_cv,
11919 				    HATLOCK_MUTEXP(hatlockp));
11920 				goto retry;
11921 			}
11922 		}
11923 
11924 		/*
11925 		 * Wait for ISM maps to be updated.
11926 		 */
11927 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
11928 			cv_wait(&sfmmup->sfmmu_tsb_cv,
11929 			    HATLOCK_MUTEXP(hatlockp));
11930 			goto retry;
11931 		}
11932 
11933 		/* Is this process joining an SCD? */
11934 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
11935 			/*
11936 			 * Flush private TSB and setup shared TSB.
11937 			 * sfmmu_finish_join_scd() does not drop the
11938 			 * hat lock.
11939 			 */
11940 			sfmmu_finish_join_scd(sfmmup);
11941 			SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
11942 		}
11943 
11944 		/*
11945 		 * If we're swapping in, get TSB(s).  Note that we must do
11946 		 * this before we get a ctx or load the MMU state.  Once
11947 		 * we swap in we have to recheck to make sure the TSB(s) and
11948 		 * ISM mappings didn't change while we slept.
11949 		 */
11950 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
11951 			sfmmu_tsb_swapin(sfmmup, hatlockp);
11952 			goto retry;
11953 		}
11954 
11955 		sfmmu_get_ctx(sfmmup);
11956 
11957 		sfmmu_hat_exit(hatlockp);
11958 		/*
11959 		 * Must restore lwp_state if not calling
11960 		 * trap() for further processing. Restore
11961 		 * it anyway.
11962 		 */
11963 		lwp->lwp_state = lwp_save_state;
11964 		return;
11965 	}
11966 	trap(rp, (caddr_t)tagaccess, traptype, 0);
11967 }
11968 
11969 static void
11970 sfmmu_tsb_chk_reloc(sfmmu_t *sfmmup, hatlock_t *hatlockp)
11971 {
11972 	struct tsb_info *tp;
11973 
11974 	ASSERT(sfmmu_hat_lock_held(sfmmup));
11975 
11976 	for (tp = sfmmup->sfmmu_tsb; tp != NULL; tp = tp->tsb_next) {
11977 		if (tp->tsb_flags & TSB_RELOC_FLAG) {
11978 			cv_wait(&sfmmup->sfmmu_tsb_cv,
11979 			    HATLOCK_MUTEXP(hatlockp));
11980 			break;
11981 		}
11982 	}
11983 }
11984 
11985 /*
11986  * sfmmu_vatopfn_suspended is called from GET_TTE when TL=0 and
11987  * TTE_SUSPENDED bit set in tte we block on aquiring a page lock
11988  * rather than spinning to avoid send mondo timeouts with
11989  * interrupts enabled. When the lock is acquired it is immediately
11990  * released and we return back to sfmmu_vatopfn just after
11991  * the GET_TTE call.
11992  */
11993 void
11994 sfmmu_vatopfn_suspended(caddr_t vaddr, sfmmu_t *sfmmu, tte_t *ttep)
11995 {
11996 	struct page	**pp;
11997 
11998 	(void) as_pagelock(sfmmu->sfmmu_as, &pp, vaddr, TTE_CSZ(ttep), S_WRITE);
11999 	as_pageunlock(sfmmu->sfmmu_as, pp, vaddr, TTE_CSZ(ttep), S_WRITE);
12000 }
12001 
12002 /*
12003  * sfmmu_tsbmiss_suspended is called from GET_TTE when TL>0 and
12004  * TTE_SUSPENDED bit set in tte. We do this so that we can handle
12005  * cross traps which cannot be handled while spinning in the
12006  * trap handlers. Simply enter and exit the kpr_suspendlock spin
12007  * mutex, which is held by the holder of the suspend bit, and then
12008  * retry the trapped instruction after unwinding.
12009  */
12010 /*ARGSUSED*/
12011 void
12012 sfmmu_tsbmiss_suspended(struct regs *rp, uintptr_t tagacc, uint_t traptype)
12013 {
12014 	ASSERT(curthread != kreloc_thread);
12015 	mutex_enter(&kpr_suspendlock);
12016 	mutex_exit(&kpr_suspendlock);
12017 }
12018 
12019 /*
12020  * This routine could be optimized to reduce the number of xcalls by flushing
12021  * the entire TLBs if region reference count is above some threshold but the
12022  * tradeoff will depend on the size of the TLB. So for now flush the specific
12023  * page a context at a time.
12024  *
12025  * If uselocks is 0 then it's called after all cpus were captured and all the
12026  * hat locks were taken. In this case don't take the region lock by relying on
12027  * the order of list region update operations in hat_join_region(),
12028  * hat_leave_region() and hat_dup_region(). The ordering in those routines
12029  * guarantees that list is always forward walkable and reaches active sfmmus
12030  * regardless of where xc_attention() captures a cpu.
12031  */
12032 cpuset_t
12033 sfmmu_rgntlb_demap(caddr_t addr, sf_region_t *rgnp,
12034     struct hme_blk *hmeblkp, int uselocks)
12035 {
12036 	sfmmu_t	*sfmmup;
12037 	cpuset_t cpuset;
12038 	cpuset_t rcpuset;
12039 	hatlock_t *hatlockp;
12040 	uint_t rid = rgnp->rgn_id;
12041 	sf_rgn_link_t *rlink;
12042 	sf_scd_t *scdp;
12043 
12044 	ASSERT(hmeblkp->hblk_shared);
12045 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
12046 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
12047 
12048 	CPUSET_ZERO(rcpuset);
12049 	if (uselocks) {
12050 		mutex_enter(&rgnp->rgn_mutex);
12051 	}
12052 	sfmmup = rgnp->rgn_sfmmu_head;
12053 	while (sfmmup != NULL) {
12054 		if (uselocks) {
12055 			hatlockp = sfmmu_hat_enter(sfmmup);
12056 		}
12057 
12058 		/*
12059 		 * When an SCD is created the SCD hat is linked on the sfmmu
12060 		 * region lists for each hme region which is part of the
12061 		 * SCD. If we find an SCD hat, when walking these lists,
12062 		 * then we flush the shared TSBs, if we find a private hat,
12063 		 * which is part of an SCD, but where the region
12064 		 * is not part of the SCD then we flush the private TSBs.
12065 		 */
12066 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12067 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
12068 			scdp = sfmmup->sfmmu_scdp;
12069 			if (SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
12070 				if (uselocks) {
12071 					sfmmu_hat_exit(hatlockp);
12072 				}
12073 				goto next;
12074 			}
12075 		}
12076 
12077 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12078 
12079 		kpreempt_disable();
12080 		cpuset = sfmmup->sfmmu_cpusran;
12081 		CPUSET_AND(cpuset, cpu_ready_set);
12082 		CPUSET_DEL(cpuset, CPU->cpu_id);
12083 		SFMMU_XCALL_STATS(sfmmup);
12084 		xt_some(cpuset, vtag_flushpage_tl1,
12085 		    (uint64_t)addr, (uint64_t)sfmmup);
12086 		vtag_flushpage(addr, (uint64_t)sfmmup);
12087 		if (uselocks) {
12088 			sfmmu_hat_exit(hatlockp);
12089 		}
12090 		kpreempt_enable();
12091 		CPUSET_OR(rcpuset, cpuset);
12092 
12093 next:
12094 		/* LINTED: constant in conditional context */
12095 		SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
12096 		ASSERT(rlink != NULL);
12097 		sfmmup = rlink->next;
12098 	}
12099 	if (uselocks) {
12100 		mutex_exit(&rgnp->rgn_mutex);
12101 	}
12102 	return (rcpuset);
12103 }
12104 
12105 /*
12106  * This routine takes an sfmmu pointer and the va for an adddress in an
12107  * ISM region as input and returns the corresponding region id in ism_rid.
12108  * The return value of 1 indicates that a region has been found and ism_rid
12109  * is valid, otherwise 0 is returned.
12110  */
12111 static int
12112 find_ism_rid(sfmmu_t *sfmmup, sfmmu_t *ism_sfmmup, caddr_t va, uint_t *ism_rid)
12113 {
12114 	ism_blk_t	*ism_blkp;
12115 	int		i;
12116 	ism_map_t	*ism_map;
12117 #ifdef DEBUG
12118 	struct hat	*ism_hatid;
12119 #endif
12120 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12121 
12122 	ism_blkp = sfmmup->sfmmu_iblk;
12123 	while (ism_blkp != NULL) {
12124 		ism_map = ism_blkp->iblk_maps;
12125 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
12126 			if ((va >= ism_start(ism_map[i])) &&
12127 			    (va < ism_end(ism_map[i]))) {
12128 
12129 				*ism_rid = ism_map[i].imap_rid;
12130 #ifdef DEBUG
12131 				ism_hatid = ism_map[i].imap_ismhat;
12132 				ASSERT(ism_hatid == ism_sfmmup);
12133 				ASSERT(ism_hatid->sfmmu_ismhat);
12134 #endif
12135 				return (1);
12136 			}
12137 		}
12138 		ism_blkp = ism_blkp->iblk_next;
12139 	}
12140 	return (0);
12141 }
12142 
12143 /*
12144  * Special routine to flush out ism mappings- TSBs, TLBs and D-caches.
12145  * This routine may be called with all cpu's captured. Therefore, the
12146  * caller is responsible for holding all locks and disabling kernel
12147  * preemption.
12148  */
12149 /* ARGSUSED */
12150 static void
12151 sfmmu_ismtlbcache_demap(caddr_t addr, sfmmu_t *ism_sfmmup,
12152 	struct hme_blk *hmeblkp, pfn_t pfnum, int cache_flush_flag)
12153 {
12154 	cpuset_t 	cpuset;
12155 	caddr_t 	va;
12156 	ism_ment_t	*ment;
12157 	sfmmu_t		*sfmmup;
12158 #ifdef VAC
12159 	int 		vcolor;
12160 #endif
12161 
12162 	sf_scd_t	*scdp;
12163 	uint_t		ism_rid;
12164 
12165 	ASSERT(!hmeblkp->hblk_shared);
12166 	/*
12167 	 * Walk the ism_hat's mapping list and flush the page
12168 	 * from every hat sharing this ism_hat. This routine
12169 	 * may be called while all cpu's have been captured.
12170 	 * Therefore we can't attempt to grab any locks. For now
12171 	 * this means we will protect the ism mapping list under
12172 	 * a single lock which will be grabbed by the caller.
12173 	 * If hat_share/unshare scalibility becomes a performance
12174 	 * problem then we may need to re-think ism mapping list locking.
12175 	 */
12176 	ASSERT(ism_sfmmup->sfmmu_ismhat);
12177 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
12178 	addr = addr - ISMID_STARTADDR;
12179 
12180 	for (ment = ism_sfmmup->sfmmu_iment; ment; ment = ment->iment_next) {
12181 
12182 		sfmmup = ment->iment_hat;
12183 
12184 		va = ment->iment_base_va;
12185 		va = (caddr_t)((uintptr_t)va  + (uintptr_t)addr);
12186 
12187 		/*
12188 		 * When an SCD is created the SCD hat is linked on the ism
12189 		 * mapping lists for each ISM segment which is part of the
12190 		 * SCD. If we find an SCD hat, when walking these lists,
12191 		 * then we flush the shared TSBs, if we find a private hat,
12192 		 * which is part of an SCD, but where the region
12193 		 * corresponding to this va is not part of the SCD then we
12194 		 * flush the private TSBs.
12195 		 */
12196 		if (!sfmmup->sfmmu_scdhat && sfmmup->sfmmu_scdp != NULL &&
12197 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD) &&
12198 		    !SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
12199 			if (!find_ism_rid(sfmmup, ism_sfmmup, va,
12200 			    &ism_rid)) {
12201 				cmn_err(CE_PANIC,
12202 				    "can't find matching ISM rid!");
12203 			}
12204 
12205 			scdp = sfmmup->sfmmu_scdp;
12206 			if (SFMMU_IS_ISMRID_VALID(ism_rid) &&
12207 			    SF_RGNMAP_TEST(scdp->scd_ismregion_map,
12208 			    ism_rid)) {
12209 				continue;
12210 			}
12211 		}
12212 		SFMMU_UNLOAD_TSB(va, sfmmup, hmeblkp, 1);
12213 
12214 		cpuset = sfmmup->sfmmu_cpusran;
12215 		CPUSET_AND(cpuset, cpu_ready_set);
12216 		CPUSET_DEL(cpuset, CPU->cpu_id);
12217 		SFMMU_XCALL_STATS(sfmmup);
12218 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)va,
12219 		    (uint64_t)sfmmup);
12220 		vtag_flushpage(va, (uint64_t)sfmmup);
12221 
12222 #ifdef VAC
12223 		/*
12224 		 * Flush D$
12225 		 * When flushing D$ we must flush all
12226 		 * cpu's. See sfmmu_cache_flush().
12227 		 */
12228 		if (cache_flush_flag == CACHE_FLUSH) {
12229 			cpuset = cpu_ready_set;
12230 			CPUSET_DEL(cpuset, CPU->cpu_id);
12231 
12232 			SFMMU_XCALL_STATS(sfmmup);
12233 			vcolor = addr_to_vcolor(va);
12234 			xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12235 			vac_flushpage(pfnum, vcolor);
12236 		}
12237 #endif	/* VAC */
12238 	}
12239 }
12240 
12241 /*
12242  * Demaps the TSB, CPU caches, and flushes all TLBs on all CPUs of
12243  * a particular virtual address and ctx.  If noflush is set we do not
12244  * flush the TLB/TSB.  This function may or may not be called with the
12245  * HAT lock held.
12246  */
12247 static void
12248 sfmmu_tlbcache_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12249 	pfn_t pfnum, int tlb_noflush, int cpu_flag, int cache_flush_flag,
12250 	int hat_lock_held)
12251 {
12252 #ifdef VAC
12253 	int vcolor;
12254 #endif
12255 	cpuset_t cpuset;
12256 	hatlock_t *hatlockp;
12257 
12258 	ASSERT(!hmeblkp->hblk_shared);
12259 
12260 #if defined(lint) && !defined(VAC)
12261 	pfnum = pfnum;
12262 	cpu_flag = cpu_flag;
12263 	cache_flush_flag = cache_flush_flag;
12264 #endif
12265 
12266 	/*
12267 	 * There is no longer a need to protect against ctx being
12268 	 * stolen here since we don't store the ctx in the TSB anymore.
12269 	 */
12270 #ifdef VAC
12271 	vcolor = addr_to_vcolor(addr);
12272 #endif
12273 
12274 	/*
12275 	 * We must hold the hat lock during the flush of TLB,
12276 	 * to avoid a race with sfmmu_invalidate_ctx(), where
12277 	 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12278 	 * causing TLB demap routine to skip flush on that MMU.
12279 	 * If the context on a MMU has already been set to
12280 	 * INVALID_CONTEXT, we just get an extra flush on
12281 	 * that MMU.
12282 	 */
12283 	if (!hat_lock_held && !tlb_noflush)
12284 		hatlockp = sfmmu_hat_enter(sfmmup);
12285 
12286 	kpreempt_disable();
12287 	if (!tlb_noflush) {
12288 		/*
12289 		 * Flush the TSB and TLB.
12290 		 */
12291 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12292 
12293 		cpuset = sfmmup->sfmmu_cpusran;
12294 		CPUSET_AND(cpuset, cpu_ready_set);
12295 		CPUSET_DEL(cpuset, CPU->cpu_id);
12296 
12297 		SFMMU_XCALL_STATS(sfmmup);
12298 
12299 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
12300 		    (uint64_t)sfmmup);
12301 
12302 		vtag_flushpage(addr, (uint64_t)sfmmup);
12303 	}
12304 
12305 	if (!hat_lock_held && !tlb_noflush)
12306 		sfmmu_hat_exit(hatlockp);
12307 
12308 #ifdef VAC
12309 	/*
12310 	 * Flush the D$
12311 	 *
12312 	 * Even if the ctx is stolen, we need to flush the
12313 	 * cache. Our ctx stealer only flushes the TLBs.
12314 	 */
12315 	if (cache_flush_flag == CACHE_FLUSH) {
12316 		if (cpu_flag & FLUSH_ALL_CPUS) {
12317 			cpuset = cpu_ready_set;
12318 		} else {
12319 			cpuset = sfmmup->sfmmu_cpusran;
12320 			CPUSET_AND(cpuset, cpu_ready_set);
12321 		}
12322 		CPUSET_DEL(cpuset, CPU->cpu_id);
12323 		SFMMU_XCALL_STATS(sfmmup);
12324 		xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12325 		vac_flushpage(pfnum, vcolor);
12326 	}
12327 #endif	/* VAC */
12328 	kpreempt_enable();
12329 }
12330 
12331 /*
12332  * Demaps the TSB and flushes all TLBs on all cpus for a particular virtual
12333  * address and ctx.  If noflush is set we do not currently do anything.
12334  * This function may or may not be called with the HAT lock held.
12335  */
12336 static void
12337 sfmmu_tlb_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
12338 	int tlb_noflush, int hat_lock_held)
12339 {
12340 	cpuset_t cpuset;
12341 	hatlock_t *hatlockp;
12342 
12343 	ASSERT(!hmeblkp->hblk_shared);
12344 
12345 	/*
12346 	 * If the process is exiting we have nothing to do.
12347 	 */
12348 	if (tlb_noflush)
12349 		return;
12350 
12351 	/*
12352 	 * Flush TSB.
12353 	 */
12354 	if (!hat_lock_held)
12355 		hatlockp = sfmmu_hat_enter(sfmmup);
12356 	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp, 0);
12357 
12358 	kpreempt_disable();
12359 
12360 	cpuset = sfmmup->sfmmu_cpusran;
12361 	CPUSET_AND(cpuset, cpu_ready_set);
12362 	CPUSET_DEL(cpuset, CPU->cpu_id);
12363 
12364 	SFMMU_XCALL_STATS(sfmmup);
12365 	xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, (uint64_t)sfmmup);
12366 
12367 	vtag_flushpage(addr, (uint64_t)sfmmup);
12368 
12369 	if (!hat_lock_held)
12370 		sfmmu_hat_exit(hatlockp);
12371 
12372 	kpreempt_enable();
12373 
12374 }
12375 
12376 /*
12377  * Special case of sfmmu_tlb_demap for MMU_PAGESIZE hblks. Use the xcall
12378  * call handler that can flush a range of pages to save on xcalls.
12379  */
12380 static int sfmmu_xcall_save;
12381 
12382 /*
12383  * this routine is never used for demaping addresses backed by SRD hmeblks.
12384  */
12385 static void
12386 sfmmu_tlb_range_demap(demap_range_t *dmrp)
12387 {
12388 	sfmmu_t *sfmmup = dmrp->dmr_sfmmup;
12389 	hatlock_t *hatlockp;
12390 	cpuset_t cpuset;
12391 	uint64_t sfmmu_pgcnt;
12392 	pgcnt_t pgcnt = 0;
12393 	int pgunload = 0;
12394 	int dirtypg = 0;
12395 	caddr_t addr = dmrp->dmr_addr;
12396 	caddr_t eaddr;
12397 	uint64_t bitvec = dmrp->dmr_bitvec;
12398 
12399 	ASSERT(bitvec & 1);
12400 
12401 	/*
12402 	 * Flush TSB and calculate number of pages to flush.
12403 	 */
12404 	while (bitvec != 0) {
12405 		dirtypg = 0;
12406 		/*
12407 		 * Find the first page to flush and then count how many
12408 		 * pages there are after it that also need to be flushed.
12409 		 * This way the number of TSB flushes is minimized.
12410 		 */
12411 		while ((bitvec & 1) == 0) {
12412 			pgcnt++;
12413 			addr += MMU_PAGESIZE;
12414 			bitvec >>= 1;
12415 		}
12416 		while (bitvec & 1) {
12417 			dirtypg++;
12418 			bitvec >>= 1;
12419 		}
12420 		eaddr = addr + ptob(dirtypg);
12421 		hatlockp = sfmmu_hat_enter(sfmmup);
12422 		sfmmu_unload_tsb_range(sfmmup, addr, eaddr, TTE8K);
12423 		sfmmu_hat_exit(hatlockp);
12424 		pgunload += dirtypg;
12425 		addr = eaddr;
12426 		pgcnt += dirtypg;
12427 	}
12428 
12429 	ASSERT((pgcnt<<MMU_PAGESHIFT) <= dmrp->dmr_endaddr - dmrp->dmr_addr);
12430 	if (sfmmup->sfmmu_free == 0) {
12431 		addr = dmrp->dmr_addr;
12432 		bitvec = dmrp->dmr_bitvec;
12433 
12434 		/*
12435 		 * make sure it has SFMMU_PGCNT_SHIFT bits only,
12436 		 * as it will be used to pack argument for xt_some
12437 		 */
12438 		ASSERT((pgcnt > 0) &&
12439 		    (pgcnt <= (1 << SFMMU_PGCNT_SHIFT)));
12440 
12441 		/*
12442 		 * Encode pgcnt as (pgcnt -1 ), and pass (pgcnt - 1) in
12443 		 * the low 6 bits of sfmmup. This is doable since pgcnt
12444 		 * always >= 1.
12445 		 */
12446 		ASSERT(!((uint64_t)sfmmup & SFMMU_PGCNT_MASK));
12447 		sfmmu_pgcnt = (uint64_t)sfmmup |
12448 		    ((pgcnt - 1) & SFMMU_PGCNT_MASK);
12449 
12450 		/*
12451 		 * We must hold the hat lock during the flush of TLB,
12452 		 * to avoid a race with sfmmu_invalidate_ctx(), where
12453 		 * sfmmu_cnum on a MMU could be set to INVALID_CONTEXT,
12454 		 * causing TLB demap routine to skip flush on that MMU.
12455 		 * If the context on a MMU has already been set to
12456 		 * INVALID_CONTEXT, we just get an extra flush on
12457 		 * that MMU.
12458 		 */
12459 		hatlockp = sfmmu_hat_enter(sfmmup);
12460 		kpreempt_disable();
12461 
12462 		cpuset = sfmmup->sfmmu_cpusran;
12463 		CPUSET_AND(cpuset, cpu_ready_set);
12464 		CPUSET_DEL(cpuset, CPU->cpu_id);
12465 
12466 		SFMMU_XCALL_STATS(sfmmup);
12467 		xt_some(cpuset, vtag_flush_pgcnt_tl1, (uint64_t)addr,
12468 		    sfmmu_pgcnt);
12469 
12470 		for (; bitvec != 0; bitvec >>= 1) {
12471 			if (bitvec & 1)
12472 				vtag_flushpage(addr, (uint64_t)sfmmup);
12473 			addr += MMU_PAGESIZE;
12474 		}
12475 		kpreempt_enable();
12476 		sfmmu_hat_exit(hatlockp);
12477 
12478 		sfmmu_xcall_save += (pgunload-1);
12479 	}
12480 	dmrp->dmr_bitvec = 0;
12481 }
12482 
12483 /*
12484  * In cases where we need to synchronize with TLB/TSB miss trap
12485  * handlers, _and_ need to flush the TLB, it's a lot easier to
12486  * throw away the context from the process than to do a
12487  * special song and dance to keep things consistent for the
12488  * handlers.
12489  *
12490  * Since the process suddenly ends up without a context and our caller
12491  * holds the hat lock, threads that fault after this function is called
12492  * will pile up on the lock.  We can then do whatever we need to
12493  * atomically from the context of the caller.  The first blocked thread
12494  * to resume executing will get the process a new context, and the
12495  * process will resume executing.
12496  *
12497  * One added advantage of this approach is that on MMUs that
12498  * support a "flush all" operation, we will delay the flush until
12499  * cnum wrap-around, and then flush the TLB one time.  This
12500  * is rather rare, so it's a lot less expensive than making 8000
12501  * x-calls to flush the TLB 8000 times.
12502  *
12503  * A per-process (PP) lock is used to synchronize ctx allocations in
12504  * resume() and ctx invalidations here.
12505  */
12506 static void
12507 sfmmu_invalidate_ctx(sfmmu_t *sfmmup)
12508 {
12509 	cpuset_t cpuset;
12510 	int cnum, currcnum;
12511 	mmu_ctx_t *mmu_ctxp;
12512 	int i;
12513 	uint_t pstate_save;
12514 
12515 	SFMMU_STAT(sf_ctx_inv);
12516 
12517 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12518 	ASSERT(sfmmup != ksfmmup);
12519 
12520 	kpreempt_disable();
12521 
12522 	mmu_ctxp = CPU_MMU_CTXP(CPU);
12523 	ASSERT(mmu_ctxp);
12524 	ASSERT(mmu_ctxp->mmu_idx < max_mmu_ctxdoms);
12525 	ASSERT(mmu_ctxp == mmu_ctxs_tbl[mmu_ctxp->mmu_idx]);
12526 
12527 	currcnum = sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum;
12528 
12529 	pstate_save = sfmmu_disable_intrs();
12530 
12531 	lock_set(&sfmmup->sfmmu_ctx_lock);	/* acquire PP lock */
12532 	/* set HAT cnum invalid across all context domains. */
12533 	for (i = 0; i < max_mmu_ctxdoms; i++) {
12534 
12535 		cnum = 	sfmmup->sfmmu_ctxs[i].cnum;
12536 		if (cnum == INVALID_CONTEXT) {
12537 			continue;
12538 		}
12539 
12540 		sfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
12541 	}
12542 	membar_enter();	/* make sure globally visible to all CPUs */
12543 	lock_clear(&sfmmup->sfmmu_ctx_lock);	/* release PP lock */
12544 
12545 	sfmmu_enable_intrs(pstate_save);
12546 
12547 	cpuset = sfmmup->sfmmu_cpusran;
12548 	CPUSET_DEL(cpuset, CPU->cpu_id);
12549 	CPUSET_AND(cpuset, cpu_ready_set);
12550 	if (!CPUSET_ISNULL(cpuset)) {
12551 		SFMMU_XCALL_STATS(sfmmup);
12552 		xt_some(cpuset, sfmmu_raise_tsb_exception,
12553 		    (uint64_t)sfmmup, INVALID_CONTEXT);
12554 		xt_sync(cpuset);
12555 		SFMMU_STAT(sf_tsb_raise_exception);
12556 		SFMMU_MMU_STAT(mmu_tsb_raise_exception);
12557 	}
12558 
12559 	/*
12560 	 * If the hat to-be-invalidated is the same as the current
12561 	 * process on local CPU we need to invalidate
12562 	 * this CPU context as well.
12563 	 */
12564 	if ((sfmmu_getctx_sec() == currcnum) &&
12565 	    (currcnum != INVALID_CONTEXT)) {
12566 		/* sets shared context to INVALID too */
12567 		sfmmu_setctx_sec(INVALID_CONTEXT);
12568 		sfmmu_clear_utsbinfo();
12569 	}
12570 
12571 	SFMMU_FLAGS_SET(sfmmup, HAT_ALLCTX_INVALID);
12572 
12573 	kpreempt_enable();
12574 
12575 	/*
12576 	 * we hold the hat lock, so nobody should allocate a context
12577 	 * for us yet
12578 	 */
12579 	ASSERT(sfmmup->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum == INVALID_CONTEXT);
12580 }
12581 
12582 #ifdef VAC
12583 /*
12584  * We need to flush the cache in all cpus.  It is possible that
12585  * a process referenced a page as cacheable but has sinced exited
12586  * and cleared the mapping list.  We still to flush it but have no
12587  * state so all cpus is the only alternative.
12588  */
12589 void
12590 sfmmu_cache_flush(pfn_t pfnum, int vcolor)
12591 {
12592 	cpuset_t cpuset;
12593 
12594 	kpreempt_disable();
12595 	cpuset = cpu_ready_set;
12596 	CPUSET_DEL(cpuset, CPU->cpu_id);
12597 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12598 	xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
12599 	xt_sync(cpuset);
12600 	vac_flushpage(pfnum, vcolor);
12601 	kpreempt_enable();
12602 }
12603 
12604 void
12605 sfmmu_cache_flushcolor(int vcolor, pfn_t pfnum)
12606 {
12607 	cpuset_t cpuset;
12608 
12609 	ASSERT(vcolor >= 0);
12610 
12611 	kpreempt_disable();
12612 	cpuset = cpu_ready_set;
12613 	CPUSET_DEL(cpuset, CPU->cpu_id);
12614 	SFMMU_XCALL_STATS(NULL);	/* account to any ctx */
12615 	xt_some(cpuset, vac_flushcolor_tl1, vcolor, pfnum);
12616 	xt_sync(cpuset);
12617 	vac_flushcolor(vcolor, pfnum);
12618 	kpreempt_enable();
12619 }
12620 #endif	/* VAC */
12621 
12622 /*
12623  * We need to prevent processes from accessing the TSB using a cached physical
12624  * address.  It's alright if they try to access the TSB via virtual address
12625  * since they will just fault on that virtual address once the mapping has
12626  * been suspended.
12627  */
12628 #pragma weak sendmondo_in_recover
12629 
12630 /* ARGSUSED */
12631 static int
12632 sfmmu_tsb_pre_relocator(caddr_t va, uint_t tsbsz, uint_t flags, void *tsbinfo)
12633 {
12634 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12635 	sfmmu_t *sfmmup = tsbinfop->tsb_sfmmu;
12636 	hatlock_t *hatlockp;
12637 	sf_scd_t *scdp;
12638 
12639 	if (flags != HAT_PRESUSPEND)
12640 		return (0);
12641 
12642 	/*
12643 	 * If tsb is a shared TSB with TSB_SHAREDCTX set, sfmmup must
12644 	 * be a shared hat, then set SCD's tsbinfo's flag.
12645 	 * If tsb is not shared, sfmmup is a private hat, then set
12646 	 * its private tsbinfo's flag.
12647 	 */
12648 	hatlockp = sfmmu_hat_enter(sfmmup);
12649 	tsbinfop->tsb_flags |= TSB_RELOC_FLAG;
12650 
12651 	if (!(tsbinfop->tsb_flags & TSB_SHAREDCTX)) {
12652 		sfmmu_tsb_inv_ctx(sfmmup);
12653 		sfmmu_hat_exit(hatlockp);
12654 	} else {
12655 		/* release lock on the shared hat */
12656 		sfmmu_hat_exit(hatlockp);
12657 		/* sfmmup is a shared hat */
12658 		ASSERT(sfmmup->sfmmu_scdhat);
12659 		scdp = sfmmup->sfmmu_scdp;
12660 		ASSERT(scdp != NULL);
12661 		/* get private hat from the scd list */
12662 		mutex_enter(&scdp->scd_mutex);
12663 		sfmmup = scdp->scd_sf_list;
12664 		while (sfmmup != NULL) {
12665 			hatlockp = sfmmu_hat_enter(sfmmup);
12666 			/*
12667 			 * We do not call sfmmu_tsb_inv_ctx here because
12668 			 * sendmondo_in_recover check is only needed for
12669 			 * sun4u.
12670 			 */
12671 			sfmmu_invalidate_ctx(sfmmup);
12672 			sfmmu_hat_exit(hatlockp);
12673 			sfmmup = sfmmup->sfmmu_scd_link.next;
12674 
12675 		}
12676 		mutex_exit(&scdp->scd_mutex);
12677 	}
12678 	return (0);
12679 }
12680 
12681 static void
12682 sfmmu_tsb_inv_ctx(sfmmu_t *sfmmup)
12683 {
12684 	extern uint32_t sendmondo_in_recover;
12685 
12686 	ASSERT(sfmmu_hat_lock_held(sfmmup));
12687 
12688 	/*
12689 	 * For Cheetah+ Erratum 25:
12690 	 * Wait for any active recovery to finish.  We can't risk
12691 	 * relocating the TSB of the thread running mondo_recover_proc()
12692 	 * since, if we did that, we would deadlock.  The scenario we are
12693 	 * trying to avoid is as follows:
12694 	 *
12695 	 * THIS CPU			RECOVER CPU
12696 	 * --------			-----------
12697 	 *				Begins recovery, walking through TSB
12698 	 * hat_pagesuspend() TSB TTE
12699 	 *				TLB miss on TSB TTE, spins at TL1
12700 	 * xt_sync()
12701 	 *	send_mondo_timeout()
12702 	 *	mondo_recover_proc()
12703 	 *	((deadlocked))
12704 	 *
12705 	 * The second half of the workaround is that mondo_recover_proc()
12706 	 * checks to see if the tsb_info has the RELOC flag set, and if it
12707 	 * does, it skips over that TSB without ever touching tsbinfop->tsb_va
12708 	 * and hence avoiding the TLB miss that could result in a deadlock.
12709 	 */
12710 	if (&sendmondo_in_recover) {
12711 		membar_enter();	/* make sure RELOC flag visible */
12712 		while (sendmondo_in_recover) {
12713 			drv_usecwait(1);
12714 			membar_consumer();
12715 		}
12716 	}
12717 
12718 	sfmmu_invalidate_ctx(sfmmup);
12719 }
12720 
12721 /* ARGSUSED */
12722 static int
12723 sfmmu_tsb_post_relocator(caddr_t va, uint_t tsbsz, uint_t flags,
12724 	void *tsbinfo, pfn_t newpfn)
12725 {
12726 	hatlock_t *hatlockp;
12727 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
12728 	sfmmu_t	*sfmmup = tsbinfop->tsb_sfmmu;
12729 
12730 	if (flags != HAT_POSTUNSUSPEND)
12731 		return (0);
12732 
12733 	hatlockp = sfmmu_hat_enter(sfmmup);
12734 
12735 	SFMMU_STAT(sf_tsb_reloc);
12736 
12737 	/*
12738 	 * The process may have swapped out while we were relocating one
12739 	 * of its TSBs.  If so, don't bother doing the setup since the
12740 	 * process can't be using the memory anymore.
12741 	 */
12742 	if ((tsbinfop->tsb_flags & TSB_SWAPPED) == 0) {
12743 		ASSERT(va == tsbinfop->tsb_va);
12744 		sfmmu_tsbinfo_setup_phys(tsbinfop, newpfn);
12745 
12746 		if (tsbinfop->tsb_flags & TSB_FLUSH_NEEDED) {
12747 			sfmmu_inv_tsb(tsbinfop->tsb_va,
12748 			    TSB_BYTES(tsbinfop->tsb_szc));
12749 			tsbinfop->tsb_flags &= ~TSB_FLUSH_NEEDED;
12750 		}
12751 	}
12752 
12753 	membar_exit();
12754 	tsbinfop->tsb_flags &= ~TSB_RELOC_FLAG;
12755 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
12756 
12757 	sfmmu_hat_exit(hatlockp);
12758 
12759 	return (0);
12760 }
12761 
12762 /*
12763  * Allocate and initialize a tsb_info structure.  Note that we may or may not
12764  * allocate a TSB here, depending on the flags passed in.
12765  */
12766 static int
12767 sfmmu_tsbinfo_alloc(struct tsb_info **tsbinfopp, int tsb_szc, int tte_sz_mask,
12768 	uint_t flags, sfmmu_t *sfmmup)
12769 {
12770 	int err;
12771 
12772 	*tsbinfopp = (struct tsb_info *)kmem_cache_alloc(
12773 	    sfmmu_tsbinfo_cache, KM_SLEEP);
12774 
12775 	if ((err = sfmmu_init_tsbinfo(*tsbinfopp, tte_sz_mask,
12776 	    tsb_szc, flags, sfmmup)) != 0) {
12777 		kmem_cache_free(sfmmu_tsbinfo_cache, *tsbinfopp);
12778 		SFMMU_STAT(sf_tsb_allocfail);
12779 		*tsbinfopp = NULL;
12780 		return (err);
12781 	}
12782 	SFMMU_STAT(sf_tsb_alloc);
12783 
12784 	/*
12785 	 * Bump the TSB size counters for this TSB size.
12786 	 */
12787 	(*(((int *)&sfmmu_tsbsize_stat) + tsb_szc))++;
12788 	return (0);
12789 }
12790 
12791 static void
12792 sfmmu_tsb_free(struct tsb_info *tsbinfo)
12793 {
12794 	caddr_t tsbva = tsbinfo->tsb_va;
12795 	uint_t tsb_size = TSB_BYTES(tsbinfo->tsb_szc);
12796 	struct kmem_cache *kmem_cachep = tsbinfo->tsb_cache;
12797 	vmem_t	*vmp = tsbinfo->tsb_vmp;
12798 
12799 	/*
12800 	 * If we allocated this TSB from relocatable kernel memory, then we
12801 	 * need to uninstall the callback handler.
12802 	 */
12803 	if (tsbinfo->tsb_cache != sfmmu_tsb8k_cache) {
12804 		uintptr_t slab_mask;
12805 		caddr_t slab_vaddr;
12806 		page_t **ppl;
12807 		int ret;
12808 
12809 		ASSERT(tsb_size <= MMU_PAGESIZE4M || use_bigtsb_arena);
12810 		if (tsb_size > MMU_PAGESIZE4M)
12811 			slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
12812 		else
12813 			slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
12814 		slab_vaddr = (caddr_t)((uintptr_t)tsbva & slab_mask);
12815 
12816 		ret = as_pagelock(&kas, &ppl, slab_vaddr, PAGESIZE, S_WRITE);
12817 		ASSERT(ret == 0);
12818 		hat_delete_callback(tsbva, (uint_t)tsb_size, (void *)tsbinfo,
12819 		    0, NULL);
12820 		as_pageunlock(&kas, ppl, slab_vaddr, PAGESIZE, S_WRITE);
12821 	}
12822 
12823 	if (kmem_cachep != NULL) {
12824 		kmem_cache_free(kmem_cachep, tsbva);
12825 	} else {
12826 		vmem_xfree(vmp, (void *)tsbva, tsb_size);
12827 	}
12828 	tsbinfo->tsb_va = (caddr_t)0xbad00bad;
12829 	atomic_add_64(&tsb_alloc_bytes, -(int64_t)tsb_size);
12830 }
12831 
12832 static void
12833 sfmmu_tsbinfo_free(struct tsb_info *tsbinfo)
12834 {
12835 	if ((tsbinfo->tsb_flags & TSB_SWAPPED) == 0) {
12836 		sfmmu_tsb_free(tsbinfo);
12837 	}
12838 	kmem_cache_free(sfmmu_tsbinfo_cache, tsbinfo);
12839 
12840 }
12841 
12842 /*
12843  * Setup all the references to physical memory for this tsbinfo.
12844  * The underlying page(s) must be locked.
12845  */
12846 static void
12847 sfmmu_tsbinfo_setup_phys(struct tsb_info *tsbinfo, pfn_t pfn)
12848 {
12849 	ASSERT(pfn != PFN_INVALID);
12850 	ASSERT(pfn == va_to_pfn(tsbinfo->tsb_va));
12851 
12852 #ifndef sun4v
12853 	if (tsbinfo->tsb_szc == 0) {
12854 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn,
12855 		    PROT_WRITE|PROT_READ, TTE8K);
12856 	} else {
12857 		/*
12858 		 * Round down PA and use a large mapping; the handlers will
12859 		 * compute the TSB pointer at the correct offset into the
12860 		 * big virtual page.  NOTE: this assumes all TSBs larger
12861 		 * than 8K must come from physically contiguous slabs of
12862 		 * size tsb_slab_size.
12863 		 */
12864 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn & ~tsb_slab_mask,
12865 		    PROT_WRITE|PROT_READ, tsb_slab_ttesz);
12866 	}
12867 	tsbinfo->tsb_pa = ptob(pfn);
12868 
12869 	TTE_SET_LOCKED(&tsbinfo->tsb_tte); /* lock the tte into dtlb */
12870 	TTE_SET_MOD(&tsbinfo->tsb_tte);    /* enable writes */
12871 
12872 	ASSERT(TTE_IS_PRIVILEGED(&tsbinfo->tsb_tte));
12873 	ASSERT(TTE_IS_LOCKED(&tsbinfo->tsb_tte));
12874 #else /* sun4v */
12875 	tsbinfo->tsb_pa = ptob(pfn);
12876 #endif /* sun4v */
12877 }
12878 
12879 
12880 /*
12881  * Returns zero on success, ENOMEM if over the high water mark,
12882  * or EAGAIN if the caller needs to retry with a smaller TSB
12883  * size (or specify TSB_FORCEALLOC if the allocation can't fail).
12884  *
12885  * This call cannot fail to allocate a TSB if TSB_FORCEALLOC
12886  * is specified and the TSB requested is PAGESIZE, though it
12887  * may sleep waiting for memory if sufficient memory is not
12888  * available.
12889  */
12890 static int
12891 sfmmu_init_tsbinfo(struct tsb_info *tsbinfo, int tteszmask,
12892     int tsbcode, uint_t flags, sfmmu_t *sfmmup)
12893 {
12894 	caddr_t vaddr = NULL;
12895 	caddr_t slab_vaddr;
12896 	uintptr_t slab_mask;
12897 	int tsbbytes = TSB_BYTES(tsbcode);
12898 	int lowmem = 0;
12899 	struct kmem_cache *kmem_cachep = NULL;
12900 	vmem_t *vmp = NULL;
12901 	lgrp_id_t lgrpid = LGRP_NONE;
12902 	pfn_t pfn;
12903 	uint_t cbflags = HAC_SLEEP;
12904 	page_t **pplist;
12905 	int ret;
12906 
12907 	ASSERT(tsbbytes <= MMU_PAGESIZE4M || use_bigtsb_arena);
12908 	if (tsbbytes > MMU_PAGESIZE4M)
12909 		slab_mask = ~((uintptr_t)bigtsb_slab_mask) << PAGESHIFT;
12910 	else
12911 		slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
12912 
12913 	if (flags & (TSB_FORCEALLOC | TSB_SWAPIN | TSB_GROW | TSB_SHRINK))
12914 		flags |= TSB_ALLOC;
12915 
12916 	ASSERT((flags & TSB_FORCEALLOC) == 0 || tsbcode == TSB_MIN_SZCODE);
12917 
12918 	tsbinfo->tsb_sfmmu = sfmmup;
12919 
12920 	/*
12921 	 * If not allocating a TSB, set up the tsbinfo, set TSB_SWAPPED, and
12922 	 * return.
12923 	 */
12924 	if ((flags & TSB_ALLOC) == 0) {
12925 		tsbinfo->tsb_szc = tsbcode;
12926 		tsbinfo->tsb_ttesz_mask = tteszmask;
12927 		tsbinfo->tsb_va = (caddr_t)0xbadbadbeef;
12928 		tsbinfo->tsb_pa = -1;
12929 		tsbinfo->tsb_tte.ll = 0;
12930 		tsbinfo->tsb_next = NULL;
12931 		tsbinfo->tsb_flags = TSB_SWAPPED;
12932 		tsbinfo->tsb_cache = NULL;
12933 		tsbinfo->tsb_vmp = NULL;
12934 		return (0);
12935 	}
12936 
12937 #ifdef DEBUG
12938 	/*
12939 	 * For debugging:
12940 	 * Randomly force allocation failures every tsb_alloc_mtbf
12941 	 * tries if TSB_FORCEALLOC is not specified.  This will
12942 	 * return ENOMEM if tsb_alloc_mtbf is odd, or EAGAIN if
12943 	 * it is even, to allow testing of both failure paths...
12944 	 */
12945 	if (tsb_alloc_mtbf && ((flags & TSB_FORCEALLOC) == 0) &&
12946 	    (tsb_alloc_count++ == tsb_alloc_mtbf)) {
12947 		tsb_alloc_count = 0;
12948 		tsb_alloc_fail_mtbf++;
12949 		return ((tsb_alloc_mtbf & 1)? ENOMEM : EAGAIN);
12950 	}
12951 #endif	/* DEBUG */
12952 
12953 	/*
12954 	 * Enforce high water mark if we are not doing a forced allocation
12955 	 * and are not shrinking a process' TSB.
12956 	 */
12957 	if ((flags & TSB_SHRINK) == 0 &&
12958 	    (tsbbytes + tsb_alloc_bytes) > tsb_alloc_hiwater) {
12959 		if ((flags & TSB_FORCEALLOC) == 0)
12960 			return (ENOMEM);
12961 		lowmem = 1;
12962 	}
12963 
12964 	/*
12965 	 * Allocate from the correct location based upon the size of the TSB
12966 	 * compared to the base page size, and what memory conditions dictate.
12967 	 * Note we always do nonblocking allocations from the TSB arena since
12968 	 * we don't want memory fragmentation to cause processes to block
12969 	 * indefinitely waiting for memory; until the kernel algorithms that
12970 	 * coalesce large pages are improved this is our best option.
12971 	 *
12972 	 * Algorithm:
12973 	 *	If allocating a "large" TSB (>8K), allocate from the
12974 	 *		appropriate kmem_tsb_default_arena vmem arena
12975 	 *	else if low on memory or the TSB_FORCEALLOC flag is set or
12976 	 *	tsb_forceheap is set
12977 	 *		Allocate from kernel heap via sfmmu_tsb8k_cache with
12978 	 *		KM_SLEEP (never fails)
12979 	 *	else
12980 	 *		Allocate from appropriate sfmmu_tsb_cache with
12981 	 *		KM_NOSLEEP
12982 	 *	endif
12983 	 */
12984 	if (tsb_lgrp_affinity)
12985 		lgrpid = lgrp_home_id(curthread);
12986 	if (lgrpid == LGRP_NONE)
12987 		lgrpid = 0;	/* use lgrp of boot CPU */
12988 
12989 	if (tsbbytes > MMU_PAGESIZE) {
12990 		if (tsbbytes > MMU_PAGESIZE4M) {
12991 			vmp = kmem_bigtsb_default_arena[lgrpid];
12992 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
12993 			    0, 0, NULL, NULL, VM_NOSLEEP);
12994 		} else {
12995 			vmp = kmem_tsb_default_arena[lgrpid];
12996 			vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes,
12997 			    0, 0, NULL, NULL, VM_NOSLEEP);
12998 		}
12999 #ifdef	DEBUG
13000 	} else if (lowmem || (flags & TSB_FORCEALLOC) || tsb_forceheap) {
13001 #else	/* !DEBUG */
13002 	} else if (lowmem || (flags & TSB_FORCEALLOC)) {
13003 #endif	/* DEBUG */
13004 		kmem_cachep = sfmmu_tsb8k_cache;
13005 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_SLEEP);
13006 		ASSERT(vaddr != NULL);
13007 	} else {
13008 		kmem_cachep = sfmmu_tsb_cache[lgrpid];
13009 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_NOSLEEP);
13010 	}
13011 
13012 	tsbinfo->tsb_cache = kmem_cachep;
13013 	tsbinfo->tsb_vmp = vmp;
13014 
13015 	if (vaddr == NULL) {
13016 		return (EAGAIN);
13017 	}
13018 
13019 	atomic_add_64(&tsb_alloc_bytes, (int64_t)tsbbytes);
13020 	kmem_cachep = tsbinfo->tsb_cache;
13021 
13022 	/*
13023 	 * If we are allocating from outside the cage, then we need to
13024 	 * register a relocation callback handler.  Note that for now
13025 	 * since pseudo mappings always hang off of the slab's root page,
13026 	 * we need only lock the first 8K of the TSB slab.  This is a bit
13027 	 * hacky but it is good for performance.
13028 	 */
13029 	if (kmem_cachep != sfmmu_tsb8k_cache) {
13030 		slab_vaddr = (caddr_t)((uintptr_t)vaddr & slab_mask);
13031 		ret = as_pagelock(&kas, &pplist, slab_vaddr, PAGESIZE, S_WRITE);
13032 		ASSERT(ret == 0);
13033 		ret = hat_add_callback(sfmmu_tsb_cb_id, vaddr, (uint_t)tsbbytes,
13034 		    cbflags, (void *)tsbinfo, &pfn, NULL);
13035 
13036 		/*
13037 		 * Need to free up resources if we could not successfully
13038 		 * add the callback function and return an error condition.
13039 		 */
13040 		if (ret != 0) {
13041 			if (kmem_cachep) {
13042 				kmem_cache_free(kmem_cachep, vaddr);
13043 			} else {
13044 				vmem_xfree(vmp, (void *)vaddr, tsbbytes);
13045 			}
13046 			as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE,
13047 			    S_WRITE);
13048 			return (EAGAIN);
13049 		}
13050 	} else {
13051 		/*
13052 		 * Since allocation of 8K TSBs from heap is rare and occurs
13053 		 * during memory pressure we allocate them from permanent
13054 		 * memory rather than using callbacks to get the PFN.
13055 		 */
13056 		pfn = hat_getpfnum(kas.a_hat, vaddr);
13057 	}
13058 
13059 	tsbinfo->tsb_va = vaddr;
13060 	tsbinfo->tsb_szc = tsbcode;
13061 	tsbinfo->tsb_ttesz_mask = tteszmask;
13062 	tsbinfo->tsb_next = NULL;
13063 	tsbinfo->tsb_flags = 0;
13064 
13065 	sfmmu_tsbinfo_setup_phys(tsbinfo, pfn);
13066 
13067 	sfmmu_inv_tsb(vaddr, tsbbytes);
13068 
13069 	if (kmem_cachep != sfmmu_tsb8k_cache) {
13070 		as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE, S_WRITE);
13071 	}
13072 
13073 	return (0);
13074 }
13075 
13076 /*
13077  * Initialize per cpu tsb and per cpu tsbmiss_area
13078  */
13079 void
13080 sfmmu_init_tsbs(void)
13081 {
13082 	int i;
13083 	struct tsbmiss	*tsbmissp;
13084 	struct kpmtsbm	*kpmtsbmp;
13085 #ifndef sun4v
13086 	extern int	dcache_line_mask;
13087 #endif /* sun4v */
13088 	extern uint_t	vac_colors;
13089 
13090 	/*
13091 	 * Init. tsb miss area.
13092 	 */
13093 	tsbmissp = tsbmiss_area;
13094 
13095 	for (i = 0; i < NCPU; tsbmissp++, i++) {
13096 		/*
13097 		 * initialize the tsbmiss area.
13098 		 * Do this for all possible CPUs as some may be added
13099 		 * while the system is running. There is no cost to this.
13100 		 */
13101 		tsbmissp->ksfmmup = ksfmmup;
13102 #ifndef sun4v
13103 		tsbmissp->dcache_line_mask = (uint16_t)dcache_line_mask;
13104 #endif /* sun4v */
13105 		tsbmissp->khashstart =
13106 		    (struct hmehash_bucket *)va_to_pa((caddr_t)khme_hash);
13107 		tsbmissp->uhashstart =
13108 		    (struct hmehash_bucket *)va_to_pa((caddr_t)uhme_hash);
13109 		tsbmissp->khashsz = khmehash_num;
13110 		tsbmissp->uhashsz = uhmehash_num;
13111 	}
13112 
13113 	sfmmu_tsb_cb_id = hat_register_callback('T'<<16 | 'S' << 8 | 'B',
13114 	    sfmmu_tsb_pre_relocator, sfmmu_tsb_post_relocator, NULL, 0);
13115 
13116 	if (kpm_enable == 0)
13117 		return;
13118 
13119 	/* -- Begin KPM specific init -- */
13120 
13121 	if (kpm_smallpages) {
13122 		/*
13123 		 * If we're using base pagesize pages for seg_kpm
13124 		 * mappings, we use the kernel TSB since we can't afford
13125 		 * to allocate a second huge TSB for these mappings.
13126 		 */
13127 		kpm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13128 		kpm_tsbsz = ktsb_szcode;
13129 		kpmsm_tsbbase = kpm_tsbbase;
13130 		kpmsm_tsbsz = kpm_tsbsz;
13131 	} else {
13132 		/*
13133 		 * In VAC conflict case, just put the entries in the
13134 		 * kernel 8K indexed TSB for now so we can find them.
13135 		 * This could really be changed in the future if we feel
13136 		 * the need...
13137 		 */
13138 		kpmsm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
13139 		kpmsm_tsbsz = ktsb_szcode;
13140 		kpm_tsbbase = ktsb_phys? ktsb4m_pbase : (uint64_t)ktsb4m_base;
13141 		kpm_tsbsz = ktsb4m_szcode;
13142 	}
13143 
13144 	kpmtsbmp = kpmtsbm_area;
13145 	for (i = 0; i < NCPU; kpmtsbmp++, i++) {
13146 		/*
13147 		 * Initialize the kpmtsbm area.
13148 		 * Do this for all possible CPUs as some may be added
13149 		 * while the system is running. There is no cost to this.
13150 		 */
13151 		kpmtsbmp->vbase = kpm_vbase;
13152 		kpmtsbmp->vend = kpm_vbase + kpm_size * vac_colors;
13153 		kpmtsbmp->sz_shift = kpm_size_shift;
13154 		kpmtsbmp->kpmp_shift = kpmp_shift;
13155 		kpmtsbmp->kpmp2pshft = (uchar_t)kpmp2pshft;
13156 		if (kpm_smallpages == 0) {
13157 			kpmtsbmp->kpmp_table_sz = kpmp_table_sz;
13158 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_table);
13159 		} else {
13160 			kpmtsbmp->kpmp_table_sz = kpmp_stable_sz;
13161 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_stable);
13162 		}
13163 		kpmtsbmp->msegphashpa = va_to_pa(memseg_phash);
13164 		kpmtsbmp->flags = KPMTSBM_ENABLE_FLAG;
13165 #ifdef	DEBUG
13166 		kpmtsbmp->flags |= (kpm_tsbmtl) ?  KPMTSBM_TLTSBM_FLAG : 0;
13167 #endif	/* DEBUG */
13168 		if (ktsb_phys)
13169 			kpmtsbmp->flags |= KPMTSBM_TSBPHYS_FLAG;
13170 	}
13171 
13172 	/* -- End KPM specific init -- */
13173 }
13174 
13175 /* Avoid using sfmmu_tsbinfo_alloc() to avoid kmem_alloc - no real reason */
13176 struct tsb_info ktsb_info[2];
13177 
13178 /*
13179  * Called from hat_kern_setup() to setup the tsb_info for ksfmmup.
13180  */
13181 void
13182 sfmmu_init_ktsbinfo()
13183 {
13184 	ASSERT(ksfmmup != NULL);
13185 	ASSERT(ksfmmup->sfmmu_tsb == NULL);
13186 	/*
13187 	 * Allocate tsbinfos for kernel and copy in data
13188 	 * to make debug easier and sun4v setup easier.
13189 	 */
13190 	ktsb_info[0].tsb_sfmmu = ksfmmup;
13191 	ktsb_info[0].tsb_szc = ktsb_szcode;
13192 	ktsb_info[0].tsb_ttesz_mask = TSB8K|TSB64K|TSB512K;
13193 	ktsb_info[0].tsb_va = ktsb_base;
13194 	ktsb_info[0].tsb_pa = ktsb_pbase;
13195 	ktsb_info[0].tsb_flags = 0;
13196 	ktsb_info[0].tsb_tte.ll = 0;
13197 	ktsb_info[0].tsb_cache = NULL;
13198 
13199 	ktsb_info[1].tsb_sfmmu = ksfmmup;
13200 	ktsb_info[1].tsb_szc = ktsb4m_szcode;
13201 	ktsb_info[1].tsb_ttesz_mask = TSB4M;
13202 	ktsb_info[1].tsb_va = ktsb4m_base;
13203 	ktsb_info[1].tsb_pa = ktsb4m_pbase;
13204 	ktsb_info[1].tsb_flags = 0;
13205 	ktsb_info[1].tsb_tte.ll = 0;
13206 	ktsb_info[1].tsb_cache = NULL;
13207 
13208 	/* Link them into ksfmmup. */
13209 	ktsb_info[0].tsb_next = &ktsb_info[1];
13210 	ktsb_info[1].tsb_next = NULL;
13211 	ksfmmup->sfmmu_tsb = &ktsb_info[0];
13212 
13213 	sfmmu_setup_tsbinfo(ksfmmup);
13214 }
13215 
13216 /*
13217  * Cache the last value returned from va_to_pa().  If the VA specified
13218  * in the current call to cached_va_to_pa() maps to the same Page (as the
13219  * previous call to cached_va_to_pa()), then compute the PA using
13220  * cached info, else call va_to_pa().
13221  *
13222  * Note: this function is neither MT-safe nor consistent in the presence
13223  * of multiple, interleaved threads.  This function was created to enable
13224  * an optimization used during boot (at a point when there's only one thread
13225  * executing on the "boot CPU", and before startup_vm() has been called).
13226  */
13227 static uint64_t
13228 cached_va_to_pa(void *vaddr)
13229 {
13230 	static uint64_t prev_vaddr_base = 0;
13231 	static uint64_t prev_pfn = 0;
13232 
13233 	if ((((uint64_t)vaddr) & MMU_PAGEMASK) == prev_vaddr_base) {
13234 		return (prev_pfn | ((uint64_t)vaddr & MMU_PAGEOFFSET));
13235 	} else {
13236 		uint64_t pa = va_to_pa(vaddr);
13237 
13238 		if (pa != ((uint64_t)-1)) {
13239 			/*
13240 			 * Computed physical address is valid.  Cache its
13241 			 * related info for the next cached_va_to_pa() call.
13242 			 */
13243 			prev_pfn = pa & MMU_PAGEMASK;
13244 			prev_vaddr_base = ((uint64_t)vaddr) & MMU_PAGEMASK;
13245 		}
13246 
13247 		return (pa);
13248 	}
13249 }
13250 
13251 /*
13252  * Carve up our nucleus hblk region.  We may allocate more hblks than
13253  * asked due to rounding errors but we are guaranteed to have at least
13254  * enough space to allocate the requested number of hblk8's and hblk1's.
13255  */
13256 void
13257 sfmmu_init_nucleus_hblks(caddr_t addr, size_t size, int nhblk8, int nhblk1)
13258 {
13259 	struct hme_blk *hmeblkp;
13260 	size_t hme8blk_sz, hme1blk_sz;
13261 	size_t i;
13262 	size_t hblk8_bound;
13263 	ulong_t j = 0, k = 0;
13264 
13265 	ASSERT(addr != NULL && size != 0);
13266 
13267 	/* Need to use proper structure alignment */
13268 	hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t));
13269 	hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t));
13270 
13271 	nucleus_hblk8.list = (void *)addr;
13272 	nucleus_hblk8.index = 0;
13273 
13274 	/*
13275 	 * Use as much memory as possible for hblk8's since we
13276 	 * expect all bop_alloc'ed memory to be allocated in 8k chunks.
13277 	 * We need to hold back enough space for the hblk1's which
13278 	 * we'll allocate next.
13279 	 */
13280 	hblk8_bound = size - (nhblk1 * hme1blk_sz) - hme8blk_sz;
13281 	for (i = 0; i <= hblk8_bound; i += hme8blk_sz, j++) {
13282 		hmeblkp = (struct hme_blk *)addr;
13283 		addr += hme8blk_sz;
13284 		hmeblkp->hblk_nuc_bit = 1;
13285 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13286 	}
13287 	nucleus_hblk8.len = j;
13288 	ASSERT(j >= nhblk8);
13289 	SFMMU_STAT_ADD(sf_hblk8_ncreate, j);
13290 
13291 	nucleus_hblk1.list = (void *)addr;
13292 	nucleus_hblk1.index = 0;
13293 	for (; i <= (size - hme1blk_sz); i += hme1blk_sz, k++) {
13294 		hmeblkp = (struct hme_blk *)addr;
13295 		addr += hme1blk_sz;
13296 		hmeblkp->hblk_nuc_bit = 1;
13297 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
13298 	}
13299 	ASSERT(k >= nhblk1);
13300 	nucleus_hblk1.len = k;
13301 	SFMMU_STAT_ADD(sf_hblk1_ncreate, k);
13302 }
13303 
13304 /*
13305  * This function is currently not supported on this platform. For what
13306  * it's supposed to do, see hat.c and hat_srmmu.c
13307  */
13308 /* ARGSUSED */
13309 faultcode_t
13310 hat_softlock(struct hat *hat, caddr_t addr, size_t *lenp, page_t **ppp,
13311     uint_t flags)
13312 {
13313 	ASSERT(hat->sfmmu_xhat_provider == NULL);
13314 	return (FC_NOSUPPORT);
13315 }
13316 
13317 /*
13318  * Searchs the mapping list of the page for a mapping of the same size. If not
13319  * found the corresponding bit is cleared in the p_index field. When large
13320  * pages are more prevalent in the system, we can maintain the mapping list
13321  * in order and we don't have to traverse the list each time. Just check the
13322  * next and prev entries, and if both are of different size, we clear the bit.
13323  */
13324 static void
13325 sfmmu_rm_large_mappings(page_t *pp, int ttesz)
13326 {
13327 	struct sf_hment *sfhmep;
13328 	struct hme_blk *hmeblkp;
13329 	int	index;
13330 	pgcnt_t	npgs;
13331 
13332 	ASSERT(ttesz > TTE8K);
13333 
13334 	ASSERT(sfmmu_mlist_held(pp));
13335 
13336 	ASSERT(PP_ISMAPPED_LARGE(pp));
13337 
13338 	/*
13339 	 * Traverse mapping list looking for another mapping of same size.
13340 	 * since we only want to clear index field if all mappings of
13341 	 * that size are gone.
13342 	 */
13343 
13344 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
13345 		if (IS_PAHME(sfhmep))
13346 			continue;
13347 		hmeblkp = sfmmu_hmetohblk(sfhmep);
13348 		if (hmeblkp->hblk_xhat_bit)
13349 			continue;
13350 		if (hme_size(sfhmep) == ttesz) {
13351 			/*
13352 			 * another mapping of the same size. don't clear index.
13353 			 */
13354 			return;
13355 		}
13356 	}
13357 
13358 	/*
13359 	 * Clear the p_index bit for large page.
13360 	 */
13361 	index = PAGESZ_TO_INDEX(ttesz);
13362 	npgs = TTEPAGES(ttesz);
13363 	while (npgs-- > 0) {
13364 		ASSERT(pp->p_index & index);
13365 		pp->p_index &= ~index;
13366 		pp = PP_PAGENEXT(pp);
13367 	}
13368 }
13369 
13370 /*
13371  * return supported features
13372  */
13373 /* ARGSUSED */
13374 int
13375 hat_supported(enum hat_features feature, void *arg)
13376 {
13377 	switch (feature) {
13378 	case    HAT_SHARED_PT:
13379 	case	HAT_DYNAMIC_ISM_UNMAP:
13380 	case	HAT_VMODSORT:
13381 		return (1);
13382 	case	HAT_SHARED_REGIONS:
13383 		if (shctx_on)
13384 			return (1);
13385 		else
13386 			return (0);
13387 	default:
13388 		return (0);
13389 	}
13390 }
13391 
13392 void
13393 hat_enter(struct hat *hat)
13394 {
13395 	hatlock_t	*hatlockp;
13396 
13397 	if (hat != ksfmmup) {
13398 		hatlockp = TSB_HASH(hat);
13399 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
13400 	}
13401 }
13402 
13403 void
13404 hat_exit(struct hat *hat)
13405 {
13406 	hatlock_t	*hatlockp;
13407 
13408 	if (hat != ksfmmup) {
13409 		hatlockp = TSB_HASH(hat);
13410 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
13411 	}
13412 }
13413 
13414 /*ARGSUSED*/
13415 void
13416 hat_reserve(struct as *as, caddr_t addr, size_t len)
13417 {
13418 }
13419 
13420 static void
13421 hat_kstat_init(void)
13422 {
13423 	kstat_t *ksp;
13424 
13425 	ksp = kstat_create("unix", 0, "sfmmu_global_stat", "hat",
13426 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_global_stat),
13427 	    KSTAT_FLAG_VIRTUAL);
13428 	if (ksp) {
13429 		ksp->ks_data = (void *) &sfmmu_global_stat;
13430 		kstat_install(ksp);
13431 	}
13432 	ksp = kstat_create("unix", 0, "sfmmu_tsbsize_stat", "hat",
13433 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_tsbsize_stat),
13434 	    KSTAT_FLAG_VIRTUAL);
13435 	if (ksp) {
13436 		ksp->ks_data = (void *) &sfmmu_tsbsize_stat;
13437 		kstat_install(ksp);
13438 	}
13439 	ksp = kstat_create("unix", 0, "sfmmu_percpu_stat", "hat",
13440 	    KSTAT_TYPE_RAW, sizeof (struct sfmmu_percpu_stat) * NCPU,
13441 	    KSTAT_FLAG_WRITABLE);
13442 	if (ksp) {
13443 		ksp->ks_update = sfmmu_kstat_percpu_update;
13444 		kstat_install(ksp);
13445 	}
13446 }
13447 
13448 /* ARGSUSED */
13449 static int
13450 sfmmu_kstat_percpu_update(kstat_t *ksp, int rw)
13451 {
13452 	struct sfmmu_percpu_stat *cpu_kstat = ksp->ks_data;
13453 	struct tsbmiss *tsbm = tsbmiss_area;
13454 	struct kpmtsbm *kpmtsbm = kpmtsbm_area;
13455 	int i;
13456 
13457 	ASSERT(cpu_kstat);
13458 	if (rw == KSTAT_READ) {
13459 		for (i = 0; i < NCPU; cpu_kstat++, tsbm++, kpmtsbm++, i++) {
13460 			cpu_kstat->sf_itlb_misses = 0;
13461 			cpu_kstat->sf_dtlb_misses = 0;
13462 			cpu_kstat->sf_utsb_misses = tsbm->utsb_misses -
13463 			    tsbm->uprot_traps;
13464 			cpu_kstat->sf_ktsb_misses = tsbm->ktsb_misses +
13465 			    kpmtsbm->kpm_tsb_misses - tsbm->kprot_traps;
13466 			cpu_kstat->sf_tsb_hits = 0;
13467 			cpu_kstat->sf_umod_faults = tsbm->uprot_traps;
13468 			cpu_kstat->sf_kmod_faults = tsbm->kprot_traps;
13469 		}
13470 	} else {
13471 		/* KSTAT_WRITE is used to clear stats */
13472 		for (i = 0; i < NCPU; tsbm++, kpmtsbm++, i++) {
13473 			tsbm->utsb_misses = 0;
13474 			tsbm->ktsb_misses = 0;
13475 			tsbm->uprot_traps = 0;
13476 			tsbm->kprot_traps = 0;
13477 			kpmtsbm->kpm_dtlb_misses = 0;
13478 			kpmtsbm->kpm_tsb_misses = 0;
13479 		}
13480 	}
13481 	return (0);
13482 }
13483 
13484 #ifdef	DEBUG
13485 
13486 tte_t  *gorig[NCPU], *gcur[NCPU], *gnew[NCPU];
13487 
13488 /*
13489  * A tte checker. *orig_old is the value we read before cas.
13490  *	*cur is the value returned by cas.
13491  *	*new is the desired value when we do the cas.
13492  *
13493  *	*hmeblkp is currently unused.
13494  */
13495 
13496 /* ARGSUSED */
13497 void
13498 chk_tte(tte_t *orig_old, tte_t *cur, tte_t *new, struct hme_blk *hmeblkp)
13499 {
13500 	pfn_t i, j, k;
13501 	int cpuid = CPU->cpu_id;
13502 
13503 	gorig[cpuid] = orig_old;
13504 	gcur[cpuid] = cur;
13505 	gnew[cpuid] = new;
13506 
13507 #ifdef lint
13508 	hmeblkp = hmeblkp;
13509 #endif
13510 
13511 	if (TTE_IS_VALID(orig_old)) {
13512 		if (TTE_IS_VALID(cur)) {
13513 			i = TTE_TO_TTEPFN(orig_old);
13514 			j = TTE_TO_TTEPFN(cur);
13515 			k = TTE_TO_TTEPFN(new);
13516 			if (i != j) {
13517 				/* remap error? */
13518 				panic("chk_tte: bad pfn, 0x%lx, 0x%lx", i, j);
13519 			}
13520 
13521 			if (i != k) {
13522 				/* remap error? */
13523 				panic("chk_tte: bad pfn2, 0x%lx, 0x%lx", i, k);
13524 			}
13525 		} else {
13526 			if (TTE_IS_VALID(new)) {
13527 				panic("chk_tte: invalid cur? ");
13528 			}
13529 
13530 			i = TTE_TO_TTEPFN(orig_old);
13531 			k = TTE_TO_TTEPFN(new);
13532 			if (i != k) {
13533 				panic("chk_tte: bad pfn3, 0x%lx, 0x%lx", i, k);
13534 			}
13535 		}
13536 	} else {
13537 		if (TTE_IS_VALID(cur)) {
13538 			j = TTE_TO_TTEPFN(cur);
13539 			if (TTE_IS_VALID(new)) {
13540 				k = TTE_TO_TTEPFN(new);
13541 				if (j != k) {
13542 					panic("chk_tte: bad pfn4, 0x%lx, 0x%lx",
13543 					    j, k);
13544 				}
13545 			} else {
13546 				panic("chk_tte: why here?");
13547 			}
13548 		} else {
13549 			if (!TTE_IS_VALID(new)) {
13550 				panic("chk_tte: why here2 ?");
13551 			}
13552 		}
13553 	}
13554 }
13555 
13556 #endif /* DEBUG */
13557 
13558 extern void prefetch_tsbe_read(struct tsbe *);
13559 extern void prefetch_tsbe_write(struct tsbe *);
13560 
13561 
13562 /*
13563  * We want to prefetch 7 cache lines ahead for our read prefetch.  This gives
13564  * us optimal performance on Cheetah+.  You can only have 8 outstanding
13565  * prefetches at any one time, so we opted for 7 read prefetches and 1 write
13566  * prefetch to make the most utilization of the prefetch capability.
13567  */
13568 #define	TSBE_PREFETCH_STRIDE (7)
13569 
13570 void
13571 sfmmu_copy_tsb(struct tsb_info *old_tsbinfo, struct tsb_info *new_tsbinfo)
13572 {
13573 	int old_bytes = TSB_BYTES(old_tsbinfo->tsb_szc);
13574 	int new_bytes = TSB_BYTES(new_tsbinfo->tsb_szc);
13575 	int old_entries = TSB_ENTRIES(old_tsbinfo->tsb_szc);
13576 	int new_entries = TSB_ENTRIES(new_tsbinfo->tsb_szc);
13577 	struct tsbe *old;
13578 	struct tsbe *new;
13579 	struct tsbe *new_base = (struct tsbe *)new_tsbinfo->tsb_va;
13580 	uint64_t va;
13581 	int new_offset;
13582 	int i;
13583 	int vpshift;
13584 	int last_prefetch;
13585 
13586 	if (old_bytes == new_bytes) {
13587 		bcopy(old_tsbinfo->tsb_va, new_tsbinfo->tsb_va, new_bytes);
13588 	} else {
13589 
13590 		/*
13591 		 * A TSBE is 16 bytes which means there are four TSBE's per
13592 		 * P$ line (64 bytes), thus every 4 TSBE's we prefetch.
13593 		 */
13594 		old = (struct tsbe *)old_tsbinfo->tsb_va;
13595 		last_prefetch = old_entries - (4*(TSBE_PREFETCH_STRIDE+1));
13596 		for (i = 0; i < old_entries; i++, old++) {
13597 			if (((i & (4-1)) == 0) && (i < last_prefetch))
13598 				prefetch_tsbe_read(old);
13599 			if (!old->tte_tag.tag_invalid) {
13600 				/*
13601 				 * We have a valid TTE to remap.  Check the
13602 				 * size.  We won't remap 64K or 512K TTEs
13603 				 * because they span more than one TSB entry
13604 				 * and are indexed using an 8K virt. page.
13605 				 * Ditto for 32M and 256M TTEs.
13606 				 */
13607 				if (TTE_CSZ(&old->tte_data) == TTE64K ||
13608 				    TTE_CSZ(&old->tte_data) == TTE512K)
13609 					continue;
13610 				if (mmu_page_sizes == max_mmu_page_sizes) {
13611 					if (TTE_CSZ(&old->tte_data) == TTE32M ||
13612 					    TTE_CSZ(&old->tte_data) == TTE256M)
13613 						continue;
13614 				}
13615 
13616 				/* clear the lower 22 bits of the va */
13617 				va = *(uint64_t *)old << 22;
13618 				/* turn va into a virtual pfn */
13619 				va >>= 22 - TSB_START_SIZE;
13620 				/*
13621 				 * or in bits from the offset in the tsb
13622 				 * to get the real virtual pfn. These
13623 				 * correspond to bits [21:13] in the va
13624 				 */
13625 				vpshift =
13626 				    TTE_BSZS_SHIFT(TTE_CSZ(&old->tte_data)) &
13627 				    0x1ff;
13628 				va |= (i << vpshift);
13629 				va >>= vpshift;
13630 				new_offset = va & (new_entries - 1);
13631 				new = new_base + new_offset;
13632 				prefetch_tsbe_write(new);
13633 				*new = *old;
13634 			}
13635 		}
13636 	}
13637 }
13638 
13639 /*
13640  * unused in sfmmu
13641  */
13642 void
13643 hat_dump(void)
13644 {
13645 }
13646 
13647 /*
13648  * Called when a thread is exiting and we have switched to the kernel address
13649  * space.  Perform the same VM initialization resume() uses when switching
13650  * processes.
13651  *
13652  * Note that sfmmu_load_mmustate() is currently a no-op for kernel threads, but
13653  * we call it anyway in case the semantics change in the future.
13654  */
13655 /*ARGSUSED*/
13656 void
13657 hat_thread_exit(kthread_t *thd)
13658 {
13659 	uint_t pgsz_cnum;
13660 	uint_t pstate_save;
13661 
13662 	ASSERT(thd->t_procp->p_as == &kas);
13663 
13664 	pgsz_cnum = KCONTEXT;
13665 #ifdef sun4u
13666 	pgsz_cnum |= (ksfmmup->sfmmu_cext << CTXREG_EXT_SHIFT);
13667 #endif
13668 
13669 	/*
13670 	 * Note that sfmmu_load_mmustate() is currently a no-op for
13671 	 * kernel threads. We need to disable interrupts here,
13672 	 * simply because otherwise sfmmu_load_mmustate() would panic
13673 	 * if the caller does not disable interrupts.
13674 	 */
13675 	pstate_save = sfmmu_disable_intrs();
13676 
13677 	/* Compatibility Note: hw takes care of MMU_SCONTEXT1 */
13678 	sfmmu_setctx_sec(pgsz_cnum);
13679 	sfmmu_load_mmustate(ksfmmup);
13680 	sfmmu_enable_intrs(pstate_save);
13681 }
13682 
13683 
13684 /*
13685  * SRD support
13686  */
13687 #define	SRD_HASH_FUNCTION(vp)	(((((uintptr_t)(vp)) >> 4) ^ \
13688 				    (((uintptr_t)(vp)) >> 11)) & \
13689 				    srd_hashmask)
13690 
13691 /*
13692  * Attach the process to the srd struct associated with the exec vnode
13693  * from which the process is started.
13694  */
13695 void
13696 hat_join_srd(struct hat *sfmmup, vnode_t *evp)
13697 {
13698 	uint_t hash = SRD_HASH_FUNCTION(evp);
13699 	sf_srd_t *srdp;
13700 	sf_srd_t *newsrdp;
13701 
13702 	ASSERT(sfmmup != ksfmmup);
13703 	ASSERT(sfmmup->sfmmu_srdp == NULL);
13704 
13705 	if (!shctx_on) {
13706 		return;
13707 	}
13708 
13709 	VN_HOLD(evp);
13710 
13711 	if (srd_buckets[hash].srdb_srdp != NULL) {
13712 		mutex_enter(&srd_buckets[hash].srdb_lock);
13713 		for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13714 		    srdp = srdp->srd_hash) {
13715 			if (srdp->srd_evp == evp) {
13716 				ASSERT(srdp->srd_refcnt >= 0);
13717 				sfmmup->sfmmu_srdp = srdp;
13718 				atomic_add_32(
13719 				    (volatile uint_t *)&srdp->srd_refcnt, 1);
13720 				mutex_exit(&srd_buckets[hash].srdb_lock);
13721 				return;
13722 			}
13723 		}
13724 		mutex_exit(&srd_buckets[hash].srdb_lock);
13725 	}
13726 	newsrdp = kmem_cache_alloc(srd_cache, KM_SLEEP);
13727 	ASSERT(newsrdp->srd_next_ismrid == 0 && newsrdp->srd_next_hmerid == 0);
13728 
13729 	newsrdp->srd_evp = evp;
13730 	newsrdp->srd_refcnt = 1;
13731 	newsrdp->srd_hmergnfree = NULL;
13732 	newsrdp->srd_ismrgnfree = NULL;
13733 
13734 	mutex_enter(&srd_buckets[hash].srdb_lock);
13735 	for (srdp = srd_buckets[hash].srdb_srdp; srdp != NULL;
13736 	    srdp = srdp->srd_hash) {
13737 		if (srdp->srd_evp == evp) {
13738 			ASSERT(srdp->srd_refcnt >= 0);
13739 			sfmmup->sfmmu_srdp = srdp;
13740 			atomic_add_32((volatile uint_t *)&srdp->srd_refcnt, 1);
13741 			mutex_exit(&srd_buckets[hash].srdb_lock);
13742 			kmem_cache_free(srd_cache, newsrdp);
13743 			return;
13744 		}
13745 	}
13746 	newsrdp->srd_hash = srd_buckets[hash].srdb_srdp;
13747 	srd_buckets[hash].srdb_srdp = newsrdp;
13748 	sfmmup->sfmmu_srdp = newsrdp;
13749 
13750 	mutex_exit(&srd_buckets[hash].srdb_lock);
13751 
13752 }
13753 
13754 static void
13755 sfmmu_leave_srd(sfmmu_t *sfmmup)
13756 {
13757 	vnode_t *evp;
13758 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
13759 	uint_t hash;
13760 	sf_srd_t **prev_srdpp;
13761 	sf_region_t *rgnp;
13762 	sf_region_t *nrgnp;
13763 #ifdef DEBUG
13764 	int rgns = 0;
13765 #endif
13766 	int i;
13767 
13768 	ASSERT(sfmmup != ksfmmup);
13769 	ASSERT(srdp != NULL);
13770 	ASSERT(srdp->srd_refcnt > 0);
13771 	ASSERT(sfmmup->sfmmu_scdp == NULL);
13772 	ASSERT(sfmmup->sfmmu_free == 1);
13773 
13774 	sfmmup->sfmmu_srdp = NULL;
13775 	evp = srdp->srd_evp;
13776 	ASSERT(evp != NULL);
13777 	if (atomic_add_32_nv(
13778 	    (volatile uint_t *)&srdp->srd_refcnt, -1)) {
13779 		VN_RELE(evp);
13780 		return;
13781 	}
13782 
13783 	hash = SRD_HASH_FUNCTION(evp);
13784 	mutex_enter(&srd_buckets[hash].srdb_lock);
13785 	for (prev_srdpp = &srd_buckets[hash].srdb_srdp;
13786 	    (srdp = *prev_srdpp) != NULL; prev_srdpp = &srdp->srd_hash) {
13787 		if (srdp->srd_evp == evp) {
13788 			break;
13789 		}
13790 	}
13791 	if (srdp == NULL || srdp->srd_refcnt) {
13792 		mutex_exit(&srd_buckets[hash].srdb_lock);
13793 		VN_RELE(evp);
13794 		return;
13795 	}
13796 	*prev_srdpp = srdp->srd_hash;
13797 	mutex_exit(&srd_buckets[hash].srdb_lock);
13798 
13799 	ASSERT(srdp->srd_refcnt == 0);
13800 	VN_RELE(evp);
13801 
13802 #ifdef DEBUG
13803 	for (i = 0; i < SFMMU_MAX_REGION_BUCKETS; i++) {
13804 		ASSERT(srdp->srd_rgnhash[i] == NULL);
13805 	}
13806 #endif /* DEBUG */
13807 
13808 	/* free each hme regions in the srd */
13809 	for (rgnp = srdp->srd_hmergnfree; rgnp != NULL; rgnp = nrgnp) {
13810 		nrgnp = rgnp->rgn_next;
13811 		ASSERT(rgnp->rgn_id < srdp->srd_next_hmerid);
13812 		ASSERT(rgnp->rgn_refcnt == 0);
13813 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13814 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13815 		ASSERT(rgnp->rgn_hmeflags == 0);
13816 		ASSERT(srdp->srd_hmergnp[rgnp->rgn_id] == rgnp);
13817 #ifdef DEBUG
13818 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13819 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13820 		}
13821 		rgns++;
13822 #endif /* DEBUG */
13823 		kmem_cache_free(region_cache, rgnp);
13824 	}
13825 	ASSERT(rgns == srdp->srd_next_hmerid);
13826 
13827 #ifdef DEBUG
13828 	rgns = 0;
13829 #endif
13830 	/* free each ism rgns in the srd */
13831 	for (rgnp = srdp->srd_ismrgnfree; rgnp != NULL; rgnp = nrgnp) {
13832 		nrgnp = rgnp->rgn_next;
13833 		ASSERT(rgnp->rgn_id < srdp->srd_next_ismrid);
13834 		ASSERT(rgnp->rgn_refcnt == 0);
13835 		ASSERT(rgnp->rgn_sfmmu_head == NULL);
13836 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
13837 		ASSERT(srdp->srd_ismrgnp[rgnp->rgn_id] == rgnp);
13838 #ifdef DEBUG
13839 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
13840 			ASSERT(rgnp->rgn_ttecnt[i] == 0);
13841 		}
13842 		rgns++;
13843 #endif /* DEBUG */
13844 		kmem_cache_free(region_cache, rgnp);
13845 	}
13846 	ASSERT(rgns == srdp->srd_next_ismrid);
13847 	ASSERT(srdp->srd_ismbusyrgns == 0);
13848 	ASSERT(srdp->srd_hmebusyrgns == 0);
13849 
13850 	srdp->srd_next_ismrid = 0;
13851 	srdp->srd_next_hmerid = 0;
13852 
13853 	bzero((void *)srdp->srd_ismrgnp,
13854 	    sizeof (sf_region_t *) * SFMMU_MAX_ISM_REGIONS);
13855 	bzero((void *)srdp->srd_hmergnp,
13856 	    sizeof (sf_region_t *) * SFMMU_MAX_HME_REGIONS);
13857 
13858 	ASSERT(srdp->srd_scdp == NULL);
13859 	kmem_cache_free(srd_cache, srdp);
13860 }
13861 
13862 /* ARGSUSED */
13863 static int
13864 sfmmu_srdcache_constructor(void *buf, void *cdrarg, int kmflags)
13865 {
13866 	sf_srd_t *srdp = (sf_srd_t *)buf;
13867 	bzero(buf, sizeof (*srdp));
13868 
13869 	mutex_init(&srdp->srd_mutex, NULL, MUTEX_DEFAULT, NULL);
13870 	mutex_init(&srdp->srd_scd_mutex, NULL, MUTEX_DEFAULT, NULL);
13871 	return (0);
13872 }
13873 
13874 /* ARGSUSED */
13875 static void
13876 sfmmu_srdcache_destructor(void *buf, void *cdrarg)
13877 {
13878 	sf_srd_t *srdp = (sf_srd_t *)buf;
13879 
13880 	mutex_destroy(&srdp->srd_mutex);
13881 	mutex_destroy(&srdp->srd_scd_mutex);
13882 }
13883 
13884 /*
13885  * The caller makes sure hat_join_region()/hat_leave_region() can't be called
13886  * at the same time for the same process and address range. This is ensured by
13887  * the fact that address space is locked as writer when a process joins the
13888  * regions. Therefore there's no need to hold an srd lock during the entire
13889  * execution of hat_join_region()/hat_leave_region().
13890  */
13891 
13892 #define	RGN_HASH_FUNCTION(obj)	(((((uintptr_t)(obj)) >> 4) ^ \
13893 				    (((uintptr_t)(obj)) >> 11)) & \
13894 					srd_rgn_hashmask)
13895 /*
13896  * This routine implements the shared context functionality required when
13897  * attaching a segment to an address space. It must be called from
13898  * hat_share() for D(ISM) segments and from segvn_create() for segments
13899  * with the MAP_PRIVATE and MAP_TEXT flags set. It returns a region_cookie
13900  * which is saved in the private segment data for hme segments and
13901  * the ism_map structure for ism segments.
13902  */
13903 hat_region_cookie_t
13904 hat_join_region(struct hat *sfmmup,
13905 	caddr_t r_saddr,
13906 	size_t r_size,
13907 	void *r_obj,
13908 	u_offset_t r_objoff,
13909 	uchar_t r_perm,
13910 	uchar_t r_pgszc,
13911 	hat_rgn_cb_func_t r_cb_function,
13912 	uint_t flags)
13913 {
13914 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
13915 	uint_t rhash;
13916 	uint_t rid;
13917 	hatlock_t *hatlockp;
13918 	sf_region_t *rgnp;
13919 	sf_region_t *new_rgnp = NULL;
13920 	int i;
13921 	uint16_t *nextidp;
13922 	sf_region_t **freelistp;
13923 	int maxids;
13924 	sf_region_t **rarrp;
13925 	uint16_t *busyrgnsp;
13926 	ulong_t rttecnt;
13927 	uchar_t tteflag;
13928 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
13929 	int text = (r_type == HAT_REGION_TEXT);
13930 
13931 	if (srdp == NULL || r_size == 0) {
13932 		return (HAT_INVALID_REGION_COOKIE);
13933 	}
13934 
13935 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
13936 	ASSERT(sfmmup != ksfmmup);
13937 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
13938 	ASSERT(srdp->srd_refcnt > 0);
13939 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
13940 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
13941 	ASSERT(r_pgszc < mmu_page_sizes);
13942 	if (!IS_P2ALIGNED(r_saddr, TTEBYTES(r_pgszc)) ||
13943 	    !IS_P2ALIGNED(r_size, TTEBYTES(r_pgszc))) {
13944 		panic("hat_join_region: region addr or size is not aligned\n");
13945 	}
13946 
13947 
13948 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
13949 	    SFMMU_REGION_HME;
13950 	/*
13951 	 * Currently only support shared hmes for the read only main text
13952 	 * region.
13953 	 */
13954 	if (r_type == SFMMU_REGION_HME && ((r_obj != srdp->srd_evp) ||
13955 	    (r_perm & PROT_WRITE))) {
13956 		return (HAT_INVALID_REGION_COOKIE);
13957 	}
13958 
13959 	rhash = RGN_HASH_FUNCTION(r_obj);
13960 
13961 	if (r_type == SFMMU_REGION_ISM) {
13962 		nextidp = &srdp->srd_next_ismrid;
13963 		freelistp = &srdp->srd_ismrgnfree;
13964 		maxids = SFMMU_MAX_ISM_REGIONS;
13965 		rarrp = srdp->srd_ismrgnp;
13966 		busyrgnsp = &srdp->srd_ismbusyrgns;
13967 	} else {
13968 		nextidp = &srdp->srd_next_hmerid;
13969 		freelistp = &srdp->srd_hmergnfree;
13970 		maxids = SFMMU_MAX_HME_REGIONS;
13971 		rarrp = srdp->srd_hmergnp;
13972 		busyrgnsp = &srdp->srd_hmebusyrgns;
13973 	}
13974 
13975 	mutex_enter(&srdp->srd_mutex);
13976 
13977 	for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
13978 	    rgnp = rgnp->rgn_hash) {
13979 		if (rgnp->rgn_saddr == r_saddr && rgnp->rgn_size == r_size &&
13980 		    rgnp->rgn_obj == r_obj && rgnp->rgn_objoff == r_objoff &&
13981 		    rgnp->rgn_perm == r_perm && rgnp->rgn_pgszc == r_pgszc) {
13982 			break;
13983 		}
13984 	}
13985 
13986 rfound:
13987 	if (rgnp != NULL) {
13988 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
13989 		ASSERT(rgnp->rgn_cb_function == r_cb_function);
13990 		ASSERT(rgnp->rgn_refcnt >= 0);
13991 		rid = rgnp->rgn_id;
13992 		ASSERT(rid < maxids);
13993 		ASSERT(rarrp[rid] == rgnp);
13994 		ASSERT(rid < *nextidp);
13995 		atomic_add_32((volatile uint_t *)&rgnp->rgn_refcnt, 1);
13996 		mutex_exit(&srdp->srd_mutex);
13997 		if (new_rgnp != NULL) {
13998 			kmem_cache_free(region_cache, new_rgnp);
13999 		}
14000 		if (r_type == SFMMU_REGION_HME) {
14001 			int myjoin =
14002 			    (sfmmup == astosfmmu(curthread->t_procp->p_as));
14003 
14004 			sfmmu_link_to_hmeregion(sfmmup, rgnp);
14005 			/*
14006 			 * bitmap should be updated after linking sfmmu on
14007 			 * region list so that pageunload() doesn't skip
14008 			 * TSB/TLB flush. As soon as bitmap is updated another
14009 			 * thread in this process can already start accessing
14010 			 * this region.
14011 			 */
14012 			/*
14013 			 * Normally ttecnt accounting is done as part of
14014 			 * pagefault handling. But a process may not take any
14015 			 * pagefaults on shared hmeblks created by some other
14016 			 * process. To compensate for this assume that the
14017 			 * entire region will end up faulted in using
14018 			 * the region's pagesize.
14019 			 *
14020 			 */
14021 			if (r_pgszc > TTE8K) {
14022 				tteflag = 1 << r_pgszc;
14023 				if (disable_large_pages & tteflag) {
14024 					tteflag = 0;
14025 				}
14026 			} else {
14027 				tteflag = 0;
14028 			}
14029 			if (tteflag && !(sfmmup->sfmmu_rtteflags & tteflag)) {
14030 				hatlockp = sfmmu_hat_enter(sfmmup);
14031 				sfmmup->sfmmu_rtteflags |= tteflag;
14032 				sfmmu_hat_exit(hatlockp);
14033 			}
14034 			hatlockp = sfmmu_hat_enter(sfmmup);
14035 
14036 			/*
14037 			 * Preallocate 1/4 of ttecnt's in 8K TSB for >= 4M
14038 			 * region to allow for large page allocation failure.
14039 			 */
14040 			if (r_pgszc >= TTE4M) {
14041 				sfmmup->sfmmu_tsb0_4minflcnt +=
14042 				    r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14043 			}
14044 
14045 			/* update sfmmu_ttecnt with the shme rgn ttecnt */
14046 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14047 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14048 			    rttecnt);
14049 
14050 			if (text && r_pgszc >= TTE4M &&
14051 			    (tteflag || ((disable_large_pages >> TTE4M) &
14052 			    ((1 << (r_pgszc - TTE4M + 1)) - 1))) &&
14053 			    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
14054 				SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
14055 			}
14056 
14057 			sfmmu_hat_exit(hatlockp);
14058 			/*
14059 			 * On Panther we need to make sure TLB is programmed
14060 			 * to accept 32M/256M pages.  Call
14061 			 * sfmmu_check_page_sizes() now to make sure TLB is
14062 			 * setup before making hmeregions visible to other
14063 			 * threads.
14064 			 */
14065 			sfmmu_check_page_sizes(sfmmup, 1);
14066 			hatlockp = sfmmu_hat_enter(sfmmup);
14067 			SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14068 
14069 			/*
14070 			 * if context is invalid tsb miss exception code will
14071 			 * call sfmmu_check_page_sizes() and update tsbmiss
14072 			 * area later.
14073 			 */
14074 			kpreempt_disable();
14075 			if (myjoin &&
14076 			    (sfmmup->sfmmu_ctxs[CPU_MMU_IDX(CPU)].cnum
14077 			    != INVALID_CONTEXT)) {
14078 				struct tsbmiss *tsbmp;
14079 
14080 				tsbmp = &tsbmiss_area[CPU->cpu_id];
14081 				ASSERT(sfmmup == tsbmp->usfmmup);
14082 				BT_SET(tsbmp->shmermap, rid);
14083 				if (r_pgszc > TTE64K) {
14084 					tsbmp->uhat_rtteflags |= tteflag;
14085 				}
14086 
14087 			}
14088 			kpreempt_enable();
14089 
14090 			sfmmu_hat_exit(hatlockp);
14091 			ASSERT((hat_region_cookie_t)((uint64_t)rid) !=
14092 			    HAT_INVALID_REGION_COOKIE);
14093 		} else {
14094 			hatlockp = sfmmu_hat_enter(sfmmup);
14095 			SF_RGNMAP_ADD(sfmmup->sfmmu_ismregion_map, rid);
14096 			sfmmu_hat_exit(hatlockp);
14097 		}
14098 		ASSERT(rid < maxids);
14099 
14100 		if (r_type == SFMMU_REGION_ISM) {
14101 			sfmmu_find_scd(sfmmup);
14102 		}
14103 		return ((hat_region_cookie_t)((uint64_t)rid));
14104 	}
14105 
14106 	ASSERT(new_rgnp == NULL);
14107 
14108 	if (*busyrgnsp >= maxids) {
14109 		mutex_exit(&srdp->srd_mutex);
14110 		return (HAT_INVALID_REGION_COOKIE);
14111 	}
14112 
14113 	ASSERT(MUTEX_HELD(&srdp->srd_mutex));
14114 	if (*freelistp != NULL) {
14115 		rgnp = *freelistp;
14116 		*freelistp = rgnp->rgn_next;
14117 		ASSERT(rgnp->rgn_id < *nextidp);
14118 		ASSERT(rgnp->rgn_id < maxids);
14119 		ASSERT(rgnp->rgn_flags & SFMMU_REGION_FREE);
14120 		ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK)
14121 		    == r_type);
14122 		ASSERT(rarrp[rgnp->rgn_id] == rgnp);
14123 		ASSERT(rgnp->rgn_hmeflags == 0);
14124 	} else {
14125 		/*
14126 		 * release local locks before memory allocation.
14127 		 */
14128 		mutex_exit(&srdp->srd_mutex);
14129 
14130 		new_rgnp = kmem_cache_alloc(region_cache, KM_SLEEP);
14131 
14132 		mutex_enter(&srdp->srd_mutex);
14133 		for (rgnp = srdp->srd_rgnhash[rhash]; rgnp != NULL;
14134 		    rgnp = rgnp->rgn_hash) {
14135 			if (rgnp->rgn_saddr == r_saddr &&
14136 			    rgnp->rgn_size == r_size &&
14137 			    rgnp->rgn_obj == r_obj &&
14138 			    rgnp->rgn_objoff == r_objoff &&
14139 			    rgnp->rgn_perm == r_perm &&
14140 			    rgnp->rgn_pgszc == r_pgszc) {
14141 				break;
14142 			}
14143 		}
14144 		if (rgnp != NULL) {
14145 			goto rfound;
14146 		}
14147 
14148 		if (*nextidp >= maxids) {
14149 			mutex_exit(&srdp->srd_mutex);
14150 			goto fail;
14151 		}
14152 		rgnp = new_rgnp;
14153 		new_rgnp = NULL;
14154 		rgnp->rgn_id = (*nextidp)++;
14155 		ASSERT(rgnp->rgn_id < maxids);
14156 		ASSERT(rarrp[rgnp->rgn_id] == NULL);
14157 		rarrp[rgnp->rgn_id] = rgnp;
14158 	}
14159 
14160 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14161 	ASSERT(rgnp->rgn_hmeflags == 0);
14162 #ifdef DEBUG
14163 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14164 		ASSERT(rgnp->rgn_ttecnt[i] == 0);
14165 	}
14166 #endif
14167 	rgnp->rgn_saddr = r_saddr;
14168 	rgnp->rgn_size = r_size;
14169 	rgnp->rgn_obj = r_obj;
14170 	rgnp->rgn_objoff = r_objoff;
14171 	rgnp->rgn_perm = r_perm;
14172 	rgnp->rgn_pgszc = r_pgszc;
14173 	rgnp->rgn_flags = r_type;
14174 	rgnp->rgn_refcnt = 0;
14175 	rgnp->rgn_cb_function = r_cb_function;
14176 	rgnp->rgn_hash = srdp->srd_rgnhash[rhash];
14177 	srdp->srd_rgnhash[rhash] = rgnp;
14178 	(*busyrgnsp)++;
14179 	ASSERT(*busyrgnsp <= maxids);
14180 	goto rfound;
14181 
14182 fail:
14183 	ASSERT(new_rgnp != NULL);
14184 	kmem_cache_free(region_cache, new_rgnp);
14185 	return (HAT_INVALID_REGION_COOKIE);
14186 }
14187 
14188 /*
14189  * This function implements the shared context functionality required
14190  * when detaching a segment from an address space. It must be called
14191  * from hat_unshare() for all D(ISM) segments and from segvn_unmap(),
14192  * for segments with a valid region_cookie.
14193  * It will also be called from all seg_vn routines which change a
14194  * segment's attributes such as segvn_setprot(), segvn_setpagesize(),
14195  * segvn_clrszc() & segvn_advise(), as well as in the case of COW fault
14196  * from segvn_fault().
14197  */
14198 void
14199 hat_leave_region(struct hat *sfmmup, hat_region_cookie_t rcookie, uint_t flags)
14200 {
14201 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14202 	sf_scd_t *scdp;
14203 	uint_t rhash;
14204 	uint_t rid = (uint_t)((uint64_t)rcookie);
14205 	hatlock_t *hatlockp = NULL;
14206 	sf_region_t *rgnp;
14207 	sf_region_t **prev_rgnpp;
14208 	sf_region_t *cur_rgnp;
14209 	void *r_obj;
14210 	int i;
14211 	caddr_t	r_saddr;
14212 	caddr_t r_eaddr;
14213 	size_t	r_size;
14214 	uchar_t	r_pgszc;
14215 	uchar_t r_type = flags & HAT_REGION_TYPE_MASK;
14216 
14217 	ASSERT(sfmmup != ksfmmup);
14218 	ASSERT(srdp != NULL);
14219 	ASSERT(srdp->srd_refcnt > 0);
14220 	ASSERT(!(flags & ~HAT_REGION_TYPE_MASK));
14221 	ASSERT(flags == HAT_REGION_TEXT || flags == HAT_REGION_ISM);
14222 	ASSERT(!sfmmup->sfmmu_free || sfmmup->sfmmu_scdp == NULL);
14223 
14224 	r_type = (r_type == HAT_REGION_ISM) ? SFMMU_REGION_ISM :
14225 	    SFMMU_REGION_HME;
14226 
14227 	if (r_type == SFMMU_REGION_ISM) {
14228 		ASSERT(SFMMU_IS_ISMRID_VALID(rid));
14229 		ASSERT(rid < SFMMU_MAX_ISM_REGIONS);
14230 		rgnp = srdp->srd_ismrgnp[rid];
14231 	} else {
14232 		ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14233 		ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14234 		rgnp = srdp->srd_hmergnp[rid];
14235 	}
14236 	ASSERT(rgnp != NULL);
14237 	ASSERT(rgnp->rgn_id == rid);
14238 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14239 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14240 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
14241 
14242 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
14243 	if (r_type == SFMMU_REGION_HME && sfmmup->sfmmu_as->a_xhat != NULL) {
14244 		xhat_unload_callback_all(sfmmup->sfmmu_as, rgnp->rgn_saddr,
14245 		    rgnp->rgn_size, 0, NULL);
14246 	}
14247 
14248 	if (sfmmup->sfmmu_free) {
14249 		ulong_t rttecnt;
14250 		r_pgszc = rgnp->rgn_pgszc;
14251 		r_size = rgnp->rgn_size;
14252 
14253 		ASSERT(sfmmup->sfmmu_scdp == NULL);
14254 		if (r_type == SFMMU_REGION_ISM) {
14255 			SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14256 		} else {
14257 			/* update shme rgns ttecnt in sfmmu_ttecnt */
14258 			rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14259 			ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14260 
14261 			atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc],
14262 			    -rttecnt);
14263 
14264 			SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14265 		}
14266 	} else if (r_type == SFMMU_REGION_ISM) {
14267 		hatlockp = sfmmu_hat_enter(sfmmup);
14268 		ASSERT(rid < srdp->srd_next_ismrid);
14269 		SF_RGNMAP_DEL(sfmmup->sfmmu_ismregion_map, rid);
14270 		scdp = sfmmup->sfmmu_scdp;
14271 		if (scdp != NULL &&
14272 		    SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid)) {
14273 			sfmmu_leave_scd(sfmmup, r_type);
14274 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14275 		}
14276 		sfmmu_hat_exit(hatlockp);
14277 	} else {
14278 		ulong_t rttecnt;
14279 		r_pgszc = rgnp->rgn_pgszc;
14280 		r_saddr = rgnp->rgn_saddr;
14281 		r_size = rgnp->rgn_size;
14282 		r_eaddr = r_saddr + r_size;
14283 
14284 		ASSERT(r_type == SFMMU_REGION_HME);
14285 		hatlockp = sfmmu_hat_enter(sfmmup);
14286 		ASSERT(rid < srdp->srd_next_hmerid);
14287 		SF_RGNMAP_DEL(sfmmup->sfmmu_hmeregion_map, rid);
14288 
14289 		/*
14290 		 * If region is part of an SCD call sfmmu_leave_scd().
14291 		 * Otherwise if process is not exiting and has valid context
14292 		 * just drop the context on the floor to lose stale TLB
14293 		 * entries and force the update of tsb miss area to reflect
14294 		 * the new region map. After that clean our TSB entries.
14295 		 */
14296 		scdp = sfmmup->sfmmu_scdp;
14297 		if (scdp != NULL &&
14298 		    SF_RGNMAP_TEST(scdp->scd_hmeregion_map, rid)) {
14299 			sfmmu_leave_scd(sfmmup, r_type);
14300 			ASSERT(sfmmu_hat_lock_held(sfmmup));
14301 		}
14302 		sfmmu_invalidate_ctx(sfmmup);
14303 
14304 		i = TTE8K;
14305 		while (i < mmu_page_sizes) {
14306 			if (rgnp->rgn_ttecnt[i] != 0) {
14307 				sfmmu_unload_tsb_range(sfmmup, r_saddr,
14308 				    r_eaddr, i);
14309 				if (i < TTE4M) {
14310 					i = TTE4M;
14311 					continue;
14312 				} else {
14313 					break;
14314 				}
14315 			}
14316 			i++;
14317 		}
14318 		/* Remove the preallocated 1/4 8k ttecnt for 4M regions. */
14319 		if (r_pgszc >= TTE4M) {
14320 			rttecnt = r_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14321 			ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
14322 			    rttecnt);
14323 			sfmmup->sfmmu_tsb0_4minflcnt -= rttecnt;
14324 		}
14325 
14326 		/* update shme rgns ttecnt in sfmmu_ttecnt */
14327 		rttecnt = r_size >> TTE_PAGE_SHIFT(r_pgszc);
14328 		ASSERT(sfmmup->sfmmu_ttecnt[r_pgszc] >= rttecnt);
14329 		atomic_add_long(&sfmmup->sfmmu_ttecnt[r_pgszc], -rttecnt);
14330 
14331 		sfmmu_hat_exit(hatlockp);
14332 		if (scdp != NULL && sfmmup->sfmmu_scdp == NULL) {
14333 			/* sfmmup left the scd, grow private tsb */
14334 			sfmmu_check_page_sizes(sfmmup, 1);
14335 		} else {
14336 			sfmmu_check_page_sizes(sfmmup, 0);
14337 		}
14338 	}
14339 
14340 	if (r_type == SFMMU_REGION_HME) {
14341 		sfmmu_unlink_from_hmeregion(sfmmup, rgnp);
14342 	}
14343 
14344 	r_obj = rgnp->rgn_obj;
14345 	if (atomic_add_32_nv((volatile uint_t *)&rgnp->rgn_refcnt, -1)) {
14346 		return;
14347 	}
14348 
14349 	/*
14350 	 * looks like nobody uses this region anymore. Free it.
14351 	 */
14352 	rhash = RGN_HASH_FUNCTION(r_obj);
14353 	mutex_enter(&srdp->srd_mutex);
14354 	for (prev_rgnpp = &srdp->srd_rgnhash[rhash];
14355 	    (cur_rgnp = *prev_rgnpp) != NULL;
14356 	    prev_rgnpp = &cur_rgnp->rgn_hash) {
14357 		if (cur_rgnp == rgnp && cur_rgnp->rgn_refcnt == 0) {
14358 			break;
14359 		}
14360 	}
14361 
14362 	if (cur_rgnp == NULL) {
14363 		mutex_exit(&srdp->srd_mutex);
14364 		return;
14365 	}
14366 
14367 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == r_type);
14368 	*prev_rgnpp = rgnp->rgn_hash;
14369 	if (r_type == SFMMU_REGION_ISM) {
14370 		rgnp->rgn_flags |= SFMMU_REGION_FREE;
14371 		ASSERT(rid < srdp->srd_next_ismrid);
14372 		rgnp->rgn_next = srdp->srd_ismrgnfree;
14373 		srdp->srd_ismrgnfree = rgnp;
14374 		ASSERT(srdp->srd_ismbusyrgns > 0);
14375 		srdp->srd_ismbusyrgns--;
14376 		mutex_exit(&srdp->srd_mutex);
14377 		return;
14378 	}
14379 	mutex_exit(&srdp->srd_mutex);
14380 
14381 	/*
14382 	 * Destroy region's hmeblks.
14383 	 */
14384 	sfmmu_unload_hmeregion(srdp, rgnp);
14385 
14386 	rgnp->rgn_hmeflags = 0;
14387 
14388 	ASSERT(rgnp->rgn_sfmmu_head == NULL);
14389 	ASSERT(rgnp->rgn_id == rid);
14390 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
14391 		rgnp->rgn_ttecnt[i] = 0;
14392 	}
14393 	rgnp->rgn_flags |= SFMMU_REGION_FREE;
14394 	mutex_enter(&srdp->srd_mutex);
14395 	ASSERT(rid < srdp->srd_next_hmerid);
14396 	rgnp->rgn_next = srdp->srd_hmergnfree;
14397 	srdp->srd_hmergnfree = rgnp;
14398 	ASSERT(srdp->srd_hmebusyrgns > 0);
14399 	srdp->srd_hmebusyrgns--;
14400 	mutex_exit(&srdp->srd_mutex);
14401 }
14402 
14403 /*
14404  * For now only called for hmeblk regions and not for ISM regions.
14405  */
14406 void
14407 hat_dup_region(struct hat *sfmmup, hat_region_cookie_t rcookie)
14408 {
14409 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14410 	uint_t rid = (uint_t)((uint64_t)rcookie);
14411 	sf_region_t *rgnp;
14412 	sf_rgn_link_t *rlink;
14413 	sf_rgn_link_t *hrlink;
14414 	ulong_t	rttecnt;
14415 
14416 	ASSERT(sfmmup != ksfmmup);
14417 	ASSERT(srdp != NULL);
14418 	ASSERT(srdp->srd_refcnt > 0);
14419 
14420 	ASSERT(rid < srdp->srd_next_hmerid);
14421 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14422 	ASSERT(rid < SFMMU_MAX_HME_REGIONS);
14423 
14424 	rgnp = srdp->srd_hmergnp[rid];
14425 	ASSERT(rgnp->rgn_refcnt > 0);
14426 	ASSERT(rgnp->rgn_id == rid);
14427 	ASSERT((rgnp->rgn_flags & SFMMU_REGION_TYPE_MASK) == SFMMU_REGION_HME);
14428 	ASSERT(!(rgnp->rgn_flags & SFMMU_REGION_FREE));
14429 
14430 	atomic_add_32((volatile uint_t *)&rgnp->rgn_refcnt, 1);
14431 
14432 	/* LINTED: constant in conditional context */
14433 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 0);
14434 	ASSERT(rlink != NULL);
14435 	mutex_enter(&rgnp->rgn_mutex);
14436 	ASSERT(rgnp->rgn_sfmmu_head != NULL);
14437 	/* LINTED: constant in conditional context */
14438 	SFMMU_HMERID2RLINKP(rgnp->rgn_sfmmu_head, rid, hrlink, 0, 0);
14439 	ASSERT(hrlink != NULL);
14440 	ASSERT(hrlink->prev == NULL);
14441 	rlink->next = rgnp->rgn_sfmmu_head;
14442 	rlink->prev = NULL;
14443 	hrlink->prev = sfmmup;
14444 	/*
14445 	 * make sure rlink's next field is correct
14446 	 * before making this link visible.
14447 	 */
14448 	membar_stst();
14449 	rgnp->rgn_sfmmu_head = sfmmup;
14450 	mutex_exit(&rgnp->rgn_mutex);
14451 
14452 	/* update sfmmu_ttecnt with the shme rgn ttecnt */
14453 	rttecnt = rgnp->rgn_size >> TTE_PAGE_SHIFT(rgnp->rgn_pgszc);
14454 	atomic_add_long(&sfmmup->sfmmu_ttecnt[rgnp->rgn_pgszc], rttecnt);
14455 	/* update tsb0 inflation count */
14456 	if (rgnp->rgn_pgszc >= TTE4M) {
14457 		sfmmup->sfmmu_tsb0_4minflcnt +=
14458 		    rgnp->rgn_size >> (TTE_PAGE_SHIFT(TTE8K) + 2);
14459 	}
14460 	/*
14461 	 * Update regionid bitmask without hat lock since no other thread
14462 	 * can update this region bitmask right now.
14463 	 */
14464 	SF_RGNMAP_ADD(sfmmup->sfmmu_hmeregion_map, rid);
14465 }
14466 
14467 /* ARGSUSED */
14468 static int
14469 sfmmu_rgncache_constructor(void *buf, void *cdrarg, int kmflags)
14470 {
14471 	sf_region_t *rgnp = (sf_region_t *)buf;
14472 	bzero(buf, sizeof (*rgnp));
14473 
14474 	mutex_init(&rgnp->rgn_mutex, NULL, MUTEX_DEFAULT, NULL);
14475 
14476 	return (0);
14477 }
14478 
14479 /* ARGSUSED */
14480 static void
14481 sfmmu_rgncache_destructor(void *buf, void *cdrarg)
14482 {
14483 	sf_region_t *rgnp = (sf_region_t *)buf;
14484 	mutex_destroy(&rgnp->rgn_mutex);
14485 }
14486 
14487 static int
14488 sfrgnmap_isnull(sf_region_map_t *map)
14489 {
14490 	int i;
14491 
14492 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14493 		if (map->bitmap[i] != 0) {
14494 			return (0);
14495 		}
14496 	}
14497 	return (1);
14498 }
14499 
14500 static int
14501 sfhmergnmap_isnull(sf_hmeregion_map_t *map)
14502 {
14503 	int i;
14504 
14505 	for (i = 0; i < SFMMU_HMERGNMAP_WORDS; i++) {
14506 		if (map->bitmap[i] != 0) {
14507 			return (0);
14508 		}
14509 	}
14510 	return (1);
14511 }
14512 
14513 #ifdef DEBUG
14514 static void
14515 check_scd_sfmmu_list(sfmmu_t **headp, sfmmu_t *sfmmup, int onlist)
14516 {
14517 	sfmmu_t *sp;
14518 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
14519 
14520 	for (sp = *headp; sp != NULL; sp = sp->sfmmu_scd_link.next) {
14521 		ASSERT(srdp == sp->sfmmu_srdp);
14522 		if (sp == sfmmup) {
14523 			if (onlist) {
14524 				return;
14525 			} else {
14526 				panic("shctx: sfmmu 0x%p found on scd"
14527 				    "list 0x%p", sfmmup, *headp);
14528 			}
14529 		}
14530 	}
14531 	if (onlist) {
14532 		panic("shctx: sfmmu 0x%p not found on scd list 0x%p",
14533 		    sfmmup, *headp);
14534 	} else {
14535 		return;
14536 	}
14537 }
14538 #else /* DEBUG */
14539 #define	check_scd_sfmmu_list(headp, sfmmup, onlist)
14540 #endif /* DEBUG */
14541 
14542 /*
14543  * Removes an sfmmu from the SCD sfmmu list.
14544  */
14545 static void
14546 sfmmu_from_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14547 {
14548 	ASSERT(sfmmup->sfmmu_srdp != NULL);
14549 	check_scd_sfmmu_list(headp, sfmmup, 1);
14550 	if (sfmmup->sfmmu_scd_link.prev != NULL) {
14551 		ASSERT(*headp != sfmmup);
14552 		sfmmup->sfmmu_scd_link.prev->sfmmu_scd_link.next =
14553 		    sfmmup->sfmmu_scd_link.next;
14554 	} else {
14555 		ASSERT(*headp == sfmmup);
14556 		*headp = sfmmup->sfmmu_scd_link.next;
14557 	}
14558 	if (sfmmup->sfmmu_scd_link.next != NULL) {
14559 		sfmmup->sfmmu_scd_link.next->sfmmu_scd_link.prev =
14560 		    sfmmup->sfmmu_scd_link.prev;
14561 	}
14562 }
14563 
14564 
14565 /*
14566  * Adds an sfmmu to the start of the queue.
14567  */
14568 static void
14569 sfmmu_to_scd_list(sfmmu_t **headp, sfmmu_t *sfmmup)
14570 {
14571 	check_scd_sfmmu_list(headp, sfmmup, 0);
14572 	sfmmup->sfmmu_scd_link.prev = NULL;
14573 	sfmmup->sfmmu_scd_link.next = *headp;
14574 	if (*headp != NULL)
14575 		(*headp)->sfmmu_scd_link.prev = sfmmup;
14576 	*headp = sfmmup;
14577 }
14578 
14579 /*
14580  * Remove an scd from the start of the queue.
14581  */
14582 static void
14583 sfmmu_remove_scd(sf_scd_t **headp, sf_scd_t *scdp)
14584 {
14585 	if (scdp->scd_prev != NULL) {
14586 		ASSERT(*headp != scdp);
14587 		scdp->scd_prev->scd_next = scdp->scd_next;
14588 	} else {
14589 		ASSERT(*headp == scdp);
14590 		*headp = scdp->scd_next;
14591 	}
14592 
14593 	if (scdp->scd_next != NULL) {
14594 		scdp->scd_next->scd_prev = scdp->scd_prev;
14595 	}
14596 }
14597 
14598 /*
14599  * Add an scd to the start of the queue.
14600  */
14601 static void
14602 sfmmu_add_scd(sf_scd_t **headp, sf_scd_t *scdp)
14603 {
14604 	scdp->scd_prev = NULL;
14605 	scdp->scd_next = *headp;
14606 	if (*headp != NULL) {
14607 		(*headp)->scd_prev = scdp;
14608 	}
14609 	*headp = scdp;
14610 }
14611 
14612 static int
14613 sfmmu_alloc_scd_tsbs(sf_srd_t *srdp, sf_scd_t *scdp)
14614 {
14615 	uint_t rid;
14616 	uint_t i;
14617 	uint_t j;
14618 	ulong_t w;
14619 	sf_region_t *rgnp;
14620 	ulong_t tte8k_cnt = 0;
14621 	ulong_t tte4m_cnt = 0;
14622 	uint_t tsb_szc;
14623 	sfmmu_t *scsfmmup = scdp->scd_sfmmup;
14624 	sfmmu_t	*ism_hatid;
14625 	struct tsb_info *newtsb;
14626 	int szc;
14627 
14628 	ASSERT(srdp != NULL);
14629 
14630 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14631 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14632 			continue;
14633 		}
14634 		j = 0;
14635 		while (w) {
14636 			if (!(w & 0x1)) {
14637 				j++;
14638 				w >>= 1;
14639 				continue;
14640 			}
14641 			rid = (i << BT_ULSHIFT) | j;
14642 			j++;
14643 			w >>= 1;
14644 
14645 			if (rid < SFMMU_MAX_HME_REGIONS) {
14646 				rgnp = srdp->srd_hmergnp[rid];
14647 				ASSERT(rgnp->rgn_id == rid);
14648 				ASSERT(rgnp->rgn_refcnt > 0);
14649 
14650 				if (rgnp->rgn_pgszc < TTE4M) {
14651 					tte8k_cnt += rgnp->rgn_size >>
14652 					    TTE_PAGE_SHIFT(TTE8K);
14653 				} else {
14654 					ASSERT(rgnp->rgn_pgszc >= TTE4M);
14655 					tte4m_cnt += rgnp->rgn_size >>
14656 					    TTE_PAGE_SHIFT(TTE4M);
14657 					/*
14658 					 * Inflate SCD tsb0 by preallocating
14659 					 * 1/4 8k ttecnt for 4M regions to
14660 					 * allow for lgpg alloc failure.
14661 					 */
14662 					tte8k_cnt += rgnp->rgn_size >>
14663 					    (TTE_PAGE_SHIFT(TTE8K) + 2);
14664 				}
14665 			} else {
14666 				rid -= SFMMU_MAX_HME_REGIONS;
14667 				rgnp = srdp->srd_ismrgnp[rid];
14668 				ASSERT(rgnp->rgn_id == rid);
14669 				ASSERT(rgnp->rgn_refcnt > 0);
14670 
14671 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14672 				ASSERT(ism_hatid->sfmmu_ismhat);
14673 
14674 				for (szc = 0; szc < TTE4M; szc++) {
14675 					tte8k_cnt +=
14676 					    ism_hatid->sfmmu_ttecnt[szc] <<
14677 					    TTE_BSZS_SHIFT(szc);
14678 				}
14679 
14680 				ASSERT(rgnp->rgn_pgszc >= TTE4M);
14681 				if (rgnp->rgn_pgszc >= TTE4M) {
14682 					tte4m_cnt += rgnp->rgn_size >>
14683 					    TTE_PAGE_SHIFT(TTE4M);
14684 				}
14685 			}
14686 		}
14687 	}
14688 
14689 	tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
14690 
14691 	/* Allocate both the SCD TSBs here. */
14692 	if (sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14693 	    tsb_szc, TSB8K|TSB64K|TSB512K, TSB_ALLOC, scsfmmup) &&
14694 	    (tsb_szc <= TSB_4M_SZCODE ||
14695 	    sfmmu_tsbinfo_alloc(&scsfmmup->sfmmu_tsb,
14696 	    TSB_4M_SZCODE, TSB8K|TSB64K|TSB512K,
14697 	    TSB_ALLOC, scsfmmup))) {
14698 
14699 		SFMMU_STAT(sf_scd_1sttsb_allocfail);
14700 		return (TSB_ALLOCFAIL);
14701 	} else {
14702 		scsfmmup->sfmmu_tsb->tsb_flags |= TSB_SHAREDCTX;
14703 
14704 		if (tte4m_cnt) {
14705 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
14706 			if (sfmmu_tsbinfo_alloc(&newtsb, tsb_szc,
14707 			    TSB4M|TSB32M|TSB256M, TSB_ALLOC, scsfmmup) &&
14708 			    (tsb_szc <= TSB_4M_SZCODE ||
14709 			    sfmmu_tsbinfo_alloc(&newtsb, TSB_4M_SZCODE,
14710 			    TSB4M|TSB32M|TSB256M,
14711 			    TSB_ALLOC, scsfmmup))) {
14712 				/*
14713 				 * If we fail to allocate the 2nd shared tsb,
14714 				 * just free the 1st tsb, return failure.
14715 				 */
14716 				sfmmu_tsbinfo_free(scsfmmup->sfmmu_tsb);
14717 				SFMMU_STAT(sf_scd_2ndtsb_allocfail);
14718 				return (TSB_ALLOCFAIL);
14719 			} else {
14720 				ASSERT(scsfmmup->sfmmu_tsb->tsb_next == NULL);
14721 				newtsb->tsb_flags |= TSB_SHAREDCTX;
14722 				scsfmmup->sfmmu_tsb->tsb_next = newtsb;
14723 				SFMMU_STAT(sf_scd_2ndtsb_alloc);
14724 			}
14725 		}
14726 		SFMMU_STAT(sf_scd_1sttsb_alloc);
14727 	}
14728 	return (TSB_SUCCESS);
14729 }
14730 
14731 static void
14732 sfmmu_free_scd_tsbs(sfmmu_t *scd_sfmmu)
14733 {
14734 	while (scd_sfmmu->sfmmu_tsb != NULL) {
14735 		struct tsb_info *next = scd_sfmmu->sfmmu_tsb->tsb_next;
14736 		sfmmu_tsbinfo_free(scd_sfmmu->sfmmu_tsb);
14737 		scd_sfmmu->sfmmu_tsb = next;
14738 	}
14739 }
14740 
14741 /*
14742  * Link the sfmmu onto the hme region list.
14743  */
14744 void
14745 sfmmu_link_to_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14746 {
14747 	uint_t rid;
14748 	sf_rgn_link_t *rlink;
14749 	sfmmu_t *head;
14750 	sf_rgn_link_t *hrlink;
14751 
14752 	rid = rgnp->rgn_id;
14753 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14754 
14755 	/* LINTED: constant in conditional context */
14756 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 1, 1);
14757 	ASSERT(rlink != NULL);
14758 	mutex_enter(&rgnp->rgn_mutex);
14759 	if ((head = rgnp->rgn_sfmmu_head) == NULL) {
14760 		rlink->next = NULL;
14761 		rlink->prev = NULL;
14762 		/*
14763 		 * make sure rlink's next field is NULL
14764 		 * before making this link visible.
14765 		 */
14766 		membar_stst();
14767 		rgnp->rgn_sfmmu_head = sfmmup;
14768 	} else {
14769 		/* LINTED: constant in conditional context */
14770 		SFMMU_HMERID2RLINKP(head, rid, hrlink, 0, 0);
14771 		ASSERT(hrlink != NULL);
14772 		ASSERT(hrlink->prev == NULL);
14773 		rlink->next = head;
14774 		rlink->prev = NULL;
14775 		hrlink->prev = sfmmup;
14776 		/*
14777 		 * make sure rlink's next field is correct
14778 		 * before making this link visible.
14779 		 */
14780 		membar_stst();
14781 		rgnp->rgn_sfmmu_head = sfmmup;
14782 	}
14783 	mutex_exit(&rgnp->rgn_mutex);
14784 }
14785 
14786 /*
14787  * Unlink the sfmmu from the hme region list.
14788  */
14789 void
14790 sfmmu_unlink_from_hmeregion(sfmmu_t *sfmmup, sf_region_t *rgnp)
14791 {
14792 	uint_t rid;
14793 	sf_rgn_link_t *rlink;
14794 
14795 	rid = rgnp->rgn_id;
14796 	ASSERT(SFMMU_IS_SHMERID_VALID(rid));
14797 
14798 	/* LINTED: constant in conditional context */
14799 	SFMMU_HMERID2RLINKP(sfmmup, rid, rlink, 0, 0);
14800 	ASSERT(rlink != NULL);
14801 	mutex_enter(&rgnp->rgn_mutex);
14802 	if (rgnp->rgn_sfmmu_head == sfmmup) {
14803 		sfmmu_t *next = rlink->next;
14804 		rgnp->rgn_sfmmu_head = next;
14805 		/*
14806 		 * if we are stopped by xc_attention() after this
14807 		 * point the forward link walking in
14808 		 * sfmmu_rgntlb_demap() will work correctly since the
14809 		 * head correctly points to the next element.
14810 		 */
14811 		membar_stst();
14812 		rlink->next = NULL;
14813 		ASSERT(rlink->prev == NULL);
14814 		if (next != NULL) {
14815 			sf_rgn_link_t *nrlink;
14816 			/* LINTED: constant in conditional context */
14817 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14818 			ASSERT(nrlink != NULL);
14819 			ASSERT(nrlink->prev == sfmmup);
14820 			nrlink->prev = NULL;
14821 		}
14822 	} else {
14823 		sfmmu_t *next = rlink->next;
14824 		sfmmu_t *prev = rlink->prev;
14825 		sf_rgn_link_t *prlink;
14826 
14827 		ASSERT(prev != NULL);
14828 		/* LINTED: constant in conditional context */
14829 		SFMMU_HMERID2RLINKP(prev, rid, prlink, 0, 0);
14830 		ASSERT(prlink != NULL);
14831 		ASSERT(prlink->next == sfmmup);
14832 		prlink->next = next;
14833 		/*
14834 		 * if we are stopped by xc_attention()
14835 		 * after this point the forward link walking
14836 		 * will work correctly since the prev element
14837 		 * correctly points to the next element.
14838 		 */
14839 		membar_stst();
14840 		rlink->next = NULL;
14841 		rlink->prev = NULL;
14842 		if (next != NULL) {
14843 			sf_rgn_link_t *nrlink;
14844 			/* LINTED: constant in conditional context */
14845 			SFMMU_HMERID2RLINKP(next, rid, nrlink, 0, 0);
14846 			ASSERT(nrlink != NULL);
14847 			ASSERT(nrlink->prev == sfmmup);
14848 			nrlink->prev = prev;
14849 		}
14850 	}
14851 	mutex_exit(&rgnp->rgn_mutex);
14852 }
14853 
14854 /*
14855  * Link scd sfmmu onto ism or hme region list for each region in the
14856  * scd region map.
14857  */
14858 void
14859 sfmmu_link_scd_to_regions(sf_srd_t *srdp, sf_scd_t *scdp)
14860 {
14861 	uint_t rid;
14862 	uint_t i;
14863 	uint_t j;
14864 	ulong_t w;
14865 	sf_region_t *rgnp;
14866 	sfmmu_t *scsfmmup;
14867 
14868 	scsfmmup = scdp->scd_sfmmup;
14869 	ASSERT(scsfmmup->sfmmu_scdhat);
14870 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14871 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14872 			continue;
14873 		}
14874 		j = 0;
14875 		while (w) {
14876 			if (!(w & 0x1)) {
14877 				j++;
14878 				w >>= 1;
14879 				continue;
14880 			}
14881 			rid = (i << BT_ULSHIFT) | j;
14882 			j++;
14883 			w >>= 1;
14884 
14885 			if (rid < SFMMU_MAX_HME_REGIONS) {
14886 				rgnp = srdp->srd_hmergnp[rid];
14887 				ASSERT(rgnp->rgn_id == rid);
14888 				ASSERT(rgnp->rgn_refcnt > 0);
14889 				sfmmu_link_to_hmeregion(scsfmmup, rgnp);
14890 			} else {
14891 				sfmmu_t *ism_hatid = NULL;
14892 				ism_ment_t *ism_ment;
14893 				rid -= SFMMU_MAX_HME_REGIONS;
14894 				rgnp = srdp->srd_ismrgnp[rid];
14895 				ASSERT(rgnp->rgn_id == rid);
14896 				ASSERT(rgnp->rgn_refcnt > 0);
14897 
14898 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14899 				ASSERT(ism_hatid->sfmmu_ismhat);
14900 				ism_ment = &scdp->scd_ism_links[rid];
14901 				ism_ment->iment_hat = scsfmmup;
14902 				ism_ment->iment_base_va = rgnp->rgn_saddr;
14903 				mutex_enter(&ism_mlist_lock);
14904 				iment_add(ism_ment, ism_hatid);
14905 				mutex_exit(&ism_mlist_lock);
14906 
14907 			}
14908 		}
14909 	}
14910 }
14911 /*
14912  * Unlink scd sfmmu from ism or hme region list for each region in the
14913  * scd region map.
14914  */
14915 void
14916 sfmmu_unlink_scd_from_regions(sf_srd_t *srdp, sf_scd_t *scdp)
14917 {
14918 	uint_t rid;
14919 	uint_t i;
14920 	uint_t j;
14921 	ulong_t w;
14922 	sf_region_t *rgnp;
14923 	sfmmu_t *scsfmmup;
14924 
14925 	scsfmmup = scdp->scd_sfmmup;
14926 	for (i = 0; i < SFMMU_RGNMAP_WORDS; i++) {
14927 		if ((w = scdp->scd_region_map.bitmap[i]) == 0) {
14928 			continue;
14929 		}
14930 		j = 0;
14931 		while (w) {
14932 			if (!(w & 0x1)) {
14933 				j++;
14934 				w >>= 1;
14935 				continue;
14936 			}
14937 			rid = (i << BT_ULSHIFT) | j;
14938 			j++;
14939 			w >>= 1;
14940 
14941 			if (rid < SFMMU_MAX_HME_REGIONS) {
14942 				rgnp = srdp->srd_hmergnp[rid];
14943 				ASSERT(rgnp->rgn_id == rid);
14944 				ASSERT(rgnp->rgn_refcnt > 0);
14945 				sfmmu_unlink_from_hmeregion(scsfmmup,
14946 				    rgnp);
14947 
14948 			} else {
14949 				sfmmu_t *ism_hatid = NULL;
14950 				ism_ment_t *ism_ment;
14951 				rid -= SFMMU_MAX_HME_REGIONS;
14952 				rgnp = srdp->srd_ismrgnp[rid];
14953 				ASSERT(rgnp->rgn_id == rid);
14954 				ASSERT(rgnp->rgn_refcnt > 0);
14955 
14956 				ism_hatid = (sfmmu_t *)rgnp->rgn_obj;
14957 				ASSERT(ism_hatid->sfmmu_ismhat);
14958 				ism_ment = &scdp->scd_ism_links[rid];
14959 				ASSERT(ism_ment->iment_hat == scdp->scd_sfmmup);
14960 				ASSERT(ism_ment->iment_base_va ==
14961 				    rgnp->rgn_saddr);
14962 				ism_ment->iment_hat = NULL;
14963 				ism_ment->iment_base_va = 0;
14964 				mutex_enter(&ism_mlist_lock);
14965 				iment_sub(ism_ment, ism_hatid);
14966 				mutex_exit(&ism_mlist_lock);
14967 
14968 			}
14969 		}
14970 	}
14971 }
14972 /*
14973  * Allocates and initialises a new SCD structure, this is called with
14974  * the srd_scd_mutex held and returns with the reference count
14975  * initialised to 1.
14976  */
14977 static sf_scd_t *
14978 sfmmu_alloc_scd(sf_srd_t *srdp, sf_region_map_t *new_map)
14979 {
14980 	sf_scd_t *new_scdp;
14981 	sfmmu_t *scsfmmup;
14982 	int i;
14983 
14984 	ASSERT(MUTEX_HELD(&srdp->srd_scd_mutex));
14985 	new_scdp = kmem_cache_alloc(scd_cache, KM_SLEEP);
14986 
14987 	scsfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
14988 	new_scdp->scd_sfmmup = scsfmmup;
14989 	scsfmmup->sfmmu_srdp = srdp;
14990 	scsfmmup->sfmmu_scdp = new_scdp;
14991 	scsfmmup->sfmmu_tsb0_4minflcnt = 0;
14992 	scsfmmup->sfmmu_scdhat = 1;
14993 	CPUSET_ALL(scsfmmup->sfmmu_cpusran);
14994 	bzero(scsfmmup->sfmmu_hmeregion_links, SFMMU_L1_HMERLINKS_SIZE);
14995 
14996 	ASSERT(max_mmu_ctxdoms > 0);
14997 	for (i = 0; i < max_mmu_ctxdoms; i++) {
14998 		scsfmmup->sfmmu_ctxs[i].cnum = INVALID_CONTEXT;
14999 		scsfmmup->sfmmu_ctxs[i].gnum = 0;
15000 	}
15001 
15002 	for (i = 0; i < MMU_PAGE_SIZES; i++) {
15003 		new_scdp->scd_rttecnt[i] = 0;
15004 	}
15005 
15006 	new_scdp->scd_region_map = *new_map;
15007 	new_scdp->scd_refcnt = 1;
15008 	if (sfmmu_alloc_scd_tsbs(srdp, new_scdp) != TSB_SUCCESS) {
15009 		kmem_cache_free(scd_cache, new_scdp);
15010 		kmem_cache_free(sfmmuid_cache, scsfmmup);
15011 		return (NULL);
15012 	}
15013 	return (new_scdp);
15014 }
15015 
15016 /*
15017  * The first phase of a process joining an SCD. The hat structure is
15018  * linked to the SCD queue and then the HAT_JOIN_SCD sfmmu flag is set
15019  * and a cross-call with context invalidation is used to cause the
15020  * remaining work to be carried out in the sfmmu_tsbmiss_exception()
15021  * routine.
15022  */
15023 static void
15024 sfmmu_join_scd(sf_scd_t *scdp, sfmmu_t *sfmmup)
15025 {
15026 	hatlock_t *hatlockp;
15027 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15028 	int i;
15029 	sf_scd_t *old_scdp;
15030 
15031 	ASSERT(srdp != NULL);
15032 	ASSERT(scdp != NULL);
15033 	ASSERT(scdp->scd_refcnt > 0);
15034 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15035 
15036 	if ((old_scdp = sfmmup->sfmmu_scdp) != NULL) {
15037 		ASSERT(old_scdp != scdp);
15038 
15039 		mutex_enter(&old_scdp->scd_mutex);
15040 		sfmmu_from_scd_list(&old_scdp->scd_sf_list, sfmmup);
15041 		mutex_exit(&old_scdp->scd_mutex);
15042 		/*
15043 		 * sfmmup leaves the old scd. Update sfmmu_ttecnt to
15044 		 * include the shme rgn ttecnt for rgns that
15045 		 * were in the old SCD
15046 		 */
15047 		for (i = 0; i < mmu_page_sizes; i++) {
15048 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15049 			    old_scdp->scd_rttecnt[i]);
15050 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15051 			    sfmmup->sfmmu_scdrttecnt[i]);
15052 		}
15053 	}
15054 
15055 	/*
15056 	 * Move sfmmu to the scd lists.
15057 	 */
15058 	mutex_enter(&scdp->scd_mutex);
15059 	sfmmu_to_scd_list(&scdp->scd_sf_list, sfmmup);
15060 	mutex_exit(&scdp->scd_mutex);
15061 	SF_SCD_INCR_REF(scdp);
15062 
15063 	hatlockp = sfmmu_hat_enter(sfmmup);
15064 	/*
15065 	 * For a multi-thread process, we must stop
15066 	 * all the other threads before joining the scd.
15067 	 */
15068 
15069 	SFMMU_FLAGS_SET(sfmmup, HAT_JOIN_SCD);
15070 
15071 	sfmmu_invalidate_ctx(sfmmup);
15072 	sfmmup->sfmmu_scdp = scdp;
15073 
15074 	/*
15075 	 * Copy scd_rttecnt into sfmmup's sfmmu_scdrttecnt, and update
15076 	 * sfmmu_ttecnt to not include the rgn ttecnt just joined in SCD.
15077 	 */
15078 	for (i = 0; i < mmu_page_sizes; i++) {
15079 		sfmmup->sfmmu_scdrttecnt[i] = scdp->scd_rttecnt[i];
15080 		ASSERT(sfmmup->sfmmu_ttecnt[i] >= scdp->scd_rttecnt[i]);
15081 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15082 		    -sfmmup->sfmmu_scdrttecnt[i]);
15083 	}
15084 	/* update tsb0 inflation count */
15085 	if (old_scdp != NULL) {
15086 		sfmmup->sfmmu_tsb0_4minflcnt +=
15087 		    old_scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15088 	}
15089 	ASSERT(sfmmup->sfmmu_tsb0_4minflcnt >=
15090 	    scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt);
15091 	sfmmup->sfmmu_tsb0_4minflcnt -= scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15092 
15093 	sfmmu_hat_exit(hatlockp);
15094 
15095 	if (old_scdp != NULL) {
15096 		SF_SCD_DECR_REF(srdp, old_scdp);
15097 	}
15098 
15099 }
15100 
15101 /*
15102  * This routine is called by a process to become part of an SCD. It is called
15103  * from sfmmu_tsbmiss_exception() once most of the initial work has been
15104  * done by sfmmu_join_scd(). This routine must not drop the hat lock.
15105  */
15106 static void
15107 sfmmu_finish_join_scd(sfmmu_t *sfmmup)
15108 {
15109 	struct tsb_info	*tsbinfop;
15110 
15111 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15112 	ASSERT(sfmmup->sfmmu_scdp != NULL);
15113 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD));
15114 	ASSERT(!SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15115 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ALLCTX_INVALID));
15116 
15117 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
15118 	    tsbinfop = tsbinfop->tsb_next) {
15119 		if (tsbinfop->tsb_flags & TSB_SWAPPED) {
15120 			continue;
15121 		}
15122 		ASSERT(!(tsbinfop->tsb_flags & TSB_RELOC_FLAG));
15123 
15124 		sfmmu_inv_tsb(tsbinfop->tsb_va,
15125 		    TSB_BYTES(tsbinfop->tsb_szc));
15126 	}
15127 
15128 	/* Set HAT_CTX1_FLAG for all SCD ISMs */
15129 	sfmmu_ism_hatflags(sfmmup, 1);
15130 
15131 	SFMMU_STAT(sf_join_scd);
15132 }
15133 
15134 /*
15135  * This routine is called in order to check if there is an SCD which matches
15136  * the process's region map if not then a new SCD may be created.
15137  */
15138 static void
15139 sfmmu_find_scd(sfmmu_t *sfmmup)
15140 {
15141 	sf_srd_t *srdp = sfmmup->sfmmu_srdp;
15142 	sf_scd_t *scdp, *new_scdp;
15143 	int ret;
15144 
15145 	ASSERT(srdp != NULL);
15146 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15147 
15148 	mutex_enter(&srdp->srd_scd_mutex);
15149 	for (scdp = srdp->srd_scdp; scdp != NULL;
15150 	    scdp = scdp->scd_next) {
15151 		SF_RGNMAP_EQUAL(&scdp->scd_region_map,
15152 		    &sfmmup->sfmmu_region_map, ret);
15153 		if (ret == 1) {
15154 			SF_SCD_INCR_REF(scdp);
15155 			mutex_exit(&srdp->srd_scd_mutex);
15156 			sfmmu_join_scd(scdp, sfmmup);
15157 			ASSERT(scdp->scd_refcnt >= 2);
15158 			atomic_add_32((volatile uint32_t *)
15159 			    &scdp->scd_refcnt, -1);
15160 			return;
15161 		} else {
15162 			/*
15163 			 * If the sfmmu region map is a subset of the scd
15164 			 * region map, then the assumption is that this process
15165 			 * will continue attaching to ISM segments until the
15166 			 * region maps are equal.
15167 			 */
15168 			SF_RGNMAP_IS_SUBSET(&scdp->scd_region_map,
15169 			    &sfmmup->sfmmu_region_map, ret);
15170 			if (ret == 1) {
15171 				mutex_exit(&srdp->srd_scd_mutex);
15172 				return;
15173 			}
15174 		}
15175 	}
15176 
15177 	ASSERT(scdp == NULL);
15178 	/*
15179 	 * No matching SCD has been found, create a new one.
15180 	 */
15181 	if ((new_scdp = sfmmu_alloc_scd(srdp, &sfmmup->sfmmu_region_map)) ==
15182 	    NULL) {
15183 		mutex_exit(&srdp->srd_scd_mutex);
15184 		return;
15185 	}
15186 
15187 	/*
15188 	 * sfmmu_alloc_scd() returns with a ref count of 1 on the scd.
15189 	 */
15190 
15191 	/* Set scd_rttecnt for shme rgns in SCD */
15192 	sfmmu_set_scd_rttecnt(srdp, new_scdp);
15193 
15194 	/*
15195 	 * Link scd onto srd_scdp list and scd sfmmu onto region/iment lists.
15196 	 */
15197 	sfmmu_link_scd_to_regions(srdp, new_scdp);
15198 	sfmmu_add_scd(&srdp->srd_scdp, new_scdp);
15199 	SFMMU_STAT_ADD(sf_create_scd, 1);
15200 
15201 	mutex_exit(&srdp->srd_scd_mutex);
15202 	sfmmu_join_scd(new_scdp, sfmmup);
15203 	ASSERT(new_scdp->scd_refcnt >= 2);
15204 	atomic_add_32((volatile uint32_t *)&new_scdp->scd_refcnt, -1);
15205 }
15206 
15207 /*
15208  * This routine is called by a process to remove itself from an SCD. It is
15209  * either called when the processes has detached from a segment or from
15210  * hat_free_start() as a result of calling exit.
15211  */
15212 static void
15213 sfmmu_leave_scd(sfmmu_t *sfmmup, uchar_t r_type)
15214 {
15215 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15216 	sf_srd_t *srdp =  sfmmup->sfmmu_srdp;
15217 	hatlock_t *hatlockp = TSB_HASH(sfmmup);
15218 	int i;
15219 
15220 	ASSERT(scdp != NULL);
15221 	ASSERT(srdp != NULL);
15222 
15223 	if (sfmmup->sfmmu_free) {
15224 		/*
15225 		 * If the process is part of an SCD the sfmmu is unlinked
15226 		 * from scd_sf_list.
15227 		 */
15228 		mutex_enter(&scdp->scd_mutex);
15229 		sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15230 		mutex_exit(&scdp->scd_mutex);
15231 		/*
15232 		 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15233 		 * are about to leave the SCD
15234 		 */
15235 		for (i = 0; i < mmu_page_sizes; i++) {
15236 			ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15237 			    scdp->scd_rttecnt[i]);
15238 			atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15239 			    sfmmup->sfmmu_scdrttecnt[i]);
15240 			sfmmup->sfmmu_scdrttecnt[i] = 0;
15241 		}
15242 		sfmmup->sfmmu_scdp = NULL;
15243 
15244 		SF_SCD_DECR_REF(srdp, scdp);
15245 		return;
15246 	}
15247 
15248 	ASSERT(r_type != SFMMU_REGION_ISM ||
15249 	    SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15250 	ASSERT(scdp->scd_refcnt);
15251 	ASSERT(!sfmmup->sfmmu_free);
15252 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15253 	ASSERT(AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
15254 
15255 	/*
15256 	 * Wait for ISM maps to be updated.
15257 	 */
15258 	if (r_type != SFMMU_REGION_ISM) {
15259 		while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY) &&
15260 		    sfmmup->sfmmu_scdp != NULL) {
15261 			cv_wait(&sfmmup->sfmmu_tsb_cv,
15262 			    HATLOCK_MUTEXP(hatlockp));
15263 		}
15264 
15265 		if (sfmmup->sfmmu_scdp == NULL) {
15266 			sfmmu_hat_exit(hatlockp);
15267 			return;
15268 		}
15269 		SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
15270 	}
15271 
15272 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_JOIN_SCD)) {
15273 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_JOIN_SCD);
15274 		/*
15275 		 * Since HAT_JOIN_SCD was set our context
15276 		 * is still invalid.
15277 		 */
15278 	} else {
15279 		/*
15280 		 * For a multi-thread process, we must stop
15281 		 * all the other threads before leaving the scd.
15282 		 */
15283 
15284 		sfmmu_invalidate_ctx(sfmmup);
15285 	}
15286 
15287 	/* Clear all the rid's for ISM, delete flags, etc */
15288 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
15289 	sfmmu_ism_hatflags(sfmmup, 0);
15290 
15291 	/*
15292 	 * Update sfmmu_ttecnt to include the rgn ttecnt for rgns that
15293 	 * are in SCD before this sfmmup leaves the SCD.
15294 	 */
15295 	for (i = 0; i < mmu_page_sizes; i++) {
15296 		ASSERT(sfmmup->sfmmu_scdrttecnt[i] ==
15297 		    scdp->scd_rttecnt[i]);
15298 		atomic_add_long(&sfmmup->sfmmu_ttecnt[i],
15299 		    sfmmup->sfmmu_scdrttecnt[i]);
15300 		sfmmup->sfmmu_scdrttecnt[i] = 0;
15301 		/* update ismttecnt to include SCD ism before hat leaves SCD */
15302 		sfmmup->sfmmu_ismttecnt[i] += sfmmup->sfmmu_scdismttecnt[i];
15303 		sfmmup->sfmmu_scdismttecnt[i] = 0;
15304 	}
15305 	/* update tsb0 inflation count */
15306 	sfmmup->sfmmu_tsb0_4minflcnt += scdp->scd_sfmmup->sfmmu_tsb0_4minflcnt;
15307 
15308 	if (r_type != SFMMU_REGION_ISM) {
15309 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
15310 	}
15311 	sfmmup->sfmmu_scdp = NULL;
15312 
15313 	sfmmu_hat_exit(hatlockp);
15314 
15315 	/*
15316 	 * Unlink sfmmu from scd_sf_list this can be done without holding
15317 	 * the hat lock as we hold the sfmmu_as lock which prevents
15318 	 * hat_join_region from adding this thread to the scd again. Other
15319 	 * threads check if sfmmu_scdp is NULL under hat lock and if it's NULL
15320 	 * they won't get here, since sfmmu_leave_scd() clears sfmmu_scdp
15321 	 * while holding the hat lock.
15322 	 */
15323 	mutex_enter(&scdp->scd_mutex);
15324 	sfmmu_from_scd_list(&scdp->scd_sf_list, sfmmup);
15325 	mutex_exit(&scdp->scd_mutex);
15326 	SFMMU_STAT(sf_leave_scd);
15327 
15328 	SF_SCD_DECR_REF(srdp, scdp);
15329 	hatlockp = sfmmu_hat_enter(sfmmup);
15330 
15331 }
15332 
15333 /*
15334  * Unlink and free up an SCD structure with a reference count of 0.
15335  */
15336 static void
15337 sfmmu_destroy_scd(sf_srd_t *srdp, sf_scd_t *scdp, sf_region_map_t *scd_rmap)
15338 {
15339 	sfmmu_t *scsfmmup;
15340 	sf_scd_t *sp;
15341 	hatlock_t *shatlockp;
15342 	int i, ret;
15343 
15344 	mutex_enter(&srdp->srd_scd_mutex);
15345 	for (sp = srdp->srd_scdp; sp != NULL; sp = sp->scd_next) {
15346 		if (sp == scdp)
15347 			break;
15348 	}
15349 	if (sp == NULL || sp->scd_refcnt) {
15350 		mutex_exit(&srdp->srd_scd_mutex);
15351 		return;
15352 	}
15353 
15354 	/*
15355 	 * It is possible that the scd has been freed and reallocated with a
15356 	 * different region map while we've been waiting for the srd_scd_mutex.
15357 	 */
15358 	SF_RGNMAP_EQUAL(scd_rmap, &sp->scd_region_map, ret);
15359 	if (ret != 1) {
15360 		mutex_exit(&srdp->srd_scd_mutex);
15361 		return;
15362 	}
15363 
15364 	ASSERT(scdp->scd_sf_list == NULL);
15365 	/*
15366 	 * Unlink scd from srd_scdp list.
15367 	 */
15368 	sfmmu_remove_scd(&srdp->srd_scdp, scdp);
15369 	mutex_exit(&srdp->srd_scd_mutex);
15370 
15371 	sfmmu_unlink_scd_from_regions(srdp, scdp);
15372 
15373 	/* Clear shared context tsb and release ctx */
15374 	scsfmmup = scdp->scd_sfmmup;
15375 
15376 	/*
15377 	 * create a barrier so that scd will not be destroyed
15378 	 * if other thread still holds the same shared hat lock.
15379 	 * E.g., sfmmu_tsbmiss_exception() needs to acquire the
15380 	 * shared hat lock before checking the shared tsb reloc flag.
15381 	 */
15382 	shatlockp = sfmmu_hat_enter(scsfmmup);
15383 	sfmmu_hat_exit(shatlockp);
15384 
15385 	sfmmu_free_scd_tsbs(scsfmmup);
15386 
15387 	for (i = 0; i < SFMMU_L1_HMERLINKS; i++) {
15388 		if (scsfmmup->sfmmu_hmeregion_links[i] != NULL) {
15389 			kmem_free(scsfmmup->sfmmu_hmeregion_links[i],
15390 			    SFMMU_L2_HMERLINKS_SIZE);
15391 			scsfmmup->sfmmu_hmeregion_links[i] = NULL;
15392 		}
15393 	}
15394 	kmem_cache_free(sfmmuid_cache, scsfmmup);
15395 	kmem_cache_free(scd_cache, scdp);
15396 	SFMMU_STAT(sf_destroy_scd);
15397 }
15398 
15399 /*
15400  * Modifies the HAT_CTX1_FLAG for each of the ISM segments which correspond to
15401  * bits which are set in the ism_region_map parameter. This flag indicates to
15402  * the tsbmiss handler that mapping for these segments should be loaded using
15403  * the shared context.
15404  */
15405 static void
15406 sfmmu_ism_hatflags(sfmmu_t *sfmmup, int addflag)
15407 {
15408 	sf_scd_t *scdp = sfmmup->sfmmu_scdp;
15409 	ism_blk_t *ism_blkp;
15410 	ism_map_t *ism_map;
15411 	int i, rid;
15412 
15413 	ASSERT(sfmmup->sfmmu_iblk != NULL);
15414 	ASSERT(scdp != NULL);
15415 	/*
15416 	 * Note that the caller either set HAT_ISMBUSY flag or checked
15417 	 * under hat lock that HAT_ISMBUSY was not set by another thread.
15418 	 */
15419 	ASSERT(sfmmu_hat_lock_held(sfmmup));
15420 
15421 	ism_blkp = sfmmup->sfmmu_iblk;
15422 	while (ism_blkp != NULL) {
15423 		ism_map = ism_blkp->iblk_maps;
15424 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
15425 			rid = ism_map[i].imap_rid;
15426 			if (rid == SFMMU_INVALID_ISMRID) {
15427 				continue;
15428 			}
15429 			ASSERT(rid >= 0 && rid < SFMMU_MAX_ISM_REGIONS);
15430 			if (SF_RGNMAP_TEST(scdp->scd_ismregion_map, rid) &&
15431 			    addflag) {
15432 				ism_map[i].imap_hatflags |=
15433 				    HAT_CTX1_FLAG;
15434 			} else {
15435 				ism_map[i].imap_hatflags &=
15436 				    ~HAT_CTX1_FLAG;
15437 			}
15438 		}
15439 		ism_blkp = ism_blkp->iblk_next;
15440 	}
15441 }
15442 
15443 static int
15444 sfmmu_srd_lock_held(sf_srd_t *srdp)
15445 {
15446 	return (MUTEX_HELD(&srdp->srd_mutex));
15447 }
15448 
15449 /* ARGSUSED */
15450 static int
15451 sfmmu_scdcache_constructor(void *buf, void *cdrarg, int kmflags)
15452 {
15453 	sf_scd_t *scdp = (sf_scd_t *)buf;
15454 
15455 	bzero(buf, sizeof (sf_scd_t));
15456 	mutex_init(&scdp->scd_mutex, NULL, MUTEX_DEFAULT, NULL);
15457 	return (0);
15458 }
15459 
15460 /* ARGSUSED */
15461 static void
15462 sfmmu_scdcache_destructor(void *buf, void *cdrarg)
15463 {
15464 	sf_scd_t *scdp = (sf_scd_t *)buf;
15465 
15466 	mutex_destroy(&scdp->scd_mutex);
15467 }
15468