xref: /titanic_52/usr/src/uts/sfmmu/vm/hat_sfmmu.c (revision bdfc6d18da790deeec2e0eb09c625902defe2498)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * VM - Hardware Address Translation management for Spitfire MMU.
31  *
32  * This file implements the machine specific hardware translation
33  * needed by the VM system.  The machine independent interface is
34  * described in <vm/hat.h> while the machine dependent interface
35  * and data structures are described in <vm/hat_sfmmu.h>.
36  *
37  * The hat layer manages the address translation hardware as a cache
38  * driven by calls from the higher levels in the VM system.
39  */
40 
41 #include <sys/types.h>
42 #include <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 <sys/dtrace.h>
84 #include <vm/vm_dep.h>
85 #include <vm/xhat_sfmmu.h>
86 #include <sys/fpu/fpusystm.h>
87 
88 #if defined(SF_ERRATA_57)
89 extern caddr_t errata57_limit;
90 #endif
91 
92 #define	HME8BLK_SZ_RND		((roundup(HME8BLK_SZ, sizeof (int64_t))) /  \
93 				(sizeof (int64_t)))
94 #define	HBLK_RESERVE		((struct hme_blk *)hblk_reserve)
95 
96 #define	HBLK_RESERVE_CNT	128
97 #define	HBLK_RESERVE_MIN	20
98 
99 static struct hme_blk		*freehblkp;
100 static kmutex_t			freehblkp_lock;
101 static int			freehblkcnt;
102 
103 static int64_t			hblk_reserve[HME8BLK_SZ_RND];
104 static kmutex_t			hblk_reserve_lock;
105 static kthread_t		*hblk_reserve_thread;
106 
107 static nucleus_hblk8_info_t	nucleus_hblk8;
108 static nucleus_hblk1_info_t	nucleus_hblk1;
109 
110 /*
111  * SFMMU specific hat functions
112  */
113 void	hat_pagecachectl(struct page *, int);
114 
115 /* flags for hat_pagecachectl */
116 #define	HAT_CACHE	0x1
117 #define	HAT_UNCACHE	0x2
118 #define	HAT_TMPNC	0x4
119 
120 /*
121  * Flag to allow the creation of non-cacheable translations
122  * to system memory. It is off by default. At the moment this
123  * flag is used by the ecache error injector. The error injector
124  * will turn it on when creating such a translation then shut it
125  * off when it's finished.
126  */
127 
128 int	sfmmu_allow_nc_trans = 0;
129 
130 /*
131  * Flag to disable large page support.
132  * 	value of 1 => disable all large pages.
133  *	bits 1, 2, and 3 are to disable 64K, 512K and 4M pages respectively.
134  *
135  * For example, use the value 0x4 to disable 512K pages.
136  *
137  */
138 #define	LARGE_PAGES_OFF		0x1
139 
140 /*
141  * WARNING: 512K pages MUST be disabled for ISM/DISM. If not
142  * a process would page fault indefinitely if it tried to
143  * access a 512K page.
144  */
145 int	disable_ism_large_pages = (1 << TTE512K);
146 int	disable_large_pages = 0;
147 int	disable_auto_large_pages = 0;
148 
149 /*
150  * Private sfmmu data structures for hat management
151  */
152 static struct kmem_cache *sfmmuid_cache;
153 
154 /*
155  * Private sfmmu data structures for ctx management
156  */
157 static struct ctx	*ctxhand;	/* hand used while stealing ctxs */
158 static struct ctx	*ctxfree;	/* head of free ctx list */
159 static struct ctx	*ctxdirty;	/* head of dirty ctx list */
160 
161 /*
162  * Private sfmmu data structures for tsb management
163  */
164 static struct kmem_cache *sfmmu_tsbinfo_cache;
165 static struct kmem_cache *sfmmu_tsb8k_cache;
166 static struct kmem_cache *sfmmu_tsb_cache[NLGRPS_MAX];
167 static vmem_t *kmem_tsb_arena;
168 
169 /*
170  * sfmmu static variables for hmeblk resource management.
171  */
172 static vmem_t *hat_memload1_arena; /* HAT translation arena for sfmmu1_cache */
173 static struct kmem_cache *sfmmu8_cache;
174 static struct kmem_cache *sfmmu1_cache;
175 static struct kmem_cache *pa_hment_cache;
176 
177 static kmutex_t 	ctx_list_lock;	/* mutex for ctx free/dirty lists */
178 static kmutex_t 	ism_mlist_lock;	/* mutex for ism mapping list */
179 /*
180  * private data for ism
181  */
182 static struct kmem_cache *ism_blk_cache;
183 static struct kmem_cache *ism_ment_cache;
184 #define	ISMID_STARTADDR	NULL
185 
186 /*
187  * Whether to delay TLB flushes and use Cheetah's flush-all support
188  * when removing contexts from the dirty list.
189  */
190 int delay_tlb_flush;
191 int disable_delay_tlb_flush;
192 
193 /*
194  * ``hat_lock'' is a hashed mutex lock for protecting sfmmu TSB lists,
195  * HAT flags, synchronizing TLB/TSB coherency, and context management.
196  * The lock is hashed on the sfmmup since the case where we need to lock
197  * all processes is rare but does occur (e.g. we need to unload a shared
198  * mapping from all processes using the mapping).  We have a lot of buckets,
199  * and each slab of sfmmu_t's can use about a quarter of them, giving us
200  * a fairly good distribution without wasting too much space and overhead
201  * when we have to grab them all.
202  */
203 #define	SFMMU_NUM_LOCK	128		/* must be power of two */
204 hatlock_t	hat_lock[SFMMU_NUM_LOCK];
205 
206 /*
207  * Hash algorithm optimized for a small number of slabs.
208  *  7 is (highbit((sizeof sfmmu_t)) - 1)
209  * This hash algorithm is based upon the knowledge that sfmmu_t's come from a
210  * kmem_cache, and thus they will be sequential within that cache.  In
211  * addition, each new slab will have a different "color" up to cache_maxcolor
212  * which will skew the hashing for each successive slab which is allocated.
213  * If the size of sfmmu_t changed to a larger size, this algorithm may need
214  * to be revisited.
215  */
216 #define	TSB_HASH_SHIFT_BITS (7)
217 #define	PTR_HASH(x) ((uintptr_t)x >> TSB_HASH_SHIFT_BITS)
218 
219 #ifdef DEBUG
220 int tsb_hash_debug = 0;
221 #define	TSB_HASH(sfmmup)	\
222 	(tsb_hash_debug ? &hat_lock[0] : \
223 	&hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)])
224 #else	/* DEBUG */
225 #define	TSB_HASH(sfmmup)	&hat_lock[PTR_HASH(sfmmup) & (SFMMU_NUM_LOCK-1)]
226 #endif	/* DEBUG */
227 
228 
229 /* sfmmu_replace_tsb() return codes. */
230 typedef enum tsb_replace_rc {
231 	TSB_SUCCESS,
232 	TSB_ALLOCFAIL,
233 	TSB_LOSTRACE,
234 	TSB_ALREADY_SWAPPED,
235 	TSB_CANTGROW
236 } tsb_replace_rc_t;
237 
238 /*
239  * Flags for TSB allocation routines.
240  */
241 #define	TSB_ALLOC	0x01
242 #define	TSB_FORCEALLOC	0x02
243 #define	TSB_GROW	0x04
244 #define	TSB_SHRINK	0x08
245 #define	TSB_SWAPIN	0x10
246 
247 /*
248  * Support for HAT callbacks.
249  */
250 #define	SFMMU_MAX_RELOC_CALLBACKS	10
251 int sfmmu_max_cb_id = SFMMU_MAX_RELOC_CALLBACKS;
252 static id_t sfmmu_cb_nextid = 0;
253 static id_t sfmmu_tsb_cb_id;
254 struct sfmmu_callback *sfmmu_cb_table;
255 
256 /*
257  * Kernel page relocation is enabled by default for non-caged
258  * kernel pages.  This has little effect unless segkmem_reloc is
259  * set, since by default kernel memory comes from inside the
260  * kernel cage.
261  */
262 int hat_kpr_enabled = 1;
263 
264 kmutex_t	kpr_mutex;
265 kmutex_t	kpr_suspendlock;
266 kthread_t	*kreloc_thread;
267 
268 /*
269  * Enable VA->PA translation sanity checking on DEBUG kernels.
270  * Disabled by default.  This is incompatible with some
271  * drivers (error injector, RSM) so if it breaks you get
272  * to keep both pieces.
273  */
274 int hat_check_vtop = 0;
275 
276 /*
277  * Private sfmmu routines (prototypes)
278  */
279 static struct hme_blk *sfmmu_shadow_hcreate(sfmmu_t *, caddr_t, int, uint_t);
280 static struct 	hme_blk *sfmmu_hblk_alloc(sfmmu_t *, caddr_t,
281 			struct hmehash_bucket *, uint_t, hmeblk_tag, uint_t);
282 static caddr_t	sfmmu_hblk_unload(struct hat *, struct hme_blk *, caddr_t,
283 			caddr_t, demap_range_t *, uint_t);
284 static caddr_t	sfmmu_hblk_sync(struct hat *, struct hme_blk *, caddr_t,
285 			caddr_t, int);
286 static void	sfmmu_hblk_free(struct hmehash_bucket *, struct hme_blk *,
287 			uint64_t, struct hme_blk **);
288 static void	sfmmu_hblks_list_purge(struct hme_blk **);
289 static uint_t	sfmmu_get_free_hblk(struct hme_blk **, uint_t);
290 static uint_t	sfmmu_put_free_hblk(struct hme_blk *, uint_t);
291 static struct hme_blk *sfmmu_hblk_steal(int);
292 static int	sfmmu_steal_this_hblk(struct hmehash_bucket *,
293 			struct hme_blk *, uint64_t, uint64_t,
294 			struct hme_blk *);
295 static caddr_t	sfmmu_hblk_unlock(struct hme_blk *, caddr_t, caddr_t);
296 
297 static void	sfmmu_memload_batchsmall(struct hat *, caddr_t, page_t **,
298 		    uint_t, uint_t, pgcnt_t);
299 void		sfmmu_tteload(struct hat *, tte_t *, caddr_t, page_t *,
300 			uint_t);
301 static int	sfmmu_tteload_array(sfmmu_t *, tte_t *, caddr_t, page_t **,
302 			uint_t);
303 static struct hmehash_bucket *sfmmu_tteload_acquire_hashbucket(sfmmu_t *,
304 					caddr_t, int);
305 static struct hme_blk *sfmmu_tteload_find_hmeblk(sfmmu_t *,
306 			struct hmehash_bucket *, caddr_t, uint_t, uint_t);
307 static int	sfmmu_tteload_addentry(sfmmu_t *, struct hme_blk *, tte_t *,
308 			caddr_t, page_t **, uint_t);
309 static void	sfmmu_tteload_release_hashbucket(struct hmehash_bucket *);
310 
311 static int	sfmmu_pagearray_setup(caddr_t, page_t **, tte_t *, int);
312 pfn_t		sfmmu_uvatopfn(caddr_t, sfmmu_t *);
313 void		sfmmu_memtte(tte_t *, pfn_t, uint_t, int);
314 static void	sfmmu_vac_conflict(struct hat *, caddr_t, page_t *);
315 static int	sfmmu_vacconflict_array(caddr_t, page_t *, int *);
316 static int	tst_tnc(page_t *pp, pgcnt_t);
317 static void	conv_tnc(page_t *pp, int);
318 
319 static struct ctx *sfmmu_get_ctx(sfmmu_t *);
320 static void	sfmmu_free_ctx(sfmmu_t *, struct ctx *);
321 static void	sfmmu_free_sfmmu(sfmmu_t *);
322 
323 static void	sfmmu_gettte(struct hat *, caddr_t, tte_t *);
324 static void	sfmmu_ttesync(struct hat *, caddr_t, tte_t *, page_t *);
325 static void	sfmmu_chgattr(struct hat *, caddr_t, size_t, uint_t, int);
326 
327 static cpuset_t	sfmmu_pageunload(page_t *, struct sf_hment *, int);
328 static void	hat_pagereload(struct page *, struct page *);
329 static cpuset_t	sfmmu_pagesync(page_t *, struct sf_hment *, uint_t);
330 static void	sfmmu_page_cache_array(page_t *, int, int, pgcnt_t);
331 static void	sfmmu_page_cache(page_t *, int, int, int);
332 
333 static void	sfmmu_tlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *,
334 			pfn_t, int, int, int, int);
335 static void	sfmmu_ismtlbcache_demap(caddr_t, sfmmu_t *, struct hme_blk *,
336 			pfn_t, int);
337 static void	sfmmu_tlb_demap(caddr_t, sfmmu_t *, struct hme_blk *, int, int);
338 static void	sfmmu_tlb_range_demap(demap_range_t *);
339 static void	sfmmu_tlb_ctx_demap(sfmmu_t *);
340 static void	sfmmu_tlb_all_demap(void);
341 static void	sfmmu_tlb_swap_ctx(sfmmu_t *, struct ctx *);
342 static void	sfmmu_sync_mmustate(sfmmu_t *);
343 
344 static void 	sfmmu_tsbinfo_setup_phys(struct tsb_info *, pfn_t);
345 static int	sfmmu_tsbinfo_alloc(struct tsb_info **, int, int, uint_t,
346 			sfmmu_t *);
347 static void	sfmmu_tsb_free(struct tsb_info *);
348 static void	sfmmu_tsbinfo_free(struct tsb_info *);
349 static int	sfmmu_init_tsbinfo(struct tsb_info *, int, int, uint_t,
350 			sfmmu_t *);
351 
352 static void	sfmmu_tsb_swapin(sfmmu_t *, hatlock_t *);
353 static int	sfmmu_select_tsb_szc(pgcnt_t);
354 static void	sfmmu_mod_tsb(sfmmu_t *, caddr_t, tte_t *, int);
355 #define		sfmmu_load_tsb(sfmmup, vaddr, tte, szc) \
356 	sfmmu_mod_tsb(sfmmup, vaddr, tte, szc)
357 #define		sfmmu_unload_tsb(sfmmup, vaddr, szc)    \
358 	sfmmu_mod_tsb(sfmmup, vaddr, NULL, szc)
359 static void	sfmmu_copy_tsb(struct tsb_info *, struct tsb_info *);
360 static tsb_replace_rc_t sfmmu_replace_tsb(sfmmu_t *, struct tsb_info *, uint_t,
361     hatlock_t *, uint_t);
362 static void	sfmmu_size_tsb(sfmmu_t *, int, uint64_t, uint64_t);
363 
364 static void	sfmmu_cache_flush(pfn_t, int);
365 void		sfmmu_cache_flushcolor(int, pfn_t);
366 static caddr_t	sfmmu_hblk_chgattr(sfmmu_t *, struct hme_blk *, caddr_t,
367 			caddr_t, demap_range_t *, uint_t, int);
368 
369 static uint64_t	sfmmu_vtop_attr(uint_t, int mode, tte_t *);
370 static uint_t	sfmmu_ptov_attr(tte_t *);
371 static caddr_t	sfmmu_hblk_chgprot(sfmmu_t *, struct hme_blk *, caddr_t,
372 			caddr_t, demap_range_t *, uint_t);
373 static uint_t	sfmmu_vtop_prot(uint_t, uint_t *);
374 static int	sfmmu_idcache_constructor(void *, void *, int);
375 static void	sfmmu_idcache_destructor(void *, void *);
376 static int	sfmmu_hblkcache_constructor(void *, void *, int);
377 static void	sfmmu_hblkcache_destructor(void *, void *);
378 static void	sfmmu_hblkcache_reclaim(void *);
379 static void	sfmmu_shadow_hcleanup(sfmmu_t *, struct hme_blk *,
380 			struct hmehash_bucket *);
381 static void	sfmmu_free_hblks(sfmmu_t *, caddr_t, caddr_t, int);
382 
383 static void	sfmmu_reuse_ctx(struct ctx *, sfmmu_t *);
384 static void	sfmmu_disallow_ctx_steal(sfmmu_t *);
385 static void	sfmmu_allow_ctx_steal(sfmmu_t *);
386 
387 static void	sfmmu_rm_large_mappings(page_t *, int);
388 
389 static void	hat_lock_init(void);
390 static void	hat_kstat_init(void);
391 static int	sfmmu_kstat_percpu_update(kstat_t *ksp, int rw);
392 static void	sfmmu_check_page_sizes(sfmmu_t *, int);
393 static int	fnd_mapping_sz(page_t *);
394 static void	iment_add(struct ism_ment *,  struct hat *);
395 static void	iment_sub(struct ism_ment *, struct hat *);
396 static pgcnt_t	ism_tsb_entries(sfmmu_t *, int szc);
397 extern void	sfmmu_setup_tsbinfo(sfmmu_t *);
398 extern void	sfmmu_clear_utsbinfo(void);
399 
400 /* kpm prototypes */
401 static caddr_t	sfmmu_kpm_mapin(page_t *);
402 static void	sfmmu_kpm_mapout(page_t *, caddr_t);
403 static int	sfmmu_kpme_lookup(struct kpme *, page_t *);
404 static void	sfmmu_kpme_add(struct kpme *, page_t *);
405 static void	sfmmu_kpme_sub(struct kpme *, page_t *);
406 static caddr_t	sfmmu_kpm_getvaddr(page_t *, int *);
407 static int	sfmmu_kpm_fault(caddr_t, struct memseg *, page_t *);
408 static int	sfmmu_kpm_fault_small(caddr_t, struct memseg *, page_t *);
409 static void	sfmmu_kpm_vac_conflict(page_t *, caddr_t);
410 static void	sfmmu_kpm_pageunload(page_t *);
411 static void	sfmmu_kpm_vac_unload(page_t *, caddr_t);
412 static void	sfmmu_kpm_demap_large(caddr_t);
413 static void	sfmmu_kpm_demap_small(caddr_t);
414 static void	sfmmu_kpm_demap_tlbs(caddr_t, int);
415 static void	sfmmu_kpm_hme_unload(page_t *);
416 static kpm_hlk_t *sfmmu_kpm_kpmp_enter(page_t *, pgcnt_t);
417 static void	sfmmu_kpm_kpmp_exit(kpm_hlk_t *kpmp);
418 static void	sfmmu_kpm_page_cache(page_t *, int, int);
419 
420 /* kpm globals */
421 #ifdef	DEBUG
422 /*
423  * Enable trap level tsbmiss handling
424  */
425 int	kpm_tsbmtl = 1;
426 
427 /*
428  * Flush the TLB on kpm mapout. Note: Xcalls are used (again) for the
429  * required TLB shootdowns in this case, so handle w/ care. Off by default.
430  */
431 int	kpm_tlb_flush;
432 #endif	/* DEBUG */
433 
434 static void	*sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *, size_t, int);
435 
436 #ifdef DEBUG
437 static void	sfmmu_check_hblk_flist();
438 #endif
439 
440 /*
441  * Semi-private sfmmu data structures.  Some of them are initialize in
442  * startup or in hat_init. Some of them are private but accessed by
443  * assembly code or mach_sfmmu.c
444  */
445 struct hmehash_bucket *uhme_hash;	/* user hmeblk hash table */
446 struct hmehash_bucket *khme_hash;	/* kernel hmeblk hash table */
447 uint64_t	uhme_hash_pa;		/* PA of uhme_hash */
448 uint64_t	khme_hash_pa;		/* PA of khme_hash */
449 int 		uhmehash_num;		/* # of buckets in user hash table */
450 int 		khmehash_num;		/* # of buckets in kernel hash table */
451 struct ctx	*ctxs;			/* used by <machine/mmu.c> */
452 uint_t		nctxs;			/* total number of contexts */
453 
454 int		cache;			/* describes system cache */
455 
456 caddr_t		ktsb_base;		/* kernel 8k-indexed tsb base address */
457 uint64_t	ktsb_pbase;		/* kernel 8k-indexed tsb phys address */
458 int		ktsb_szcode;		/* kernel 8k-indexed tsb size code */
459 int		ktsb_sz;		/* kernel 8k-indexed tsb size */
460 
461 caddr_t		ktsb4m_base;		/* kernel 4m-indexed tsb base address */
462 uint64_t	ktsb4m_pbase;		/* kernel 4m-indexed tsb phys address */
463 int		ktsb4m_szcode;		/* kernel 4m-indexed tsb size code */
464 int		ktsb4m_sz;		/* kernel 4m-indexed tsb size */
465 
466 uint64_t	kpm_tsbbase;		/* kernel seg_kpm 4M TSB base address */
467 int		kpm_tsbsz;		/* kernel seg_kpm 4M TSB size code */
468 uint64_t	kpmsm_tsbbase;		/* kernel seg_kpm 8K TSB base address */
469 int		kpmsm_tsbsz;		/* kernel seg_kpm 8K TSB size code */
470 
471 #ifndef sun4v
472 int		utsb_dtlb_ttenum = -1;	/* index in TLB for utsb locked TTE */
473 int		utsb4m_dtlb_ttenum = -1; /* index in TLB for 4M TSB TTE */
474 int		dtlb_resv_ttenum;	/* index in TLB of first reserved TTE */
475 caddr_t		utsb_vabase;		/* reserved kernel virtual memory */
476 caddr_t		utsb4m_vabase;		/* for trap handler TSB accesses */
477 #endif /* sun4v */
478 uint64_t	tsb_alloc_bytes = 0;	/* bytes allocated to TSBs */
479 vmem_t		*kmem_tsb_default_arena[NLGRPS_MAX];	/* For dynamic TSBs */
480 
481 /*
482  * Size to use for TSB slabs.  Future platforms that support page sizes
483  * larger than 4M may wish to change these values, and provide their own
484  * assembly macros for building and decoding the TSB base register contents.
485  */
486 uint_t	tsb_slab_size = MMU_PAGESIZE4M;
487 uint_t	tsb_slab_shift = MMU_PAGESHIFT4M;
488 uint_t	tsb_slab_ttesz = TTE4M;
489 uint_t	tsb_slab_mask = 0x1ff;	/* 4M page alignment for 8K pfn */
490 
491 /* largest TSB size to grow to, will be smaller on smaller memory systems */
492 int	tsb_max_growsize = UTSB_MAX_SZCODE;
493 
494 /*
495  * Tunable parameters dealing with TSB policies.
496  */
497 
498 /*
499  * This undocumented tunable forces all 8K TSBs to be allocated from
500  * the kernel heap rather than from the kmem_tsb_default_arena arenas.
501  */
502 #ifdef	DEBUG
503 int	tsb_forceheap = 0;
504 #endif	/* DEBUG */
505 
506 /*
507  * Decide whether to use per-lgroup arenas, or one global set of
508  * TSB arenas.  The default is not to break up per-lgroup, since
509  * most platforms don't recognize any tangible benefit from it.
510  */
511 int	tsb_lgrp_affinity = 0;
512 
513 /*
514  * Used for growing the TSB based on the process RSS.
515  * tsb_rss_factor is based on the smallest TSB, and is
516  * shifted by the TSB size to determine if we need to grow.
517  * The default will grow the TSB if the number of TTEs for
518  * this page size exceeds 75% of the number of TSB entries,
519  * which should _almost_ eliminate all conflict misses
520  * (at the expense of using up lots and lots of memory).
521  */
522 #define	TSB_RSS_FACTOR		(TSB_ENTRIES(TSB_MIN_SZCODE) * 0.75)
523 #define	SFMMU_RSS_TSBSIZE(tsbszc)	(tsb_rss_factor << tsbszc)
524 #define	SELECT_TSB_SIZECODE(pgcnt) ( \
525 	(enable_tsb_rss_sizing)? sfmmu_select_tsb_szc(pgcnt) : \
526 	default_tsb_size)
527 #define	TSB_OK_SHRINK()	\
528 	(tsb_alloc_bytes > tsb_alloc_hiwater || freemem < desfree)
529 #define	TSB_OK_GROW()	\
530 	(tsb_alloc_bytes < tsb_alloc_hiwater && freemem > desfree)
531 
532 int	enable_tsb_rss_sizing = 1;
533 int	tsb_rss_factor	= (int)TSB_RSS_FACTOR;
534 
535 /* which TSB size code to use for new address spaces or if rss sizing off */
536 int default_tsb_size = TSB_8K_SZCODE;
537 
538 static uint64_t tsb_alloc_hiwater; /* limit TSB reserved memory */
539 uint64_t tsb_alloc_hiwater_factor; /* tsb_alloc_hiwater = physmem / this */
540 #define	TSB_ALLOC_HIWATER_FACTOR_DEFAULT	32
541 
542 #ifdef DEBUG
543 static int tsb_random_size = 0;	/* set to 1 to test random tsb sizes on alloc */
544 static int tsb_grow_stress = 0;	/* if set to 1, keep replacing TSB w/ random */
545 static int tsb_alloc_mtbf = 0;	/* fail allocation every n attempts */
546 static int tsb_alloc_fail_mtbf = 0;
547 static int tsb_alloc_count = 0;
548 #endif /* DEBUG */
549 
550 /* if set to 1, will remap valid TTEs when growing TSB. */
551 int tsb_remap_ttes = 1;
552 
553 /* if we have more than this many mappings, allocate a second TSB. */
554 int tsb_sectsb_threshold = 1;
555 
556 /*
557  * kstat data
558  */
559 struct sfmmu_global_stat sfmmu_global_stat;
560 struct sfmmu_tsbsize_stat sfmmu_tsbsize_stat;
561 
562 /*
563  * Global data
564  */
565 sfmmu_t 	*ksfmmup;		/* kernel's hat id */
566 struct ctx 	*kctx;			/* kernel's context */
567 
568 #ifdef DEBUG
569 static void	chk_tte(tte_t *, tte_t *, tte_t *, struct hme_blk *);
570 #endif
571 
572 /* sfmmu locking operations */
573 static kmutex_t *sfmmu_mlspl_enter(struct page *, int);
574 static int	sfmmu_mlspl_held(struct page *, int);
575 
576 static kmutex_t *sfmmu_page_enter(page_t *);
577 static void	sfmmu_page_exit(kmutex_t *);
578 static int	sfmmu_page_spl_held(struct page *);
579 
580 /* sfmmu internal locking operations - accessed directly */
581 static void	sfmmu_mlist_reloc_enter(page_t *, page_t *,
582 				kmutex_t **, kmutex_t **);
583 static void	sfmmu_mlist_reloc_exit(kmutex_t *, kmutex_t *);
584 static hatlock_t *
585 		sfmmu_hat_enter(sfmmu_t *);
586 static hatlock_t *
587 		sfmmu_hat_tryenter(sfmmu_t *);
588 static void	sfmmu_hat_exit(hatlock_t *);
589 static void	sfmmu_hat_lock_all(void);
590 static void	sfmmu_hat_unlock_all(void);
591 static void	sfmmu_ismhat_enter(sfmmu_t *, int);
592 static void	sfmmu_ismhat_exit(sfmmu_t *, int);
593 
594 /*
595  * Array of mutexes protecting a page's mapping list and p_nrm field.
596  *
597  * The hash function looks complicated, but is made up so that:
598  *
599  * "pp" not shifted, so adjacent pp values will hash to different cache lines
600  *  (8 byte alignment * 8 bytes/mutes == 64 byte coherency subblock)
601  *
602  * "pp" >> mml_shift, incorporates more source bits into the hash result
603  *
604  *  "& (mml_table_size - 1), should be faster than using remainder "%"
605  *
606  * Hopefully, mml_table, mml_table_size and mml_shift are all in the same
607  * cacheline, since they get declared next to each other below. We'll trust
608  * ld not to do something random.
609  */
610 #ifdef	DEBUG
611 int mlist_hash_debug = 0;
612 #define	MLIST_HASH(pp)	(mlist_hash_debug ? &mml_table[0] : \
613 	&mml_table[((uintptr_t)(pp) + \
614 	((uintptr_t)(pp) >> mml_shift)) & (mml_table_sz - 1)])
615 #else	/* !DEBUG */
616 #define	MLIST_HASH(pp)   &mml_table[ \
617 	((uintptr_t)(pp) + ((uintptr_t)(pp) >> mml_shift)) & (mml_table_sz - 1)]
618 #endif	/* !DEBUG */
619 
620 kmutex_t		*mml_table;
621 uint_t			mml_table_sz;	/* must be a power of 2 */
622 uint_t			mml_shift;	/* log2(mml_table_sz) + 3 for align */
623 
624 /*
625  * kpm_page lock hash.
626  * All slots should be used equally and 2 adjacent kpm_page_t's
627  * shouldn't have their mutexes in the same cache line.
628  */
629 #ifdef	DEBUG
630 int kpmp_hash_debug = 0;
631 #define	KPMP_HASH(kpp)	(kpmp_hash_debug ? &kpmp_table[0] : &kpmp_table[ \
632 	((uintptr_t)(kpp) + ((uintptr_t)(kpp) >> kpmp_shift)) \
633 	& (kpmp_table_sz - 1)])
634 #else	/* !DEBUG */
635 #define	KPMP_HASH(kpp)	&kpmp_table[ \
636 	((uintptr_t)(kpp) + ((uintptr_t)(kpp) >> kpmp_shift)) \
637 	& (kpmp_table_sz - 1)]
638 #endif	/* DEBUG */
639 
640 kpm_hlk_t	*kpmp_table;
641 uint_t		kpmp_table_sz;	/* must be a power of 2 */
642 uchar_t		kpmp_shift;
643 
644 #ifdef	DEBUG
645 #define	KPMP_SHASH(kpp)	(kpmp_hash_debug ? &kpmp_stable[0] : &kpmp_stable[ \
646 	(((uintptr_t)(kpp) << kpmp_shift) + (uintptr_t)(kpp)) \
647 	& (kpmp_stable_sz - 1)])
648 #else	/* !DEBUG */
649 #define	KPMP_SHASH(kpp)	&kpmp_stable[ \
650 	(((uintptr_t)(kpp) << kpmp_shift) + (uintptr_t)(kpp)) \
651 	& (kpmp_stable_sz - 1)]
652 #endif	/* DEBUG */
653 
654 kpm_shlk_t	*kpmp_stable;
655 uint_t		kpmp_stable_sz;	/* must be a power of 2 */
656 
657 /*
658  * SPL_HASH was improved to avoid false cache line sharing
659  */
660 #define	SPL_TABLE_SIZE	128
661 #define	SPL_MASK	(SPL_TABLE_SIZE - 1)
662 #define	SPL_SHIFT	7		/* log2(SPL_TABLE_SIZE) */
663 
664 #define	SPL_INDEX(pp) \
665 	((((uintptr_t)(pp) >> SPL_SHIFT) ^ \
666 	((uintptr_t)(pp) >> (SPL_SHIFT << 1))) & \
667 	(SPL_TABLE_SIZE - 1))
668 
669 #define	SPL_HASH(pp)    \
670 	(&sfmmu_page_lock[SPL_INDEX(pp) & SPL_MASK].pad_mutex)
671 
672 static	pad_mutex_t	sfmmu_page_lock[SPL_TABLE_SIZE];
673 
674 
675 /*
676  * hat_unload_callback() will group together callbacks in order
677  * to avoid xt_sync() calls.  This is the maximum size of the group.
678  */
679 #define	MAX_CB_ADDR	32
680 
681 #ifdef DEBUG
682 
683 /*
684  * Debugging trace ring buffer for stolen and freed ctxs.  The
685  * stolen_ctxs[] array is protected by the ctx_trace_mutex.
686  */
687 struct ctx_trace stolen_ctxs[TRSIZE];
688 struct ctx_trace *ctx_trace_first = &stolen_ctxs[0];
689 struct ctx_trace *ctx_trace_last = &stolen_ctxs[TRSIZE-1];
690 struct ctx_trace *ctx_trace_ptr = &stolen_ctxs[0];
691 kmutex_t ctx_trace_mutex;
692 uint_t	num_ctx_stolen = 0;
693 
694 int	ism_debug = 0;
695 
696 #endif /* DEBUG */
697 
698 tte_t	hw_tte;
699 static ulong_t sfmmu_dmr_maxbit = DMR_MAXBIT;
700 
701 /*
702  * kpm virtual address to physical address
703  */
704 #define	SFMMU_KPM_VTOP(vaddr, paddr) {					\
705 	uintptr_t r, v;							\
706 									\
707 	r = ((vaddr) - kpm_vbase) >> (uintptr_t)kpm_size_shift;		\
708 	(paddr) = (vaddr) - kpm_vbase;					\
709 	if (r != 0) {							\
710 		v = ((uintptr_t)(vaddr) >> MMU_PAGESHIFT) &		\
711 		    vac_colors_mask;					\
712 		(paddr) -= r << kpm_size_shift;				\
713 		if (r > v)						\
714 			(paddr) += (r - v) << MMU_PAGESHIFT;		\
715 		else							\
716 			(paddr) -= r << MMU_PAGESHIFT;			\
717 	}								\
718 }
719 
720 /*
721  * Wrapper for vmem_xalloc since vmem_create only allows limited
722  * parameters for vm_source_alloc functions.  This function allows us
723  * to specify alignment consistent with the size of the object being
724  * allocated.
725  */
726 static void *
727 sfmmu_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag)
728 {
729 	return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag));
730 }
731 
732 /* Common code for setting tsb_alloc_hiwater. */
733 #define	SFMMU_SET_TSB_ALLOC_HIWATER(pages)	tsb_alloc_hiwater = \
734 		ptob(pages) / tsb_alloc_hiwater_factor
735 
736 /*
737  * Set tsb_max_growsize to allow at most all of physical memory to be mapped by
738  * a single TSB.  physmem is the number of physical pages so we need physmem 8K
739  * TTEs to represent all those physical pages.  We round this up by using
740  * 1<<highbit().  To figure out which size code to use, remember that the size
741  * code is just an amount to shift the smallest TSB size to get the size of
742  * this TSB.  So we subtract that size, TSB_START_SIZE, from highbit() (or
743  * highbit() - 1) to get the size code for the smallest TSB that can represent
744  * all of physical memory, while erring on the side of too much.
745  *
746  * If the computed size code is less than the current tsb_max_growsize, we set
747  * tsb_max_growsize to the computed size code.  In the case where the computed
748  * size code is greater than tsb_max_growsize, we have these restrictions that
749  * apply to increasing tsb_max_growsize:
750  *	1) TSBs can't grow larger than the TSB slab size
751  *	2) TSBs can't grow larger than UTSB_MAX_SZCODE.
752  */
753 #define	SFMMU_SET_TSB_MAX_GROWSIZE(pages) {				\
754 	int	i, szc;							\
755 									\
756 	i = highbit(pages);						\
757 	if ((1 << (i - 1)) == (pages))					\
758 		i--;		/* 2^n case, round down */		\
759 	szc = i - TSB_START_SIZE;					\
760 	if (szc < tsb_max_growsize)					\
761 		tsb_max_growsize = szc;					\
762 	else if ((szc > tsb_max_growsize) &&				\
763 	    (szc <= tsb_slab_shift - (TSB_START_SIZE + TSB_ENTRY_SHIFT))) \
764 		tsb_max_growsize = MIN(szc, UTSB_MAX_SZCODE);		\
765 }
766 
767 /*
768  * Given a pointer to an sfmmu and a TTE size code, return a pointer to the
769  * tsb_info which handles that TTE size.
770  */
771 #define	SFMMU_GET_TSBINFO(tsbinfop, sfmmup, tte_szc)			\
772 	(tsbinfop) = (sfmmup)->sfmmu_tsb;				\
773 	ASSERT(sfmmu_hat_lock_held(sfmmup));				\
774 	if ((tte_szc) >= TTE4M)						\
775 		(tsbinfop) = (tsbinfop)->tsb_next;
776 
777 /*
778  * Return the number of mappings present in the HAT
779  * for a particular process and page size.
780  */
781 #define	SFMMU_TTE_CNT(sfmmup, szc)					\
782 	(sfmmup)->sfmmu_iblk?						\
783 	    (sfmmup)->sfmmu_ismttecnt[(szc)] +				\
784 	    (sfmmup)->sfmmu_ttecnt[(szc)] :				\
785 	    (sfmmup)->sfmmu_ttecnt[(szc)];
786 
787 /*
788  * Macro to use to unload entries from the TSB.
789  * It has knowledge of which page sizes get replicated in the TSB
790  * and will call the appropriate unload routine for the appropriate size.
791  */
792 #define	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp)				\
793 {									\
794 	int ttesz = get_hblk_ttesz(hmeblkp);				\
795 	if (ttesz == TTE8K || ttesz == TTE4M) {				\
796 		sfmmu_unload_tsb(sfmmup, addr, ttesz);			\
797 	} else {							\
798 		caddr_t sva = (caddr_t)get_hblk_base(hmeblkp);		\
799 		caddr_t eva = sva + get_hblk_span(hmeblkp);		\
800 		ASSERT(addr >= sva && addr < eva);			\
801 		sfmmu_unload_tsb_range(sfmmup, sva, eva, ttesz);	\
802 	}								\
803 }
804 
805 
806 /* Update tsb_alloc_hiwater after memory is configured. */
807 /*ARGSUSED*/
808 static void
809 sfmmu_update_tsb_post_add(void *arg, pgcnt_t delta_pages)
810 {
811 	/* Assumes physmem has already been updated. */
812 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
813 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
814 }
815 
816 /*
817  * Update tsb_alloc_hiwater before memory is deleted.  We'll do nothing here
818  * and update tsb_alloc_hiwater and tsb_max_growsize after the memory is
819  * deleted.
820  */
821 /*ARGSUSED*/
822 static int
823 sfmmu_update_tsb_pre_del(void *arg, pgcnt_t delta_pages)
824 {
825 	return (0);
826 }
827 
828 /* Update tsb_alloc_hiwater after memory fails to be unconfigured. */
829 /*ARGSUSED*/
830 static void
831 sfmmu_update_tsb_post_del(void *arg, pgcnt_t delta_pages, int cancelled)
832 {
833 	/*
834 	 * Whether the delete was cancelled or not, just go ahead and update
835 	 * tsb_alloc_hiwater and tsb_max_growsize.
836 	 */
837 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
838 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
839 }
840 
841 static kphysm_setup_vector_t sfmmu_update_tsb_vec = {
842 	KPHYSM_SETUP_VECTOR_VERSION,	/* version */
843 	sfmmu_update_tsb_post_add,	/* post_add */
844 	sfmmu_update_tsb_pre_del,	/* pre_del */
845 	sfmmu_update_tsb_post_del	/* post_del */
846 };
847 
848 
849 /*
850  * HME_BLK HASH PRIMITIVES
851  */
852 
853 /*
854  * Enter a hme on the mapping list for page pp.
855  * When large pages are more prevalent in the system we might want to
856  * keep the mapping list in ascending order by the hment size. For now,
857  * small pages are more frequent, so don't slow it down.
858  */
859 #define	HME_ADD(hme, pp)					\
860 {								\
861 	ASSERT(sfmmu_mlist_held(pp));				\
862 								\
863 	hme->hme_prev = NULL;					\
864 	hme->hme_next = pp->p_mapping;				\
865 	hme->hme_page = pp;					\
866 	if (pp->p_mapping) {					\
867 		((struct sf_hment *)(pp->p_mapping))->hme_prev = hme;\
868 		ASSERT(pp->p_share > 0);			\
869 	} else  {						\
870 		/* EMPTY */					\
871 		ASSERT(pp->p_share == 0);			\
872 	}							\
873 	pp->p_mapping = hme;					\
874 	pp->p_share++;						\
875 }
876 
877 /*
878  * Enter a hme on the mapping list for page pp.
879  * If we are unmapping a large translation, we need to make sure that the
880  * change is reflect in the corresponding bit of the p_index field.
881  */
882 #define	HME_SUB(hme, pp)					\
883 {								\
884 	ASSERT(sfmmu_mlist_held(pp));				\
885 	ASSERT(hme->hme_page == pp || IS_PAHME(hme));		\
886 								\
887 	if (pp->p_mapping == NULL) {				\
888 		panic("hme_remove - no mappings");		\
889 	}							\
890 								\
891 	membar_stst();	/* ensure previous stores finish */	\
892 								\
893 	ASSERT(pp->p_share > 0);				\
894 	pp->p_share--;						\
895 								\
896 	if (hme->hme_prev) {					\
897 		ASSERT(pp->p_mapping != hme);			\
898 		ASSERT(hme->hme_prev->hme_page == pp ||		\
899 			IS_PAHME(hme->hme_prev));		\
900 		hme->hme_prev->hme_next = hme->hme_next;	\
901 	} else {						\
902 		ASSERT(pp->p_mapping == hme);			\
903 		pp->p_mapping = hme->hme_next;			\
904 		ASSERT((pp->p_mapping == NULL) ?		\
905 			(pp->p_share == 0) : 1);		\
906 	}							\
907 								\
908 	if (hme->hme_next) {					\
909 		ASSERT(hme->hme_next->hme_page == pp ||		\
910 			IS_PAHME(hme->hme_next));		\
911 		hme->hme_next->hme_prev = hme->hme_prev;	\
912 	}							\
913 								\
914 	/* zero out the entry */				\
915 	hme->hme_next = NULL;					\
916 	hme->hme_prev = NULL;					\
917 	hme->hme_page = NULL;					\
918 								\
919 	if (hme_size(hme) > TTE8K) {				\
920 		/* remove mappings for remainder of large pg */	\
921 		sfmmu_rm_large_mappings(pp, hme_size(hme));	\
922 	}							\
923 }
924 
925 /*
926  * This function returns the hment given the hme_blk and a vaddr.
927  * It assumes addr has already been checked to belong to hme_blk's
928  * range.
929  */
930 #define	HBLKTOHME(hment, hmeblkp, addr)					\
931 {									\
932 	int index;							\
933 	HBLKTOHME_IDX(hment, hmeblkp, addr, index)			\
934 }
935 
936 /*
937  * Version of HBLKTOHME that also returns the index in hmeblkp
938  * of the hment.
939  */
940 #define	HBLKTOHME_IDX(hment, hmeblkp, addr, idx)			\
941 {									\
942 	ASSERT(in_hblk_range((hmeblkp), (addr)));			\
943 									\
944 	if (get_hblk_ttesz(hmeblkp) == TTE8K) {				\
945 		idx = (((uintptr_t)(addr) >> MMU_PAGESHIFT) & (NHMENTS-1)); \
946 	} else								\
947 		idx = 0;						\
948 									\
949 	(hment) = &(hmeblkp)->hblk_hme[idx];				\
950 }
951 
952 /*
953  * Disable any page sizes not supported by the CPU
954  */
955 void
956 hat_init_pagesizes()
957 {
958 	int 		i;
959 
960 	mmu_exported_page_sizes = 0;
961 	for (i = TTE8K; i < max_mmu_page_sizes; i++) {
962 		extern int	disable_text_largepages;
963 		extern int	disable_initdata_largepages;
964 
965 		szc_2_userszc[i] = (uint_t)-1;
966 		userszc_2_szc[i] = (uint_t)-1;
967 
968 		if ((mmu_exported_pagesize_mask & (1 << i)) == 0) {
969 			disable_large_pages |= (1 << i);
970 			disable_ism_large_pages |= (1 << i);
971 			disable_text_largepages |= (1 << i);
972 			disable_initdata_largepages |= (1 << i);
973 		} else {
974 			szc_2_userszc[i] = mmu_exported_page_sizes;
975 			userszc_2_szc[mmu_exported_page_sizes] = i;
976 			mmu_exported_page_sizes++;
977 		}
978 	}
979 
980 	disable_auto_large_pages = disable_large_pages;
981 
982 	/*
983 	 * Initialize mmu-specific large page sizes.
984 	 */
985 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
986 	    (&mmu_large_pages_disabled)) {
987 		disable_large_pages |= mmu_large_pages_disabled(HAT_LOAD);
988 		disable_ism_large_pages |=
989 		    mmu_large_pages_disabled(HAT_LOAD_SHARE);
990 		disable_auto_large_pages |=
991 		    mmu_large_pages_disabled(HAT_LOAD_AUTOLPG);
992 	}
993 
994 }
995 
996 /*
997  * Initialize the hardware address translation structures.
998  */
999 void
1000 hat_init(void)
1001 {
1002 	struct ctx	*ctx;
1003 	struct ctx	*cur_ctx = NULL;
1004 	int 		i;
1005 
1006 	hat_lock_init();
1007 	hat_kstat_init();
1008 
1009 	/*
1010 	 * Hardware-only bits in a TTE
1011 	 */
1012 	MAKE_TTE_MASK(&hw_tte);
1013 
1014 	hat_init_pagesizes();
1015 
1016 	/* Initialize the hash locks */
1017 	for (i = 0; i < khmehash_num; i++) {
1018 		mutex_init(&khme_hash[i].hmehash_mutex, NULL,
1019 		    MUTEX_DEFAULT, NULL);
1020 	}
1021 	for (i = 0; i < uhmehash_num; i++) {
1022 		mutex_init(&uhme_hash[i].hmehash_mutex, NULL,
1023 		    MUTEX_DEFAULT, NULL);
1024 	}
1025 	khmehash_num--;		/* make sure counter starts from 0 */
1026 	uhmehash_num--;		/* make sure counter starts from 0 */
1027 
1028 	/*
1029 	 * Initialize ctx structures and list lock.
1030 	 * We keep two lists of ctxs. The "free" list contains contexts
1031 	 * ready to use.  The "dirty" list contains contexts that are OK
1032 	 * to use after flushing the TLBs of any stale mappings.
1033 	 */
1034 	mutex_init(&ctx_list_lock, NULL, MUTEX_DEFAULT, NULL);
1035 	kctx = &ctxs[KCONTEXT];
1036 	ctx = &ctxs[NUM_LOCKED_CTXS];
1037 	ctxhand = ctxfree = ctx;		/* head of free list */
1038 	ctxdirty = NULL;
1039 	for (i = NUM_LOCKED_CTXS; i < nctxs; i++) {
1040 		cur_ctx = &ctxs[i];
1041 		cur_ctx->ctx_flags = CTX_FREE_FLAG;
1042 		cur_ctx->ctx_free = &ctxs[i + 1];
1043 	}
1044 	cur_ctx->ctx_free = NULL;		/* tail of free list */
1045 
1046 	/*
1047 	 * Intialize ism mapping list lock.
1048 	 */
1049 	mutex_init(&ism_mlist_lock, NULL, MUTEX_DEFAULT, NULL);
1050 
1051 	sfmmuid_cache = kmem_cache_create("sfmmuid_cache", sizeof (sfmmu_t),
1052 	    0, sfmmu_idcache_constructor, sfmmu_idcache_destructor,
1053 	    NULL, NULL, NULL, 0);
1054 
1055 	sfmmu_tsbinfo_cache = kmem_cache_create("sfmmu_tsbinfo_cache",
1056 	    sizeof (struct tsb_info), 0, NULL, NULL, NULL, NULL, NULL, 0);
1057 
1058 	/*
1059 	 * Since we only use the tsb8k cache to "borrow" pages for TSBs
1060 	 * from the heap when low on memory or when TSB_FORCEALLOC is
1061 	 * specified, don't use magazines to cache them--we want to return
1062 	 * them to the system as quickly as possible.
1063 	 */
1064 	sfmmu_tsb8k_cache = kmem_cache_create("sfmmu_tsb8k_cache",
1065 	    MMU_PAGESIZE, MMU_PAGESIZE, NULL, NULL, NULL, NULL,
1066 	    static_arena, KMC_NOMAGAZINE);
1067 
1068 	/*
1069 	 * Set tsb_alloc_hiwater to 1/tsb_alloc_hiwater_factor of physical
1070 	 * memory, which corresponds to the old static reserve for TSBs.
1071 	 * tsb_alloc_hiwater_factor defaults to 32.  This caps the amount of
1072 	 * memory we'll allocate for TSB slabs; beyond this point TSB
1073 	 * allocations will be taken from the kernel heap (via
1074 	 * sfmmu_tsb8k_cache) and will be throttled as would any other kmem
1075 	 * consumer.
1076 	 */
1077 	if (tsb_alloc_hiwater_factor == 0) {
1078 		tsb_alloc_hiwater_factor = TSB_ALLOC_HIWATER_FACTOR_DEFAULT;
1079 	}
1080 	SFMMU_SET_TSB_ALLOC_HIWATER(physmem);
1081 
1082 	/* Set tsb_max_growsize. */
1083 	SFMMU_SET_TSB_MAX_GROWSIZE(physmem);
1084 
1085 	/*
1086 	 * On smaller memory systems, allocate TSB memory in 512K chunks
1087 	 * instead of the default 4M slab size.  The trap handlers need to
1088 	 * be patched with the final slab shift since they need to be able
1089 	 * to construct the TSB pointer at runtime.
1090 	 */
1091 	if ((tsb_max_growsize <= TSB_512K_SZCODE) &&
1092 	    !(disable_large_pages & (1 << TTE512K))) {
1093 		tsb_slab_size = MMU_PAGESIZE512K;
1094 		tsb_slab_shift = MMU_PAGESHIFT512K;
1095 		tsb_slab_ttesz = TTE512K;
1096 		tsb_slab_mask = 0x3f;	/* 512K page alignment for 8K pfn */
1097 	}
1098 
1099 	/*
1100 	 * Set up memory callback to update tsb_alloc_hiwater and
1101 	 * tsb_max_growsize.
1102 	 */
1103 	i = kphysm_setup_func_register(&sfmmu_update_tsb_vec, (void *) 0);
1104 	ASSERT(i == 0);
1105 
1106 	/*
1107 	 * kmem_tsb_arena is the source from which large TSB slabs are
1108 	 * drawn.  The quantum of this arena corresponds to the largest
1109 	 * TSB size we can dynamically allocate for user processes.
1110 	 * Currently it must also be a supported page size since we
1111 	 * use exactly one translation entry to map each slab page.
1112 	 *
1113 	 * The per-lgroup kmem_tsb_default_arena arenas are the arenas from
1114 	 * which most TSBs are allocated.  Since most TSB allocations are
1115 	 * typically 8K we have a kmem cache we stack on top of each
1116 	 * kmem_tsb_default_arena to speed up those allocations.
1117 	 *
1118 	 * Note the two-level scheme of arenas is required only
1119 	 * because vmem_create doesn't allow us to specify alignment
1120 	 * requirements.  If this ever changes the code could be
1121 	 * simplified to use only one level of arenas.
1122 	 */
1123 	kmem_tsb_arena = vmem_create("kmem_tsb", NULL, 0, tsb_slab_size,
1124 	    sfmmu_vmem_xalloc_aligned_wrapper, vmem_xfree, heap_arena,
1125 	    0, VM_SLEEP);
1126 
1127 	if (tsb_lgrp_affinity) {
1128 		char s[50];
1129 		for (i = 0; i < NLGRPS_MAX; i++) {
1130 			(void) sprintf(s, "kmem_tsb_lgrp%d", i);
1131 			kmem_tsb_default_arena[i] =
1132 			    vmem_create(s, NULL, 0, PAGESIZE,
1133 			    sfmmu_tsb_segkmem_alloc, sfmmu_tsb_segkmem_free,
1134 			    kmem_tsb_arena, 0, VM_SLEEP | VM_BESTFIT);
1135 			(void) sprintf(s, "sfmmu_tsb_lgrp%d_cache", i);
1136 			sfmmu_tsb_cache[i] = kmem_cache_create(s, PAGESIZE,
1137 			    PAGESIZE, NULL, NULL, NULL, NULL,
1138 			    kmem_tsb_default_arena[i], 0);
1139 		}
1140 	} else {
1141 		kmem_tsb_default_arena[0] = vmem_create("kmem_tsb_default",
1142 		    NULL, 0, PAGESIZE, sfmmu_tsb_segkmem_alloc,
1143 		    sfmmu_tsb_segkmem_free, kmem_tsb_arena, 0,
1144 		    VM_SLEEP | VM_BESTFIT);
1145 
1146 		sfmmu_tsb_cache[0] = kmem_cache_create("sfmmu_tsb_cache",
1147 		    PAGESIZE, PAGESIZE, NULL, NULL, NULL, NULL,
1148 		    kmem_tsb_default_arena[0], 0);
1149 	}
1150 
1151 	sfmmu8_cache = kmem_cache_create("sfmmu8_cache", HME8BLK_SZ,
1152 		HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1153 		sfmmu_hblkcache_destructor,
1154 		sfmmu_hblkcache_reclaim, (void *)HME8BLK_SZ,
1155 		hat_memload_arena, KMC_NOHASH);
1156 
1157 	hat_memload1_arena = vmem_create("hat_memload1", NULL, 0, PAGESIZE,
1158 	    segkmem_alloc_permanent, segkmem_free, heap_arena, 0, VM_SLEEP);
1159 
1160 	sfmmu1_cache = kmem_cache_create("sfmmu1_cache", HME1BLK_SZ,
1161 		HMEBLK_ALIGN, sfmmu_hblkcache_constructor,
1162 		sfmmu_hblkcache_destructor,
1163 		NULL, (void *)HME1BLK_SZ,
1164 		hat_memload1_arena, KMC_NOHASH);
1165 
1166 	pa_hment_cache = kmem_cache_create("pa_hment_cache", PAHME_SZ,
1167 		0, NULL, NULL, NULL, NULL, static_arena, KMC_NOHASH);
1168 
1169 	ism_blk_cache = kmem_cache_create("ism_blk_cache",
1170 		sizeof (ism_blk_t), ecache_alignsize, NULL, NULL,
1171 		NULL, NULL, static_arena, KMC_NOHASH);
1172 
1173 	ism_ment_cache = kmem_cache_create("ism_ment_cache",
1174 		sizeof (ism_ment_t), 0, NULL, NULL,
1175 		NULL, NULL, NULL, 0);
1176 
1177 	/*
1178 	 * We grab the first hat for the kernel,
1179 	 */
1180 	AS_LOCK_ENTER(&kas, &kas.a_lock, RW_WRITER);
1181 	kas.a_hat = hat_alloc(&kas);
1182 	AS_LOCK_EXIT(&kas, &kas.a_lock);
1183 
1184 	/*
1185 	 * Initialize hblk_reserve.
1186 	 */
1187 	((struct hme_blk *)hblk_reserve)->hblk_nextpa =
1188 				va_to_pa((caddr_t)hblk_reserve);
1189 
1190 #ifndef sun4v
1191 	/*
1192 	 * Reserve some kernel virtual address space for the locked TTEs
1193 	 * that allow us to probe the TSB from TL>0.
1194 	 */
1195 	utsb_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1196 		0, 0, NULL, NULL, VM_SLEEP);
1197 	utsb4m_vabase = vmem_xalloc(heap_arena, tsb_slab_size, tsb_slab_size,
1198 		0, 0, NULL, NULL, VM_SLEEP);
1199 #endif
1200 
1201 	/*
1202 	 * The big page VAC handling code assumes VAC
1203 	 * will not be bigger than the smallest big
1204 	 * page- which is 64K.
1205 	 */
1206 	if (TTEPAGES(TTE64K) < CACHE_NUM_COLOR) {
1207 		cmn_err(CE_PANIC, "VAC too big!");
1208 	}
1209 
1210 	(void) xhat_init();
1211 
1212 	uhme_hash_pa = va_to_pa(uhme_hash);
1213 	khme_hash_pa = va_to_pa(khme_hash);
1214 
1215 	/*
1216 	 * Initialize relocation locks. kpr_suspendlock is held
1217 	 * at PIL_MAX to prevent interrupts from pinning the holder
1218 	 * of a suspended TTE which may access it leading to a
1219 	 * deadlock condition.
1220 	 */
1221 	mutex_init(&kpr_mutex, NULL, MUTEX_DEFAULT, NULL);
1222 	mutex_init(&kpr_suspendlock, NULL, MUTEX_SPIN, (void *)PIL_MAX);
1223 }
1224 
1225 /*
1226  * Initialize locking for the hat layer, called early during boot.
1227  */
1228 static void
1229 hat_lock_init()
1230 {
1231 	int i;
1232 	struct ctx *ctx;
1233 
1234 	/*
1235 	 * initialize the array of mutexes protecting a page's mapping
1236 	 * list and p_nrm field.
1237 	 */
1238 	for (i = 0; i < mml_table_sz; i++)
1239 		mutex_init(&mml_table[i], NULL, MUTEX_DEFAULT, NULL);
1240 
1241 	if (kpm_enable) {
1242 		for (i = 0; i < kpmp_table_sz; i++) {
1243 			mutex_init(&kpmp_table[i].khl_mutex, NULL,
1244 			    MUTEX_DEFAULT, NULL);
1245 		}
1246 	}
1247 
1248 	/*
1249 	 * Initialize array of mutex locks that protects sfmmu fields and
1250 	 * TSB lists.
1251 	 */
1252 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
1253 		mutex_init(HATLOCK_MUTEXP(&hat_lock[i]), NULL, MUTEX_DEFAULT,
1254 		    NULL);
1255 
1256 #ifdef	DEBUG
1257 	mutex_init(&ctx_trace_mutex, NULL, MUTEX_DEFAULT, NULL);
1258 #endif	/* DEBUG */
1259 
1260 	for (ctx = ctxs, i = 0; i < nctxs; i++, ctx++) {
1261 		rw_init(&ctx->ctx_rwlock, NULL, RW_DEFAULT, NULL);
1262 	}
1263 }
1264 
1265 extern caddr_t kmem64_base, kmem64_end;
1266 
1267 #define	SFMMU_KERNEL_MAXVA \
1268 	(kmem64_base ? (uintptr_t)kmem64_end : (SYSLIMIT))
1269 
1270 /*
1271  * Allocate a hat structure.
1272  * Called when an address space first uses a hat.
1273  */
1274 struct hat *
1275 hat_alloc(struct as *as)
1276 {
1277 	sfmmu_t *sfmmup;
1278 	struct ctx *ctx;
1279 	int i;
1280 	extern uint_t get_color_start(struct as *);
1281 
1282 	ASSERT(AS_WRITE_HELD(as, &as->a_lock));
1283 	sfmmup = kmem_cache_alloc(sfmmuid_cache, KM_SLEEP);
1284 	sfmmup->sfmmu_as = as;
1285 	sfmmup->sfmmu_flags = 0;
1286 
1287 	if (as == &kas) {
1288 		ctx = kctx;
1289 		ksfmmup = sfmmup;
1290 		sfmmup->sfmmu_cnum = ctxtoctxnum(ctx);
1291 		ASSERT(sfmmup->sfmmu_cnum == KCONTEXT);
1292 		sfmmup->sfmmu_cext = 0;
1293 		ctx->ctx_sfmmu = sfmmup;
1294 		ctx->ctx_flags = 0;
1295 		sfmmup->sfmmu_clrstart = 0;
1296 		sfmmup->sfmmu_tsb = NULL;
1297 		/*
1298 		 * hat_kern_setup() will call sfmmu_init_ktsbinfo()
1299 		 * to setup tsb_info for ksfmmup.
1300 		 */
1301 	} else {
1302 
1303 		/*
1304 		 * Just set to invalid ctx. When it faults, it will
1305 		 * get a valid ctx. This would avoid the situation
1306 		 * where we get a ctx, but it gets stolen and then
1307 		 * we fault when we try to run and so have to get
1308 		 * another ctx.
1309 		 */
1310 		sfmmup->sfmmu_cnum = INVALID_CONTEXT;
1311 		sfmmup->sfmmu_cext = 0;
1312 		/* initialize original physical page coloring bin */
1313 		sfmmup->sfmmu_clrstart = get_color_start(as);
1314 #ifdef DEBUG
1315 		if (tsb_random_size) {
1316 			uint32_t randval = (uint32_t)gettick() >> 4;
1317 			int size = randval % (tsb_max_growsize + 1);
1318 
1319 			/* chose a random tsb size for stress testing */
1320 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb, size,
1321 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1322 		} else
1323 #endif /* DEBUG */
1324 			(void) sfmmu_tsbinfo_alloc(&sfmmup->sfmmu_tsb,
1325 			    default_tsb_size,
1326 			    TSB8K|TSB64K|TSB512K, 0, sfmmup);
1327 		sfmmup->sfmmu_flags = HAT_SWAPPED;
1328 		ASSERT(sfmmup->sfmmu_tsb != NULL);
1329 	}
1330 	sfmmu_setup_tsbinfo(sfmmup);
1331 	for (i = 0; i < max_mmu_page_sizes; i++) {
1332 		sfmmup->sfmmu_ttecnt[i] = 0;
1333 		sfmmup->sfmmu_ismttecnt[i] = 0;
1334 		sfmmup->sfmmu_pgsz[i] = TTE8K;
1335 	}
1336 
1337 	sfmmup->sfmmu_iblk = NULL;
1338 	sfmmup->sfmmu_ismhat = 0;
1339 	sfmmup->sfmmu_ismblkpa = (uint64_t)-1;
1340 	if (sfmmup == ksfmmup) {
1341 		CPUSET_ALL(sfmmup->sfmmu_cpusran);
1342 	} else {
1343 		CPUSET_ZERO(sfmmup->sfmmu_cpusran);
1344 	}
1345 	sfmmup->sfmmu_free = 0;
1346 	sfmmup->sfmmu_rmstat = 0;
1347 	sfmmup->sfmmu_clrbin = sfmmup->sfmmu_clrstart;
1348 	sfmmup->sfmmu_xhat_provider = NULL;
1349 	cv_init(&sfmmup->sfmmu_tsb_cv, NULL, CV_DEFAULT, NULL);
1350 	return (sfmmup);
1351 }
1352 
1353 /*
1354  * Hat_setup, makes an address space context the current active one.
1355  * In sfmmu this translates to setting the secondary context with the
1356  * corresponding context.
1357  */
1358 void
1359 hat_setup(struct hat *sfmmup, int allocflag)
1360 {
1361 	struct ctx *ctx;
1362 	uint_t ctx_num;
1363 	hatlock_t *hatlockp;
1364 
1365 	/* Init needs some special treatment. */
1366 	if (allocflag == HAT_INIT) {
1367 		/*
1368 		 * Make sure that we have
1369 		 * 1. a TSB
1370 		 * 2. a valid ctx that doesn't get stolen after this point.
1371 		 */
1372 		hatlockp = sfmmu_hat_enter(sfmmup);
1373 
1374 		/*
1375 		 * Swap in the TSB.  hat_init() allocates tsbinfos without
1376 		 * TSBs, but we need one for init, since the kernel does some
1377 		 * special things to set up its stack and needs the TSB to
1378 		 * resolve page faults.
1379 		 */
1380 		sfmmu_tsb_swapin(sfmmup, hatlockp);
1381 
1382 		sfmmu_disallow_ctx_steal(sfmmup);
1383 
1384 		kpreempt_disable();
1385 
1386 		ctx = sfmmutoctx(sfmmup);
1387 		CPUSET_ADD(sfmmup->sfmmu_cpusran, CPU->cpu_id);
1388 		ctx_num = ctxtoctxnum(ctx);
1389 		ASSERT(sfmmup == ctx->ctx_sfmmu);
1390 		ASSERT(ctx_num >= NUM_LOCKED_CTXS);
1391 		sfmmu_setctx_sec(ctx_num);
1392 		sfmmu_load_mmustate(sfmmup);
1393 
1394 		kpreempt_enable();
1395 
1396 		/*
1397 		 * Allow ctx to be stolen.
1398 		 */
1399 		sfmmu_allow_ctx_steal(sfmmup);
1400 		sfmmu_hat_exit(hatlockp);
1401 	} else {
1402 		ASSERT(allocflag == HAT_ALLOC);
1403 
1404 		hatlockp = sfmmu_hat_enter(sfmmup);
1405 		kpreempt_disable();
1406 
1407 		CPUSET_ADD(sfmmup->sfmmu_cpusran, CPU->cpu_id);
1408 		sfmmu_setctx_sec(INVALID_CONTEXT);
1409 		sfmmu_clear_utsbinfo();
1410 
1411 		kpreempt_enable();
1412 		sfmmu_hat_exit(hatlockp);
1413 	}
1414 }
1415 
1416 /*
1417  * Free all the translation resources for the specified address space.
1418  * Called from as_free when an address space is being destroyed.
1419  */
1420 void
1421 hat_free_start(struct hat *sfmmup)
1422 {
1423 	ASSERT(AS_WRITE_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
1424 	ASSERT(sfmmup != ksfmmup);
1425 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1426 
1427 	sfmmup->sfmmu_free = 1;
1428 }
1429 
1430 void
1431 hat_free_end(struct hat *sfmmup)
1432 {
1433 	int i;
1434 
1435 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1436 	if (sfmmup->sfmmu_ismhat) {
1437 		for (i = 0; i < mmu_page_sizes; i++) {
1438 			sfmmup->sfmmu_ttecnt[i] = 0;
1439 			sfmmup->sfmmu_ismttecnt[i] = 0;
1440 		}
1441 	} else {
1442 		/* EMPTY */
1443 		ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
1444 		ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
1445 		ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
1446 		ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
1447 		ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
1448 		ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
1449 	}
1450 
1451 	if (sfmmup->sfmmu_rmstat) {
1452 		hat_freestat(sfmmup->sfmmu_as, NULL);
1453 	}
1454 	if (!delay_tlb_flush) {
1455 		sfmmu_tlb_ctx_demap(sfmmup);
1456 		xt_sync(sfmmup->sfmmu_cpusran);
1457 	} else {
1458 		SFMMU_STAT(sf_tlbflush_deferred);
1459 	}
1460 	sfmmu_free_ctx(sfmmup, sfmmutoctx(sfmmup));
1461 	while (sfmmup->sfmmu_tsb != NULL) {
1462 		struct tsb_info *next = sfmmup->sfmmu_tsb->tsb_next;
1463 		sfmmu_tsbinfo_free(sfmmup->sfmmu_tsb);
1464 		sfmmup->sfmmu_tsb = next;
1465 	}
1466 	sfmmu_free_sfmmu(sfmmup);
1467 
1468 	kmem_cache_free(sfmmuid_cache, sfmmup);
1469 }
1470 
1471 /*
1472  * Set up any translation structures, for the specified address space,
1473  * that are needed or preferred when the process is being swapped in.
1474  */
1475 /* ARGSUSED */
1476 void
1477 hat_swapin(struct hat *hat)
1478 {
1479 	ASSERT(hat->sfmmu_xhat_provider == NULL);
1480 }
1481 
1482 /*
1483  * Free all of the translation resources, for the specified address space,
1484  * that can be freed while the process is swapped out. Called from as_swapout.
1485  * Also, free up the ctx that this process was using.
1486  */
1487 void
1488 hat_swapout(struct hat *sfmmup)
1489 {
1490 	struct hmehash_bucket *hmebp;
1491 	struct hme_blk *hmeblkp;
1492 	struct hme_blk *pr_hblk = NULL;
1493 	struct hme_blk *nx_hblk;
1494 	struct ctx *ctx;
1495 	int cnum;
1496 	int i;
1497 	uint64_t hblkpa, prevpa, nx_pa;
1498 	struct hme_blk *list = NULL;
1499 	hatlock_t *hatlockp;
1500 	struct tsb_info *tsbinfop;
1501 	struct free_tsb {
1502 		struct free_tsb *next;
1503 		struct tsb_info *tsbinfop;
1504 	};			/* free list of TSBs */
1505 	struct free_tsb *freelist, *last, *next;
1506 
1507 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
1508 	SFMMU_STAT(sf_swapout);
1509 
1510 	/*
1511 	 * There is no way to go from an as to all its translations in sfmmu.
1512 	 * Here is one of the times when we take the big hit and traverse
1513 	 * the hash looking for hme_blks to free up.  Not only do we free up
1514 	 * this as hme_blks but all those that are free.  We are obviously
1515 	 * swapping because we need memory so let's free up as much
1516 	 * as we can.
1517 	 *
1518 	 * Note that we don't flush TLB/TSB here -- it's not necessary
1519 	 * because:
1520 	 *  1) we free the ctx we're using and throw away the TSB(s);
1521 	 *  2) processes aren't runnable while being swapped out.
1522 	 */
1523 	ASSERT(sfmmup != KHATID);
1524 	for (i = 0; i <= UHMEHASH_SZ; i++) {
1525 		hmebp = &uhme_hash[i];
1526 		SFMMU_HASH_LOCK(hmebp);
1527 		hmeblkp = hmebp->hmeblkp;
1528 		hblkpa = hmebp->hmeh_nextpa;
1529 		prevpa = 0;
1530 		pr_hblk = NULL;
1531 		while (hmeblkp) {
1532 
1533 			ASSERT(!hmeblkp->hblk_xhat_bit);
1534 
1535 			if ((hmeblkp->hblk_tag.htag_id == sfmmup) &&
1536 			    !hmeblkp->hblk_shw_bit && !hmeblkp->hblk_lckcnt) {
1537 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
1538 					(caddr_t)get_hblk_base(hmeblkp),
1539 					get_hblk_endaddr(hmeblkp),
1540 					NULL, HAT_UNLOAD);
1541 			}
1542 			nx_hblk = hmeblkp->hblk_next;
1543 			nx_pa = hmeblkp->hblk_nextpa;
1544 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
1545 				ASSERT(!hmeblkp->hblk_lckcnt);
1546 				sfmmu_hblk_hash_rm(hmebp, hmeblkp,
1547 					prevpa, pr_hblk);
1548 				sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
1549 			} else {
1550 				pr_hblk = hmeblkp;
1551 				prevpa = hblkpa;
1552 			}
1553 			hmeblkp = nx_hblk;
1554 			hblkpa = nx_pa;
1555 		}
1556 		SFMMU_HASH_UNLOCK(hmebp);
1557 	}
1558 
1559 	sfmmu_hblks_list_purge(&list);
1560 
1561 	/*
1562 	 * Now free up the ctx so that others can reuse it.
1563 	 */
1564 	hatlockp = sfmmu_hat_enter(sfmmup);
1565 	ctx = sfmmutoctx(sfmmup);
1566 	cnum = ctxtoctxnum(ctx);
1567 
1568 	if (cnum != INVALID_CONTEXT) {
1569 		rw_enter(&ctx->ctx_rwlock, RW_WRITER);
1570 		if (sfmmup->sfmmu_cnum == cnum) {
1571 			sfmmu_reuse_ctx(ctx, sfmmup);
1572 			/*
1573 			 * Put ctx back to the free list.
1574 			 */
1575 			mutex_enter(&ctx_list_lock);
1576 			CTX_SET_FLAGS(ctx, CTX_FREE_FLAG);
1577 			ctx->ctx_free = ctxfree;
1578 			ctxfree = ctx;
1579 			mutex_exit(&ctx_list_lock);
1580 		}
1581 		rw_exit(&ctx->ctx_rwlock);
1582 	}
1583 
1584 	/*
1585 	 * Free TSBs, but not tsbinfos, and set SWAPPED flag.
1586 	 * If TSBs were never swapped in, just return.
1587 	 * This implies that we don't support partial swapping
1588 	 * of TSBs -- either all are swapped out, or none are.
1589 	 *
1590 	 * We must hold the HAT lock here to prevent racing with another
1591 	 * thread trying to unmap TTEs from the TSB or running the post-
1592 	 * relocator after relocating the TSB's memory.  Unfortunately, we
1593 	 * can't free memory while holding the HAT lock or we could
1594 	 * deadlock, so we build a list of TSBs to be freed after marking
1595 	 * the tsbinfos as swapped out and free them after dropping the
1596 	 * lock.
1597 	 */
1598 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
1599 		sfmmu_hat_exit(hatlockp);
1600 		return;
1601 	}
1602 
1603 	SFMMU_FLAGS_SET(sfmmup, HAT_SWAPPED);
1604 	last = freelist = NULL;
1605 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
1606 	    tsbinfop = tsbinfop->tsb_next) {
1607 		ASSERT((tsbinfop->tsb_flags & TSB_SWAPPED) == 0);
1608 
1609 		/*
1610 		 * Cast the TSB into a struct free_tsb and put it on the free
1611 		 * list.
1612 		 */
1613 		if (freelist == NULL) {
1614 			last = freelist = (struct free_tsb *)tsbinfop->tsb_va;
1615 		} else {
1616 			last->next = (struct free_tsb *)tsbinfop->tsb_va;
1617 			last = last->next;
1618 		}
1619 		last->next = NULL;
1620 		last->tsbinfop = tsbinfop;
1621 		tsbinfop->tsb_flags |= TSB_SWAPPED;
1622 		/*
1623 		 * Zero out the TTE to clear the valid bit.
1624 		 * Note we can't use a value like 0xbad because we want to
1625 		 * ensure diagnostic bits are NEVER set on TTEs that might
1626 		 * be loaded.  The intent is to catch any invalid access
1627 		 * to the swapped TSB, such as a thread running with a valid
1628 		 * context without first calling sfmmu_tsb_swapin() to
1629 		 * allocate TSB memory.
1630 		 */
1631 		tsbinfop->tsb_tte.ll = 0;
1632 	}
1633 
1634 	/* Now we can drop the lock and free the TSB memory. */
1635 	sfmmu_hat_exit(hatlockp);
1636 	for (; freelist != NULL; freelist = next) {
1637 		next = freelist->next;
1638 		sfmmu_tsb_free(freelist->tsbinfop);
1639 	}
1640 }
1641 
1642 /*
1643  * Duplicate the translations of an as into another newas
1644  */
1645 /* ARGSUSED */
1646 int
1647 hat_dup(struct hat *hat, struct hat *newhat, caddr_t addr, size_t len,
1648 	uint_t flag)
1649 {
1650 	ASSERT(hat->sfmmu_xhat_provider == NULL);
1651 	ASSERT((flag == 0) || (flag == HAT_DUP_ALL) || (flag == HAT_DUP_COW));
1652 
1653 	if (flag == HAT_DUP_COW) {
1654 		panic("hat_dup: HAT_DUP_COW not supported");
1655 	}
1656 	return (0);
1657 }
1658 
1659 /*
1660  * Set up addr to map to page pp with protection prot.
1661  * As an optimization we also load the TSB with the
1662  * corresponding tte but it is no big deal if  the tte gets kicked out.
1663  */
1664 void
1665 hat_memload(struct hat *hat, caddr_t addr, struct page *pp,
1666 	uint_t attr, uint_t flags)
1667 {
1668 	tte_t tte;
1669 
1670 
1671 	ASSERT(hat != NULL);
1672 	ASSERT(PAGE_LOCKED(pp));
1673 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
1674 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
1675 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
1676 
1677 	if (PP_ISFREE(pp)) {
1678 		panic("hat_memload: loading a mapping to free page %p",
1679 		    (void *)pp);
1680 	}
1681 
1682 	if (hat->sfmmu_xhat_provider) {
1683 		XHAT_MEMLOAD(hat, addr, pp, attr, flags);
1684 		return;
1685 	}
1686 
1687 	ASSERT((hat == ksfmmup) ||
1688 		AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock));
1689 
1690 	if (flags & ~SFMMU_LOAD_ALLFLAG)
1691 		cmn_err(CE_NOTE, "hat_memload: unsupported flags %d",
1692 		    flags & ~SFMMU_LOAD_ALLFLAG);
1693 
1694 	if (hat->sfmmu_rmstat)
1695 		hat_resvstat(MMU_PAGESIZE, hat->sfmmu_as, addr);
1696 
1697 #if defined(SF_ERRATA_57)
1698 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
1699 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
1700 	    !(flags & HAT_LOAD_SHARE)) {
1701 		cmn_err(CE_WARN, "hat_memload: illegal attempt to make user "
1702 		    " page executable");
1703 		attr &= ~PROT_EXEC;
1704 	}
1705 #endif
1706 
1707 	sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
1708 	(void) sfmmu_tteload_array(hat, &tte, addr, &pp, flags);
1709 
1710 	/*
1711 	 * Check TSB and TLB page sizes.
1712 	 */
1713 	if ((flags & HAT_LOAD_SHARE) == 0) {
1714 		sfmmu_check_page_sizes(hat, 1);
1715 	}
1716 }
1717 
1718 /*
1719  * hat_devload can be called to map real memory (e.g.
1720  * /dev/kmem) and even though hat_devload will determine pf is
1721  * for memory, it will be unable to get a shared lock on the
1722  * page (because someone else has it exclusively) and will
1723  * pass dp = NULL.  If tteload doesn't get a non-NULL
1724  * page pointer it can't cache memory.
1725  */
1726 void
1727 hat_devload(struct hat *hat, caddr_t addr, size_t len, pfn_t pfn,
1728 	uint_t attr, int flags)
1729 {
1730 	tte_t tte;
1731 	struct page *pp = NULL;
1732 	int use_lgpg = 0;
1733 
1734 	ASSERT(hat != NULL);
1735 
1736 	if (hat->sfmmu_xhat_provider) {
1737 		XHAT_DEVLOAD(hat, addr, len, pfn, attr, flags);
1738 		return;
1739 	}
1740 
1741 	ASSERT(!(flags & ~SFMMU_LOAD_ALLFLAG));
1742 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
1743 	ASSERT((hat == ksfmmup) ||
1744 		AS_LOCK_HELD(hat->sfmmu_as, &hat->sfmmu_as->a_lock));
1745 	if (len == 0)
1746 		panic("hat_devload: zero len");
1747 	if (flags & ~SFMMU_LOAD_ALLFLAG)
1748 		cmn_err(CE_NOTE, "hat_devload: unsupported flags %d",
1749 		    flags & ~SFMMU_LOAD_ALLFLAG);
1750 
1751 #if defined(SF_ERRATA_57)
1752 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
1753 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
1754 	    !(flags & HAT_LOAD_SHARE)) {
1755 		cmn_err(CE_WARN, "hat_devload: illegal attempt to make user "
1756 		    " page executable");
1757 		attr &= ~PROT_EXEC;
1758 	}
1759 #endif
1760 
1761 	/*
1762 	 * If it's a memory page find its pp
1763 	 */
1764 	if (!(flags & HAT_LOAD_NOCONSIST) && pf_is_memory(pfn)) {
1765 		pp = page_numtopp_nolock(pfn);
1766 		if (pp == NULL) {
1767 			flags |= HAT_LOAD_NOCONSIST;
1768 		} else {
1769 			if (PP_ISFREE(pp)) {
1770 				panic("hat_memload: loading "
1771 				    "a mapping to free page %p",
1772 				    (void *)pp);
1773 			}
1774 			if (!PAGE_LOCKED(pp) && !PP_ISNORELOC(pp)) {
1775 				panic("hat_memload: loading a mapping "
1776 				    "to unlocked relocatable page %p",
1777 				    (void *)pp);
1778 			}
1779 			ASSERT(len == MMU_PAGESIZE);
1780 		}
1781 	}
1782 
1783 	if (hat->sfmmu_rmstat)
1784 		hat_resvstat(len, hat->sfmmu_as, addr);
1785 
1786 	if (flags & HAT_LOAD_NOCONSIST) {
1787 		attr |= SFMMU_UNCACHEVTTE;
1788 		use_lgpg = 1;
1789 	}
1790 	if (!pf_is_memory(pfn)) {
1791 		attr |= SFMMU_UNCACHEPTTE | HAT_NOSYNC;
1792 		use_lgpg = 1;
1793 		switch (attr & HAT_ORDER_MASK) {
1794 			case HAT_STRICTORDER:
1795 			case HAT_UNORDERED_OK:
1796 				/*
1797 				 * we set the side effect bit for all non
1798 				 * memory mappings unless merging is ok
1799 				 */
1800 				attr |= SFMMU_SIDEFFECT;
1801 				break;
1802 			case HAT_MERGING_OK:
1803 			case HAT_LOADCACHING_OK:
1804 			case HAT_STORECACHING_OK:
1805 				break;
1806 			default:
1807 				panic("hat_devload: bad attr");
1808 				break;
1809 		}
1810 	}
1811 	while (len) {
1812 		if (!use_lgpg) {
1813 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
1814 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
1815 			    flags);
1816 			len -= MMU_PAGESIZE;
1817 			addr += MMU_PAGESIZE;
1818 			pfn++;
1819 			continue;
1820 		}
1821 		/*
1822 		 *  try to use large pages, check va/pa alignments
1823 		 *  Note that 32M/256M page sizes are not (yet) supported.
1824 		 */
1825 		if ((len >= MMU_PAGESIZE4M) &&
1826 		    !((uintptr_t)addr & MMU_PAGEOFFSET4M) &&
1827 		    !(disable_large_pages & (1 << TTE4M)) &&
1828 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET4M)) {
1829 			sfmmu_memtte(&tte, pfn, attr, TTE4M);
1830 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
1831 			    flags);
1832 			len -= MMU_PAGESIZE4M;
1833 			addr += MMU_PAGESIZE4M;
1834 			pfn += MMU_PAGESIZE4M / MMU_PAGESIZE;
1835 		} else if ((len >= MMU_PAGESIZE512K) &&
1836 		    !((uintptr_t)addr & MMU_PAGEOFFSET512K) &&
1837 		    !(disable_large_pages & (1 << TTE512K)) &&
1838 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET512K)) {
1839 			sfmmu_memtte(&tte, pfn, attr, TTE512K);
1840 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
1841 			    flags);
1842 			len -= MMU_PAGESIZE512K;
1843 			addr += MMU_PAGESIZE512K;
1844 			pfn += MMU_PAGESIZE512K / MMU_PAGESIZE;
1845 		} else if ((len >= MMU_PAGESIZE64K) &&
1846 		    !((uintptr_t)addr & MMU_PAGEOFFSET64K) &&
1847 		    !(disable_large_pages & (1 << TTE64K)) &&
1848 		    !(mmu_ptob(pfn) & MMU_PAGEOFFSET64K)) {
1849 			sfmmu_memtte(&tte, pfn, attr, TTE64K);
1850 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
1851 			    flags);
1852 			len -= MMU_PAGESIZE64K;
1853 			addr += MMU_PAGESIZE64K;
1854 			pfn += MMU_PAGESIZE64K / MMU_PAGESIZE;
1855 		} else {
1856 			sfmmu_memtte(&tte, pfn, attr, TTE8K);
1857 			(void) sfmmu_tteload_array(hat, &tte, addr, &pp,
1858 			    flags);
1859 			len -= MMU_PAGESIZE;
1860 			addr += MMU_PAGESIZE;
1861 			pfn++;
1862 		}
1863 	}
1864 
1865 	/*
1866 	 * Check TSB and TLB page sizes.
1867 	 */
1868 	if ((flags & HAT_LOAD_SHARE) == 0) {
1869 		sfmmu_check_page_sizes(hat, 1);
1870 	}
1871 }
1872 
1873 /*
1874  * Map the largest extend possible out of the page array. The array may NOT
1875  * be in order.  The largest possible mapping a page can have
1876  * is specified in the p_szc field.  The p_szc field
1877  * cannot change as long as there any mappings (large or small)
1878  * to any of the pages that make up the large page. (ie. any
1879  * promotion/demotion of page size is not up to the hat but up to
1880  * the page free list manager).  The array
1881  * should consist of properly aligned contigous pages that are
1882  * part of a big page for a large mapping to be created.
1883  */
1884 void
1885 hat_memload_array(struct hat *hat, caddr_t addr, size_t len,
1886 	struct page **pps, uint_t attr, uint_t flags)
1887 {
1888 	int  ttesz;
1889 	size_t mapsz;
1890 	pgcnt_t	numpg, npgs;
1891 	tte_t tte;
1892 	page_t *pp;
1893 	int large_pages_disable;
1894 
1895 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
1896 
1897 	if (hat->sfmmu_xhat_provider) {
1898 		XHAT_MEMLOAD_ARRAY(hat, addr, len, pps, attr, flags);
1899 		return;
1900 	}
1901 
1902 	if (hat->sfmmu_rmstat)
1903 		hat_resvstat(len, hat->sfmmu_as, addr);
1904 
1905 #if defined(SF_ERRATA_57)
1906 	if ((hat != ksfmmup) && AS_TYPE_64BIT(hat->sfmmu_as) &&
1907 	    (addr < errata57_limit) && (attr & PROT_EXEC) &&
1908 	    !(flags & HAT_LOAD_SHARE)) {
1909 		cmn_err(CE_WARN, "hat_memload_array: illegal attempt to make "
1910 		    "user page executable");
1911 		attr &= ~PROT_EXEC;
1912 	}
1913 #endif
1914 
1915 	/* Get number of pages */
1916 	npgs = len >> MMU_PAGESHIFT;
1917 
1918 	if (flags & HAT_LOAD_SHARE) {
1919 		large_pages_disable = disable_ism_large_pages;
1920 	} else {
1921 		large_pages_disable = disable_large_pages;
1922 	}
1923 
1924 	if (npgs < NHMENTS || large_pages_disable == LARGE_PAGES_OFF) {
1925 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs);
1926 		return;
1927 	}
1928 
1929 	while (npgs >= NHMENTS) {
1930 		pp = *pps;
1931 		for (ttesz = pp->p_szc; ttesz != TTE8K; ttesz--) {
1932 			/*
1933 			 * Check if this page size is disabled.
1934 			 */
1935 			if (large_pages_disable & (1 << ttesz))
1936 				continue;
1937 
1938 			numpg = TTEPAGES(ttesz);
1939 			mapsz = numpg << MMU_PAGESHIFT;
1940 			if ((npgs >= numpg) &&
1941 			    IS_P2ALIGNED(addr, mapsz) &&
1942 			    IS_P2ALIGNED(pp->p_pagenum, numpg)) {
1943 				/*
1944 				 * At this point we have enough pages and
1945 				 * we know the virtual address and the pfn
1946 				 * are properly aligned.  We still need
1947 				 * to check for physical contiguity but since
1948 				 * it is very likely that this is the case
1949 				 * we will assume they are so and undo
1950 				 * the request if necessary.  It would
1951 				 * be great if we could get a hint flag
1952 				 * like HAT_CONTIG which would tell us
1953 				 * the pages are contigous for sure.
1954 				 */
1955 				sfmmu_memtte(&tte, (*pps)->p_pagenum,
1956 					attr, ttesz);
1957 				if (!sfmmu_tteload_array(hat, &tte, addr,
1958 				    pps, flags)) {
1959 					break;
1960 				}
1961 			}
1962 		}
1963 		if (ttesz == TTE8K) {
1964 			/*
1965 			 * We were not able to map array using a large page
1966 			 * batch a hmeblk or fraction at a time.
1967 			 */
1968 			numpg = ((uintptr_t)addr >> MMU_PAGESHIFT)
1969 				& (NHMENTS-1);
1970 			numpg = NHMENTS - numpg;
1971 			ASSERT(numpg <= npgs);
1972 			mapsz = numpg * MMU_PAGESIZE;
1973 			sfmmu_memload_batchsmall(hat, addr, pps, attr, flags,
1974 							numpg);
1975 		}
1976 		addr += mapsz;
1977 		npgs -= numpg;
1978 		pps += numpg;
1979 	}
1980 
1981 	if (npgs) {
1982 		sfmmu_memload_batchsmall(hat, addr, pps, attr, flags, npgs);
1983 	}
1984 
1985 	/*
1986 	 * Check TSB and TLB page sizes.
1987 	 */
1988 	if ((flags & HAT_LOAD_SHARE) == 0) {
1989 		sfmmu_check_page_sizes(hat, 1);
1990 	}
1991 }
1992 
1993 /*
1994  * Function tries to batch 8K pages into the same hme blk.
1995  */
1996 static void
1997 sfmmu_memload_batchsmall(struct hat *hat, caddr_t vaddr, page_t **pps,
1998 		    uint_t attr, uint_t flags, pgcnt_t npgs)
1999 {
2000 	tte_t	tte;
2001 	page_t *pp;
2002 	struct hmehash_bucket *hmebp;
2003 	struct hme_blk *hmeblkp;
2004 	int	index;
2005 
2006 	while (npgs) {
2007 		/*
2008 		 * Acquire the hash bucket.
2009 		 */
2010 		hmebp = sfmmu_tteload_acquire_hashbucket(hat, vaddr, TTE8K);
2011 		ASSERT(hmebp);
2012 
2013 		/*
2014 		 * Find the hment block.
2015 		 */
2016 		hmeblkp = sfmmu_tteload_find_hmeblk(hat, hmebp, vaddr,
2017 				TTE8K, flags);
2018 		ASSERT(hmeblkp);
2019 
2020 		do {
2021 			/*
2022 			 * Make the tte.
2023 			 */
2024 			pp = *pps;
2025 			sfmmu_memtte(&tte, pp->p_pagenum, attr, TTE8K);
2026 
2027 			/*
2028 			 * Add the translation.
2029 			 */
2030 			(void) sfmmu_tteload_addentry(hat, hmeblkp, &tte,
2031 					vaddr, pps, flags);
2032 
2033 			/*
2034 			 * Goto next page.
2035 			 */
2036 			pps++;
2037 			npgs--;
2038 
2039 			/*
2040 			 * Goto next address.
2041 			 */
2042 			vaddr += MMU_PAGESIZE;
2043 
2044 			/*
2045 			 * Don't crossover into a different hmentblk.
2046 			 */
2047 			index = (int)(((uintptr_t)vaddr >> MMU_PAGESHIFT) &
2048 			    (NHMENTS-1));
2049 
2050 		} while (index != 0 && npgs != 0);
2051 
2052 		/*
2053 		 * Release the hash bucket.
2054 		 */
2055 
2056 		sfmmu_tteload_release_hashbucket(hmebp);
2057 	}
2058 }
2059 
2060 /*
2061  * Construct a tte for a page:
2062  *
2063  * tte_valid = 1
2064  * tte_size2 = size & TTE_SZ2_BITS (Panther-only)
2065  * tte_size = size
2066  * tte_nfo = attr & HAT_NOFAULT
2067  * tte_ie = attr & HAT_STRUCTURE_LE
2068  * tte_hmenum = hmenum
2069  * tte_pahi = pp->p_pagenum >> TTE_PASHIFT;
2070  * tte_palo = pp->p_pagenum & TTE_PALOMASK;
2071  * tte_ref = 1 (optimization)
2072  * tte_wr_perm = attr & PROT_WRITE;
2073  * tte_no_sync = attr & HAT_NOSYNC
2074  * tte_lock = attr & SFMMU_LOCKTTE
2075  * tte_cp = !(attr & SFMMU_UNCACHEPTTE)
2076  * tte_cv = !(attr & SFMMU_UNCACHEVTTE)
2077  * tte_e = attr & SFMMU_SIDEFFECT
2078  * tte_priv = !(attr & PROT_USER)
2079  * tte_hwwr = if nosync is set and it is writable we set the mod bit (opt)
2080  * tte_glb = 0
2081  */
2082 void
2083 sfmmu_memtte(tte_t *ttep, pfn_t pfn, uint_t attr, int tte_sz)
2084 {
2085 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
2086 
2087 	ttep->tte_inthi = MAKE_TTE_INTHI(pfn, attr, tte_sz, 0 /* hmenum */);
2088 	ttep->tte_intlo = MAKE_TTE_INTLO(pfn, attr, tte_sz, 0 /* hmenum */);
2089 
2090 	if (TTE_IS_NOSYNC(ttep)) {
2091 		TTE_SET_REF(ttep);
2092 		if (TTE_IS_WRITABLE(ttep)) {
2093 			TTE_SET_MOD(ttep);
2094 		}
2095 	}
2096 	if (TTE_IS_NFO(ttep) && TTE_IS_EXECUTABLE(ttep)) {
2097 		panic("sfmmu_memtte: can't set both NFO and EXEC bits");
2098 	}
2099 }
2100 
2101 /*
2102  * This function will add a translation to the hme_blk and allocate the
2103  * hme_blk if one does not exist.
2104  * If a page structure is specified then it will add the
2105  * corresponding hment to the mapping list.
2106  * It will also update the hmenum field for the tte.
2107  */
2108 void
2109 sfmmu_tteload(struct hat *sfmmup, tte_t *ttep, caddr_t vaddr, page_t *pp,
2110 	uint_t flags)
2111 {
2112 	(void) sfmmu_tteload_array(sfmmup, ttep, vaddr, &pp, flags);
2113 }
2114 
2115 /*
2116  * Load (ttep != NULL) or unload (ttep == NULL) one entry in the TSB.
2117  * Assumes that a particular page size may only be resident in one TSB.
2118  */
2119 static void
2120 sfmmu_mod_tsb(sfmmu_t *sfmmup, caddr_t vaddr, tte_t *ttep, int ttesz)
2121 {
2122 	struct tsb_info *tsbinfop = NULL;
2123 	uint64_t tag;
2124 	struct tsbe *tsbe_addr;
2125 	uint64_t tsb_base;
2126 	uint_t tsb_size;
2127 	int vpshift = MMU_PAGESHIFT;
2128 	int phys = 0;
2129 
2130 	if (sfmmup == ksfmmup) { /* No support for 32/256M ksfmmu pages */
2131 		phys = ktsb_phys;
2132 		if (ttesz >= TTE4M) {
2133 #ifndef sun4v
2134 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2135 #endif
2136 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2137 			tsb_size = ktsb4m_szcode;
2138 		} else {
2139 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2140 			tsb_size = ktsb_szcode;
2141 		}
2142 	} else {
2143 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2144 
2145 		/*
2146 		 * If there isn't a TSB for this page size, or the TSB is
2147 		 * swapped out, there is nothing to do.  Note that the latter
2148 		 * case seems impossible but can occur if hat_pageunload()
2149 		 * is called on an ISM mapping while the process is swapped
2150 		 * out.
2151 		 */
2152 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2153 			return;
2154 
2155 		/*
2156 		 * If another thread is in the middle of relocating a TSB
2157 		 * we can't unload the entry so set a flag so that the
2158 		 * TSB will be flushed before it can be accessed by the
2159 		 * process.
2160 		 */
2161 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2162 			if (ttep == NULL)
2163 				tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2164 			return;
2165 		}
2166 #if defined(UTSB_PHYS)
2167 		phys = 1;
2168 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2169 #else
2170 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2171 #endif
2172 		tsb_size = tsbinfop->tsb_szc;
2173 	}
2174 	if (ttesz >= TTE4M)
2175 		vpshift = MMU_PAGESHIFT4M;
2176 
2177 	tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2178 	tag = sfmmu_make_tsbtag(vaddr);
2179 
2180 	if (ttep == NULL) {
2181 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2182 	} else {
2183 		if (ttesz >= TTE4M) {
2184 			SFMMU_STAT(sf_tsb_load4m);
2185 		} else {
2186 			SFMMU_STAT(sf_tsb_load8k);
2187 		}
2188 
2189 		sfmmu_load_tsbe(tsbe_addr, tag, ttep, phys);
2190 	}
2191 }
2192 
2193 /*
2194  * Unmap all entries from [start, end) matching the given page size.
2195  *
2196  * This function is used primarily to unmap replicated 64K or 512K entries
2197  * from the TSB that are inserted using the base page size TSB pointer, but
2198  * it may also be called to unmap a range of addresses from the TSB.
2199  */
2200 void
2201 sfmmu_unload_tsb_range(sfmmu_t *sfmmup, caddr_t start, caddr_t end, int ttesz)
2202 {
2203 	struct tsb_info *tsbinfop;
2204 	uint64_t tag;
2205 	struct tsbe *tsbe_addr;
2206 	caddr_t vaddr;
2207 	uint64_t tsb_base;
2208 	int vpshift, vpgsz;
2209 	uint_t tsb_size;
2210 	int phys = 0;
2211 
2212 	/*
2213 	 * Assumptions:
2214 	 *  If ttesz == 8K, 64K or 512K, we walk through the range 8K
2215 	 *  at a time shooting down any valid entries we encounter.
2216 	 *
2217 	 *  If ttesz >= 4M we walk the range 4M at a time shooting
2218 	 *  down any valid mappings we find.
2219 	 */
2220 	if (sfmmup == ksfmmup) {
2221 		phys = ktsb_phys;
2222 		if (ttesz >= TTE4M) {
2223 #ifndef sun4v
2224 			ASSERT((ttesz != TTE32M) && (ttesz != TTE256M));
2225 #endif
2226 			tsb_base = (phys)? ktsb4m_pbase : (uint64_t)ktsb4m_base;
2227 			tsb_size = ktsb4m_szcode;
2228 		} else {
2229 			tsb_base = (phys)? ktsb_pbase : (uint64_t)ktsb_base;
2230 			tsb_size = ktsb_szcode;
2231 		}
2232 	} else {
2233 		SFMMU_GET_TSBINFO(tsbinfop, sfmmup, ttesz);
2234 
2235 		/*
2236 		 * If there isn't a TSB for this page size, or the TSB is
2237 		 * swapped out, there is nothing to do.  Note that the latter
2238 		 * case seems impossible but can occur if hat_pageunload()
2239 		 * is called on an ISM mapping while the process is swapped
2240 		 * out.
2241 		 */
2242 		if (tsbinfop == NULL || (tsbinfop->tsb_flags & TSB_SWAPPED))
2243 			return;
2244 
2245 		/*
2246 		 * If another thread is in the middle of relocating a TSB
2247 		 * we can't unload the entry so set a flag so that the
2248 		 * TSB will be flushed before it can be accessed by the
2249 		 * process.
2250 		 */
2251 		if ((tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) {
2252 			tsbinfop->tsb_flags |= TSB_FLUSH_NEEDED;
2253 			return;
2254 		}
2255 #if defined(UTSB_PHYS)
2256 		phys = 1;
2257 		tsb_base = (uint64_t)tsbinfop->tsb_pa;
2258 #else
2259 		tsb_base = (uint64_t)tsbinfop->tsb_va;
2260 #endif
2261 		tsb_size = tsbinfop->tsb_szc;
2262 	}
2263 	if (ttesz >= TTE4M) {
2264 		vpshift = MMU_PAGESHIFT4M;
2265 		vpgsz = MMU_PAGESIZE4M;
2266 	} else {
2267 		vpshift = MMU_PAGESHIFT;
2268 		vpgsz = MMU_PAGESIZE;
2269 	}
2270 
2271 	for (vaddr = start; vaddr < end; vaddr += vpgsz) {
2272 		tag = sfmmu_make_tsbtag(vaddr);
2273 		tsbe_addr = sfmmu_get_tsbe(tsb_base, vaddr, vpshift, tsb_size);
2274 		sfmmu_unload_tsbe(tsbe_addr, tag, phys);
2275 	}
2276 }
2277 
2278 /*
2279  * Select the optimum TSB size given the number of mappings
2280  * that need to be cached.
2281  */
2282 static int
2283 sfmmu_select_tsb_szc(pgcnt_t pgcnt)
2284 {
2285 	int szc = 0;
2286 
2287 #ifdef DEBUG
2288 	if (tsb_grow_stress) {
2289 		uint32_t randval = (uint32_t)gettick() >> 4;
2290 		return (randval % (tsb_max_growsize + 1));
2291 	}
2292 #endif	/* DEBUG */
2293 
2294 	while ((szc < tsb_max_growsize) && (pgcnt > SFMMU_RSS_TSBSIZE(szc)))
2295 		szc++;
2296 	return (szc);
2297 }
2298 
2299 /*
2300  * This function will add a translation to the hme_blk and allocate the
2301  * hme_blk if one does not exist.
2302  * If a page structure is specified then it will add the
2303  * corresponding hment to the mapping list.
2304  * It will also update the hmenum field for the tte.
2305  * Furthermore, it attempts to create a large page translation
2306  * for <addr,hat> at page array pps.  It assumes addr and first
2307  * pp is correctly aligned.  It returns 0 if successful and 1 otherwise.
2308  */
2309 static int
2310 sfmmu_tteload_array(sfmmu_t *sfmmup, tte_t *ttep, caddr_t vaddr,
2311 	page_t **pps, uint_t flags)
2312 {
2313 	struct hmehash_bucket *hmebp;
2314 	struct hme_blk *hmeblkp;
2315 	int 	ret;
2316 	uint_t	size;
2317 
2318 	/*
2319 	 * Get mapping size.
2320 	 */
2321 	size = TTE_CSZ(ttep);
2322 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
2323 
2324 	/*
2325 	 * Acquire the hash bucket.
2326 	 */
2327 	hmebp = sfmmu_tteload_acquire_hashbucket(sfmmup, vaddr, size);
2328 	ASSERT(hmebp);
2329 
2330 	/*
2331 	 * Find the hment block.
2332 	 */
2333 	hmeblkp = sfmmu_tteload_find_hmeblk(sfmmup, hmebp, vaddr, size, flags);
2334 	ASSERT(hmeblkp);
2335 
2336 	/*
2337 	 * Add the translation.
2338 	 */
2339 	ret = sfmmu_tteload_addentry(sfmmup, hmeblkp, ttep, vaddr, pps, flags);
2340 
2341 	/*
2342 	 * Release the hash bucket.
2343 	 */
2344 	sfmmu_tteload_release_hashbucket(hmebp);
2345 
2346 	return (ret);
2347 }
2348 
2349 /*
2350  * Function locks and returns a pointer to the hash bucket for vaddr and size.
2351  */
2352 static struct hmehash_bucket *
2353 sfmmu_tteload_acquire_hashbucket(sfmmu_t *sfmmup, caddr_t vaddr, int size)
2354 {
2355 	struct hmehash_bucket *hmebp;
2356 	int hmeshift;
2357 
2358 	hmeshift = HME_HASH_SHIFT(size);
2359 
2360 	hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
2361 
2362 	SFMMU_HASH_LOCK(hmebp);
2363 
2364 	return (hmebp);
2365 }
2366 
2367 /*
2368  * Function returns a pointer to an hmeblk in the hash bucket, hmebp. If the
2369  * hmeblk doesn't exists for the [sfmmup, vaddr & size] signature, a hmeblk is
2370  * allocated.
2371  */
2372 static struct hme_blk *
2373 sfmmu_tteload_find_hmeblk(sfmmu_t *sfmmup, struct hmehash_bucket *hmebp,
2374 	caddr_t vaddr, uint_t size, uint_t flags)
2375 {
2376 	hmeblk_tag hblktag;
2377 	int hmeshift;
2378 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
2379 	uint64_t hblkpa, prevpa;
2380 	struct kmem_cache *sfmmu_cache;
2381 	uint_t forcefree;
2382 
2383 	hblktag.htag_id = sfmmup;
2384 	hmeshift = HME_HASH_SHIFT(size);
2385 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
2386 	hblktag.htag_rehash = HME_HASH_REHASH(size);
2387 
2388 ttearray_realloc:
2389 
2390 	HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, hblkpa,
2391 	    pr_hblk, prevpa, &list);
2392 
2393 	/*
2394 	 * We block until hblk_reserve_lock is released; it's held by
2395 	 * the thread, temporarily using hblk_reserve, until hblk_reserve is
2396 	 * replaced by a hblk from sfmmu8_cache.
2397 	 */
2398 	if (hmeblkp == (struct hme_blk *)hblk_reserve &&
2399 	    hblk_reserve_thread != curthread) {
2400 		SFMMU_HASH_UNLOCK(hmebp);
2401 		mutex_enter(&hblk_reserve_lock);
2402 		mutex_exit(&hblk_reserve_lock);
2403 		SFMMU_STAT(sf_hblk_reserve_hit);
2404 		SFMMU_HASH_LOCK(hmebp);
2405 		goto ttearray_realloc;
2406 	}
2407 
2408 	if (hmeblkp == NULL) {
2409 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
2410 		    hblktag, flags);
2411 	} else {
2412 		/*
2413 		 * It is possible for 8k and 64k hblks to collide since they
2414 		 * have the same rehash value. This is because we
2415 		 * lazily free hblks and 8K/64K blks could be lingering.
2416 		 * If we find size mismatch we free the block and & try again.
2417 		 */
2418 		if (get_hblk_ttesz(hmeblkp) != size) {
2419 			ASSERT(!hmeblkp->hblk_vcnt);
2420 			ASSERT(!hmeblkp->hblk_hmecnt);
2421 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa, pr_hblk);
2422 			sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
2423 			goto ttearray_realloc;
2424 		}
2425 		if (hmeblkp->hblk_shw_bit) {
2426 			/*
2427 			 * if the hblk was previously used as a shadow hblk then
2428 			 * we will change it to a normal hblk
2429 			 */
2430 			if (hmeblkp->hblk_shw_mask) {
2431 				sfmmu_shadow_hcleanup(sfmmup, hmeblkp, hmebp);
2432 				ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
2433 				goto ttearray_realloc;
2434 			} else {
2435 				hmeblkp->hblk_shw_bit = 0;
2436 			}
2437 		}
2438 		SFMMU_STAT(sf_hblk_hit);
2439 	}
2440 
2441 	/*
2442 	 * hat_memload() should never call kmem_cache_free(); see block
2443 	 * comment showing the stacktrace in sfmmu_hblk_alloc();
2444 	 * enqueue each hblk in the list to reserve list if it's created
2445 	 * from sfmmu8_cache *and* sfmmup == KHATID.
2446 	 */
2447 	forcefree = (sfmmup == KHATID) ? 1 : 0;
2448 	while ((pr_hblk = list) != NULL) {
2449 		list = pr_hblk->hblk_next;
2450 		sfmmu_cache = get_hblk_cache(pr_hblk);
2451 		if ((sfmmu_cache == sfmmu8_cache) &&
2452 		    sfmmu_put_free_hblk(pr_hblk, forcefree))
2453 			continue;
2454 
2455 		ASSERT(sfmmup != KHATID);
2456 		kmem_cache_free(sfmmu_cache, pr_hblk);
2457 	}
2458 
2459 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
2460 	ASSERT(!hmeblkp->hblk_shw_bit);
2461 
2462 	return (hmeblkp);
2463 }
2464 
2465 /*
2466  * Function adds a tte entry into the hmeblk. It returns 0 if successful and 1
2467  * otherwise.
2468  */
2469 static int
2470 sfmmu_tteload_addentry(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, tte_t *ttep,
2471 	caddr_t vaddr, page_t **pps, uint_t flags)
2472 {
2473 	page_t *pp = *pps;
2474 	int hmenum, size, remap;
2475 	tte_t tteold, flush_tte;
2476 #ifdef DEBUG
2477 	tte_t orig_old;
2478 #endif /* DEBUG */
2479 	struct sf_hment *sfhme;
2480 	kmutex_t *pml, *pmtx;
2481 	hatlock_t *hatlockp;
2482 
2483 	/*
2484 	 * remove this panic when we decide to let user virtual address
2485 	 * space be >= USERLIMIT.
2486 	 */
2487 	if (!TTE_IS_PRIVILEGED(ttep) && vaddr >= (caddr_t)USERLIMIT)
2488 		panic("user addr %p in kernel space", vaddr);
2489 #if defined(TTE_IS_GLOBAL)
2490 	if (TTE_IS_GLOBAL(ttep))
2491 		panic("sfmmu_tteload: creating global tte");
2492 #endif
2493 
2494 #ifdef DEBUG
2495 	if (pf_is_memory(sfmmu_ttetopfn(ttep, vaddr)) &&
2496 	    !TTE_IS_PCACHEABLE(ttep) && !sfmmu_allow_nc_trans)
2497 		panic("sfmmu_tteload: non cacheable memory tte");
2498 #endif /* DEBUG */
2499 
2500 	if ((flags & HAT_LOAD_SHARE) || !TTE_IS_REF(ttep) ||
2501 	    !TTE_IS_MOD(ttep)) {
2502 		/*
2503 		 * Don't load TSB for dummy as in ISM.  Also don't preload
2504 		 * the TSB if the TTE isn't writable since we're likely to
2505 		 * fault on it again -- preloading can be fairly expensive.
2506 		 */
2507 		flags |= SFMMU_NO_TSBLOAD;
2508 	}
2509 
2510 	size = TTE_CSZ(ttep);
2511 	switch (size) {
2512 	case TTE8K:
2513 		SFMMU_STAT(sf_tteload8k);
2514 		break;
2515 	case TTE64K:
2516 		SFMMU_STAT(sf_tteload64k);
2517 		break;
2518 	case TTE512K:
2519 		SFMMU_STAT(sf_tteload512k);
2520 		break;
2521 	case TTE4M:
2522 		SFMMU_STAT(sf_tteload4m);
2523 		break;
2524 	case (TTE32M):
2525 		SFMMU_STAT(sf_tteload32m);
2526 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
2527 		break;
2528 	case (TTE256M):
2529 		SFMMU_STAT(sf_tteload256m);
2530 		ASSERT(mmu_page_sizes == max_mmu_page_sizes);
2531 		break;
2532 	}
2533 
2534 	ASSERT(!((uintptr_t)vaddr & TTE_PAGE_OFFSET(size)));
2535 
2536 	HBLKTOHME_IDX(sfhme, hmeblkp, vaddr, hmenum);
2537 
2538 	/*
2539 	 * Need to grab mlist lock here so that pageunload
2540 	 * will not change tte behind us.
2541 	 */
2542 	if (pp) {
2543 		pml = sfmmu_mlist_enter(pp);
2544 	}
2545 
2546 	sfmmu_copytte(&sfhme->hme_tte, &tteold);
2547 	/*
2548 	 * Look for corresponding hment and if valid verify
2549 	 * pfns are equal.
2550 	 */
2551 	remap = TTE_IS_VALID(&tteold);
2552 	if (remap) {
2553 		pfn_t	new_pfn, old_pfn;
2554 
2555 		old_pfn = TTE_TO_PFN(vaddr, &tteold);
2556 		new_pfn = TTE_TO_PFN(vaddr, ttep);
2557 
2558 		if (flags & HAT_LOAD_REMAP) {
2559 			/* make sure we are remapping same type of pages */
2560 			if (pf_is_memory(old_pfn) != pf_is_memory(new_pfn)) {
2561 				panic("sfmmu_tteload - tte remap io<->memory");
2562 			}
2563 			if (old_pfn != new_pfn &&
2564 			    (pp != NULL || sfhme->hme_page != NULL)) {
2565 				panic("sfmmu_tteload - tte remap pp != NULL");
2566 			}
2567 		} else if (old_pfn != new_pfn) {
2568 			panic("sfmmu_tteload - tte remap, hmeblkp 0x%p",
2569 			    (void *)hmeblkp);
2570 		}
2571 		ASSERT(TTE_CSZ(&tteold) == TTE_CSZ(ttep));
2572 	}
2573 
2574 	if (pp) {
2575 		if (size == TTE8K) {
2576 			/*
2577 			 * Handle VAC consistency
2578 			 */
2579 			if (!remap && (cache & CACHE_VAC) && !PP_ISNC(pp)) {
2580 				sfmmu_vac_conflict(sfmmup, vaddr, pp);
2581 			}
2582 
2583 			if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
2584 				pmtx = sfmmu_page_enter(pp);
2585 				PP_CLRRO(pp);
2586 				sfmmu_page_exit(pmtx);
2587 			} else if (!PP_ISMAPPED(pp) &&
2588 			    (!TTE_IS_WRITABLE(ttep)) && !(PP_ISMOD(pp))) {
2589 				pmtx = sfmmu_page_enter(pp);
2590 				if (!(PP_ISMOD(pp))) {
2591 					PP_SETRO(pp);
2592 				}
2593 				sfmmu_page_exit(pmtx);
2594 			}
2595 
2596 		} else if (sfmmu_pagearray_setup(vaddr, pps, ttep, remap)) {
2597 			/*
2598 			 * sfmmu_pagearray_setup failed so return
2599 			 */
2600 			sfmmu_mlist_exit(pml);
2601 			return (1);
2602 		}
2603 	}
2604 
2605 	/*
2606 	 * Make sure hment is not on a mapping list.
2607 	 */
2608 	ASSERT(remap || (sfhme->hme_page == NULL));
2609 
2610 	/* if it is not a remap then hme->next better be NULL */
2611 	ASSERT((!remap) ? sfhme->hme_next == NULL : 1);
2612 
2613 	if (flags & HAT_LOAD_LOCK) {
2614 		if (((int)hmeblkp->hblk_lckcnt + 1) >= MAX_HBLK_LCKCNT) {
2615 			panic("too high lckcnt-hmeblk %p",
2616 			    (void *)hmeblkp);
2617 		}
2618 		atomic_add_16(&hmeblkp->hblk_lckcnt, 1);
2619 
2620 		HBLK_STACK_TRACE(hmeblkp, HBLK_LOCK);
2621 	}
2622 
2623 	if (pp && PP_ISNC(pp)) {
2624 		/*
2625 		 * If the physical page is marked to be uncacheable, like
2626 		 * by a vac conflict, make sure the new mapping is also
2627 		 * uncacheable.
2628 		 */
2629 		TTE_CLR_VCACHEABLE(ttep);
2630 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
2631 	}
2632 	ttep->tte_hmenum = hmenum;
2633 
2634 #ifdef DEBUG
2635 	orig_old = tteold;
2636 #endif /* DEBUG */
2637 
2638 	while (sfmmu_modifytte_try(&tteold, ttep, &sfhme->hme_tte) < 0) {
2639 		if ((sfmmup == KHATID) &&
2640 		    (flags & (HAT_LOAD_LOCK | HAT_LOAD_REMAP))) {
2641 			sfmmu_copytte(&sfhme->hme_tte, &tteold);
2642 		}
2643 #ifdef DEBUG
2644 		chk_tte(&orig_old, &tteold, ttep, hmeblkp);
2645 #endif /* DEBUG */
2646 	}
2647 
2648 	if (!TTE_IS_VALID(&tteold)) {
2649 
2650 		atomic_add_16(&hmeblkp->hblk_vcnt, 1);
2651 		atomic_add_long(&sfmmup->sfmmu_ttecnt[size], 1);
2652 
2653 		/*
2654 		 * HAT_RELOAD_SHARE has been deprecated with lpg DISM.
2655 		 */
2656 
2657 		if (size > TTE8K && (flags & HAT_LOAD_SHARE) == 0 &&
2658 		    sfmmup != ksfmmup) {
2659 			/*
2660 			 * If this is the first large mapping for the process
2661 			 * we must force any CPUs running this process to TL=0
2662 			 * where they will reload the HAT flags from the
2663 			 * tsbmiss area.  This is necessary to make the large
2664 			 * mappings we are about to load visible to those CPUs;
2665 			 * otherwise they'll loop forever calling pagefault()
2666 			 * since we don't search large hash chains by default.
2667 			 */
2668 			hatlockp = sfmmu_hat_enter(sfmmup);
2669 			if (size == TTE512K &&
2670 			    !SFMMU_FLAGS_ISSET(sfmmup, HAT_512K_FLAG)) {
2671 				SFMMU_FLAGS_SET(sfmmup, HAT_512K_FLAG);
2672 				sfmmu_sync_mmustate(sfmmup);
2673 			} else if (size == TTE4M &&
2674 			    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4M_FLAG)) {
2675 				SFMMU_FLAGS_SET(sfmmup, HAT_4M_FLAG);
2676 				sfmmu_sync_mmustate(sfmmup);
2677 			} else if (size == TTE64K &&
2678 			    !SFMMU_FLAGS_ISSET(sfmmup, HAT_64K_FLAG)) {
2679 				SFMMU_FLAGS_SET(sfmmup, HAT_64K_FLAG);
2680 				/* no sync mmustate; 64K shares 8K hashes */
2681 			} else if (mmu_page_sizes == max_mmu_page_sizes) {
2682 			    if (size == TTE32M &&
2683 				!SFMMU_FLAGS_ISSET(sfmmup, HAT_32M_FLAG)) {
2684 				SFMMU_FLAGS_SET(sfmmup, HAT_32M_FLAG);
2685 				sfmmu_sync_mmustate(sfmmup);
2686 			    } else if (size == TTE256M &&
2687 				!SFMMU_FLAGS_ISSET(sfmmup, HAT_256M_FLAG)) {
2688 				SFMMU_FLAGS_SET(sfmmup, HAT_256M_FLAG);
2689 				sfmmu_sync_mmustate(sfmmup);
2690 			    }
2691 			}
2692 			if (size >= TTE4M && (flags & HAT_LOAD_TEXT) &&
2693 			    !SFMMU_FLAGS_ISSET(sfmmup, HAT_4MTEXT_FLAG)) {
2694 				SFMMU_FLAGS_SET(sfmmup, HAT_4MTEXT_FLAG);
2695 			}
2696 			sfmmu_hat_exit(hatlockp);
2697 		}
2698 	}
2699 	ASSERT(TTE_IS_VALID(&sfhme->hme_tte));
2700 
2701 	flush_tte.tte_intlo = (tteold.tte_intlo ^ ttep->tte_intlo) &
2702 	    hw_tte.tte_intlo;
2703 	flush_tte.tte_inthi = (tteold.tte_inthi ^ ttep->tte_inthi) &
2704 	    hw_tte.tte_inthi;
2705 
2706 	if (remap && (flush_tte.tte_inthi || flush_tte.tte_intlo)) {
2707 		/*
2708 		 * If remap and new tte differs from old tte we need
2709 		 * to sync the mod bit and flush TLB/TSB.  We don't
2710 		 * need to sync ref bit because we currently always set
2711 		 * ref bit in tteload.
2712 		 */
2713 		ASSERT(TTE_IS_REF(ttep));
2714 		if (TTE_IS_MOD(&tteold)) {
2715 			sfmmu_ttesync(sfmmup, vaddr, &tteold, pp);
2716 		}
2717 		sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 0);
2718 		xt_sync(sfmmup->sfmmu_cpusran);
2719 	}
2720 
2721 	if ((flags & SFMMU_NO_TSBLOAD) == 0) {
2722 		/*
2723 		 * We only preload 8K and 4M mappings into the TSB, since
2724 		 * 64K and 512K mappings are replicated and hence don't
2725 		 * have a single, unique TSB entry. Ditto for 32M/256M.
2726 		 */
2727 		if (size == TTE8K || size == TTE4M) {
2728 			hatlockp = sfmmu_hat_enter(sfmmup);
2729 			sfmmu_load_tsb(sfmmup, vaddr, &sfhme->hme_tte, size);
2730 			sfmmu_hat_exit(hatlockp);
2731 		}
2732 	}
2733 	if (pp) {
2734 		if (!remap) {
2735 			HME_ADD(sfhme, pp);
2736 			atomic_add_16(&hmeblkp->hblk_hmecnt, 1);
2737 			ASSERT(hmeblkp->hblk_hmecnt > 0);
2738 
2739 			/*
2740 			 * Cannot ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
2741 			 * see pageunload() for comment.
2742 			 */
2743 		}
2744 		sfmmu_mlist_exit(pml);
2745 	}
2746 
2747 	return (0);
2748 }
2749 /*
2750  * Function unlocks hash bucket.
2751  */
2752 static void
2753 sfmmu_tteload_release_hashbucket(struct hmehash_bucket *hmebp)
2754 {
2755 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
2756 	SFMMU_HASH_UNLOCK(hmebp);
2757 }
2758 
2759 /*
2760  * function which checks and sets up page array for a large
2761  * translation.  Will set p_vcolor, p_index, p_ro fields.
2762  * Assumes addr and pfnum of first page are properly aligned.
2763  * Will check for physical contiguity. If check fails it return
2764  * non null.
2765  */
2766 static int
2767 sfmmu_pagearray_setup(caddr_t addr, page_t **pps, tte_t *ttep, int remap)
2768 {
2769 	int 	i, index, ttesz, osz;
2770 	pfn_t	pfnum;
2771 	pgcnt_t	npgs;
2772 	int cflags = 0;
2773 	page_t *pp, *pp1;
2774 	kmutex_t *pmtx;
2775 	int vac_err = 0;
2776 	int newidx = 0;
2777 
2778 	ttesz = TTE_CSZ(ttep);
2779 
2780 	ASSERT(ttesz > TTE8K);
2781 
2782 	npgs = TTEPAGES(ttesz);
2783 	index = PAGESZ_TO_INDEX(ttesz);
2784 
2785 	pfnum = (*pps)->p_pagenum;
2786 	ASSERT(IS_P2ALIGNED(pfnum, npgs));
2787 
2788 	/*
2789 	 * Save the first pp so we can do HAT_TMPNC at the end.
2790 	 */
2791 	pp1 = *pps;
2792 	osz = fnd_mapping_sz(pp1);
2793 
2794 	for (i = 0; i < npgs; i++, pps++) {
2795 		pp = *pps;
2796 		ASSERT(PAGE_LOCKED(pp));
2797 		ASSERT(pp->p_szc >= ttesz);
2798 		ASSERT(pp->p_szc == pp1->p_szc);
2799 		ASSERT(sfmmu_mlist_held(pp));
2800 
2801 		/*
2802 		 * XXX is it possible to maintain P_RO on the root only?
2803 		 */
2804 		if (TTE_IS_WRITABLE(ttep) && PP_ISRO(pp)) {
2805 			pmtx = sfmmu_page_enter(pp);
2806 			PP_CLRRO(pp);
2807 			sfmmu_page_exit(pmtx);
2808 		} else if (!PP_ISMAPPED(pp) && !TTE_IS_WRITABLE(ttep) &&
2809 		    !PP_ISMOD(pp)) {
2810 			pmtx = sfmmu_page_enter(pp);
2811 			if (!(PP_ISMOD(pp))) {
2812 				PP_SETRO(pp);
2813 			}
2814 			sfmmu_page_exit(pmtx);
2815 		}
2816 
2817 		/*
2818 		 * If this is a remap we skip vac & contiguity checks.
2819 		 */
2820 		if (remap)
2821 			continue;
2822 
2823 		/*
2824 		 * set p_vcolor and detect any vac conflicts.
2825 		 */
2826 		if (vac_err == 0) {
2827 			vac_err = sfmmu_vacconflict_array(addr, pp, &cflags);
2828 
2829 		}
2830 
2831 		/*
2832 		 * Save current index in case we need to undo it.
2833 		 * Note: "PAGESZ_TO_INDEX(sz)	(1 << (sz))"
2834 		 *	"SFMMU_INDEX_SHIFT	6"
2835 		 *	 "SFMMU_INDEX_MASK	((1 << SFMMU_INDEX_SHIFT) - 1)"
2836 		 *	 "PP_MAPINDEX(p_index)	(p_index & SFMMU_INDEX_MASK)"
2837 		 *
2838 		 * So:	index = PAGESZ_TO_INDEX(ttesz);
2839 		 *	if ttesz == 1 then index = 0x2
2840 		 *		    2 then index = 0x4
2841 		 *		    3 then index = 0x8
2842 		 *		    4 then index = 0x10
2843 		 *		    5 then index = 0x20
2844 		 * The code below checks if it's a new pagesize (ie, newidx)
2845 		 * in case we need to take it back out of p_index,
2846 		 * and then or's the new index into the existing index.
2847 		 */
2848 		if ((PP_MAPINDEX(pp) & index) == 0)
2849 			newidx = 1;
2850 		pp->p_index = (PP_MAPINDEX(pp) | index);
2851 
2852 		/*
2853 		 * contiguity check
2854 		 */
2855 		if (pp->p_pagenum != pfnum) {
2856 			/*
2857 			 * If we fail the contiguity test then
2858 			 * the only thing we need to fix is the p_index field.
2859 			 * We might get a few extra flushes but since this
2860 			 * path is rare that is ok.  The p_ro field will
2861 			 * get automatically fixed on the next tteload to
2862 			 * the page.  NO TNC bit is set yet.
2863 			 */
2864 			while (i >= 0) {
2865 				pp = *pps;
2866 				if (newidx)
2867 					pp->p_index = (PP_MAPINDEX(pp) &
2868 					    ~index);
2869 				pps--;
2870 				i--;
2871 			}
2872 			return (1);
2873 		}
2874 		pfnum++;
2875 		addr += MMU_PAGESIZE;
2876 	}
2877 
2878 	if (vac_err) {
2879 		if (ttesz > osz) {
2880 			/*
2881 			 * There are some smaller mappings that causes vac
2882 			 * conflicts. Convert all existing small mappings to
2883 			 * TNC.
2884 			 */
2885 			SFMMU_STAT_ADD(sf_uncache_conflict, npgs);
2886 			sfmmu_page_cache_array(pp1, HAT_TMPNC, CACHE_FLUSH,
2887 				npgs);
2888 		} else {
2889 			/* EMPTY */
2890 			/*
2891 			 * If there exists an big page mapping,
2892 			 * that means the whole existing big page
2893 			 * has TNC setting already. No need to covert to
2894 			 * TNC again.
2895 			 */
2896 			ASSERT(PP_ISTNC(pp1));
2897 		}
2898 	}
2899 
2900 	return (0);
2901 }
2902 
2903 /*
2904  * Routine that detects vac consistency for a large page. It also
2905  * sets virtual color for all pp's for this big mapping.
2906  */
2907 static int
2908 sfmmu_vacconflict_array(caddr_t addr, page_t *pp, int *cflags)
2909 {
2910 	int vcolor, ocolor;
2911 
2912 	ASSERT(sfmmu_mlist_held(pp));
2913 
2914 	if (PP_ISNC(pp)) {
2915 		return (HAT_TMPNC);
2916 	}
2917 
2918 	vcolor = addr_to_vcolor(addr);
2919 	if (PP_NEWPAGE(pp)) {
2920 		PP_SET_VCOLOR(pp, vcolor);
2921 		return (0);
2922 	}
2923 
2924 	ocolor = PP_GET_VCOLOR(pp);
2925 	if (ocolor == vcolor) {
2926 		return (0);
2927 	}
2928 
2929 	if (!PP_ISMAPPED(pp)) {
2930 		/*
2931 		 * Previous user of page had a differnet color
2932 		 * but since there are no current users
2933 		 * we just flush the cache and change the color.
2934 		 * As an optimization for large pages we flush the
2935 		 * entire cache of that color and set a flag.
2936 		 */
2937 		SFMMU_STAT(sf_pgcolor_conflict);
2938 		if (!CacheColor_IsFlushed(*cflags, ocolor)) {
2939 			CacheColor_SetFlushed(*cflags, ocolor);
2940 			sfmmu_cache_flushcolor(ocolor, pp->p_pagenum);
2941 		}
2942 		PP_SET_VCOLOR(pp, vcolor);
2943 		return (0);
2944 	}
2945 
2946 	/*
2947 	 * We got a real conflict with a current mapping.
2948 	 * set flags to start unencaching all mappings
2949 	 * and return failure so we restart looping
2950 	 * the pp array from the beginning.
2951 	 */
2952 	return (HAT_TMPNC);
2953 }
2954 
2955 /*
2956  * creates a large page shadow hmeblk for a tte.
2957  * The purpose of this routine is to allow us to do quick unloads because
2958  * the vm layer can easily pass a very large but sparsely populated range.
2959  */
2960 static struct hme_blk *
2961 sfmmu_shadow_hcreate(sfmmu_t *sfmmup, caddr_t vaddr, int ttesz, uint_t flags)
2962 {
2963 	struct hmehash_bucket *hmebp;
2964 	hmeblk_tag hblktag;
2965 	int hmeshift, size, vshift;
2966 	uint_t shw_mask, newshw_mask;
2967 	struct hme_blk *hmeblkp;
2968 
2969 	ASSERT(sfmmup != KHATID);
2970 	if (mmu_page_sizes == max_mmu_page_sizes) {
2971 		ASSERT(ttesz < TTE256M);
2972 	} else {
2973 		ASSERT(ttesz < TTE4M);
2974 		ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
2975 		ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
2976 	}
2977 
2978 	if (ttesz == TTE8K) {
2979 		size = TTE512K;
2980 	} else {
2981 		size = ++ttesz;
2982 	}
2983 
2984 	hblktag.htag_id = sfmmup;
2985 	hmeshift = HME_HASH_SHIFT(size);
2986 	hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
2987 	hblktag.htag_rehash = HME_HASH_REHASH(size);
2988 	hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
2989 
2990 	SFMMU_HASH_LOCK(hmebp);
2991 
2992 	HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
2993 	ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
2994 	if (hmeblkp == NULL) {
2995 		hmeblkp = sfmmu_hblk_alloc(sfmmup, vaddr, hmebp, size,
2996 			hblktag, flags);
2997 	}
2998 	ASSERT(hmeblkp);
2999 	if (!hmeblkp->hblk_shw_mask) {
3000 		/*
3001 		 * if this is a unused hblk it was just allocated or could
3002 		 * potentially be a previous large page hblk so we need to
3003 		 * set the shadow bit.
3004 		 */
3005 		hmeblkp->hblk_shw_bit = 1;
3006 	}
3007 	ASSERT(hmeblkp->hblk_shw_bit == 1);
3008 	vshift = vaddr_to_vshift(hblktag, vaddr, size);
3009 	ASSERT(vshift < 8);
3010 	/*
3011 	 * Atomically set shw mask bit
3012 	 */
3013 	do {
3014 		shw_mask = hmeblkp->hblk_shw_mask;
3015 		newshw_mask = shw_mask | (1 << vshift);
3016 		newshw_mask = cas32(&hmeblkp->hblk_shw_mask, shw_mask,
3017 		    newshw_mask);
3018 	} while (newshw_mask != shw_mask);
3019 
3020 	SFMMU_HASH_UNLOCK(hmebp);
3021 
3022 	return (hmeblkp);
3023 }
3024 
3025 /*
3026  * This routine cleanup a previous shadow hmeblk and changes it to
3027  * a regular hblk.  This happens rarely but it is possible
3028  * when a process wants to use large pages and there are hblks still
3029  * lying around from the previous as that used these hmeblks.
3030  * The alternative was to cleanup the shadow hblks at unload time
3031  * but since so few user processes actually use large pages, it is
3032  * better to be lazy and cleanup at this time.
3033  */
3034 static void
3035 sfmmu_shadow_hcleanup(sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
3036 	struct hmehash_bucket *hmebp)
3037 {
3038 	caddr_t addr, endaddr;
3039 	int hashno, size;
3040 
3041 	ASSERT(hmeblkp->hblk_shw_bit);
3042 
3043 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
3044 
3045 	if (!hmeblkp->hblk_shw_mask) {
3046 		hmeblkp->hblk_shw_bit = 0;
3047 		return;
3048 	}
3049 	addr = (caddr_t)get_hblk_base(hmeblkp);
3050 	endaddr = get_hblk_endaddr(hmeblkp);
3051 	size = get_hblk_ttesz(hmeblkp);
3052 	hashno = size - 1;
3053 	ASSERT(hashno > 0);
3054 	SFMMU_HASH_UNLOCK(hmebp);
3055 
3056 	sfmmu_free_hblks(sfmmup, addr, endaddr, hashno);
3057 
3058 	SFMMU_HASH_LOCK(hmebp);
3059 }
3060 
3061 static void
3062 sfmmu_free_hblks(sfmmu_t *sfmmup, caddr_t addr, caddr_t endaddr,
3063 	int hashno)
3064 {
3065 	int hmeshift, shadow = 0;
3066 	hmeblk_tag hblktag;
3067 	struct hmehash_bucket *hmebp;
3068 	struct hme_blk *hmeblkp;
3069 	struct hme_blk *nx_hblk, *pr_hblk, *list = NULL;
3070 	uint64_t hblkpa, prevpa, nx_pa;
3071 
3072 	ASSERT(hashno > 0);
3073 	hblktag.htag_id = sfmmup;
3074 	hblktag.htag_rehash = hashno;
3075 
3076 	hmeshift = HME_HASH_SHIFT(hashno);
3077 
3078 	while (addr < endaddr) {
3079 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3080 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3081 		SFMMU_HASH_LOCK(hmebp);
3082 		/* inline HME_HASH_SEARCH */
3083 		hmeblkp = hmebp->hmeblkp;
3084 		hblkpa = hmebp->hmeh_nextpa;
3085 		prevpa = 0;
3086 		pr_hblk = NULL;
3087 		while (hmeblkp) {
3088 			ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
3089 			if (HTAGS_EQ(hmeblkp->hblk_tag, hblktag)) {
3090 				/* found hme_blk */
3091 				if (hmeblkp->hblk_shw_bit) {
3092 					if (hmeblkp->hblk_shw_mask) {
3093 						shadow = 1;
3094 						sfmmu_shadow_hcleanup(sfmmup,
3095 						    hmeblkp, hmebp);
3096 						break;
3097 					} else {
3098 						hmeblkp->hblk_shw_bit = 0;
3099 					}
3100 				}
3101 
3102 				/*
3103 				 * Hblk_hmecnt and hblk_vcnt could be non zero
3104 				 * since hblk_unload() does not gurantee that.
3105 				 *
3106 				 * XXX - this could cause tteload() to spin
3107 				 * where sfmmu_shadow_hcleanup() is called.
3108 				 */
3109 			}
3110 
3111 			nx_hblk = hmeblkp->hblk_next;
3112 			nx_pa = hmeblkp->hblk_nextpa;
3113 			if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
3114 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa,
3115 					pr_hblk);
3116 				sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
3117 			} else {
3118 				pr_hblk = hmeblkp;
3119 				prevpa = hblkpa;
3120 			}
3121 			hmeblkp = nx_hblk;
3122 			hblkpa = nx_pa;
3123 		}
3124 
3125 		SFMMU_HASH_UNLOCK(hmebp);
3126 
3127 		if (shadow) {
3128 			/*
3129 			 * We found another shadow hblk so cleaned its
3130 			 * children.  We need to go back and cleanup
3131 			 * the original hblk so we don't change the
3132 			 * addr.
3133 			 */
3134 			shadow = 0;
3135 		} else {
3136 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
3137 				(1 << hmeshift));
3138 		}
3139 	}
3140 	sfmmu_hblks_list_purge(&list);
3141 }
3142 
3143 /*
3144  * Release one hardware address translation lock on the given address range.
3145  */
3146 void
3147 hat_unlock(struct hat *sfmmup, caddr_t addr, size_t len)
3148 {
3149 	struct hmehash_bucket *hmebp;
3150 	hmeblk_tag hblktag;
3151 	int hmeshift, hashno = 1;
3152 	struct hme_blk *hmeblkp, *list = NULL;
3153 	caddr_t endaddr;
3154 
3155 	ASSERT(sfmmup != NULL);
3156 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3157 
3158 	ASSERT((sfmmup == ksfmmup) ||
3159 		AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
3160 	ASSERT((len & MMU_PAGEOFFSET) == 0);
3161 	endaddr = addr + len;
3162 	hblktag.htag_id = sfmmup;
3163 
3164 	/*
3165 	 * Spitfire supports 4 page sizes.
3166 	 * Most pages are expected to be of the smallest page size (8K) and
3167 	 * these will not need to be rehashed. 64K pages also don't need to be
3168 	 * rehashed because an hmeblk spans 64K of address space. 512K pages
3169 	 * might need 1 rehash and and 4M pages might need 2 rehashes.
3170 	 */
3171 	while (addr < endaddr) {
3172 		hmeshift = HME_HASH_SHIFT(hashno);
3173 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3174 		hblktag.htag_rehash = hashno;
3175 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3176 
3177 		SFMMU_HASH_LOCK(hmebp);
3178 
3179 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
3180 		if (hmeblkp != NULL) {
3181 			/*
3182 			 * If we encounter a shadow hmeblk then
3183 			 * we know there are no valid hmeblks mapping
3184 			 * this address at this size or larger.
3185 			 * Just increment address by the smallest
3186 			 * page size.
3187 			 */
3188 			if (hmeblkp->hblk_shw_bit) {
3189 				addr += MMU_PAGESIZE;
3190 			} else {
3191 				addr = sfmmu_hblk_unlock(hmeblkp, addr,
3192 				    endaddr);
3193 			}
3194 			SFMMU_HASH_UNLOCK(hmebp);
3195 			hashno = 1;
3196 			continue;
3197 		}
3198 		SFMMU_HASH_UNLOCK(hmebp);
3199 
3200 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
3201 			/*
3202 			 * We have traversed the whole list and rehashed
3203 			 * if necessary without finding the address to unlock
3204 			 * which should never happen.
3205 			 */
3206 			panic("sfmmu_unlock: addr not found. "
3207 			    "addr %p hat %p", (void *)addr, (void *)sfmmup);
3208 		} else {
3209 			hashno++;
3210 		}
3211 	}
3212 
3213 	sfmmu_hblks_list_purge(&list);
3214 }
3215 
3216 /*
3217  * Function to unlock a range of addresses in an hmeblk.  It returns the
3218  * next address that needs to be unlocked.
3219  * Should be called with the hash lock held.
3220  */
3221 static caddr_t
3222 sfmmu_hblk_unlock(struct hme_blk *hmeblkp, caddr_t addr, caddr_t endaddr)
3223 {
3224 	struct sf_hment *sfhme;
3225 	tte_t tteold, ttemod;
3226 	int ttesz, ret;
3227 
3228 	ASSERT(in_hblk_range(hmeblkp, addr));
3229 	ASSERT(hmeblkp->hblk_shw_bit == 0);
3230 
3231 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
3232 	ttesz = get_hblk_ttesz(hmeblkp);
3233 
3234 	HBLKTOHME(sfhme, hmeblkp, addr);
3235 	while (addr < endaddr) {
3236 readtte:
3237 		sfmmu_copytte(&sfhme->hme_tte, &tteold);
3238 		if (TTE_IS_VALID(&tteold)) {
3239 
3240 			ttemod = tteold;
3241 
3242 			ret = sfmmu_modifytte_try(&tteold, &ttemod,
3243 			    &sfhme->hme_tte);
3244 
3245 			if (ret < 0)
3246 				goto readtte;
3247 
3248 			if (hmeblkp->hblk_lckcnt == 0)
3249 				panic("zero hblk lckcnt");
3250 
3251 			if (((uintptr_t)addr + TTEBYTES(ttesz)) >
3252 			    (uintptr_t)endaddr)
3253 				panic("can't unlock large tte");
3254 
3255 			ASSERT(hmeblkp->hblk_lckcnt > 0);
3256 			atomic_add_16(&hmeblkp->hblk_lckcnt, -1);
3257 			HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
3258 		} else {
3259 			panic("sfmmu_hblk_unlock: invalid tte");
3260 		}
3261 		addr += TTEBYTES(ttesz);
3262 		sfhme++;
3263 	}
3264 	return (addr);
3265 }
3266 
3267 /*
3268  * Physical Address Mapping Framework
3269  *
3270  * General rules:
3271  *
3272  * (1) Applies only to seg_kmem memory pages. To make things easier,
3273  *     seg_kpm addresses are also accepted by the routines, but nothing
3274  *     is done with them since by definition their PA mappings are static.
3275  * (2) hat_add_callback() may only be called while holding the page lock
3276  *     SE_SHARED or SE_EXCL of the underlying page (e.g., as_pagelock()).
3277  * (3) prehandler() and posthandler() may not call hat_add_callback() or
3278  *     hat_delete_callback(), nor should they allocate memory. Post quiesce
3279  *     callbacks may not sleep or acquire adaptive mutex locks.
3280  * (4) Either prehandler() or posthandler() (but not both) may be specified
3281  *     as being NULL.  Specifying an errhandler() is optional.
3282  *
3283  * Details of using the framework:
3284  *
3285  * registering a callback (hat_register_callback())
3286  *
3287  *	Pass prehandler, posthandler, errhandler addresses
3288  *	as described below. If capture_cpus argument is nonzero,
3289  *	suspend callback to the prehandler will occur with CPUs
3290  *	captured and executing xc_loop() and CPUs will remain
3291  *	captured until after the posthandler suspend callback
3292  *	occurs.
3293  *
3294  * adding a callback (hat_add_callback())
3295  *
3296  *      as_pagelock();
3297  *	hat_add_callback();
3298  *      save returned pfn in private data structures or program registers;
3299  *      as_pageunlock();
3300  *
3301  * prehandler()
3302  *
3303  *	Stop all accesses by physical address to this memory page.
3304  *	Called twice: the first, PRESUSPEND, is a context safe to acquire
3305  *	adaptive locks. The second, SUSPEND, is called at high PIL with
3306  *	CPUs captured so adaptive locks may NOT be acquired (and all spin
3307  *	locks must be XCALL_PIL or higher locks).
3308  *
3309  *	May return the following errors:
3310  *		EIO:	A fatal error has occurred. This will result in panic.
3311  *		EAGAIN:	The page cannot be suspended. This will fail the
3312  *			relocation.
3313  *		0:	Success.
3314  *
3315  * posthandler()
3316  *
3317  *      Save new pfn in private data structures or program registers;
3318  *	not allowed to fail (non-zero return values will result in panic).
3319  *
3320  * errhandler()
3321  *
3322  *	called when an error occurs related to the callback.  Currently
3323  *	the only such error is HAT_CB_ERR_LEAKED which indicates that
3324  *	a page is being freed, but there are still outstanding callback(s)
3325  *	registered on the page.
3326  *
3327  * removing a callback (hat_delete_callback(); e.g., prior to freeing memory)
3328  *
3329  *	stop using physical address
3330  *	hat_delete_callback();
3331  *
3332  */
3333 
3334 /*
3335  * Register a callback class.  Each subsystem should do this once and
3336  * cache the id_t returned for use in setting up and tearing down callbacks.
3337  * There is no facility for removing callback IDs once they are created,
3338  * so this should only be called from modules which cannot be unloaded.
3339  */
3340 id_t
3341 hat_register_callback(int (*prehandler)(caddr_t, uint_t, uint_t, void *),
3342 	int (*posthandler)(caddr_t, uint_t, uint_t, void *, pfn_t),
3343 	int (*errhandler)(caddr_t, uint_t, uint_t, void *),
3344 	int capture_cpus)
3345 {
3346 	id_t id;
3347 
3348 	/*
3349 	 * If this callback has already been registered just return the
3350 	 * ID for it.
3351 	 */
3352 	for (id = 0; id < sfmmu_max_cb_id; id++) {
3353 		if (sfmmu_cb_table[id].prehandler == prehandler &&
3354 		    sfmmu_cb_table[id].posthandler == posthandler &&
3355 		    sfmmu_cb_table[id].errhandler == errhandler &&
3356 		    sfmmu_cb_table[id].capture_cpus == capture_cpus) {
3357 			return (id);
3358 		}
3359 	}
3360 
3361 	id = sfmmu_cb_nextid++;
3362 
3363 	ASSERT(prehandler != NULL || posthandler != NULL);
3364 
3365 	if (id >= sfmmu_max_cb_id)
3366 		panic("hat_register_callback: out of callback IDs");
3367 
3368 	sfmmu_cb_table[id].prehandler = prehandler;
3369 	sfmmu_cb_table[id].posthandler = posthandler;
3370 	sfmmu_cb_table[id].errhandler = errhandler;
3371 	sfmmu_cb_table[id].capture_cpus = capture_cpus;
3372 
3373 	return (id);
3374 }
3375 
3376 /*
3377  * Add relocation callbacks to the specified addr/len which will be called
3378  * when relocating the associated page.  See the description of pre and
3379  * posthandler above for more details.  IMPT: this operation is only valid
3380  * on seg_kmem pages!!
3381  *
3382  * If HAC_PAGELOCK is included in flags, the underlying memory page is
3383  * locked internally so the caller must be able to deal with the callback
3384  * running even before this function has returned.  If HAC_PAGELOCK is not
3385  * set, it is assumed that the underlying memory pages are locked.
3386  *
3387  * Since the caller must track the individual page boundaries anyway,
3388  * we only allow a callback to be added to a single page (large
3389  * or small).  Thus [addr, addr + len) MUST be contained within a single
3390  * page.
3391  *
3392  * Registering multiple callbacks on the same [addr, addr+len) is supported,
3393  * in which case the corresponding callback will be called once with each
3394  * unique parameter specified. The number of subsequent deletes must match
3395  * since reference counts are held.  If a callback is desired for each
3396  * virtual object with the same parameter specified for multiple callbacks,
3397  * a different virtual address should be specified at the time of
3398  * callback registration.
3399  *
3400  * Returns the pfn of the underlying kernel page in *rpfn
3401  * on success, or PFN_INVALID on failure.
3402  *
3403  * Returns values:
3404  *    0:      success
3405  *    ENOMEM: memory allocation failure (e.g. flags was passed as HAC_NOSLEEP)
3406  *    EINVAL: callback ID is not valid
3407  *    ENXIO:  ["vaddr", "vaddr" + len) is not mapped in the kernel's address
3408  *            space, or crosses a page boundary
3409  */
3410 int
3411 hat_add_callback(id_t callback_id, caddr_t vaddr, uint_t len, uint_t flags,
3412 	void *pvt, pfn_t *rpfn)
3413 {
3414 	struct 		hmehash_bucket *hmebp;
3415 	hmeblk_tag 	hblktag;
3416 	struct hme_blk	*hmeblkp;
3417 	int 		hmeshift, hashno;
3418 	caddr_t 	saddr, eaddr, baseaddr;
3419 	struct pa_hment *pahmep, *tpahmep;
3420 	struct sf_hment *sfhmep, *osfhmep, *tsfhmep;
3421 	kmutex_t	*pml;
3422 	tte_t   	tte;
3423 	page_t		*pp, *rpp;
3424 	pfn_t		pfn;
3425 	int		kmflags = (flags & HAC_SLEEP)? KM_SLEEP : KM_NOSLEEP;
3426 	int		locked = 0;
3427 
3428 	/*
3429 	 * For KPM mappings, just return the physical address since we
3430 	 * don't need to register any callbacks.
3431 	 */
3432 	if (IS_KPM_ADDR(vaddr)) {
3433 		uint64_t paddr;
3434 		SFMMU_KPM_VTOP(vaddr, paddr);
3435 		*rpfn = btop(paddr);
3436 		return (0);
3437 	}
3438 
3439 	if (callback_id < (id_t)0 || callback_id >= sfmmu_cb_nextid) {
3440 		*rpfn = PFN_INVALID;
3441 		return (EINVAL);
3442 	}
3443 
3444 	if ((pahmep = kmem_cache_alloc(pa_hment_cache, kmflags)) == NULL) {
3445 		*rpfn = PFN_INVALID;
3446 		return (ENOMEM);
3447 	}
3448 
3449 	sfhmep = &pahmep->sfment;
3450 
3451 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
3452 	eaddr = saddr + len;
3453 
3454 rehash:
3455 	/* Find the mapping(s) for this page */
3456 	for (hashno = TTE64K, hmeblkp = NULL;
3457 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
3458 	    hashno++) {
3459 		hmeshift = HME_HASH_SHIFT(hashno);
3460 		hblktag.htag_id = ksfmmup;
3461 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
3462 		hblktag.htag_rehash = hashno;
3463 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
3464 
3465 		SFMMU_HASH_LOCK(hmebp);
3466 
3467 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
3468 
3469 		if (hmeblkp == NULL)
3470 			SFMMU_HASH_UNLOCK(hmebp);
3471 	}
3472 
3473 	if (hmeblkp == NULL) {
3474 		*rpfn = PFN_INVALID;
3475 		return (ENXIO);
3476 	}
3477 
3478 	/*
3479 	 * Make sure the boundaries for the callback fall within this
3480 	 * single mapping.
3481 	 */
3482 	baseaddr = (caddr_t)get_hblk_base(hmeblkp);
3483 	ASSERT(saddr >= baseaddr);
3484 	if (eaddr > (caddr_t)get_hblk_endaddr(hmeblkp)) {
3485 		SFMMU_HASH_UNLOCK(hmebp);
3486 		*rpfn = PFN_INVALID;
3487 		return (ENXIO);
3488 	}
3489 
3490 	HBLKTOHME(osfhmep, hmeblkp, saddr);
3491 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
3492 
3493 	ASSERT(TTE_IS_VALID(&tte));
3494 	pfn = sfmmu_ttetopfn(&tte, vaddr);
3495 
3496 	pp = osfhmep->hme_page;
3497 	pml = sfmmu_mlist_enter(pp);
3498 
3499 	if ((flags & HAC_PAGELOCK) && !locked) {
3500 		if (!page_trylock(pp, SE_SHARED)) {
3501 			/*
3502 			 * Somebody is holding SE_EXCL lock.  Drop all
3503 			 * our locks, lookup the page in &kvp, and
3504 			 * retry.
3505 			 */
3506 			sfmmu_mlist_exit(pml);
3507 			SFMMU_HASH_UNLOCK(hmebp);
3508 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
3509 			ASSERT(pp != NULL);
3510 			rpp = PP_PAGEROOT(pp);
3511 			if (rpp != pp) {
3512 				page_unlock(pp);
3513 				(void) page_lock(rpp, SE_SHARED, NULL,
3514 				    P_NO_RECLAIM);
3515 			}
3516 			locked = 1;
3517 			goto rehash;
3518 		}
3519 		locked = 1;
3520 	}
3521 
3522 	if (!PAGE_LOCKED(pp) && !panicstr)
3523 		panic("hat_add_callback: page 0x%p not locked", pp);
3524 
3525 	if (osfhmep->hme_page != pp || pp->p_vnode != &kvp ||
3526 	    pp->p_offset < (u_offset_t)baseaddr ||
3527 	    pp->p_offset > (u_offset_t)eaddr) {
3528 		/*
3529 		 * The page moved before we got our hands on it.  Drop
3530 		 * all the locks and try again.
3531 		 */
3532 		ASSERT((flags & HAC_PAGELOCK) != 0);
3533 		sfmmu_mlist_exit(pml);
3534 		SFMMU_HASH_UNLOCK(hmebp);
3535 		page_unlock(pp);
3536 		locked = 0;
3537 		goto rehash;
3538 	}
3539 
3540 	ASSERT(osfhmep->hme_page == pp);
3541 
3542 	for (tsfhmep = pp->p_mapping; tsfhmep != NULL;
3543 	    tsfhmep = tsfhmep->hme_next) {
3544 
3545 		/*
3546 		 * skip va to pa mappings
3547 		 */
3548 		if (!IS_PAHME(tsfhmep))
3549 			continue;
3550 
3551 		tpahmep = tsfhmep->hme_data;
3552 		ASSERT(tpahmep != NULL);
3553 
3554 		/*
3555 		 * See if the pahment already exists.
3556 		 */
3557 		if ((tpahmep->pvt == pvt) &&
3558 		    (tpahmep->addr == vaddr) &&
3559 		    (tpahmep->len == len)) {
3560 			ASSERT(tpahmep->cb_id == callback_id);
3561 			tpahmep->refcnt++;
3562 			pp->p_share++;
3563 
3564 			sfmmu_mlist_exit(pml);
3565 			SFMMU_HASH_UNLOCK(hmebp);
3566 
3567 			if (locked)
3568 				page_unlock(pp);
3569 
3570 			kmem_cache_free(pa_hment_cache, pahmep);
3571 
3572 			*rpfn = pfn;
3573 			return (0);
3574 		}
3575 	}
3576 
3577 	/*
3578 	 * setup this shiny new pa_hment ..
3579 	 */
3580 	pp->p_share++;
3581 	pahmep->cb_id = callback_id;
3582 	pahmep->addr = vaddr;
3583 	pahmep->len = len;
3584 	pahmep->refcnt = 1;
3585 	pahmep->flags = 0;
3586 	pahmep->pvt = pvt;
3587 
3588 	/*
3589 	 * .. and also set up the sf_hment and link to p_mapping list.
3590 	 */
3591 	sfhmep->hme_tte.ll = 0;
3592 	sfhmep->hme_data = pahmep;
3593 	sfhmep->hme_prev = osfhmep;
3594 	sfhmep->hme_next = osfhmep->hme_next;
3595 
3596 	if (osfhmep->hme_next)
3597 		osfhmep->hme_next->hme_prev = sfhmep;
3598 
3599 	osfhmep->hme_next = sfhmep;
3600 
3601 	sfmmu_mlist_exit(pml);
3602 	SFMMU_HASH_UNLOCK(hmebp);
3603 
3604 	*rpfn = pfn;
3605 	if (locked)
3606 		page_unlock(pp);
3607 
3608 	return (0);
3609 }
3610 
3611 /*
3612  * Remove the relocation callbacks from the specified addr/len.
3613  */
3614 void
3615 hat_delete_callback(caddr_t vaddr, uint_t len, void *pvt, uint_t flags)
3616 {
3617 	struct		hmehash_bucket *hmebp;
3618 	hmeblk_tag	hblktag;
3619 	struct hme_blk	*hmeblkp;
3620 	int		hmeshift, hashno;
3621 	caddr_t		saddr, eaddr, baseaddr;
3622 	struct pa_hment	*pahmep;
3623 	struct sf_hment	*sfhmep, *osfhmep;
3624 	kmutex_t	*pml;
3625 	tte_t		tte;
3626 	page_t		*pp, *rpp;
3627 	int		locked = 0;
3628 
3629 	if (IS_KPM_ADDR(vaddr))
3630 		return;
3631 
3632 	saddr = (caddr_t)((uintptr_t)vaddr & MMU_PAGEMASK);
3633 	eaddr = saddr + len;
3634 
3635 rehash:
3636 	/* Find the mapping(s) for this page */
3637 	for (hashno = TTE64K, hmeblkp = NULL;
3638 	    hmeblkp == NULL && hashno <= mmu_hashcnt;
3639 	    hashno++) {
3640 		hmeshift = HME_HASH_SHIFT(hashno);
3641 		hblktag.htag_id = ksfmmup;
3642 		hblktag.htag_bspage = HME_HASH_BSPAGE(saddr, hmeshift);
3643 		hblktag.htag_rehash = hashno;
3644 		hmebp = HME_HASH_FUNCTION(ksfmmup, saddr, hmeshift);
3645 
3646 		SFMMU_HASH_LOCK(hmebp);
3647 
3648 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
3649 
3650 		if (hmeblkp == NULL)
3651 			SFMMU_HASH_UNLOCK(hmebp);
3652 	}
3653 
3654 	if (hmeblkp == NULL) {
3655 		if (!panicstr) {
3656 			panic("hat_delete_callback: addr 0x%p not found",
3657 			    saddr);
3658 		}
3659 		return;
3660 	}
3661 
3662 	baseaddr = (caddr_t)get_hblk_base(hmeblkp);
3663 	HBLKTOHME(osfhmep, hmeblkp, saddr);
3664 
3665 	sfmmu_copytte(&osfhmep->hme_tte, &tte);
3666 	ASSERT(TTE_IS_VALID(&tte));
3667 
3668 	pp = osfhmep->hme_page;
3669 	pml = sfmmu_mlist_enter(pp);
3670 
3671 	if ((flags & HAC_PAGELOCK) && !locked) {
3672 		if (!page_trylock(pp, SE_SHARED)) {
3673 			/*
3674 			 * Somebody is holding SE_EXCL lock.  Drop all
3675 			 * our locks, lookup the page in &kvp, and
3676 			 * retry.
3677 			 */
3678 			sfmmu_mlist_exit(pml);
3679 			SFMMU_HASH_UNLOCK(hmebp);
3680 			pp = page_lookup(&kvp, (u_offset_t)saddr, SE_SHARED);
3681 			ASSERT(pp != NULL);
3682 			rpp = PP_PAGEROOT(pp);
3683 			if (rpp != pp) {
3684 				page_unlock(pp);
3685 				(void) page_lock(rpp, SE_SHARED, NULL,
3686 				    P_NO_RECLAIM);
3687 			}
3688 			locked = 1;
3689 			goto rehash;
3690 		}
3691 		locked = 1;
3692 	}
3693 
3694 	ASSERT(PAGE_LOCKED(pp));
3695 
3696 	if (osfhmep->hme_page != pp || pp->p_vnode != &kvp ||
3697 	    pp->p_offset < (u_offset_t)baseaddr ||
3698 	    pp->p_offset > (u_offset_t)eaddr) {
3699 		/*
3700 		 * The page moved before we got our hands on it.  Drop
3701 		 * all the locks and try again.
3702 		 */
3703 		ASSERT((flags & HAC_PAGELOCK) != 0);
3704 		sfmmu_mlist_exit(pml);
3705 		SFMMU_HASH_UNLOCK(hmebp);
3706 		page_unlock(pp);
3707 		locked = 0;
3708 		goto rehash;
3709 	}
3710 
3711 	ASSERT(osfhmep->hme_page == pp);
3712 
3713 	for (sfhmep = pp->p_mapping; sfhmep != NULL;
3714 	    sfhmep = sfhmep->hme_next) {
3715 
3716 		/*
3717 		 * skip va<->pa mappings
3718 		 */
3719 		if (!IS_PAHME(sfhmep))
3720 			continue;
3721 
3722 		pahmep = sfhmep->hme_data;
3723 		ASSERT(pahmep != NULL);
3724 
3725 		/*
3726 		 * if pa_hment matches, remove it
3727 		 */
3728 		if ((pahmep->pvt == pvt) &&
3729 		    (pahmep->addr == vaddr) &&
3730 		    (pahmep->len == len)) {
3731 			break;
3732 		}
3733 	}
3734 
3735 	if (sfhmep == NULL) {
3736 		if (!panicstr) {
3737 			panic("hat_delete_callback: pa_hment not found, pp %p",
3738 			    (void *)pp);
3739 		}
3740 		return;
3741 	}
3742 
3743 	/*
3744 	 * Note: at this point a valid kernel mapping must still be
3745 	 * present on this page.
3746 	 */
3747 	pp->p_share--;
3748 	if (pp->p_share <= 0)
3749 		panic("hat_delete_callback: zero p_share");
3750 
3751 	if (--pahmep->refcnt == 0) {
3752 		if (pahmep->flags != 0)
3753 			panic("hat_delete_callback: pa_hment is busy");
3754 
3755 		/*
3756 		 * Remove sfhmep from the mapping list for the page.
3757 		 */
3758 		if (sfhmep->hme_prev) {
3759 			sfhmep->hme_prev->hme_next = sfhmep->hme_next;
3760 		} else {
3761 			pp->p_mapping = sfhmep->hme_next;
3762 		}
3763 
3764 		if (sfhmep->hme_next)
3765 			sfhmep->hme_next->hme_prev = sfhmep->hme_prev;
3766 
3767 		sfmmu_mlist_exit(pml);
3768 		SFMMU_HASH_UNLOCK(hmebp);
3769 
3770 		if (locked)
3771 			page_unlock(pp);
3772 
3773 		kmem_cache_free(pa_hment_cache, pahmep);
3774 		return;
3775 	}
3776 
3777 	sfmmu_mlist_exit(pml);
3778 	SFMMU_HASH_UNLOCK(hmebp);
3779 	if (locked)
3780 		page_unlock(pp);
3781 }
3782 
3783 /*
3784  * hat_probe returns 1 if the translation for the address 'addr' is
3785  * loaded, zero otherwise.
3786  *
3787  * hat_probe should be used only for advisorary purposes because it may
3788  * occasionally return the wrong value. The implementation must guarantee that
3789  * returning the wrong value is a very rare event. hat_probe is used
3790  * to implement optimizations in the segment drivers.
3791  *
3792  */
3793 int
3794 hat_probe(struct hat *sfmmup, caddr_t addr)
3795 {
3796 	pfn_t pfn;
3797 	tte_t tte;
3798 
3799 	ASSERT(sfmmup != NULL);
3800 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3801 
3802 	ASSERT((sfmmup == ksfmmup) ||
3803 		AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
3804 
3805 	if (sfmmup == ksfmmup) {
3806 		while ((pfn = sfmmu_vatopfn(addr, sfmmup, &tte))
3807 		    == PFN_SUSPENDED) {
3808 			sfmmu_vatopfn_suspended(addr, sfmmup, &tte);
3809 		}
3810 	} else {
3811 		pfn = sfmmu_uvatopfn(addr, sfmmup);
3812 	}
3813 
3814 	if (pfn != PFN_INVALID)
3815 		return (1);
3816 	else
3817 		return (0);
3818 }
3819 
3820 ssize_t
3821 hat_getpagesize(struct hat *sfmmup, caddr_t addr)
3822 {
3823 	tte_t tte;
3824 
3825 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3826 
3827 	sfmmu_gettte(sfmmup, addr, &tte);
3828 	if (TTE_IS_VALID(&tte)) {
3829 		return (TTEBYTES(TTE_CSZ(&tte)));
3830 	}
3831 	return (-1);
3832 }
3833 
3834 static void
3835 sfmmu_gettte(struct hat *sfmmup, caddr_t addr, tte_t *ttep)
3836 {
3837 	struct hmehash_bucket *hmebp;
3838 	hmeblk_tag hblktag;
3839 	int hmeshift, hashno = 1;
3840 	struct hme_blk *hmeblkp, *list = NULL;
3841 	struct sf_hment *sfhmep;
3842 
3843 	/* support for ISM */
3844 	ism_map_t	*ism_map;
3845 	ism_blk_t	*ism_blkp;
3846 	int		i;
3847 	sfmmu_t		*ism_hatid = NULL;
3848 	sfmmu_t		*locked_hatid = NULL;
3849 
3850 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
3851 
3852 	ism_blkp = sfmmup->sfmmu_iblk;
3853 	if (ism_blkp) {
3854 		sfmmu_ismhat_enter(sfmmup, 0);
3855 		locked_hatid = sfmmup;
3856 	}
3857 	while (ism_blkp && ism_hatid == NULL) {
3858 		ism_map = ism_blkp->iblk_maps;
3859 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
3860 			if (addr >= ism_start(ism_map[i]) &&
3861 			    addr < ism_end(ism_map[i])) {
3862 				sfmmup = ism_hatid = ism_map[i].imap_ismhat;
3863 				addr = (caddr_t)(addr -
3864 					ism_start(ism_map[i]));
3865 				break;
3866 			}
3867 		}
3868 		ism_blkp = ism_blkp->iblk_next;
3869 	}
3870 	if (locked_hatid) {
3871 		sfmmu_ismhat_exit(locked_hatid, 0);
3872 	}
3873 
3874 	hblktag.htag_id = sfmmup;
3875 	ttep->ll = 0;
3876 
3877 	do {
3878 		hmeshift = HME_HASH_SHIFT(hashno);
3879 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
3880 		hblktag.htag_rehash = hashno;
3881 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
3882 
3883 		SFMMU_HASH_LOCK(hmebp);
3884 
3885 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
3886 		if (hmeblkp != NULL) {
3887 			HBLKTOHME(sfhmep, hmeblkp, addr);
3888 			sfmmu_copytte(&sfhmep->hme_tte, ttep);
3889 			SFMMU_HASH_UNLOCK(hmebp);
3890 			break;
3891 		}
3892 		SFMMU_HASH_UNLOCK(hmebp);
3893 		hashno++;
3894 	} while (HME_REHASH(sfmmup) && (hashno <= mmu_hashcnt));
3895 
3896 	sfmmu_hblks_list_purge(&list);
3897 }
3898 
3899 uint_t
3900 hat_getattr(struct hat *sfmmup, caddr_t addr, uint_t *attr)
3901 {
3902 	tte_t tte;
3903 
3904 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
3905 
3906 	sfmmu_gettte(sfmmup, addr, &tte);
3907 	if (TTE_IS_VALID(&tte)) {
3908 		*attr = sfmmu_ptov_attr(&tte);
3909 		return (0);
3910 	}
3911 	*attr = 0;
3912 	return ((uint_t)0xffffffff);
3913 }
3914 
3915 /*
3916  * Enables more attributes on specified address range (ie. logical OR)
3917  */
3918 void
3919 hat_setattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
3920 {
3921 	if (hat->sfmmu_xhat_provider) {
3922 		XHAT_SETATTR(hat, addr, len, attr);
3923 		return;
3924 	} else {
3925 		/*
3926 		 * This must be a CPU HAT. If the address space has
3927 		 * XHATs attached, change attributes for all of them,
3928 		 * just in case
3929 		 */
3930 		ASSERT(hat->sfmmu_as != NULL);
3931 		if (hat->sfmmu_as->a_xhat != NULL)
3932 			xhat_setattr_all(hat->sfmmu_as, addr, len, attr);
3933 	}
3934 
3935 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_SETATTR);
3936 }
3937 
3938 /*
3939  * Assigns attributes to the specified address range.  All the attributes
3940  * are specified.
3941  */
3942 void
3943 hat_chgattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
3944 {
3945 	if (hat->sfmmu_xhat_provider) {
3946 		XHAT_CHGATTR(hat, addr, len, attr);
3947 		return;
3948 	} else {
3949 		/*
3950 		 * This must be a CPU HAT. If the address space has
3951 		 * XHATs attached, change attributes for all of them,
3952 		 * just in case
3953 		 */
3954 		ASSERT(hat->sfmmu_as != NULL);
3955 		if (hat->sfmmu_as->a_xhat != NULL)
3956 			xhat_chgattr_all(hat->sfmmu_as, addr, len, attr);
3957 	}
3958 
3959 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CHGATTR);
3960 }
3961 
3962 /*
3963  * Remove attributes on the specified address range (ie. loginal NAND)
3964  */
3965 void
3966 hat_clrattr(struct hat *hat, caddr_t addr, size_t len, uint_t attr)
3967 {
3968 	if (hat->sfmmu_xhat_provider) {
3969 		XHAT_CLRATTR(hat, addr, len, attr);
3970 		return;
3971 	} else {
3972 		/*
3973 		 * This must be a CPU HAT. If the address space has
3974 		 * XHATs attached, change attributes for all of them,
3975 		 * just in case
3976 		 */
3977 		ASSERT(hat->sfmmu_as != NULL);
3978 		if (hat->sfmmu_as->a_xhat != NULL)
3979 			xhat_clrattr_all(hat->sfmmu_as, addr, len, attr);
3980 	}
3981 
3982 	sfmmu_chgattr(hat, addr, len, attr, SFMMU_CLRATTR);
3983 }
3984 
3985 /*
3986  * Change attributes on an address range to that specified by attr and mode.
3987  */
3988 static void
3989 sfmmu_chgattr(struct hat *sfmmup, caddr_t addr, size_t len, uint_t attr,
3990 	int mode)
3991 {
3992 	struct hmehash_bucket *hmebp;
3993 	hmeblk_tag hblktag;
3994 	int hmeshift, hashno = 1;
3995 	struct hme_blk *hmeblkp, *list = NULL;
3996 	caddr_t endaddr;
3997 	cpuset_t cpuset;
3998 	demap_range_t dmr;
3999 
4000 	CPUSET_ZERO(cpuset);
4001 
4002 	ASSERT((sfmmup == ksfmmup) ||
4003 		AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
4004 	ASSERT((len & MMU_PAGEOFFSET) == 0);
4005 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
4006 
4007 	if ((attr & PROT_USER) && (mode != SFMMU_CLRATTR) &&
4008 	    ((addr + len) > (caddr_t)USERLIMIT)) {
4009 		panic("user addr %p in kernel space",
4010 		    (void *)addr);
4011 	}
4012 
4013 	endaddr = addr + len;
4014 	hblktag.htag_id = sfmmup;
4015 	DEMAP_RANGE_INIT(sfmmup, &dmr);
4016 
4017 	while (addr < endaddr) {
4018 		hmeshift = HME_HASH_SHIFT(hashno);
4019 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
4020 		hblktag.htag_rehash = hashno;
4021 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
4022 
4023 		SFMMU_HASH_LOCK(hmebp);
4024 
4025 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
4026 		if (hmeblkp != NULL) {
4027 			/*
4028 			 * We've encountered a shadow hmeblk so skip the range
4029 			 * of the next smaller mapping size.
4030 			 */
4031 			if (hmeblkp->hblk_shw_bit) {
4032 				ASSERT(sfmmup != ksfmmup);
4033 				ASSERT(hashno > 1);
4034 				addr = (caddr_t)P2END((uintptr_t)addr,
4035 					    TTEBYTES(hashno - 1));
4036 			} else {
4037 				addr = sfmmu_hblk_chgattr(sfmmup,
4038 				    hmeblkp, addr, endaddr, &dmr, attr, mode);
4039 			}
4040 			SFMMU_HASH_UNLOCK(hmebp);
4041 			hashno = 1;
4042 			continue;
4043 		}
4044 		SFMMU_HASH_UNLOCK(hmebp);
4045 
4046 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
4047 			/*
4048 			 * We have traversed the whole list and rehashed
4049 			 * if necessary without finding the address to chgattr.
4050 			 * This is ok, so we increment the address by the
4051 			 * smallest hmeblk range for kernel mappings or for
4052 			 * user mappings with no large pages, and the largest
4053 			 * hmeblk range, to account for shadow hmeblks, for
4054 			 * user mappings with large pages and continue.
4055 			 */
4056 			if (sfmmup == ksfmmup)
4057 				addr = (caddr_t)P2END((uintptr_t)addr,
4058 					    TTEBYTES(1));
4059 			else
4060 				addr = (caddr_t)P2END((uintptr_t)addr,
4061 					    TTEBYTES(hashno));
4062 			hashno = 1;
4063 		} else {
4064 			hashno++;
4065 		}
4066 	}
4067 
4068 	sfmmu_hblks_list_purge(&list);
4069 	DEMAP_RANGE_FLUSH(&dmr);
4070 	cpuset = sfmmup->sfmmu_cpusran;
4071 	xt_sync(cpuset);
4072 }
4073 
4074 /*
4075  * This function chgattr on a range of addresses in an hmeblk.  It returns the
4076  * next addres that needs to be chgattr.
4077  * It should be called with the hash lock held.
4078  * XXX It should be possible to optimize chgattr by not flushing every time but
4079  * on the other hand:
4080  * 1. do one flush crosscall.
4081  * 2. only flush if we are increasing permissions (make sure this will work)
4082  */
4083 static caddr_t
4084 sfmmu_hblk_chgattr(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
4085 	caddr_t endaddr, demap_range_t *dmrp, uint_t attr, int mode)
4086 {
4087 	tte_t tte, tteattr, tteflags, ttemod;
4088 	struct sf_hment *sfhmep;
4089 	int ttesz;
4090 	struct page *pp = NULL;
4091 	kmutex_t *pml, *pmtx;
4092 	int ret;
4093 	int use_demap_range;
4094 #if defined(SF_ERRATA_57)
4095 	int check_exec;
4096 #endif
4097 
4098 	ASSERT(in_hblk_range(hmeblkp, addr));
4099 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4100 
4101 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4102 	ttesz = get_hblk_ttesz(hmeblkp);
4103 
4104 	/*
4105 	 * Flush the current demap region if addresses have been
4106 	 * skipped or the page size doesn't match.
4107 	 */
4108 	use_demap_range = (TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp));
4109 	if (use_demap_range) {
4110 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
4111 	} else {
4112 		DEMAP_RANGE_FLUSH(dmrp);
4113 	}
4114 
4115 	tteattr.ll = sfmmu_vtop_attr(attr, mode, &tteflags);
4116 #if defined(SF_ERRATA_57)
4117 	check_exec = (sfmmup != ksfmmup) &&
4118 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
4119 	    TTE_IS_EXECUTABLE(&tteattr);
4120 #endif
4121 	HBLKTOHME(sfhmep, hmeblkp, addr);
4122 	while (addr < endaddr) {
4123 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
4124 		if (TTE_IS_VALID(&tte)) {
4125 			if ((tte.ll & tteflags.ll) == tteattr.ll) {
4126 				/*
4127 				 * if the new attr is the same as old
4128 				 * continue
4129 				 */
4130 				goto next_addr;
4131 			}
4132 			if (!TTE_IS_WRITABLE(&tteattr)) {
4133 				/*
4134 				 * make sure we clear hw modify bit if we
4135 				 * removing write protections
4136 				 */
4137 				tteflags.tte_intlo |= TTE_HWWR_INT;
4138 			}
4139 
4140 			pml = NULL;
4141 			pp = sfhmep->hme_page;
4142 			if (pp) {
4143 				pml = sfmmu_mlist_enter(pp);
4144 			}
4145 
4146 			if (pp != sfhmep->hme_page) {
4147 				/*
4148 				 * tte must have been unloaded.
4149 				 */
4150 				ASSERT(pml);
4151 				sfmmu_mlist_exit(pml);
4152 				continue;
4153 			}
4154 
4155 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
4156 
4157 			ttemod = tte;
4158 			ttemod.ll = (ttemod.ll & ~tteflags.ll) | tteattr.ll;
4159 			ASSERT(TTE_TO_TTEPFN(&ttemod) == TTE_TO_TTEPFN(&tte));
4160 
4161 #if defined(SF_ERRATA_57)
4162 			if (check_exec && addr < errata57_limit)
4163 				ttemod.tte_exec_perm = 0;
4164 #endif
4165 			ret = sfmmu_modifytte_try(&tte, &ttemod,
4166 			    &sfhmep->hme_tte);
4167 
4168 			if (ret < 0) {
4169 				/* tte changed underneath us */
4170 				if (pml) {
4171 					sfmmu_mlist_exit(pml);
4172 				}
4173 				continue;
4174 			}
4175 
4176 			if (tteflags.tte_intlo & TTE_HWWR_INT) {
4177 				/*
4178 				 * need to sync if we are clearing modify bit.
4179 				 */
4180 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
4181 			}
4182 
4183 			if (pp && PP_ISRO(pp)) {
4184 				if (tteattr.tte_intlo & TTE_WRPRM_INT) {
4185 					pmtx = sfmmu_page_enter(pp);
4186 					PP_CLRRO(pp);
4187 					sfmmu_page_exit(pmtx);
4188 				}
4189 			}
4190 
4191 			if (ret > 0 && use_demap_range) {
4192 				DEMAP_RANGE_MARKPG(dmrp, addr);
4193 			} else if (ret > 0) {
4194 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
4195 			}
4196 
4197 			if (pml) {
4198 				sfmmu_mlist_exit(pml);
4199 			}
4200 		}
4201 next_addr:
4202 		addr += TTEBYTES(ttesz);
4203 		sfhmep++;
4204 		DEMAP_RANGE_NEXTPG(dmrp);
4205 	}
4206 	return (addr);
4207 }
4208 
4209 /*
4210  * This routine converts virtual attributes to physical ones.  It will
4211  * update the tteflags field with the tte mask corresponding to the attributes
4212  * affected and it returns the new attributes.  It will also clear the modify
4213  * bit if we are taking away write permission.  This is necessary since the
4214  * modify bit is the hardware permission bit and we need to clear it in order
4215  * to detect write faults.
4216  */
4217 static uint64_t
4218 sfmmu_vtop_attr(uint_t attr, int mode, tte_t *ttemaskp)
4219 {
4220 	tte_t ttevalue;
4221 
4222 	ASSERT(!(attr & ~SFMMU_LOAD_ALLATTR));
4223 
4224 	switch (mode) {
4225 	case SFMMU_CHGATTR:
4226 		/* all attributes specified */
4227 		ttevalue.tte_inthi = MAKE_TTEATTR_INTHI(attr);
4228 		ttevalue.tte_intlo = MAKE_TTEATTR_INTLO(attr);
4229 		ttemaskp->tte_inthi = TTEINTHI_ATTR;
4230 		ttemaskp->tte_intlo = TTEINTLO_ATTR;
4231 		break;
4232 	case SFMMU_SETATTR:
4233 		ASSERT(!(attr & ~HAT_PROT_MASK));
4234 		ttemaskp->ll = 0;
4235 		ttevalue.ll = 0;
4236 		/*
4237 		 * a valid tte implies exec and read for sfmmu
4238 		 * so no need to do anything about them.
4239 		 * since priviledged access implies user access
4240 		 * PROT_USER doesn't make sense either.
4241 		 */
4242 		if (attr & PROT_WRITE) {
4243 			ttemaskp->tte_intlo |= TTE_WRPRM_INT;
4244 			ttevalue.tte_intlo |= TTE_WRPRM_INT;
4245 		}
4246 		break;
4247 	case SFMMU_CLRATTR:
4248 		/* attributes will be nand with current ones */
4249 		if (attr & ~(PROT_WRITE | PROT_USER)) {
4250 			panic("sfmmu: attr %x not supported", attr);
4251 		}
4252 		ttemaskp->ll = 0;
4253 		ttevalue.ll = 0;
4254 		if (attr & PROT_WRITE) {
4255 			/* clear both writable and modify bit */
4256 			ttemaskp->tte_intlo |= TTE_WRPRM_INT | TTE_HWWR_INT;
4257 		}
4258 		if (attr & PROT_USER) {
4259 			ttemaskp->tte_intlo |= TTE_PRIV_INT;
4260 			ttevalue.tte_intlo |= TTE_PRIV_INT;
4261 		}
4262 		break;
4263 	default:
4264 		panic("sfmmu_vtop_attr: bad mode %x", mode);
4265 	}
4266 	ASSERT(TTE_TO_TTEPFN(&ttevalue) == 0);
4267 	return (ttevalue.ll);
4268 }
4269 
4270 static uint_t
4271 sfmmu_ptov_attr(tte_t *ttep)
4272 {
4273 	uint_t attr;
4274 
4275 	ASSERT(TTE_IS_VALID(ttep));
4276 
4277 	attr = PROT_READ;
4278 
4279 	if (TTE_IS_WRITABLE(ttep)) {
4280 		attr |= PROT_WRITE;
4281 	}
4282 	if (TTE_IS_EXECUTABLE(ttep)) {
4283 		attr |= PROT_EXEC;
4284 	}
4285 	if (!TTE_IS_PRIVILEGED(ttep)) {
4286 		attr |= PROT_USER;
4287 	}
4288 	if (TTE_IS_NFO(ttep)) {
4289 		attr |= HAT_NOFAULT;
4290 	}
4291 	if (TTE_IS_NOSYNC(ttep)) {
4292 		attr |= HAT_NOSYNC;
4293 	}
4294 	if (TTE_IS_SIDEFFECT(ttep)) {
4295 		attr |= SFMMU_SIDEFFECT;
4296 	}
4297 	if (!TTE_IS_VCACHEABLE(ttep)) {
4298 		attr |= SFMMU_UNCACHEVTTE;
4299 	}
4300 	if (!TTE_IS_PCACHEABLE(ttep)) {
4301 		attr |= SFMMU_UNCACHEPTTE;
4302 	}
4303 	return (attr);
4304 }
4305 
4306 /*
4307  * hat_chgprot is a deprecated hat call.  New segment drivers
4308  * should store all attributes and use hat_*attr calls.
4309  *
4310  * Change the protections in the virtual address range
4311  * given to the specified virtual protection.  If vprot is ~PROT_WRITE,
4312  * then remove write permission, leaving the other
4313  * permissions unchanged.  If vprot is ~PROT_USER, remove user permissions.
4314  *
4315  */
4316 void
4317 hat_chgprot(struct hat *sfmmup, caddr_t addr, size_t len, uint_t vprot)
4318 {
4319 	struct hmehash_bucket *hmebp;
4320 	hmeblk_tag hblktag;
4321 	int hmeshift, hashno = 1;
4322 	struct hme_blk *hmeblkp, *list = NULL;
4323 	caddr_t endaddr;
4324 	cpuset_t cpuset;
4325 	demap_range_t dmr;
4326 
4327 	ASSERT((len & MMU_PAGEOFFSET) == 0);
4328 	ASSERT(((uintptr_t)addr & MMU_PAGEOFFSET) == 0);
4329 
4330 	if (sfmmup->sfmmu_xhat_provider) {
4331 		XHAT_CHGPROT(sfmmup, addr, len, vprot);
4332 		return;
4333 	} else {
4334 		/*
4335 		 * This must be a CPU HAT. If the address space has
4336 		 * XHATs attached, change attributes for all of them,
4337 		 * just in case
4338 		 */
4339 		ASSERT(sfmmup->sfmmu_as != NULL);
4340 		if (sfmmup->sfmmu_as->a_xhat != NULL)
4341 			xhat_chgprot_all(sfmmup->sfmmu_as, addr, len, vprot);
4342 	}
4343 
4344 	CPUSET_ZERO(cpuset);
4345 
4346 	if ((vprot != (uint_t)~PROT_WRITE) && (vprot & PROT_USER) &&
4347 	    ((addr + len) > (caddr_t)USERLIMIT)) {
4348 		panic("user addr %p vprot %x in kernel space",
4349 		    (void *)addr, vprot);
4350 	}
4351 	endaddr = addr + len;
4352 	hblktag.htag_id = sfmmup;
4353 	DEMAP_RANGE_INIT(sfmmup, &dmr);
4354 
4355 	while (addr < endaddr) {
4356 		hmeshift = HME_HASH_SHIFT(hashno);
4357 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
4358 		hblktag.htag_rehash = hashno;
4359 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
4360 
4361 		SFMMU_HASH_LOCK(hmebp);
4362 
4363 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
4364 		if (hmeblkp != NULL) {
4365 			/*
4366 			 * We've encountered a shadow hmeblk so skip the range
4367 			 * of the next smaller mapping size.
4368 			 */
4369 			if (hmeblkp->hblk_shw_bit) {
4370 				ASSERT(sfmmup != ksfmmup);
4371 				ASSERT(hashno > 1);
4372 				addr = (caddr_t)P2END((uintptr_t)addr,
4373 					    TTEBYTES(hashno - 1));
4374 			} else {
4375 				addr = sfmmu_hblk_chgprot(sfmmup, hmeblkp,
4376 					addr, endaddr, &dmr, vprot);
4377 			}
4378 			SFMMU_HASH_UNLOCK(hmebp);
4379 			hashno = 1;
4380 			continue;
4381 		}
4382 		SFMMU_HASH_UNLOCK(hmebp);
4383 
4384 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
4385 			/*
4386 			 * We have traversed the whole list and rehashed
4387 			 * if necessary without finding the address to chgprot.
4388 			 * This is ok so we increment the address by the
4389 			 * smallest hmeblk range for kernel mappings and the
4390 			 * largest hmeblk range, to account for shadow hmeblks,
4391 			 * for user mappings and continue.
4392 			 */
4393 			if (sfmmup == ksfmmup)
4394 				addr = (caddr_t)P2END((uintptr_t)addr,
4395 					    TTEBYTES(1));
4396 			else
4397 				addr = (caddr_t)P2END((uintptr_t)addr,
4398 					    TTEBYTES(hashno));
4399 			hashno = 1;
4400 		} else {
4401 			hashno++;
4402 		}
4403 	}
4404 
4405 	sfmmu_hblks_list_purge(&list);
4406 	DEMAP_RANGE_FLUSH(&dmr);
4407 	cpuset = sfmmup->sfmmu_cpusran;
4408 	xt_sync(cpuset);
4409 }
4410 
4411 /*
4412  * This function chgprots a range of addresses in an hmeblk.  It returns the
4413  * next addres that needs to be chgprot.
4414  * It should be called with the hash lock held.
4415  * XXX It shold be possible to optimize chgprot by not flushing every time but
4416  * on the other hand:
4417  * 1. do one flush crosscall.
4418  * 2. only flush if we are increasing permissions (make sure this will work)
4419  */
4420 static caddr_t
4421 sfmmu_hblk_chgprot(sfmmu_t *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
4422 	caddr_t endaddr, demap_range_t *dmrp, uint_t vprot)
4423 {
4424 	uint_t pprot;
4425 	tte_t tte, ttemod;
4426 	struct sf_hment *sfhmep;
4427 	uint_t tteflags;
4428 	int ttesz;
4429 	struct page *pp = NULL;
4430 	kmutex_t *pml, *pmtx;
4431 	int ret;
4432 	int use_demap_range;
4433 #if defined(SF_ERRATA_57)
4434 	int check_exec;
4435 #endif
4436 
4437 	ASSERT(in_hblk_range(hmeblkp, addr));
4438 	ASSERT(hmeblkp->hblk_shw_bit == 0);
4439 
4440 #ifdef DEBUG
4441 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
4442 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
4443 		panic("sfmmu_hblk_chgprot: partial chgprot of large page");
4444 	}
4445 #endif /* DEBUG */
4446 
4447 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
4448 	ttesz = get_hblk_ttesz(hmeblkp);
4449 
4450 	pprot = sfmmu_vtop_prot(vprot, &tteflags);
4451 #if defined(SF_ERRATA_57)
4452 	check_exec = (sfmmup != ksfmmup) &&
4453 	    AS_TYPE_64BIT(sfmmup->sfmmu_as) &&
4454 	    ((vprot & PROT_EXEC) == PROT_EXEC);
4455 #endif
4456 	HBLKTOHME(sfhmep, hmeblkp, addr);
4457 
4458 	/*
4459 	 * Flush the current demap region if addresses have been
4460 	 * skipped or the page size doesn't match.
4461 	 */
4462 	use_demap_range = (TTEBYTES(ttesz) == MMU_PAGESIZE);
4463 	if (use_demap_range) {
4464 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
4465 	} else {
4466 		DEMAP_RANGE_FLUSH(dmrp);
4467 	}
4468 
4469 	while (addr < endaddr) {
4470 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
4471 		if (TTE_IS_VALID(&tte)) {
4472 			if (TTE_GET_LOFLAGS(&tte, tteflags) == pprot) {
4473 				/*
4474 				 * if the new protection is the same as old
4475 				 * continue
4476 				 */
4477 				goto next_addr;
4478 			}
4479 			pml = NULL;
4480 			pp = sfhmep->hme_page;
4481 			if (pp) {
4482 				pml = sfmmu_mlist_enter(pp);
4483 			}
4484 			if (pp != sfhmep->hme_page) {
4485 				/*
4486 				 * tte most have been unloaded
4487 				 * underneath us.  Recheck
4488 				 */
4489 				ASSERT(pml);
4490 				sfmmu_mlist_exit(pml);
4491 				continue;
4492 			}
4493 
4494 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
4495 
4496 			ttemod = tte;
4497 			TTE_SET_LOFLAGS(&ttemod, tteflags, pprot);
4498 #if defined(SF_ERRATA_57)
4499 			if (check_exec && addr < errata57_limit)
4500 				ttemod.tte_exec_perm = 0;
4501 #endif
4502 			ret = sfmmu_modifytte_try(&tte, &ttemod,
4503 			    &sfhmep->hme_tte);
4504 
4505 			if (ret < 0) {
4506 				/* tte changed underneath us */
4507 				if (pml) {
4508 					sfmmu_mlist_exit(pml);
4509 				}
4510 				continue;
4511 			}
4512 
4513 			if (tteflags & TTE_HWWR_INT) {
4514 				/*
4515 				 * need to sync if we are clearing modify bit.
4516 				 */
4517 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
4518 			}
4519 
4520 			if (pp && PP_ISRO(pp)) {
4521 				if (pprot & TTE_WRPRM_INT) {
4522 					pmtx = sfmmu_page_enter(pp);
4523 					PP_CLRRO(pp);
4524 					sfmmu_page_exit(pmtx);
4525 				}
4526 			}
4527 
4528 			if (ret > 0 && use_demap_range) {
4529 				DEMAP_RANGE_MARKPG(dmrp, addr);
4530 			} else if (ret > 0) {
4531 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
4532 			}
4533 
4534 			if (pml) {
4535 				sfmmu_mlist_exit(pml);
4536 			}
4537 		}
4538 next_addr:
4539 		addr += TTEBYTES(ttesz);
4540 		sfhmep++;
4541 		DEMAP_RANGE_NEXTPG(dmrp);
4542 	}
4543 	return (addr);
4544 }
4545 
4546 /*
4547  * This routine is deprecated and should only be used by hat_chgprot.
4548  * The correct routine is sfmmu_vtop_attr.
4549  * This routine converts virtual page protections to physical ones.  It will
4550  * update the tteflags field with the tte mask corresponding to the protections
4551  * affected and it returns the new protections.  It will also clear the modify
4552  * bit if we are taking away write permission.  This is necessary since the
4553  * modify bit is the hardware permission bit and we need to clear it in order
4554  * to detect write faults.
4555  * It accepts the following special protections:
4556  * ~PROT_WRITE = remove write permissions.
4557  * ~PROT_USER = remove user permissions.
4558  */
4559 static uint_t
4560 sfmmu_vtop_prot(uint_t vprot, uint_t *tteflagsp)
4561 {
4562 	if (vprot == (uint_t)~PROT_WRITE) {
4563 		*tteflagsp = TTE_WRPRM_INT | TTE_HWWR_INT;
4564 		return (0);		/* will cause wrprm to be cleared */
4565 	}
4566 	if (vprot == (uint_t)~PROT_USER) {
4567 		*tteflagsp = TTE_PRIV_INT;
4568 		return (0);		/* will cause privprm to be cleared */
4569 	}
4570 	if ((vprot == 0) || (vprot == PROT_USER) ||
4571 		((vprot & PROT_ALL) != vprot)) {
4572 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
4573 	}
4574 
4575 	switch (vprot) {
4576 	case (PROT_READ):
4577 	case (PROT_EXEC):
4578 	case (PROT_EXEC | PROT_READ):
4579 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
4580 		return (TTE_PRIV_INT); 		/* set prv and clr wrt */
4581 	case (PROT_WRITE):
4582 	case (PROT_WRITE | PROT_READ):
4583 	case (PROT_EXEC | PROT_WRITE):
4584 	case (PROT_EXEC | PROT_WRITE | PROT_READ):
4585 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
4586 		return (TTE_PRIV_INT | TTE_WRPRM_INT); 	/* set prv and wrt */
4587 	case (PROT_USER | PROT_READ):
4588 	case (PROT_USER | PROT_EXEC):
4589 	case (PROT_USER | PROT_EXEC | PROT_READ):
4590 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT | TTE_HWWR_INT;
4591 		return (0); 			/* clr prv and wrt */
4592 	case (PROT_USER | PROT_WRITE):
4593 	case (PROT_USER | PROT_WRITE | PROT_READ):
4594 	case (PROT_USER | PROT_EXEC | PROT_WRITE):
4595 	case (PROT_USER | PROT_EXEC | PROT_WRITE | PROT_READ):
4596 		*tteflagsp = TTE_PRIV_INT | TTE_WRPRM_INT;
4597 		return (TTE_WRPRM_INT); 	/* clr prv and set wrt */
4598 	default:
4599 		panic("sfmmu_vtop_prot -- bad prot %x", vprot);
4600 	}
4601 	return (0);
4602 }
4603 
4604 /*
4605  * Alternate unload for very large virtual ranges. With a true 64 bit VA,
4606  * the normal algorithm would take too long for a very large VA range with
4607  * few real mappings. This routine just walks thru all HMEs in the global
4608  * hash table to find and remove mappings.
4609  */
4610 static void
4611 hat_unload_large_virtual(
4612 	struct hat		*sfmmup,
4613 	caddr_t			startaddr,
4614 	size_t			len,
4615 	uint_t			flags,
4616 	hat_callback_t		*callback)
4617 {
4618 	struct hmehash_bucket *hmebp;
4619 	struct hme_blk *hmeblkp;
4620 	struct hme_blk *pr_hblk = NULL;
4621 	struct hme_blk *nx_hblk;
4622 	struct hme_blk *list = NULL;
4623 	int i;
4624 	uint64_t hblkpa, prevpa, nx_pa;
4625 	hatlock_t	*hatlockp;
4626 	struct tsb_info	*tsbinfop;
4627 	struct ctx	*ctx;
4628 	caddr_t	endaddr = startaddr + len;
4629 	caddr_t	sa;
4630 	caddr_t	ea;
4631 	caddr_t	cb_sa[MAX_CB_ADDR];
4632 	caddr_t	cb_ea[MAX_CB_ADDR];
4633 	int	addr_cnt = 0;
4634 	int	a = 0;
4635 	int	cnum;
4636 
4637 	hatlockp = sfmmu_hat_enter(sfmmup);
4638 
4639 	/*
4640 	 * Since we know we're unmapping a huge range of addresses,
4641 	 * just throw away the context and switch to another.  It's
4642 	 * cheaper than trying to unmap all of the TTEs we may find
4643 	 * from the TLB individually, which is too expensive in terms
4644 	 * of xcalls.  Better yet, if we're exiting, no need to flush
4645 	 * anything at all!
4646 	 */
4647 	if (!sfmmup->sfmmu_free) {
4648 		ctx = sfmmutoctx(sfmmup);
4649 		rw_enter(&ctx->ctx_rwlock, RW_WRITER);
4650 		cnum = sfmmutoctxnum(sfmmup);
4651 		if (cnum != INVALID_CONTEXT) {
4652 			sfmmu_tlb_swap_ctx(sfmmup, ctx);
4653 		}
4654 		rw_exit(&ctx->ctx_rwlock);
4655 
4656 		for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
4657 		    tsbinfop = tsbinfop->tsb_next) {
4658 			if (tsbinfop->tsb_flags & TSB_SWAPPED)
4659 				continue;
4660 			sfmmu_inv_tsb(tsbinfop->tsb_va,
4661 			    TSB_BYTES(tsbinfop->tsb_szc));
4662 		}
4663 	}
4664 
4665 	/*
4666 	 * Loop through all the hash buckets of HME blocks looking for matches.
4667 	 */
4668 	for (i = 0; i <= UHMEHASH_SZ; i++) {
4669 		hmebp = &uhme_hash[i];
4670 		SFMMU_HASH_LOCK(hmebp);
4671 		hmeblkp = hmebp->hmeblkp;
4672 		hblkpa = hmebp->hmeh_nextpa;
4673 		prevpa = 0;
4674 		pr_hblk = NULL;
4675 		while (hmeblkp) {
4676 			nx_hblk = hmeblkp->hblk_next;
4677 			nx_pa = hmeblkp->hblk_nextpa;
4678 
4679 			/*
4680 			 * skip if not this context, if a shadow block or
4681 			 * if the mapping is not in the requested range
4682 			 */
4683 			if (hmeblkp->hblk_tag.htag_id != sfmmup ||
4684 			    hmeblkp->hblk_shw_bit ||
4685 			    (sa = (caddr_t)get_hblk_base(hmeblkp)) >= endaddr ||
4686 			    (ea = get_hblk_endaddr(hmeblkp)) <= startaddr) {
4687 				pr_hblk = hmeblkp;
4688 				prevpa = hblkpa;
4689 				goto next_block;
4690 			}
4691 
4692 			/*
4693 			 * unload if there are any current valid mappings
4694 			 */
4695 			if (hmeblkp->hblk_vcnt != 0 ||
4696 			    hmeblkp->hblk_hmecnt != 0)
4697 				(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
4698 				    sa, ea, NULL, flags);
4699 
4700 			/*
4701 			 * on unmap we also release the HME block itself, once
4702 			 * all mappings are gone.
4703 			 */
4704 			if ((flags & HAT_UNLOAD_UNMAP) != 0 &&
4705 			    !hmeblkp->hblk_vcnt &&
4706 			    !hmeblkp->hblk_hmecnt) {
4707 				ASSERT(!hmeblkp->hblk_lckcnt);
4708 				sfmmu_hblk_hash_rm(hmebp, hmeblkp,
4709 					prevpa, pr_hblk);
4710 				sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
4711 			} else {
4712 				pr_hblk = hmeblkp;
4713 				prevpa = hblkpa;
4714 			}
4715 
4716 			if (callback == NULL)
4717 				goto next_block;
4718 
4719 			/*
4720 			 * HME blocks may span more than one page, but we may be
4721 			 * unmapping only one page, so check for a smaller range
4722 			 * for the callback
4723 			 */
4724 			if (sa < startaddr)
4725 				sa = startaddr;
4726 			if (--ea > endaddr)
4727 				ea = endaddr - 1;
4728 
4729 			cb_sa[addr_cnt] = sa;
4730 			cb_ea[addr_cnt] = ea;
4731 			if (++addr_cnt == MAX_CB_ADDR) {
4732 				for (a = 0; a < MAX_CB_ADDR; ++a) {
4733 					callback->hcb_start_addr = cb_sa[a];
4734 					callback->hcb_end_addr = cb_ea[a];
4735 					callback->hcb_function(callback);
4736 				}
4737 				addr_cnt = 0;
4738 			}
4739 
4740 next_block:
4741 			hmeblkp = nx_hblk;
4742 			hblkpa = nx_pa;
4743 		}
4744 		SFMMU_HASH_UNLOCK(hmebp);
4745 	}
4746 
4747 	sfmmu_hblks_list_purge(&list);
4748 
4749 	for (a = 0; a < addr_cnt; ++a) {
4750 		callback->hcb_start_addr = cb_sa[a];
4751 		callback->hcb_end_addr = cb_ea[a];
4752 		callback->hcb_function(callback);
4753 	}
4754 
4755 	sfmmu_hat_exit(hatlockp);
4756 
4757 	/*
4758 	 * Check TSB and TLB page sizes if the process isn't exiting.
4759 	 */
4760 	if (!sfmmup->sfmmu_free)
4761 		sfmmu_check_page_sizes(sfmmup, 0);
4762 }
4763 
4764 
4765 /*
4766  * Unload all the mappings in the range [addr..addr+len). addr and len must
4767  * be MMU_PAGESIZE aligned.
4768  */
4769 
4770 extern struct seg *segkmap;
4771 #define	ISSEGKMAP(sfmmup, addr) (sfmmup == ksfmmup && \
4772 segkmap->s_base <= (addr) && (addr) < (segkmap->s_base + segkmap->s_size))
4773 
4774 
4775 void
4776 hat_unload_callback(
4777 	struct hat *sfmmup,
4778 	caddr_t addr,
4779 	size_t len,
4780 	uint_t flags,
4781 	hat_callback_t *callback)
4782 {
4783 	struct hmehash_bucket *hmebp;
4784 	hmeblk_tag hblktag;
4785 	int hmeshift, hashno, iskernel;
4786 	struct hme_blk *hmeblkp, *pr_hblk, *list = NULL;
4787 	caddr_t endaddr;
4788 	cpuset_t cpuset;
4789 	uint64_t hblkpa, prevpa;
4790 	int addr_count = 0;
4791 	int a;
4792 	caddr_t cb_start_addr[MAX_CB_ADDR];
4793 	caddr_t cb_end_addr[MAX_CB_ADDR];
4794 	int issegkmap = ISSEGKMAP(sfmmup, addr);
4795 	demap_range_t dmr, *dmrp;
4796 
4797 	if (sfmmup->sfmmu_xhat_provider) {
4798 		XHAT_UNLOAD_CALLBACK(sfmmup, addr, len, flags, callback);
4799 		return;
4800 	} else {
4801 		/*
4802 		 * This must be a CPU HAT. If the address space has
4803 		 * XHATs attached, unload the mappings for all of them,
4804 		 * just in case
4805 		 */
4806 		ASSERT(sfmmup->sfmmu_as != NULL);
4807 		if (sfmmup->sfmmu_as->a_xhat != NULL)
4808 			xhat_unload_callback_all(sfmmup->sfmmu_as, addr,
4809 			    len, flags, callback);
4810 	}
4811 
4812 	ASSERT((sfmmup == ksfmmup) || (flags & HAT_UNLOAD_OTHER) || \
4813 	    AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
4814 
4815 	ASSERT(sfmmup != NULL);
4816 	ASSERT((len & MMU_PAGEOFFSET) == 0);
4817 	ASSERT(!((uintptr_t)addr & MMU_PAGEOFFSET));
4818 
4819 	/*
4820 	 * Probing through a large VA range (say 63 bits) will be slow, even
4821 	 * at 4 Meg steps between the probes. So, when the virtual address range
4822 	 * is very large, search the HME entries for what to unload.
4823 	 *
4824 	 *	len >> TTE_PAGE_SHIFT(TTE4M) is the # of 4Meg probes we'd need
4825 	 *
4826 	 *	UHMEHASH_SZ is number of hash buckets to examine
4827 	 *
4828 	 */
4829 	if (sfmmup != KHATID && (len >> TTE_PAGE_SHIFT(TTE4M)) > UHMEHASH_SZ) {
4830 		hat_unload_large_virtual(sfmmup, addr, len, flags, callback);
4831 		return;
4832 	}
4833 
4834 	CPUSET_ZERO(cpuset);
4835 
4836 	/*
4837 	 * If the process is exiting, we can save a lot of fuss since
4838 	 * we'll flush the TLB when we free the ctx anyway.
4839 	 */
4840 	if (sfmmup->sfmmu_free)
4841 		dmrp = NULL;
4842 	else
4843 		dmrp = &dmr;
4844 
4845 	DEMAP_RANGE_INIT(sfmmup, dmrp);
4846 	endaddr = addr + len;
4847 	hblktag.htag_id = sfmmup;
4848 
4849 	/*
4850 	 * It is likely for the vm to call unload over a wide range of
4851 	 * addresses that are actually very sparsely populated by
4852 	 * translations.  In order to speed this up the sfmmu hat supports
4853 	 * the concept of shadow hmeblks. Dummy large page hmeblks that
4854 	 * correspond to actual small translations are allocated at tteload
4855 	 * time and are referred to as shadow hmeblks.  Now, during unload
4856 	 * time, we first check if we have a shadow hmeblk for that
4857 	 * translation.  The absence of one means the corresponding address
4858 	 * range is empty and can be skipped.
4859 	 *
4860 	 * The kernel is an exception to above statement and that is why
4861 	 * we don't use shadow hmeblks and hash starting from the smallest
4862 	 * page size.
4863 	 */
4864 	if (sfmmup == KHATID) {
4865 		iskernel = 1;
4866 		hashno = TTE64K;
4867 	} else {
4868 		iskernel = 0;
4869 		if (mmu_page_sizes == max_mmu_page_sizes) {
4870 			hashno = TTE256M;
4871 		} else {
4872 			hashno = TTE4M;
4873 		}
4874 	}
4875 	while (addr < endaddr) {
4876 		hmeshift = HME_HASH_SHIFT(hashno);
4877 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
4878 		hblktag.htag_rehash = hashno;
4879 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
4880 
4881 		SFMMU_HASH_LOCK(hmebp);
4882 
4883 		HME_HASH_SEARCH_PREV(hmebp, hblktag, hmeblkp, hblkpa, pr_hblk,
4884 			prevpa, &list);
4885 		if (hmeblkp == NULL) {
4886 			/*
4887 			 * didn't find an hmeblk. skip the appropiate
4888 			 * address range.
4889 			 */
4890 			SFMMU_HASH_UNLOCK(hmebp);
4891 			if (iskernel) {
4892 				if (hashno < mmu_hashcnt) {
4893 					hashno++;
4894 					continue;
4895 				} else {
4896 					hashno = TTE64K;
4897 					addr = (caddr_t)roundup((uintptr_t)addr
4898 						+ 1, MMU_PAGESIZE64K);
4899 					continue;
4900 				}
4901 			}
4902 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
4903 				(1 << hmeshift));
4904 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
4905 				ASSERT(hashno == TTE64K);
4906 				continue;
4907 			}
4908 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
4909 				hashno = TTE512K;
4910 				continue;
4911 			}
4912 			if (mmu_page_sizes == max_mmu_page_sizes) {
4913 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
4914 					hashno = TTE4M;
4915 					continue;
4916 				}
4917 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
4918 					hashno = TTE32M;
4919 					continue;
4920 				}
4921 				hashno = TTE256M;
4922 				continue;
4923 			} else {
4924 				hashno = TTE4M;
4925 				continue;
4926 			}
4927 		}
4928 		ASSERT(hmeblkp);
4929 		if (!hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
4930 			/*
4931 			 * If the valid count is zero we can skip the range
4932 			 * mapped by this hmeblk.
4933 			 * We free hblks in the case of HAT_UNMAP.  HAT_UNMAP
4934 			 * is used by segment drivers as a hint
4935 			 * that the mapping resource won't be used any longer.
4936 			 * The best example of this is during exit().
4937 			 */
4938 			addr = (caddr_t)roundup((uintptr_t)addr + 1,
4939 				get_hblk_span(hmeblkp));
4940 			if ((flags & HAT_UNLOAD_UNMAP) ||
4941 			    (iskernel && !issegkmap)) {
4942 				sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa,
4943 				    pr_hblk);
4944 				sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
4945 			}
4946 			SFMMU_HASH_UNLOCK(hmebp);
4947 
4948 			if (iskernel) {
4949 				hashno = TTE64K;
4950 				continue;
4951 			}
4952 			if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
4953 				ASSERT(hashno == TTE64K);
4954 				continue;
4955 			}
4956 			if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
4957 				hashno = TTE512K;
4958 				continue;
4959 			}
4960 			if (mmu_page_sizes == max_mmu_page_sizes) {
4961 				if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
4962 					hashno = TTE4M;
4963 					continue;
4964 				}
4965 				if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
4966 					hashno = TTE32M;
4967 					continue;
4968 				}
4969 				hashno = TTE256M;
4970 				continue;
4971 			} else {
4972 				hashno = TTE4M;
4973 				continue;
4974 			}
4975 		}
4976 		if (hmeblkp->hblk_shw_bit) {
4977 			/*
4978 			 * If we encounter a shadow hmeblk we know there is
4979 			 * smaller sized hmeblks mapping the same address space.
4980 			 * Decrement the hash size and rehash.
4981 			 */
4982 			ASSERT(sfmmup != KHATID);
4983 			hashno--;
4984 			SFMMU_HASH_UNLOCK(hmebp);
4985 			continue;
4986 		}
4987 
4988 		/*
4989 		 * track callback address ranges.
4990 		 * only start a new range when it's not contiguous
4991 		 */
4992 		if (callback != NULL) {
4993 			if (addr_count > 0 &&
4994 			    addr == cb_end_addr[addr_count - 1])
4995 				--addr_count;
4996 			else
4997 				cb_start_addr[addr_count] = addr;
4998 		}
4999 
5000 		addr = sfmmu_hblk_unload(sfmmup, hmeblkp, addr, endaddr,
5001 				dmrp, flags);
5002 
5003 		if (callback != NULL)
5004 			cb_end_addr[addr_count++] = addr;
5005 
5006 		if (((flags & HAT_UNLOAD_UNMAP) || (iskernel && !issegkmap)) &&
5007 		    !hmeblkp->hblk_vcnt && !hmeblkp->hblk_hmecnt) {
5008 			sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa,
5009 			    pr_hblk);
5010 			sfmmu_hblk_free(hmebp, hmeblkp, hblkpa, &list);
5011 		}
5012 		SFMMU_HASH_UNLOCK(hmebp);
5013 
5014 		/*
5015 		 * Notify our caller as to exactly which pages
5016 		 * have been unloaded. We do these in clumps,
5017 		 * to minimize the number of xt_sync()s that need to occur.
5018 		 */
5019 		if (callback != NULL && addr_count == MAX_CB_ADDR) {
5020 			DEMAP_RANGE_FLUSH(dmrp);
5021 			if (dmrp != NULL) {
5022 				cpuset = sfmmup->sfmmu_cpusran;
5023 				xt_sync(cpuset);
5024 			}
5025 
5026 			for (a = 0; a < MAX_CB_ADDR; ++a) {
5027 				callback->hcb_start_addr = cb_start_addr[a];
5028 				callback->hcb_end_addr = cb_end_addr[a];
5029 				callback->hcb_function(callback);
5030 			}
5031 			addr_count = 0;
5032 		}
5033 		if (iskernel) {
5034 			hashno = TTE64K;
5035 			continue;
5036 		}
5037 		if ((uintptr_t)addr & MMU_PAGEOFFSET512K) {
5038 			ASSERT(hashno == TTE64K);
5039 			continue;
5040 		}
5041 		if ((uintptr_t)addr & MMU_PAGEOFFSET4M) {
5042 			hashno = TTE512K;
5043 			continue;
5044 		}
5045 		if (mmu_page_sizes == max_mmu_page_sizes) {
5046 			if ((uintptr_t)addr & MMU_PAGEOFFSET32M) {
5047 				hashno = TTE4M;
5048 				continue;
5049 			}
5050 			if ((uintptr_t)addr & MMU_PAGEOFFSET256M) {
5051 				hashno = TTE32M;
5052 				continue;
5053 			}
5054 			hashno = TTE256M;
5055 		} else {
5056 			hashno = TTE4M;
5057 		}
5058 	}
5059 
5060 	sfmmu_hblks_list_purge(&list);
5061 	DEMAP_RANGE_FLUSH(dmrp);
5062 	if (dmrp != NULL) {
5063 		cpuset = sfmmup->sfmmu_cpusran;
5064 		xt_sync(cpuset);
5065 	}
5066 	if (callback && addr_count != 0) {
5067 		for (a = 0; a < addr_count; ++a) {
5068 			callback->hcb_start_addr = cb_start_addr[a];
5069 			callback->hcb_end_addr = cb_end_addr[a];
5070 			callback->hcb_function(callback);
5071 		}
5072 	}
5073 
5074 	/*
5075 	 * Check TSB and TLB page sizes if the process isn't exiting.
5076 	 */
5077 	if (!sfmmup->sfmmu_free)
5078 		sfmmu_check_page_sizes(sfmmup, 0);
5079 }
5080 
5081 /*
5082  * Unload all the mappings in the range [addr..addr+len). addr and len must
5083  * be MMU_PAGESIZE aligned.
5084  */
5085 void
5086 hat_unload(struct hat *sfmmup, caddr_t addr, size_t len, uint_t flags)
5087 {
5088 	if (sfmmup->sfmmu_xhat_provider) {
5089 		XHAT_UNLOAD(sfmmup, addr, len, flags);
5090 		return;
5091 	}
5092 	hat_unload_callback(sfmmup, addr, len, flags, NULL);
5093 }
5094 
5095 
5096 /*
5097  * Find the largest mapping size for this page.
5098  */
5099 static int
5100 fnd_mapping_sz(page_t *pp)
5101 {
5102 	int sz;
5103 	int p_index;
5104 
5105 	p_index = PP_MAPINDEX(pp);
5106 
5107 	sz = 0;
5108 	p_index >>= 1;	/* don't care about 8K bit */
5109 	for (; p_index; p_index >>= 1) {
5110 		sz++;
5111 	}
5112 
5113 	return (sz);
5114 }
5115 
5116 /*
5117  * This function unloads a range of addresses for an hmeblk.
5118  * It returns the next address to be unloaded.
5119  * It should be called with the hash lock held.
5120  */
5121 static caddr_t
5122 sfmmu_hblk_unload(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5123 	caddr_t endaddr, demap_range_t *dmrp, uint_t flags)
5124 {
5125 	tte_t	tte, ttemod;
5126 	struct	sf_hment *sfhmep;
5127 	int	ttesz;
5128 	long	ttecnt;
5129 	page_t *pp;
5130 	kmutex_t *pml;
5131 	int ret;
5132 	int use_demap_range;
5133 
5134 	ASSERT(in_hblk_range(hmeblkp, addr));
5135 	ASSERT(!hmeblkp->hblk_shw_bit);
5136 #ifdef DEBUG
5137 	if (get_hblk_ttesz(hmeblkp) != TTE8K &&
5138 	    (endaddr < get_hblk_endaddr(hmeblkp))) {
5139 		panic("sfmmu_hblk_unload: partial unload of large page");
5140 	}
5141 #endif /* DEBUG */
5142 
5143 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5144 	ttesz = get_hblk_ttesz(hmeblkp);
5145 
5146 	use_demap_range = (do_virtual_coloring &&
5147 				TTEBYTES(ttesz) == DEMAP_RANGE_PGSZ(dmrp));
5148 	if (use_demap_range) {
5149 		DEMAP_RANGE_CONTINUE(dmrp, addr, endaddr);
5150 	} else {
5151 		DEMAP_RANGE_FLUSH(dmrp);
5152 	}
5153 	ttecnt = 0;
5154 	HBLKTOHME(sfhmep, hmeblkp, addr);
5155 
5156 	while (addr < endaddr) {
5157 		pml = NULL;
5158 again:
5159 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5160 		if (TTE_IS_VALID(&tte)) {
5161 			pp = sfhmep->hme_page;
5162 			if (pp && pml == NULL) {
5163 				pml = sfmmu_mlist_enter(pp);
5164 			}
5165 
5166 			/*
5167 			 * Verify if hme still points to 'pp' now that
5168 			 * we have p_mapping lock.
5169 			 */
5170 			if (sfhmep->hme_page != pp) {
5171 				if (pp != NULL && sfhmep->hme_page != NULL) {
5172 					if (pml) {
5173 						sfmmu_mlist_exit(pml);
5174 					}
5175 					/* Re-start this iteration. */
5176 					continue;
5177 				}
5178 				ASSERT((pp != NULL) &&
5179 				    (sfhmep->hme_page == NULL));
5180 				goto tte_unloaded;
5181 			}
5182 
5183 			/*
5184 			 * This point on we have both HASH and p_mapping
5185 			 * lock.
5186 			 */
5187 			ASSERT(pp == sfhmep->hme_page);
5188 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5189 
5190 			/*
5191 			 * We need to loop on modify tte because it is
5192 			 * possible for pagesync to come along and
5193 			 * change the software bits beneath us.
5194 			 *
5195 			 * Page_unload can also invalidate the tte after
5196 			 * we read tte outside of p_mapping lock.
5197 			 */
5198 			ttemod = tte;
5199 
5200 			TTE_SET_INVALID(&ttemod);
5201 			ret = sfmmu_modifytte_try(&tte, &ttemod,
5202 			    &sfhmep->hme_tte);
5203 
5204 			if (ret <= 0) {
5205 				if (TTE_IS_VALID(&tte)) {
5206 					goto again;
5207 				} else {
5208 					/*
5209 					 * We read in a valid pte, but it
5210 					 * is unloaded by page_unload.
5211 					 * hme_page has become NULL and
5212 					 * we hold no p_mapping lock.
5213 					 */
5214 					ASSERT(pp == NULL && pml == NULL);
5215 					goto tte_unloaded;
5216 				}
5217 			}
5218 
5219 			if (!(flags & HAT_UNLOAD_NOSYNC)) {
5220 				sfmmu_ttesync(sfmmup, addr, &tte, pp);
5221 			}
5222 
5223 			/*
5224 			 * Ok- we invalidated the tte. Do the rest of the job.
5225 			 */
5226 			ttecnt++;
5227 
5228 			if (flags & HAT_UNLOAD_UNLOCK) {
5229 				ASSERT(hmeblkp->hblk_lckcnt > 0);
5230 				atomic_add_16(&hmeblkp->hblk_lckcnt, -1);
5231 				HBLK_STACK_TRACE(hmeblkp, HBLK_UNLOCK);
5232 			}
5233 
5234 			/*
5235 			 * Normally we would need to flush the page
5236 			 * from the virtual cache at this point in
5237 			 * order to prevent a potential cache alias
5238 			 * inconsistency.
5239 			 * The particular scenario we need to worry
5240 			 * about is:
5241 			 * Given:  va1 and va2 are two virtual address
5242 			 * that alias and map the same physical
5243 			 * address.
5244 			 * 1.	mapping exists from va1 to pa and data
5245 			 * has been read into the cache.
5246 			 * 2.	unload va1.
5247 			 * 3.	load va2 and modify data using va2.
5248 			 * 4	unload va2.
5249 			 * 5.	load va1 and reference data.  Unless we
5250 			 * flush the data cache when we unload we will
5251 			 * get stale data.
5252 			 * Fortunately, page coloring eliminates the
5253 			 * above scenario by remembering the color a
5254 			 * physical page was last or is currently
5255 			 * mapped to.  Now, we delay the flush until
5256 			 * the loading of translations.  Only when the
5257 			 * new translation is of a different color
5258 			 * are we forced to flush.
5259 			 */
5260 			if (use_demap_range) {
5261 				/*
5262 				 * Mark this page as needing a demap.
5263 				 */
5264 				DEMAP_RANGE_MARKPG(dmrp, addr);
5265 			} else {
5266 				if (do_virtual_coloring) {
5267 					sfmmu_tlb_demap(addr, sfmmup, hmeblkp,
5268 					    sfmmup->sfmmu_free, 0);
5269 				} else {
5270 					pfn_t pfnum;
5271 
5272 					pfnum = TTE_TO_PFN(addr, &tte);
5273 					sfmmu_tlbcache_demap(addr, sfmmup,
5274 					    hmeblkp, pfnum, sfmmup->sfmmu_free,
5275 					    FLUSH_NECESSARY_CPUS,
5276 					    CACHE_FLUSH, 0);
5277 				}
5278 			}
5279 
5280 			if (pp) {
5281 				/*
5282 				 * Remove the hment from the mapping list
5283 				 */
5284 				ASSERT(hmeblkp->hblk_hmecnt > 0);
5285 
5286 				/*
5287 				 * Again, we cannot
5288 				 * ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS);
5289 				 */
5290 				HME_SUB(sfhmep, pp);
5291 				membar_stst();
5292 				atomic_add_16(&hmeblkp->hblk_hmecnt, -1);
5293 			}
5294 
5295 			ASSERT(hmeblkp->hblk_vcnt > 0);
5296 			atomic_add_16(&hmeblkp->hblk_vcnt, -1);
5297 
5298 			ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
5299 			    !hmeblkp->hblk_lckcnt);
5300 
5301 			if (pp && (pp->p_nrm & (P_KPMC | P_KPMS | P_TNC))) {
5302 				if (PP_ISTNC(pp)) {
5303 					/*
5304 					 * If page was temporary
5305 					 * uncached, try to recache
5306 					 * it. Note that HME_SUB() was
5307 					 * called above so p_index and
5308 					 * mlist had been updated.
5309 					 */
5310 					conv_tnc(pp, ttesz);
5311 				} else if (pp->p_mapping == NULL) {
5312 					ASSERT(kpm_enable);
5313 					/*
5314 					 * Page is marked to be in VAC conflict
5315 					 * to an existing kpm mapping and/or is
5316 					 * kpm mapped using only the regular
5317 					 * pagesize.
5318 					 */
5319 					sfmmu_kpm_hme_unload(pp);
5320 				}
5321 			}
5322 		} else if ((pp = sfhmep->hme_page) != NULL) {
5323 				/*
5324 				 * TTE is invalid but the hme
5325 				 * still exists. let pageunload
5326 				 * complete its job.
5327 				 */
5328 				ASSERT(pml == NULL);
5329 				pml = sfmmu_mlist_enter(pp);
5330 				if (sfhmep->hme_page != NULL) {
5331 					sfmmu_mlist_exit(pml);
5332 					pml = NULL;
5333 					goto again;
5334 				}
5335 				ASSERT(sfhmep->hme_page == NULL);
5336 		} else if (hmeblkp->hblk_hmecnt != 0) {
5337 			/*
5338 			 * pageunload may have not finished decrementing
5339 			 * hblk_vcnt and hblk_hmecnt. Find page_t if any and
5340 			 * wait for pageunload to finish. Rely on pageunload
5341 			 * to decrement hblk_hmecnt after hblk_vcnt.
5342 			 */
5343 			pfn_t pfn = TTE_TO_TTEPFN(&tte);
5344 			ASSERT(pml == NULL);
5345 			if (pf_is_memory(pfn)) {
5346 				pp = page_numtopp_nolock(pfn);
5347 				if (pp != NULL) {
5348 					pml = sfmmu_mlist_enter(pp);
5349 					sfmmu_mlist_exit(pml);
5350 					pml = NULL;
5351 				}
5352 			}
5353 		}
5354 
5355 tte_unloaded:
5356 		/*
5357 		 * At this point, the tte we are looking at
5358 		 * should be unloaded, and hme has been unlinked
5359 		 * from page too. This is important because in
5360 		 * pageunload, it does ttesync() then HME_SUB.
5361 		 * We need to make sure HME_SUB has been completed
5362 		 * so we know ttesync() has been completed. Otherwise,
5363 		 * at exit time, after return from hat layer, VM will
5364 		 * release as structure which hat_setstat() (called
5365 		 * by ttesync()) needs.
5366 		 */
5367 #ifdef DEBUG
5368 		{
5369 			tte_t	dtte;
5370 
5371 			ASSERT(sfhmep->hme_page == NULL);
5372 
5373 			sfmmu_copytte(&sfhmep->hme_tte, &dtte);
5374 			ASSERT(!TTE_IS_VALID(&dtte));
5375 		}
5376 #endif
5377 
5378 		if (pml) {
5379 			sfmmu_mlist_exit(pml);
5380 		}
5381 
5382 		addr += TTEBYTES(ttesz);
5383 		sfhmep++;
5384 		DEMAP_RANGE_NEXTPG(dmrp);
5385 	}
5386 	if (ttecnt > 0)
5387 		atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -ttecnt);
5388 	return (addr);
5389 }
5390 
5391 /*
5392  * Synchronize all the mappings in the range [addr..addr+len).
5393  * Can be called with clearflag having two states:
5394  * HAT_SYNC_DONTZERO means just return the rm stats
5395  * HAT_SYNC_ZERORM means zero rm bits in the tte and return the stats
5396  */
5397 void
5398 hat_sync(struct hat *sfmmup, caddr_t addr, size_t len, uint_t clearflag)
5399 {
5400 	struct hmehash_bucket *hmebp;
5401 	hmeblk_tag hblktag;
5402 	int hmeshift, hashno = 1;
5403 	struct hme_blk *hmeblkp, *list = NULL;
5404 	caddr_t endaddr;
5405 	cpuset_t cpuset;
5406 
5407 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
5408 	ASSERT((sfmmup == ksfmmup) ||
5409 		AS_LOCK_HELD(sfmmup->sfmmu_as, &sfmmup->sfmmu_as->a_lock));
5410 	ASSERT((len & MMU_PAGEOFFSET) == 0);
5411 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
5412 		(clearflag == HAT_SYNC_ZERORM));
5413 
5414 	CPUSET_ZERO(cpuset);
5415 
5416 	endaddr = addr + len;
5417 	hblktag.htag_id = sfmmup;
5418 	/*
5419 	 * Spitfire supports 4 page sizes.
5420 	 * Most pages are expected to be of the smallest page
5421 	 * size (8K) and these will not need to be rehashed. 64K
5422 	 * pages also don't need to be rehashed because the an hmeblk
5423 	 * spans 64K of address space. 512K pages might need 1 rehash and
5424 	 * and 4M pages 2 rehashes.
5425 	 */
5426 	while (addr < endaddr) {
5427 		hmeshift = HME_HASH_SHIFT(hashno);
5428 		hblktag.htag_bspage = HME_HASH_BSPAGE(addr, hmeshift);
5429 		hblktag.htag_rehash = hashno;
5430 		hmebp = HME_HASH_FUNCTION(sfmmup, addr, hmeshift);
5431 
5432 		SFMMU_HASH_LOCK(hmebp);
5433 
5434 		HME_HASH_SEARCH(hmebp, hblktag, hmeblkp, &list);
5435 		if (hmeblkp != NULL) {
5436 			/*
5437 			 * We've encountered a shadow hmeblk so skip the range
5438 			 * of the next smaller mapping size.
5439 			 */
5440 			if (hmeblkp->hblk_shw_bit) {
5441 				ASSERT(sfmmup != ksfmmup);
5442 				ASSERT(hashno > 1);
5443 				addr = (caddr_t)P2END((uintptr_t)addr,
5444 					    TTEBYTES(hashno - 1));
5445 			} else {
5446 				addr = sfmmu_hblk_sync(sfmmup, hmeblkp,
5447 				    addr, endaddr, clearflag);
5448 			}
5449 			SFMMU_HASH_UNLOCK(hmebp);
5450 			hashno = 1;
5451 			continue;
5452 		}
5453 		SFMMU_HASH_UNLOCK(hmebp);
5454 
5455 		if (!HME_REHASH(sfmmup) || (hashno >= mmu_hashcnt)) {
5456 			/*
5457 			 * We have traversed the whole list and rehashed
5458 			 * if necessary without finding the address to sync.
5459 			 * This is ok so we increment the address by the
5460 			 * smallest hmeblk range for kernel mappings and the
5461 			 * largest hmeblk range, to account for shadow hmeblks,
5462 			 * for user mappings and continue.
5463 			 */
5464 			if (sfmmup == ksfmmup)
5465 				addr = (caddr_t)P2END((uintptr_t)addr,
5466 					    TTEBYTES(1));
5467 			else
5468 				addr = (caddr_t)P2END((uintptr_t)addr,
5469 					    TTEBYTES(hashno));
5470 			hashno = 1;
5471 		} else {
5472 			hashno++;
5473 		}
5474 	}
5475 	sfmmu_hblks_list_purge(&list);
5476 	cpuset = sfmmup->sfmmu_cpusran;
5477 	xt_sync(cpuset);
5478 }
5479 
5480 static caddr_t
5481 sfmmu_hblk_sync(struct hat *sfmmup, struct hme_blk *hmeblkp, caddr_t addr,
5482 	caddr_t endaddr, int clearflag)
5483 {
5484 	tte_t	tte, ttemod;
5485 	struct sf_hment *sfhmep;
5486 	int ttesz;
5487 	struct page *pp;
5488 	kmutex_t *pml;
5489 	int ret;
5490 
5491 	ASSERT(hmeblkp->hblk_shw_bit == 0);
5492 
5493 	endaddr = MIN(endaddr, get_hblk_endaddr(hmeblkp));
5494 
5495 	ttesz = get_hblk_ttesz(hmeblkp);
5496 	HBLKTOHME(sfhmep, hmeblkp, addr);
5497 
5498 	while (addr < endaddr) {
5499 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5500 		if (TTE_IS_VALID(&tte)) {
5501 			pml = NULL;
5502 			pp = sfhmep->hme_page;
5503 			if (pp) {
5504 				pml = sfmmu_mlist_enter(pp);
5505 			}
5506 			if (pp != sfhmep->hme_page) {
5507 				/*
5508 				 * tte most have been unloaded
5509 				 * underneath us.  Recheck
5510 				 */
5511 				ASSERT(pml);
5512 				sfmmu_mlist_exit(pml);
5513 				continue;
5514 			}
5515 
5516 			ASSERT(pp == NULL || sfmmu_mlist_held(pp));
5517 
5518 			if (clearflag == HAT_SYNC_ZERORM) {
5519 				ttemod = tte;
5520 				TTE_CLR_RM(&ttemod);
5521 				ret = sfmmu_modifytte_try(&tte, &ttemod,
5522 				    &sfhmep->hme_tte);
5523 				if (ret < 0) {
5524 					if (pml) {
5525 						sfmmu_mlist_exit(pml);
5526 					}
5527 					continue;
5528 				}
5529 
5530 				if (ret > 0) {
5531 					sfmmu_tlb_demap(addr, sfmmup,
5532 						hmeblkp, 0, 0);
5533 				}
5534 			}
5535 			sfmmu_ttesync(sfmmup, addr, &tte, pp);
5536 			if (pml) {
5537 				sfmmu_mlist_exit(pml);
5538 			}
5539 		}
5540 		addr += TTEBYTES(ttesz);
5541 		sfhmep++;
5542 	}
5543 	return (addr);
5544 }
5545 
5546 /*
5547  * This function will sync a tte to the page struct and it will
5548  * update the hat stats. Currently it allows us to pass a NULL pp
5549  * and we will simply update the stats.  We may want to change this
5550  * so we only keep stats for pages backed by pp's.
5551  */
5552 static void
5553 sfmmu_ttesync(struct hat *sfmmup, caddr_t addr, tte_t *ttep, page_t *pp)
5554 {
5555 	uint_t rm = 0;
5556 	int   	sz;
5557 	pgcnt_t	npgs;
5558 
5559 	ASSERT(TTE_IS_VALID(ttep));
5560 
5561 	if (TTE_IS_NOSYNC(ttep)) {
5562 		return;
5563 	}
5564 
5565 	if (TTE_IS_REF(ttep))  {
5566 		rm = P_REF;
5567 	}
5568 	if (TTE_IS_MOD(ttep))  {
5569 		rm |= P_MOD;
5570 	}
5571 
5572 	if (rm == 0) {
5573 		return;
5574 	}
5575 
5576 	sz = TTE_CSZ(ttep);
5577 	if (sfmmup->sfmmu_rmstat) {
5578 		int i;
5579 		caddr_t	vaddr = addr;
5580 
5581 		for (i = 0; i < TTEPAGES(sz); i++, vaddr += MMU_PAGESIZE) {
5582 			hat_setstat(sfmmup->sfmmu_as, vaddr, MMU_PAGESIZE, rm);
5583 		}
5584 
5585 	}
5586 
5587 	/*
5588 	 * XXX I want to use cas to update nrm bits but they
5589 	 * currently belong in common/vm and not in hat where
5590 	 * they should be.
5591 	 * The nrm bits are protected by the same mutex as
5592 	 * the one that protects the page's mapping list.
5593 	 */
5594 	if (!pp)
5595 		return;
5596 	ASSERT(sfmmu_mlist_held(pp));
5597 	/*
5598 	 * If the tte is for a large page, we need to sync all the
5599 	 * pages covered by the tte.
5600 	 */
5601 	if (sz != TTE8K) {
5602 		ASSERT(pp->p_szc != 0);
5603 		pp = PP_GROUPLEADER(pp, sz);
5604 		ASSERT(sfmmu_mlist_held(pp));
5605 	}
5606 
5607 	/* Get number of pages from tte size. */
5608 	npgs = TTEPAGES(sz);
5609 
5610 	do {
5611 		ASSERT(pp);
5612 		ASSERT(sfmmu_mlist_held(pp));
5613 		if (((rm & P_REF) != 0 && !PP_ISREF(pp)) ||
5614 		    ((rm & P_MOD) != 0 && !PP_ISMOD(pp)))
5615 			hat_page_setattr(pp, rm);
5616 
5617 		/*
5618 		 * Are we done? If not, we must have a large mapping.
5619 		 * For large mappings we need to sync the rest of the pages
5620 		 * covered by this tte; goto the next page.
5621 		 */
5622 	} while (--npgs > 0 && (pp = PP_PAGENEXT(pp)));
5623 }
5624 
5625 /*
5626  * Execute pre-callback handler of each pa_hment linked to pp
5627  *
5628  * Inputs:
5629  *   flag: either HAT_PRESUSPEND or HAT_SUSPEND.
5630  *   capture_cpus: pointer to return value (below)
5631  *
5632  * Returns:
5633  *   Propagates the subsystem callback return values back to the caller;
5634  *   returns 0 on success.  If capture_cpus is non-NULL, the value returned
5635  *   is zero if all of the pa_hments are of a type that do not require
5636  *   capturing CPUs prior to suspending the mapping, else it is 1.
5637  */
5638 static int
5639 hat_pageprocess_precallbacks(struct page *pp, uint_t flag, int *capture_cpus)
5640 {
5641 	struct sf_hment	*sfhmep;
5642 	struct pa_hment *pahmep;
5643 	int (*f)(caddr_t, uint_t, uint_t, void *);
5644 	int		ret;
5645 	id_t		id;
5646 	int		locked = 0;
5647 	kmutex_t	*pml;
5648 
5649 	ASSERT(PAGE_EXCL(pp));
5650 	if (!sfmmu_mlist_held(pp)) {
5651 		pml = sfmmu_mlist_enter(pp);
5652 		locked = 1;
5653 	}
5654 
5655 	if (capture_cpus)
5656 		*capture_cpus = 0;
5657 
5658 top:
5659 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
5660 		/*
5661 		 * skip sf_hments corresponding to VA<->PA mappings;
5662 		 * for pa_hment's, hme_tte.ll is zero
5663 		 */
5664 		if (!IS_PAHME(sfhmep))
5665 			continue;
5666 
5667 		pahmep = sfhmep->hme_data;
5668 		ASSERT(pahmep != NULL);
5669 
5670 		/*
5671 		 * skip if pre-handler has been called earlier in this loop
5672 		 */
5673 		if (pahmep->flags & flag)
5674 			continue;
5675 
5676 		id = pahmep->cb_id;
5677 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
5678 		if (capture_cpus && sfmmu_cb_table[id].capture_cpus != 0)
5679 			*capture_cpus = 1;
5680 		if ((f = sfmmu_cb_table[id].prehandler) == NULL) {
5681 			pahmep->flags |= flag;
5682 			continue;
5683 		}
5684 
5685 		/*
5686 		 * Drop the mapping list lock to avoid locking order issues.
5687 		 */
5688 		if (locked)
5689 			sfmmu_mlist_exit(pml);
5690 
5691 		ret = f(pahmep->addr, pahmep->len, flag, pahmep->pvt);
5692 		if (ret != 0)
5693 			return (ret);	/* caller must do the cleanup */
5694 
5695 		if (locked) {
5696 			pml = sfmmu_mlist_enter(pp);
5697 			pahmep->flags |= flag;
5698 			goto top;
5699 		}
5700 
5701 		pahmep->flags |= flag;
5702 	}
5703 
5704 	if (locked)
5705 		sfmmu_mlist_exit(pml);
5706 
5707 	return (0);
5708 }
5709 
5710 /*
5711  * Execute post-callback handler of each pa_hment linked to pp
5712  *
5713  * Same overall assumptions and restrictions apply as for
5714  * hat_pageprocess_precallbacks().
5715  */
5716 static void
5717 hat_pageprocess_postcallbacks(struct page *pp, uint_t flag)
5718 {
5719 	pfn_t pgpfn = pp->p_pagenum;
5720 	pfn_t pgmask = btop(page_get_pagesize(pp->p_szc)) - 1;
5721 	pfn_t newpfn;
5722 	struct sf_hment *sfhmep;
5723 	struct pa_hment *pahmep;
5724 	int (*f)(caddr_t, uint_t, uint_t, void *, pfn_t);
5725 	id_t	id;
5726 	int	locked = 0;
5727 	kmutex_t *pml;
5728 
5729 	ASSERT(PAGE_EXCL(pp));
5730 	if (!sfmmu_mlist_held(pp)) {
5731 		pml = sfmmu_mlist_enter(pp);
5732 		locked = 1;
5733 	}
5734 
5735 top:
5736 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
5737 		/*
5738 		 * skip sf_hments corresponding to VA<->PA mappings;
5739 		 * for pa_hment's, hme_tte.ll is zero
5740 		 */
5741 		if (!IS_PAHME(sfhmep))
5742 			continue;
5743 
5744 		pahmep = sfhmep->hme_data;
5745 		ASSERT(pahmep != NULL);
5746 
5747 		if ((pahmep->flags & flag) == 0)
5748 			continue;
5749 
5750 		pahmep->flags &= ~flag;
5751 
5752 		id = pahmep->cb_id;
5753 		ASSERT(id >= (id_t)0 && id < sfmmu_cb_nextid);
5754 		if ((f = sfmmu_cb_table[id].posthandler) == NULL)
5755 			continue;
5756 
5757 		/*
5758 		 * Convert the base page PFN into the constituent PFN
5759 		 * which is needed by the callback handler.
5760 		 */
5761 		newpfn = pgpfn | (btop((uintptr_t)pahmep->addr) & pgmask);
5762 
5763 		/*
5764 		 * Drop the mapping list lock to avoid locking order issues.
5765 		 */
5766 		if (locked)
5767 			sfmmu_mlist_exit(pml);
5768 
5769 		if (f(pahmep->addr, pahmep->len, flag, pahmep->pvt, newpfn)
5770 		    != 0)
5771 			panic("sfmmu: posthandler failed");
5772 
5773 		if (locked) {
5774 			pml = sfmmu_mlist_enter(pp);
5775 			goto top;
5776 		}
5777 	}
5778 
5779 	if (locked)
5780 		sfmmu_mlist_exit(pml);
5781 }
5782 
5783 /*
5784  * Suspend locked kernel mapping
5785  */
5786 void
5787 hat_pagesuspend(struct page *pp)
5788 {
5789 	struct sf_hment *sfhmep;
5790 	sfmmu_t *sfmmup;
5791 	tte_t tte, ttemod;
5792 	struct hme_blk *hmeblkp;
5793 	caddr_t addr;
5794 	int index, cons;
5795 	cpuset_t cpuset;
5796 
5797 	ASSERT(PAGE_EXCL(pp));
5798 	ASSERT(sfmmu_mlist_held(pp));
5799 
5800 	mutex_enter(&kpr_suspendlock);
5801 
5802 	/*
5803 	 * Call into dtrace to tell it we're about to suspend a
5804 	 * kernel mapping. This prevents us from running into issues
5805 	 * with probe context trying to touch a suspended page
5806 	 * in the relocation codepath itself.
5807 	 */
5808 	if (dtrace_kreloc_init)
5809 		(*dtrace_kreloc_init)();
5810 
5811 	index = PP_MAPINDEX(pp);
5812 	cons = TTE8K;
5813 
5814 retry:
5815 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
5816 
5817 		if (IS_PAHME(sfhmep))
5818 			continue;
5819 
5820 		if (get_hblk_ttesz(sfmmu_hmetohblk(sfhmep)) != cons)
5821 			continue;
5822 
5823 		/*
5824 		 * Loop until we successfully set the suspend bit in
5825 		 * the TTE.
5826 		 */
5827 again:
5828 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
5829 		ASSERT(TTE_IS_VALID(&tte));
5830 
5831 		ttemod = tte;
5832 		TTE_SET_SUSPEND(&ttemod);
5833 		if (sfmmu_modifytte_try(&tte, &ttemod,
5834 		    &sfhmep->hme_tte) < 0)
5835 			goto again;
5836 
5837 		/*
5838 		 * Invalidate TSB entry
5839 		 */
5840 		hmeblkp = sfmmu_hmetohblk(sfhmep);
5841 
5842 		sfmmup = hblktosfmmu(hmeblkp);
5843 		ASSERT(sfmmup == ksfmmup);
5844 
5845 		addr = tte_to_vaddr(hmeblkp, tte);
5846 
5847 		/*
5848 		 * No need to make sure that the TSB for this sfmmu is
5849 		 * not being relocated since it is ksfmmup and thus it
5850 		 * will never be relocated.
5851 		 */
5852 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp);
5853 
5854 		/*
5855 		 * Update xcall stats
5856 		 */
5857 		cpuset = cpu_ready_set;
5858 		CPUSET_DEL(cpuset, CPU->cpu_id);
5859 
5860 		/* LINTED: constant in conditional context */
5861 		SFMMU_XCALL_STATS(KCONTEXT);
5862 
5863 		/*
5864 		 * Flush TLB entry on remote CPU's
5865 		 */
5866 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, KCONTEXT);
5867 		xt_sync(cpuset);
5868 
5869 		/*
5870 		 * Flush TLB entry on local CPU
5871 		 */
5872 		vtag_flushpage(addr, KCONTEXT);
5873 	}
5874 
5875 	while (index != 0) {
5876 		index = index >> 1;
5877 		if (index != 0)
5878 			cons++;
5879 		if (index & 0x1) {
5880 			pp = PP_GROUPLEADER(pp, cons);
5881 			goto retry;
5882 		}
5883 	}
5884 }
5885 
5886 #ifdef	DEBUG
5887 
5888 #define	N_PRLE	1024
5889 struct prle {
5890 	page_t *targ;
5891 	page_t *repl;
5892 	int status;
5893 	int pausecpus;
5894 	hrtime_t whence;
5895 };
5896 
5897 static struct prle page_relocate_log[N_PRLE];
5898 static int prl_entry;
5899 static kmutex_t prl_mutex;
5900 
5901 #define	PAGE_RELOCATE_LOG(t, r, s, p)					\
5902 	mutex_enter(&prl_mutex);					\
5903 	page_relocate_log[prl_entry].targ = *(t);			\
5904 	page_relocate_log[prl_entry].repl = *(r);			\
5905 	page_relocate_log[prl_entry].status = (s);			\
5906 	page_relocate_log[prl_entry].pausecpus = (p);			\
5907 	page_relocate_log[prl_entry].whence = gethrtime();		\
5908 	prl_entry = (prl_entry == (N_PRLE - 1))? 0 : prl_entry + 1;	\
5909 	mutex_exit(&prl_mutex);
5910 
5911 #else	/* !DEBUG */
5912 #define	PAGE_RELOCATE_LOG(t, r, s, p)
5913 #endif
5914 
5915 /*
5916  * Core Kernel Page Relocation Algorithm
5917  *
5918  * Input:
5919  *
5920  * target : 	constituent pages are SE_EXCL locked.
5921  * replacement:	constituent pages are SE_EXCL locked.
5922  *
5923  * Output:
5924  *
5925  * nrelocp:	number of pages relocated
5926  */
5927 int
5928 hat_page_relocate(page_t **target, page_t **replacement, spgcnt_t *nrelocp)
5929 {
5930 	page_t		*targ, *repl;
5931 	page_t		*tpp, *rpp;
5932 	kmutex_t	*low, *high;
5933 	spgcnt_t	npages, i;
5934 	page_t		*pl = NULL;
5935 	int		old_pil;
5936 	cpuset_t	cpuset;
5937 	int		cap_cpus;
5938 	int		ret;
5939 
5940 	if (hat_kpr_enabled == 0 || !kcage_on || PP_ISNORELOC(*target)) {
5941 		PAGE_RELOCATE_LOG(target, replacement, EAGAIN, -1);
5942 		return (EAGAIN);
5943 	}
5944 
5945 	mutex_enter(&kpr_mutex);
5946 	kreloc_thread = curthread;
5947 
5948 	targ = *target;
5949 	repl = *replacement;
5950 	ASSERT(repl != NULL);
5951 	ASSERT(targ->p_szc == repl->p_szc);
5952 
5953 	npages = page_get_pagecnt(targ->p_szc);
5954 
5955 	/*
5956 	 * unload VA<->PA mappings that are not locked
5957 	 */
5958 	tpp = targ;
5959 	for (i = 0; i < npages; i++) {
5960 		(void) hat_pageunload(tpp, SFMMU_KERNEL_RELOC);
5961 		tpp++;
5962 	}
5963 
5964 	/*
5965 	 * Do "presuspend" callbacks, in a context from which we can still
5966 	 * block as needed. Note that we don't hold the mapping list lock
5967 	 * of "targ" at this point due to potential locking order issues;
5968 	 * we assume that between the hat_pageunload() above and holding
5969 	 * the SE_EXCL lock that the mapping list *cannot* change at this
5970 	 * point.
5971 	 */
5972 	ret = hat_pageprocess_precallbacks(targ, HAT_PRESUSPEND, &cap_cpus);
5973 	if (ret != 0) {
5974 		/*
5975 		 * EIO translates to fatal error, for all others cleanup
5976 		 * and return EAGAIN.
5977 		 */
5978 		ASSERT(ret != EIO);
5979 		hat_pageprocess_postcallbacks(targ, HAT_POSTUNSUSPEND);
5980 		PAGE_RELOCATE_LOG(target, replacement, ret, -1);
5981 		kreloc_thread = NULL;
5982 		mutex_exit(&kpr_mutex);
5983 		return (EAGAIN);
5984 	}
5985 
5986 	/*
5987 	 * acquire p_mapping list lock for both the target and replacement
5988 	 * root pages.
5989 	 *
5990 	 * low and high refer to the need to grab the mlist locks in a
5991 	 * specific order in order to prevent race conditions.  Thus the
5992 	 * lower lock must be grabbed before the higher lock.
5993 	 *
5994 	 * This will block hat_unload's accessing p_mapping list.  Since
5995 	 * we have SE_EXCL lock, hat_memload and hat_pageunload will be
5996 	 * blocked.  Thus, no one else will be accessing the p_mapping list
5997 	 * while we suspend and reload the locked mapping below.
5998 	 */
5999 	tpp = targ;
6000 	rpp = repl;
6001 	sfmmu_mlist_reloc_enter(tpp, rpp, &low, &high);
6002 
6003 	kpreempt_disable();
6004 
6005 	/*
6006 	 * If the replacement page is of a different virtual color
6007 	 * than the page it is replacing, we need to handle the VAC
6008 	 * consistency for it just as we would if we were setting up
6009 	 * a new mapping to a page.
6010 	 */
6011 	if ((tpp->p_szc == 0) && (PP_GET_VCOLOR(rpp) != NO_VCOLOR)) {
6012 		if (tpp->p_vcolor != rpp->p_vcolor) {
6013 			sfmmu_cache_flushcolor(PP_GET_VCOLOR(rpp),
6014 			    rpp->p_pagenum);
6015 		}
6016 	}
6017 
6018 	/*
6019 	 * We raise our PIL to 13 so that we don't get captured by
6020 	 * another CPU or pinned by an interrupt thread.  We can't go to
6021 	 * PIL 14 since the nexus driver(s) may need to interrupt at
6022 	 * that level in the case of IOMMU pseudo mappings.
6023 	 */
6024 	cpuset = cpu_ready_set;
6025 	CPUSET_DEL(cpuset, CPU->cpu_id);
6026 	if (!cap_cpus || CPUSET_ISNULL(cpuset)) {
6027 		old_pil = splr(XCALL_PIL);
6028 	} else {
6029 		old_pil = -1;
6030 		xc_attention(cpuset);
6031 	}
6032 	ASSERT(getpil() == XCALL_PIL);
6033 
6034 	/*
6035 	 * Now do suspend callbacks. In the case of an IOMMU mapping
6036 	 * this will suspend all DMA activity to the page while it is
6037 	 * being relocated. Since we are well above LOCK_LEVEL and CPUs
6038 	 * may be captured at this point we should have acquired any needed
6039 	 * locks in the presuspend callback.
6040 	 */
6041 	ret = hat_pageprocess_precallbacks(targ, HAT_SUSPEND, NULL);
6042 	if (ret != 0) {
6043 		repl = targ;
6044 		goto suspend_fail;
6045 	}
6046 
6047 	/*
6048 	 * Raise the PIL yet again, this time to block all high-level
6049 	 * interrupts on this CPU. This is necessary to prevent an
6050 	 * interrupt routine from pinning the thread which holds the
6051 	 * mapping suspended and then touching the suspended page.
6052 	 *
6053 	 * Once the page is suspended we also need to be careful to
6054 	 * avoid calling any functions which touch any seg_kmem memory
6055 	 * since that memory may be backed by the very page we are
6056 	 * relocating in here!
6057 	 */
6058 	hat_pagesuspend(targ);
6059 
6060 	/*
6061 	 * Now that we are confident everybody has stopped using this page,
6062 	 * copy the page contents.  Note we use a physical copy to prevent
6063 	 * locking issues and to avoid fpRAS because we can't handle it in
6064 	 * this context.
6065 	 */
6066 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6067 		/*
6068 		 * Copy the contents of the page.
6069 		 */
6070 		ppcopy_kernel(tpp, rpp);
6071 	}
6072 
6073 	tpp = targ;
6074 	rpp = repl;
6075 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6076 		/*
6077 		 * Copy attributes.  VAC consistency was handled above,
6078 		 * if required.
6079 		 */
6080 		rpp->p_nrm = tpp->p_nrm;
6081 		tpp->p_nrm = 0;
6082 		rpp->p_index = tpp->p_index;
6083 		tpp->p_index = 0;
6084 		rpp->p_vcolor = tpp->p_vcolor;
6085 	}
6086 
6087 	/*
6088 	 * First, unsuspend the page, if we set the suspend bit, and transfer
6089 	 * the mapping list from the target page to the replacement page.
6090 	 * Next process postcallbacks; since pa_hment's are linked only to the
6091 	 * p_mapping list of root page, we don't iterate over the constituent
6092 	 * pages.
6093 	 */
6094 	hat_pagereload(targ, repl);
6095 
6096 suspend_fail:
6097 	hat_pageprocess_postcallbacks(repl, HAT_UNSUSPEND);
6098 
6099 	/*
6100 	 * Now lower our PIL and release any captured CPUs since we
6101 	 * are out of the "danger zone".  After this it will again be
6102 	 * safe to acquire adaptive mutex locks, or to drop them...
6103 	 */
6104 	if (old_pil != -1) {
6105 		splx(old_pil);
6106 	} else {
6107 		xc_dismissed(cpuset);
6108 	}
6109 
6110 	kpreempt_enable();
6111 
6112 	sfmmu_mlist_reloc_exit(low, high);
6113 
6114 	/*
6115 	 * Postsuspend callbacks should drop any locks held across
6116 	 * the suspend callbacks.  As before, we don't hold the mapping
6117 	 * list lock at this point.. our assumption is that the mapping
6118 	 * list still can't change due to our holding SE_EXCL lock and
6119 	 * there being no unlocked mappings left. Hence the restriction
6120 	 * on calling context to hat_delete_callback()
6121 	 */
6122 	hat_pageprocess_postcallbacks(repl, HAT_POSTUNSUSPEND);
6123 	if (ret != 0) {
6124 		/*
6125 		 * The second presuspend call failed: we got here through
6126 		 * the suspend_fail label above.
6127 		 */
6128 		ASSERT(ret != EIO);
6129 		PAGE_RELOCATE_LOG(target, replacement, ret, cap_cpus);
6130 		kreloc_thread = NULL;
6131 		mutex_exit(&kpr_mutex);
6132 		return (EAGAIN);
6133 	}
6134 
6135 	/*
6136 	 * Now that we're out of the performance critical section we can
6137 	 * take care of updating the hash table, since we still
6138 	 * hold all the pages locked SE_EXCL at this point we
6139 	 * needn't worry about things changing out from under us.
6140 	 */
6141 	tpp = targ;
6142 	rpp = repl;
6143 	for (i = 0; i < npages; i++, tpp++, rpp++) {
6144 
6145 		/*
6146 		 * replace targ with replacement in page_hash table
6147 		 */
6148 		targ = tpp;
6149 		page_relocate_hash(rpp, targ);
6150 
6151 		/*
6152 		 * concatenate target; caller of platform_page_relocate()
6153 		 * expects target to be concatenated after returning.
6154 		 */
6155 		ASSERT(targ->p_next == targ);
6156 		ASSERT(targ->p_prev == targ);
6157 		page_list_concat(&pl, &targ);
6158 	}
6159 
6160 	ASSERT(*target == pl);
6161 	*nrelocp = npages;
6162 	PAGE_RELOCATE_LOG(target, replacement, 0, cap_cpus);
6163 	kreloc_thread = NULL;
6164 	mutex_exit(&kpr_mutex);
6165 	return (0);
6166 }
6167 
6168 /*
6169  * Called when stray pa_hments are found attached to a page which is
6170  * being freed.  Notify the subsystem which attached the pa_hment of
6171  * the error if it registered a suitable handler, else panic.
6172  */
6173 static void
6174 sfmmu_pahment_leaked(struct pa_hment *pahmep)
6175 {
6176 	id_t cb_id = pahmep->cb_id;
6177 
6178 	ASSERT(cb_id >= (id_t)0 && cb_id < sfmmu_cb_nextid);
6179 	if (sfmmu_cb_table[cb_id].errhandler != NULL) {
6180 		if (sfmmu_cb_table[cb_id].errhandler(pahmep->addr, pahmep->len,
6181 		    HAT_CB_ERR_LEAKED, pahmep->pvt) == 0)
6182 			return;		/* non-fatal */
6183 	}
6184 	panic("pa_hment leaked: 0x%p", pahmep);
6185 }
6186 
6187 /*
6188  * Remove all mappings to page 'pp'.
6189  */
6190 int
6191 hat_pageunload(struct page *pp, uint_t forceflag)
6192 {
6193 	struct page *origpp = pp;
6194 	struct sf_hment *sfhme, *tmphme;
6195 	struct hme_blk *hmeblkp;
6196 	kmutex_t *pml, *pmtx;
6197 	cpuset_t cpuset, tset;
6198 	int index, cons;
6199 	int xhme_blks;
6200 	int pa_hments;
6201 
6202 	ASSERT(PAGE_EXCL(pp));
6203 
6204 retry_xhat:
6205 	tmphme = NULL;
6206 	xhme_blks = 0;
6207 	pa_hments = 0;
6208 	CPUSET_ZERO(cpuset);
6209 
6210 	pml = sfmmu_mlist_enter(pp);
6211 
6212 	if (pp->p_kpmref)
6213 		sfmmu_kpm_pageunload(pp);
6214 	ASSERT(!PP_ISMAPPED_KPM(pp));
6215 
6216 	index = PP_MAPINDEX(pp);
6217 	cons = TTE8K;
6218 retry:
6219 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
6220 		tmphme = sfhme->hme_next;
6221 
6222 		if (IS_PAHME(sfhme)) {
6223 			ASSERT(sfhme->hme_data != NULL);
6224 			pa_hments++;
6225 			continue;
6226 		}
6227 
6228 		hmeblkp = sfmmu_hmetohblk(sfhme);
6229 		if (hmeblkp->hblk_xhat_bit) {
6230 			struct xhat_hme_blk *xblk =
6231 			    (struct xhat_hme_blk *)hmeblkp;
6232 
6233 			(void) XHAT_PAGEUNLOAD(xblk->xhat_hme_blk_hat,
6234 			    pp, forceflag, XBLK2PROVBLK(xblk));
6235 
6236 			xhme_blks = 1;
6237 			continue;
6238 		}
6239 
6240 		/*
6241 		 * If there are kernel mappings don't unload them, they will
6242 		 * be suspended.
6243 		 */
6244 		if (forceflag == SFMMU_KERNEL_RELOC && hmeblkp->hblk_lckcnt &&
6245 		    hmeblkp->hblk_tag.htag_id == ksfmmup)
6246 			continue;
6247 
6248 		tset = sfmmu_pageunload(pp, sfhme, cons);
6249 		CPUSET_OR(cpuset, tset);
6250 	}
6251 
6252 	while (index != 0) {
6253 		index = index >> 1;
6254 		if (index != 0)
6255 			cons++;
6256 		if (index & 0x1) {
6257 			/* Go to leading page */
6258 			pp = PP_GROUPLEADER(pp, cons);
6259 			ASSERT(sfmmu_mlist_held(pp));
6260 			goto retry;
6261 		}
6262 	}
6263 
6264 	/*
6265 	 * cpuset may be empty if the page was only mapped by segkpm,
6266 	 * in which case we won't actually cross-trap.
6267 	 */
6268 	xt_sync(cpuset);
6269 
6270 	/*
6271 	 * The page should have no mappings at this point, unless
6272 	 * we were called from hat_page_relocate() in which case we
6273 	 * leave the locked mappings which will be suspended later.
6274 	 */
6275 	ASSERT(!PP_ISMAPPED(origpp) || xhme_blks || pa_hments ||
6276 	    (forceflag == SFMMU_KERNEL_RELOC));
6277 
6278 	if (PP_ISTNC(pp)) {
6279 		if (cons == TTE8K) {
6280 			pmtx = sfmmu_page_enter(pp);
6281 			PP_CLRTNC(pp);
6282 			sfmmu_page_exit(pmtx);
6283 		} else {
6284 			conv_tnc(pp, cons);
6285 		}
6286 	}
6287 
6288 	if (pa_hments && forceflag != SFMMU_KERNEL_RELOC) {
6289 		/*
6290 		 * Unlink any pa_hments and free them, calling back
6291 		 * the responsible subsystem to notify it of the error.
6292 		 * This can occur in situations such as drivers leaking
6293 		 * DMA handles: naughty, but common enough that we'd like
6294 		 * to keep the system running rather than bringing it
6295 		 * down with an obscure error like "pa_hment leaked"
6296 		 * which doesn't aid the user in debugging their driver.
6297 		 */
6298 		for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
6299 			tmphme = sfhme->hme_next;
6300 			if (IS_PAHME(sfhme)) {
6301 				struct pa_hment *pahmep = sfhme->hme_data;
6302 				sfmmu_pahment_leaked(pahmep);
6303 				HME_SUB(sfhme, pp);
6304 				kmem_cache_free(pa_hment_cache, pahmep);
6305 			}
6306 		}
6307 
6308 		ASSERT(!PP_ISMAPPED(origpp) || xhme_blks);
6309 	}
6310 
6311 	sfmmu_mlist_exit(pml);
6312 
6313 	/*
6314 	 * XHAT may not have finished unloading pages
6315 	 * because some other thread was waiting for
6316 	 * mlist lock and XHAT_PAGEUNLOAD let it do
6317 	 * the job.
6318 	 */
6319 	if (xhme_blks) {
6320 		pp = origpp;
6321 		goto retry_xhat;
6322 	}
6323 
6324 	return (0);
6325 }
6326 
6327 static cpuset_t
6328 sfmmu_pageunload(page_t *pp, struct sf_hment *sfhme, int cons)
6329 {
6330 	struct hme_blk *hmeblkp;
6331 	sfmmu_t *sfmmup;
6332 	tte_t tte, ttemod;
6333 #ifdef DEBUG
6334 	tte_t orig_old;
6335 #endif /* DEBUG */
6336 	caddr_t addr;
6337 	int ttesz;
6338 	int ret;
6339 	cpuset_t cpuset;
6340 
6341 	ASSERT(pp != NULL);
6342 	ASSERT(sfmmu_mlist_held(pp));
6343 	ASSERT(pp->p_vnode != &kvp);
6344 
6345 	CPUSET_ZERO(cpuset);
6346 
6347 	hmeblkp = sfmmu_hmetohblk(sfhme);
6348 
6349 readtte:
6350 	sfmmu_copytte(&sfhme->hme_tte, &tte);
6351 	if (TTE_IS_VALID(&tte)) {
6352 		sfmmup = hblktosfmmu(hmeblkp);
6353 		ttesz = get_hblk_ttesz(hmeblkp);
6354 		/*
6355 		 * Only unload mappings of 'cons' size.
6356 		 */
6357 		if (ttesz != cons)
6358 			return (cpuset);
6359 
6360 		/*
6361 		 * Note that we have p_mapping lock, but no hash lock here.
6362 		 * hblk_unload() has to have both hash lock AND p_mapping
6363 		 * lock before it tries to modify tte. So, the tte could
6364 		 * not become invalid in the sfmmu_modifytte_try() below.
6365 		 */
6366 		ttemod = tte;
6367 #ifdef DEBUG
6368 		orig_old = tte;
6369 #endif /* DEBUG */
6370 
6371 		TTE_SET_INVALID(&ttemod);
6372 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
6373 		if (ret < 0) {
6374 #ifdef DEBUG
6375 			/* only R/M bits can change. */
6376 			chk_tte(&orig_old, &tte, &ttemod, hmeblkp);
6377 #endif /* DEBUG */
6378 			goto readtte;
6379 		}
6380 
6381 		if (ret == 0) {
6382 			panic("pageunload: cas failed?");
6383 		}
6384 
6385 		addr = tte_to_vaddr(hmeblkp, tte);
6386 
6387 		sfmmu_ttesync(sfmmup, addr, &tte, pp);
6388 
6389 		atomic_add_long(&sfmmup->sfmmu_ttecnt[ttesz], -1);
6390 
6391 		/*
6392 		 * We need to flush the page from the virtual cache
6393 		 * in order to prevent a virtual cache alias
6394 		 * inconsistency. The particular scenario we need
6395 		 * to worry about is:
6396 		 * Given:  va1 and va2 are two virtual address that
6397 		 * alias and will map the same physical address.
6398 		 * 1.	mapping exists from va1 to pa and data has
6399 		 *	been read into the cache.
6400 		 * 2.	unload va1.
6401 		 * 3.	load va2 and modify data using va2.
6402 		 * 4	unload va2.
6403 		 * 5.	load va1 and reference data.  Unless we flush
6404 		 *	the data cache when we unload we will get
6405 		 *	stale data.
6406 		 * This scenario is taken care of by using virtual
6407 		 * page coloring.
6408 		 */
6409 		if (sfmmup->sfmmu_ismhat) {
6410 			/*
6411 			 * Flush TSBs, TLBs and caches
6412 			 * of every process
6413 			 * sharing this ism segment.
6414 			 */
6415 			sfmmu_hat_lock_all();
6416 			mutex_enter(&ism_mlist_lock);
6417 			kpreempt_disable();
6418 			if (do_virtual_coloring)
6419 				sfmmu_ismtlbcache_demap(addr, sfmmup, hmeblkp,
6420 					pp->p_pagenum, CACHE_NO_FLUSH);
6421 			else
6422 				sfmmu_ismtlbcache_demap(addr, sfmmup, hmeblkp,
6423 					pp->p_pagenum, CACHE_FLUSH);
6424 			kpreempt_enable();
6425 			mutex_exit(&ism_mlist_lock);
6426 			sfmmu_hat_unlock_all();
6427 			cpuset = cpu_ready_set;
6428 		} else if (do_virtual_coloring) {
6429 			sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
6430 			cpuset = sfmmup->sfmmu_cpusran;
6431 		} else {
6432 			sfmmu_tlbcache_demap(addr, sfmmup, hmeblkp,
6433 				pp->p_pagenum, 0, FLUSH_NECESSARY_CPUS,
6434 				CACHE_FLUSH, 0);
6435 			cpuset = sfmmup->sfmmu_cpusran;
6436 		}
6437 
6438 		/*
6439 		 * Hme_sub has to run after ttesync() and a_rss update.
6440 		 * See hblk_unload().
6441 		 */
6442 		HME_SUB(sfhme, pp);
6443 		membar_stst();
6444 
6445 		/*
6446 		 * We can not make ASSERT(hmeblkp->hblk_hmecnt <= NHMENTS)
6447 		 * since pteload may have done a HME_ADD() right after
6448 		 * we did the HME_SUB() above. Hmecnt is now maintained
6449 		 * by cas only. no lock guranteed its value. The only
6450 		 * gurantee we have is the hmecnt should not be less than
6451 		 * what it should be so the hblk will not be taken away.
6452 		 * It's also important that we decremented the hmecnt after
6453 		 * we are done with hmeblkp so that this hmeblk won't be
6454 		 * stolen.
6455 		 */
6456 		ASSERT(hmeblkp->hblk_hmecnt > 0);
6457 		ASSERT(hmeblkp->hblk_vcnt > 0);
6458 		atomic_add_16(&hmeblkp->hblk_vcnt, -1);
6459 		atomic_add_16(&hmeblkp->hblk_hmecnt, -1);
6460 		/*
6461 		 * This is bug 4063182.
6462 		 * XXX: fixme
6463 		 * ASSERT(hmeblkp->hblk_hmecnt || hmeblkp->hblk_vcnt ||
6464 		 *	!hmeblkp->hblk_lckcnt);
6465 		 */
6466 	} else {
6467 		panic("invalid tte? pp %p &tte %p",
6468 		    (void *)pp, (void *)&tte);
6469 	}
6470 
6471 	return (cpuset);
6472 }
6473 
6474 /*
6475  * While relocating a kernel page, this function will move the mappings
6476  * from tpp to dpp and modify any associated data with these mappings.
6477  * It also unsuspends the suspended kernel mapping.
6478  */
6479 static void
6480 hat_pagereload(struct page *tpp, struct page *dpp)
6481 {
6482 	struct sf_hment *sfhme;
6483 	tte_t tte, ttemod;
6484 	int index, cons;
6485 
6486 	ASSERT(getpil() == PIL_MAX);
6487 	ASSERT(sfmmu_mlist_held(tpp));
6488 	ASSERT(sfmmu_mlist_held(dpp));
6489 
6490 	index = PP_MAPINDEX(tpp);
6491 	cons = TTE8K;
6492 
6493 	/* Update real mappings to the page */
6494 retry:
6495 	for (sfhme = tpp->p_mapping; sfhme != NULL; sfhme = sfhme->hme_next) {
6496 		if (IS_PAHME(sfhme))
6497 			continue;
6498 		sfmmu_copytte(&sfhme->hme_tte, &tte);
6499 		ttemod = tte;
6500 
6501 		/*
6502 		 * replace old pfn with new pfn in TTE
6503 		 */
6504 		PFN_TO_TTE(ttemod, dpp->p_pagenum);
6505 
6506 		/*
6507 		 * clear suspend bit
6508 		 */
6509 		ASSERT(TTE_IS_SUSPEND(&ttemod));
6510 		TTE_CLR_SUSPEND(&ttemod);
6511 
6512 		if (sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte) < 0)
6513 			panic("hat_pagereload(): sfmmu_modifytte_try() failed");
6514 
6515 		/*
6516 		 * set hme_page point to new page
6517 		 */
6518 		sfhme->hme_page = dpp;
6519 	}
6520 
6521 	/*
6522 	 * move p_mapping list from old page to new page
6523 	 */
6524 	dpp->p_mapping = tpp->p_mapping;
6525 	tpp->p_mapping = NULL;
6526 	dpp->p_share = tpp->p_share;
6527 	tpp->p_share = 0;
6528 
6529 	while (index != 0) {
6530 		index = index >> 1;
6531 		if (index != 0)
6532 			cons++;
6533 		if (index & 0x1) {
6534 			tpp = PP_GROUPLEADER(tpp, cons);
6535 			dpp = PP_GROUPLEADER(dpp, cons);
6536 			goto retry;
6537 		}
6538 	}
6539 
6540 	if (dtrace_kreloc_fini)
6541 		(*dtrace_kreloc_fini)();
6542 	mutex_exit(&kpr_suspendlock);
6543 }
6544 
6545 uint_t
6546 hat_pagesync(struct page *pp, uint_t clearflag)
6547 {
6548 	struct sf_hment *sfhme, *tmphme = NULL;
6549 	struct hme_blk *hmeblkp;
6550 	kmutex_t *pml;
6551 	cpuset_t cpuset, tset;
6552 	int	index, cons;
6553 	extern	ulong_t po_share;
6554 	page_t	*save_pp = pp;
6555 
6556 	CPUSET_ZERO(cpuset);
6557 
6558 	if (PP_ISRO(pp) && (clearflag & HAT_SYNC_STOPON_MOD)) {
6559 		return (PP_GENERIC_ATTR(pp));
6560 	}
6561 
6562 	if ((clearflag == (HAT_SYNC_STOPON_REF | HAT_SYNC_DONTZERO)) &&
6563 	    PP_ISREF(pp)) {
6564 		return (PP_GENERIC_ATTR(pp));
6565 	}
6566 
6567 	if ((clearflag == (HAT_SYNC_STOPON_MOD | HAT_SYNC_DONTZERO)) &&
6568 	    PP_ISMOD(pp)) {
6569 		return (PP_GENERIC_ATTR(pp));
6570 	}
6571 
6572 	if ((clearflag & HAT_SYNC_STOPON_SHARED) != 0 &&
6573 	    (pp->p_share > po_share) &&
6574 	    !(clearflag & HAT_SYNC_ZERORM)) {
6575 		if (PP_ISRO(pp))
6576 			hat_page_setattr(pp, P_REF);
6577 		return (PP_GENERIC_ATTR(pp));
6578 	}
6579 
6580 	clearflag &= ~HAT_SYNC_STOPON_SHARED;
6581 	pml = sfmmu_mlist_enter(pp);
6582 	index = PP_MAPINDEX(pp);
6583 	cons = TTE8K;
6584 retry:
6585 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
6586 		/*
6587 		 * We need to save the next hment on the list since
6588 		 * it is possible for pagesync to remove an invalid hment
6589 		 * from the list.
6590 		 */
6591 		tmphme = sfhme->hme_next;
6592 		/*
6593 		 * If we are looking for large mappings and this hme doesn't
6594 		 * reach the range we are seeking, just ignore its.
6595 		 */
6596 		hmeblkp = sfmmu_hmetohblk(sfhme);
6597 		if (hmeblkp->hblk_xhat_bit)
6598 			continue;
6599 
6600 		if (hme_size(sfhme) < cons)
6601 			continue;
6602 		tset = sfmmu_pagesync(pp, sfhme,
6603 			clearflag & ~HAT_SYNC_STOPON_RM);
6604 		CPUSET_OR(cpuset, tset);
6605 		/*
6606 		 * If clearflag is HAT_SYNC_DONTZERO, break out as soon
6607 		 * as the "ref" or "mod" is set.
6608 		 */
6609 		if ((clearflag & ~HAT_SYNC_STOPON_RM) == HAT_SYNC_DONTZERO &&
6610 		    ((clearflag & HAT_SYNC_STOPON_MOD) && PP_ISMOD(save_pp)) ||
6611 		    ((clearflag & HAT_SYNC_STOPON_REF) && PP_ISREF(save_pp))) {
6612 			index = 0;
6613 			break;
6614 		}
6615 	}
6616 
6617 	while (index) {
6618 		index = index >> 1;
6619 		cons++;
6620 		if (index & 0x1) {
6621 			/* Go to leading page */
6622 			pp = PP_GROUPLEADER(pp, cons);
6623 			goto retry;
6624 		}
6625 	}
6626 
6627 	xt_sync(cpuset);
6628 	sfmmu_mlist_exit(pml);
6629 	return (PP_GENERIC_ATTR(save_pp));
6630 }
6631 
6632 /*
6633  * Get all the hardware dependent attributes for a page struct
6634  */
6635 static cpuset_t
6636 sfmmu_pagesync(struct page *pp, struct sf_hment *sfhme,
6637 	uint_t clearflag)
6638 {
6639 	caddr_t addr;
6640 	tte_t tte, ttemod;
6641 	struct hme_blk *hmeblkp;
6642 	int ret;
6643 	sfmmu_t *sfmmup;
6644 	cpuset_t cpuset;
6645 
6646 	ASSERT(pp != NULL);
6647 	ASSERT(sfmmu_mlist_held(pp));
6648 	ASSERT((clearflag == HAT_SYNC_DONTZERO) ||
6649 		(clearflag == HAT_SYNC_ZERORM));
6650 
6651 	SFMMU_STAT(sf_pagesync);
6652 
6653 	CPUSET_ZERO(cpuset);
6654 
6655 sfmmu_pagesync_retry:
6656 
6657 	sfmmu_copytte(&sfhme->hme_tte, &tte);
6658 	if (TTE_IS_VALID(&tte)) {
6659 		hmeblkp = sfmmu_hmetohblk(sfhme);
6660 		sfmmup = hblktosfmmu(hmeblkp);
6661 		addr = tte_to_vaddr(hmeblkp, tte);
6662 		if (clearflag == HAT_SYNC_ZERORM) {
6663 			ttemod = tte;
6664 			TTE_CLR_RM(&ttemod);
6665 			ret = sfmmu_modifytte_try(&tte, &ttemod,
6666 				&sfhme->hme_tte);
6667 			if (ret < 0) {
6668 				/*
6669 				 * cas failed and the new value is not what
6670 				 * we want.
6671 				 */
6672 				goto sfmmu_pagesync_retry;
6673 			}
6674 
6675 			if (ret > 0) {
6676 				/* we win the cas */
6677 				sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
6678 				cpuset = sfmmup->sfmmu_cpusran;
6679 			}
6680 		}
6681 
6682 		sfmmu_ttesync(sfmmup, addr, &tte, pp);
6683 	}
6684 	return (cpuset);
6685 }
6686 
6687 /*
6688  * Remove write permission from a mappings to a page, so that
6689  * we can detect the next modification of it. This requires modifying
6690  * the TTE then invalidating (demap) any TLB entry using that TTE.
6691  * This code is similar to sfmmu_pagesync().
6692  */
6693 static cpuset_t
6694 sfmmu_pageclrwrt(struct page *pp, struct sf_hment *sfhme)
6695 {
6696 	caddr_t addr;
6697 	tte_t tte;
6698 	tte_t ttemod;
6699 	struct hme_blk *hmeblkp;
6700 	int ret;
6701 	sfmmu_t *sfmmup;
6702 	cpuset_t cpuset;
6703 
6704 	ASSERT(pp != NULL);
6705 	ASSERT(sfmmu_mlist_held(pp));
6706 
6707 	CPUSET_ZERO(cpuset);
6708 	SFMMU_STAT(sf_clrwrt);
6709 
6710 retry:
6711 
6712 	sfmmu_copytte(&sfhme->hme_tte, &tte);
6713 	if (TTE_IS_VALID(&tte) && TTE_IS_WRITABLE(&tte)) {
6714 		hmeblkp = sfmmu_hmetohblk(sfhme);
6715 
6716 		/*
6717 		 * xhat mappings should never be to a VMODSORT page.
6718 		 */
6719 		ASSERT(hmeblkp->hblk_xhat_bit == 0);
6720 
6721 		sfmmup = hblktosfmmu(hmeblkp);
6722 		addr = tte_to_vaddr(hmeblkp, tte);
6723 
6724 		ttemod = tte;
6725 		TTE_CLR_WRT(&ttemod);
6726 		TTE_CLR_MOD(&ttemod);
6727 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
6728 
6729 		/*
6730 		 * if cas failed and the new value is not what
6731 		 * we want retry
6732 		 */
6733 		if (ret < 0)
6734 			goto retry;
6735 
6736 		/* we win the cas */
6737 		if (ret > 0) {
6738 			sfmmu_tlb_demap(addr, sfmmup, hmeblkp, 0, 0);
6739 			cpuset = sfmmup->sfmmu_cpusran;
6740 		}
6741 	}
6742 
6743 	return (cpuset);
6744 }
6745 
6746 /*
6747  * Walk all mappings of a page, removing write permission and clearing the
6748  * ref/mod bits. This code is similar to hat_pagesync()
6749  */
6750 static void
6751 hat_page_clrwrt(page_t *pp)
6752 {
6753 	struct sf_hment *sfhme;
6754 	struct sf_hment *tmphme = NULL;
6755 	kmutex_t *pml;
6756 	cpuset_t cpuset;
6757 	cpuset_t tset;
6758 	int	index;
6759 	int	 cons;
6760 
6761 	CPUSET_ZERO(cpuset);
6762 
6763 	pml = sfmmu_mlist_enter(pp);
6764 	index = PP_MAPINDEX(pp);
6765 	cons = TTE8K;
6766 retry:
6767 	for (sfhme = pp->p_mapping; sfhme; sfhme = tmphme) {
6768 		tmphme = sfhme->hme_next;
6769 
6770 		/*
6771 		 * If we are looking for large mappings and this hme doesn't
6772 		 * reach the range we are seeking, just ignore its.
6773 		 */
6774 
6775 		if (hme_size(sfhme) < cons)
6776 			continue;
6777 
6778 		tset = sfmmu_pageclrwrt(pp, sfhme);
6779 		CPUSET_OR(cpuset, tset);
6780 	}
6781 
6782 	while (index) {
6783 		index = index >> 1;
6784 		cons++;
6785 		if (index & 0x1) {
6786 			/* Go to leading page */
6787 			pp = PP_GROUPLEADER(pp, cons);
6788 			goto retry;
6789 		}
6790 	}
6791 
6792 	xt_sync(cpuset);
6793 	sfmmu_mlist_exit(pml);
6794 }
6795 
6796 /*
6797  * Set the given REF/MOD/RO bits for the given page.
6798  * For a vnode with a sorted v_pages list, we need to change
6799  * the attributes and the v_pages list together under page_vnode_mutex.
6800  */
6801 void
6802 hat_page_setattr(page_t *pp, uint_t flag)
6803 {
6804 	vnode_t		*vp = pp->p_vnode;
6805 	page_t		**listp;
6806 	kmutex_t	*pmtx;
6807 	kmutex_t	*vphm = NULL;
6808 
6809 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
6810 
6811 	/*
6812 	 * nothing to do if attribute already set
6813 	 */
6814 	if ((pp->p_nrm & flag) == flag)
6815 		return;
6816 
6817 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) {
6818 		vphm = page_vnode_mutex(vp);
6819 		mutex_enter(vphm);
6820 	}
6821 
6822 	pmtx = sfmmu_page_enter(pp);
6823 	pp->p_nrm |= flag;
6824 	sfmmu_page_exit(pmtx);
6825 
6826 	if (vphm != NULL) {
6827 		/*
6828 		 * Some File Systems examine v_pages for NULL w/o
6829 		 * grabbing the vphm mutex. Must not let it become NULL when
6830 		 * pp is the only page on the list.
6831 		 */
6832 		if (pp->p_vpnext != pp) {
6833 			page_vpsub(&vp->v_pages, pp);
6834 			if (vp->v_pages != NULL)
6835 				listp = &vp->v_pages->p_vpprev->p_vpnext;
6836 			else
6837 				listp = &vp->v_pages;
6838 			page_vpadd(listp, pp);
6839 		}
6840 		mutex_exit(vphm);
6841 	}
6842 }
6843 
6844 void
6845 hat_page_clrattr(page_t *pp, uint_t flag)
6846 {
6847 	vnode_t		*vp = pp->p_vnode;
6848 	kmutex_t	*vphm = NULL;
6849 	kmutex_t	*pmtx;
6850 
6851 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
6852 
6853 	/*
6854 	 * For vnode with a sorted v_pages list, we need to change
6855 	 * the attributes and the v_pages list together under page_vnode_mutex.
6856 	 */
6857 	if ((flag & P_MOD) != 0 && vp != NULL && IS_VMODSORT(vp)) {
6858 		vphm = page_vnode_mutex(vp);
6859 		mutex_enter(vphm);
6860 	}
6861 
6862 	pmtx = sfmmu_page_enter(pp);
6863 	pp->p_nrm &= ~flag;
6864 	sfmmu_page_exit(pmtx);
6865 
6866 	if (vphm != NULL) {
6867 		/*
6868 		 * Some File Systems examine v_pages for NULL w/o
6869 		 * grabbing the vphm mutex. Must not let it become NULL when
6870 		 * pp is the only page on the list.
6871 		 */
6872 		if (pp->p_vpnext != pp) {
6873 			page_vpsub(&vp->v_pages, pp);
6874 			page_vpadd(&vp->v_pages, pp);
6875 		}
6876 		mutex_exit(vphm);
6877 
6878 		/*
6879 		 * VMODSORT works by removing write permissions and getting
6880 		 * a fault when a page is made dirty. At this point
6881 		 * we need to remove write permission from all mappings
6882 		 * to this page.
6883 		 */
6884 		hat_page_clrwrt(pp);
6885 	}
6886 }
6887 
6888 
6889 uint_t
6890 hat_page_getattr(page_t *pp, uint_t flag)
6891 {
6892 	ASSERT(!(flag & ~(P_MOD | P_REF | P_RO)));
6893 	return ((uint_t)(pp->p_nrm & flag));
6894 }
6895 
6896 /*
6897  * DEBUG kernels: verify that a kernel va<->pa translation
6898  * is safe by checking the underlying page_t is in a page
6899  * relocation-safe state.
6900  */
6901 #ifdef	DEBUG
6902 void
6903 sfmmu_check_kpfn(pfn_t pfn)
6904 {
6905 	page_t *pp;
6906 	int index, cons;
6907 
6908 	if (hat_check_vtop == 0)
6909 		return;
6910 
6911 	if (hat_kpr_enabled == 0 || kvseg.s_base == NULL || panicstr)
6912 		return;
6913 
6914 	pp = page_numtopp_nolock(pfn);
6915 	if (!pp)
6916 		return;
6917 
6918 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
6919 		return;
6920 
6921 	/*
6922 	 * Handed a large kernel page, we dig up the root page since we
6923 	 * know the root page might have the lock also.
6924 	 */
6925 	if (pp->p_szc != 0) {
6926 		index = PP_MAPINDEX(pp);
6927 		cons = TTE8K;
6928 again:
6929 		while (index != 0) {
6930 			index >>= 1;
6931 			if (index != 0)
6932 				cons++;
6933 			if (index & 0x1) {
6934 				pp = PP_GROUPLEADER(pp, cons);
6935 				goto again;
6936 			}
6937 		}
6938 	}
6939 
6940 	if (PAGE_LOCKED(pp) || PP_ISNORELOC(pp))
6941 		return;
6942 
6943 	/*
6944 	 * Pages need to be locked or allocated "permanent" (either from
6945 	 * static_arena arena or explicitly setting PG_NORELOC when calling
6946 	 * page_create_va()) for VA->PA translations to be valid.
6947 	 */
6948 	if (!PP_ISNORELOC(pp))
6949 		panic("Illegal VA->PA translation, pp 0x%p not permanent", pp);
6950 	else
6951 		panic("Illegal VA->PA translation, pp 0x%p not locked", pp);
6952 }
6953 #endif	/* DEBUG */
6954 
6955 /*
6956  * Returns a page frame number for a given virtual address.
6957  * Returns PFN_INVALID to indicate an invalid mapping
6958  */
6959 pfn_t
6960 hat_getpfnum(struct hat *hat, caddr_t addr)
6961 {
6962 	pfn_t pfn;
6963 	tte_t tte;
6964 
6965 	/*
6966 	 * We would like to
6967 	 * ASSERT(AS_LOCK_HELD(as, &as->a_lock));
6968 	 * but we can't because the iommu driver will call this
6969 	 * routine at interrupt time and it can't grab the as lock
6970 	 * or it will deadlock: A thread could have the as lock
6971 	 * and be waiting for io.  The io can't complete
6972 	 * because the interrupt thread is blocked trying to grab
6973 	 * the as lock.
6974 	 */
6975 
6976 	ASSERT(hat->sfmmu_xhat_provider == NULL);
6977 
6978 	if (hat == ksfmmup) {
6979 		if (segkpm && IS_KPM_ADDR(addr))
6980 			return (sfmmu_kpm_vatopfn(addr));
6981 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
6982 		    == PFN_SUSPENDED) {
6983 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
6984 		}
6985 		sfmmu_check_kpfn(pfn);
6986 		return (pfn);
6987 	} else {
6988 		return (sfmmu_uvatopfn(addr, hat));
6989 	}
6990 }
6991 
6992 /*
6993  * hat_getkpfnum() is an obsolete DDI routine, and its use is discouraged.
6994  * Use hat_getpfnum(kas.a_hat, ...) instead.
6995  *
6996  * We'd like to return PFN_INVALID if the mappings have underlying page_t's
6997  * but can't right now due to the fact that some software has grown to use
6998  * this interface incorrectly. So for now when the interface is misused,
6999  * return a warning to the user that in the future it won't work in the
7000  * way they're abusing it, and carry on (after disabling page relocation).
7001  */
7002 pfn_t
7003 hat_getkpfnum(caddr_t addr)
7004 {
7005 	pfn_t pfn;
7006 	tte_t tte;
7007 	int badcaller = 0;
7008 	extern int segkmem_reloc;
7009 
7010 	if (segkpm && IS_KPM_ADDR(addr)) {
7011 		badcaller = 1;
7012 		pfn = sfmmu_kpm_vatopfn(addr);
7013 	} else {
7014 		while ((pfn = sfmmu_vatopfn(addr, ksfmmup, &tte))
7015 		    == PFN_SUSPENDED) {
7016 			sfmmu_vatopfn_suspended(addr, ksfmmup, &tte);
7017 		}
7018 		badcaller = pf_is_memory(pfn);
7019 	}
7020 
7021 	if (badcaller) {
7022 		/*
7023 		 * We can't return PFN_INVALID or the caller may panic
7024 		 * or corrupt the system.  The only alternative is to
7025 		 * disable page relocation at this point for all kernel
7026 		 * memory.  This will impact any callers of page_relocate()
7027 		 * such as FMA or DR.
7028 		 *
7029 		 * RFE: Add junk here to spit out an ereport so the sysadmin
7030 		 * can be advised that he should upgrade his device driver
7031 		 * so that this doesn't happen.
7032 		 */
7033 		hat_getkpfnum_badcall(caller());
7034 		if (hat_kpr_enabled && segkmem_reloc) {
7035 			hat_kpr_enabled = 0;
7036 			segkmem_reloc = 0;
7037 			cmn_err(CE_WARN, "Kernel Page Relocation is DISABLED");
7038 		}
7039 	}
7040 	return (pfn);
7041 }
7042 
7043 pfn_t
7044 sfmmu_uvatopfn(caddr_t vaddr, struct hat *sfmmup)
7045 {
7046 	struct hmehash_bucket *hmebp;
7047 	hmeblk_tag hblktag;
7048 	int hmeshift, hashno = 1;
7049 	struct hme_blk *hmeblkp = NULL;
7050 
7051 	struct sf_hment *sfhmep;
7052 	tte_t tte;
7053 	pfn_t pfn;
7054 
7055 	/* support for ISM */
7056 	ism_map_t	*ism_map;
7057 	ism_blk_t	*ism_blkp;
7058 	int		i;
7059 	sfmmu_t *ism_hatid = NULL;
7060 	sfmmu_t *locked_hatid = NULL;
7061 
7062 
7063 	ASSERT(sfmmup != ksfmmup);
7064 	SFMMU_STAT(sf_user_vtop);
7065 	/*
7066 	 * Set ism_hatid if vaddr falls in a ISM segment.
7067 	 */
7068 	ism_blkp = sfmmup->sfmmu_iblk;
7069 	if (ism_blkp) {
7070 		sfmmu_ismhat_enter(sfmmup, 0);
7071 		locked_hatid = sfmmup;
7072 	}
7073 	while (ism_blkp && ism_hatid == NULL) {
7074 		ism_map = ism_blkp->iblk_maps;
7075 		for (i = 0; ism_map[i].imap_ismhat && i < ISM_MAP_SLOTS; i++) {
7076 			if (vaddr >= ism_start(ism_map[i]) &&
7077 			    vaddr < ism_end(ism_map[i])) {
7078 				sfmmup = ism_hatid = ism_map[i].imap_ismhat;
7079 				vaddr = (caddr_t)(vaddr -
7080 					ism_start(ism_map[i]));
7081 				break;
7082 			}
7083 		}
7084 		ism_blkp = ism_blkp->iblk_next;
7085 	}
7086 	if (locked_hatid) {
7087 		sfmmu_ismhat_exit(locked_hatid, 0);
7088 	}
7089 
7090 	hblktag.htag_id = sfmmup;
7091 	do {
7092 		hmeshift = HME_HASH_SHIFT(hashno);
7093 		hblktag.htag_bspage = HME_HASH_BSPAGE(vaddr, hmeshift);
7094 		hblktag.htag_rehash = hashno;
7095 		hmebp = HME_HASH_FUNCTION(sfmmup, vaddr, hmeshift);
7096 
7097 		SFMMU_HASH_LOCK(hmebp);
7098 
7099 		HME_HASH_FAST_SEARCH(hmebp, hblktag, hmeblkp);
7100 		if (hmeblkp != NULL) {
7101 			HBLKTOHME(sfhmep, hmeblkp, vaddr);
7102 			sfmmu_copytte(&sfhmep->hme_tte, &tte);
7103 			if (TTE_IS_VALID(&tte)) {
7104 				pfn = TTE_TO_PFN(vaddr, &tte);
7105 			} else {
7106 				pfn = PFN_INVALID;
7107 			}
7108 			SFMMU_HASH_UNLOCK(hmebp);
7109 			return (pfn);
7110 		}
7111 		SFMMU_HASH_UNLOCK(hmebp);
7112 		hashno++;
7113 	} while (HME_REHASH(sfmmup) && (hashno <= mmu_hashcnt));
7114 	return (PFN_INVALID);
7115 }
7116 
7117 
7118 /*
7119  * For compatability with AT&T and later optimizations
7120  */
7121 /* ARGSUSED */
7122 void
7123 hat_map(struct hat *hat, caddr_t addr, size_t len, uint_t flags)
7124 {
7125 	ASSERT(hat != NULL);
7126 	ASSERT(hat->sfmmu_xhat_provider == NULL);
7127 }
7128 
7129 /*
7130  * Return the number of mappings to a particular page.
7131  * This number is an approximation of the number of
7132  * number of people sharing the page.
7133  */
7134 ulong_t
7135 hat_page_getshare(page_t *pp)
7136 {
7137 	page_t *spp = pp;	/* start page */
7138 	kmutex_t *pml;
7139 	ulong_t	cnt;
7140 	int index, sz = TTE64K;
7141 
7142 	/*
7143 	 * We need to grab the mlist lock to make sure any outstanding
7144 	 * load/unloads complete.  Otherwise we could return zero
7145 	 * even though the unload(s) hasn't finished yet.
7146 	 */
7147 	pml = sfmmu_mlist_enter(spp);
7148 	cnt = spp->p_share;
7149 
7150 	if (kpm_enable)
7151 		cnt += spp->p_kpmref;
7152 
7153 	/*
7154 	 * If we have any large mappings, we count the number of
7155 	 * mappings that this large page is part of.
7156 	 */
7157 	index = PP_MAPINDEX(spp);
7158 	index >>= 1;
7159 	while (index) {
7160 		pp = PP_GROUPLEADER(spp, sz);
7161 		if ((index & 0x1) && pp != spp) {
7162 			cnt += pp->p_share;
7163 			spp = pp;
7164 		}
7165 		index >>= 1;
7166 		sz++;
7167 	}
7168 	sfmmu_mlist_exit(pml);
7169 	return (cnt);
7170 }
7171 
7172 /*
7173  * Unload all large mappings to the pp and reset the p_szc field of every
7174  * constituent page according to the remaining mappings.
7175  *
7176  * pp must be locked SE_EXCL. Even though no other constituent pages are
7177  * locked it's legal to unload the large mappings to the pp because all
7178  * constituent pages of large locked mappings have to be locked SE_SHARED.
7179  * This means if we have SE_EXCL lock on one of constituent pages none of the
7180  * large mappings to pp are locked.
7181  *
7182  * Decrease p_szc field starting from the last constituent page and ending
7183  * with the root page. This method is used because other threads rely on the
7184  * root's p_szc to find the lock to syncronize on. After a root page_t's p_szc
7185  * is demoted then other threads will succeed in sfmmu_mlspl_enter(). This
7186  * ensures that p_szc changes of the constituent pages appears atomic for all
7187  * threads that use sfmmu_mlspl_enter() to examine p_szc field.
7188  *
7189  * This mechanism is only used for file system pages where it's not always
7190  * possible to get SE_EXCL locks on all constituent pages to demote the size
7191  * code (as is done for anonymous or kernel large pages).
7192  *
7193  * See more comments in front of sfmmu_mlspl_enter().
7194  */
7195 void
7196 hat_page_demote(page_t *pp)
7197 {
7198 	int index;
7199 	int sz;
7200 	cpuset_t cpuset;
7201 	int sync = 0;
7202 	page_t *rootpp;
7203 	struct sf_hment *sfhme;
7204 	struct sf_hment *tmphme = NULL;
7205 	struct hme_blk *hmeblkp;
7206 	uint_t pszc;
7207 	page_t *lastpp;
7208 	cpuset_t tset;
7209 	pgcnt_t npgs;
7210 	kmutex_t *pml;
7211 	kmutex_t *pmtx;
7212 
7213 	ASSERT(PAGE_EXCL(pp));
7214 	ASSERT(!PP_ISFREE(pp));
7215 	ASSERT(page_szc_lock_assert(pp));
7216 	pml = sfmmu_mlist_enter(pp);
7217 	pmtx = sfmmu_page_enter(pp);
7218 
7219 	pszc = pp->p_szc;
7220 	if (pszc == 0) {
7221 		goto out;
7222 	}
7223 
7224 	index = PP_MAPINDEX(pp) >> 1;
7225 
7226 	if (index) {
7227 		CPUSET_ZERO(cpuset);
7228 		sz = TTE64K;
7229 		sync = 1;
7230 	}
7231 
7232 	while (index) {
7233 		if (!(index & 0x1)) {
7234 			index >>= 1;
7235 			sz++;
7236 			continue;
7237 		}
7238 		ASSERT(sz <= pszc);
7239 		rootpp = PP_GROUPLEADER(pp, sz);
7240 		for (sfhme = rootpp->p_mapping; sfhme; sfhme = tmphme) {
7241 			tmphme = sfhme->hme_next;
7242 			hmeblkp = sfmmu_hmetohblk(sfhme);
7243 			if (hme_size(sfhme) != sz) {
7244 				continue;
7245 			}
7246 			if (hmeblkp->hblk_xhat_bit) {
7247 				cmn_err(CE_PANIC,
7248 				    "hat_page_demote: xhat hmeblk");
7249 			}
7250 			tset = sfmmu_pageunload(rootpp, sfhme, sz);
7251 			CPUSET_OR(cpuset, tset);
7252 		}
7253 		if (index >>= 1) {
7254 			sz++;
7255 		}
7256 	}
7257 
7258 	ASSERT(!PP_ISMAPPED_LARGE(pp));
7259 
7260 	if (sync) {
7261 		xt_sync(cpuset);
7262 		if (PP_ISTNC(pp)) {
7263 			conv_tnc(rootpp, sz);
7264 		}
7265 	}
7266 
7267 	ASSERT(pp->p_szc == pszc);
7268 	rootpp = PP_PAGEROOT(pp);
7269 	ASSERT(rootpp->p_szc == pszc);
7270 	lastpp = PP_PAGENEXT_N(rootpp, TTEPAGES(pszc) - 1);
7271 
7272 	while (lastpp != rootpp) {
7273 		sz = PP_MAPINDEX(lastpp) ? fnd_mapping_sz(lastpp) : 0;
7274 		ASSERT(sz < pszc);
7275 		npgs = (sz == 0) ? 1 : TTEPAGES(sz);
7276 		ASSERT(P2PHASE(lastpp->p_pagenum, npgs) == npgs - 1);
7277 		while (--npgs > 0) {
7278 			lastpp->p_szc = (uchar_t)sz;
7279 			lastpp = PP_PAGEPREV(lastpp);
7280 		}
7281 		if (sz) {
7282 			/*
7283 			 * make sure before current root's pszc
7284 			 * is updated all updates to constituent pages pszc
7285 			 * fields are globally visible.
7286 			 */
7287 			membar_producer();
7288 		}
7289 		lastpp->p_szc = sz;
7290 		ASSERT(IS_P2ALIGNED(lastpp->p_pagenum, TTEPAGES(sz)));
7291 		if (lastpp != rootpp) {
7292 			lastpp = PP_PAGEPREV(lastpp);
7293 		}
7294 	}
7295 	if (sz == 0) {
7296 		/* the loop above doesn't cover this case */
7297 		rootpp->p_szc = 0;
7298 	}
7299 out:
7300 	ASSERT(pp->p_szc == 0);
7301 	sfmmu_page_exit(pmtx);
7302 	sfmmu_mlist_exit(pml);
7303 }
7304 
7305 /*
7306  * Refresh the HAT ismttecnt[] element for size szc.
7307  * Caller must have set ISM busy flag to prevent mapping
7308  * lists from changing while we're traversing them.
7309  */
7310 pgcnt_t
7311 ism_tsb_entries(sfmmu_t *sfmmup, int szc)
7312 {
7313 	ism_blk_t	*ism_blkp = sfmmup->sfmmu_iblk;
7314 	ism_map_t	*ism_map;
7315 	pgcnt_t		npgs = 0;
7316 	int		j;
7317 
7318 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
7319 	for (; ism_blkp != NULL; ism_blkp = ism_blkp->iblk_next) {
7320 		ism_map = ism_blkp->iblk_maps;
7321 		for (j = 0; ism_map[j].imap_ismhat && j < ISM_MAP_SLOTS; j++)
7322 			npgs += ism_map[j].imap_ismhat->sfmmu_ttecnt[szc];
7323 	}
7324 	sfmmup->sfmmu_ismttecnt[szc] = npgs;
7325 	return (npgs);
7326 }
7327 
7328 /*
7329  * Yield the memory claim requirement for an address space.
7330  *
7331  * This is currently implemented as the number of bytes that have active
7332  * hardware translations that have page structures.  Therefore, it can
7333  * underestimate the traditional resident set size, eg, if the
7334  * physical page is present and the hardware translation is missing;
7335  * and it can overestimate the rss, eg, if there are active
7336  * translations to a frame buffer with page structs.
7337  * Also, it does not take sharing into account.
7338  *
7339  * Note that we don't acquire locks here since this function is most often
7340  * called from the clock thread.
7341  */
7342 size_t
7343 hat_get_mapped_size(struct hat *hat)
7344 {
7345 	size_t		assize = 0;
7346 	int 		i;
7347 
7348 	if (hat == NULL)
7349 		return (0);
7350 
7351 	ASSERT(hat->sfmmu_xhat_provider == NULL);
7352 
7353 	for (i = 0; i < mmu_page_sizes; i++)
7354 		assize += (pgcnt_t)hat->sfmmu_ttecnt[i] * TTEBYTES(i);
7355 
7356 	if (hat->sfmmu_iblk == NULL)
7357 		return (assize);
7358 
7359 	for (i = 0; i < mmu_page_sizes; i++)
7360 		assize += (pgcnt_t)hat->sfmmu_ismttecnt[i] * TTEBYTES(i);
7361 
7362 	return (assize);
7363 }
7364 
7365 int
7366 hat_stats_enable(struct hat *hat)
7367 {
7368 	hatlock_t	*hatlockp;
7369 
7370 	ASSERT(hat->sfmmu_xhat_provider == NULL);
7371 
7372 	hatlockp = sfmmu_hat_enter(hat);
7373 	hat->sfmmu_rmstat++;
7374 	sfmmu_hat_exit(hatlockp);
7375 	return (1);
7376 }
7377 
7378 void
7379 hat_stats_disable(struct hat *hat)
7380 {
7381 	hatlock_t	*hatlockp;
7382 
7383 	ASSERT(hat->sfmmu_xhat_provider == NULL);
7384 
7385 	hatlockp = sfmmu_hat_enter(hat);
7386 	hat->sfmmu_rmstat--;
7387 	sfmmu_hat_exit(hatlockp);
7388 }
7389 
7390 /*
7391  * Routines for entering or removing  ourselves from the
7392  * ism_hat's mapping list.
7393  */
7394 static void
7395 iment_add(struct ism_ment *iment,  struct hat *ism_hat)
7396 {
7397 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
7398 
7399 	iment->iment_prev = NULL;
7400 	iment->iment_next = ism_hat->sfmmu_iment;
7401 	if (ism_hat->sfmmu_iment) {
7402 		ism_hat->sfmmu_iment->iment_prev = iment;
7403 	}
7404 	ism_hat->sfmmu_iment = iment;
7405 }
7406 
7407 static void
7408 iment_sub(struct ism_ment *iment, struct hat *ism_hat)
7409 {
7410 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
7411 
7412 	if (ism_hat->sfmmu_iment == NULL) {
7413 		panic("ism map entry remove - no entries");
7414 	}
7415 
7416 	if (iment->iment_prev) {
7417 		ASSERT(ism_hat->sfmmu_iment != iment);
7418 		iment->iment_prev->iment_next = iment->iment_next;
7419 	} else {
7420 		ASSERT(ism_hat->sfmmu_iment == iment);
7421 		ism_hat->sfmmu_iment = iment->iment_next;
7422 	}
7423 
7424 	if (iment->iment_next) {
7425 		iment->iment_next->iment_prev = iment->iment_prev;
7426 	}
7427 
7428 	/*
7429 	 * zero out the entry
7430 	 */
7431 	iment->iment_next = NULL;
7432 	iment->iment_prev = NULL;
7433 	iment->iment_hat =  NULL;
7434 }
7435 
7436 /*
7437  * Hat_share()/unshare() return an (non-zero) error
7438  * when saddr and daddr are not properly aligned.
7439  *
7440  * The top level mapping element determines the alignment
7441  * requirement for saddr and daddr, depending on different
7442  * architectures.
7443  *
7444  * When hat_share()/unshare() are not supported,
7445  * HATOP_SHARE()/UNSHARE() return 0
7446  */
7447 int
7448 hat_share(struct hat *sfmmup, caddr_t addr,
7449 	struct hat *ism_hatid, caddr_t sptaddr, size_t len, uint_t ismszc)
7450 {
7451 	ism_blk_t	*ism_blkp;
7452 	ism_blk_t	*new_iblk;
7453 	ism_map_t 	*ism_map;
7454 	ism_ment_t	*ism_ment;
7455 	int		i, added;
7456 	hatlock_t	*hatlockp;
7457 	int		reload_mmu = 0;
7458 	uint_t		ismshift = page_get_shift(ismszc);
7459 	size_t		ismpgsz = page_get_pagesize(ismszc);
7460 	uint_t		ismmask = (uint_t)ismpgsz - 1;
7461 	size_t		sh_size = ISM_SHIFT(ismshift, len);
7462 	ushort_t	ismhatflag;
7463 
7464 #ifdef DEBUG
7465 	caddr_t		eaddr = addr + len;
7466 #endif /* DEBUG */
7467 
7468 	ASSERT(ism_hatid != NULL && sfmmup != NULL);
7469 	ASSERT(sptaddr == ISMID_STARTADDR);
7470 	/*
7471 	 * Check the alignment.
7472 	 */
7473 	if (!ISM_ALIGNED(ismshift, addr) || !ISM_ALIGNED(ismshift, sptaddr))
7474 		return (EINVAL);
7475 
7476 	/*
7477 	 * Check size alignment.
7478 	 */
7479 	if (!ISM_ALIGNED(ismshift, len))
7480 		return (EINVAL);
7481 
7482 	ASSERT(sfmmup->sfmmu_xhat_provider == NULL);
7483 
7484 	/*
7485 	 * Allocate ism_ment for the ism_hat's mapping list, and an
7486 	 * ism map blk in case we need one.  We must do our
7487 	 * allocations before acquiring locks to prevent a deadlock
7488 	 * in the kmem allocator on the mapping list lock.
7489 	 */
7490 	new_iblk = kmem_cache_alloc(ism_blk_cache, KM_SLEEP);
7491 	ism_ment = kmem_cache_alloc(ism_ment_cache, KM_SLEEP);
7492 
7493 	/*
7494 	 * Serialize ISM mappings with the ISM busy flag, and also the
7495 	 * trap handlers.
7496 	 */
7497 	sfmmu_ismhat_enter(sfmmup, 0);
7498 
7499 	/*
7500 	 * Allocate an ism map blk if necessary.
7501 	 */
7502 	if (sfmmup->sfmmu_iblk == NULL) {
7503 		sfmmup->sfmmu_iblk = new_iblk;
7504 		bzero(new_iblk, sizeof (*new_iblk));
7505 		new_iblk->iblk_nextpa = (uint64_t)-1;
7506 		membar_stst();	/* make sure next ptr visible to all CPUs */
7507 		sfmmup->sfmmu_ismblkpa = va_to_pa((caddr_t)new_iblk);
7508 		reload_mmu = 1;
7509 		new_iblk = NULL;
7510 	}
7511 
7512 #ifdef DEBUG
7513 	/*
7514 	 * Make sure mapping does not already exist.
7515 	 */
7516 	ism_blkp = sfmmup->sfmmu_iblk;
7517 	while (ism_blkp) {
7518 		ism_map = ism_blkp->iblk_maps;
7519 		for (i = 0; i < ISM_MAP_SLOTS && ism_map[i].imap_ismhat; i++) {
7520 			if ((addr >= ism_start(ism_map[i]) &&
7521 			    addr < ism_end(ism_map[i])) ||
7522 			    eaddr > ism_start(ism_map[i]) &&
7523 			    eaddr <= ism_end(ism_map[i])) {
7524 				panic("sfmmu_share: Already mapped!");
7525 			}
7526 		}
7527 		ism_blkp = ism_blkp->iblk_next;
7528 	}
7529 #endif /* DEBUG */
7530 
7531 	ASSERT(ismszc >= TTE4M);
7532 	if (ismszc == TTE4M) {
7533 		ismhatflag = HAT_4M_FLAG;
7534 	} else if (ismszc == TTE32M) {
7535 		ismhatflag = HAT_32M_FLAG;
7536 	} else if (ismszc == TTE256M) {
7537 		ismhatflag = HAT_256M_FLAG;
7538 	}
7539 	/*
7540 	 * Add mapping to first available mapping slot.
7541 	 */
7542 	ism_blkp = sfmmup->sfmmu_iblk;
7543 	added = 0;
7544 	while (!added) {
7545 		ism_map = ism_blkp->iblk_maps;
7546 		for (i = 0; i < ISM_MAP_SLOTS; i++)  {
7547 			if (ism_map[i].imap_ismhat == NULL) {
7548 
7549 				ism_map[i].imap_ismhat = ism_hatid;
7550 				ism_map[i].imap_vb_shift = (ushort_t)ismshift;
7551 				ism_map[i].imap_hatflags = ismhatflag;
7552 				ism_map[i].imap_sz_mask = ismmask;
7553 				/*
7554 				 * imap_seg is checked in ISM_CHECK to see if
7555 				 * non-NULL, then other info assumed valid.
7556 				 */
7557 				membar_stst();
7558 				ism_map[i].imap_seg = (uintptr_t)addr | sh_size;
7559 				ism_map[i].imap_ment = ism_ment;
7560 
7561 				/*
7562 				 * Now add ourselves to the ism_hat's
7563 				 * mapping list.
7564 				 */
7565 				ism_ment->iment_hat = sfmmup;
7566 				ism_ment->iment_base_va = addr;
7567 				ism_hatid->sfmmu_ismhat = 1;
7568 				ism_hatid->sfmmu_flags = 0;
7569 				mutex_enter(&ism_mlist_lock);
7570 				iment_add(ism_ment, ism_hatid);
7571 				mutex_exit(&ism_mlist_lock);
7572 				added = 1;
7573 				break;
7574 			}
7575 		}
7576 		if (!added && ism_blkp->iblk_next == NULL) {
7577 			ism_blkp->iblk_next = new_iblk;
7578 			new_iblk = NULL;
7579 			bzero(ism_blkp->iblk_next,
7580 			    sizeof (*ism_blkp->iblk_next));
7581 			ism_blkp->iblk_next->iblk_nextpa = (uint64_t)-1;
7582 			membar_stst();
7583 			ism_blkp->iblk_nextpa =
7584 				va_to_pa((caddr_t)ism_blkp->iblk_next);
7585 		}
7586 		ism_blkp = ism_blkp->iblk_next;
7587 	}
7588 
7589 	/*
7590 	 * Update our counters for this sfmmup's ism mappings.
7591 	 */
7592 	for (i = 0; i <= ismszc; i++) {
7593 		if (!(disable_ism_large_pages & (1 << i)))
7594 			(void) ism_tsb_entries(sfmmup, i);
7595 	}
7596 
7597 	hatlockp = sfmmu_hat_enter(sfmmup);
7598 
7599 	/*
7600 	 * For ISM and DISM we do not support 512K pages, so we only
7601 	 * only search the 4M and 8K/64K hashes for 4 pagesize cpus, and search
7602 	 * the 256M or 32M, and 4M and 8K/64K hashes for 6 pagesize cpus.
7603 	 */
7604 	ASSERT((disable_ism_large_pages & (1 << TTE512K)) != 0);
7605 
7606 	if (ismszc > TTE4M && !SFMMU_FLAGS_ISSET(sfmmup, HAT_4M_FLAG))
7607 		SFMMU_FLAGS_SET(sfmmup, HAT_4M_FLAG);
7608 
7609 	if (!SFMMU_FLAGS_ISSET(sfmmup, HAT_64K_FLAG))
7610 		SFMMU_FLAGS_SET(sfmmup, HAT_64K_FLAG);
7611 
7612 	/*
7613 	 * If we updated the ismblkpa for this HAT or we need
7614 	 * to start searching the 256M or 32M or 4M hash, we must
7615 	 * make sure all CPUs running this process reload their
7616 	 * tsbmiss area.  Otherwise they will fail to load the mappings
7617 	 * in the tsbmiss handler and will loop calling pagefault().
7618 	 */
7619 	switch (ismszc) {
7620 	case TTE256M:
7621 		if (reload_mmu || !SFMMU_FLAGS_ISSET(sfmmup, HAT_256M_FLAG)) {
7622 			SFMMU_FLAGS_SET(sfmmup, HAT_256M_FLAG);
7623 			sfmmu_sync_mmustate(sfmmup);
7624 		}
7625 		break;
7626 	case TTE32M:
7627 		if (reload_mmu || !SFMMU_FLAGS_ISSET(sfmmup, HAT_32M_FLAG)) {
7628 			SFMMU_FLAGS_SET(sfmmup, HAT_32M_FLAG);
7629 			sfmmu_sync_mmustate(sfmmup);
7630 		}
7631 		break;
7632 	case TTE4M:
7633 		if (reload_mmu || !SFMMU_FLAGS_ISSET(sfmmup, HAT_4M_FLAG)) {
7634 			SFMMU_FLAGS_SET(sfmmup, HAT_4M_FLAG);
7635 			sfmmu_sync_mmustate(sfmmup);
7636 		}
7637 		break;
7638 	default:
7639 		break;
7640 	}
7641 
7642 	/*
7643 	 * Now we can drop the locks.
7644 	 */
7645 	sfmmu_ismhat_exit(sfmmup, 1);
7646 	sfmmu_hat_exit(hatlockp);
7647 
7648 	/*
7649 	 * Free up ismblk if we didn't use it.
7650 	 */
7651 	if (new_iblk != NULL)
7652 		kmem_cache_free(ism_blk_cache, new_iblk);
7653 
7654 	/*
7655 	 * Check TSB and TLB page sizes.
7656 	 */
7657 	sfmmu_check_page_sizes(sfmmup, 1);
7658 
7659 	return (0);
7660 }
7661 
7662 /*
7663  * hat_unshare removes exactly one ism_map from
7664  * this process's as.  It expects multiple calls
7665  * to hat_unshare for multiple shm segments.
7666  */
7667 void
7668 hat_unshare(struct hat *sfmmup, caddr_t addr, size_t len, uint_t ismszc)
7669 {
7670 	ism_map_t 	*ism_map;
7671 	ism_ment_t	*free_ment = NULL;
7672 	ism_blk_t	*ism_blkp;
7673 	struct hat	*ism_hatid;
7674 	struct ctx	*ctx;
7675 	int 		cnum, found, i;
7676 	hatlock_t	*hatlockp;
7677 	struct tsb_info	*tsbinfo;
7678 	uint_t		ismshift = page_get_shift(ismszc);
7679 	size_t		sh_size = ISM_SHIFT(ismshift, len);
7680 
7681 	ASSERT(ISM_ALIGNED(ismshift, addr));
7682 	ASSERT(ISM_ALIGNED(ismshift, len));
7683 	ASSERT(sfmmup != NULL);
7684 	ASSERT(sfmmup != ksfmmup);
7685 
7686 	if (sfmmup->sfmmu_xhat_provider) {
7687 		XHAT_UNSHARE(sfmmup, addr, len);
7688 		return;
7689 	} else {
7690 		/*
7691 		 * This must be a CPU HAT. If the address space has
7692 		 * XHATs attached, inform all XHATs that ISM segment
7693 		 * is going away
7694 		 */
7695 		ASSERT(sfmmup->sfmmu_as != NULL);
7696 		if (sfmmup->sfmmu_as->a_xhat != NULL)
7697 			xhat_unshare_all(sfmmup->sfmmu_as, addr, len);
7698 	}
7699 
7700 	/*
7701 	 * Make sure that during the entire time ISM mappings are removed,
7702 	 * the trap handlers serialize behind us, and that no one else
7703 	 * can be mucking with ISM mappings.  This also lets us get away
7704 	 * with not doing expensive cross calls to flush the TLB -- we
7705 	 * just discard the context, flush the entire TSB, and call it
7706 	 * a day.
7707 	 */
7708 	sfmmu_ismhat_enter(sfmmup, 0);
7709 
7710 	/*
7711 	 * Remove the mapping.
7712 	 *
7713 	 * We can't have any holes in the ism map.
7714 	 * The tsb miss code while searching the ism map will
7715 	 * stop on an empty map slot.  So we must move
7716 	 * everyone past the hole up 1 if any.
7717 	 *
7718 	 * Also empty ism map blks are not freed until the
7719 	 * process exits. This is to prevent a MT race condition
7720 	 * between sfmmu_unshare() and sfmmu_tsbmiss_exception().
7721 	 */
7722 	found = 0;
7723 	ism_blkp = sfmmup->sfmmu_iblk;
7724 	while (!found && ism_blkp) {
7725 		ism_map = ism_blkp->iblk_maps;
7726 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
7727 			if (addr == ism_start(ism_map[i]) &&
7728 			    sh_size == (size_t)(ism_size(ism_map[i]))) {
7729 				found = 1;
7730 				break;
7731 			}
7732 		}
7733 		if (!found)
7734 			ism_blkp = ism_blkp->iblk_next;
7735 	}
7736 
7737 	if (found) {
7738 		ism_hatid = ism_map[i].imap_ismhat;
7739 		ASSERT(ism_hatid != NULL);
7740 		ASSERT(ism_hatid->sfmmu_ismhat == 1);
7741 		ASSERT(ism_hatid->sfmmu_cnum == INVALID_CONTEXT);
7742 
7743 		/*
7744 		 * First remove ourselves from the ism mapping list.
7745 		 */
7746 		mutex_enter(&ism_mlist_lock);
7747 		iment_sub(ism_map[i].imap_ment, ism_hatid);
7748 		mutex_exit(&ism_mlist_lock);
7749 		free_ment = ism_map[i].imap_ment;
7750 
7751 		/*
7752 		 * Now gurantee that any other cpu
7753 		 * that tries to process an ISM miss
7754 		 * will go to tl=0.
7755 		 */
7756 		hatlockp = sfmmu_hat_enter(sfmmup);
7757 		ctx = sfmmutoctx(sfmmup);
7758 		rw_enter(&ctx->ctx_rwlock, RW_WRITER);
7759 		cnum = sfmmutoctxnum(sfmmup);
7760 
7761 		if (cnum != INVALID_CONTEXT) {
7762 			sfmmu_tlb_swap_ctx(sfmmup, ctx);
7763 		}
7764 		rw_exit(&ctx->ctx_rwlock);
7765 		sfmmu_hat_exit(hatlockp);
7766 
7767 		/*
7768 		 * We delete the ism map by copying
7769 		 * the next map over the current one.
7770 		 * We will take the next one in the maps
7771 		 * array or from the next ism_blk.
7772 		 */
7773 		while (ism_blkp) {
7774 			ism_map = ism_blkp->iblk_maps;
7775 			while (i < (ISM_MAP_SLOTS - 1)) {
7776 				ism_map[i] = ism_map[i + 1];
7777 				i++;
7778 			}
7779 			/* i == (ISM_MAP_SLOTS - 1) */
7780 			ism_blkp = ism_blkp->iblk_next;
7781 			if (ism_blkp) {
7782 				ism_map[i] = ism_blkp->iblk_maps[0];
7783 				i = 0;
7784 			} else {
7785 				ism_map[i].imap_seg = 0;
7786 				ism_map[i].imap_vb_shift = 0;
7787 				ism_map[i].imap_hatflags = 0;
7788 				ism_map[i].imap_sz_mask = 0;
7789 				ism_map[i].imap_ismhat = NULL;
7790 				ism_map[i].imap_ment = NULL;
7791 			}
7792 		}
7793 
7794 		/*
7795 		 * Now flush entire TSB for the process, since
7796 		 * demapping page by page can be too expensive.
7797 		 * We don't have to flush the TLB here anymore
7798 		 * since we switch to a new TLB ctx instead.
7799 		 * Also, there is no need to flush if the process
7800 		 * is exiting since the TSB will be freed later.
7801 		 */
7802 		if (!sfmmup->sfmmu_free) {
7803 			hatlockp = sfmmu_hat_enter(sfmmup);
7804 			for (tsbinfo = sfmmup->sfmmu_tsb; tsbinfo != NULL;
7805 			    tsbinfo = tsbinfo->tsb_next) {
7806 				if (tsbinfo->tsb_flags & TSB_SWAPPED)
7807 					continue;
7808 				sfmmu_inv_tsb(tsbinfo->tsb_va,
7809 				    TSB_BYTES(tsbinfo->tsb_szc));
7810 			}
7811 			sfmmu_hat_exit(hatlockp);
7812 		}
7813 	}
7814 
7815 	/*
7816 	 * Update our counters for this sfmmup's ism mappings.
7817 	 */
7818 	for (i = 0; i <= ismszc; i++) {
7819 		if (!(disable_ism_large_pages & (1 << i)))
7820 			(void) ism_tsb_entries(sfmmup, i);
7821 	}
7822 
7823 	sfmmu_ismhat_exit(sfmmup, 0);
7824 
7825 	/*
7826 	 * We must do our freeing here after dropping locks
7827 	 * to prevent a deadlock in the kmem allocator on the
7828 	 * mapping list lock.
7829 	 */
7830 	if (free_ment != NULL)
7831 		kmem_cache_free(ism_ment_cache, free_ment);
7832 
7833 	/*
7834 	 * Check TSB and TLB page sizes if the process isn't exiting.
7835 	 */
7836 	if (!sfmmup->sfmmu_free)
7837 		sfmmu_check_page_sizes(sfmmup, 0);
7838 }
7839 
7840 /* ARGSUSED */
7841 static int
7842 sfmmu_idcache_constructor(void *buf, void *cdrarg, int kmflags)
7843 {
7844 	/* void *buf is sfmmu_t pointer */
7845 	return (0);
7846 }
7847 
7848 /* ARGSUSED */
7849 static void
7850 sfmmu_idcache_destructor(void *buf, void *cdrarg)
7851 {
7852 	/* void *buf is sfmmu_t pointer */
7853 }
7854 
7855 /*
7856  * setup kmem hmeblks by bzeroing all members and initializing the nextpa
7857  * field to be the pa of this hmeblk
7858  */
7859 /* ARGSUSED */
7860 static int
7861 sfmmu_hblkcache_constructor(void *buf, void *cdrarg, int kmflags)
7862 {
7863 	struct hme_blk *hmeblkp;
7864 
7865 	bzero(buf, (size_t)cdrarg);
7866 	hmeblkp = (struct hme_blk *)buf;
7867 	hmeblkp->hblk_nextpa = va_to_pa((caddr_t)hmeblkp);
7868 
7869 #ifdef	HBLK_TRACE
7870 	mutex_init(&hmeblkp->hblk_audit_lock, NULL, MUTEX_DEFAULT, NULL);
7871 #endif	/* HBLK_TRACE */
7872 
7873 	return (0);
7874 }
7875 
7876 /* ARGSUSED */
7877 static void
7878 sfmmu_hblkcache_destructor(void *buf, void *cdrarg)
7879 {
7880 
7881 #ifdef	HBLK_TRACE
7882 
7883 	struct hme_blk *hmeblkp;
7884 
7885 	hmeblkp = (struct hme_blk *)buf;
7886 	mutex_destroy(&hmeblkp->hblk_audit_lock);
7887 
7888 #endif	/* HBLK_TRACE */
7889 }
7890 
7891 #define	SFMMU_CACHE_RECLAIM_SCAN_RATIO 8
7892 static int sfmmu_cache_reclaim_scan_ratio = SFMMU_CACHE_RECLAIM_SCAN_RATIO;
7893 /*
7894  * The kmem allocator will callback into our reclaim routine when the system
7895  * is running low in memory.  We traverse the hash and free up all unused but
7896  * still cached hme_blks.  We also traverse the free list and free them up
7897  * as well.
7898  */
7899 /*ARGSUSED*/
7900 static void
7901 sfmmu_hblkcache_reclaim(void *cdrarg)
7902 {
7903 	int i;
7904 	uint64_t hblkpa, prevpa, nx_pa;
7905 	struct hmehash_bucket *hmebp;
7906 	struct hme_blk *hmeblkp, *nx_hblk, *pr_hblk = NULL;
7907 	static struct hmehash_bucket *uhmehash_reclaim_hand;
7908 	static struct hmehash_bucket *khmehash_reclaim_hand;
7909 	struct hme_blk *list = NULL;
7910 
7911 	hmebp = uhmehash_reclaim_hand;
7912 	if (hmebp == NULL || hmebp > &uhme_hash[UHMEHASH_SZ])
7913 		uhmehash_reclaim_hand = hmebp = uhme_hash;
7914 	uhmehash_reclaim_hand += UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
7915 
7916 	for (i = UHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
7917 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
7918 			hmeblkp = hmebp->hmeblkp;
7919 			hblkpa = hmebp->hmeh_nextpa;
7920 			prevpa = 0;
7921 			pr_hblk = NULL;
7922 			while (hmeblkp) {
7923 				nx_hblk = hmeblkp->hblk_next;
7924 				nx_pa = hmeblkp->hblk_nextpa;
7925 				if (!hmeblkp->hblk_vcnt &&
7926 				    !hmeblkp->hblk_hmecnt) {
7927 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
7928 						prevpa, pr_hblk);
7929 					sfmmu_hblk_free(hmebp, hmeblkp,
7930 					    hblkpa, &list);
7931 				} else {
7932 					pr_hblk = hmeblkp;
7933 					prevpa = hblkpa;
7934 				}
7935 				hmeblkp = nx_hblk;
7936 				hblkpa = nx_pa;
7937 			}
7938 			SFMMU_HASH_UNLOCK(hmebp);
7939 		}
7940 		if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
7941 			hmebp = uhme_hash;
7942 	}
7943 
7944 	hmebp = khmehash_reclaim_hand;
7945 	if (hmebp == NULL || hmebp > &khme_hash[KHMEHASH_SZ])
7946 		khmehash_reclaim_hand = hmebp = khme_hash;
7947 	khmehash_reclaim_hand += KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio;
7948 
7949 	for (i = KHMEHASH_SZ / sfmmu_cache_reclaim_scan_ratio; i; i--) {
7950 		if (SFMMU_HASH_LOCK_TRYENTER(hmebp) != 0) {
7951 			hmeblkp = hmebp->hmeblkp;
7952 			hblkpa = hmebp->hmeh_nextpa;
7953 			prevpa = 0;
7954 			pr_hblk = NULL;
7955 			while (hmeblkp) {
7956 				nx_hblk = hmeblkp->hblk_next;
7957 				nx_pa = hmeblkp->hblk_nextpa;
7958 				if (!hmeblkp->hblk_vcnt &&
7959 				    !hmeblkp->hblk_hmecnt) {
7960 					sfmmu_hblk_hash_rm(hmebp, hmeblkp,
7961 						prevpa, pr_hblk);
7962 					sfmmu_hblk_free(hmebp, hmeblkp,
7963 					    hblkpa, &list);
7964 				} else {
7965 					pr_hblk = hmeblkp;
7966 					prevpa = hblkpa;
7967 				}
7968 				hmeblkp = nx_hblk;
7969 				hblkpa = nx_pa;
7970 			}
7971 			SFMMU_HASH_UNLOCK(hmebp);
7972 		}
7973 		if (hmebp++ == &khme_hash[KHMEHASH_SZ])
7974 			hmebp = khme_hash;
7975 	}
7976 	sfmmu_hblks_list_purge(&list);
7977 }
7978 
7979 /*
7980  * sfmmu_get_ppvcolor should become a vm_machdep or hatop interface.
7981  * same goes for sfmmu_get_addrvcolor().
7982  *
7983  * This function will return the virtual color for the specified page. The
7984  * virtual color corresponds to this page current mapping or its last mapping.
7985  * It is used by memory allocators to choose addresses with the correct
7986  * alignment so vac consistency is automatically maintained.  If the page
7987  * has no color it returns -1.
7988  */
7989 int
7990 sfmmu_get_ppvcolor(struct page *pp)
7991 {
7992 	int color;
7993 
7994 	if (!(cache & CACHE_VAC) || PP_NEWPAGE(pp)) {
7995 		return (-1);
7996 	}
7997 	color = PP_GET_VCOLOR(pp);
7998 	ASSERT(color < mmu_btop(shm_alignment));
7999 	return (color);
8000 }
8001 
8002 /*
8003  * This function will return the desired alignment for vac consistency
8004  * (vac color) given a virtual address.  If no vac is present it returns -1.
8005  */
8006 int
8007 sfmmu_get_addrvcolor(caddr_t vaddr)
8008 {
8009 	if (cache & CACHE_VAC) {
8010 		return (addr_to_vcolor(vaddr));
8011 	} else {
8012 		return (-1);
8013 	}
8014 
8015 }
8016 
8017 /*
8018  * Check for conflicts.
8019  * A conflict exists if the new and existent mappings do not match in
8020  * their "shm_alignment fields. If conflicts exist, the existant mappings
8021  * are flushed unless one of them is locked. If one of them is locked, then
8022  * the mappings are flushed and converted to non-cacheable mappings.
8023  */
8024 static void
8025 sfmmu_vac_conflict(struct hat *hat, caddr_t addr, page_t *pp)
8026 {
8027 	struct hat *tmphat;
8028 	struct sf_hment *sfhmep, *tmphme = NULL;
8029 	struct hme_blk *hmeblkp;
8030 	int vcolor;
8031 	tte_t tte;
8032 
8033 	ASSERT(sfmmu_mlist_held(pp));
8034 	ASSERT(!PP_ISNC(pp));		/* page better be cacheable */
8035 
8036 	vcolor = addr_to_vcolor(addr);
8037 	if (PP_NEWPAGE(pp)) {
8038 		PP_SET_VCOLOR(pp, vcolor);
8039 		return;
8040 	}
8041 
8042 	if (PP_GET_VCOLOR(pp) == vcolor) {
8043 		return;
8044 	}
8045 
8046 	if (!PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp)) {
8047 		/*
8048 		 * Previous user of page had a different color
8049 		 * but since there are no current users
8050 		 * we just flush the cache and change the color.
8051 		 */
8052 		SFMMU_STAT(sf_pgcolor_conflict);
8053 		sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
8054 		PP_SET_VCOLOR(pp, vcolor);
8055 		return;
8056 	}
8057 
8058 	/*
8059 	 * If we get here we have a vac conflict with a current
8060 	 * mapping.  VAC conflict policy is as follows.
8061 	 * - The default is to unload the other mappings unless:
8062 	 * - If we have a large mapping we uncache the page.
8063 	 * We need to uncache the rest of the large page too.
8064 	 * - If any of the mappings are locked we uncache the page.
8065 	 * - If the requested mapping is inconsistent
8066 	 * with another mapping and that mapping
8067 	 * is in the same address space we have to
8068 	 * make it non-cached.  The default thing
8069 	 * to do is unload the inconsistent mapping
8070 	 * but if they are in the same address space
8071 	 * we run the risk of unmapping the pc or the
8072 	 * stack which we will use as we return to the user,
8073 	 * in which case we can then fault on the thing
8074 	 * we just unloaded and get into an infinite loop.
8075 	 */
8076 	if (PP_ISMAPPED_LARGE(pp)) {
8077 		int sz;
8078 
8079 		/*
8080 		 * Existing mapping is for big pages. We don't unload
8081 		 * existing big mappings to satisfy new mappings.
8082 		 * Always convert all mappings to TNC.
8083 		 */
8084 		sz = fnd_mapping_sz(pp);
8085 		pp = PP_GROUPLEADER(pp, sz);
8086 		SFMMU_STAT_ADD(sf_uncache_conflict, TTEPAGES(sz));
8087 		sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH,
8088 			TTEPAGES(sz));
8089 
8090 		return;
8091 	}
8092 
8093 	/*
8094 	 * check if any mapping is in same as or if it is locked
8095 	 * since in that case we need to uncache.
8096 	 */
8097 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
8098 		tmphme = sfhmep->hme_next;
8099 		hmeblkp = sfmmu_hmetohblk(sfhmep);
8100 		if (hmeblkp->hblk_xhat_bit)
8101 			continue;
8102 		tmphat = hblktosfmmu(hmeblkp);
8103 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
8104 		ASSERT(TTE_IS_VALID(&tte));
8105 		if ((tmphat == hat) || hmeblkp->hblk_lckcnt) {
8106 			/*
8107 			 * We have an uncache conflict
8108 			 */
8109 			SFMMU_STAT(sf_uncache_conflict);
8110 			sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH, 1);
8111 			return;
8112 		}
8113 	}
8114 
8115 	/*
8116 	 * We have an unload conflict
8117 	 * We have already checked for LARGE mappings, therefore
8118 	 * the remaining mapping(s) must be TTE8K.
8119 	 */
8120 	SFMMU_STAT(sf_unload_conflict);
8121 
8122 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
8123 		tmphme = sfhmep->hme_next;
8124 		hmeblkp = sfmmu_hmetohblk(sfhmep);
8125 		if (hmeblkp->hblk_xhat_bit)
8126 			continue;
8127 		(void) sfmmu_pageunload(pp, sfhmep, TTE8K);
8128 	}
8129 
8130 	if (PP_ISMAPPED_KPM(pp))
8131 		sfmmu_kpm_vac_unload(pp, addr);
8132 
8133 	/*
8134 	 * Unloads only do TLB flushes so we need to flush the
8135 	 * cache here.
8136 	 */
8137 	sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
8138 	PP_SET_VCOLOR(pp, vcolor);
8139 }
8140 
8141 /*
8142  * Whenever a mapping is unloaded and the page is in TNC state,
8143  * we see if the page can be made cacheable again. 'pp' is
8144  * the page that we just unloaded a mapping from, the size
8145  * of mapping that was unloaded is 'ottesz'.
8146  * Remark:
8147  * The recache policy for mpss pages can leave a performance problem
8148  * under the following circumstances:
8149  * . A large page in uncached mode has just been unmapped.
8150  * . All constituent pages are TNC due to a conflicting small mapping.
8151  * . There are many other, non conflicting, small mappings around for
8152  *   a lot of the constituent pages.
8153  * . We're called w/ the "old" groupleader page and the old ottesz,
8154  *   but this is irrelevant, since we're no more "PP_ISMAPPED_LARGE", so
8155  *   we end up w/ TTE8K or npages == 1.
8156  * . We call tst_tnc w/ the old groupleader only, and if there is no
8157  *   conflict, we re-cache only this page.
8158  * . All other small mappings are not checked and will be left in TNC mode.
8159  * The problem is not very serious because:
8160  * . mpss is actually only defined for heap and stack, so the probability
8161  *   is not very high that a large page mapping exists in parallel to a small
8162  *   one (this is possible, but seems to be bad programming style in the
8163  *   appl).
8164  * . The problem gets a little bit more serious, when those TNC pages
8165  *   have to be mapped into kernel space, e.g. for networking.
8166  * . When VAC alias conflicts occur in applications, this is regarded
8167  *   as an application bug. So if kstat's show them, the appl should
8168  *   be changed anyway.
8169  */
8170 static void
8171 conv_tnc(page_t *pp, int ottesz)
8172 {
8173 	int cursz, dosz;
8174 	pgcnt_t curnpgs, dopgs;
8175 	pgcnt_t pg64k;
8176 	page_t *pp2;
8177 
8178 	/*
8179 	 * Determine how big a range we check for TNC and find
8180 	 * leader page. cursz is the size of the biggest
8181 	 * mapping that still exist on 'pp'.
8182 	 */
8183 	if (PP_ISMAPPED_LARGE(pp)) {
8184 		cursz = fnd_mapping_sz(pp);
8185 	} else {
8186 		cursz = TTE8K;
8187 	}
8188 
8189 	if (ottesz >= cursz) {
8190 		dosz = ottesz;
8191 		pp2 = pp;
8192 	} else {
8193 		dosz = cursz;
8194 		pp2 = PP_GROUPLEADER(pp, dosz);
8195 	}
8196 
8197 	pg64k = TTEPAGES(TTE64K);
8198 	dopgs = TTEPAGES(dosz);
8199 
8200 	ASSERT(dopgs == 1 || ((dopgs & (pg64k - 1)) == 0));
8201 
8202 	while (dopgs != 0) {
8203 		curnpgs = TTEPAGES(cursz);
8204 		if (tst_tnc(pp2, curnpgs)) {
8205 			SFMMU_STAT_ADD(sf_recache, curnpgs);
8206 			sfmmu_page_cache_array(pp2, HAT_CACHE, CACHE_NO_FLUSH,
8207 				curnpgs);
8208 		}
8209 
8210 		ASSERT(dopgs >= curnpgs);
8211 		dopgs -= curnpgs;
8212 
8213 		if (dopgs == 0) {
8214 			break;
8215 		}
8216 
8217 		pp2 = PP_PAGENEXT_N(pp2, curnpgs);
8218 		if (((dopgs & (pg64k - 1)) == 0) && PP_ISMAPPED_LARGE(pp2)) {
8219 			cursz = fnd_mapping_sz(pp2);
8220 		} else {
8221 			cursz = TTE8K;
8222 		}
8223 	}
8224 }
8225 
8226 /*
8227  * Returns 1 if page(s) can be converted from TNC to cacheable setting,
8228  * returns 0 otherwise. Note that oaddr argument is valid for only
8229  * 8k pages.
8230  */
8231 static int
8232 tst_tnc(page_t *pp, pgcnt_t npages)
8233 {
8234 	struct	sf_hment *sfhme;
8235 	struct	hme_blk *hmeblkp;
8236 	tte_t	tte;
8237 	caddr_t	vaddr;
8238 	int	clr_valid = 0;
8239 	int 	color, color1, bcolor;
8240 	int	i, ncolors;
8241 
8242 	ASSERT(pp != NULL);
8243 	ASSERT(!(cache & CACHE_WRITEBACK));
8244 
8245 	if (npages > 1) {
8246 		ncolors = CACHE_NUM_COLOR;
8247 	}
8248 
8249 	for (i = 0; i < npages; i++) {
8250 		ASSERT(sfmmu_mlist_held(pp));
8251 		ASSERT(PP_ISTNC(pp));
8252 		ASSERT(PP_GET_VCOLOR(pp) == NO_VCOLOR);
8253 
8254 		if (PP_ISPNC(pp)) {
8255 			return (0);
8256 		}
8257 
8258 		clr_valid = 0;
8259 		if (PP_ISMAPPED_KPM(pp)) {
8260 			caddr_t kpmvaddr;
8261 
8262 			ASSERT(kpm_enable);
8263 			kpmvaddr = hat_kpm_page2va(pp, 1);
8264 			ASSERT(!(npages > 1 && IS_KPM_ALIAS_RANGE(kpmvaddr)));
8265 			color1 = addr_to_vcolor(kpmvaddr);
8266 			clr_valid = 1;
8267 		}
8268 
8269 		for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
8270 			hmeblkp = sfmmu_hmetohblk(sfhme);
8271 			if (hmeblkp->hblk_xhat_bit)
8272 				continue;
8273 
8274 			sfmmu_copytte(&sfhme->hme_tte, &tte);
8275 			ASSERT(TTE_IS_VALID(&tte));
8276 
8277 			vaddr = tte_to_vaddr(hmeblkp, tte);
8278 			color = addr_to_vcolor(vaddr);
8279 
8280 			if (npages > 1) {
8281 				/*
8282 				 * If there is a big mapping, make sure
8283 				 * 8K mapping is consistent with the big
8284 				 * mapping.
8285 				 */
8286 				bcolor = i % ncolors;
8287 				if (color != bcolor) {
8288 					return (0);
8289 				}
8290 			}
8291 			if (!clr_valid) {
8292 				clr_valid = 1;
8293 				color1 = color;
8294 			}
8295 
8296 			if (color1 != color) {
8297 				return (0);
8298 			}
8299 		}
8300 
8301 		pp = PP_PAGENEXT(pp);
8302 	}
8303 
8304 	return (1);
8305 }
8306 
8307 static void
8308 sfmmu_page_cache_array(page_t *pp, int flags, int cache_flush_flag,
8309 	pgcnt_t npages)
8310 {
8311 	kmutex_t *pmtx;
8312 	int i, ncolors, bcolor;
8313 	kpm_hlk_t *kpmp;
8314 	cpuset_t cpuset;
8315 
8316 	ASSERT(pp != NULL);
8317 	ASSERT(!(cache & CACHE_WRITEBACK));
8318 
8319 	kpmp = sfmmu_kpm_kpmp_enter(pp, npages);
8320 	pmtx = sfmmu_page_enter(pp);
8321 
8322 	/*
8323 	 * Fast path caching single unmapped page
8324 	 */
8325 	if (npages == 1 && !PP_ISMAPPED(pp) && !PP_ISMAPPED_KPM(pp) &&
8326 	    flags == HAT_CACHE) {
8327 		PP_CLRTNC(pp);
8328 		PP_CLRPNC(pp);
8329 		sfmmu_page_exit(pmtx);
8330 		sfmmu_kpm_kpmp_exit(kpmp);
8331 		return;
8332 	}
8333 
8334 	/*
8335 	 * We need to capture all cpus in order to change cacheability
8336 	 * because we can't allow one cpu to access the same physical
8337 	 * page using a cacheable and a non-cachebale mapping at the same
8338 	 * time. Since we may end up walking the ism mapping list
8339 	 * have to grab it's lock now since we can't after all the
8340 	 * cpus have been captured.
8341 	 */
8342 	sfmmu_hat_lock_all();
8343 	mutex_enter(&ism_mlist_lock);
8344 	kpreempt_disable();
8345 	cpuset = cpu_ready_set;
8346 	xc_attention(cpuset);
8347 
8348 	if (npages > 1) {
8349 		/*
8350 		 * Make sure all colors are flushed since the
8351 		 * sfmmu_page_cache() only flushes one color-
8352 		 * it does not know big pages.
8353 		 */
8354 		ncolors = CACHE_NUM_COLOR;
8355 		if (flags & HAT_TMPNC) {
8356 			for (i = 0; i < ncolors; i++) {
8357 				sfmmu_cache_flushcolor(i, pp->p_pagenum);
8358 			}
8359 			cache_flush_flag = CACHE_NO_FLUSH;
8360 		}
8361 	}
8362 
8363 	for (i = 0; i < npages; i++) {
8364 
8365 		ASSERT(sfmmu_mlist_held(pp));
8366 
8367 		if (!(flags == HAT_TMPNC && PP_ISTNC(pp))) {
8368 
8369 			if (npages > 1) {
8370 				bcolor = i % ncolors;
8371 			} else {
8372 				bcolor = NO_VCOLOR;
8373 			}
8374 
8375 			sfmmu_page_cache(pp, flags, cache_flush_flag,
8376 			    bcolor);
8377 		}
8378 
8379 		pp = PP_PAGENEXT(pp);
8380 	}
8381 
8382 	xt_sync(cpuset);
8383 	xc_dismissed(cpuset);
8384 	mutex_exit(&ism_mlist_lock);
8385 	sfmmu_hat_unlock_all();
8386 	sfmmu_page_exit(pmtx);
8387 	sfmmu_kpm_kpmp_exit(kpmp);
8388 	kpreempt_enable();
8389 }
8390 
8391 /*
8392  * This function changes the virtual cacheability of all mappings to a
8393  * particular page.  When changing from uncache to cacheable the mappings will
8394  * only be changed if all of them have the same virtual color.
8395  * We need to flush the cache in all cpus.  It is possible that
8396  * a process referenced a page as cacheable but has sinced exited
8397  * and cleared the mapping list.  We still to flush it but have no
8398  * state so all cpus is the only alternative.
8399  */
8400 static void
8401 sfmmu_page_cache(page_t *pp, int flags, int cache_flush_flag, int bcolor)
8402 {
8403 	struct	sf_hment *sfhme;
8404 	struct	hme_blk *hmeblkp;
8405 	sfmmu_t *sfmmup;
8406 	tte_t	tte, ttemod;
8407 	caddr_t	vaddr;
8408 	int	ret, color;
8409 	pfn_t	pfn;
8410 
8411 	color = bcolor;
8412 	pfn = pp->p_pagenum;
8413 
8414 	for (sfhme = pp->p_mapping; sfhme; sfhme = sfhme->hme_next) {
8415 
8416 		hmeblkp = sfmmu_hmetohblk(sfhme);
8417 
8418 		if (hmeblkp->hblk_xhat_bit)
8419 			continue;
8420 
8421 		sfmmu_copytte(&sfhme->hme_tte, &tte);
8422 		ASSERT(TTE_IS_VALID(&tte));
8423 		vaddr = tte_to_vaddr(hmeblkp, tte);
8424 		color = addr_to_vcolor(vaddr);
8425 
8426 #ifdef DEBUG
8427 		if ((flags & HAT_CACHE) && bcolor != NO_VCOLOR) {
8428 			ASSERT(color == bcolor);
8429 		}
8430 #endif
8431 
8432 		ASSERT(flags != HAT_TMPNC || color == PP_GET_VCOLOR(pp));
8433 
8434 		ttemod = tte;
8435 		if (flags & (HAT_UNCACHE | HAT_TMPNC)) {
8436 			TTE_CLR_VCACHEABLE(&ttemod);
8437 		} else {	/* flags & HAT_CACHE */
8438 			TTE_SET_VCACHEABLE(&ttemod);
8439 		}
8440 		ret = sfmmu_modifytte_try(&tte, &ttemod, &sfhme->hme_tte);
8441 		if (ret < 0) {
8442 			/*
8443 			 * Since all cpus are captured modifytte should not
8444 			 * fail.
8445 			 */
8446 			panic("sfmmu_page_cache: write to tte failed");
8447 		}
8448 
8449 		sfmmup = hblktosfmmu(hmeblkp);
8450 		if (cache_flush_flag == CACHE_FLUSH) {
8451 			/*
8452 			 * Flush TSBs, TLBs and caches
8453 			 */
8454 			if (sfmmup->sfmmu_ismhat) {
8455 				if (flags & HAT_CACHE) {
8456 					SFMMU_STAT(sf_ism_recache);
8457 				} else {
8458 					SFMMU_STAT(sf_ism_uncache);
8459 				}
8460 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
8461 				    pfn, CACHE_FLUSH);
8462 			} else {
8463 				sfmmu_tlbcache_demap(vaddr, sfmmup, hmeblkp,
8464 				    pfn, 0, FLUSH_ALL_CPUS, CACHE_FLUSH, 1);
8465 			}
8466 
8467 			/*
8468 			 * all cache entries belonging to this pfn are
8469 			 * now flushed.
8470 			 */
8471 			cache_flush_flag = CACHE_NO_FLUSH;
8472 		} else {
8473 
8474 			/*
8475 			 * Flush only TSBs and TLBs.
8476 			 */
8477 			if (sfmmup->sfmmu_ismhat) {
8478 				if (flags & HAT_CACHE) {
8479 					SFMMU_STAT(sf_ism_recache);
8480 				} else {
8481 					SFMMU_STAT(sf_ism_uncache);
8482 				}
8483 				sfmmu_ismtlbcache_demap(vaddr, sfmmup, hmeblkp,
8484 				    pfn, CACHE_NO_FLUSH);
8485 			} else {
8486 				sfmmu_tlb_demap(vaddr, sfmmup, hmeblkp, 0, 1);
8487 			}
8488 		}
8489 	}
8490 
8491 	if (PP_ISMAPPED_KPM(pp))
8492 		sfmmu_kpm_page_cache(pp, flags, cache_flush_flag);
8493 
8494 	switch (flags) {
8495 
8496 		default:
8497 			panic("sfmmu_pagecache: unknown flags");
8498 			break;
8499 
8500 		case HAT_CACHE:
8501 			PP_CLRTNC(pp);
8502 			PP_CLRPNC(pp);
8503 			PP_SET_VCOLOR(pp, color);
8504 			break;
8505 
8506 		case HAT_TMPNC:
8507 			PP_SETTNC(pp);
8508 			PP_SET_VCOLOR(pp, NO_VCOLOR);
8509 			break;
8510 
8511 		case HAT_UNCACHE:
8512 			PP_SETPNC(pp);
8513 			PP_CLRTNC(pp);
8514 			PP_SET_VCOLOR(pp, NO_VCOLOR);
8515 			break;
8516 	}
8517 }
8518 
8519 /*
8520  * This routine gets called when the system has run out of free contexts.
8521  * This will simply choose context passed to it to be stolen and reused.
8522  */
8523 /* ARGSUSED */
8524 static void
8525 sfmmu_reuse_ctx(struct ctx *ctx, sfmmu_t *sfmmup)
8526 {
8527 	sfmmu_t *stolen_sfmmup;
8528 	cpuset_t cpuset;
8529 	ushort_t	cnum = ctxtoctxnum(ctx);
8530 
8531 	ASSERT(cnum != KCONTEXT);
8532 	ASSERT(rw_read_locked(&ctx->ctx_rwlock) == 0);	/* write locked */
8533 
8534 	/*
8535 	 * simply steal and reuse the ctx passed to us.
8536 	 */
8537 	stolen_sfmmup = ctx->ctx_sfmmu;
8538 	ASSERT(sfmmu_hat_lock_held(sfmmup));
8539 	ASSERT(stolen_sfmmup->sfmmu_cnum == cnum);
8540 	ASSERT(stolen_sfmmup != ksfmmup);
8541 
8542 	TRACE_CTXS(&ctx_trace_mutex, ctx_trace_ptr, cnum, stolen_sfmmup,
8543 	    sfmmup, CTX_TRC_STEAL);
8544 	SFMMU_STAT(sf_ctxsteal);
8545 
8546 	/*
8547 	 * Update sfmmu and ctx structs. After this point all threads
8548 	 * belonging to this hat/proc will fault and not use the ctx
8549 	 * being stolen.
8550 	 */
8551 	kpreempt_disable();
8552 	/*
8553 	 * Enforce reverse order of assignments from sfmmu_get_ctx().  This
8554 	 * is done to prevent a race where a thread faults with the context
8555 	 * but the TSB has changed.
8556 	 */
8557 	stolen_sfmmup->sfmmu_cnum = INVALID_CONTEXT;
8558 	membar_enter();
8559 	ctx->ctx_sfmmu = NULL;
8560 
8561 	/*
8562 	 * 1. flush TLB in all CPUs that ran the process whose ctx
8563 	 * we are stealing.
8564 	 * 2. change context for all other CPUs to INVALID_CONTEXT,
8565 	 * if they are running in the context that we are going to steal.
8566 	 */
8567 	cpuset = stolen_sfmmup->sfmmu_cpusran;
8568 	CPUSET_DEL(cpuset, CPU->cpu_id);
8569 	CPUSET_AND(cpuset, cpu_ready_set);
8570 	SFMMU_XCALL_STATS(cnum);
8571 	xt_some(cpuset, sfmmu_ctx_steal_tl1, cnum, INVALID_CONTEXT);
8572 	xt_sync(cpuset);
8573 
8574 	/*
8575 	 * flush TLB of local processor
8576 	 */
8577 	vtag_flushctx(cnum);
8578 
8579 	/*
8580 	 * If we just stole the ctx from the current process
8581 	 * on local cpu then we also invalidate his context
8582 	 * here.
8583 	 */
8584 	if (sfmmu_getctx_sec() == cnum) {
8585 		sfmmu_setctx_sec(INVALID_CONTEXT);
8586 		sfmmu_clear_utsbinfo();
8587 	}
8588 
8589 	kpreempt_enable();
8590 	SFMMU_STAT(sf_tlbflush_ctx);
8591 }
8592 
8593 /*
8594  * Returns a context with the reader lock held.
8595  *
8596  * We maintain 2 different list of contexts.  The first list
8597  * is the free list and it is headed by ctxfree.  These contexts
8598  * are ready to use.  The second list is the dirty list and is
8599  * headed by ctxdirty. These contexts have been freed but haven't
8600  * been flushed from the TLB.
8601  *
8602  * It's the responsibility of the caller to guarantee that the
8603  * process serializes on calls here by taking the HAT lock for
8604  * the hat.
8605  *
8606  * Changing the page size is a rather complicated process, so
8607  * rather than jump through lots of hoops to special case it,
8608  * the easiest way to go about it is to tell the MMU we want
8609  * to change page sizes and then switch to using a different
8610  * context.  When we program the context registers for the
8611  * process, we can take care of setting up the (new) page size
8612  * for that context at that point.
8613  */
8614 
8615 static struct ctx *
8616 sfmmu_get_ctx(sfmmu_t *sfmmup)
8617 {
8618 	struct ctx *ctx;
8619 	ushort_t cnum;
8620 	struct ctx *lastctx = &ctxs[nctxs-1];
8621 	struct ctx *firstctx = &ctxs[NUM_LOCKED_CTXS];
8622 	uint_t	found_stealable_ctx;
8623 	uint_t	retry_count = 0;
8624 
8625 #define	NEXT_CTX(ctx)   (((ctx) >= lastctx) ? firstctx : ((ctx) + 1))
8626 
8627 retry:
8628 
8629 	ASSERT(sfmmup->sfmmu_cnum != KCONTEXT);
8630 	/*
8631 	 * Check to see if this process has already got a ctx.
8632 	 * In that case just set the sec-ctx, grab a readers lock, and
8633 	 * return.
8634 	 *
8635 	 * We have to double check after we get the readers lock on the
8636 	 * context, since it could be stolen in this short window.
8637 	 */
8638 	if (sfmmup->sfmmu_cnum >= NUM_LOCKED_CTXS) {
8639 		ctx = sfmmutoctx(sfmmup);
8640 		rw_enter(&ctx->ctx_rwlock, RW_READER);
8641 		if (ctx->ctx_sfmmu == sfmmup) {
8642 			return (ctx);
8643 		} else {
8644 			rw_exit(&ctx->ctx_rwlock);
8645 		}
8646 	}
8647 
8648 	found_stealable_ctx = 0;
8649 	mutex_enter(&ctx_list_lock);
8650 	if ((ctx = ctxfree) != NULL) {
8651 		/*
8652 		 * Found a ctx in free list. Delete it from the list and
8653 		 * use it.  There's a short window where the stealer can
8654 		 * look at the context before we grab the lock on the
8655 		 * context, so we have to handle that with the free flag.
8656 		 */
8657 		SFMMU_STAT(sf_ctxfree);
8658 		ctxfree = ctx->ctx_free;
8659 		ctx->ctx_sfmmu = NULL;
8660 		mutex_exit(&ctx_list_lock);
8661 		rw_enter(&ctx->ctx_rwlock, RW_WRITER);
8662 		ASSERT(ctx->ctx_sfmmu == NULL);
8663 		ASSERT((ctx->ctx_flags & CTX_FREE_FLAG) != 0);
8664 	} else if ((ctx = ctxdirty) != NULL) {
8665 		/*
8666 		 * No free contexts.  If we have at least one dirty ctx
8667 		 * then flush the TLBs on all cpus if necessary and move
8668 		 * the dirty list to the free list.
8669 		 */
8670 		SFMMU_STAT(sf_ctxdirty);
8671 		ctxdirty = NULL;
8672 		if (delay_tlb_flush)
8673 			sfmmu_tlb_all_demap();
8674 		ctxfree = ctx->ctx_free;
8675 		ctx->ctx_sfmmu = NULL;
8676 		mutex_exit(&ctx_list_lock);
8677 		rw_enter(&ctx->ctx_rwlock, RW_WRITER);
8678 		ASSERT(ctx->ctx_sfmmu == NULL);
8679 		ASSERT((ctx->ctx_flags & CTX_FREE_FLAG) != 0);
8680 	} else {
8681 		/*
8682 		 * No free context available, so steal one.
8683 		 *
8684 		 * The policy to choose the appropriate context is simple;
8685 		 * just sweep all the ctxs using ctxhand. This will steal
8686 		 * the LRU ctx.
8687 		 *
8688 		 * We however only steal a non-free context that can be
8689 		 * write locked.  Keep searching till we find a stealable
8690 		 * ctx.
8691 		 */
8692 		mutex_exit(&ctx_list_lock);
8693 		ctx = ctxhand;
8694 		do {
8695 			/*
8696 			 * If you get the writers lock, and the ctx isn't
8697 			 * a free ctx, THEN you can steal this ctx.
8698 			 */
8699 			if ((ctx->ctx_flags & CTX_FREE_FLAG) == 0 &&
8700 			    rw_tryenter(&ctx->ctx_rwlock, RW_WRITER) != 0) {
8701 				if (ctx->ctx_flags & CTX_FREE_FLAG) {
8702 					/* let the first guy have it */
8703 					rw_exit(&ctx->ctx_rwlock);
8704 				} else {
8705 					found_stealable_ctx = 1;
8706 					break;
8707 				}
8708 			}
8709 			ctx = NEXT_CTX(ctx);
8710 		} while (ctx != ctxhand);
8711 
8712 		if (found_stealable_ctx) {
8713 			/*
8714 			 * Try and reuse the ctx.
8715 			 */
8716 			sfmmu_reuse_ctx(ctx, sfmmup);
8717 
8718 		} else if (retry_count++ < GET_CTX_RETRY_CNT) {
8719 			goto retry;
8720 
8721 		} else {
8722 			panic("Can't find any stealable context");
8723 		}
8724 	}
8725 
8726 	ASSERT(rw_read_locked(&ctx->ctx_rwlock) == 0);	/* write locked */
8727 	ctx->ctx_sfmmu = sfmmup;
8728 
8729 	/*
8730 	 * Clear the ctx_flags field.
8731 	 */
8732 	ctx->ctx_flags = 0;
8733 
8734 	cnum = ctxtoctxnum(ctx);
8735 	membar_exit();
8736 	sfmmup->sfmmu_cnum = cnum;
8737 
8738 	/*
8739 	 * Let the MMU set up the page sizes to use for
8740 	 * this context in the TLB. Don't program 2nd dtlb for ism hat.
8741 	 */
8742 	if ((&mmu_set_ctx_page_sizes) && (sfmmup->sfmmu_ismhat == 0))
8743 		mmu_set_ctx_page_sizes(sfmmup);
8744 
8745 	/*
8746 	 * Downgrade to reader's lock.
8747 	 */
8748 	rw_downgrade(&ctx->ctx_rwlock);
8749 
8750 	/*
8751 	 * If this value doesn't get set to what we want
8752 	 * it won't matter, so don't worry about locking.
8753 	 */
8754 	ctxhand = NEXT_CTX(ctx);
8755 
8756 	/*
8757 	 * Better not have been stolen while we held the ctx'
8758 	 * lock or we're hosed.
8759 	 */
8760 	ASSERT(sfmmup == sfmmutoctx(sfmmup)->ctx_sfmmu);
8761 
8762 	return (ctx);
8763 
8764 #undef NEXT_CTX
8765 }
8766 
8767 
8768 /*
8769  * Set the process context to INVALID_CONTEXT (but
8770  * without stealing the ctx) so that it faults and
8771  * reloads the MMU state from TL=0.  Caller must
8772  * hold the hat lock since we don't acquire it here.
8773  */
8774 static void
8775 sfmmu_sync_mmustate(sfmmu_t *sfmmup)
8776 {
8777 	int cnum;
8778 	cpuset_t cpuset;
8779 
8780 	ASSERT(sfmmup != ksfmmup);
8781 	ASSERT(sfmmu_hat_lock_held(sfmmup));
8782 
8783 	kpreempt_disable();
8784 
8785 	cnum = sfmmutoctxnum(sfmmup);
8786 	if (cnum != INVALID_CONTEXT) {
8787 		cpuset = sfmmup->sfmmu_cpusran;
8788 		CPUSET_DEL(cpuset, CPU->cpu_id);
8789 		CPUSET_AND(cpuset, cpu_ready_set);
8790 		SFMMU_XCALL_STATS(cnum);
8791 
8792 		xt_some(cpuset, sfmmu_raise_tsb_exception,
8793 		    cnum, INVALID_CONTEXT);
8794 		xt_sync(cpuset);
8795 
8796 		/*
8797 		 * If the process is running on the local CPU
8798 		 * we need to update the MMU state here as well.
8799 		 */
8800 		if (sfmmu_getctx_sec() == cnum)
8801 			sfmmu_load_mmustate(sfmmup);
8802 
8803 		SFMMU_STAT(sf_tsb_raise_exception);
8804 	}
8805 
8806 	kpreempt_enable();
8807 }
8808 
8809 
8810 /*
8811  * Replace the specified TSB with a new TSB.  This function gets called when
8812  * we grow, shrink or swapin a TSB.  When swapping in a TSB (TSB_SWAPIN), the
8813  * TSB_FORCEALLOC flag may be used to force allocation of a minimum-sized TSB
8814  * (8K).
8815  *
8816  * Caller must hold the HAT lock, but should assume any tsb_info
8817  * pointers it has are no longer valid after calling this function.
8818  *
8819  * Return values:
8820  *	TSB_ALLOCFAIL	Failed to allocate a TSB, due to memory constraints
8821  *	TSB_LOSTRACE	HAT is busy, i.e. another thread is already doing
8822  *			something to this tsbinfo/TSB
8823  *	TSB_SUCCESS	Operation succeeded
8824  */
8825 static tsb_replace_rc_t
8826 sfmmu_replace_tsb(sfmmu_t *sfmmup, struct tsb_info *old_tsbinfo, uint_t szc,
8827     hatlock_t *hatlockp, uint_t flags)
8828 {
8829 	struct tsb_info *new_tsbinfo = NULL;
8830 	struct tsb_info *curtsb, *prevtsb;
8831 	uint_t tte_sz_mask;
8832 	cpuset_t cpuset;
8833 	struct ctx *ctx = NULL;
8834 	int ctxnum;
8835 
8836 	ASSERT(sfmmup != ksfmmup);
8837 	ASSERT(sfmmup->sfmmu_ismhat == 0);
8838 	ASSERT(sfmmu_hat_lock_held(sfmmup));
8839 	ASSERT(szc <= tsb_max_growsize);
8840 
8841 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_BUSY))
8842 		return (TSB_LOSTRACE);
8843 
8844 	/*
8845 	 * Find the tsb_info ahead of this one in the list, and
8846 	 * also make sure that the tsb_info passed in really
8847 	 * exists!
8848 	 */
8849 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
8850 	    curtsb != old_tsbinfo && curtsb != NULL;
8851 	    prevtsb = curtsb, curtsb = curtsb->tsb_next);
8852 	ASSERT(curtsb != NULL);
8853 
8854 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
8855 		/*
8856 		 * The process is swapped out, so just set the new size
8857 		 * code.  When it swaps back in, we'll allocate a new one
8858 		 * of the new chosen size.
8859 		 */
8860 		curtsb->tsb_szc = szc;
8861 		return (TSB_SUCCESS);
8862 	}
8863 	SFMMU_FLAGS_SET(sfmmup, HAT_BUSY);
8864 
8865 	tte_sz_mask = old_tsbinfo->tsb_ttesz_mask;
8866 
8867 	/*
8868 	 * All initialization is done inside of sfmmu_tsbinfo_alloc().
8869 	 * If we fail to allocate a TSB, exit.
8870 	 */
8871 	sfmmu_hat_exit(hatlockp);
8872 	if (sfmmu_tsbinfo_alloc(&new_tsbinfo, szc, tte_sz_mask,
8873 	    flags, sfmmup)) {
8874 		(void) sfmmu_hat_enter(sfmmup);
8875 		if (!(flags & TSB_SWAPIN))
8876 			SFMMU_STAT(sf_tsb_resize_failures);
8877 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
8878 		return (TSB_ALLOCFAIL);
8879 	}
8880 	(void) sfmmu_hat_enter(sfmmup);
8881 
8882 	/*
8883 	 * Re-check to make sure somebody else didn't muck with us while we
8884 	 * didn't hold the HAT lock.  If the process swapped out, fine, just
8885 	 * exit; this can happen if we try to shrink the TSB from the context
8886 	 * of another process (such as on an ISM unmap), though it is rare.
8887 	 */
8888 	if (!(flags & TSB_SWAPIN) && SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
8889 		SFMMU_STAT(sf_tsb_resize_failures);
8890 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
8891 		sfmmu_hat_exit(hatlockp);
8892 		sfmmu_tsbinfo_free(new_tsbinfo);
8893 		(void) sfmmu_hat_enter(sfmmup);
8894 		return (TSB_LOSTRACE);
8895 	}
8896 
8897 #ifdef	DEBUG
8898 	/* Reverify that the tsb_info still exists.. for debugging only */
8899 	for (prevtsb = NULL, curtsb = sfmmup->sfmmu_tsb;
8900 	    curtsb != old_tsbinfo && curtsb != NULL;
8901 	    prevtsb = curtsb, curtsb = curtsb->tsb_next);
8902 	ASSERT(curtsb != NULL);
8903 #endif	/* DEBUG */
8904 
8905 	/*
8906 	 * Quiesce any CPUs running this process on their next TLB miss
8907 	 * so they atomically see the new tsb_info.  We temporarily set the
8908 	 * context to invalid context so new threads that come on processor
8909 	 * after we do the xcall to cpusran will also serialize behind the
8910 	 * HAT lock on TLB miss and will see the new TSB.  Since this short
8911 	 * race with a new thread coming on processor is relatively rare,
8912 	 * this synchronization mechanism should be cheaper than always
8913 	 * pausing all CPUs for the duration of the setup, which is what
8914 	 * the old implementation did.  This is particuarly true if we are
8915 	 * copying a huge chunk of memory around during that window.
8916 	 *
8917 	 * The memory barriers are to make sure things stay consistent
8918 	 * with resume() since it does not hold the HAT lock while
8919 	 * walking the list of tsb_info structures.
8920 	 */
8921 	if ((flags & TSB_SWAPIN) != TSB_SWAPIN) {
8922 		/* The TSB is either growing or shrinking. */
8923 		ctx = sfmmutoctx(sfmmup);
8924 		rw_enter(&ctx->ctx_rwlock, RW_WRITER);
8925 
8926 		ctxnum = sfmmutoctxnum(sfmmup);
8927 		sfmmup->sfmmu_cnum = INVALID_CONTEXT;
8928 		membar_enter();	/* make sure visible on all CPUs */
8929 
8930 		kpreempt_disable();
8931 		if (ctxnum != INVALID_CONTEXT) {
8932 			cpuset = sfmmup->sfmmu_cpusran;
8933 			CPUSET_DEL(cpuset, CPU->cpu_id);
8934 			CPUSET_AND(cpuset, cpu_ready_set);
8935 			SFMMU_XCALL_STATS(ctxnum);
8936 
8937 			xt_some(cpuset, sfmmu_raise_tsb_exception,
8938 			    ctxnum, INVALID_CONTEXT);
8939 			xt_sync(cpuset);
8940 
8941 			SFMMU_STAT(sf_tsb_raise_exception);
8942 		}
8943 		kpreempt_enable();
8944 	} else {
8945 		/*
8946 		 * It is illegal to swap in TSBs from a process other
8947 		 * than a process being swapped in.  This in turn
8948 		 * implies we do not have a valid MMU context here
8949 		 * since a process needs one to resolve translation
8950 		 * misses.
8951 		 */
8952 		ASSERT(curthread->t_procp->p_as->a_hat == sfmmup);
8953 		ASSERT(sfmmutoctxnum(sfmmup) == INVALID_CONTEXT);
8954 	}
8955 
8956 	new_tsbinfo->tsb_next = old_tsbinfo->tsb_next;
8957 	membar_stst();	/* strict ordering required */
8958 	if (prevtsb)
8959 		prevtsb->tsb_next = new_tsbinfo;
8960 	else
8961 		sfmmup->sfmmu_tsb = new_tsbinfo;
8962 	membar_enter();	/* make sure new TSB globally visible */
8963 	sfmmu_setup_tsbinfo(sfmmup);
8964 
8965 	/*
8966 	 * We need to migrate TSB entries from the old TSB to the new TSB
8967 	 * if tsb_remap_ttes is set and the TSB is growing.
8968 	 */
8969 	if (tsb_remap_ttes && ((flags & TSB_GROW) == TSB_GROW))
8970 		sfmmu_copy_tsb(old_tsbinfo, new_tsbinfo);
8971 
8972 	if ((flags & TSB_SWAPIN) != TSB_SWAPIN) {
8973 		kpreempt_disable();
8974 		membar_exit();
8975 		sfmmup->sfmmu_cnum = ctxnum;
8976 		if (ctxnum != INVALID_CONTEXT &&
8977 		    sfmmu_getctx_sec() == ctxnum) {
8978 			sfmmu_load_mmustate(sfmmup);
8979 		}
8980 		kpreempt_enable();
8981 		rw_exit(&ctx->ctx_rwlock);
8982 	}
8983 
8984 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_BUSY);
8985 
8986 	/*
8987 	 * Drop the HAT lock to free our old tsb_info.
8988 	 */
8989 	sfmmu_hat_exit(hatlockp);
8990 
8991 	if ((flags & TSB_GROW) == TSB_GROW) {
8992 		SFMMU_STAT(sf_tsb_grow);
8993 	} else if ((flags & TSB_SHRINK) == TSB_SHRINK) {
8994 		SFMMU_STAT(sf_tsb_shrink);
8995 	}
8996 
8997 	sfmmu_tsbinfo_free(old_tsbinfo);
8998 
8999 	(void) sfmmu_hat_enter(sfmmup);
9000 	return (TSB_SUCCESS);
9001 }
9002 
9003 /*
9004  * Steal context from process, forcing the process to switch to another
9005  * context on the next TLB miss, and therefore start using the TLB that
9006  * is reprogrammed for the new page sizes.
9007  */
9008 void
9009 sfmmu_steal_context(sfmmu_t *sfmmup, uint8_t *tmp_pgsz)
9010 {
9011 	struct ctx *ctx;
9012 	int i, cnum;
9013 	hatlock_t *hatlockp = NULL;
9014 
9015 	hatlockp = sfmmu_hat_enter(sfmmup);
9016 	/* USIII+-IV+ optimization, requires hat lock */
9017 	if (tmp_pgsz) {
9018 		for (i = 0; i < mmu_page_sizes; i++)
9019 			sfmmup->sfmmu_pgsz[i] = tmp_pgsz[i];
9020 	}
9021 	SFMMU_STAT(sf_tlb_reprog_pgsz);
9022 	ctx = sfmmutoctx(sfmmup);
9023 	rw_enter(&ctx->ctx_rwlock, RW_WRITER);
9024 	cnum = sfmmutoctxnum(sfmmup);
9025 
9026 	if (cnum != INVALID_CONTEXT) {
9027 		sfmmu_tlb_swap_ctx(sfmmup, ctx);
9028 	}
9029 	rw_exit(&ctx->ctx_rwlock);
9030 	sfmmu_hat_exit(hatlockp);
9031 }
9032 
9033 /*
9034  * This function assumes that there are either four or six supported page
9035  * sizes and at most two programmable TLBs, so we need to decide which
9036  * page sizes are most important and then tell the MMU layer so it
9037  * can adjust the TLB page sizes accordingly (if supported).
9038  *
9039  * If these assumptions change, this function will need to be
9040  * updated to support whatever the new limits are.
9041  *
9042  * The growing flag is nonzero if we are growing the address space,
9043  * and zero if it is shrinking.  This allows us to decide whether
9044  * to grow or shrink our TSB, depending upon available memory
9045  * conditions.
9046  */
9047 static void
9048 sfmmu_check_page_sizes(sfmmu_t *sfmmup, int growing)
9049 {
9050 	uint64_t ttecnt[MMU_PAGE_SIZES];
9051 	uint64_t tte8k_cnt, tte4m_cnt;
9052 	uint8_t i;
9053 
9054 	/*
9055 	 * Kernel threads, processes with small address spaces not using
9056 	 * large pages, and dummy ISM HATs need not apply.
9057 	 */
9058 	if (sfmmup == ksfmmup || sfmmup->sfmmu_ismhat != NULL)
9059 		return;
9060 
9061 	if ((sfmmup->sfmmu_flags & HAT_LGPG_FLAGS) == 0 &&
9062 	    sfmmup->sfmmu_ttecnt[TTE8K] <= tsb_rss_factor)
9063 		return;
9064 
9065 	for (i = 0; i < mmu_page_sizes; i++) {
9066 		ttecnt[i] = SFMMU_TTE_CNT(sfmmup, i);
9067 	}
9068 
9069 	/* Check pagesizes in use, and possibly reprogram DTLB. */
9070 	if (&mmu_check_page_sizes)
9071 		mmu_check_page_sizes(sfmmup, ttecnt);
9072 
9073 	/*
9074 	 * Calculate the number of 8k ttes to represent the span of these
9075 	 * pages.
9076 	 */
9077 	tte8k_cnt = ttecnt[TTE8K] +
9078 	    (ttecnt[TTE64K] << (MMU_PAGESHIFT64K - MMU_PAGESHIFT)) +
9079 	    (ttecnt[TTE512K] << (MMU_PAGESHIFT512K - MMU_PAGESHIFT));
9080 	if (mmu_page_sizes == max_mmu_page_sizes) {
9081 		tte4m_cnt = ttecnt[TTE4M] +
9082 		    (ttecnt[TTE32M] << (MMU_PAGESHIFT32M - MMU_PAGESHIFT4M)) +
9083 		    (ttecnt[TTE256M] << (MMU_PAGESHIFT256M - MMU_PAGESHIFT4M));
9084 	} else {
9085 		tte4m_cnt = ttecnt[TTE4M];
9086 	}
9087 
9088 	/*
9089 	 * Inflate TSB sizes by a factor of 2 if this process
9090 	 * uses 4M text pages to minimize extra conflict misses
9091 	 * in the first TSB since without counting text pages
9092 	 * 8K TSB may become too small.
9093 	 *
9094 	 * Also double the size of the second TSB to minimize
9095 	 * extra conflict misses due to competition between 4M text pages
9096 	 * and data pages.
9097 	 */
9098 	if (sfmmup->sfmmu_flags & HAT_4MTEXT_FLAG) {
9099 		tte8k_cnt <<= 1;
9100 		tte4m_cnt <<= 1;
9101 	}
9102 
9103 	/*
9104 	 * Check to see if our TSB is the right size; we may need to
9105 	 * grow or shrink it.  If the process is small, our work is
9106 	 * finished at this point.
9107 	 */
9108 	if (tte8k_cnt <= tsb_rss_factor && tte4m_cnt <= tsb_sectsb_threshold) {
9109 		return;
9110 	}
9111 	sfmmu_size_tsb(sfmmup, growing, tte8k_cnt, tte4m_cnt);
9112 }
9113 
9114 static void
9115 sfmmu_size_tsb(sfmmu_t *sfmmup, int growing, uint64_t tte8k_cnt,
9116 	uint64_t tte4m_cnt)
9117 {
9118 	int tsb_bits;
9119 	uint_t tsb_szc;
9120 	struct tsb_info *tsbinfop;
9121 	hatlock_t *hatlockp = NULL;
9122 
9123 	hatlockp = sfmmu_hat_enter(sfmmup);
9124 	ASSERT(hatlockp != NULL);
9125 	tsbinfop = sfmmup->sfmmu_tsb;
9126 	ASSERT(tsbinfop != NULL);
9127 
9128 	/*
9129 	 * If we're growing, select the size based on RSS.  If we're
9130 	 * shrinking, leave some room so we don't have to turn around and
9131 	 * grow again immediately.
9132 	 */
9133 	if (growing)
9134 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt);
9135 	else
9136 		tsb_szc = SELECT_TSB_SIZECODE(tte8k_cnt << 1);
9137 
9138 	if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
9139 	    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
9140 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
9141 		    hatlockp, TSB_SHRINK);
9142 	} else if (growing && tsb_szc > tsbinfop->tsb_szc && TSB_OK_GROW()) {
9143 		(void) sfmmu_replace_tsb(sfmmup, tsbinfop, tsb_szc,
9144 		    hatlockp, TSB_GROW);
9145 	}
9146 	tsbinfop = sfmmup->sfmmu_tsb;
9147 
9148 	/*
9149 	 * With the TLB and first TSB out of the way, we need to see if
9150 	 * we need a second TSB for 4M pages.  If we managed to reprogram
9151 	 * the TLB page sizes above, the process will start using this new
9152 	 * TSB right away; otherwise, it will start using it on the next
9153 	 * context switch.  Either way, it's no big deal so there's no
9154 	 * synchronization with the trap handlers here unless we grow the
9155 	 * TSB (in which case it's required to prevent using the old one
9156 	 * after it's freed). Note: second tsb is required for 32M/256M
9157 	 * page sizes.
9158 	 */
9159 	if (tte4m_cnt > tsb_sectsb_threshold) {
9160 		/*
9161 		 * If we're growing, select the size based on RSS.  If we're
9162 		 * shrinking, leave some room so we don't have to turn
9163 		 * around and grow again immediately.
9164 		 */
9165 		if (growing)
9166 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt);
9167 		else
9168 			tsb_szc = SELECT_TSB_SIZECODE(tte4m_cnt << 1);
9169 		if (tsbinfop->tsb_next == NULL) {
9170 			struct tsb_info *newtsb;
9171 			int allocflags = SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)?
9172 			    0 : TSB_ALLOC;
9173 
9174 			sfmmu_hat_exit(hatlockp);
9175 
9176 			/*
9177 			 * Try to allocate a TSB for 4[32|256]M pages.  If we
9178 			 * can't get the size we want, retry w/a minimum sized
9179 			 * TSB.  If that still didn't work, give up; we can
9180 			 * still run without one.
9181 			 */
9182 			tsb_bits = (mmu_page_sizes == max_mmu_page_sizes)?
9183 			    TSB4M|TSB32M|TSB256M:TSB4M;
9184 			if ((sfmmu_tsbinfo_alloc(&newtsb, tsb_szc, tsb_bits,
9185 			    allocflags, sfmmup) != 0) &&
9186 			    (sfmmu_tsbinfo_alloc(&newtsb, TSB_MIN_SZCODE,
9187 			    tsb_bits, allocflags, sfmmup) != 0)) {
9188 				return;
9189 			}
9190 
9191 			hatlockp = sfmmu_hat_enter(sfmmup);
9192 
9193 			if (sfmmup->sfmmu_tsb->tsb_next == NULL) {
9194 				sfmmup->sfmmu_tsb->tsb_next = newtsb;
9195 				SFMMU_STAT(sf_tsb_sectsb_create);
9196 				sfmmu_setup_tsbinfo(sfmmup);
9197 				sfmmu_hat_exit(hatlockp);
9198 				return;
9199 			} else {
9200 				/*
9201 				 * It's annoying, but possible for us
9202 				 * to get here.. we dropped the HAT lock
9203 				 * because of locking order in the kmem
9204 				 * allocator, and while we were off getting
9205 				 * our memory, some other thread decided to
9206 				 * do us a favor and won the race to get a
9207 				 * second TSB for this process.  Sigh.
9208 				 */
9209 				sfmmu_hat_exit(hatlockp);
9210 				sfmmu_tsbinfo_free(newtsb);
9211 				return;
9212 			}
9213 		}
9214 
9215 		/*
9216 		 * We have a second TSB, see if it's big enough.
9217 		 */
9218 		tsbinfop = tsbinfop->tsb_next;
9219 
9220 		/*
9221 		 * Check to see if our second TSB is the right size;
9222 		 * we may need to grow or shrink it.
9223 		 * To prevent thrashing (e.g. growing the TSB on a
9224 		 * subsequent map operation), only try to shrink if
9225 		 * the TSB reach exceeds twice the virtual address
9226 		 * space size.
9227 		 */
9228 		if (!growing && (tsb_szc < tsbinfop->tsb_szc) &&
9229 		    (tsb_szc >= default_tsb_size) && TSB_OK_SHRINK()) {
9230 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
9231 			    tsb_szc, hatlockp, TSB_SHRINK);
9232 		} else if (growing && tsb_szc > tsbinfop->tsb_szc &&
9233 		    TSB_OK_GROW()) {
9234 			(void) sfmmu_replace_tsb(sfmmup, tsbinfop,
9235 			    tsb_szc, hatlockp, TSB_GROW);
9236 		}
9237 	}
9238 
9239 	sfmmu_hat_exit(hatlockp);
9240 }
9241 
9242 /*
9243  * Get the preferred page size code for a hat.
9244  * This is only advice, so locking is not done;
9245  * this transitory information could change
9246  * following the call anyway.  This interface is
9247  * sun4 private.
9248  */
9249 /*ARGSUSED*/
9250 uint_t
9251 hat_preferred_pgsz(struct hat *hat, caddr_t vaddr, size_t maplen, int maptype)
9252 {
9253 	sfmmu_t *sfmmup = (sfmmu_t *)hat;
9254 	uint_t szc, maxszc = mmu_page_sizes - 1;
9255 	size_t pgsz;
9256 
9257 	if (maptype == MAPPGSZ_ISM) {
9258 		for (szc = maxszc; szc >= TTE4M; szc--) {
9259 			if (disable_ism_large_pages & (1 << szc))
9260 				continue;
9261 
9262 			pgsz = hw_page_array[szc].hp_size;
9263 			if ((maplen >= pgsz) && IS_P2ALIGNED(vaddr, pgsz))
9264 				return (szc);
9265 		}
9266 		return (TTE4M);
9267 	} else if (&mmu_preferred_pgsz) { /* USIII+-USIV+ */
9268 		return (mmu_preferred_pgsz(sfmmup, vaddr, maplen));
9269 	} else {	/* USIII, USII, Niagara */
9270 		for (szc = maxszc; szc > TTE8K; szc--) {
9271 			if (disable_large_pages & (1 << szc))
9272 				continue;
9273 
9274 			pgsz = hw_page_array[szc].hp_size;
9275 			if ((maplen >= pgsz) && IS_P2ALIGNED(vaddr, pgsz))
9276 				return (szc);
9277 		}
9278 		return (TTE8K);
9279 	}
9280 }
9281 
9282 /*
9283  * Free up a ctx
9284  */
9285 static void
9286 sfmmu_free_ctx(sfmmu_t *sfmmup, struct ctx *ctx)
9287 {
9288 	int ctxnum;
9289 
9290 	rw_enter(&ctx->ctx_rwlock, RW_WRITER);
9291 
9292 	TRACE_CTXS(&ctx_trace_mutex, ctx_trace_ptr, sfmmup->sfmmu_cnum,
9293 	    sfmmup, 0, CTX_TRC_FREE);
9294 
9295 	if (sfmmup->sfmmu_cnum == INVALID_CONTEXT) {
9296 		CPUSET_ZERO(sfmmup->sfmmu_cpusran);
9297 		rw_exit(&ctx->ctx_rwlock);
9298 		return;
9299 	}
9300 
9301 	ASSERT(sfmmup == ctx->ctx_sfmmu);
9302 
9303 	ctx->ctx_sfmmu = NULL;
9304 	ctx->ctx_flags = 0;
9305 	sfmmup->sfmmu_cnum = INVALID_CONTEXT;
9306 	membar_enter();
9307 	CPUSET_ZERO(sfmmup->sfmmu_cpusran);
9308 	ctxnum = sfmmu_getctx_sec();
9309 	if (ctxnum == ctxtoctxnum(ctx)) {
9310 		sfmmu_setctx_sec(INVALID_CONTEXT);
9311 		sfmmu_clear_utsbinfo();
9312 	}
9313 
9314 	/*
9315 	 * Put the freed ctx on the dirty list
9316 	 */
9317 	mutex_enter(&ctx_list_lock);
9318 	CTX_SET_FLAGS(ctx, CTX_FREE_FLAG);
9319 	ctx->ctx_free = ctxdirty;
9320 	ctxdirty = ctx;
9321 	mutex_exit(&ctx_list_lock);
9322 
9323 	rw_exit(&ctx->ctx_rwlock);
9324 }
9325 
9326 /*
9327  * Free up a sfmmu
9328  * Since the sfmmu is currently embedded in the hat struct we simply zero
9329  * out our fields and free up the ism map blk list if any.
9330  */
9331 static void
9332 sfmmu_free_sfmmu(sfmmu_t *sfmmup)
9333 {
9334 	ism_blk_t	*blkp, *nx_blkp;
9335 #ifdef	DEBUG
9336 	ism_map_t	*map;
9337 	int 		i;
9338 #endif
9339 
9340 	ASSERT(sfmmup->sfmmu_ttecnt[TTE8K] == 0);
9341 	ASSERT(sfmmup->sfmmu_ttecnt[TTE64K] == 0);
9342 	ASSERT(sfmmup->sfmmu_ttecnt[TTE512K] == 0);
9343 	ASSERT(sfmmup->sfmmu_ttecnt[TTE4M] == 0);
9344 	ASSERT(sfmmup->sfmmu_ttecnt[TTE32M] == 0);
9345 	ASSERT(sfmmup->sfmmu_ttecnt[TTE256M] == 0);
9346 	ASSERT(sfmmup->sfmmu_cnum == INVALID_CONTEXT);
9347 	sfmmup->sfmmu_free = 0;
9348 	sfmmup->sfmmu_ismhat = 0;
9349 
9350 	blkp = sfmmup->sfmmu_iblk;
9351 	sfmmup->sfmmu_iblk = NULL;
9352 
9353 	while (blkp) {
9354 #ifdef	DEBUG
9355 		map = blkp->iblk_maps;
9356 		for (i = 0; i < ISM_MAP_SLOTS; i++) {
9357 			ASSERT(map[i].imap_seg == 0);
9358 			ASSERT(map[i].imap_ismhat == NULL);
9359 			ASSERT(map[i].imap_ment == NULL);
9360 		}
9361 #endif
9362 		nx_blkp = blkp->iblk_next;
9363 		blkp->iblk_next = NULL;
9364 		blkp->iblk_nextpa = (uint64_t)-1;
9365 		kmem_cache_free(ism_blk_cache, blkp);
9366 		blkp = nx_blkp;
9367 	}
9368 }
9369 
9370 /*
9371  * Locking primitves accessed by HATLOCK macros
9372  */
9373 
9374 #define	SFMMU_SPL_MTX	(0x0)
9375 #define	SFMMU_ML_MTX	(0x1)
9376 
9377 #define	SFMMU_MLSPL_MTX(type, pg)	(((type) == SFMMU_SPL_MTX) ? \
9378 					    SPL_HASH(pg) : MLIST_HASH(pg))
9379 
9380 kmutex_t *
9381 sfmmu_page_enter(struct page *pp)
9382 {
9383 	return (sfmmu_mlspl_enter(pp, SFMMU_SPL_MTX));
9384 }
9385 
9386 static void
9387 sfmmu_page_exit(kmutex_t *spl)
9388 {
9389 	mutex_exit(spl);
9390 }
9391 
9392 static int
9393 sfmmu_page_spl_held(struct page *pp)
9394 {
9395 	return (sfmmu_mlspl_held(pp, SFMMU_SPL_MTX));
9396 }
9397 
9398 kmutex_t *
9399 sfmmu_mlist_enter(struct page *pp)
9400 {
9401 	return (sfmmu_mlspl_enter(pp, SFMMU_ML_MTX));
9402 }
9403 
9404 void
9405 sfmmu_mlist_exit(kmutex_t *mml)
9406 {
9407 	mutex_exit(mml);
9408 }
9409 
9410 int
9411 sfmmu_mlist_held(struct page *pp)
9412 {
9413 
9414 	return (sfmmu_mlspl_held(pp, SFMMU_ML_MTX));
9415 }
9416 
9417 /*
9418  * Common code for sfmmu_mlist_enter() and sfmmu_page_enter().  For
9419  * sfmmu_mlist_enter() case mml_table lock array is used and for
9420  * sfmmu_page_enter() sfmmu_page_lock lock array is used.
9421  *
9422  * The lock is taken on a root page so that it protects an operation on all
9423  * constituent pages of a large page pp belongs to.
9424  *
9425  * The routine takes a lock from the appropriate array. The lock is determined
9426  * by hashing the root page. After taking the lock this routine checks if the
9427  * root page has the same size code that was used to determine the root (i.e
9428  * that root hasn't changed).  If root page has the expected p_szc field we
9429  * have the right lock and it's returned to the caller. If root's p_szc
9430  * decreased we release the lock and retry from the beginning.  This case can
9431  * happen due to hat_page_demote() decreasing p_szc between our load of p_szc
9432  * value and taking the lock. The number of retries due to p_szc decrease is
9433  * limited by the maximum p_szc value. If p_szc is 0 we return the lock
9434  * determined by hashing pp itself.
9435  *
9436  * If our caller doesn't hold a SE_SHARED or SE_EXCL lock on pp it's also
9437  * possible that p_szc can increase. To increase p_szc a thread has to lock
9438  * all constituent pages EXCL and do hat_pageunload() on all of them. All the
9439  * callers that don't hold a page locked recheck if hmeblk through which pp
9440  * was found still maps this pp.  If it doesn't map it anymore returned lock
9441  * is immediately dropped. Therefore if sfmmu_mlspl_enter() hits the case of
9442  * p_szc increase after taking the lock it returns this lock without further
9443  * retries because in this case the caller doesn't care about which lock was
9444  * taken. The caller will drop it right away.
9445  *
9446  * After the routine returns it's guaranteed that hat_page_demote() can't
9447  * change p_szc field of any of constituent pages of a large page pp belongs
9448  * to as long as pp was either locked at least SHARED prior to this call or
9449  * the caller finds that hment that pointed to this pp still references this
9450  * pp (this also assumes that the caller holds hme hash bucket lock so that
9451  * the same pp can't be remapped into the same hmeblk after it was unmapped by
9452  * hat_pageunload()).
9453  */
9454 static kmutex_t *
9455 sfmmu_mlspl_enter(struct page *pp, int type)
9456 {
9457 	kmutex_t	*mtx;
9458 	uint_t		prev_rszc = UINT_MAX;
9459 	page_t		*rootpp;
9460 	uint_t		szc;
9461 	uint_t		rszc;
9462 	uint_t		pszc = pp->p_szc;
9463 
9464 	ASSERT(pp != NULL);
9465 
9466 again:
9467 	if (pszc == 0) {
9468 		mtx = SFMMU_MLSPL_MTX(type, pp);
9469 		mutex_enter(mtx);
9470 		return (mtx);
9471 	}
9472 
9473 	/* The lock lives in the root page */
9474 	rootpp = PP_GROUPLEADER(pp, pszc);
9475 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
9476 	mutex_enter(mtx);
9477 
9478 	/*
9479 	 * Return mml in the following 3 cases:
9480 	 *
9481 	 * 1) If pp itself is root since if its p_szc decreased before we took
9482 	 * the lock pp is still the root of smaller szc page. And if its p_szc
9483 	 * increased it doesn't matter what lock we return (see comment in
9484 	 * front of this routine).
9485 	 *
9486 	 * 2) If pp's not root but rootpp is the root of a rootpp->p_szc size
9487 	 * large page we have the right lock since any previous potential
9488 	 * hat_page_demote() is done demoting from greater than current root's
9489 	 * p_szc because hat_page_demote() changes root's p_szc last. No
9490 	 * further hat_page_demote() can start or be in progress since it
9491 	 * would need the same lock we currently hold.
9492 	 *
9493 	 * 3) If rootpp's p_szc increased since previous iteration it doesn't
9494 	 * matter what lock we return (see comment in front of this routine).
9495 	 */
9496 	if (pp == rootpp || (rszc = rootpp->p_szc) == pszc ||
9497 	    rszc >= prev_rszc) {
9498 		return (mtx);
9499 	}
9500 
9501 	/*
9502 	 * hat_page_demote() could have decreased root's p_szc.
9503 	 * In this case pp's p_szc must also be smaller than pszc.
9504 	 * Retry.
9505 	 */
9506 	if (rszc < pszc) {
9507 		szc = pp->p_szc;
9508 		if (szc < pszc) {
9509 			mutex_exit(mtx);
9510 			pszc = szc;
9511 			goto again;
9512 		}
9513 		/*
9514 		 * pp's p_szc increased after it was decreased.
9515 		 * page cannot be mapped. Return current lock. The caller
9516 		 * will drop it right away.
9517 		 */
9518 		return (mtx);
9519 	}
9520 
9521 	/*
9522 	 * root's p_szc is greater than pp's p_szc.
9523 	 * hat_page_demote() is not done with all pages
9524 	 * yet. Wait for it to complete.
9525 	 */
9526 	mutex_exit(mtx);
9527 	rootpp = PP_GROUPLEADER(rootpp, rszc);
9528 	mtx = SFMMU_MLSPL_MTX(type, rootpp);
9529 	mutex_enter(mtx);
9530 	mutex_exit(mtx);
9531 	prev_rszc = rszc;
9532 	goto again;
9533 }
9534 
9535 static int
9536 sfmmu_mlspl_held(struct page *pp, int type)
9537 {
9538 	kmutex_t	*mtx;
9539 
9540 	ASSERT(pp != NULL);
9541 	/* The lock lives in the root page */
9542 	pp = PP_PAGEROOT(pp);
9543 	ASSERT(pp != NULL);
9544 
9545 	mtx = SFMMU_MLSPL_MTX(type, pp);
9546 	return (MUTEX_HELD(mtx));
9547 }
9548 
9549 static uint_t
9550 sfmmu_get_free_hblk(struct hme_blk **hmeblkpp, uint_t critical)
9551 {
9552 	struct  hme_blk *hblkp;
9553 
9554 	if (freehblkp != NULL) {
9555 		mutex_enter(&freehblkp_lock);
9556 		if (freehblkp != NULL) {
9557 			/*
9558 			 * If the current thread is owning hblk_reserve,
9559 			 * let it succede even if freehblkcnt is really low.
9560 			 */
9561 			if (freehblkcnt <= HBLK_RESERVE_MIN && !critical) {
9562 				SFMMU_STAT(sf_get_free_throttle);
9563 				mutex_exit(&freehblkp_lock);
9564 				return (0);
9565 			}
9566 			freehblkcnt--;
9567 			*hmeblkpp = freehblkp;
9568 			hblkp = *hmeblkpp;
9569 			freehblkp = hblkp->hblk_next;
9570 			mutex_exit(&freehblkp_lock);
9571 			hblkp->hblk_next = NULL;
9572 			SFMMU_STAT(sf_get_free_success);
9573 			return (1);
9574 		}
9575 		mutex_exit(&freehblkp_lock);
9576 	}
9577 	SFMMU_STAT(sf_get_free_fail);
9578 	return (0);
9579 }
9580 
9581 static uint_t
9582 sfmmu_put_free_hblk(struct hme_blk *hmeblkp, uint_t critical)
9583 {
9584 	struct  hme_blk *hblkp;
9585 
9586 	/*
9587 	 * If the current thread is mapping into kernel space,
9588 	 * let it succede even if freehblkcnt is max
9589 	 * so that it will avoid freeing it to kmem.
9590 	 * This will prevent stack overflow due to
9591 	 * possible recursion since kmem_cache_free()
9592 	 * might require creation of a slab which
9593 	 * in turn needs an hmeblk to map that slab;
9594 	 * let's break this vicious chain at the first
9595 	 * opportunity.
9596 	 */
9597 	if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
9598 		mutex_enter(&freehblkp_lock);
9599 		if (freehblkcnt < HBLK_RESERVE_CNT || critical) {
9600 			SFMMU_STAT(sf_put_free_success);
9601 			freehblkcnt++;
9602 			hmeblkp->hblk_next = freehblkp;
9603 			freehblkp = hmeblkp;
9604 			mutex_exit(&freehblkp_lock);
9605 			return (1);
9606 		}
9607 		mutex_exit(&freehblkp_lock);
9608 	}
9609 
9610 	/*
9611 	 * Bring down freehblkcnt to HBLK_RESERVE_CNT. We are here
9612 	 * only if freehblkcnt is at least HBLK_RESERVE_CNT *and*
9613 	 * we are not in the process of mapping into kernel space.
9614 	 */
9615 	ASSERT(!critical);
9616 	while (freehblkcnt > HBLK_RESERVE_CNT) {
9617 		mutex_enter(&freehblkp_lock);
9618 		if (freehblkcnt > HBLK_RESERVE_CNT) {
9619 			freehblkcnt--;
9620 			hblkp = freehblkp;
9621 			freehblkp = hblkp->hblk_next;
9622 			mutex_exit(&freehblkp_lock);
9623 			ASSERT(get_hblk_cache(hblkp) == sfmmu8_cache);
9624 			kmem_cache_free(sfmmu8_cache, hblkp);
9625 			continue;
9626 		}
9627 		mutex_exit(&freehblkp_lock);
9628 	}
9629 	SFMMU_STAT(sf_put_free_fail);
9630 	return (0);
9631 }
9632 
9633 static void
9634 sfmmu_hblk_swap(struct hme_blk *new)
9635 {
9636 	struct hme_blk *old, *hblkp, *prev;
9637 	uint64_t hblkpa, prevpa, newpa;
9638 	caddr_t	base, vaddr, endaddr;
9639 	struct hmehash_bucket *hmebp;
9640 	struct sf_hment *osfhme, *nsfhme;
9641 	page_t *pp;
9642 	kmutex_t *pml;
9643 	tte_t tte;
9644 
9645 #ifdef	DEBUG
9646 	hmeblk_tag		hblktag;
9647 	struct hme_blk		*found;
9648 #endif
9649 	old = HBLK_RESERVE;
9650 
9651 	/*
9652 	 * save pa before bcopy clobbers it
9653 	 */
9654 	newpa = new->hblk_nextpa;
9655 
9656 	base = (caddr_t)get_hblk_base(old);
9657 	endaddr = base + get_hblk_span(old);
9658 
9659 	/*
9660 	 * acquire hash bucket lock.
9661 	 */
9662 	hmebp = sfmmu_tteload_acquire_hashbucket(ksfmmup, base, TTE8K);
9663 
9664 	/*
9665 	 * copy contents from old to new
9666 	 */
9667 	bcopy((void *)old, (void *)new, HME8BLK_SZ);
9668 
9669 	/*
9670 	 * add new to hash chain
9671 	 */
9672 	sfmmu_hblk_hash_add(hmebp, new, newpa);
9673 
9674 	/*
9675 	 * search hash chain for hblk_reserve; this needs to be performed
9676 	 * after adding new, otherwise prevpa and prev won't correspond
9677 	 * to the hblk which is prior to old in hash chain when we call
9678 	 * sfmmu_hblk_hash_rm to remove old later.
9679 	 */
9680 	for (prevpa = 0, prev = NULL,
9681 	    hblkpa = hmebp->hmeh_nextpa, hblkp = hmebp->hmeblkp;
9682 	    hblkp != NULL && hblkp != old;
9683 	    prevpa = hblkpa, prev = hblkp,
9684 	    hblkpa = hblkp->hblk_nextpa, hblkp = hblkp->hblk_next);
9685 
9686 	if (hblkp != old)
9687 		panic("sfmmu_hblk_swap: hblk_reserve not found");
9688 
9689 	/*
9690 	 * p_mapping list is still pointing to hments in hblk_reserve;
9691 	 * fix up p_mapping list so that they point to hments in new.
9692 	 *
9693 	 * Since all these mappings are created by hblk_reserve_thread
9694 	 * on the way and it's using at least one of the buffers from each of
9695 	 * the newly minted slabs, there is no danger of any of these
9696 	 * mappings getting unloaded by another thread.
9697 	 *
9698 	 * tsbmiss could only modify ref/mod bits of hments in old/new.
9699 	 * Since all of these hments hold mappings established by segkmem
9700 	 * and mappings in segkmem are setup with HAT_NOSYNC, ref/mod bits
9701 	 * have no meaning for the mappings in hblk_reserve.  hments in
9702 	 * old and new are identical except for ref/mod bits.
9703 	 */
9704 	for (vaddr = base; vaddr < endaddr; vaddr += TTEBYTES(TTE8K)) {
9705 
9706 		HBLKTOHME(osfhme, old, vaddr);
9707 		sfmmu_copytte(&osfhme->hme_tte, &tte);
9708 
9709 		if (TTE_IS_VALID(&tte)) {
9710 			if ((pp = osfhme->hme_page) == NULL)
9711 				panic("sfmmu_hblk_swap: page not mapped");
9712 
9713 			pml = sfmmu_mlist_enter(pp);
9714 
9715 			if (pp != osfhme->hme_page)
9716 				panic("sfmmu_hblk_swap: mapping changed");
9717 
9718 			HBLKTOHME(nsfhme, new, vaddr);
9719 
9720 			HME_ADD(nsfhme, pp);
9721 			HME_SUB(osfhme, pp);
9722 
9723 			sfmmu_mlist_exit(pml);
9724 		}
9725 	}
9726 
9727 	/*
9728 	 * remove old from hash chain
9729 	 */
9730 	sfmmu_hblk_hash_rm(hmebp, old, prevpa, prev);
9731 
9732 #ifdef	DEBUG
9733 
9734 	hblktag.htag_id = ksfmmup;
9735 	hblktag.htag_bspage = HME_HASH_BSPAGE(base, HME_HASH_SHIFT(TTE8K));
9736 	hblktag.htag_rehash = HME_HASH_REHASH(TTE8K);
9737 	HME_HASH_FAST_SEARCH(hmebp, hblktag, found);
9738 
9739 	if (found != new)
9740 		panic("sfmmu_hblk_swap: new hblk not found");
9741 #endif
9742 
9743 	SFMMU_HASH_UNLOCK(hmebp);
9744 
9745 	/*
9746 	 * Reset hblk_reserve
9747 	 */
9748 	bzero((void *)old, HME8BLK_SZ);
9749 	old->hblk_nextpa = va_to_pa((caddr_t)old);
9750 }
9751 
9752 /*
9753  * Grab the mlist mutex for both pages passed in.
9754  *
9755  * low and high will be returned as pointers to the mutexes for these pages.
9756  * low refers to the mutex residing in the lower bin of the mlist hash, while
9757  * high refers to the mutex residing in the higher bin of the mlist hash.  This
9758  * is due to the locking order restrictions on the same thread grabbing
9759  * multiple mlist mutexes.  The low lock must be acquired before the high lock.
9760  *
9761  * If both pages hash to the same mutex, only grab that single mutex, and
9762  * high will be returned as NULL
9763  * If the pages hash to different bins in the hash, grab the lower addressed
9764  * lock first and then the higher addressed lock in order to follow the locking
9765  * rules involved with the same thread grabbing multiple mlist mutexes.
9766  * low and high will both have non-NULL values.
9767  */
9768 static void
9769 sfmmu_mlist_reloc_enter(struct page *targ, struct page *repl,
9770     kmutex_t **low, kmutex_t **high)
9771 {
9772 	kmutex_t	*mml_targ, *mml_repl;
9773 
9774 	/*
9775 	 * no need to do the dance around szc as in sfmmu_mlist_enter()
9776 	 * because this routine is only called by hat_page_relocate() and all
9777 	 * targ and repl pages are already locked EXCL so szc can't change.
9778 	 */
9779 
9780 	mml_targ = MLIST_HASH(PP_PAGEROOT(targ));
9781 	mml_repl = MLIST_HASH(PP_PAGEROOT(repl));
9782 
9783 	if (mml_targ == mml_repl) {
9784 		*low = mml_targ;
9785 		*high = NULL;
9786 	} else {
9787 		if (mml_targ < mml_repl) {
9788 			*low = mml_targ;
9789 			*high = mml_repl;
9790 		} else {
9791 			*low = mml_repl;
9792 			*high = mml_targ;
9793 		}
9794 	}
9795 
9796 	mutex_enter(*low);
9797 	if (*high)
9798 		mutex_enter(*high);
9799 }
9800 
9801 static void
9802 sfmmu_mlist_reloc_exit(kmutex_t *low, kmutex_t *high)
9803 {
9804 	if (high)
9805 		mutex_exit(high);
9806 	mutex_exit(low);
9807 }
9808 
9809 static hatlock_t *
9810 sfmmu_hat_enter(sfmmu_t *sfmmup)
9811 {
9812 	hatlock_t	*hatlockp;
9813 
9814 	if (sfmmup != ksfmmup) {
9815 		hatlockp = TSB_HASH(sfmmup);
9816 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
9817 		return (hatlockp);
9818 	}
9819 	return (NULL);
9820 }
9821 
9822 static hatlock_t *
9823 sfmmu_hat_tryenter(sfmmu_t *sfmmup)
9824 {
9825 	hatlock_t	*hatlockp;
9826 
9827 	if (sfmmup != ksfmmup) {
9828 		hatlockp = TSB_HASH(sfmmup);
9829 		if (mutex_tryenter(HATLOCK_MUTEXP(hatlockp)) == 0)
9830 			return (NULL);
9831 		return (hatlockp);
9832 	}
9833 	return (NULL);
9834 }
9835 
9836 static void
9837 sfmmu_hat_exit(hatlock_t *hatlockp)
9838 {
9839 	if (hatlockp != NULL)
9840 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
9841 }
9842 
9843 static void
9844 sfmmu_hat_lock_all(void)
9845 {
9846 	int i;
9847 	for (i = 0; i < SFMMU_NUM_LOCK; i++)
9848 		mutex_enter(HATLOCK_MUTEXP(&hat_lock[i]));
9849 }
9850 
9851 static void
9852 sfmmu_hat_unlock_all(void)
9853 {
9854 	int i;
9855 	for (i = SFMMU_NUM_LOCK - 1; i >= 0; i--)
9856 		mutex_exit(HATLOCK_MUTEXP(&hat_lock[i]));
9857 }
9858 
9859 int
9860 sfmmu_hat_lock_held(sfmmu_t *sfmmup)
9861 {
9862 	ASSERT(sfmmup != ksfmmup);
9863 	return (MUTEX_HELD(HATLOCK_MUTEXP(TSB_HASH(sfmmup))));
9864 }
9865 
9866 /*
9867  * Locking primitives to provide consistency between ISM unmap
9868  * and other operations.  Since ISM unmap can take a long time, we
9869  * use HAT_ISMBUSY flag (protected by the hatlock) to avoid creating
9870  * contention on the hatlock buckets while ISM segments are being
9871  * unmapped.  The tradeoff is that the flags don't prevent priority
9872  * inversion from occurring, so we must request kernel priority in
9873  * case we have to sleep to keep from getting buried while holding
9874  * the HAT_ISMBUSY flag set, which in turn could block other kernel
9875  * threads from running (for example, in sfmmu_uvatopfn()).
9876  */
9877 static void
9878 sfmmu_ismhat_enter(sfmmu_t *sfmmup, int hatlock_held)
9879 {
9880 	hatlock_t *hatlockp;
9881 
9882 	THREAD_KPRI_REQUEST();
9883 	if (!hatlock_held)
9884 		hatlockp = sfmmu_hat_enter(sfmmup);
9885 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY))
9886 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
9887 	SFMMU_FLAGS_SET(sfmmup, HAT_ISMBUSY);
9888 	if (!hatlock_held)
9889 		sfmmu_hat_exit(hatlockp);
9890 }
9891 
9892 static void
9893 sfmmu_ismhat_exit(sfmmu_t *sfmmup, int hatlock_held)
9894 {
9895 	hatlock_t *hatlockp;
9896 
9897 	if (!hatlock_held)
9898 		hatlockp = sfmmu_hat_enter(sfmmup);
9899 	ASSERT(SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY));
9900 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_ISMBUSY);
9901 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
9902 	if (!hatlock_held)
9903 		sfmmu_hat_exit(hatlockp);
9904 	THREAD_KPRI_RELEASE();
9905 }
9906 
9907 /*
9908  *
9909  * Algorithm:
9910  *
9911  * (1) if segkmem is not ready, allocate hblk from an array of pre-alloc'ed
9912  *	hblks.
9913  *
9914  * (2) if we are allocating an hblk for mapping a slab in sfmmu_cache,
9915  *
9916  * 		(a) try to return an hblk from reserve pool of free hblks;
9917  *		(b) if the reserve pool is empty, acquire hblk_reserve_lock
9918  *		    and return hblk_reserve.
9919  *
9920  * (3) call kmem_cache_alloc() to allocate hblk;
9921  *
9922  *		(a) if hblk_reserve_lock is held by the current thread,
9923  *		    atomically replace hblk_reserve by the hblk that is
9924  *		    returned by kmem_cache_alloc; release hblk_reserve_lock
9925  *		    and call kmem_cache_alloc() again.
9926  *		(b) if reserve pool is not full, add the hblk that is
9927  *		    returned by kmem_cache_alloc to reserve pool and
9928  *		    call kmem_cache_alloc again.
9929  *
9930  */
9931 static struct hme_blk *
9932 sfmmu_hblk_alloc(sfmmu_t *sfmmup, caddr_t vaddr,
9933 	struct hmehash_bucket *hmebp, uint_t size, hmeblk_tag hblktag,
9934 	uint_t flags)
9935 {
9936 	struct hme_blk *hmeblkp = NULL;
9937 	struct hme_blk *newhblkp;
9938 	struct hme_blk *shw_hblkp = NULL;
9939 	struct kmem_cache *sfmmu_cache = NULL;
9940 	uint64_t hblkpa;
9941 	ulong_t index;
9942 	uint_t owner;		/* set to 1 if using hblk_reserve */
9943 	uint_t forcefree;
9944 	int sleep;
9945 
9946 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
9947 
9948 	/*
9949 	 * If segkmem is not created yet, allocate from static hmeblks
9950 	 * created at the end of startup_modules().  See the block comment
9951 	 * in startup_modules() describing how we estimate the number of
9952 	 * static hmeblks that will be needed during re-map.
9953 	 */
9954 	if (!hblk_alloc_dynamic) {
9955 
9956 		if (size == TTE8K) {
9957 			index = nucleus_hblk8.index;
9958 			if (index >= nucleus_hblk8.len) {
9959 				/*
9960 				 * If we panic here, see startup_modules() to
9961 				 * make sure that we are calculating the
9962 				 * number of hblk8's that we need correctly.
9963 				 */
9964 				panic("no nucleus hblk8 to allocate");
9965 			}
9966 			hmeblkp =
9967 			    (struct hme_blk *)&nucleus_hblk8.list[index];
9968 			nucleus_hblk8.index++;
9969 			SFMMU_STAT(sf_hblk8_nalloc);
9970 		} else {
9971 			index = nucleus_hblk1.index;
9972 			if (nucleus_hblk1.index >= nucleus_hblk1.len) {
9973 				/*
9974 				 * If we panic here, see startup_modules()
9975 				 * and H8TOH1; most likely you need to
9976 				 * update the calculation of the number
9977 				 * of hblk1's the kernel needs to boot.
9978 				 */
9979 				panic("no nucleus hblk1 to allocate");
9980 			}
9981 			hmeblkp =
9982 			    (struct hme_blk *)&nucleus_hblk1.list[index];
9983 			nucleus_hblk1.index++;
9984 			SFMMU_STAT(sf_hblk1_nalloc);
9985 		}
9986 
9987 		goto hblk_init;
9988 	}
9989 
9990 	SFMMU_HASH_UNLOCK(hmebp);
9991 
9992 	if (sfmmup != KHATID) {
9993 		if (mmu_page_sizes == max_mmu_page_sizes) {
9994 			if (size < TTE256M)
9995 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
9996 				    size, flags);
9997 		} else {
9998 			if (size < TTE4M)
9999 				shw_hblkp = sfmmu_shadow_hcreate(sfmmup, vaddr,
10000 				    size, flags);
10001 		}
10002 	}
10003 
10004 fill_hblk:
10005 	owner = (hblk_reserve_thread == curthread) ? 1 : 0;
10006 
10007 	if (owner && size == TTE8K) {
10008 
10009 		/*
10010 		 * We are really in a tight spot. We already own
10011 		 * hblk_reserve and we need another hblk.  In anticipation
10012 		 * of this kind of scenario, we specifically set aside
10013 		 * HBLK_RESERVE_MIN number of hblks to be used exclusively
10014 		 * by owner of hblk_reserve.
10015 		 */
10016 		SFMMU_STAT(sf_hblk_recurse_cnt);
10017 
10018 		if (!sfmmu_get_free_hblk(&hmeblkp, 1))
10019 			panic("sfmmu_hblk_alloc: reserve list is empty");
10020 
10021 		goto hblk_verify;
10022 	}
10023 
10024 	ASSERT(!owner);
10025 
10026 	if ((flags & HAT_NO_KALLOC) == 0) {
10027 
10028 		sfmmu_cache = ((size == TTE8K) ? sfmmu8_cache : sfmmu1_cache);
10029 		sleep = ((sfmmup == KHATID) ? KM_NOSLEEP : KM_SLEEP);
10030 
10031 		if ((hmeblkp = kmem_cache_alloc(sfmmu_cache, sleep)) == NULL) {
10032 			hmeblkp = sfmmu_hblk_steal(size);
10033 		} else {
10034 			/*
10035 			 * if we are the owner of hblk_reserve,
10036 			 * swap hblk_reserve with hmeblkp and
10037 			 * start a fresh life.  Hope things go
10038 			 * better this time.
10039 			 */
10040 			if (hblk_reserve_thread == curthread) {
10041 				ASSERT(sfmmu_cache == sfmmu8_cache);
10042 				sfmmu_hblk_swap(hmeblkp);
10043 				hblk_reserve_thread = NULL;
10044 				mutex_exit(&hblk_reserve_lock);
10045 				goto fill_hblk;
10046 			}
10047 			/*
10048 			 * let's donate this hblk to our reserve list if
10049 			 * we are not mapping kernel range
10050 			 */
10051 			if (size == TTE8K && sfmmup != KHATID)
10052 				if (sfmmu_put_free_hblk(hmeblkp, 0))
10053 					goto fill_hblk;
10054 		}
10055 	} else {
10056 		/*
10057 		 * We are here to map the slab in sfmmu8_cache; let's
10058 		 * check if we could tap our reserve list; if successful,
10059 		 * this will avoid the pain of going thru sfmmu_hblk_swap
10060 		 */
10061 		SFMMU_STAT(sf_hblk_slab_cnt);
10062 		if (!sfmmu_get_free_hblk(&hmeblkp, 0)) {
10063 			/*
10064 			 * let's start hblk_reserve dance
10065 			 */
10066 			SFMMU_STAT(sf_hblk_reserve_cnt);
10067 			owner = 1;
10068 			mutex_enter(&hblk_reserve_lock);
10069 			hmeblkp = HBLK_RESERVE;
10070 			hblk_reserve_thread = curthread;
10071 		}
10072 	}
10073 
10074 hblk_verify:
10075 	ASSERT(hmeblkp != NULL);
10076 	set_hblk_sz(hmeblkp, size);
10077 	ASSERT(hmeblkp->hblk_nextpa == va_to_pa((caddr_t)hmeblkp));
10078 	SFMMU_HASH_LOCK(hmebp);
10079 	HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
10080 	if (newhblkp != NULL) {
10081 		SFMMU_HASH_UNLOCK(hmebp);
10082 		if (hmeblkp != HBLK_RESERVE) {
10083 			/*
10084 			 * This is really tricky!
10085 			 *
10086 			 * vmem_alloc(vmem_seg_arena)
10087 			 *  vmem_alloc(vmem_internal_arena)
10088 			 *   segkmem_alloc(heap_arena)
10089 			 *    vmem_alloc(heap_arena)
10090 			 *    page_create()
10091 			 *    hat_memload()
10092 			 *	kmem_cache_free()
10093 			 *	 kmem_cache_alloc()
10094 			 *	  kmem_slab_create()
10095 			 *	   vmem_alloc(kmem_internal_arena)
10096 			 *	    segkmem_alloc(heap_arena)
10097 			 *		vmem_alloc(heap_arena)
10098 			 *		page_create()
10099 			 *		hat_memload()
10100 			 *		  kmem_cache_free()
10101 			 *		...
10102 			 *
10103 			 * Thus, hat_memload() could call kmem_cache_free
10104 			 * for enough number of times that we could easily
10105 			 * hit the bottom of the stack or run out of reserve
10106 			 * list of vmem_seg structs.  So, we must donate
10107 			 * this hblk to reserve list if it's allocated
10108 			 * from sfmmu8_cache *and* mapping kernel range.
10109 			 * We don't need to worry about freeing hmeblk1's
10110 			 * to kmem since they don't map any kmem slabs.
10111 			 *
10112 			 * Note: When segkmem supports largepages, we must
10113 			 * free hmeblk1's to reserve list as well.
10114 			 */
10115 			forcefree = (sfmmup == KHATID) ? 1 : 0;
10116 			if (size == TTE8K &&
10117 			    sfmmu_put_free_hblk(hmeblkp, forcefree)) {
10118 				goto re_verify;
10119 			}
10120 			ASSERT(sfmmup != KHATID);
10121 			kmem_cache_free(get_hblk_cache(hmeblkp), hmeblkp);
10122 		} else {
10123 			/*
10124 			 * Hey! we don't need hblk_reserve any more.
10125 			 */
10126 			ASSERT(owner);
10127 			hblk_reserve_thread = NULL;
10128 			mutex_exit(&hblk_reserve_lock);
10129 			owner = 0;
10130 		}
10131 re_verify:
10132 		/*
10133 		 * let's check if the goodies are still present
10134 		 */
10135 		SFMMU_HASH_LOCK(hmebp);
10136 		HME_HASH_FAST_SEARCH(hmebp, hblktag, newhblkp);
10137 		if (newhblkp != NULL) {
10138 			/*
10139 			 * return newhblkp if it's not hblk_reserve;
10140 			 * if newhblkp is hblk_reserve, return it
10141 			 * _only if_ we are the owner of hblk_reserve.
10142 			 */
10143 			if (newhblkp != HBLK_RESERVE || owner) {
10144 				return (newhblkp);
10145 			} else {
10146 				/*
10147 				 * we just hit hblk_reserve in the hash and
10148 				 * we are not the owner of that;
10149 				 *
10150 				 * block until hblk_reserve_thread completes
10151 				 * swapping hblk_reserve and try the dance
10152 				 * once again.
10153 				 */
10154 				SFMMU_HASH_UNLOCK(hmebp);
10155 				mutex_enter(&hblk_reserve_lock);
10156 				mutex_exit(&hblk_reserve_lock);
10157 				SFMMU_STAT(sf_hblk_reserve_hit);
10158 				goto fill_hblk;
10159 			}
10160 		} else {
10161 			/*
10162 			 * it's no more! try the dance once again.
10163 			 */
10164 			SFMMU_HASH_UNLOCK(hmebp);
10165 			goto fill_hblk;
10166 		}
10167 	}
10168 
10169 hblk_init:
10170 	set_hblk_sz(hmeblkp, size);
10171 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
10172 	hmeblkp->hblk_next = (struct hme_blk *)NULL;
10173 	hmeblkp->hblk_tag = hblktag;
10174 	hmeblkp->hblk_shadow = shw_hblkp;
10175 	hblkpa = hmeblkp->hblk_nextpa;
10176 	hmeblkp->hblk_nextpa = 0;
10177 
10178 	ASSERT(get_hblk_ttesz(hmeblkp) == size);
10179 	ASSERT(get_hblk_span(hmeblkp) == HMEBLK_SPAN(size));
10180 	ASSERT(hmeblkp->hblk_hmecnt == 0);
10181 	ASSERT(hmeblkp->hblk_vcnt == 0);
10182 	ASSERT(hmeblkp->hblk_lckcnt == 0);
10183 	ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
10184 	sfmmu_hblk_hash_add(hmebp, hmeblkp, hblkpa);
10185 	return (hmeblkp);
10186 }
10187 
10188 /*
10189  * This function performs any cleanup required on the hme_blk
10190  * and returns it to the free list.
10191  */
10192 /* ARGSUSED */
10193 static void
10194 sfmmu_hblk_free(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
10195 	uint64_t hblkpa, struct hme_blk **listp)
10196 {
10197 	int shw_size, vshift;
10198 	struct hme_blk *shw_hblkp;
10199 	uint_t		shw_mask, newshw_mask;
10200 	uintptr_t	vaddr;
10201 	int		size;
10202 	uint_t		critical;
10203 
10204 	ASSERT(hmeblkp);
10205 	ASSERT(!hmeblkp->hblk_hmecnt);
10206 	ASSERT(!hmeblkp->hblk_vcnt);
10207 	ASSERT(!hmeblkp->hblk_lckcnt);
10208 	ASSERT(hblkpa == va_to_pa((caddr_t)hmeblkp));
10209 	ASSERT(hmeblkp != (struct hme_blk *)hblk_reserve);
10210 
10211 	critical = (hblktosfmmu(hmeblkp) == KHATID) ? 1 : 0;
10212 
10213 	size = get_hblk_ttesz(hmeblkp);
10214 	shw_hblkp = hmeblkp->hblk_shadow;
10215 	if (shw_hblkp) {
10216 		ASSERT(hblktosfmmu(hmeblkp) != KHATID);
10217 		if (mmu_page_sizes == max_mmu_page_sizes) {
10218 			ASSERT(size < TTE256M);
10219 		} else {
10220 			ASSERT(size < TTE4M);
10221 		}
10222 
10223 		shw_size = get_hblk_ttesz(shw_hblkp);
10224 		vaddr = get_hblk_base(hmeblkp);
10225 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
10226 		ASSERT(vshift < 8);
10227 		/*
10228 		 * Atomically clear shadow mask bit
10229 		 */
10230 		do {
10231 			shw_mask = shw_hblkp->hblk_shw_mask;
10232 			ASSERT(shw_mask & (1 << vshift));
10233 			newshw_mask = shw_mask & ~(1 << vshift);
10234 			newshw_mask = cas32(&shw_hblkp->hblk_shw_mask,
10235 				shw_mask, newshw_mask);
10236 		} while (newshw_mask != shw_mask);
10237 		hmeblkp->hblk_shadow = NULL;
10238 	}
10239 	hmeblkp->hblk_next = NULL;
10240 	hmeblkp->hblk_nextpa = hblkpa;
10241 	hmeblkp->hblk_shw_bit = 0;
10242 
10243 	if (hmeblkp->hblk_nuc_bit == 0) {
10244 
10245 		if (size == TTE8K && sfmmu_put_free_hblk(hmeblkp, critical))
10246 			return;
10247 
10248 		hmeblkp->hblk_next = *listp;
10249 		*listp = hmeblkp;
10250 	}
10251 }
10252 
10253 static void
10254 sfmmu_hblks_list_purge(struct hme_blk **listp)
10255 {
10256 	struct hme_blk	*hmeblkp;
10257 
10258 	while ((hmeblkp = *listp) != NULL) {
10259 		*listp = hmeblkp->hblk_next;
10260 		kmem_cache_free(get_hblk_cache(hmeblkp), hmeblkp);
10261 	}
10262 }
10263 
10264 #define	BUCKETS_TO_SEARCH_BEFORE_UNLOAD	30
10265 
10266 static uint_t sfmmu_hblk_steal_twice;
10267 static uint_t sfmmu_hblk_steal_count, sfmmu_hblk_steal_unload_count;
10268 
10269 /*
10270  * Steal a hmeblk
10271  * Enough hmeblks were allocated at startup (nucleus hmeblks) and also
10272  * hmeblks were added dynamically. We should never ever not be able to
10273  * find one. Look for an unused/unlocked hmeblk in user hash table.
10274  */
10275 static struct hme_blk *
10276 sfmmu_hblk_steal(int size)
10277 {
10278 	static struct hmehash_bucket *uhmehash_steal_hand = NULL;
10279 	struct hmehash_bucket *hmebp;
10280 	struct hme_blk *hmeblkp = NULL, *pr_hblk;
10281 	uint64_t hblkpa, prevpa;
10282 	int i;
10283 
10284 	for (;;) {
10285 		hmebp = (uhmehash_steal_hand == NULL) ? uhme_hash :
10286 			uhmehash_steal_hand;
10287 		ASSERT(hmebp >= uhme_hash && hmebp <= &uhme_hash[UHMEHASH_SZ]);
10288 
10289 		for (i = 0; hmeblkp == NULL && i <= UHMEHASH_SZ +
10290 		    BUCKETS_TO_SEARCH_BEFORE_UNLOAD; i++) {
10291 			SFMMU_HASH_LOCK(hmebp);
10292 			hmeblkp = hmebp->hmeblkp;
10293 			hblkpa = hmebp->hmeh_nextpa;
10294 			prevpa = 0;
10295 			pr_hblk = NULL;
10296 			while (hmeblkp) {
10297 				/*
10298 				 * check if it is a hmeblk that is not locked
10299 				 * and not shared. skip shadow hmeblks with
10300 				 * shadow_mask set i.e valid count non zero.
10301 				 */
10302 				if ((get_hblk_ttesz(hmeblkp) == size) &&
10303 				    (hmeblkp->hblk_shw_bit == 0 ||
10304 					hmeblkp->hblk_vcnt == 0) &&
10305 				    (hmeblkp->hblk_lckcnt == 0)) {
10306 					/*
10307 					 * there is a high probability that we
10308 					 * will find a free one. search some
10309 					 * buckets for a free hmeblk initially
10310 					 * before unloading a valid hmeblk.
10311 					 */
10312 					if ((hmeblkp->hblk_vcnt == 0 &&
10313 					    hmeblkp->hblk_hmecnt == 0) || (i >=
10314 					    BUCKETS_TO_SEARCH_BEFORE_UNLOAD)) {
10315 						if (sfmmu_steal_this_hblk(hmebp,
10316 						    hmeblkp, hblkpa, prevpa,
10317 						    pr_hblk)) {
10318 							/*
10319 							 * Hblk is unloaded
10320 							 * successfully
10321 							 */
10322 							break;
10323 						}
10324 					}
10325 				}
10326 				pr_hblk = hmeblkp;
10327 				prevpa = hblkpa;
10328 				hblkpa = hmeblkp->hblk_nextpa;
10329 				hmeblkp = hmeblkp->hblk_next;
10330 			}
10331 
10332 			SFMMU_HASH_UNLOCK(hmebp);
10333 			if (hmebp++ == &uhme_hash[UHMEHASH_SZ])
10334 				hmebp = uhme_hash;
10335 		}
10336 		uhmehash_steal_hand = hmebp;
10337 
10338 		if (hmeblkp != NULL)
10339 			break;
10340 
10341 		/*
10342 		 * in the worst case, look for a free one in the kernel
10343 		 * hash table.
10344 		 */
10345 		for (i = 0, hmebp = khme_hash; i <= KHMEHASH_SZ; i++) {
10346 			SFMMU_HASH_LOCK(hmebp);
10347 			hmeblkp = hmebp->hmeblkp;
10348 			hblkpa = hmebp->hmeh_nextpa;
10349 			prevpa = 0;
10350 			pr_hblk = NULL;
10351 			while (hmeblkp) {
10352 				/*
10353 				 * check if it is free hmeblk
10354 				 */
10355 				if ((get_hblk_ttesz(hmeblkp) == size) &&
10356 				    (hmeblkp->hblk_lckcnt == 0) &&
10357 				    (hmeblkp->hblk_vcnt == 0) &&
10358 				    (hmeblkp->hblk_hmecnt == 0)) {
10359 					if (sfmmu_steal_this_hblk(hmebp,
10360 					    hmeblkp, hblkpa, prevpa, pr_hblk)) {
10361 						break;
10362 					} else {
10363 						/*
10364 						 * Cannot fail since we have
10365 						 * hash lock.
10366 						 */
10367 						panic("fail to steal?");
10368 					}
10369 				}
10370 
10371 				pr_hblk = hmeblkp;
10372 				prevpa = hblkpa;
10373 				hblkpa = hmeblkp->hblk_nextpa;
10374 				hmeblkp = hmeblkp->hblk_next;
10375 			}
10376 
10377 			SFMMU_HASH_UNLOCK(hmebp);
10378 			if (hmebp++ == &khme_hash[KHMEHASH_SZ])
10379 				hmebp = khme_hash;
10380 		}
10381 
10382 		if (hmeblkp != NULL)
10383 			break;
10384 		sfmmu_hblk_steal_twice++;
10385 	}
10386 	return (hmeblkp);
10387 }
10388 
10389 /*
10390  * This routine does real work to prepare a hblk to be "stolen" by
10391  * unloading the mappings, updating shadow counts ....
10392  * It returns 1 if the block is ready to be reused (stolen), or 0
10393  * means the block cannot be stolen yet- pageunload is still working
10394  * on this hblk.
10395  */
10396 static int
10397 sfmmu_steal_this_hblk(struct hmehash_bucket *hmebp, struct hme_blk *hmeblkp,
10398 	uint64_t hblkpa, uint64_t prevpa, struct hme_blk *pr_hblk)
10399 {
10400 	int shw_size, vshift;
10401 	struct hme_blk *shw_hblkp;
10402 	uintptr_t vaddr;
10403 	uint_t shw_mask, newshw_mask;
10404 
10405 	ASSERT(SFMMU_HASH_LOCK_ISHELD(hmebp));
10406 
10407 	/*
10408 	 * check if the hmeblk is free, unload if necessary
10409 	 */
10410 	if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
10411 		sfmmu_t *sfmmup;
10412 		demap_range_t dmr;
10413 
10414 		sfmmup = hblktosfmmu(hmeblkp);
10415 		DEMAP_RANGE_INIT(sfmmup, &dmr);
10416 		(void) sfmmu_hblk_unload(sfmmup, hmeblkp,
10417 		    (caddr_t)get_hblk_base(hmeblkp),
10418 		    get_hblk_endaddr(hmeblkp), &dmr, HAT_UNLOAD);
10419 		DEMAP_RANGE_FLUSH(&dmr);
10420 		if (hmeblkp->hblk_vcnt || hmeblkp->hblk_hmecnt) {
10421 			/*
10422 			 * Pageunload is working on the same hblk.
10423 			 */
10424 			return (0);
10425 		}
10426 
10427 		sfmmu_hblk_steal_unload_count++;
10428 	}
10429 
10430 	ASSERT(hmeblkp->hblk_lckcnt == 0);
10431 	ASSERT(hmeblkp->hblk_vcnt == 0 && hmeblkp->hblk_hmecnt == 0);
10432 
10433 	sfmmu_hblk_hash_rm(hmebp, hmeblkp, prevpa, pr_hblk);
10434 	hmeblkp->hblk_nextpa = hblkpa;
10435 
10436 	shw_hblkp = hmeblkp->hblk_shadow;
10437 	if (shw_hblkp) {
10438 		shw_size = get_hblk_ttesz(shw_hblkp);
10439 		vaddr = get_hblk_base(hmeblkp);
10440 		vshift = vaddr_to_vshift(shw_hblkp->hblk_tag, vaddr, shw_size);
10441 		ASSERT(vshift < 8);
10442 		/*
10443 		 * Atomically clear shadow mask bit
10444 		 */
10445 		do {
10446 			shw_mask = shw_hblkp->hblk_shw_mask;
10447 			ASSERT(shw_mask & (1 << vshift));
10448 			newshw_mask = shw_mask & ~(1 << vshift);
10449 			newshw_mask = cas32(&shw_hblkp->hblk_shw_mask,
10450 				shw_mask, newshw_mask);
10451 		} while (newshw_mask != shw_mask);
10452 		hmeblkp->hblk_shadow = NULL;
10453 	}
10454 
10455 	/*
10456 	 * remove shadow bit if we are stealing an unused shadow hmeblk.
10457 	 * sfmmu_hblk_alloc needs it that way, will set shadow bit later if
10458 	 * we are indeed allocating a shadow hmeblk.
10459 	 */
10460 	hmeblkp->hblk_shw_bit = 0;
10461 
10462 	sfmmu_hblk_steal_count++;
10463 	SFMMU_STAT(sf_steal_count);
10464 
10465 	return (1);
10466 }
10467 
10468 struct hme_blk *
10469 sfmmu_hmetohblk(struct sf_hment *sfhme)
10470 {
10471 	struct hme_blk *hmeblkp;
10472 	struct sf_hment *sfhme0;
10473 	struct hme_blk *hblk_dummy = 0;
10474 
10475 	/*
10476 	 * No dummy sf_hments, please.
10477 	 */
10478 	ASSERT(sfhme->hme_tte.ll != 0);
10479 
10480 	sfhme0 = sfhme - sfhme->hme_tte.tte_hmenum;
10481 	hmeblkp = (struct hme_blk *)((uintptr_t)sfhme0 -
10482 		(uintptr_t)&hblk_dummy->hblk_hme[0]);
10483 
10484 	return (hmeblkp);
10485 }
10486 
10487 /*
10488  * Make sure that there is a valid ctx, if not get a ctx.
10489  * Also, get a readers lock on the ctx, so that the ctx cannot
10490  * be stolen underneath us.
10491  */
10492 static void
10493 sfmmu_disallow_ctx_steal(sfmmu_t *sfmmup)
10494 {
10495 	struct	ctx *ctx;
10496 
10497 	ASSERT(sfmmup != ksfmmup);
10498 	ASSERT(sfmmup->sfmmu_ismhat == 0);
10499 
10500 	/*
10501 	 * If ctx has been stolen, get a ctx.
10502 	 */
10503 	if (sfmmup->sfmmu_cnum == INVALID_CONTEXT) {
10504 		/*
10505 		 * Our ctx was stolen. Get a ctx with rlock.
10506 		 */
10507 		ctx = sfmmu_get_ctx(sfmmup);
10508 		return;
10509 	} else {
10510 		ctx = sfmmutoctx(sfmmup);
10511 	}
10512 
10513 	/*
10514 	 * Get the reader lock.
10515 	 */
10516 	rw_enter(&ctx->ctx_rwlock, RW_READER);
10517 	if (ctx->ctx_sfmmu != sfmmup) {
10518 		/*
10519 		 * The ctx got stolen, so spin again.
10520 		 */
10521 		rw_exit(&ctx->ctx_rwlock);
10522 		ctx = sfmmu_get_ctx(sfmmup);
10523 	}
10524 
10525 	ASSERT(sfmmup->sfmmu_cnum >= NUM_LOCKED_CTXS);
10526 }
10527 
10528 /*
10529  * Decrement reference count for our ctx. If the reference count
10530  * becomes 0, our ctx can be stolen by someone.
10531  */
10532 static void
10533 sfmmu_allow_ctx_steal(sfmmu_t *sfmmup)
10534 {
10535 	struct	ctx *ctx;
10536 
10537 	ASSERT(sfmmup != ksfmmup);
10538 	ASSERT(sfmmup->sfmmu_ismhat == 0);
10539 	ctx = sfmmutoctx(sfmmup);
10540 
10541 	ASSERT(sfmmup == ctx->ctx_sfmmu);
10542 	ASSERT(sfmmup->sfmmu_cnum != INVALID_CONTEXT);
10543 	rw_exit(&ctx->ctx_rwlock);
10544 }
10545 
10546 /*
10547  * On swapin, get appropriately sized TSB(s) and clear the HAT_SWAPPED flag.
10548  * If we can't get appropriately sized TSB(s), try for 8K TSB(s) using
10549  * KM_SLEEP allocation.
10550  *
10551  * Return 0 on success, -1 otherwise.
10552  */
10553 static void
10554 sfmmu_tsb_swapin(sfmmu_t *sfmmup, hatlock_t *hatlockp)
10555 {
10556 	struct tsb_info *tsbinfop, *next;
10557 	tsb_replace_rc_t rc;
10558 	boolean_t gotfirst = B_FALSE;
10559 
10560 	ASSERT(sfmmup != ksfmmup);
10561 	ASSERT(sfmmu_hat_lock_held(sfmmup));
10562 
10563 	while (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPIN)) {
10564 		cv_wait(&sfmmup->sfmmu_tsb_cv, HATLOCK_MUTEXP(hatlockp));
10565 	}
10566 
10567 	if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
10568 		SFMMU_FLAGS_SET(sfmmup, HAT_SWAPIN);
10569 	} else {
10570 		return;
10571 	}
10572 
10573 	ASSERT(sfmmup->sfmmu_tsb != NULL);
10574 
10575 	/*
10576 	 * Loop over all tsbinfo's replacing them with ones that actually have
10577 	 * a TSB.  If any of the replacements ever fail, bail out of the loop.
10578 	 */
10579 	for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL; tsbinfop = next) {
10580 		ASSERT(tsbinfop->tsb_flags & TSB_SWAPPED);
10581 		next = tsbinfop->tsb_next;
10582 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, tsbinfop->tsb_szc,
10583 		    hatlockp, TSB_SWAPIN);
10584 		if (rc != TSB_SUCCESS) {
10585 			break;
10586 		}
10587 		gotfirst = B_TRUE;
10588 	}
10589 
10590 	switch (rc) {
10591 	case TSB_SUCCESS:
10592 		SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
10593 		cv_broadcast(&sfmmup->sfmmu_tsb_cv);
10594 		return;
10595 	case TSB_ALLOCFAIL:
10596 		break;
10597 	default:
10598 		panic("sfmmu_replace_tsb returned unrecognized failure code "
10599 		    "%d", rc);
10600 	}
10601 
10602 	/*
10603 	 * In this case, we failed to get one of our TSBs.  If we failed to
10604 	 * get the first TSB, get one of minimum size (8KB).  Walk the list
10605 	 * and throw away the tsbinfos, starting where the allocation failed;
10606 	 * we can get by with just one TSB as long as we don't leave the
10607 	 * SWAPPED tsbinfo structures lying around.
10608 	 */
10609 	tsbinfop = sfmmup->sfmmu_tsb;
10610 	next = tsbinfop->tsb_next;
10611 	tsbinfop->tsb_next = NULL;
10612 
10613 	sfmmu_hat_exit(hatlockp);
10614 	for (tsbinfop = next; tsbinfop != NULL; tsbinfop = next) {
10615 		next = tsbinfop->tsb_next;
10616 		sfmmu_tsbinfo_free(tsbinfop);
10617 	}
10618 	hatlockp = sfmmu_hat_enter(sfmmup);
10619 
10620 	/*
10621 	 * If we don't have any TSBs, get a single 8K TSB for 8K, 64K and 512K
10622 	 * pages.
10623 	 */
10624 	if (!gotfirst) {
10625 		tsbinfop = sfmmup->sfmmu_tsb;
10626 		rc = sfmmu_replace_tsb(sfmmup, tsbinfop, TSB_MIN_SZCODE,
10627 		    hatlockp, TSB_SWAPIN | TSB_FORCEALLOC);
10628 		ASSERT(rc == TSB_SUCCESS);
10629 	}
10630 
10631 	SFMMU_FLAGS_CLEAR(sfmmup, HAT_SWAPPED|HAT_SWAPIN);
10632 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
10633 }
10634 
10635 /*
10636  * Handle exceptions for low level tsb_handler.
10637  *
10638  * There are many scenarios that could land us here:
10639  *
10640  *	1) Process has no context.  In this case, ctx is
10641  *         INVALID_CONTEXT and sfmmup->sfmmu_cnum == 1 so
10642  *         we will acquire a context before returning.
10643  *      2) Need to re-load our MMU state.  In this case,
10644  *         ctx is INVALID_CONTEXT and sfmmup->sfmmu_cnum != 1.
10645  *      3) ISM mappings are being updated.  This is handled
10646  *         just like case #2.
10647  *      4) We wish to program a new page size into the TLB.
10648  *         This is handled just like case #1, since changing
10649  *         TLB page size requires us to flush the TLB.
10650  *	5) Window fault and no valid translation found.
10651  *
10652  * Cases 1-4, ctx is INVALID_CONTEXT so we handle it and then
10653  * exit which will retry the trapped instruction.  Case #5 we
10654  * punt to trap() which will raise us a trap level and handle
10655  * the fault before unwinding.
10656  *
10657  * Note that the process will run in INVALID_CONTEXT before
10658  * faulting into here and subsequently loading the MMU registers
10659  * (including the TSB base register) associated with this process.
10660  * For this reason, the trap handlers must all test for
10661  * INVALID_CONTEXT before attempting to access any registers other
10662  * than the context registers.
10663  */
10664 void
10665 sfmmu_tsbmiss_exception(struct regs *rp, uintptr_t tagaccess, uint_t traptype)
10666 {
10667 	sfmmu_t *sfmmup;
10668 	uint_t ctxnum;
10669 	klwp_id_t lwp;
10670 	char lwp_save_state;
10671 	hatlock_t *hatlockp;
10672 	struct tsb_info *tsbinfop;
10673 
10674 	SFMMU_STAT(sf_tsb_exceptions);
10675 	sfmmup = astosfmmu(curthread->t_procp->p_as);
10676 	ctxnum = tagaccess & TAGACC_CTX_MASK;
10677 
10678 	ASSERT(sfmmup != ksfmmup && ctxnum != KCONTEXT);
10679 	ASSERT(sfmmup->sfmmu_ismhat == 0);
10680 	/*
10681 	 * First, make sure we come out of here with a valid ctx,
10682 	 * since if we don't get one we'll simply loop on the
10683 	 * faulting instruction.
10684 	 *
10685 	 * If the ISM mappings are changing, the TSB is being relocated, or
10686 	 * the process is swapped out we serialize behind the controlling
10687 	 * thread with the sfmmu_flags and sfmmu_tsb_cv condition variable.
10688 	 * Otherwise we synchronize with the context stealer or the thread
10689 	 * that required us to change out our MMU registers (such
10690 	 * as a thread changing out our TSB while we were running) by
10691 	 * locking the HAT and grabbing the rwlock on the context as a
10692 	 * reader temporarily.
10693 	 */
10694 	if (ctxnum == INVALID_CONTEXT ||
10695 	    SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
10696 		/*
10697 		 * Must set lwp state to LWP_SYS before
10698 		 * trying to acquire any adaptive lock
10699 		 */
10700 		lwp = ttolwp(curthread);
10701 		ASSERT(lwp);
10702 		lwp_save_state = lwp->lwp_state;
10703 		lwp->lwp_state = LWP_SYS;
10704 
10705 		hatlockp = sfmmu_hat_enter(sfmmup);
10706 retry:
10707 		for (tsbinfop = sfmmup->sfmmu_tsb; tsbinfop != NULL;
10708 		    tsbinfop = tsbinfop->tsb_next) {
10709 			if (tsbinfop->tsb_flags & TSB_RELOC_FLAG) {
10710 				cv_wait(&sfmmup->sfmmu_tsb_cv,
10711 				    HATLOCK_MUTEXP(hatlockp));
10712 				goto retry;
10713 			}
10714 		}
10715 
10716 		/*
10717 		 * Wait for ISM maps to be updated.
10718 		 */
10719 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_ISMBUSY)) {
10720 			cv_wait(&sfmmup->sfmmu_tsb_cv,
10721 				    HATLOCK_MUTEXP(hatlockp));
10722 			goto retry;
10723 		}
10724 
10725 		/*
10726 		 * If we're swapping in, get TSB(s).  Note that we must do
10727 		 * this before we get a ctx or load the MMU state.  Once
10728 		 * we swap in we have to recheck to make sure the TSB(s) and
10729 		 * ISM mappings didn't change while we slept.
10730 		 */
10731 		if (SFMMU_FLAGS_ISSET(sfmmup, HAT_SWAPPED)) {
10732 			sfmmu_tsb_swapin(sfmmup, hatlockp);
10733 			goto retry;
10734 		}
10735 
10736 		sfmmu_disallow_ctx_steal(sfmmup);
10737 		ctxnum = sfmmup->sfmmu_cnum;
10738 		kpreempt_disable();
10739 		sfmmu_setctx_sec(ctxnum);
10740 		sfmmu_load_mmustate(sfmmup);
10741 		kpreempt_enable();
10742 		sfmmu_allow_ctx_steal(sfmmup);
10743 		sfmmu_hat_exit(hatlockp);
10744 		/*
10745 		 * Must restore lwp_state if not calling
10746 		 * trap() for further processing. Restore
10747 		 * it anyway.
10748 		 */
10749 		lwp->lwp_state = lwp_save_state;
10750 		if (sfmmup->sfmmu_ttecnt[TTE8K] != 0 ||
10751 		    sfmmup->sfmmu_ttecnt[TTE64K] != 0 ||
10752 		    sfmmup->sfmmu_ttecnt[TTE512K] != 0 ||
10753 		    sfmmup->sfmmu_ttecnt[TTE4M] != 0 ||
10754 		    sfmmup->sfmmu_ttecnt[TTE32M] != 0 ||
10755 		    sfmmup->sfmmu_ttecnt[TTE256M] != 0) {
10756 			return;
10757 		}
10758 		if (traptype == T_DATA_PROT) {
10759 			traptype = T_DATA_MMU_MISS;
10760 		}
10761 	}
10762 	trap(rp, (caddr_t)tagaccess, traptype, 0);
10763 }
10764 
10765 /*
10766  * sfmmu_vatopfn_suspended is called from GET_TTE when TL=0 and
10767  * TTE_SUSPENDED bit set in tte we block on aquiring a page lock
10768  * rather than spinning to avoid send mondo timeouts with
10769  * interrupts enabled. When the lock is acquired it is immediately
10770  * released and we return back to sfmmu_vatopfn just after
10771  * the GET_TTE call.
10772  */
10773 void
10774 sfmmu_vatopfn_suspended(caddr_t vaddr, sfmmu_t *sfmmu, tte_t *ttep)
10775 {
10776 	struct page	**pp;
10777 
10778 	(void) as_pagelock(sfmmu->sfmmu_as, &pp, vaddr, TTE_CSZ(ttep), S_WRITE);
10779 	as_pageunlock(sfmmu->sfmmu_as, pp, vaddr, TTE_CSZ(ttep), S_WRITE);
10780 }
10781 
10782 /*
10783  * sfmmu_tsbmiss_suspended is called from GET_TTE when TL>0 and
10784  * TTE_SUSPENDED bit set in tte. We do this so that we can handle
10785  * cross traps which cannot be handled while spinning in the
10786  * trap handlers. Simply enter and exit the kpr_suspendlock spin
10787  * mutex, which is held by the holder of the suspend bit, and then
10788  * retry the trapped instruction after unwinding.
10789  */
10790 /*ARGSUSED*/
10791 void
10792 sfmmu_tsbmiss_suspended(struct regs *rp, uintptr_t tagacc, uint_t traptype)
10793 {
10794 	ASSERT(curthread != kreloc_thread);
10795 	mutex_enter(&kpr_suspendlock);
10796 	mutex_exit(&kpr_suspendlock);
10797 }
10798 
10799 /*
10800  * Special routine to flush out ism mappings- TSBs, TLBs and D-caches.
10801  * This routine may be called with all cpu's captured. Therefore, the
10802  * caller is responsible for holding all locks and disabling kernel
10803  * preemption.
10804  */
10805 /* ARGSUSED */
10806 static void
10807 sfmmu_ismtlbcache_demap(caddr_t addr, sfmmu_t *ism_sfmmup,
10808 	struct hme_blk *hmeblkp, pfn_t pfnum, int cache_flush_flag)
10809 {
10810 	cpuset_t 	cpuset;
10811 	caddr_t 	va;
10812 	ism_ment_t	*ment;
10813 	sfmmu_t		*sfmmup;
10814 	int 		ctxnum;
10815 	int 		vcolor;
10816 	int		ttesz;
10817 
10818 	/*
10819 	 * Walk the ism_hat's mapping list and flush the page
10820 	 * from every hat sharing this ism_hat. This routine
10821 	 * may be called while all cpu's have been captured.
10822 	 * Therefore we can't attempt to grab any locks. For now
10823 	 * this means we will protect the ism mapping list under
10824 	 * a single lock which will be grabbed by the caller.
10825 	 * If hat_share/unshare scalibility becomes a performance
10826 	 * problem then we may need to re-think ism mapping list locking.
10827 	 */
10828 	ASSERT(ism_sfmmup->sfmmu_ismhat);
10829 	ASSERT(MUTEX_HELD(&ism_mlist_lock));
10830 	addr = addr - ISMID_STARTADDR;
10831 	for (ment = ism_sfmmup->sfmmu_iment; ment; ment = ment->iment_next) {
10832 
10833 		sfmmup = ment->iment_hat;
10834 		ctxnum = sfmmup->sfmmu_cnum;
10835 		va = ment->iment_base_va;
10836 		va = (caddr_t)((uintptr_t)va  + (uintptr_t)addr);
10837 
10838 		/*
10839 		 * Flush TSB of ISM mappings.
10840 		 */
10841 		ttesz = get_hblk_ttesz(hmeblkp);
10842 		if (ttesz == TTE8K || ttesz == TTE4M) {
10843 			sfmmu_unload_tsb(sfmmup, va, ttesz);
10844 		} else {
10845 			caddr_t sva = va;
10846 			caddr_t eva;
10847 			ASSERT(addr == (caddr_t)get_hblk_base(hmeblkp));
10848 			eva = sva + get_hblk_span(hmeblkp);
10849 			sfmmu_unload_tsb_range(sfmmup, sva, eva, ttesz);
10850 		}
10851 
10852 		if (ctxnum != INVALID_CONTEXT) {
10853 			/*
10854 			 * Flush TLBs.  We don't need to do this for
10855 			 * invalid context since the flushing is already
10856 			 * done as part of context stealing.
10857 			 */
10858 			cpuset = sfmmup->sfmmu_cpusran;
10859 			CPUSET_AND(cpuset, cpu_ready_set);
10860 			CPUSET_DEL(cpuset, CPU->cpu_id);
10861 			SFMMU_XCALL_STATS(ctxnum);
10862 			xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)va,
10863 			    ctxnum);
10864 			vtag_flushpage(va, ctxnum);
10865 		}
10866 
10867 		/*
10868 		 * Flush D$
10869 		 * When flushing D$ we must flush all
10870 		 * cpu's. See sfmmu_cache_flush().
10871 		 */
10872 		if (cache_flush_flag == CACHE_FLUSH) {
10873 			cpuset = cpu_ready_set;
10874 			CPUSET_DEL(cpuset, CPU->cpu_id);
10875 			SFMMU_XCALL_STATS(ctxnum);
10876 			vcolor = addr_to_vcolor(va);
10877 			xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
10878 			vac_flushpage(pfnum, vcolor);
10879 		}
10880 	}
10881 }
10882 
10883 /*
10884  * Demaps the TSB, CPU caches, and flushes all TLBs on all CPUs of
10885  * a particular virtual address and ctx.  If noflush is set we do not
10886  * flush the TLB/TSB.  This function may or may not be called with the
10887  * HAT lock held.
10888  */
10889 static void
10890 sfmmu_tlbcache_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
10891 	pfn_t pfnum, int tlb_noflush, int cpu_flag, int cache_flush_flag,
10892 	int hat_lock_held)
10893 {
10894 	int ctxnum, vcolor;
10895 	cpuset_t cpuset;
10896 	hatlock_t *hatlockp;
10897 
10898 	/*
10899 	 * There is no longer a need to protect against ctx being
10900 	 * stolen here since we don't store the ctx in the TSB anymore.
10901 	 */
10902 	vcolor = addr_to_vcolor(addr);
10903 
10904 	kpreempt_disable();
10905 	if (!tlb_noflush) {
10906 		/*
10907 		 * Flush the TSB.
10908 		 */
10909 		if (!hat_lock_held)
10910 			hatlockp = sfmmu_hat_enter(sfmmup);
10911 		SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp);
10912 		ctxnum = (int)sfmmutoctxnum(sfmmup);
10913 		if (!hat_lock_held)
10914 			sfmmu_hat_exit(hatlockp);
10915 
10916 		if (ctxnum != INVALID_CONTEXT) {
10917 			/*
10918 			 * Flush TLBs.  We don't need to do this if our
10919 			 * context is invalid context.  Since we hold the
10920 			 * HAT lock the context must have been stolen and
10921 			 * hence will be flushed before re-use.
10922 			 */
10923 			cpuset = sfmmup->sfmmu_cpusran;
10924 			CPUSET_AND(cpuset, cpu_ready_set);
10925 			CPUSET_DEL(cpuset, CPU->cpu_id);
10926 			SFMMU_XCALL_STATS(ctxnum);
10927 			xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr,
10928 				ctxnum);
10929 			vtag_flushpage(addr, ctxnum);
10930 		}
10931 	}
10932 
10933 	/*
10934 	 * Flush the D$
10935 	 *
10936 	 * Even if the ctx is stolen, we need to flush the
10937 	 * cache. Our ctx stealer only flushes the TLBs.
10938 	 */
10939 	if (cache_flush_flag == CACHE_FLUSH) {
10940 		if (cpu_flag & FLUSH_ALL_CPUS) {
10941 			cpuset = cpu_ready_set;
10942 		} else {
10943 			cpuset = sfmmup->sfmmu_cpusran;
10944 			CPUSET_AND(cpuset, cpu_ready_set);
10945 		}
10946 		CPUSET_DEL(cpuset, CPU->cpu_id);
10947 		SFMMU_XCALL_STATS(sfmmutoctxnum(sfmmup));
10948 		xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
10949 		vac_flushpage(pfnum, vcolor);
10950 	}
10951 	kpreempt_enable();
10952 }
10953 
10954 /*
10955  * Demaps the TSB and flushes all TLBs on all cpus for a particular virtual
10956  * address and ctx.  If noflush is set we do not currently do anything.
10957  * This function may or may not be called with the HAT lock held.
10958  */
10959 static void
10960 sfmmu_tlb_demap(caddr_t addr, sfmmu_t *sfmmup, struct hme_blk *hmeblkp,
10961 	int tlb_noflush, int hat_lock_held)
10962 {
10963 	int ctxnum;
10964 	cpuset_t cpuset;
10965 	hatlock_t *hatlockp;
10966 
10967 	/*
10968 	 * If the process is exiting we have nothing to do.
10969 	 */
10970 	if (tlb_noflush)
10971 		return;
10972 
10973 	/*
10974 	 * Flush TSB.
10975 	 */
10976 	if (!hat_lock_held)
10977 		hatlockp = sfmmu_hat_enter(sfmmup);
10978 	SFMMU_UNLOAD_TSB(addr, sfmmup, hmeblkp);
10979 	ctxnum = sfmmutoctxnum(sfmmup);
10980 	if (!hat_lock_held)
10981 		sfmmu_hat_exit(hatlockp);
10982 
10983 	/*
10984 	 * Flush TLBs.  We don't need to do this if our context is invalid
10985 	 * context.  Since we hold the HAT lock the context must have been
10986 	 * stolen and hence will be flushed before re-use.
10987 	 */
10988 	if (ctxnum != INVALID_CONTEXT) {
10989 		/*
10990 		 * There is no need to protect against ctx being stolen.
10991 		 * If the ctx is stolen we will simply get an extra flush.
10992 		 */
10993 		kpreempt_disable();
10994 		cpuset = sfmmup->sfmmu_cpusran;
10995 		CPUSET_AND(cpuset, cpu_ready_set);
10996 		CPUSET_DEL(cpuset, CPU->cpu_id);
10997 		SFMMU_XCALL_STATS(ctxnum);
10998 		xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)addr, ctxnum);
10999 		vtag_flushpage(addr, ctxnum);
11000 		kpreempt_enable();
11001 	}
11002 }
11003 
11004 /*
11005  * Special case of sfmmu_tlb_demap for MMU_PAGESIZE hblks. Use the xcall
11006  * call handler that can flush a range of pages to save on xcalls.
11007  */
11008 static int sfmmu_xcall_save;
11009 
11010 static void
11011 sfmmu_tlb_range_demap(demap_range_t *dmrp)
11012 {
11013 	sfmmu_t *sfmmup = dmrp->dmr_sfmmup;
11014 	int ctxnum;
11015 	hatlock_t *hatlockp;
11016 	cpuset_t cpuset;
11017 	uint64_t ctx_pgcnt;
11018 	pgcnt_t pgcnt = 0;
11019 	int pgunload = 0;
11020 	int dirtypg = 0;
11021 	caddr_t addr = dmrp->dmr_addr;
11022 	caddr_t eaddr;
11023 	uint64_t bitvec = dmrp->dmr_bitvec;
11024 
11025 	ASSERT(bitvec & 1);
11026 
11027 	/*
11028 	 * Flush TSB and calculate number of pages to flush.
11029 	 */
11030 	while (bitvec != 0) {
11031 		dirtypg = 0;
11032 		/*
11033 		 * Find the first page to flush and then count how many
11034 		 * pages there are after it that also need to be flushed.
11035 		 * This way the number of TSB flushes is minimized.
11036 		 */
11037 		while ((bitvec & 1) == 0) {
11038 			pgcnt++;
11039 			addr += MMU_PAGESIZE;
11040 			bitvec >>= 1;
11041 		}
11042 		while (bitvec & 1) {
11043 			dirtypg++;
11044 			bitvec >>= 1;
11045 		}
11046 		eaddr = addr + ptob(dirtypg);
11047 		hatlockp = sfmmu_hat_enter(sfmmup);
11048 		sfmmu_unload_tsb_range(sfmmup, addr, eaddr, TTE8K);
11049 		sfmmu_hat_exit(hatlockp);
11050 		pgunload += dirtypg;
11051 		addr = eaddr;
11052 		pgcnt += dirtypg;
11053 	}
11054 
11055 	/*
11056 	 * In the case where context is invalid context, bail.
11057 	 * We hold the hat lock while checking the ctx to prevent
11058 	 * a race with sfmmu_replace_tsb() which temporarily sets
11059 	 * the ctx to INVALID_CONTEXT to force processes to enter
11060 	 * sfmmu_tsbmiss_exception().
11061 	 */
11062 	hatlockp = sfmmu_hat_enter(sfmmup);
11063 	ctxnum = sfmmutoctxnum(sfmmup);
11064 	sfmmu_hat_exit(hatlockp);
11065 	if (ctxnum == INVALID_CONTEXT) {
11066 		dmrp->dmr_bitvec = 0;
11067 		return;
11068 	}
11069 
11070 	ASSERT((pgcnt<<MMU_PAGESHIFT) <= dmrp->dmr_endaddr - dmrp->dmr_addr);
11071 	if (sfmmup->sfmmu_free == 0) {
11072 		addr = dmrp->dmr_addr;
11073 		bitvec = dmrp->dmr_bitvec;
11074 		ctx_pgcnt = (uint64_t)((ctxnum << 16) | pgcnt);
11075 		kpreempt_disable();
11076 		cpuset = sfmmup->sfmmu_cpusran;
11077 		CPUSET_AND(cpuset, cpu_ready_set);
11078 		CPUSET_DEL(cpuset, CPU->cpu_id);
11079 		SFMMU_XCALL_STATS(ctxnum);
11080 		xt_some(cpuset, vtag_flush_pgcnt_tl1, (uint64_t)addr,
11081 			ctx_pgcnt);
11082 		for (; bitvec != 0; bitvec >>= 1) {
11083 			if (bitvec & 1)
11084 				vtag_flushpage(addr, ctxnum);
11085 			addr += MMU_PAGESIZE;
11086 		}
11087 		kpreempt_enable();
11088 		sfmmu_xcall_save += (pgunload-1);
11089 	}
11090 	dmrp->dmr_bitvec = 0;
11091 }
11092 
11093 /*
11094  * Flushes only TLB.
11095  */
11096 static void
11097 sfmmu_tlb_ctx_demap(sfmmu_t *sfmmup)
11098 {
11099 	int ctxnum;
11100 	cpuset_t cpuset;
11101 
11102 	ctxnum = (int)sfmmutoctxnum(sfmmup);
11103 	if (ctxnum == INVALID_CONTEXT) {
11104 		/*
11105 		 * if ctx was stolen then simply return
11106 		 * whoever stole ctx is responsible for flush.
11107 		 */
11108 		return;
11109 	}
11110 	ASSERT(ctxnum != KCONTEXT);
11111 	/*
11112 	 * There is no need to protect against ctx being stolen.  If the
11113 	 * ctx is stolen we will simply get an extra flush.
11114 	 */
11115 	kpreempt_disable();
11116 
11117 	cpuset = sfmmup->sfmmu_cpusran;
11118 	CPUSET_DEL(cpuset, CPU->cpu_id);
11119 	CPUSET_AND(cpuset, cpu_ready_set);
11120 	SFMMU_XCALL_STATS(ctxnum);
11121 
11122 	/*
11123 	 * Flush TLB.
11124 	 * RFE: it might be worth delaying the TLB flush as well. In that
11125 	 * case each cpu would have to traverse the dirty list and flush
11126 	 * each one of those ctx from the TLB.
11127 	 */
11128 	vtag_flushctx(ctxnum);
11129 	xt_some(cpuset, vtag_flushctx_tl1, ctxnum, 0);
11130 
11131 	kpreempt_enable();
11132 	SFMMU_STAT(sf_tlbflush_ctx);
11133 }
11134 
11135 /*
11136  * Flushes all TLBs.
11137  */
11138 static void
11139 sfmmu_tlb_all_demap(void)
11140 {
11141 	cpuset_t cpuset;
11142 
11143 	/*
11144 	 * There is no need to protect against ctx being stolen.  If the
11145 	 * ctx is stolen we will simply get an extra flush.
11146 	 */
11147 	kpreempt_disable();
11148 
11149 	cpuset = cpu_ready_set;
11150 	CPUSET_DEL(cpuset, CPU->cpu_id);
11151 	/* LINTED: constant in conditional context */
11152 	SFMMU_XCALL_STATS(INVALID_CONTEXT);
11153 
11154 	vtag_flushall();
11155 	xt_some(cpuset, vtag_flushall_tl1, 0, 0);
11156 	xt_sync(cpuset);
11157 
11158 	kpreempt_enable();
11159 	SFMMU_STAT(sf_tlbflush_all);
11160 }
11161 
11162 /*
11163  * In cases where we need to synchronize with TLB/TSB miss trap
11164  * handlers, _and_ need to flush the TLB, it's a lot easier to
11165  * steal the context from the process and free it than to do a
11166  * special song and dance to keep things consistent for the
11167  * handlers.
11168  *
11169  * Since the process suddenly ends up without a context and our caller
11170  * holds the hat lock, threads that fault after this function is called
11171  * will pile up on the lock.  We can then do whatever we need to
11172  * atomically from the context of the caller.  The first blocked thread
11173  * to resume executing will get the process a new context, and the
11174  * process will resume executing.
11175  *
11176  * One added advantage of this approach is that on MMUs that
11177  * support a "flush all" operation, we will delay the flush until
11178  * we run out of contexts, and then flush the TLB one time.  This
11179  * is rather rare, so it's a lot less expensive than making 8000
11180  * x-calls to flush the TLB 8000 times.  Another is that we can do
11181  * all of this without pausing CPUs, due to some knowledge of how
11182  * resume() loads processes onto the processor; it sets the thread
11183  * into cpusran, and _then_ looks at cnum.  Because we do things in
11184  * the reverse order here, we guarantee exactly one of the following
11185  * statements is always true:
11186  *
11187  *   1) Nobody is in resume() so we have nothing to worry about anyway.
11188  *   2) The thread in resume() isn't in cpusran when we do the xcall,
11189  *      so we know when it does set itself it'll see cnum is
11190  *      INVALID_CONTEXT.
11191  *   3) The thread in resume() is in cpusran, and already might have
11192  *      looked at the old cnum.  That's OK, because we'll xcall it
11193  *      and, if necessary, flush the TLB along with the rest of the
11194  *      crowd.
11195  */
11196 static void
11197 sfmmu_tlb_swap_ctx(sfmmu_t *sfmmup, struct ctx *ctx)
11198 {
11199 	cpuset_t cpuset;
11200 	int cnum;
11201 
11202 	if (sfmmup->sfmmu_cnum == INVALID_CONTEXT)
11203 		return;
11204 
11205 	SFMMU_STAT(sf_ctx_swap);
11206 
11207 	kpreempt_disable();
11208 
11209 	ASSERT(rw_read_locked(&ctx->ctx_rwlock) == 0);
11210 	ASSERT(ctx->ctx_sfmmu == sfmmup);
11211 
11212 	cnum = ctxtoctxnum(ctx);
11213 	ASSERT(sfmmup->sfmmu_cnum == cnum);
11214 	ASSERT(cnum >= NUM_LOCKED_CTXS);
11215 
11216 	sfmmup->sfmmu_cnum = INVALID_CONTEXT;
11217 	membar_enter();	/* make sure visible on all CPUs */
11218 	ctx->ctx_sfmmu = NULL;
11219 
11220 	cpuset = sfmmup->sfmmu_cpusran;
11221 	CPUSET_DEL(cpuset, CPU->cpu_id);
11222 	CPUSET_AND(cpuset, cpu_ready_set);
11223 	SFMMU_XCALL_STATS(cnum);
11224 
11225 	/*
11226 	 * Force anybody running this process on CPU
11227 	 * to enter sfmmu_tsbmiss_exception() on the
11228 	 * next TLB miss, synchronize behind us on
11229 	 * the HAT lock, and grab a new context.  At
11230 	 * that point the new page size will become
11231 	 * active in the TLB for the new context.
11232 	 * See sfmmu_get_ctx() for details.
11233 	 */
11234 	if (delay_tlb_flush) {
11235 		xt_some(cpuset, sfmmu_raise_tsb_exception,
11236 		    cnum, INVALID_CONTEXT);
11237 		SFMMU_STAT(sf_tlbflush_deferred);
11238 	} else {
11239 		xt_some(cpuset, sfmmu_ctx_steal_tl1, cnum, INVALID_CONTEXT);
11240 		vtag_flushctx(cnum);
11241 		SFMMU_STAT(sf_tlbflush_ctx);
11242 	}
11243 	xt_sync(cpuset);
11244 
11245 	/*
11246 	 * If we just stole the ctx from the current
11247 	 * process on local CPU we need to invalidate
11248 	 * this CPU context as well.
11249 	 */
11250 	if (sfmmu_getctx_sec() == cnum) {
11251 		sfmmu_setctx_sec(INVALID_CONTEXT);
11252 		sfmmu_clear_utsbinfo();
11253 	}
11254 
11255 	kpreempt_enable();
11256 
11257 	/*
11258 	 * Now put old ctx on the dirty list since we may not
11259 	 * have flushed the context out of the TLB.  We'll let
11260 	 * the next guy who uses this ctx flush it instead.
11261 	 */
11262 	mutex_enter(&ctx_list_lock);
11263 	CTX_SET_FLAGS(ctx, CTX_FREE_FLAG);
11264 	ctx->ctx_free = ctxdirty;
11265 	ctxdirty = ctx;
11266 	mutex_exit(&ctx_list_lock);
11267 }
11268 
11269 /*
11270  * We need to flush the cache in all cpus.  It is possible that
11271  * a process referenced a page as cacheable but has sinced exited
11272  * and cleared the mapping list.  We still to flush it but have no
11273  * state so all cpus is the only alternative.
11274  */
11275 void
11276 sfmmu_cache_flush(pfn_t pfnum, int vcolor)
11277 {
11278 	cpuset_t cpuset;
11279 	int	ctxnum = INVALID_CONTEXT;
11280 
11281 	kpreempt_disable();
11282 	cpuset = cpu_ready_set;
11283 	CPUSET_DEL(cpuset, CPU->cpu_id);
11284 	SFMMU_XCALL_STATS(ctxnum);	/* account to any ctx */
11285 	xt_some(cpuset, vac_flushpage_tl1, pfnum, vcolor);
11286 	xt_sync(cpuset);
11287 	vac_flushpage(pfnum, vcolor);
11288 	kpreempt_enable();
11289 }
11290 
11291 void
11292 sfmmu_cache_flushcolor(int vcolor, pfn_t pfnum)
11293 {
11294 	cpuset_t cpuset;
11295 	int	ctxnum = INVALID_CONTEXT;
11296 
11297 	ASSERT(vcolor >= 0);
11298 
11299 	kpreempt_disable();
11300 	cpuset = cpu_ready_set;
11301 	CPUSET_DEL(cpuset, CPU->cpu_id);
11302 	SFMMU_XCALL_STATS(ctxnum);	/* account to any ctx */
11303 	xt_some(cpuset, vac_flushcolor_tl1, vcolor, pfnum);
11304 	xt_sync(cpuset);
11305 	vac_flushcolor(vcolor, pfnum);
11306 	kpreempt_enable();
11307 }
11308 
11309 /*
11310  * We need to prevent processes from accessing the TSB using a cached physical
11311  * address.  It's alright if they try to access the TSB via virtual address
11312  * since they will just fault on that virtual address once the mapping has
11313  * been suspended.
11314  */
11315 #pragma weak sendmondo_in_recover
11316 
11317 /* ARGSUSED */
11318 static int
11319 sfmmu_tsb_pre_relocator(caddr_t va, uint_t tsbsz, uint_t flags, void *tsbinfo)
11320 {
11321 	hatlock_t *hatlockp;
11322 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
11323 	sfmmu_t *sfmmup = tsbinfop->tsb_sfmmu;
11324 	struct ctx *ctx;
11325 	int cnum;
11326 	extern uint32_t sendmondo_in_recover;
11327 
11328 	if (flags != HAT_PRESUSPEND)
11329 		return (0);
11330 
11331 	hatlockp = sfmmu_hat_enter(sfmmup);
11332 
11333 	tsbinfop->tsb_flags |= TSB_RELOC_FLAG;
11334 
11335 	/*
11336 	 * For Cheetah+ Erratum 25:
11337 	 * Wait for any active recovery to finish.  We can't risk
11338 	 * relocating the TSB of the thread running mondo_recover_proc()
11339 	 * since, if we did that, we would deadlock.  The scenario we are
11340 	 * trying to avoid is as follows:
11341 	 *
11342 	 * THIS CPU			RECOVER CPU
11343 	 * --------			-----------
11344 	 *				Begins recovery, walking through TSB
11345 	 * hat_pagesuspend() TSB TTE
11346 	 *				TLB miss on TSB TTE, spins at TL1
11347 	 * xt_sync()
11348 	 *	send_mondo_timeout()
11349 	 *	mondo_recover_proc()
11350 	 *	((deadlocked))
11351 	 *
11352 	 * The second half of the workaround is that mondo_recover_proc()
11353 	 * checks to see if the tsb_info has the RELOC flag set, and if it
11354 	 * does, it skips over that TSB without ever touching tsbinfop->tsb_va
11355 	 * and hence avoiding the TLB miss that could result in a deadlock.
11356 	 */
11357 	if (&sendmondo_in_recover) {
11358 		membar_enter();	/* make sure RELOC flag visible */
11359 		while (sendmondo_in_recover) {
11360 			drv_usecwait(1);
11361 			membar_consumer();
11362 		}
11363 	}
11364 
11365 	ctx = sfmmutoctx(sfmmup);
11366 	rw_enter(&ctx->ctx_rwlock, RW_WRITER);
11367 	cnum = sfmmutoctxnum(sfmmup);
11368 
11369 	if (cnum != INVALID_CONTEXT) {
11370 		/*
11371 		 * Force all threads for this sfmmu to sfmmu_tsbmiss_exception
11372 		 * on their next TLB miss.
11373 		 */
11374 		sfmmu_tlb_swap_ctx(sfmmup, ctx);
11375 	}
11376 
11377 	rw_exit(&ctx->ctx_rwlock);
11378 
11379 	sfmmu_hat_exit(hatlockp);
11380 
11381 	return (0);
11382 }
11383 
11384 /* ARGSUSED */
11385 static int
11386 sfmmu_tsb_post_relocator(caddr_t va, uint_t tsbsz, uint_t flags,
11387 	void *tsbinfo, pfn_t newpfn)
11388 {
11389 	hatlock_t *hatlockp;
11390 	struct tsb_info *tsbinfop = (struct tsb_info *)tsbinfo;
11391 	sfmmu_t	*sfmmup = tsbinfop->tsb_sfmmu;
11392 
11393 	if (flags != HAT_POSTUNSUSPEND)
11394 		return (0);
11395 
11396 	hatlockp = sfmmu_hat_enter(sfmmup);
11397 
11398 	SFMMU_STAT(sf_tsb_reloc);
11399 
11400 	/*
11401 	 * The process may have swapped out while we were relocating one
11402 	 * of its TSBs.  If so, don't bother doing the setup since the
11403 	 * process can't be using the memory anymore.
11404 	 */
11405 	if ((tsbinfop->tsb_flags & TSB_SWAPPED) == 0) {
11406 		ASSERT(va == tsbinfop->tsb_va);
11407 		sfmmu_tsbinfo_setup_phys(tsbinfop, newpfn);
11408 		sfmmu_setup_tsbinfo(sfmmup);
11409 
11410 		if (tsbinfop->tsb_flags & TSB_FLUSH_NEEDED) {
11411 			sfmmu_inv_tsb(tsbinfop->tsb_va,
11412 			    TSB_BYTES(tsbinfop->tsb_szc));
11413 			tsbinfop->tsb_flags &= ~TSB_FLUSH_NEEDED;
11414 		}
11415 	}
11416 
11417 	membar_exit();
11418 	tsbinfop->tsb_flags &= ~TSB_RELOC_FLAG;
11419 	cv_broadcast(&sfmmup->sfmmu_tsb_cv);
11420 
11421 	sfmmu_hat_exit(hatlockp);
11422 
11423 	return (0);
11424 }
11425 
11426 /*
11427  * Allocate and initialize a tsb_info structure.  Note that we may or may not
11428  * allocate a TSB here, depending on the flags passed in.
11429  */
11430 static int
11431 sfmmu_tsbinfo_alloc(struct tsb_info **tsbinfopp, int tsb_szc, int tte_sz_mask,
11432 	uint_t flags, sfmmu_t *sfmmup)
11433 {
11434 	int err;
11435 
11436 	*tsbinfopp = (struct tsb_info *)kmem_cache_alloc(
11437 	    sfmmu_tsbinfo_cache, KM_SLEEP);
11438 
11439 	if ((err = sfmmu_init_tsbinfo(*tsbinfopp, tte_sz_mask,
11440 	    tsb_szc, flags, sfmmup)) != 0) {
11441 		kmem_cache_free(sfmmu_tsbinfo_cache, *tsbinfopp);
11442 		SFMMU_STAT(sf_tsb_allocfail);
11443 		*tsbinfopp = NULL;
11444 		return (err);
11445 	}
11446 	SFMMU_STAT(sf_tsb_alloc);
11447 
11448 	/*
11449 	 * Bump the TSB size counters for this TSB size.
11450 	 */
11451 	(*(((int *)&sfmmu_tsbsize_stat) + tsb_szc))++;
11452 	return (0);
11453 }
11454 
11455 static void
11456 sfmmu_tsb_free(struct tsb_info *tsbinfo)
11457 {
11458 	caddr_t tsbva = tsbinfo->tsb_va;
11459 	uint_t tsb_size = TSB_BYTES(tsbinfo->tsb_szc);
11460 	struct kmem_cache *kmem_cachep = tsbinfo->tsb_cache;
11461 	vmem_t	*vmp = tsbinfo->tsb_vmp;
11462 
11463 	/*
11464 	 * If we allocated this TSB from relocatable kernel memory, then we
11465 	 * need to uninstall the callback handler.
11466 	 */
11467 	if (tsbinfo->tsb_cache != sfmmu_tsb8k_cache) {
11468 		uintptr_t slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
11469 		caddr_t slab_vaddr = (caddr_t)((uintptr_t)tsbva & slab_mask);
11470 		page_t **ppl;
11471 		int ret;
11472 
11473 		ret = as_pagelock(&kas, &ppl, slab_vaddr, PAGESIZE, S_WRITE);
11474 		ASSERT(ret == 0);
11475 		hat_delete_callback(tsbva, (uint_t)tsb_size, (void *)tsbinfo,
11476 		    0);
11477 		as_pageunlock(&kas, ppl, slab_vaddr, PAGESIZE, S_WRITE);
11478 	}
11479 
11480 	if (kmem_cachep != NULL) {
11481 		kmem_cache_free(kmem_cachep, tsbva);
11482 	} else {
11483 		vmem_xfree(vmp, (void *)tsbva, tsb_size);
11484 	}
11485 	tsbinfo->tsb_va = (caddr_t)0xbad00bad;
11486 	atomic_add_64(&tsb_alloc_bytes, -(int64_t)tsb_size);
11487 }
11488 
11489 static void
11490 sfmmu_tsbinfo_free(struct tsb_info *tsbinfo)
11491 {
11492 	if ((tsbinfo->tsb_flags & TSB_SWAPPED) == 0) {
11493 		sfmmu_tsb_free(tsbinfo);
11494 	}
11495 	kmem_cache_free(sfmmu_tsbinfo_cache, tsbinfo);
11496 
11497 }
11498 
11499 /*
11500  * Setup all the references to physical memory for this tsbinfo.
11501  * The underlying page(s) must be locked.
11502  */
11503 static void
11504 sfmmu_tsbinfo_setup_phys(struct tsb_info *tsbinfo, pfn_t pfn)
11505 {
11506 	ASSERT(pfn != PFN_INVALID);
11507 	ASSERT(pfn == va_to_pfn(tsbinfo->tsb_va));
11508 
11509 #ifndef sun4v
11510 	if (tsbinfo->tsb_szc == 0) {
11511 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn,
11512 		    PROT_WRITE|PROT_READ, TTE8K);
11513 	} else {
11514 		/*
11515 		 * Round down PA and use a large mapping; the handlers will
11516 		 * compute the TSB pointer at the correct offset into the
11517 		 * big virtual page.  NOTE: this assumes all TSBs larger
11518 		 * than 8K must come from physically contiguous slabs of
11519 		 * size tsb_slab_size.
11520 		 */
11521 		sfmmu_memtte(&tsbinfo->tsb_tte, pfn & ~tsb_slab_mask,
11522 		    PROT_WRITE|PROT_READ, tsb_slab_ttesz);
11523 	}
11524 	tsbinfo->tsb_pa = ptob(pfn);
11525 
11526 	TTE_SET_LOCKED(&tsbinfo->tsb_tte); /* lock the tte into dtlb */
11527 	TTE_SET_MOD(&tsbinfo->tsb_tte);    /* enable writes */
11528 
11529 	ASSERT(TTE_IS_PRIVILEGED(&tsbinfo->tsb_tte));
11530 	ASSERT(TTE_IS_LOCKED(&tsbinfo->tsb_tte));
11531 #else /* sun4v */
11532 	tsbinfo->tsb_pa = ptob(pfn);
11533 #endif /* sun4v */
11534 }
11535 
11536 
11537 /*
11538  * Returns zero on success, ENOMEM if over the high water mark,
11539  * or EAGAIN if the caller needs to retry with a smaller TSB
11540  * size (or specify TSB_FORCEALLOC if the allocation can't fail).
11541  *
11542  * This call cannot fail to allocate a TSB if TSB_FORCEALLOC
11543  * is specified and the TSB requested is PAGESIZE, though it
11544  * may sleep waiting for memory if sufficient memory is not
11545  * available.
11546  */
11547 static int
11548 sfmmu_init_tsbinfo(struct tsb_info *tsbinfo, int tteszmask,
11549     int tsbcode, uint_t flags, sfmmu_t *sfmmup)
11550 {
11551 	caddr_t vaddr = NULL;
11552 	caddr_t slab_vaddr;
11553 	uintptr_t slab_mask = ~((uintptr_t)tsb_slab_mask) << PAGESHIFT;
11554 	int tsbbytes = TSB_BYTES(tsbcode);
11555 	int lowmem = 0;
11556 	struct kmem_cache *kmem_cachep = NULL;
11557 	vmem_t *vmp = NULL;
11558 	lgrp_id_t lgrpid = LGRP_NONE;
11559 	pfn_t pfn;
11560 	uint_t cbflags = HAC_SLEEP;
11561 	page_t **pplist;
11562 	int ret;
11563 
11564 	if (flags & (TSB_FORCEALLOC | TSB_SWAPIN | TSB_GROW | TSB_SHRINK))
11565 		flags |= TSB_ALLOC;
11566 
11567 	ASSERT((flags & TSB_FORCEALLOC) == 0 || tsbcode == TSB_MIN_SZCODE);
11568 
11569 	tsbinfo->tsb_sfmmu = sfmmup;
11570 
11571 	/*
11572 	 * If not allocating a TSB, set up the tsbinfo, set TSB_SWAPPED, and
11573 	 * return.
11574 	 */
11575 	if ((flags & TSB_ALLOC) == 0) {
11576 		tsbinfo->tsb_szc = tsbcode;
11577 		tsbinfo->tsb_ttesz_mask = tteszmask;
11578 		tsbinfo->tsb_va = (caddr_t)0xbadbadbeef;
11579 		tsbinfo->tsb_pa = -1;
11580 		tsbinfo->tsb_tte.ll = 0;
11581 		tsbinfo->tsb_next = NULL;
11582 		tsbinfo->tsb_flags = TSB_SWAPPED;
11583 		tsbinfo->tsb_cache = NULL;
11584 		tsbinfo->tsb_vmp = NULL;
11585 		return (0);
11586 	}
11587 
11588 #ifdef DEBUG
11589 	/*
11590 	 * For debugging:
11591 	 * Randomly force allocation failures every tsb_alloc_mtbf
11592 	 * tries if TSB_FORCEALLOC is not specified.  This will
11593 	 * return ENOMEM if tsb_alloc_mtbf is odd, or EAGAIN if
11594 	 * it is even, to allow testing of both failure paths...
11595 	 */
11596 	if (tsb_alloc_mtbf && ((flags & TSB_FORCEALLOC) == 0) &&
11597 	    (tsb_alloc_count++ == tsb_alloc_mtbf)) {
11598 		tsb_alloc_count = 0;
11599 		tsb_alloc_fail_mtbf++;
11600 		return ((tsb_alloc_mtbf & 1)? ENOMEM : EAGAIN);
11601 	}
11602 #endif	/* DEBUG */
11603 
11604 	/*
11605 	 * Enforce high water mark if we are not doing a forced allocation
11606 	 * and are not shrinking a process' TSB.
11607 	 */
11608 	if ((flags & TSB_SHRINK) == 0 &&
11609 	    (tsbbytes + tsb_alloc_bytes) > tsb_alloc_hiwater) {
11610 		if ((flags & TSB_FORCEALLOC) == 0)
11611 			return (ENOMEM);
11612 		lowmem = 1;
11613 	}
11614 
11615 	/*
11616 	 * Allocate from the correct location based upon the size of the TSB
11617 	 * compared to the base page size, and what memory conditions dictate.
11618 	 * Note we always do nonblocking allocations from the TSB arena since
11619 	 * we don't want memory fragmentation to cause processes to block
11620 	 * indefinitely waiting for memory; until the kernel algorithms that
11621 	 * coalesce large pages are improved this is our best option.
11622 	 *
11623 	 * Algorithm:
11624 	 *	If allocating a "large" TSB (>8K), allocate from the
11625 	 *		appropriate kmem_tsb_default_arena vmem arena
11626 	 *	else if low on memory or the TSB_FORCEALLOC flag is set or
11627 	 *	tsb_forceheap is set
11628 	 *		Allocate from kernel heap via sfmmu_tsb8k_cache with
11629 	 *		KM_SLEEP (never fails)
11630 	 *	else
11631 	 *		Allocate from appropriate sfmmu_tsb_cache with
11632 	 *		KM_NOSLEEP
11633 	 *	endif
11634 	 */
11635 	if (tsb_lgrp_affinity)
11636 		lgrpid = lgrp_home_id(curthread);
11637 	if (lgrpid == LGRP_NONE)
11638 		lgrpid = 0;	/* use lgrp of boot CPU */
11639 
11640 	if (tsbbytes > MMU_PAGESIZE) {
11641 		vmp = kmem_tsb_default_arena[lgrpid];
11642 		vaddr = (caddr_t)vmem_xalloc(vmp, tsbbytes, tsbbytes, 0, 0,
11643 		    NULL, NULL, VM_NOSLEEP);
11644 #ifdef	DEBUG
11645 	} else if (lowmem || (flags & TSB_FORCEALLOC) || tsb_forceheap) {
11646 #else	/* !DEBUG */
11647 	} else if (lowmem || (flags & TSB_FORCEALLOC)) {
11648 #endif	/* DEBUG */
11649 		kmem_cachep = sfmmu_tsb8k_cache;
11650 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_SLEEP);
11651 		ASSERT(vaddr != NULL);
11652 	} else {
11653 		kmem_cachep = sfmmu_tsb_cache[lgrpid];
11654 		vaddr = (caddr_t)kmem_cache_alloc(kmem_cachep, KM_NOSLEEP);
11655 	}
11656 
11657 	tsbinfo->tsb_cache = kmem_cachep;
11658 	tsbinfo->tsb_vmp = vmp;
11659 
11660 	if (vaddr == NULL) {
11661 		return (EAGAIN);
11662 	}
11663 
11664 	atomic_add_64(&tsb_alloc_bytes, (int64_t)tsbbytes);
11665 	kmem_cachep = tsbinfo->tsb_cache;
11666 
11667 	/*
11668 	 * If we are allocating from outside the cage, then we need to
11669 	 * register a relocation callback handler.  Note that for now
11670 	 * since pseudo mappings always hang off of the slab's root page,
11671 	 * we need only lock the first 8K of the TSB slab.  This is a bit
11672 	 * hacky but it is good for performance.
11673 	 */
11674 	if (kmem_cachep != sfmmu_tsb8k_cache) {
11675 		slab_vaddr = (caddr_t)((uintptr_t)vaddr & slab_mask);
11676 		ret = as_pagelock(&kas, &pplist, slab_vaddr, PAGESIZE, S_WRITE);
11677 		ASSERT(ret == 0);
11678 		ret = hat_add_callback(sfmmu_tsb_cb_id, vaddr, (uint_t)tsbbytes,
11679 		    cbflags, (void *)tsbinfo, &pfn);
11680 
11681 		/*
11682 		 * Need to free up resources if we could not successfully
11683 		 * add the callback function and return an error condition.
11684 		 */
11685 		if (ret != 0) {
11686 			if (kmem_cachep) {
11687 				kmem_cache_free(kmem_cachep, vaddr);
11688 			} else {
11689 				vmem_xfree(vmp, (void *)vaddr, tsbbytes);
11690 			}
11691 			as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE,
11692 			    S_WRITE);
11693 			return (EAGAIN);
11694 		}
11695 	} else {
11696 		/*
11697 		 * Since allocation of 8K TSBs from heap is rare and occurs
11698 		 * during memory pressure we allocate them from permanent
11699 		 * memory rather than using callbacks to get the PFN.
11700 		 */
11701 		pfn = hat_getpfnum(kas.a_hat, vaddr);
11702 	}
11703 
11704 	tsbinfo->tsb_va = vaddr;
11705 	tsbinfo->tsb_szc = tsbcode;
11706 	tsbinfo->tsb_ttesz_mask = tteszmask;
11707 	tsbinfo->tsb_next = NULL;
11708 	tsbinfo->tsb_flags = 0;
11709 
11710 	sfmmu_tsbinfo_setup_phys(tsbinfo, pfn);
11711 
11712 	if (kmem_cachep != sfmmu_tsb8k_cache) {
11713 		as_pageunlock(&kas, pplist, slab_vaddr, PAGESIZE, S_WRITE);
11714 	}
11715 
11716 	sfmmu_inv_tsb(vaddr, tsbbytes);
11717 	return (0);
11718 }
11719 
11720 /*
11721  * Initialize per cpu tsb and per cpu tsbmiss_area
11722  */
11723 void
11724 sfmmu_init_tsbs(void)
11725 {
11726 	int i;
11727 	struct tsbmiss	*tsbmissp;
11728 	struct kpmtsbm	*kpmtsbmp;
11729 #ifndef sun4v
11730 	extern int	dcache_line_mask;
11731 #endif /* sun4v */
11732 	extern uint_t	vac_colors;
11733 
11734 	/*
11735 	 * Init. tsb miss area.
11736 	 */
11737 	tsbmissp = tsbmiss_area;
11738 
11739 	for (i = 0; i < NCPU; tsbmissp++, i++) {
11740 		/*
11741 		 * initialize the tsbmiss area.
11742 		 * Do this for all possible CPUs as some may be added
11743 		 * while the system is running. There is no cost to this.
11744 		 */
11745 		tsbmissp->ksfmmup = ksfmmup;
11746 #ifndef sun4v
11747 		tsbmissp->dcache_line_mask = (uint16_t)dcache_line_mask;
11748 #endif /* sun4v */
11749 		tsbmissp->khashstart =
11750 		    (struct hmehash_bucket *)va_to_pa((caddr_t)khme_hash);
11751 		tsbmissp->uhashstart =
11752 		    (struct hmehash_bucket *)va_to_pa((caddr_t)uhme_hash);
11753 		tsbmissp->khashsz = khmehash_num;
11754 		tsbmissp->uhashsz = uhmehash_num;
11755 	}
11756 
11757 	if (kpm_enable == 0)
11758 		return;
11759 
11760 	if (kpm_smallpages) {
11761 		/*
11762 		 * If we're using base pagesize pages for seg_kpm
11763 		 * mappings, we use the kernel TSB since we can't afford
11764 		 * to allocate a second huge TSB for these mappings.
11765 		 */
11766 		kpm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
11767 		kpm_tsbsz = ktsb_szcode;
11768 		kpmsm_tsbbase = kpm_tsbbase;
11769 		kpmsm_tsbsz = kpm_tsbsz;
11770 	} else {
11771 		/*
11772 		 * In VAC conflict case, just put the entries in the
11773 		 * kernel 8K indexed TSB for now so we can find them.
11774 		 * This could really be changed in the future if we feel
11775 		 * the need...
11776 		 */
11777 		kpmsm_tsbbase = ktsb_phys? ktsb_pbase : (uint64_t)ktsb_base;
11778 		kpmsm_tsbsz = ktsb_szcode;
11779 		kpm_tsbbase = ktsb_phys? ktsb4m_pbase : (uint64_t)ktsb4m_base;
11780 		kpm_tsbsz = ktsb4m_szcode;
11781 	}
11782 
11783 	kpmtsbmp = kpmtsbm_area;
11784 	for (i = 0; i < NCPU; kpmtsbmp++, i++) {
11785 		/*
11786 		 * Initialize the kpmtsbm area.
11787 		 * Do this for all possible CPUs as some may be added
11788 		 * while the system is running. There is no cost to this.
11789 		 */
11790 		kpmtsbmp->vbase = kpm_vbase;
11791 		kpmtsbmp->vend = kpm_vbase + kpm_size * vac_colors;
11792 		kpmtsbmp->sz_shift = kpm_size_shift;
11793 		kpmtsbmp->kpmp_shift = kpmp_shift;
11794 		kpmtsbmp->kpmp2pshft = (uchar_t)kpmp2pshft;
11795 		if (kpm_smallpages == 0) {
11796 			kpmtsbmp->kpmp_table_sz = kpmp_table_sz;
11797 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_table);
11798 		} else {
11799 			kpmtsbmp->kpmp_table_sz = kpmp_stable_sz;
11800 			kpmtsbmp->kpmp_tablepa = va_to_pa(kpmp_stable);
11801 		}
11802 		kpmtsbmp->msegphashpa = va_to_pa(memseg_phash);
11803 		kpmtsbmp->flags = KPMTSBM_ENABLE_FLAG;
11804 #ifdef	DEBUG
11805 		kpmtsbmp->flags |= (kpm_tsbmtl) ?  KPMTSBM_TLTSBM_FLAG : 0;
11806 #endif	/* DEBUG */
11807 		if (ktsb_phys)
11808 			kpmtsbmp->flags |= KPMTSBM_TSBPHYS_FLAG;
11809 	}
11810 
11811 	sfmmu_tsb_cb_id = hat_register_callback(sfmmu_tsb_pre_relocator,
11812 	    sfmmu_tsb_post_relocator, NULL, 0);
11813 }
11814 
11815 /* Avoid using sfmmu_tsbinfo_alloc() to avoid kmem_alloc - no real reason */
11816 struct tsb_info ktsb_info[2];
11817 
11818 /*
11819  * Called from hat_kern_setup() to setup the tsb_info for ksfmmup.
11820  */
11821 void
11822 sfmmu_init_ktsbinfo()
11823 {
11824 	ASSERT(ksfmmup != NULL);
11825 	ASSERT(ksfmmup->sfmmu_tsb == NULL);
11826 	/*
11827 	 * Allocate tsbinfos for kernel and copy in data
11828 	 * to make debug easier and sun4v setup easier.
11829 	 */
11830 	ktsb_info[0].tsb_sfmmu = ksfmmup;
11831 	ktsb_info[0].tsb_szc = ktsb_szcode;
11832 	ktsb_info[0].tsb_ttesz_mask = TSB8K|TSB64K|TSB512K;
11833 	ktsb_info[0].tsb_va = ktsb_base;
11834 	ktsb_info[0].tsb_pa = ktsb_pbase;
11835 	ktsb_info[0].tsb_flags = 0;
11836 	ktsb_info[0].tsb_tte.ll = 0;
11837 	ktsb_info[0].tsb_cache = NULL;
11838 
11839 	ktsb_info[1].tsb_sfmmu = ksfmmup;
11840 	ktsb_info[1].tsb_szc = ktsb4m_szcode;
11841 	ktsb_info[1].tsb_ttesz_mask = TSB4M;
11842 	ktsb_info[1].tsb_va = ktsb4m_base;
11843 	ktsb_info[1].tsb_pa = ktsb4m_pbase;
11844 	ktsb_info[1].tsb_flags = 0;
11845 	ktsb_info[1].tsb_tte.ll = 0;
11846 	ktsb_info[1].tsb_cache = NULL;
11847 
11848 	/* Link them into ksfmmup. */
11849 	ktsb_info[0].tsb_next = &ktsb_info[1];
11850 	ktsb_info[1].tsb_next = NULL;
11851 	ksfmmup->sfmmu_tsb = &ktsb_info[0];
11852 
11853 	sfmmu_setup_tsbinfo(ksfmmup);
11854 }
11855 
11856 /*
11857  * Cache the last value returned from va_to_pa().  If the VA specified
11858  * in the current call to cached_va_to_pa() maps to the same Page (as the
11859  * previous call to cached_va_to_pa()), then compute the PA using
11860  * cached info, else call va_to_pa().
11861  *
11862  * Note: this function is neither MT-safe nor consistent in the presence
11863  * of multiple, interleaved threads.  This function was created to enable
11864  * an optimization used during boot (at a point when there's only one thread
11865  * executing on the "boot CPU", and before startup_vm() has been called).
11866  */
11867 static uint64_t
11868 cached_va_to_pa(void *vaddr)
11869 {
11870 	static uint64_t prev_vaddr_base = 0;
11871 	static uint64_t prev_pfn = 0;
11872 
11873 	if ((((uint64_t)vaddr) & MMU_PAGEMASK) == prev_vaddr_base) {
11874 		return (prev_pfn | ((uint64_t)vaddr & MMU_PAGEOFFSET));
11875 	} else {
11876 		uint64_t pa = va_to_pa(vaddr);
11877 
11878 		if (pa != ((uint64_t)-1)) {
11879 			/*
11880 			 * Computed physical address is valid.  Cache its
11881 			 * related info for the next cached_va_to_pa() call.
11882 			 */
11883 			prev_pfn = pa & MMU_PAGEMASK;
11884 			prev_vaddr_base = ((uint64_t)vaddr) & MMU_PAGEMASK;
11885 		}
11886 
11887 		return (pa);
11888 	}
11889 }
11890 
11891 /*
11892  * Carve up our nucleus hblk region.  We may allocate more hblks than
11893  * asked due to rounding errors but we are guaranteed to have at least
11894  * enough space to allocate the requested number of hblk8's and hblk1's.
11895  */
11896 void
11897 sfmmu_init_nucleus_hblks(caddr_t addr, size_t size, int nhblk8, int nhblk1)
11898 {
11899 	struct hme_blk *hmeblkp;
11900 	size_t hme8blk_sz, hme1blk_sz;
11901 	size_t i;
11902 	size_t hblk8_bound;
11903 	ulong_t j = 0, k = 0;
11904 
11905 	ASSERT(addr != NULL && size != 0);
11906 
11907 	/* Need to use proper structure alignment */
11908 	hme8blk_sz = roundup(HME8BLK_SZ, sizeof (int64_t));
11909 	hme1blk_sz = roundup(HME1BLK_SZ, sizeof (int64_t));
11910 
11911 	nucleus_hblk8.list = (void *)addr;
11912 	nucleus_hblk8.index = 0;
11913 
11914 	/*
11915 	 * Use as much memory as possible for hblk8's since we
11916 	 * expect all bop_alloc'ed memory to be allocated in 8k chunks.
11917 	 * We need to hold back enough space for the hblk1's which
11918 	 * we'll allocate next.
11919 	 */
11920 	hblk8_bound = size - (nhblk1 * hme1blk_sz) - hme8blk_sz;
11921 	for (i = 0; i <= hblk8_bound; i += hme8blk_sz, j++) {
11922 		hmeblkp = (struct hme_blk *)addr;
11923 		addr += hme8blk_sz;
11924 		hmeblkp->hblk_nuc_bit = 1;
11925 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
11926 	}
11927 	nucleus_hblk8.len = j;
11928 	ASSERT(j >= nhblk8);
11929 	SFMMU_STAT_ADD(sf_hblk8_ncreate, j);
11930 
11931 	nucleus_hblk1.list = (void *)addr;
11932 	nucleus_hblk1.index = 0;
11933 	for (; i <= (size - hme1blk_sz); i += hme1blk_sz, k++) {
11934 		hmeblkp = (struct hme_blk *)addr;
11935 		addr += hme1blk_sz;
11936 		hmeblkp->hblk_nuc_bit = 1;
11937 		hmeblkp->hblk_nextpa = cached_va_to_pa((caddr_t)hmeblkp);
11938 	}
11939 	ASSERT(k >= nhblk1);
11940 	nucleus_hblk1.len = k;
11941 	SFMMU_STAT_ADD(sf_hblk1_ncreate, k);
11942 }
11943 
11944 /*
11945  * This function is currently not supported on this platform. For what
11946  * it's supposed to do, see hat.c and hat_srmmu.c
11947  */
11948 /* ARGSUSED */
11949 faultcode_t
11950 hat_softlock(struct hat *hat, caddr_t addr, size_t *lenp, page_t **ppp,
11951     uint_t flags)
11952 {
11953 	ASSERT(hat->sfmmu_xhat_provider == NULL);
11954 	return (FC_NOSUPPORT);
11955 }
11956 
11957 /*
11958  * Searchs the mapping list of the page for a mapping of the same size. If not
11959  * found the corresponding bit is cleared in the p_index field. When large
11960  * pages are more prevalent in the system, we can maintain the mapping list
11961  * in order and we don't have to traverse the list each time. Just check the
11962  * next and prev entries, and if both are of different size, we clear the bit.
11963  */
11964 static void
11965 sfmmu_rm_large_mappings(page_t *pp, int ttesz)
11966 {
11967 	struct sf_hment *sfhmep;
11968 	struct hme_blk *hmeblkp;
11969 	int	index;
11970 	pgcnt_t	npgs;
11971 
11972 	ASSERT(ttesz > TTE8K);
11973 
11974 	ASSERT(sfmmu_mlist_held(pp));
11975 
11976 	ASSERT(PP_ISMAPPED_LARGE(pp));
11977 
11978 	/*
11979 	 * Traverse mapping list looking for another mapping of same size.
11980 	 * since we only want to clear index field if all mappings of
11981 	 * that size are gone.
11982 	 */
11983 
11984 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = sfhmep->hme_next) {
11985 		hmeblkp = sfmmu_hmetohblk(sfhmep);
11986 		if (hmeblkp->hblk_xhat_bit)
11987 			continue;
11988 		if (hme_size(sfhmep) == ttesz) {
11989 			/*
11990 			 * another mapping of the same size. don't clear index.
11991 			 */
11992 			return;
11993 		}
11994 	}
11995 
11996 	/*
11997 	 * Clear the p_index bit for large page.
11998 	 */
11999 	index = PAGESZ_TO_INDEX(ttesz);
12000 	npgs = TTEPAGES(ttesz);
12001 	while (npgs-- > 0) {
12002 		ASSERT(pp->p_index & index);
12003 		pp->p_index &= ~index;
12004 		pp = PP_PAGENEXT(pp);
12005 	}
12006 }
12007 
12008 /*
12009  * return supported features
12010  */
12011 /* ARGSUSED */
12012 int
12013 hat_supported(enum hat_features feature, void *arg)
12014 {
12015 	switch (feature) {
12016 	case    HAT_SHARED_PT:
12017 	case	HAT_DYNAMIC_ISM_UNMAP:
12018 	case	HAT_VMODSORT:
12019 		return (1);
12020 	default:
12021 		return (0);
12022 	}
12023 }
12024 
12025 void
12026 hat_enter(struct hat *hat)
12027 {
12028 	hatlock_t	*hatlockp;
12029 
12030 	if (hat != ksfmmup) {
12031 		hatlockp = TSB_HASH(hat);
12032 		mutex_enter(HATLOCK_MUTEXP(hatlockp));
12033 	}
12034 }
12035 
12036 void
12037 hat_exit(struct hat *hat)
12038 {
12039 	hatlock_t	*hatlockp;
12040 
12041 	if (hat != ksfmmup) {
12042 		hatlockp = TSB_HASH(hat);
12043 		mutex_exit(HATLOCK_MUTEXP(hatlockp));
12044 	}
12045 }
12046 
12047 /*ARGSUSED*/
12048 void
12049 hat_reserve(struct as *as, caddr_t addr, size_t len)
12050 {
12051 }
12052 
12053 static void
12054 hat_kstat_init(void)
12055 {
12056 	kstat_t *ksp;
12057 
12058 	ksp = kstat_create("unix", 0, "sfmmu_global_stat", "hat",
12059 		KSTAT_TYPE_RAW, sizeof (struct sfmmu_global_stat),
12060 		KSTAT_FLAG_VIRTUAL);
12061 	if (ksp) {
12062 		ksp->ks_data = (void *) &sfmmu_global_stat;
12063 		kstat_install(ksp);
12064 	}
12065 	ksp = kstat_create("unix", 0, "sfmmu_tsbsize_stat", "hat",
12066 		KSTAT_TYPE_RAW, sizeof (struct sfmmu_tsbsize_stat),
12067 		KSTAT_FLAG_VIRTUAL);
12068 	if (ksp) {
12069 		ksp->ks_data = (void *) &sfmmu_tsbsize_stat;
12070 		kstat_install(ksp);
12071 	}
12072 	ksp = kstat_create("unix", 0, "sfmmu_percpu_stat", "hat",
12073 		KSTAT_TYPE_RAW, sizeof (struct sfmmu_percpu_stat) * NCPU,
12074 		KSTAT_FLAG_WRITABLE);
12075 	if (ksp) {
12076 		ksp->ks_update = sfmmu_kstat_percpu_update;
12077 		kstat_install(ksp);
12078 	}
12079 }
12080 
12081 /* ARGSUSED */
12082 static int
12083 sfmmu_kstat_percpu_update(kstat_t *ksp, int rw)
12084 {
12085 	struct sfmmu_percpu_stat *cpu_kstat = ksp->ks_data;
12086 	struct tsbmiss *tsbm = tsbmiss_area;
12087 	struct kpmtsbm *kpmtsbm = kpmtsbm_area;
12088 	int i;
12089 
12090 	ASSERT(cpu_kstat);
12091 	if (rw == KSTAT_READ) {
12092 		for (i = 0; i < NCPU; cpu_kstat++, tsbm++, kpmtsbm++, i++) {
12093 			cpu_kstat->sf_itlb_misses = tsbm->itlb_misses;
12094 			cpu_kstat->sf_dtlb_misses = tsbm->dtlb_misses;
12095 			cpu_kstat->sf_utsb_misses = tsbm->utsb_misses -
12096 				tsbm->uprot_traps;
12097 			cpu_kstat->sf_ktsb_misses = tsbm->ktsb_misses +
12098 				kpmtsbm->kpm_tsb_misses - tsbm->kprot_traps;
12099 
12100 			if (tsbm->itlb_misses > 0 && tsbm->dtlb_misses > 0) {
12101 				cpu_kstat->sf_tsb_hits =
12102 				(tsbm->itlb_misses + tsbm->dtlb_misses) -
12103 				(tsbm->utsb_misses + tsbm->ktsb_misses +
12104 				kpmtsbm->kpm_tsb_misses);
12105 			} else {
12106 				cpu_kstat->sf_tsb_hits = 0;
12107 			}
12108 			cpu_kstat->sf_umod_faults = tsbm->uprot_traps;
12109 			cpu_kstat->sf_kmod_faults = tsbm->kprot_traps;
12110 		}
12111 	} else {
12112 		/* KSTAT_WRITE is used to clear stats */
12113 		for (i = 0; i < NCPU; tsbm++, kpmtsbm++, i++) {
12114 			tsbm->itlb_misses = 0;
12115 			tsbm->dtlb_misses = 0;
12116 			tsbm->utsb_misses = 0;
12117 			tsbm->ktsb_misses = 0;
12118 			tsbm->uprot_traps = 0;
12119 			tsbm->kprot_traps = 0;
12120 			kpmtsbm->kpm_dtlb_misses = 0;
12121 			kpmtsbm->kpm_tsb_misses = 0;
12122 		}
12123 	}
12124 	return (0);
12125 }
12126 
12127 #ifdef	DEBUG
12128 
12129 tte_t  *gorig[NCPU], *gcur[NCPU], *gnew[NCPU];
12130 
12131 /*
12132  * A tte checker. *orig_old is the value we read before cas.
12133  *	*cur is the value returned by cas.
12134  *	*new is the desired value when we do the cas.
12135  *
12136  *	*hmeblkp is currently unused.
12137  */
12138 
12139 /* ARGSUSED */
12140 void
12141 chk_tte(tte_t *orig_old, tte_t *cur, tte_t *new, struct hme_blk *hmeblkp)
12142 {
12143 	uint_t i, j, k;
12144 	int cpuid = CPU->cpu_id;
12145 
12146 	gorig[cpuid] = orig_old;
12147 	gcur[cpuid] = cur;
12148 	gnew[cpuid] = new;
12149 
12150 #ifdef lint
12151 	hmeblkp = hmeblkp;
12152 #endif
12153 
12154 	if (TTE_IS_VALID(orig_old)) {
12155 		if (TTE_IS_VALID(cur)) {
12156 			i = TTE_TO_TTEPFN(orig_old);
12157 			j = TTE_TO_TTEPFN(cur);
12158 			k = TTE_TO_TTEPFN(new);
12159 			if (i != j) {
12160 				/* remap error? */
12161 				panic("chk_tte: bad pfn, 0x%x, 0x%x",
12162 					i, j);
12163 			}
12164 
12165 			if (i != k) {
12166 				/* remap error? */
12167 				panic("chk_tte: bad pfn2, 0x%x, 0x%x",
12168 					i, k);
12169 			}
12170 		} else {
12171 			if (TTE_IS_VALID(new)) {
12172 				panic("chk_tte: invalid cur? ");
12173 			}
12174 
12175 			i = TTE_TO_TTEPFN(orig_old);
12176 			k = TTE_TO_TTEPFN(new);
12177 			if (i != k) {
12178 				panic("chk_tte: bad pfn3, 0x%x, 0x%x",
12179 					i, k);
12180 			}
12181 		}
12182 	} else {
12183 		if (TTE_IS_VALID(cur)) {
12184 			j = TTE_TO_TTEPFN(cur);
12185 			if (TTE_IS_VALID(new)) {
12186 				k = TTE_TO_TTEPFN(new);
12187 				if (j != k) {
12188 					panic("chk_tte: bad pfn4, 0x%x, 0x%x",
12189 						j, k);
12190 				}
12191 			} else {
12192 				panic("chk_tte: why here?");
12193 			}
12194 		} else {
12195 			if (!TTE_IS_VALID(new)) {
12196 				panic("chk_tte: why here2 ?");
12197 			}
12198 		}
12199 	}
12200 }
12201 
12202 #endif /* DEBUG */
12203 
12204 extern void prefetch_tsbe_read(struct tsbe *);
12205 extern void prefetch_tsbe_write(struct tsbe *);
12206 
12207 
12208 /*
12209  * We want to prefetch 7 cache lines ahead for our read prefetch.  This gives
12210  * us optimal performance on Cheetah+.  You can only have 8 outstanding
12211  * prefetches at any one time, so we opted for 7 read prefetches and 1 write
12212  * prefetch to make the most utilization of the prefetch capability.
12213  */
12214 #define	TSBE_PREFETCH_STRIDE (7)
12215 
12216 void
12217 sfmmu_copy_tsb(struct tsb_info *old_tsbinfo, struct tsb_info *new_tsbinfo)
12218 {
12219 	int old_bytes = TSB_BYTES(old_tsbinfo->tsb_szc);
12220 	int new_bytes = TSB_BYTES(new_tsbinfo->tsb_szc);
12221 	int old_entries = TSB_ENTRIES(old_tsbinfo->tsb_szc);
12222 	int new_entries = TSB_ENTRIES(new_tsbinfo->tsb_szc);
12223 	struct tsbe *old;
12224 	struct tsbe *new;
12225 	struct tsbe *new_base = (struct tsbe *)new_tsbinfo->tsb_va;
12226 	uint64_t va;
12227 	int new_offset;
12228 	int i;
12229 	int vpshift;
12230 	int last_prefetch;
12231 
12232 	if (old_bytes == new_bytes) {
12233 		bcopy(old_tsbinfo->tsb_va, new_tsbinfo->tsb_va, new_bytes);
12234 	} else {
12235 
12236 		/*
12237 		 * A TSBE is 16 bytes which means there are four TSBE's per
12238 		 * P$ line (64 bytes), thus every 4 TSBE's we prefetch.
12239 		 */
12240 		old = (struct tsbe *)old_tsbinfo->tsb_va;
12241 		last_prefetch = old_entries - (4*(TSBE_PREFETCH_STRIDE+1));
12242 		for (i = 0; i < old_entries; i++, old++) {
12243 			if (((i & (4-1)) == 0) && (i < last_prefetch))
12244 				prefetch_tsbe_read(old);
12245 			if (!old->tte_tag.tag_invalid) {
12246 				/*
12247 				 * We have a valid TTE to remap.  Check the
12248 				 * size.  We won't remap 64K or 512K TTEs
12249 				 * because they span more than one TSB entry
12250 				 * and are indexed using an 8K virt. page.
12251 				 * Ditto for 32M and 256M TTEs.
12252 				 */
12253 				if (TTE_CSZ(&old->tte_data) == TTE64K ||
12254 				    TTE_CSZ(&old->tte_data) == TTE512K)
12255 					continue;
12256 				if (mmu_page_sizes == max_mmu_page_sizes) {
12257 				    if (TTE_CSZ(&old->tte_data) == TTE32M ||
12258 					TTE_CSZ(&old->tte_data) == TTE256M)
12259 					    continue;
12260 				}
12261 
12262 				/* clear the lower 22 bits of the va */
12263 				va = *(uint64_t *)old << 22;
12264 				/* turn va into a virtual pfn */
12265 				va >>= 22 - TSB_START_SIZE;
12266 				/*
12267 				 * or in bits from the offset in the tsb
12268 				 * to get the real virtual pfn. These
12269 				 * correspond to bits [21:13] in the va
12270 				 */
12271 				vpshift =
12272 				    TTE_BSZS_SHIFT(TTE_CSZ(&old->tte_data)) &
12273 				    0x1ff;
12274 				va |= (i << vpshift);
12275 				va >>= vpshift;
12276 				new_offset = va & (new_entries - 1);
12277 				new = new_base + new_offset;
12278 				prefetch_tsbe_write(new);
12279 				*new = *old;
12280 			}
12281 		}
12282 	}
12283 }
12284 
12285 /*
12286  * Kernel Physical Mapping (kpm) facility
12287  */
12288 
12289 /* -- hat_kpm interface section -- */
12290 
12291 /*
12292  * Mapin a locked page and return the vaddr.
12293  * When a kpme is provided by the caller it is added to
12294  * the page p_kpmelist. The page to be mapped in must
12295  * be at least read locked (p_selock).
12296  */
12297 caddr_t
12298 hat_kpm_mapin(struct page *pp, struct kpme *kpme)
12299 {
12300 	kmutex_t	*pml;
12301 	caddr_t		vaddr;
12302 
12303 	if (kpm_enable == 0) {
12304 		cmn_err(CE_WARN, "hat_kpm_mapin: kpm_enable not set");
12305 		return ((caddr_t)NULL);
12306 	}
12307 
12308 	if (pp == NULL || PAGE_LOCKED(pp) == 0) {
12309 		cmn_err(CE_WARN, "hat_kpm_mapin: pp zero or not locked");
12310 		return ((caddr_t)NULL);
12311 	}
12312 
12313 	pml = sfmmu_mlist_enter(pp);
12314 	ASSERT(pp->p_kpmref >= 0);
12315 
12316 	vaddr = (pp->p_kpmref == 0) ?
12317 		sfmmu_kpm_mapin(pp) : hat_kpm_page2va(pp, 1);
12318 
12319 	if (kpme != NULL) {
12320 		/*
12321 		 * Tolerate multiple mapins for the same kpme to avoid
12322 		 * the need for an extra serialization.
12323 		 */
12324 		if ((sfmmu_kpme_lookup(kpme, pp)) == 0)
12325 			sfmmu_kpme_add(kpme, pp);
12326 
12327 		ASSERT(pp->p_kpmref > 0);
12328 
12329 	} else {
12330 		pp->p_kpmref++;
12331 	}
12332 
12333 	sfmmu_mlist_exit(pml);
12334 	return (vaddr);
12335 }
12336 
12337 /*
12338  * Mapout a locked page.
12339  * When a kpme is provided by the caller it is removed from
12340  * the page p_kpmelist. The page to be mapped out must be at
12341  * least read locked (p_selock).
12342  * Note: The seg_kpm layer provides a mapout interface for the
12343  * case that a kpme is used and the underlying page is unlocked.
12344  * This can be used instead of calling this function directly.
12345  */
12346 void
12347 hat_kpm_mapout(struct page *pp, struct kpme *kpme, caddr_t vaddr)
12348 {
12349 	kmutex_t	*pml;
12350 
12351 	if (kpm_enable == 0) {
12352 		cmn_err(CE_WARN, "hat_kpm_mapout: kpm_enable not set");
12353 		return;
12354 	}
12355 
12356 	if (IS_KPM_ADDR(vaddr) == 0) {
12357 		cmn_err(CE_WARN, "hat_kpm_mapout: no kpm address");
12358 		return;
12359 	}
12360 
12361 	if (pp == NULL || PAGE_LOCKED(pp) == 0) {
12362 		cmn_err(CE_WARN, "hat_kpm_mapout: page zero or not locked");
12363 		return;
12364 	}
12365 
12366 	if (kpme != NULL) {
12367 		ASSERT(pp == kpme->kpe_page);
12368 		pp = kpme->kpe_page;
12369 		pml = sfmmu_mlist_enter(pp);
12370 
12371 		if (sfmmu_kpme_lookup(kpme, pp) == 0)
12372 			panic("hat_kpm_mapout: kpme not found pp=%p",
12373 				(void *)pp);
12374 
12375 		ASSERT(pp->p_kpmref > 0);
12376 		sfmmu_kpme_sub(kpme, pp);
12377 
12378 	} else {
12379 		pml = sfmmu_mlist_enter(pp);
12380 		pp->p_kpmref--;
12381 	}
12382 
12383 	ASSERT(pp->p_kpmref >= 0);
12384 	if (pp->p_kpmref == 0)
12385 		sfmmu_kpm_mapout(pp, vaddr);
12386 
12387 	sfmmu_mlist_exit(pml);
12388 }
12389 
12390 /*
12391  * Return the kpm virtual address for the page at pp.
12392  * If checkswap is non zero and the page is backed by a
12393  * swap vnode the physical address is used rather than
12394  * p_offset to determine the kpm region.
12395  * Note: The function has to be used w/ extreme care. The
12396  * stability of the page identity is in the responsibility
12397  * of the caller.
12398  */
12399 caddr_t
12400 hat_kpm_page2va(struct page *pp, int checkswap)
12401 {
12402 	int		vcolor, vcolor_pa;
12403 	uintptr_t	paddr, vaddr;
12404 
12405 	ASSERT(kpm_enable);
12406 
12407 	paddr = ptob(pp->p_pagenum);
12408 	vcolor_pa = addr_to_vcolor(paddr);
12409 
12410 	if (checkswap && pp->p_vnode && IS_SWAPFSVP(pp->p_vnode))
12411 		vcolor = (PP_ISNC(pp)) ? vcolor_pa : PP_GET_VCOLOR(pp);
12412 	else
12413 		vcolor = addr_to_vcolor(pp->p_offset);
12414 
12415 	vaddr = (uintptr_t)kpm_vbase + paddr;
12416 
12417 	if (vcolor_pa != vcolor) {
12418 		vaddr += ((uintptr_t)(vcolor - vcolor_pa) << MMU_PAGESHIFT);
12419 		vaddr += (vcolor_pa > vcolor) ?
12420 			((uintptr_t)vcolor_pa << kpm_size_shift) :
12421 			((uintptr_t)(vcolor - vcolor_pa) << kpm_size_shift);
12422 	}
12423 
12424 	return ((caddr_t)vaddr);
12425 }
12426 
12427 /*
12428  * Return the page for the kpm virtual address vaddr.
12429  * Caller is responsible for the kpm mapping and lock
12430  * state of the page.
12431  */
12432 page_t *
12433 hat_kpm_vaddr2page(caddr_t vaddr)
12434 {
12435 	uintptr_t	paddr;
12436 	pfn_t		pfn;
12437 
12438 	ASSERT(IS_KPM_ADDR(vaddr));
12439 
12440 	SFMMU_KPM_VTOP(vaddr, paddr);
12441 	pfn = (pfn_t)btop(paddr);
12442 
12443 	return (page_numtopp_nolock(pfn));
12444 }
12445 
12446 /* page to kpm_page */
12447 #define	PP2KPMPG(pp, kp) {						\
12448 	struct memseg	*mseg;						\
12449 	pgcnt_t		inx;						\
12450 	pfn_t		pfn;						\
12451 									\
12452 	pfn = pp->p_pagenum;						\
12453 	mseg = page_numtomemseg_nolock(pfn);				\
12454 	ASSERT(mseg);							\
12455 	inx = ptokpmp(kpmptop(ptokpmp(pfn)) - mseg->kpm_pbase);		\
12456 	ASSERT(inx < mseg->kpm_nkpmpgs);				\
12457 	kp = &mseg->kpm_pages[inx];					\
12458 }
12459 
12460 /* page to kpm_spage */
12461 #define	PP2KPMSPG(pp, ksp) {						\
12462 	struct memseg	*mseg;						\
12463 	pgcnt_t		inx;						\
12464 	pfn_t		pfn;						\
12465 									\
12466 	pfn = pp->p_pagenum;						\
12467 	mseg = page_numtomemseg_nolock(pfn);				\
12468 	ASSERT(mseg);							\
12469 	inx = pfn - mseg->kpm_pbase;					\
12470 	ksp = &mseg->kpm_spages[inx];					\
12471 }
12472 
12473 /*
12474  * hat_kpm_fault is called from segkpm_fault when a kpm tsbmiss occurred
12475  * which could not be resolved by the trap level tsbmiss handler for the
12476  * following reasons:
12477  * . The vaddr is in VAC alias range (always PAGESIZE mapping size).
12478  * . The kpm (s)page range of vaddr is in a VAC alias prevention state.
12479  * . tsbmiss handling at trap level is not desired (DEBUG kernel only,
12480  *   kpm_tsbmtl == 0).
12481  */
12482 int
12483 hat_kpm_fault(struct hat *hat, caddr_t vaddr)
12484 {
12485 	int		error;
12486 	uintptr_t	paddr;
12487 	pfn_t		pfn;
12488 	struct memseg	*mseg;
12489 	page_t	*pp;
12490 
12491 	if (kpm_enable == 0) {
12492 		cmn_err(CE_WARN, "hat_kpm_fault: kpm_enable not set");
12493 		return (ENOTSUP);
12494 	}
12495 
12496 	ASSERT(hat == ksfmmup);
12497 	ASSERT(IS_KPM_ADDR(vaddr));
12498 
12499 	SFMMU_KPM_VTOP(vaddr, paddr);
12500 	pfn = (pfn_t)btop(paddr);
12501 	mseg = page_numtomemseg_nolock(pfn);
12502 	if (mseg == NULL)
12503 		return (EFAULT);
12504 
12505 	pp = &mseg->pages[(pgcnt_t)(pfn - mseg->pages_base)];
12506 	ASSERT((pfn_t)pp->p_pagenum == pfn);
12507 
12508 	if (!PAGE_LOCKED(pp))
12509 		return (EFAULT);
12510 
12511 	if (kpm_smallpages == 0)
12512 		error = sfmmu_kpm_fault(vaddr, mseg, pp);
12513 	else
12514 		error = sfmmu_kpm_fault_small(vaddr, mseg, pp);
12515 
12516 	return (error);
12517 }
12518 
12519 extern  krwlock_t memsegslock;
12520 
12521 /*
12522  * memseg_hash[] was cleared, need to clear memseg_phash[] too.
12523  */
12524 void
12525 hat_kpm_mseghash_clear(int nentries)
12526 {
12527 	pgcnt_t i;
12528 
12529 	if (kpm_enable == 0)
12530 		return;
12531 
12532 	for (i = 0; i < nentries; i++)
12533 		memseg_phash[i] = MSEG_NULLPTR_PA;
12534 }
12535 
12536 /*
12537  * Update memseg_phash[inx] when memseg_hash[inx] was changed.
12538  */
12539 void
12540 hat_kpm_mseghash_update(pgcnt_t inx, struct memseg *msp)
12541 {
12542 	if (kpm_enable == 0)
12543 		return;
12544 
12545 	memseg_phash[inx] = (msp) ? va_to_pa(msp) : MSEG_NULLPTR_PA;
12546 }
12547 
12548 /*
12549  * Update kpm memseg members from basic memseg info.
12550  */
12551 void
12552 hat_kpm_addmem_mseg_update(struct memseg *msp, pgcnt_t nkpmpgs,
12553 	offset_t kpm_pages_off)
12554 {
12555 	if (kpm_enable == 0)
12556 		return;
12557 
12558 	msp->kpm_pages = (kpm_page_t *)((caddr_t)msp->pages + kpm_pages_off);
12559 	msp->kpm_nkpmpgs = nkpmpgs;
12560 	msp->kpm_pbase = kpmptop(ptokpmp(msp->pages_base));
12561 	msp->pagespa = va_to_pa(msp->pages);
12562 	msp->epagespa = va_to_pa(msp->epages);
12563 	msp->kpm_pagespa = va_to_pa(msp->kpm_pages);
12564 }
12565 
12566 /*
12567  * Setup nextpa when a memseg is inserted.
12568  * Assumes that the memsegslock is already held.
12569  */
12570 void
12571 hat_kpm_addmem_mseg_insert(struct memseg *msp)
12572 {
12573 	if (kpm_enable == 0)
12574 		return;
12575 
12576 	ASSERT(RW_LOCK_HELD(&memsegslock));
12577 	msp->nextpa = (memsegs) ? va_to_pa(memsegs) : MSEG_NULLPTR_PA;
12578 }
12579 
12580 /*
12581  * Setup memsegspa when a memseg is (head) inserted.
12582  * Called before memsegs is updated to complete a
12583  * memseg insert operation.
12584  * Assumes that the memsegslock is already held.
12585  */
12586 void
12587 hat_kpm_addmem_memsegs_update(struct memseg *msp)
12588 {
12589 	if (kpm_enable == 0)
12590 		return;
12591 
12592 	ASSERT(RW_LOCK_HELD(&memsegslock));
12593 	ASSERT(memsegs);
12594 	memsegspa = va_to_pa(msp);
12595 }
12596 
12597 /*
12598  * Return end of metadata for an already setup memseg.
12599  *
12600  * Note: kpm_pages and kpm_spages are aliases and the underlying
12601  * member of struct memseg is a union, therefore they always have
12602  * the same address within a memseg. They must be differentiated
12603  * when pointer arithmetic is used with them.
12604  */
12605 caddr_t
12606 hat_kpm_mseg_reuse(struct memseg *msp)
12607 {
12608 	caddr_t end;
12609 
12610 	if (kpm_smallpages == 0)
12611 		end = (caddr_t)(msp->kpm_pages + msp->kpm_nkpmpgs);
12612 	else
12613 		end = (caddr_t)(msp->kpm_spages + msp->kpm_nkpmpgs);
12614 
12615 	return (end);
12616 }
12617 
12618 /*
12619  * Update memsegspa (when first memseg in list
12620  * is deleted) or nextpa  when a memseg deleted.
12621  * Assumes that the memsegslock is already held.
12622  */
12623 void
12624 hat_kpm_delmem_mseg_update(struct memseg *msp, struct memseg **mspp)
12625 {
12626 	struct memseg *lmsp;
12627 
12628 	if (kpm_enable == 0)
12629 		return;
12630 
12631 	ASSERT(RW_LOCK_HELD(&memsegslock));
12632 
12633 	if (mspp == &memsegs) {
12634 		memsegspa = (msp->next) ?
12635 				va_to_pa(msp->next) : MSEG_NULLPTR_PA;
12636 	} else {
12637 		lmsp = (struct memseg *)
12638 			((uint64_t)mspp - offsetof(struct memseg, next));
12639 		lmsp->nextpa = (msp->next) ?
12640 				va_to_pa(msp->next) : MSEG_NULLPTR_PA;
12641 	}
12642 }
12643 
12644 /*
12645  * Update kpm members for all memseg's involved in a split operation
12646  * and do the atomic update of the physical memseg chain.
12647  *
12648  * Note: kpm_pages and kpm_spages are aliases and the underlying member
12649  * of struct memseg is a union, therefore they always have the same
12650  * address within a memseg. With that the direct assignments and
12651  * va_to_pa conversions below don't have to be distinguished wrt. to
12652  * kpm_smallpages. They must be differentiated when pointer arithmetic
12653  * is used with them.
12654  *
12655  * Assumes that the memsegslock is already held.
12656  */
12657 void
12658 hat_kpm_split_mseg_update(struct memseg *msp, struct memseg **mspp,
12659 	struct memseg *lo, struct memseg *mid, struct memseg *hi)
12660 {
12661 	pgcnt_t start, end, kbase, kstart, num;
12662 	struct memseg *lmsp;
12663 
12664 	if (kpm_enable == 0)
12665 		return;
12666 
12667 	ASSERT(RW_LOCK_HELD(&memsegslock));
12668 	ASSERT(msp && mid && msp->kpm_pages);
12669 
12670 	kbase = ptokpmp(msp->kpm_pbase);
12671 
12672 	if (lo) {
12673 		num = lo->pages_end - lo->pages_base;
12674 		start = kpmptop(ptokpmp(lo->pages_base));
12675 		/* align end to kpm page size granularity */
12676 		end = kpmptop(ptokpmp(start + num - 1)) + kpmpnpgs;
12677 		lo->kpm_pbase = start;
12678 		lo->kpm_nkpmpgs = ptokpmp(end - start);
12679 		lo->kpm_pages = msp->kpm_pages;
12680 		lo->kpm_pagespa = va_to_pa(lo->kpm_pages);
12681 		lo->pagespa = va_to_pa(lo->pages);
12682 		lo->epagespa = va_to_pa(lo->epages);
12683 		lo->nextpa = va_to_pa(lo->next);
12684 	}
12685 
12686 	/* mid */
12687 	num = mid->pages_end - mid->pages_base;
12688 	kstart = ptokpmp(mid->pages_base);
12689 	start = kpmptop(kstart);
12690 	/* align end to kpm page size granularity */
12691 	end = kpmptop(ptokpmp(start + num - 1)) + kpmpnpgs;
12692 	mid->kpm_pbase = start;
12693 	mid->kpm_nkpmpgs = ptokpmp(end - start);
12694 	if (kpm_smallpages == 0) {
12695 		mid->kpm_pages = msp->kpm_pages + (kstart - kbase);
12696 	} else {
12697 		mid->kpm_spages = msp->kpm_spages + (kstart - kbase);
12698 	}
12699 	mid->kpm_pagespa = va_to_pa(mid->kpm_pages);
12700 	mid->pagespa = va_to_pa(mid->pages);
12701 	mid->epagespa = va_to_pa(mid->epages);
12702 	mid->nextpa = (mid->next) ?  va_to_pa(mid->next) : MSEG_NULLPTR_PA;
12703 
12704 	if (hi) {
12705 		num = hi->pages_end - hi->pages_base;
12706 		kstart = ptokpmp(hi->pages_base);
12707 		start = kpmptop(kstart);
12708 		/* align end to kpm page size granularity */
12709 		end = kpmptop(ptokpmp(start + num - 1)) + kpmpnpgs;
12710 		hi->kpm_pbase = start;
12711 		hi->kpm_nkpmpgs = ptokpmp(end - start);
12712 		if (kpm_smallpages == 0) {
12713 			hi->kpm_pages = msp->kpm_pages + (kstart - kbase);
12714 		} else {
12715 			hi->kpm_spages = msp->kpm_spages + (kstart - kbase);
12716 		}
12717 		hi->kpm_pagespa = va_to_pa(hi->kpm_pages);
12718 		hi->pagespa = va_to_pa(hi->pages);
12719 		hi->epagespa = va_to_pa(hi->epages);
12720 		hi->nextpa = (hi->next) ? va_to_pa(hi->next) : MSEG_NULLPTR_PA;
12721 	}
12722 
12723 	/*
12724 	 * Atomic update of the physical memseg chain
12725 	 */
12726 	if (mspp == &memsegs) {
12727 		memsegspa = (lo) ? va_to_pa(lo) : va_to_pa(mid);
12728 	} else {
12729 		lmsp = (struct memseg *)
12730 			((uint64_t)mspp - offsetof(struct memseg, next));
12731 		lmsp->nextpa = (lo) ? va_to_pa(lo) : va_to_pa(mid);
12732 	}
12733 }
12734 
12735 /*
12736  * Walk the memsegs chain, applying func to each memseg span and vcolor.
12737  */
12738 void
12739 hat_kpm_walk(void (*func)(void *, void *, size_t), void *arg)
12740 {
12741 	pfn_t	pbase, pend;
12742 	int	vcolor;
12743 	void	*base;
12744 	size_t	size;
12745 	struct memseg *msp;
12746 	extern uint_t vac_colors;
12747 
12748 	for (msp = memsegs; msp; msp = msp->next) {
12749 		pbase = msp->pages_base;
12750 		pend = msp->pages_end;
12751 		for (vcolor = 0; vcolor < vac_colors; vcolor++) {
12752 			base = ptob(pbase) + kpm_vbase + kpm_size * vcolor;
12753 			size = ptob(pend - pbase);
12754 			func(arg, base, size);
12755 		}
12756 	}
12757 }
12758 
12759 
12760 /* -- sfmmu_kpm internal section -- */
12761 
12762 /*
12763  * Return the page frame number if a valid segkpm mapping exists
12764  * for vaddr, otherwise return PFN_INVALID. No locks are grabbed.
12765  * Should only be used by other sfmmu routines.
12766  */
12767 pfn_t
12768 sfmmu_kpm_vatopfn(caddr_t vaddr)
12769 {
12770 	uintptr_t	paddr;
12771 	pfn_t		pfn;
12772 	page_t	*pp;
12773 
12774 	ASSERT(kpm_enable && IS_KPM_ADDR(vaddr));
12775 
12776 	SFMMU_KPM_VTOP(vaddr, paddr);
12777 	pfn = (pfn_t)btop(paddr);
12778 	pp = page_numtopp_nolock(pfn);
12779 	if (pp && pp->p_kpmref)
12780 		return (pfn);
12781 	else
12782 		return ((pfn_t)PFN_INVALID);
12783 }
12784 
12785 /*
12786  * Lookup a kpme in the p_kpmelist.
12787  */
12788 static int
12789 sfmmu_kpme_lookup(struct kpme *kpme, page_t *pp)
12790 {
12791 	struct kpme	*p;
12792 
12793 	for (p = pp->p_kpmelist; p; p = p->kpe_next) {
12794 		if (p == kpme)
12795 			return (1);
12796 	}
12797 	return (0);
12798 }
12799 
12800 /*
12801  * Insert a kpme into the p_kpmelist and increment
12802  * the per page kpm reference count.
12803  */
12804 static void
12805 sfmmu_kpme_add(struct kpme *kpme, page_t *pp)
12806 {
12807 	ASSERT(pp->p_kpmref >= 0);
12808 
12809 	/* head insert */
12810 	kpme->kpe_prev = NULL;
12811 	kpme->kpe_next = pp->p_kpmelist;
12812 
12813 	if (pp->p_kpmelist)
12814 		pp->p_kpmelist->kpe_prev = kpme;
12815 
12816 	pp->p_kpmelist = kpme;
12817 	kpme->kpe_page = pp;
12818 	pp->p_kpmref++;
12819 }
12820 
12821 /*
12822  * Remove a kpme from the p_kpmelist and decrement
12823  * the per page kpm reference count.
12824  */
12825 static void
12826 sfmmu_kpme_sub(struct kpme *kpme, page_t *pp)
12827 {
12828 	ASSERT(pp->p_kpmref > 0);
12829 
12830 	if (kpme->kpe_prev) {
12831 		ASSERT(pp->p_kpmelist != kpme);
12832 		ASSERT(kpme->kpe_prev->kpe_page == pp);
12833 		kpme->kpe_prev->kpe_next = kpme->kpe_next;
12834 	} else {
12835 		ASSERT(pp->p_kpmelist == kpme);
12836 		pp->p_kpmelist = kpme->kpe_next;
12837 	}
12838 
12839 	if (kpme->kpe_next) {
12840 		ASSERT(kpme->kpe_next->kpe_page == pp);
12841 		kpme->kpe_next->kpe_prev = kpme->kpe_prev;
12842 	}
12843 
12844 	kpme->kpe_next = kpme->kpe_prev = NULL;
12845 	kpme->kpe_page = NULL;
12846 	pp->p_kpmref--;
12847 }
12848 
12849 /*
12850  * Mapin a single page, it is called every time a page changes it's state
12851  * from kpm-unmapped to kpm-mapped. It may not be called, when only a new
12852  * kpm instance does a mapin and wants to share the mapping.
12853  * Assumes that the mlist mutex is already grabbed.
12854  */
12855 static caddr_t
12856 sfmmu_kpm_mapin(page_t *pp)
12857 {
12858 	kpm_page_t	*kp;
12859 	kpm_hlk_t	*kpmp;
12860 	caddr_t		vaddr;
12861 	int		kpm_vac_range;
12862 	pfn_t		pfn;
12863 	tte_t		tte;
12864 	kmutex_t	*pmtx;
12865 	int		uncached;
12866 	kpm_spage_t	*ksp;
12867 	kpm_shlk_t	*kpmsp;
12868 	int		oldval;
12869 
12870 	ASSERT(sfmmu_mlist_held(pp));
12871 	ASSERT(pp->p_kpmref == 0);
12872 
12873 	vaddr = sfmmu_kpm_getvaddr(pp, &kpm_vac_range);
12874 
12875 	ASSERT(IS_KPM_ADDR(vaddr));
12876 	uncached = PP_ISNC(pp);
12877 	pfn = pp->p_pagenum;
12878 
12879 	if (kpm_smallpages)
12880 		goto smallpages_mapin;
12881 
12882 	PP2KPMPG(pp, kp);
12883 
12884 	kpmp = KPMP_HASH(kp);
12885 	mutex_enter(&kpmp->khl_mutex);
12886 
12887 	ASSERT(PP_ISKPMC(pp) == 0);
12888 	ASSERT(PP_ISKPMS(pp) == 0);
12889 
12890 	if (uncached) {
12891 		/* ASSERT(pp->p_share); XXX use hat_page_getshare */
12892 		if (kpm_vac_range == 0) {
12893 			if (kp->kp_refcnts == 0) {
12894 				/*
12895 				 * Must remove large page mapping if it exists.
12896 				 * Pages in uncached state can only be mapped
12897 				 * small (PAGESIZE) within the regular kpm
12898 				 * range.
12899 				 */
12900 				if (kp->kp_refcntc == -1) {
12901 					/* remove go indication */
12902 					sfmmu_kpm_tsbmtl(&kp->kp_refcntc,
12903 						&kpmp->khl_lock, KPMTSBM_STOP);
12904 				}
12905 				if (kp->kp_refcnt > 0 && kp->kp_refcntc == 0)
12906 					sfmmu_kpm_demap_large(vaddr);
12907 			}
12908 			ASSERT(kp->kp_refcntc >= 0);
12909 			kp->kp_refcntc++;
12910 		}
12911 		pmtx = sfmmu_page_enter(pp);
12912 		PP_SETKPMC(pp);
12913 		sfmmu_page_exit(pmtx);
12914 	}
12915 
12916 	if ((kp->kp_refcntc > 0 || kp->kp_refcnts > 0) && kpm_vac_range == 0) {
12917 		/*
12918 		 * Have to do a small (PAGESIZE) mapin within this kpm_page
12919 		 * range since it is marked to be in VAC conflict mode or
12920 		 * when there are still other small mappings around.
12921 		 */
12922 
12923 		/* tte assembly */
12924 		if (uncached == 0)
12925 			KPM_TTE_VCACHED(tte.ll, pfn, TTE8K);
12926 		else
12927 			KPM_TTE_VUNCACHED(tte.ll, pfn, TTE8K);
12928 
12929 		/* tsb dropin */
12930 		sfmmu_kpm_load_tsb(vaddr, &tte, MMU_PAGESHIFT);
12931 
12932 		pmtx = sfmmu_page_enter(pp);
12933 		PP_SETKPMS(pp);
12934 		sfmmu_page_exit(pmtx);
12935 
12936 		kp->kp_refcnts++;
12937 		ASSERT(kp->kp_refcnts > 0);
12938 		goto exit;
12939 	}
12940 
12941 	if (kpm_vac_range == 0) {
12942 		/*
12943 		 * Fast path / regular case, no VAC conflict handling
12944 		 * in progress within this kpm_page range.
12945 		 */
12946 		if (kp->kp_refcnt == 0) {
12947 
12948 			/* tte assembly */
12949 			KPM_TTE_VCACHED(tte.ll, pfn, TTE4M);
12950 
12951 			/* tsb dropin */
12952 			sfmmu_kpm_load_tsb(vaddr, &tte, MMU_PAGESHIFT4M);
12953 
12954 			/* Set go flag for TL tsbmiss handler */
12955 			if (kp->kp_refcntc == 0)
12956 				sfmmu_kpm_tsbmtl(&kp->kp_refcntc,
12957 						&kpmp->khl_lock, KPMTSBM_START);
12958 
12959 			ASSERT(kp->kp_refcntc == -1);
12960 		}
12961 		kp->kp_refcnt++;
12962 		ASSERT(kp->kp_refcnt);
12963 
12964 	} else {
12965 		/*
12966 		 * The page is not setup according to the common VAC
12967 		 * prevention rules for the regular and kpm mapping layer
12968 		 * E.g. the page layer was not able to deliver a right
12969 		 * vcolor'ed page for a given vaddr corresponding to
12970 		 * the wanted p_offset. It has to be mapped in small in
12971 		 * within the corresponding kpm vac range in order to
12972 		 * prevent VAC alias conflicts.
12973 		 */
12974 
12975 		/* tte assembly */
12976 		if (uncached == 0) {
12977 			KPM_TTE_VCACHED(tte.ll, pfn, TTE8K);
12978 		} else {
12979 			KPM_TTE_VUNCACHED(tte.ll, pfn, TTE8K);
12980 		}
12981 
12982 		/* tsb dropin */
12983 		sfmmu_kpm_load_tsb(vaddr, &tte, MMU_PAGESHIFT);
12984 
12985 		kp->kp_refcnta++;
12986 		if (kp->kp_refcntc == -1) {
12987 			ASSERT(kp->kp_refcnt > 0);
12988 
12989 			/* remove go indication */
12990 			sfmmu_kpm_tsbmtl(&kp->kp_refcntc, &kpmp->khl_lock,
12991 					KPMTSBM_STOP);
12992 		}
12993 		ASSERT(kp->kp_refcntc >= 0);
12994 	}
12995 exit:
12996 	mutex_exit(&kpmp->khl_mutex);
12997 	return (vaddr);
12998 
12999 smallpages_mapin:
13000 	if (uncached == 0) {
13001 		/* tte assembly */
13002 		KPM_TTE_VCACHED(tte.ll, pfn, TTE8K);
13003 	} else {
13004 		/* ASSERT(pp->p_share); XXX use hat_page_getshare */
13005 		pmtx = sfmmu_page_enter(pp);
13006 		PP_SETKPMC(pp);
13007 		sfmmu_page_exit(pmtx);
13008 		/* tte assembly */
13009 		KPM_TTE_VUNCACHED(tte.ll, pfn, TTE8K);
13010 	}
13011 
13012 	/* tsb dropin */
13013 	sfmmu_kpm_load_tsb(vaddr, &tte, MMU_PAGESHIFT);
13014 
13015 	PP2KPMSPG(pp, ksp);
13016 	kpmsp = KPMP_SHASH(ksp);
13017 
13018 	oldval = sfmmu_kpm_stsbmtl(&ksp->kp_mapped, &kpmsp->kshl_lock,
13019 				(uncached) ? KPM_MAPPEDSC : KPM_MAPPEDS);
13020 
13021 	if (oldval != 0)
13022 		panic("sfmmu_kpm_mapin: stale smallpages mapping");
13023 
13024 	return (vaddr);
13025 }
13026 
13027 /*
13028  * Mapout a single page, it is called every time a page changes it's state
13029  * from kpm-mapped to kpm-unmapped. It may not be called, when only a kpm
13030  * instance calls mapout and there are still other instances mapping the
13031  * page. Assumes that the mlist mutex is already grabbed.
13032  *
13033  * Note: In normal mode (no VAC conflict prevention pending) TLB's are
13034  * not flushed. This is the core segkpm behavior to avoid xcalls. It is
13035  * no problem because a translation from a segkpm virtual address to a
13036  * physical address is always the same. The only downside is a slighty
13037  * increased window of vulnerability for misbehaving _kernel_ modules.
13038  */
13039 static void
13040 sfmmu_kpm_mapout(page_t *pp, caddr_t vaddr)
13041 {
13042 	kpm_page_t	*kp;
13043 	kpm_hlk_t	*kpmp;
13044 	int		alias_range;
13045 	kmutex_t	*pmtx;
13046 	kpm_spage_t	*ksp;
13047 	kpm_shlk_t	*kpmsp;
13048 	int		oldval;
13049 
13050 	ASSERT(sfmmu_mlist_held(pp));
13051 	ASSERT(pp->p_kpmref == 0);
13052 
13053 	alias_range = IS_KPM_ALIAS_RANGE(vaddr);
13054 
13055 	if (kpm_smallpages)
13056 		goto smallpages_mapout;
13057 
13058 	PP2KPMPG(pp, kp);
13059 	kpmp = KPMP_HASH(kp);
13060 	mutex_enter(&kpmp->khl_mutex);
13061 
13062 	if (alias_range) {
13063 		ASSERT(PP_ISKPMS(pp) == 0);
13064 		if (kp->kp_refcnta <= 0) {
13065 			panic("sfmmu_kpm_mapout: bad refcnta kp=%p",
13066 				(void *)kp);
13067 		}
13068 
13069 		if (PP_ISTNC(pp))  {
13070 			if (PP_ISKPMC(pp) == 0) {
13071 				/*
13072 				 * Uncached kpm mappings must always have
13073 				 * forced "small page" mode.
13074 				 */
13075 				panic("sfmmu_kpm_mapout: uncached page not "
13076 					"kpm marked");
13077 			}
13078 			sfmmu_kpm_demap_small(vaddr);
13079 
13080 			pmtx = sfmmu_page_enter(pp);
13081 			PP_CLRKPMC(pp);
13082 			sfmmu_page_exit(pmtx);
13083 
13084 			/*
13085 			 * Check if we can resume cached mode. This might
13086 			 * be the case if the kpm mapping was the only
13087 			 * mapping in conflict with other non rule
13088 			 * compliant mappings. The page is no more marked
13089 			 * as kpm mapped, so the conv_tnc path will not
13090 			 * change kpm state.
13091 			 */
13092 			conv_tnc(pp, TTE8K);
13093 
13094 		} else if (PP_ISKPMC(pp) == 0) {
13095 			/* remove TSB entry only */
13096 			sfmmu_kpm_unload_tsb(vaddr, MMU_PAGESHIFT);
13097 
13098 		} else {
13099 			/* already demapped */
13100 			pmtx = sfmmu_page_enter(pp);
13101 			PP_CLRKPMC(pp);
13102 			sfmmu_page_exit(pmtx);
13103 		}
13104 		kp->kp_refcnta--;
13105 		goto exit;
13106 	}
13107 
13108 	if (kp->kp_refcntc <= 0 && kp->kp_refcnts == 0) {
13109 		/*
13110 		 * Fast path / regular case.
13111 		 */
13112 		ASSERT(kp->kp_refcntc >= -1);
13113 		ASSERT(!(pp->p_nrm & (P_KPMC | P_KPMS | P_TNC | P_PNC)));
13114 
13115 		if (kp->kp_refcnt <= 0)
13116 			panic("sfmmu_kpm_mapout: bad refcnt kp=%p", (void *)kp);
13117 
13118 		if (--kp->kp_refcnt == 0) {
13119 			/* remove go indication */
13120 			if (kp->kp_refcntc == -1) {
13121 				sfmmu_kpm_tsbmtl(&kp->kp_refcntc,
13122 					&kpmp->khl_lock, KPMTSBM_STOP);
13123 			}
13124 			ASSERT(kp->kp_refcntc == 0);
13125 
13126 			/* remove TSB entry */
13127 			sfmmu_kpm_unload_tsb(vaddr, MMU_PAGESHIFT4M);
13128 #ifdef	DEBUG
13129 			if (kpm_tlb_flush)
13130 				sfmmu_kpm_demap_tlbs(vaddr, KCONTEXT);
13131 #endif
13132 		}
13133 
13134 	} else {
13135 		/*
13136 		 * The VAC alias path.
13137 		 * We come here if the kpm vaddr is not in any alias_range
13138 		 * and we are unmapping a page within the regular kpm_page
13139 		 * range. The kpm_page either holds conflict pages and/or
13140 		 * is in "small page" mode. If the page is not marked
13141 		 * P_KPMS it couldn't have a valid PAGESIZE sized TSB
13142 		 * entry. Dcache flushing is done lazy and follows the
13143 		 * rules of the regular virtual page coloring scheme.
13144 		 *
13145 		 * Per page states and required actions:
13146 		 *   P_KPMC: remove a kpm mapping that is conflicting.
13147 		 *   P_KPMS: remove a small kpm mapping within a kpm_page.
13148 		 *   P_TNC:  check if we can re-cache the page.
13149 		 *   P_PNC:  we cannot re-cache, sorry.
13150 		 * Per kpm_page:
13151 		 *   kp_refcntc > 0: page is part of a kpm_page with conflicts.
13152 		 *   kp_refcnts > 0: rm a small mapped page within a kpm_page.
13153 		 */
13154 
13155 		if (PP_ISKPMS(pp)) {
13156 			if (kp->kp_refcnts < 1) {
13157 				panic("sfmmu_kpm_mapout: bad refcnts kp=%p",
13158 					(void *)kp);
13159 			}
13160 			sfmmu_kpm_demap_small(vaddr);
13161 
13162 			/*
13163 			 * Check if we can resume cached mode. This might
13164 			 * be the case if the kpm mapping was the only
13165 			 * mapping in conflict with other non rule
13166 			 * compliant mappings. The page is no more marked
13167 			 * as kpm mapped, so the conv_tnc path will not
13168 			 * change kpm state.
13169 			 */
13170 			if (PP_ISTNC(pp))  {
13171 				if (!PP_ISKPMC(pp)) {
13172 					/*
13173 					 * Uncached kpm mappings must always
13174 					 * have forced "small page" mode.
13175 					 */
13176 					panic("sfmmu_kpm_mapout: uncached "
13177 						"page not kpm marked");
13178 				}
13179 				conv_tnc(pp, TTE8K);
13180 			}
13181 			kp->kp_refcnts--;
13182 			kp->kp_refcnt++;
13183 			pmtx = sfmmu_page_enter(pp);
13184 			PP_CLRKPMS(pp);
13185 			sfmmu_page_exit(pmtx);
13186 		}
13187 
13188 		if (PP_ISKPMC(pp)) {
13189 			if (kp->kp_refcntc < 1) {
13190 				panic("sfmmu_kpm_mapout: bad refcntc kp=%p",
13191 					(void *)kp);
13192 			}
13193 			pmtx = sfmmu_page_enter(pp);
13194 			PP_CLRKPMC(pp);
13195 			sfmmu_page_exit(pmtx);
13196 			kp->kp_refcntc--;
13197 		}
13198 
13199 		if (kp->kp_refcnt-- < 1)
13200 			panic("sfmmu_kpm_mapout: bad refcnt kp=%p", (void *)kp);
13201 	}
13202 exit:
13203 	mutex_exit(&kpmp->khl_mutex);
13204 	return;
13205 
13206 smallpages_mapout:
13207 	PP2KPMSPG(pp, ksp);
13208 	kpmsp = KPMP_SHASH(ksp);
13209 
13210 	if (PP_ISKPMC(pp) == 0) {
13211 		oldval = sfmmu_kpm_stsbmtl(&ksp->kp_mapped,
13212 					&kpmsp->kshl_lock, 0);
13213 
13214 		if (oldval != KPM_MAPPEDS) {
13215 			/*
13216 			 * When we're called after sfmmu_kpm_hme_unload,
13217 			 * KPM_MAPPEDSC is valid too.
13218 			 */
13219 			if (oldval != KPM_MAPPEDSC)
13220 				panic("sfmmu_kpm_mapout: incorrect mapping");
13221 		}
13222 
13223 		/* remove TSB entry */
13224 		sfmmu_kpm_unload_tsb(vaddr, MMU_PAGESHIFT);
13225 #ifdef	DEBUG
13226 		if (kpm_tlb_flush)
13227 			sfmmu_kpm_demap_tlbs(vaddr, KCONTEXT);
13228 #endif
13229 
13230 	} else if (PP_ISTNC(pp)) {
13231 		oldval = sfmmu_kpm_stsbmtl(&ksp->kp_mapped,
13232 					&kpmsp->kshl_lock, 0);
13233 
13234 		if (oldval != KPM_MAPPEDSC || PP_ISKPMC(pp) == 0)
13235 			panic("sfmmu_kpm_mapout: inconsistent TNC mapping");
13236 
13237 		sfmmu_kpm_demap_small(vaddr);
13238 
13239 		pmtx = sfmmu_page_enter(pp);
13240 		PP_CLRKPMC(pp);
13241 		sfmmu_page_exit(pmtx);
13242 
13243 		/*
13244 		 * Check if we can resume cached mode. This might be
13245 		 * the case if the kpm mapping was the only mapping
13246 		 * in conflict with other non rule compliant mappings.
13247 		 * The page is no more marked as kpm mapped, so the
13248 		 * conv_tnc path will not change the kpm state.
13249 		 */
13250 		conv_tnc(pp, TTE8K);
13251 
13252 	} else {
13253 		oldval = sfmmu_kpm_stsbmtl(&ksp->kp_mapped,
13254 					&kpmsp->kshl_lock, 0);
13255 
13256 		if (oldval != KPM_MAPPEDSC)
13257 			panic("sfmmu_kpm_mapout: inconsistent mapping");
13258 
13259 		pmtx = sfmmu_page_enter(pp);
13260 		PP_CLRKPMC(pp);
13261 		sfmmu_page_exit(pmtx);
13262 	}
13263 }
13264 
13265 #define	abs(x)  ((x) < 0 ? -(x) : (x))
13266 
13267 /*
13268  * Determine appropriate kpm mapping address and handle any kpm/hme
13269  * conflicts. Page mapping list and its vcolor parts must be protected.
13270  */
13271 static caddr_t
13272 sfmmu_kpm_getvaddr(page_t *pp, int *kpm_vac_rangep)
13273 {
13274 	int		vcolor, vcolor_pa;
13275 	caddr_t		vaddr;
13276 	uintptr_t	paddr;
13277 
13278 
13279 	ASSERT(sfmmu_mlist_held(pp));
13280 
13281 	paddr = ptob(pp->p_pagenum);
13282 	vcolor_pa = addr_to_vcolor(paddr);
13283 
13284 	if (IS_SWAPFSVP(pp->p_vnode)) {
13285 		vcolor = (PP_NEWPAGE(pp) || PP_ISNC(pp)) ?
13286 		    vcolor_pa : PP_GET_VCOLOR(pp);
13287 	} else {
13288 		vcolor = addr_to_vcolor(pp->p_offset);
13289 	}
13290 
13291 	vaddr = kpm_vbase + paddr;
13292 	*kpm_vac_rangep = 0;
13293 
13294 	if (vcolor_pa != vcolor) {
13295 		*kpm_vac_rangep = abs(vcolor - vcolor_pa);
13296 		vaddr += ((uintptr_t)(vcolor - vcolor_pa) << MMU_PAGESHIFT);
13297 		vaddr += (vcolor_pa > vcolor) ?
13298 			((uintptr_t)vcolor_pa << kpm_size_shift) :
13299 			((uintptr_t)(vcolor - vcolor_pa) << kpm_size_shift);
13300 
13301 		ASSERT(!PP_ISMAPPED_LARGE(pp));
13302 	}
13303 
13304 	if (PP_ISNC(pp))
13305 		return (vaddr);
13306 
13307 	if (PP_NEWPAGE(pp)) {
13308 		PP_SET_VCOLOR(pp, vcolor);
13309 		return (vaddr);
13310 	}
13311 
13312 	if (PP_GET_VCOLOR(pp) == vcolor)
13313 		return (vaddr);
13314 
13315 	ASSERT(!PP_ISMAPPED_KPM(pp));
13316 	sfmmu_kpm_vac_conflict(pp, vaddr);
13317 
13318 	return (vaddr);
13319 }
13320 
13321 /*
13322  * VAC conflict state bit values.
13323  * The following defines are used to make the handling of the
13324  * various input states more concise. For that the kpm states
13325  * per kpm_page and per page are combined in a summary state.
13326  * Each single state has a corresponding bit value in the
13327  * summary state. These defines only apply for kpm large page
13328  * mappings. Within comments the abbreviations "kc, c, ks, s"
13329  * are used as short form of the actual state, e.g. "kc" for
13330  * "kp_refcntc > 0", etc.
13331  */
13332 #define	KPM_KC	0x00000008	/* kpm_page: kp_refcntc > 0 */
13333 #define	KPM_C	0x00000004	/* page: P_KPMC set */
13334 #define	KPM_KS	0x00000002	/* kpm_page: kp_refcnts > 0 */
13335 #define	KPM_S	0x00000001	/* page: P_KPMS set */
13336 
13337 /*
13338  * Summary states used in sfmmu_kpm_fault (KPM_TSBM_*).
13339  * See also more detailed comments within in the sfmmu_kpm_fault switch.
13340  * Abbreviations used:
13341  * CONFL: VAC conflict(s) within a kpm_page.
13342  * MAPS:  Mapped small: Page mapped in using a regular page size kpm mapping.
13343  * RASM:  Re-assembling of a large page mapping possible.
13344  * RPLS:  Replace: TSB miss due to TSB replacement only.
13345  * BRKO:  Breakup Other: A large kpm mapping has to be broken because another
13346  *        page within the kpm_page is already involved in a VAC conflict.
13347  * BRKT:  Breakup This: A large kpm mapping has to be broken, this page is
13348  *        is involved in a VAC conflict.
13349  */
13350 #define	KPM_TSBM_CONFL_GONE	(0)
13351 #define	KPM_TSBM_MAPS_RASM	(KPM_KS)
13352 #define	KPM_TSBM_RPLS_RASM	(KPM_KS | KPM_S)
13353 #define	KPM_TSBM_MAPS_BRKO	(KPM_KC)
13354 #define	KPM_TSBM_MAPS		(KPM_KC | KPM_KS)
13355 #define	KPM_TSBM_RPLS		(KPM_KC | KPM_KS | KPM_S)
13356 #define	KPM_TSBM_MAPS_BRKT	(KPM_KC | KPM_C)
13357 #define	KPM_TSBM_MAPS_CONFL	(KPM_KC | KPM_C | KPM_KS)
13358 #define	KPM_TSBM_RPLS_CONFL	(KPM_KC | KPM_C | KPM_KS | KPM_S)
13359 
13360 /*
13361  * kpm fault handler for mappings with large page size.
13362  */
13363 int
13364 sfmmu_kpm_fault(caddr_t vaddr, struct memseg *mseg, page_t *pp)
13365 {
13366 	int		error;
13367 	pgcnt_t		inx;
13368 	kpm_page_t	*kp;
13369 	tte_t		tte;
13370 	pfn_t		pfn = pp->p_pagenum;
13371 	kpm_hlk_t	*kpmp;
13372 	kmutex_t	*pml;
13373 	int		alias_range;
13374 	int		uncached = 0;
13375 	kmutex_t	*pmtx;
13376 	int		badstate;
13377 	uint_t		tsbmcase;
13378 
13379 	alias_range = IS_KPM_ALIAS_RANGE(vaddr);
13380 
13381 	inx = ptokpmp(kpmptop(ptokpmp(pfn)) - mseg->kpm_pbase);
13382 	if (inx >= mseg->kpm_nkpmpgs) {
13383 		cmn_err(CE_PANIC, "sfmmu_kpm_fault: kpm overflow in memseg "
13384 			"0x%p  pp 0x%p", (void *)mseg, (void *)pp);
13385 	}
13386 
13387 	kp = &mseg->kpm_pages[inx];
13388 	kpmp = KPMP_HASH(kp);
13389 
13390 	pml = sfmmu_mlist_enter(pp);
13391 
13392 	if (!PP_ISMAPPED_KPM(pp)) {
13393 		sfmmu_mlist_exit(pml);
13394 		return (EFAULT);
13395 	}
13396 
13397 	mutex_enter(&kpmp->khl_mutex);
13398 
13399 	if (alias_range) {
13400 		ASSERT(!PP_ISMAPPED_LARGE(pp));
13401 		if (kp->kp_refcnta > 0) {
13402 			if (PP_ISKPMC(pp)) {
13403 				pmtx = sfmmu_page_enter(pp);
13404 				PP_CLRKPMC(pp);
13405 				sfmmu_page_exit(pmtx);
13406 			}
13407 			/*
13408 			 * Check for vcolor conflicts. Return here
13409 			 * w/ either no conflict (fast path), removed hme
13410 			 * mapping chains (unload conflict) or uncached
13411 			 * (uncache conflict). VACaches are cleaned and
13412 			 * p_vcolor and PP_TNC are set accordingly for the
13413 			 * conflict cases.  Drop kpmp for uncache conflict
13414 			 * cases since it will be grabbed within
13415 			 * sfmmu_kpm_page_cache in case of an uncache
13416 			 * conflict.
13417 			 */
13418 			mutex_exit(&kpmp->khl_mutex);
13419 			sfmmu_kpm_vac_conflict(pp, vaddr);
13420 			mutex_enter(&kpmp->khl_mutex);
13421 
13422 			if (PP_ISNC(pp)) {
13423 				uncached = 1;
13424 				pmtx = sfmmu_page_enter(pp);
13425 				PP_SETKPMC(pp);
13426 				sfmmu_page_exit(pmtx);
13427 			}
13428 			goto smallexit;
13429 
13430 		} else {
13431 			/*
13432 			 * We got a tsbmiss on a not active kpm_page range.
13433 			 * Let segkpm_fault decide how to panic.
13434 			 */
13435 			error = EFAULT;
13436 		}
13437 		goto exit;
13438 	}
13439 
13440 	badstate = (kp->kp_refcnt < 0 || kp->kp_refcnts < 0);
13441 	if (kp->kp_refcntc == -1) {
13442 		/*
13443 		 * We should come here only if trap level tsb miss
13444 		 * handler is disabled.
13445 		 */
13446 		badstate |= (kp->kp_refcnt == 0 || kp->kp_refcnts > 0 ||
13447 			PP_ISKPMC(pp) || PP_ISKPMS(pp) || PP_ISNC(pp));
13448 
13449 		if (badstate == 0)
13450 			goto largeexit;
13451 	}
13452 
13453 	if (badstate || kp->kp_refcntc < 0)
13454 		goto badstate_exit;
13455 
13456 	/*
13457 	 * Combine the per kpm_page and per page kpm VAC states to
13458 	 * a summary state in order to make the kpm fault handling
13459 	 * more concise.
13460 	 */
13461 	tsbmcase = (((kp->kp_refcntc > 0) ? KPM_KC : 0) |
13462 			((kp->kp_refcnts > 0) ? KPM_KS : 0) |
13463 			(PP_ISKPMC(pp) ? KPM_C : 0) |
13464 			(PP_ISKPMS(pp) ? KPM_S : 0));
13465 
13466 	switch (tsbmcase) {
13467 	case KPM_TSBM_CONFL_GONE:		/* - - - - */
13468 		/*
13469 		 * That's fine, we either have no more vac conflict in
13470 		 * this kpm page or someone raced in and has solved the
13471 		 * vac conflict for us -- call sfmmu_kpm_vac_conflict
13472 		 * to take care for correcting the vcolor and flushing
13473 		 * the dcache if required.
13474 		 */
13475 		mutex_exit(&kpmp->khl_mutex);
13476 		sfmmu_kpm_vac_conflict(pp, vaddr);
13477 		mutex_enter(&kpmp->khl_mutex);
13478 
13479 		if (PP_ISNC(pp) || kp->kp_refcnt <= 0 ||
13480 		    addr_to_vcolor(vaddr) != PP_GET_VCOLOR(pp)) {
13481 			panic("sfmmu_kpm_fault: inconsistent CONFL_GONE "
13482 				"state, pp=%p", (void *)pp);
13483 		}
13484 		goto largeexit;
13485 
13486 	case KPM_TSBM_MAPS_RASM:		/* - - ks - */
13487 		/*
13488 		 * All conflicts in this kpm page are gone but there are
13489 		 * already small mappings around, so we also map this
13490 		 * page small. This could be the trigger case for a
13491 		 * small mapping reaper, if this is really needed.
13492 		 * For now fall thru to the KPM_TSBM_MAPS handling.
13493 		 */
13494 
13495 	case KPM_TSBM_MAPS:			/* kc - ks - */
13496 		/*
13497 		 * Large page mapping is already broken, this page is not
13498 		 * conflicting, so map it small. Call sfmmu_kpm_vac_conflict
13499 		 * to take care for correcting the vcolor and flushing
13500 		 * the dcache if required.
13501 		 */
13502 		mutex_exit(&kpmp->khl_mutex);
13503 		sfmmu_kpm_vac_conflict(pp, vaddr);
13504 		mutex_enter(&kpmp->khl_mutex);
13505 
13506 		if (PP_ISNC(pp) || kp->kp_refcnt <= 0 ||
13507 		    addr_to_vcolor(vaddr) != PP_GET_VCOLOR(pp)) {
13508 			panic("sfmmu_kpm_fault:  inconsistent MAPS state, "
13509 				"pp=%p", (void *)pp);
13510 		}
13511 		kp->kp_refcnt--;
13512 		kp->kp_refcnts++;
13513 		pmtx = sfmmu_page_enter(pp);
13514 		PP_SETKPMS(pp);
13515 		sfmmu_page_exit(pmtx);
13516 		goto smallexit;
13517 
13518 	case KPM_TSBM_RPLS_RASM:		/* - - ks s */
13519 		/*
13520 		 * All conflicts in this kpm page are gone but this page
13521 		 * is mapped small. This could be the trigger case for a
13522 		 * small mapping reaper, if this is really needed.
13523 		 * For now we drop it in small again. Fall thru to the
13524 		 * KPM_TSBM_RPLS handling.
13525 		 */
13526 
13527 	case KPM_TSBM_RPLS:			/* kc - ks s */
13528 		/*
13529 		 * Large page mapping is already broken, this page is not
13530 		 * conflicting but already mapped small, so drop it in
13531 		 * small again.
13532 		 */
13533 		if (PP_ISNC(pp) ||
13534 		    addr_to_vcolor(vaddr) != PP_GET_VCOLOR(pp)) {
13535 			panic("sfmmu_kpm_fault:  inconsistent RPLS state, "
13536 				"pp=%p", (void *)pp);
13537 		}
13538 		goto smallexit;
13539 
13540 	case KPM_TSBM_MAPS_BRKO:		/* kc - - - */
13541 		/*
13542 		 * The kpm page where we live in is marked conflicting
13543 		 * but this page is not conflicting. So we have to map it
13544 		 * in small. Call sfmmu_kpm_vac_conflict to take care for
13545 		 * correcting the vcolor and flushing the dcache if required.
13546 		 */
13547 		mutex_exit(&kpmp->khl_mutex);
13548 		sfmmu_kpm_vac_conflict(pp, vaddr);
13549 		mutex_enter(&kpmp->khl_mutex);
13550 
13551 		if (PP_ISNC(pp) || kp->kp_refcnt <= 0 ||
13552 		    addr_to_vcolor(vaddr) != PP_GET_VCOLOR(pp)) {
13553 			panic("sfmmu_kpm_fault:  inconsistent MAPS_BRKO state, "
13554 				"pp=%p", (void *)pp);
13555 		}
13556 		kp->kp_refcnt--;
13557 		kp->kp_refcnts++;
13558 		pmtx = sfmmu_page_enter(pp);
13559 		PP_SETKPMS(pp);
13560 		sfmmu_page_exit(pmtx);
13561 		goto smallexit;
13562 
13563 	case KPM_TSBM_MAPS_BRKT:		/* kc c - - */
13564 	case KPM_TSBM_MAPS_CONFL:		/* kc c ks - */
13565 		if (!PP_ISMAPPED(pp)) {
13566 			/*
13567 			 * We got a tsbmiss on kpm large page range that is
13568 			 * marked to contain vac conflicting pages introduced
13569 			 * by hme mappings. The hme mappings are all gone and
13570 			 * must have bypassed the kpm alias prevention logic.
13571 			 */
13572 			panic("sfmmu_kpm_fault: stale VAC conflict, pp=%p",
13573 				(void *)pp);
13574 		}
13575 
13576 		/*
13577 		 * Check for vcolor conflicts. Return here w/ either no
13578 		 * conflict (fast path), removed hme mapping chains
13579 		 * (unload conflict) or uncached (uncache conflict).
13580 		 * Dcache is cleaned and p_vcolor and P_TNC are set
13581 		 * accordingly. Drop kpmp for uncache conflict cases
13582 		 * since it will be grabbed within sfmmu_kpm_page_cache
13583 		 * in case of an uncache conflict.
13584 		 */
13585 		mutex_exit(&kpmp->khl_mutex);
13586 		sfmmu_kpm_vac_conflict(pp, vaddr);
13587 		mutex_enter(&kpmp->khl_mutex);
13588 
13589 		if (kp->kp_refcnt <= 0)
13590 			panic("sfmmu_kpm_fault: bad refcnt kp=%p", (void *)kp);
13591 
13592 		if (PP_ISNC(pp)) {
13593 			uncached = 1;
13594 		} else {
13595 			/*
13596 			 * When an unload conflict is solved and there are
13597 			 * no other small mappings around, we can resume
13598 			 * largepage mode. Otherwise we have to map or drop
13599 			 * in small. This could be a trigger for a small
13600 			 * mapping reaper when this was the last conflict
13601 			 * within the kpm page and when there are only
13602 			 * other small mappings around.
13603 			 */
13604 			ASSERT(addr_to_vcolor(vaddr) == PP_GET_VCOLOR(pp));
13605 			ASSERT(kp->kp_refcntc > 0);
13606 			kp->kp_refcntc--;
13607 			pmtx = sfmmu_page_enter(pp);
13608 			PP_CLRKPMC(pp);
13609 			sfmmu_page_exit(pmtx);
13610 			ASSERT(PP_ISKPMS(pp) == 0);
13611 			if (kp->kp_refcntc == 0 && kp->kp_refcnts == 0)
13612 				goto largeexit;
13613 		}
13614 
13615 		kp->kp_refcnt--;
13616 		kp->kp_refcnts++;
13617 		pmtx = sfmmu_page_enter(pp);
13618 		PP_SETKPMS(pp);
13619 		sfmmu_page_exit(pmtx);
13620 		goto smallexit;
13621 
13622 	case KPM_TSBM_RPLS_CONFL:		/* kc c ks s */
13623 		if (!PP_ISMAPPED(pp)) {
13624 			/*
13625 			 * We got a tsbmiss on kpm large page range that is
13626 			 * marked to contain vac conflicting pages introduced
13627 			 * by hme mappings. They are all gone and must have
13628 			 * somehow bypassed the kpm alias prevention logic.
13629 			 */
13630 			panic("sfmmu_kpm_fault: stale VAC conflict, pp=%p",
13631 				(void *)pp);
13632 		}
13633 
13634 		/*
13635 		 * This state is only possible for an uncached mapping.
13636 		 */
13637 		if (!PP_ISNC(pp)) {
13638 			panic("sfmmu_kpm_fault: page not uncached, pp=%p",
13639 				(void *)pp);
13640 		}
13641 		uncached = 1;
13642 		goto smallexit;
13643 
13644 	default:
13645 badstate_exit:
13646 		panic("sfmmu_kpm_fault: inconsistent VAC state, vaddr=%p kp=%p "
13647 			"pp=%p", (void *)vaddr, (void *)kp, (void *)pp);
13648 	}
13649 
13650 smallexit:
13651 	/* tte assembly */
13652 	if (uncached == 0)
13653 		KPM_TTE_VCACHED(tte.ll, pfn, TTE8K);
13654 	else
13655 		KPM_TTE_VUNCACHED(tte.ll, pfn, TTE8K);
13656 
13657 	/* tsb dropin */
13658 	sfmmu_kpm_load_tsb(vaddr, &tte, MMU_PAGESHIFT);
13659 
13660 	error = 0;
13661 	goto exit;
13662 
13663 largeexit:
13664 	if (kp->kp_refcnt > 0) {
13665 
13666 		/* tte assembly */
13667 		KPM_TTE_VCACHED(tte.ll, pfn, TTE4M);
13668 
13669 		/* tsb dropin */
13670 		sfmmu_kpm_load_tsb(vaddr, &tte, MMU_PAGESHIFT4M);
13671 
13672 		if (kp->kp_refcntc == 0) {
13673 			/* Set "go" flag for TL tsbmiss handler */
13674 			sfmmu_kpm_tsbmtl(&kp->kp_refcntc, &kpmp->khl_lock,
13675 					KPMTSBM_START);
13676 		}
13677 		ASSERT(kp->kp_refcntc == -1);
13678 		error = 0;
13679 
13680 	} else
13681 		error = EFAULT;
13682 exit:
13683 	mutex_exit(&kpmp->khl_mutex);
13684 	sfmmu_mlist_exit(pml);
13685 	return (error);
13686 }
13687 
13688 /*
13689  * kpm fault handler for mappings with small page size.
13690  */
13691 int
13692 sfmmu_kpm_fault_small(caddr_t vaddr, struct memseg *mseg, page_t *pp)
13693 {
13694 	int		error = 0;
13695 	pgcnt_t		inx;
13696 	kpm_spage_t	*ksp;
13697 	kpm_shlk_t	*kpmsp;
13698 	kmutex_t	*pml;
13699 	pfn_t		pfn = pp->p_pagenum;
13700 	tte_t		tte;
13701 	kmutex_t	*pmtx;
13702 	int		oldval;
13703 
13704 	inx = pfn - mseg->kpm_pbase;
13705 	ksp = &mseg->kpm_spages[inx];
13706 	kpmsp = KPMP_SHASH(ksp);
13707 
13708 	pml = sfmmu_mlist_enter(pp);
13709 
13710 	if (!PP_ISMAPPED_KPM(pp)) {
13711 		sfmmu_mlist_exit(pml);
13712 		return (EFAULT);
13713 	}
13714 
13715 	/*
13716 	 * kp_mapped lookup protected by mlist mutex
13717 	 */
13718 	if (ksp->kp_mapped == KPM_MAPPEDS) {
13719 		/*
13720 		 * Fast path tsbmiss
13721 		 */
13722 		ASSERT(!PP_ISKPMC(pp));
13723 		ASSERT(!PP_ISNC(pp));
13724 
13725 		/* tte assembly */
13726 		KPM_TTE_VCACHED(tte.ll, pfn, TTE8K);
13727 
13728 		/* tsb dropin */
13729 		sfmmu_kpm_load_tsb(vaddr, &tte, MMU_PAGESHIFT);
13730 
13731 	} else if (ksp->kp_mapped == KPM_MAPPEDSC) {
13732 		/*
13733 		 * Got here due to existing or gone kpm/hme VAC conflict.
13734 		 * Recheck for vcolor conflicts. Return here w/ either
13735 		 * no conflict, removed hme mapping chain (unload
13736 		 * conflict) or uncached (uncache conflict). VACaches
13737 		 * are cleaned and p_vcolor and PP_TNC are set accordingly
13738 		 * for the conflict cases.
13739 		 */
13740 		sfmmu_kpm_vac_conflict(pp, vaddr);
13741 
13742 		if (PP_ISNC(pp)) {
13743 			/* ASSERT(pp->p_share); XXX use hat_page_getshare */
13744 
13745 			/* tte assembly */
13746 			KPM_TTE_VUNCACHED(tte.ll, pfn, TTE8K);
13747 
13748 			/* tsb dropin */
13749 			sfmmu_kpm_load_tsb(vaddr, &tte, MMU_PAGESHIFT);
13750 
13751 		} else {
13752 			if (PP_ISKPMC(pp)) {
13753 				pmtx = sfmmu_page_enter(pp);
13754 				PP_CLRKPMC(pp);
13755 				sfmmu_page_exit(pmtx);
13756 			}
13757 
13758 			/* tte assembly */
13759 			KPM_TTE_VCACHED(tte.ll, pfn, TTE8K);
13760 
13761 			/* tsb dropin */
13762 			sfmmu_kpm_load_tsb(vaddr, &tte, MMU_PAGESHIFT);
13763 
13764 			oldval = sfmmu_kpm_stsbmtl(&ksp->kp_mapped,
13765 					&kpmsp->kshl_lock, KPM_MAPPEDS);
13766 
13767 			if (oldval != KPM_MAPPEDSC)
13768 				panic("sfmmu_kpm_fault_small: "
13769 					"stale smallpages mapping");
13770 		}
13771 
13772 	} else {
13773 		/*
13774 		 * We got a tsbmiss on a not active kpm_page range.
13775 		 * Let decide segkpm_fault how to panic.
13776 		 */
13777 		error = EFAULT;
13778 	}
13779 
13780 	sfmmu_mlist_exit(pml);
13781 	return (error);
13782 }
13783 
13784 /*
13785  * Check/handle potential hme/kpm mapping conflicts
13786  */
13787 static void
13788 sfmmu_kpm_vac_conflict(page_t *pp, caddr_t vaddr)
13789 {
13790 	int		vcolor;
13791 	struct sf_hment	*sfhmep;
13792 	struct hat	*tmphat;
13793 	struct sf_hment	*tmphme = NULL;
13794 	struct hme_blk	*hmeblkp;
13795 	tte_t		tte;
13796 
13797 	ASSERT(sfmmu_mlist_held(pp));
13798 
13799 	if (PP_ISNC(pp))
13800 		return;
13801 
13802 	vcolor = addr_to_vcolor(vaddr);
13803 	if (PP_GET_VCOLOR(pp) == vcolor)
13804 		return;
13805 
13806 	/*
13807 	 * There could be no vcolor conflict between a large cached
13808 	 * hme page and a non alias range kpm page (neither large nor
13809 	 * small mapped). So if a hme conflict already exists between
13810 	 * a constituent page of a large hme mapping and a shared small
13811 	 * conflicting hme mapping, both mappings must be already
13812 	 * uncached at this point.
13813 	 */
13814 	ASSERT(!PP_ISMAPPED_LARGE(pp));
13815 
13816 	if (!PP_ISMAPPED(pp)) {
13817 		/*
13818 		 * Previous hme user of page had a different color
13819 		 * but since there are no current users
13820 		 * we just flush the cache and change the color.
13821 		 */
13822 		SFMMU_STAT(sf_pgcolor_conflict);
13823 		sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
13824 		PP_SET_VCOLOR(pp, vcolor);
13825 		return;
13826 	}
13827 
13828 	/*
13829 	 * If we get here we have a vac conflict with a current hme
13830 	 * mapping. This must have been established by forcing a wrong
13831 	 * colored mapping, e.g. by using mmap(2) with MAP_FIXED.
13832 	 */
13833 
13834 	/*
13835 	 * Check if any mapping is in same as or if it is locked
13836 	 * since in that case we need to uncache.
13837 	 */
13838 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
13839 		tmphme = sfhmep->hme_next;
13840 		hmeblkp = sfmmu_hmetohblk(sfhmep);
13841 		if (hmeblkp->hblk_xhat_bit)
13842 			continue;
13843 		tmphat = hblktosfmmu(hmeblkp);
13844 		sfmmu_copytte(&sfhmep->hme_tte, &tte);
13845 		ASSERT(TTE_IS_VALID(&tte));
13846 		if ((tmphat == ksfmmup) || hmeblkp->hblk_lckcnt) {
13847 			/*
13848 			 * We have an uncache conflict
13849 			 */
13850 			SFMMU_STAT(sf_uncache_conflict);
13851 			sfmmu_page_cache_array(pp, HAT_TMPNC, CACHE_FLUSH, 1);
13852 			return;
13853 		}
13854 	}
13855 
13856 	/*
13857 	 * We have an unload conflict
13858 	 */
13859 	SFMMU_STAT(sf_unload_conflict);
13860 
13861 	for (sfhmep = pp->p_mapping; sfhmep; sfhmep = tmphme) {
13862 		tmphme = sfhmep->hme_next;
13863 		hmeblkp = sfmmu_hmetohblk(sfhmep);
13864 		if (hmeblkp->hblk_xhat_bit)
13865 			continue;
13866 		(void) sfmmu_pageunload(pp, sfhmep, TTE8K);
13867 	}
13868 
13869 	/*
13870 	 * Unloads only does tlb flushes so we need to flush the
13871 	 * dcache vcolor here.
13872 	 */
13873 	sfmmu_cache_flush(pp->p_pagenum, PP_GET_VCOLOR(pp));
13874 	PP_SET_VCOLOR(pp, vcolor);
13875 }
13876 
13877 /*
13878  * Remove all kpm mappings using kpme's for pp and check that
13879  * all kpm mappings (w/ and w/o kpme's) are gone.
13880  */
13881 static void
13882 sfmmu_kpm_pageunload(page_t *pp)
13883 {
13884 	caddr_t		vaddr;
13885 	struct kpme	*kpme, *nkpme;
13886 
13887 	ASSERT(pp != NULL);
13888 	ASSERT(pp->p_kpmref);
13889 	ASSERT(sfmmu_mlist_held(pp));
13890 
13891 	vaddr = hat_kpm_page2va(pp, 1);
13892 
13893 	for (kpme = pp->p_kpmelist; kpme; kpme = nkpme) {
13894 		ASSERT(kpme->kpe_page == pp);
13895 
13896 		if (pp->p_kpmref == 0)
13897 			panic("sfmmu_kpm_pageunload: stale p_kpmref pp=%p "
13898 				"kpme=%p", (void *)pp, (void *)kpme);
13899 
13900 		nkpme = kpme->kpe_next;
13901 
13902 		/* Add instance callback here here if needed later */
13903 		sfmmu_kpme_sub(kpme, pp);
13904 	}
13905 
13906 	/*
13907 	 * Also correct after mixed kpme/nonkpme mappings. If nonkpme
13908 	 * segkpm clients have unlocked the page and forgot to mapout
13909 	 * we panic here.
13910 	 */
13911 	if (pp->p_kpmref != 0)
13912 		panic("sfmmu_kpm_pageunload: bad refcnt pp=%p", (void *)pp);
13913 
13914 	sfmmu_kpm_mapout(pp, vaddr);
13915 }
13916 
13917 /*
13918  * Remove a large kpm mapping from kernel TSB and all TLB's.
13919  */
13920 static void
13921 sfmmu_kpm_demap_large(caddr_t vaddr)
13922 {
13923 	sfmmu_kpm_unload_tsb(vaddr, MMU_PAGESHIFT4M);
13924 	sfmmu_kpm_demap_tlbs(vaddr, KCONTEXT);
13925 }
13926 
13927 /*
13928  * Remove a small kpm mapping from kernel TSB and all TLB's.
13929  */
13930 static void
13931 sfmmu_kpm_demap_small(caddr_t vaddr)
13932 {
13933 	sfmmu_kpm_unload_tsb(vaddr, MMU_PAGESHIFT);
13934 	sfmmu_kpm_demap_tlbs(vaddr, KCONTEXT);
13935 }
13936 
13937 /*
13938  * Demap a kpm mapping in all TLB's.
13939  */
13940 static void
13941 sfmmu_kpm_demap_tlbs(caddr_t vaddr, int ctxnum)
13942 {
13943 	cpuset_t cpuset;
13944 
13945 	kpreempt_disable();
13946 	cpuset = ksfmmup->sfmmu_cpusran;
13947 	CPUSET_AND(cpuset, cpu_ready_set);
13948 	CPUSET_DEL(cpuset, CPU->cpu_id);
13949 	SFMMU_XCALL_STATS(ctxnum);
13950 	xt_some(cpuset, vtag_flushpage_tl1, (uint64_t)vaddr, ctxnum);
13951 	vtag_flushpage(vaddr, ctxnum);
13952 	kpreempt_enable();
13953 }
13954 
13955 /*
13956  * Summary states used in sfmmu_kpm_vac_unload (KPM_VUL__*).
13957  * See also more detailed comments within in the sfmmu_kpm_vac_unload switch.
13958  * Abbreviations used:
13959  * BIG:   Large page kpm mapping in use.
13960  * CONFL: VAC conflict(s) within a kpm_page.
13961  * INCR:  Count of conflicts within a kpm_page is going to be incremented.
13962  * DECR:  Count of conflicts within a kpm_page is going to be decremented.
13963  * UNMAP_SMALL: A small (regular page size) mapping is going to be unmapped.
13964  * TNC:   Temporary non cached: a kpm mapped page is mapped in TNC state.
13965  */
13966 #define	KPM_VUL_BIG		(0)
13967 #define	KPM_VUL_CONFL_INCR1	(KPM_KS)
13968 #define	KPM_VUL_UNMAP_SMALL1	(KPM_KS | KPM_S)
13969 #define	KPM_VUL_CONFL_INCR2	(KPM_KC)
13970 #define	KPM_VUL_CONFL_INCR3	(KPM_KC | KPM_KS)
13971 #define	KPM_VUL_UNMAP_SMALL2	(KPM_KC | KPM_KS | KPM_S)
13972 #define	KPM_VUL_CONFL_DECR1	(KPM_KC | KPM_C)
13973 #define	KPM_VUL_CONFL_DECR2	(KPM_KC | KPM_C | KPM_KS)
13974 #define	KPM_VUL_TNC		(KPM_KC | KPM_C | KPM_KS | KPM_S)
13975 
13976 /*
13977  * Handle VAC unload conflicts introduced by hme mappings or vice
13978  * versa when a hme conflict mapping is replaced by a non conflict
13979  * one. Perform actions and state transitions according to the
13980  * various page and kpm_page entry states. VACache flushes are in
13981  * the responsibiliy of the caller. We still hold the mlist lock.
13982  */
13983 static void
13984 sfmmu_kpm_vac_unload(page_t *pp, caddr_t vaddr)
13985 {
13986 	kpm_page_t	*kp;
13987 	kpm_hlk_t	*kpmp;
13988 	caddr_t		kpmvaddr = hat_kpm_page2va(pp, 1);
13989 	int		newcolor;
13990 	kmutex_t	*pmtx;
13991 	uint_t		vacunlcase;
13992 	int		badstate = 0;
13993 	kpm_spage_t	*ksp;
13994 	kpm_shlk_t	*kpmsp;
13995 
13996 	ASSERT(PAGE_LOCKED(pp));
13997 	ASSERT(sfmmu_mlist_held(pp));
13998 	ASSERT(!PP_ISNC(pp));
13999 
14000 	newcolor = addr_to_vcolor(kpmvaddr) != addr_to_vcolor(vaddr);
14001 	if (kpm_smallpages)
14002 		goto smallpages_vac_unload;
14003 
14004 	PP2KPMPG(pp, kp);
14005 	kpmp = KPMP_HASH(kp);
14006 	mutex_enter(&kpmp->khl_mutex);
14007 
14008 	if (IS_KPM_ALIAS_RANGE(kpmvaddr)) {
14009 		if (kp->kp_refcnta < 1) {
14010 			panic("sfmmu_kpm_vac_unload: bad refcnta kpm_page=%p\n",
14011 				(void *)kp);
14012 		}
14013 
14014 		if (PP_ISKPMC(pp) == 0) {
14015 			if (newcolor == 0)
14016 				goto exit;
14017 			sfmmu_kpm_demap_small(kpmvaddr);
14018 			pmtx = sfmmu_page_enter(pp);
14019 			PP_SETKPMC(pp);
14020 			sfmmu_page_exit(pmtx);
14021 
14022 		} else if (newcolor == 0) {
14023 			pmtx = sfmmu_page_enter(pp);
14024 			PP_CLRKPMC(pp);
14025 			sfmmu_page_exit(pmtx);
14026 
14027 		} else {
14028 			badstate++;
14029 		}
14030 
14031 		goto exit;
14032 	}
14033 
14034 	badstate = (kp->kp_refcnt < 0 || kp->kp_refcnts < 0);
14035 	if (kp->kp_refcntc == -1) {
14036 		/*
14037 		 * We should come here only if trap level tsb miss
14038 		 * handler is disabled.
14039 		 */
14040 		badstate |= (kp->kp_refcnt == 0 || kp->kp_refcnts > 0 ||
14041 			PP_ISKPMC(pp) || PP_ISKPMS(pp) || PP_ISNC(pp));
14042 	} else {
14043 		badstate |= (kp->kp_refcntc < 0);
14044 	}
14045 
14046 	if (badstate)
14047 		goto exit;
14048 
14049 	if (PP_ISKPMC(pp) == 0 && newcolor == 0) {
14050 		ASSERT(PP_ISKPMS(pp) == 0);
14051 		goto exit;
14052 	}
14053 
14054 	/*
14055 	 * Combine the per kpm_page and per page kpm VAC states
14056 	 * to a summary state in order to make the vac unload
14057 	 * handling more concise.
14058 	 */
14059 	vacunlcase = (((kp->kp_refcntc > 0) ? KPM_KC : 0) |
14060 			((kp->kp_refcnts > 0) ? KPM_KS : 0) |
14061 			(PP_ISKPMC(pp) ? KPM_C : 0) |
14062 			(PP_ISKPMS(pp) ? KPM_S : 0));
14063 
14064 	switch (vacunlcase) {
14065 	case KPM_VUL_BIG:				/* - - - - */
14066 		/*
14067 		 * Have to breakup the large page mapping to be
14068 		 * able to handle the conflicting hme vaddr.
14069 		 */
14070 		if (kp->kp_refcntc == -1) {
14071 			/* remove go indication */
14072 			sfmmu_kpm_tsbmtl(&kp->kp_refcntc,
14073 					&kpmp->khl_lock, KPMTSBM_STOP);
14074 		}
14075 		sfmmu_kpm_demap_large(kpmvaddr);
14076 
14077 		ASSERT(kp->kp_refcntc == 0);
14078 		kp->kp_refcntc++;
14079 		pmtx = sfmmu_page_enter(pp);
14080 		PP_SETKPMC(pp);
14081 		sfmmu_page_exit(pmtx);
14082 		break;
14083 
14084 	case KPM_VUL_UNMAP_SMALL1:			/* -  - ks s */
14085 	case KPM_VUL_UNMAP_SMALL2:			/* kc - ks s */
14086 		/*
14087 		 * New conflict w/ an active kpm page, actually mapped
14088 		 * in by small TSB/TLB entries. Remove the mapping and
14089 		 * update states.
14090 		 */
14091 		ASSERT(newcolor);
14092 		sfmmu_kpm_demap_small(kpmvaddr);
14093 		kp->kp_refcnts--;
14094 		kp->kp_refcnt++;
14095 		kp->kp_refcntc++;
14096 		pmtx = sfmmu_page_enter(pp);
14097 		PP_CLRKPMS(pp);
14098 		PP_SETKPMC(pp);
14099 		sfmmu_page_exit(pmtx);
14100 		break;
14101 
14102 	case KPM_VUL_CONFL_INCR1:			/* -  - ks - */
14103 	case KPM_VUL_CONFL_INCR2:			/* kc - -  - */
14104 	case KPM_VUL_CONFL_INCR3:			/* kc - ks - */
14105 		/*
14106 		 * New conflict on a active kpm mapped page not yet in
14107 		 * TSB/TLB. Mark page and increment the kpm_page conflict
14108 		 * count.
14109 		 */
14110 		ASSERT(newcolor);
14111 		kp->kp_refcntc++;
14112 		pmtx = sfmmu_page_enter(pp);
14113 		PP_SETKPMC(pp);
14114 		sfmmu_page_exit(pmtx);
14115 		break;
14116 
14117 	case KPM_VUL_CONFL_DECR1:			/* kc c -  - */
14118 	case KPM_VUL_CONFL_DECR2:			/* kc c ks - */
14119 		/*
14120 		 * A conflicting hme mapping is removed for an active
14121 		 * kpm page not yet in TSB/TLB. Unmark page and decrement
14122 		 * the kpm_page conflict count.
14123 		 */
14124 		ASSERT(newcolor == 0);
14125 		kp->kp_refcntc--;
14126 		pmtx = sfmmu_page_enter(pp);
14127 		PP_CLRKPMC(pp);
14128 		sfmmu_page_exit(pmtx);
14129 		break;
14130 
14131 	case KPM_VUL_TNC:				/* kc c ks s */
14132 		cmn_err(CE_NOTE, "sfmmu_kpm_vac_unload: "
14133 			"page not in NC state");
14134 		/* FALLTHRU */
14135 
14136 	default:
14137 		badstate++;
14138 	}
14139 exit:
14140 	if (badstate) {
14141 		panic("sfmmu_kpm_vac_unload: inconsistent VAC state, "
14142 			"kpmvaddr=%p kp=%p pp=%p",
14143 			(void *)kpmvaddr, (void *)kp, (void *)pp);
14144 	}
14145 	mutex_exit(&kpmp->khl_mutex);
14146 
14147 	return;
14148 
14149 smallpages_vac_unload:
14150 	if (newcolor == 0)
14151 		return;
14152 
14153 	PP2KPMSPG(pp, ksp);
14154 	kpmsp = KPMP_SHASH(ksp);
14155 
14156 	if (PP_ISKPMC(pp) == 0) {
14157 		if (ksp->kp_mapped == KPM_MAPPEDS) {
14158 			/*
14159 			 * Stop TL tsbmiss handling
14160 			 */
14161 			(void) sfmmu_kpm_stsbmtl(&ksp->kp_mapped,
14162 					&kpmsp->kshl_lock, KPM_MAPPEDSC);
14163 
14164 			sfmmu_kpm_demap_small(kpmvaddr);
14165 
14166 		} else if (ksp->kp_mapped != KPM_MAPPEDSC) {
14167 			panic("sfmmu_kpm_vac_unload: inconsistent mapping");
14168 		}
14169 
14170 		pmtx = sfmmu_page_enter(pp);
14171 		PP_SETKPMC(pp);
14172 		sfmmu_page_exit(pmtx);
14173 
14174 	} else {
14175 		if (ksp->kp_mapped != KPM_MAPPEDSC)
14176 			panic("sfmmu_kpm_vac_unload: inconsistent mapping");
14177 	}
14178 }
14179 
14180 /*
14181  * Page is marked to be in VAC conflict to an existing kpm mapping
14182  * or is kpm mapped using only the regular pagesize. Called from
14183  * sfmmu_hblk_unload when a mlist is completely removed.
14184  */
14185 static void
14186 sfmmu_kpm_hme_unload(page_t *pp)
14187 {
14188 	/* tte assembly */
14189 	kpm_page_t	*kp;
14190 	kpm_hlk_t	*kpmp;
14191 	caddr_t		vaddr;
14192 	kmutex_t	*pmtx;
14193 	uint_t		flags;
14194 	kpm_spage_t	*ksp;
14195 
14196 	ASSERT(sfmmu_mlist_held(pp));
14197 	ASSERT(PP_ISMAPPED_KPM(pp));
14198 
14199 	flags = pp->p_nrm & (P_KPMC | P_KPMS);
14200 	if (kpm_smallpages)
14201 		goto smallpages_hme_unload;
14202 
14203 	if (flags == (P_KPMC | P_KPMS)) {
14204 		panic("sfmmu_kpm_hme_unload: page should be uncached");
14205 
14206 	} else if (flags == P_KPMS) {
14207 		/*
14208 		 * Page mapped small but not involved in VAC conflict
14209 		 */
14210 		return;
14211 	}
14212 
14213 	vaddr = hat_kpm_page2va(pp, 1);
14214 
14215 	PP2KPMPG(pp, kp);
14216 	kpmp = KPMP_HASH(kp);
14217 	mutex_enter(&kpmp->khl_mutex);
14218 
14219 	if (IS_KPM_ALIAS_RANGE(vaddr)) {
14220 		if (kp->kp_refcnta < 1) {
14221 			panic("sfmmu_kpm_hme_unload: bad refcnta kpm_page=%p\n",
14222 				(void *)kp);
14223 		}
14224 
14225 	} else {
14226 		if (kp->kp_refcntc < 1) {
14227 			panic("sfmmu_kpm_hme_unload: bad refcntc kpm_page=%p\n",
14228 				(void *)kp);
14229 		}
14230 		kp->kp_refcntc--;
14231 	}
14232 
14233 	pmtx = sfmmu_page_enter(pp);
14234 	PP_CLRKPMC(pp);
14235 	sfmmu_page_exit(pmtx);
14236 
14237 	mutex_exit(&kpmp->khl_mutex);
14238 	return;
14239 
14240 smallpages_hme_unload:
14241 	if (flags != P_KPMC)
14242 		panic("sfmmu_kpm_hme_unload: page should be uncached");
14243 
14244 	vaddr = hat_kpm_page2va(pp, 1);
14245 	PP2KPMSPG(pp, ksp);
14246 
14247 	if (ksp->kp_mapped != KPM_MAPPEDSC)
14248 		panic("sfmmu_kpm_hme_unload: inconsistent mapping");
14249 
14250 	/*
14251 	 * Keep KPM_MAPPEDSC until the next kpm tsbmiss where it
14252 	 * prevents TL tsbmiss handling and force a hat_kpm_fault.
14253 	 * There we can start over again.
14254 	 */
14255 
14256 	pmtx = sfmmu_page_enter(pp);
14257 	PP_CLRKPMC(pp);
14258 	sfmmu_page_exit(pmtx);
14259 }
14260 
14261 /*
14262  * Special hooks for sfmmu_page_cache_array() when changing the
14263  * cacheability of a page. It is used to obey the hat_kpm lock
14264  * ordering (mlist -> kpmp -> spl, and back).
14265  */
14266 static kpm_hlk_t *
14267 sfmmu_kpm_kpmp_enter(page_t *pp, pgcnt_t npages)
14268 {
14269 	kpm_page_t	*kp;
14270 	kpm_hlk_t	*kpmp;
14271 
14272 	ASSERT(sfmmu_mlist_held(pp));
14273 
14274 	if (kpm_smallpages || PP_ISMAPPED_KPM(pp) == 0)
14275 		return (NULL);
14276 
14277 	ASSERT(npages <= kpmpnpgs);
14278 
14279 	PP2KPMPG(pp, kp);
14280 	kpmp = KPMP_HASH(kp);
14281 	mutex_enter(&kpmp->khl_mutex);
14282 
14283 	return (kpmp);
14284 }
14285 
14286 static void
14287 sfmmu_kpm_kpmp_exit(kpm_hlk_t *kpmp)
14288 {
14289 	if (kpm_smallpages || kpmp == NULL)
14290 		return;
14291 
14292 	mutex_exit(&kpmp->khl_mutex);
14293 }
14294 
14295 /*
14296  * Summary states used in sfmmu_kpm_page_cache (KPM_*).
14297  * See also more detailed comments within in the sfmmu_kpm_page_cache switch.
14298  * Abbreviations used:
14299  * UNC:     Input state for an uncache request.
14300  *   BIG:     Large page kpm mapping in use.
14301  *   SMALL:   Page has a small kpm mapping within a kpm_page range.
14302  *   NODEMAP: No demap needed.
14303  *   NOP:     No operation needed on this input state.
14304  * CACHE:   Input state for a re-cache request.
14305  *   MAPS:    Page is in TNC and kpm VAC conflict state and kpm mapped small.
14306  *   NOMAP:   Page is in TNC and kpm VAC conflict state, but not small kpm
14307  *            mapped.
14308  *   NOMAPO:  Page is in TNC and kpm VAC conflict state, but not small kpm
14309  *            mapped. There are also other small kpm mappings within this
14310  *            kpm_page.
14311  */
14312 #define	KPM_UNC_BIG		(0)
14313 #define	KPM_UNC_NODEMAP1	(KPM_KS)
14314 #define	KPM_UNC_SMALL1		(KPM_KS | KPM_S)
14315 #define	KPM_UNC_NODEMAP2	(KPM_KC)
14316 #define	KPM_UNC_NODEMAP3	(KPM_KC | KPM_KS)
14317 #define	KPM_UNC_SMALL2		(KPM_KC | KPM_KS | KPM_S)
14318 #define	KPM_UNC_NOP1		(KPM_KC | KPM_C)
14319 #define	KPM_UNC_NOP2		(KPM_KC | KPM_C | KPM_KS)
14320 #define	KPM_CACHE_NOMAP		(KPM_KC | KPM_C)
14321 #define	KPM_CACHE_NOMAPO	(KPM_KC | KPM_C | KPM_KS)
14322 #define	KPM_CACHE_MAPS		(KPM_KC | KPM_C | KPM_KS | KPM_S)
14323 
14324 /*
14325  * This function is called when the virtual cacheability of a page
14326  * is changed and the page has an actice kpm mapping. The mlist mutex,
14327  * the spl hash lock and the kpmp mutex (if needed) are already grabbed.
14328  */
14329 static void
14330 sfmmu_kpm_page_cache(page_t *pp, int flags, int cache_flush_tag)
14331 {
14332 	kpm_page_t	*kp;
14333 	kpm_hlk_t	*kpmp;
14334 	caddr_t		kpmvaddr;
14335 	int		badstate = 0;
14336 	uint_t		pgcacase;
14337 	kpm_spage_t	*ksp;
14338 	kpm_shlk_t	*kpmsp;
14339 	int		oldval;
14340 
14341 	ASSERT(PP_ISMAPPED_KPM(pp));
14342 	ASSERT(sfmmu_mlist_held(pp));
14343 	ASSERT(sfmmu_page_spl_held(pp));
14344 
14345 	if (flags != HAT_TMPNC && flags != HAT_CACHE)
14346 		panic("sfmmu_kpm_page_cache: bad flags");
14347 
14348 	kpmvaddr = hat_kpm_page2va(pp, 1);
14349 
14350 	if (flags == HAT_TMPNC && cache_flush_tag == CACHE_FLUSH) {
14351 		pfn_t pfn = pp->p_pagenum;
14352 		int vcolor = addr_to_vcolor(kpmvaddr);
14353 		cpuset_t cpuset = cpu_ready_set;
14354 
14355 		/* Flush vcolor in DCache */
14356 		CPUSET_DEL(cpuset, CPU->cpu_id);
14357 		SFMMU_XCALL_STATS(ksfmmup->sfmmu_cnum);
14358 		xt_some(cpuset, vac_flushpage_tl1, pfn, vcolor);
14359 		vac_flushpage(pfn, vcolor);
14360 	}
14361 
14362 	if (kpm_smallpages)
14363 		goto smallpages_page_cache;
14364 
14365 	PP2KPMPG(pp, kp);
14366 	kpmp = KPMP_HASH(kp);
14367 	ASSERT(MUTEX_HELD(&kpmp->khl_mutex));
14368 
14369 	if (IS_KPM_ALIAS_RANGE(kpmvaddr)) {
14370 		if (kp->kp_refcnta < 1) {
14371 			panic("sfmmu_kpm_page_cache: bad refcnta "
14372 				"kpm_page=%p\n", (void *)kp);
14373 		}
14374 		sfmmu_kpm_demap_small(kpmvaddr);
14375 		if (flags == HAT_TMPNC) {
14376 			PP_SETKPMC(pp);
14377 			ASSERT(!PP_ISKPMS(pp));
14378 		} else {
14379 			ASSERT(PP_ISKPMC(pp));
14380 			PP_CLRKPMC(pp);
14381 		}
14382 		goto exit;
14383 	}
14384 
14385 	badstate = (kp->kp_refcnt < 0 || kp->kp_refcnts < 0);
14386 	if (kp->kp_refcntc == -1) {
14387 		/*
14388 		 * We should come here only if trap level tsb miss
14389 		 * handler is disabled.
14390 		 */
14391 		badstate |= (kp->kp_refcnt == 0 || kp->kp_refcnts > 0 ||
14392 			PP_ISKPMC(pp) || PP_ISKPMS(pp) || PP_ISNC(pp));
14393 	} else {
14394 		badstate |= (kp->kp_refcntc < 0);
14395 	}
14396 
14397 	if (badstate)
14398 		goto exit;
14399 
14400 	/*
14401 	 * Combine the per kpm_page and per page kpm VAC states to
14402 	 * a summary state in order to make the VAC cache/uncache
14403 	 * handling more concise.
14404 	 */
14405 	pgcacase = (((kp->kp_refcntc > 0) ? KPM_KC : 0) |
14406 			((kp->kp_refcnts > 0) ? KPM_KS : 0) |
14407 			(PP_ISKPMC(pp) ? KPM_C : 0) |
14408 			(PP_ISKPMS(pp) ? KPM_S : 0));
14409 
14410 	if (flags == HAT_CACHE) {
14411 		switch (pgcacase) {
14412 		case KPM_CACHE_MAPS:			/* kc c ks s */
14413 			sfmmu_kpm_demap_small(kpmvaddr);
14414 			if (kp->kp_refcnts < 1) {
14415 				panic("sfmmu_kpm_page_cache: bad refcnts "
14416 				"kpm_page=%p\n", (void *)kp);
14417 			}
14418 			kp->kp_refcnts--;
14419 			kp->kp_refcnt++;
14420 			PP_CLRKPMS(pp);
14421 			/* FALLTHRU */
14422 
14423 		case KPM_CACHE_NOMAP:			/* kc c -  - */
14424 		case KPM_CACHE_NOMAPO:			/* kc c ks - */
14425 			kp->kp_refcntc--;
14426 			PP_CLRKPMC(pp);
14427 			break;
14428 
14429 		default:
14430 			badstate++;
14431 		}
14432 		goto exit;
14433 	}
14434 
14435 	switch (pgcacase) {
14436 	case KPM_UNC_BIG:				/* - - - - */
14437 		if (kp->kp_refcnt < 1) {
14438 			panic("sfmmu_kpm_page_cache: bad refcnt "
14439 				"kpm_page=%p\n", (void *)kp);
14440 		}
14441 
14442 		/*
14443 		 * Have to breakup the large page mapping in preparation
14444 		 * to the upcoming TNC mode handled by small mappings.
14445 		 * The demap can already be done due to another conflict
14446 		 * within the kpm_page.
14447 		 */
14448 		if (kp->kp_refcntc == -1) {
14449 			/* remove go indication */
14450 			sfmmu_kpm_tsbmtl(&kp->kp_refcntc,
14451 				&kpmp->khl_lock, KPMTSBM_STOP);
14452 		}
14453 		ASSERT(kp->kp_refcntc == 0);
14454 		sfmmu_kpm_demap_large(kpmvaddr);
14455 		kp->kp_refcntc++;
14456 		PP_SETKPMC(pp);
14457 		break;
14458 
14459 	case KPM_UNC_SMALL1:				/* -  - ks s */
14460 	case KPM_UNC_SMALL2:				/* kc - ks s */
14461 		/*
14462 		 * Have to demap an already small kpm mapping in preparation
14463 		 * to the upcoming TNC mode. The demap can already be done
14464 		 * due to another conflict within the kpm_page.
14465 		 */
14466 		sfmmu_kpm_demap_small(kpmvaddr);
14467 		kp->kp_refcntc++;
14468 		kp->kp_refcnts--;
14469 		kp->kp_refcnt++;
14470 		PP_CLRKPMS(pp);
14471 		PP_SETKPMC(pp);
14472 		break;
14473 
14474 	case KPM_UNC_NODEMAP1:				/* -  - ks - */
14475 		/* fallthru */
14476 
14477 	case KPM_UNC_NODEMAP2:				/* kc - -  - */
14478 	case KPM_UNC_NODEMAP3:				/* kc - ks - */
14479 		kp->kp_refcntc++;
14480 		PP_SETKPMC(pp);
14481 		break;
14482 
14483 	case KPM_UNC_NOP1:				/* kc c -  - */
14484 	case KPM_UNC_NOP2:				/* kc c ks - */
14485 		break;
14486 
14487 	default:
14488 		badstate++;
14489 	}
14490 exit:
14491 	if (badstate) {
14492 		panic("sfmmu_kpm_page_cache: inconsistent VAC state "
14493 			"kpmvaddr=%p kp=%p pp=%p", (void *)kpmvaddr,
14494 			(void *)kp, (void *)pp);
14495 	}
14496 	return;
14497 
14498 smallpages_page_cache:
14499 	PP2KPMSPG(pp, ksp);
14500 	kpmsp = KPMP_SHASH(ksp);
14501 
14502 	oldval = sfmmu_kpm_stsbmtl(&ksp->kp_mapped,
14503 				&kpmsp->kshl_lock, KPM_MAPPEDSC);
14504 
14505 	if (!(oldval == KPM_MAPPEDS || oldval == KPM_MAPPEDSC))
14506 		panic("smallpages_page_cache: inconsistent mapping");
14507 
14508 	sfmmu_kpm_demap_small(kpmvaddr);
14509 
14510 	if (flags == HAT_TMPNC) {
14511 		PP_SETKPMC(pp);
14512 		ASSERT(!PP_ISKPMS(pp));
14513 
14514 	} else {
14515 		ASSERT(PP_ISKPMC(pp));
14516 		PP_CLRKPMC(pp);
14517 	}
14518 
14519 	/*
14520 	 * Keep KPM_MAPPEDSC until the next kpm tsbmiss where it
14521 	 * prevents TL tsbmiss handling and force a hat_kpm_fault.
14522 	 * There we can start over again.
14523 	 */
14524 }
14525 
14526 /*
14527  * unused in sfmmu
14528  */
14529 void
14530 hat_dump(void)
14531 {
14532 }
14533 
14534 /*
14535  * Called when a thread is exiting and we have switched to the kernel address
14536  * space.  Perform the same VM initialization resume() uses when switching
14537  * processes.
14538  *
14539  * Note that sfmmu_load_mmustate() is currently a no-op for kernel threads, but
14540  * we call it anyway in case the semantics change in the future.
14541  */
14542 /*ARGSUSED*/
14543 void
14544 hat_thread_exit(kthread_t *thd)
14545 {
14546 	ASSERT(thd->t_procp->p_as == &kas);
14547 
14548 	sfmmu_setctx_sec(KCONTEXT);
14549 	sfmmu_load_mmustate(ksfmmup);
14550 }
14551