xref: /titanic_41/usr/src/uts/common/vm/vm_anon.c (revision f8919bdadda3ebb97bd55cc14a16e0271ed57615)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
27 /*	  All Rights Reserved  	*/
28 
29 /*
30  * University Copyright- Copyright (c) 1982, 1986, 1988
31  * The Regents of the University of California
32  * All Rights Reserved
33  *
34  * University Acknowledgment- Portions of this document are derived from
35  * software developed by the University of California, Berkeley, and its
36  * contributors.
37  */
38 
39 #pragma ident	"%Z%%M%	%I%	%E% SMI"
40 
41 /*
42  * VM - anonymous pages.
43  *
44  * This layer sits immediately above the vm_swap layer.  It manages
45  * physical pages that have no permanent identity in the file system
46  * name space, using the services of the vm_swap layer to allocate
47  * backing storage for these pages.  Since these pages have no external
48  * identity, they are discarded when the last reference is removed.
49  *
50  * An important function of this layer is to manage low-level sharing
51  * of pages that are logically distinct but that happen to be
52  * physically identical (e.g., the corresponding pages of the processes
53  * resulting from a fork before one process or the other changes their
54  * contents).  This pseudo-sharing is present only as an optimization
55  * and is not to be confused with true sharing in which multiple
56  * address spaces deliberately contain references to the same object;
57  * such sharing is managed at a higher level.
58  *
59  * The key data structure here is the anon struct, which contains a
60  * reference count for its associated physical page and a hint about
61  * the identity of that page.  Anon structs typically live in arrays,
62  * with an instance's position in its array determining where the
63  * corresponding backing storage is allocated; however, the swap_xlate()
64  * routine abstracts away this representation information so that the
65  * rest of the anon layer need not know it.  (See the swap layer for
66  * more details on anon struct layout.)
67  *
68  * In the future versions of the system, the association between an
69  * anon struct and its position on backing store will change so that
70  * we don't require backing store all anonymous pages in the system.
71  * This is important for consideration for large memory systems.
72  * We can also use this technique to delay binding physical locations
73  * to anonymous pages until pageout/swapout time where we can make
74  * smarter allocation decisions to improve anonymous klustering.
75  *
76  * Many of the routines defined here take a (struct anon **) argument,
77  * which allows the code at this level to manage anon pages directly,
78  * so that callers can regard anon structs as opaque objects and not be
79  * concerned with assigning or inspecting their contents.
80  *
81  * Clients of this layer refer to anon pages indirectly.  That is, they
82  * maintain arrays of pointers to anon structs rather than maintaining
83  * anon structs themselves.  The (struct anon **) arguments mentioned
84  * above are pointers to entries in these arrays.  It is these arrays
85  * that capture the mapping between offsets within a given segment and
86  * the corresponding anonymous backing storage address.
87  */
88 
89 #ifdef DEBUG
90 #define	ANON_DEBUG
91 #endif
92 
93 #include <sys/types.h>
94 #include <sys/t_lock.h>
95 #include <sys/param.h>
96 #include <sys/systm.h>
97 #include <sys/mman.h>
98 #include <sys/cred.h>
99 #include <sys/thread.h>
100 #include <sys/vnode.h>
101 #include <sys/cpuvar.h>
102 #include <sys/swap.h>
103 #include <sys/cmn_err.h>
104 #include <sys/vtrace.h>
105 #include <sys/kmem.h>
106 #include <sys/sysmacros.h>
107 #include <sys/bitmap.h>
108 #include <sys/vmsystm.h>
109 #include <sys/debug.h>
110 #include <sys/fs/swapnode.h>
111 #include <sys/tnf_probe.h>
112 #include <sys/lgrp.h>
113 #include <sys/policy.h>
114 #include <sys/condvar_impl.h>
115 #include <sys/mutex_impl.h>
116 #include <sys/rctl.h>
117 
118 #include <vm/as.h>
119 #include <vm/hat.h>
120 #include <vm/anon.h>
121 #include <vm/page.h>
122 #include <vm/vpage.h>
123 #include <vm/seg.h>
124 #include <vm/rm.h>
125 
126 #include <fs/fs_subr.h>
127 
128 struct vnode *anon_vp;
129 
130 int anon_debug;
131 
132 kmutex_t	anoninfo_lock;
133 struct		k_anoninfo k_anoninfo;
134 ani_free_t	ani_free_pool[ANI_MAX_POOL];
135 pad_mutex_t	anon_array_lock[ANON_LOCKSIZE];
136 kcondvar_t	anon_array_cv[ANON_LOCKSIZE];
137 
138 /*
139  * Global hash table for (vp, off) -> anon slot
140  */
141 extern	int swap_maxcontig;
142 size_t	anon_hash_size;
143 struct anon **anon_hash;
144 
145 static struct kmem_cache *anon_cache;
146 static struct kmem_cache *anonmap_cache;
147 
148 #ifdef VM_STATS
149 static struct anonvmstats_str {
150 	ulong_t getpages[30];
151 	ulong_t privatepages[10];
152 	ulong_t demotepages[9];
153 	ulong_t decrefpages[9];
154 	ulong_t	dupfillholes[4];
155 	ulong_t freepages[1];
156 } anonvmstats;
157 #endif /* VM_STATS */
158 
159 
160 /*ARGSUSED*/
161 static int
162 anonmap_cache_constructor(void *buf, void *cdrarg, int kmflags)
163 {
164 	struct anon_map *amp = buf;
165 
166 	rw_init(&amp->a_rwlock, NULL, RW_DEFAULT, NULL);
167 	return (0);
168 }
169 
170 /*ARGSUSED1*/
171 static void
172 anonmap_cache_destructor(void *buf, void *cdrarg)
173 {
174 	struct anon_map *amp = buf;
175 
176 	rw_destroy(&amp->a_rwlock);
177 }
178 
179 kmutex_t	anonhash_lock[AH_LOCK_SIZE];
180 kmutex_t	anonpages_hash_lock[AH_LOCK_SIZE];
181 
182 void
183 anon_init(void)
184 {
185 	int i;
186 
187 	anon_hash_size = 1L << highbit(physmem / ANON_HASHAVELEN);
188 
189 	for (i = 0; i < AH_LOCK_SIZE; i++) {
190 		mutex_init(&anonhash_lock[i], NULL, MUTEX_DEFAULT, NULL);
191 		mutex_init(&anonpages_hash_lock[i], NULL, MUTEX_DEFAULT, NULL);
192 	}
193 
194 	for (i = 0; i < ANON_LOCKSIZE; i++) {
195 		mutex_init(&anon_array_lock[i].pad_mutex, NULL,
196 		    MUTEX_DEFAULT, NULL);
197 		cv_init(&anon_array_cv[i], NULL, CV_DEFAULT, NULL);
198 	}
199 
200 	anon_hash = (struct anon **)
201 	    kmem_zalloc(sizeof (struct anon *) * anon_hash_size, KM_SLEEP);
202 	anon_cache = kmem_cache_create("anon_cache", sizeof (struct anon),
203 	    AN_CACHE_ALIGN, NULL, NULL, NULL, NULL, NULL, 0);
204 	anonmap_cache = kmem_cache_create("anonmap_cache",
205 	    sizeof (struct anon_map), 0,
206 	    anonmap_cache_constructor, anonmap_cache_destructor, NULL,
207 	    NULL, NULL, 0);
208 	swap_maxcontig = (1024 * 1024) >> PAGESHIFT;	/* 1MB of pages */
209 
210 	anon_vp = vn_alloc(KM_SLEEP);
211 	vn_setops(anon_vp, swap_vnodeops);
212 	anon_vp->v_type = VREG;
213 	anon_vp->v_flag |= (VISSWAP|VISSWAPFS);
214 }
215 
216 /*
217  * Global anon slot hash table manipulation.
218  */
219 
220 static void
221 anon_addhash(struct anon *ap)
222 {
223 	int index;
224 
225 	ASSERT(MUTEX_HELD(&anonhash_lock[AH_LOCK(ap->an_vp, ap->an_off)]));
226 	index = ANON_HASH(ap->an_vp, ap->an_off);
227 	ap->an_hash = anon_hash[index];
228 	anon_hash[index] = ap;
229 }
230 
231 static void
232 anon_rmhash(struct anon *ap)
233 {
234 	struct anon **app;
235 
236 	ASSERT(MUTEX_HELD(&anonhash_lock[AH_LOCK(ap->an_vp, ap->an_off)]));
237 
238 	for (app = &anon_hash[ANON_HASH(ap->an_vp, ap->an_off)];
239 	    *app; app = &((*app)->an_hash)) {
240 		if (*app == ap) {
241 			*app = ap->an_hash;
242 			break;
243 		}
244 	}
245 }
246 
247 /*
248  * The anon array interfaces. Functions allocating,
249  * freeing array of pointers, and returning/setting
250  * entries in the array of pointers for a given offset.
251  *
252  * Create the list of pointers
253  */
254 struct anon_hdr *
255 anon_create(pgcnt_t npages, int flags)
256 {
257 	struct anon_hdr *ahp;
258 	ulong_t nchunks;
259 	int kmemflags = (flags & ANON_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
260 
261 	if ((ahp = kmem_zalloc(sizeof (struct anon_hdr), kmemflags)) == NULL) {
262 		return (NULL);
263 	}
264 
265 	mutex_init(&ahp->serial_lock, NULL, MUTEX_DEFAULT, NULL);
266 	/*
267 	 * Single level case.
268 	 */
269 	ahp->size = npages;
270 	if (npages <= ANON_CHUNK_SIZE || (flags & ANON_ALLOC_FORCE)) {
271 
272 		if (flags & ANON_ALLOC_FORCE)
273 			ahp->flags |= ANON_ALLOC_FORCE;
274 
275 		ahp->array_chunk = kmem_zalloc(
276 		    ahp->size * sizeof (struct anon *), kmemflags);
277 
278 		if (ahp->array_chunk == NULL) {
279 			kmem_free(ahp, sizeof (struct anon_hdr));
280 			return (NULL);
281 		}
282 	} else {
283 		/*
284 		 * 2 Level case.
285 		 * anon hdr size needs to be rounded off  to be a multiple
286 		 * of ANON_CHUNK_SIZE. This is important as various anon
287 		 * related functions depend on this.
288 		 * NOTE -
289 		 * anon_grow()  makes anon hdr size a multiple of
290 		 * ANON_CHUNK_SIZE.
291 		 * amp size is <= anon hdr size.
292 		 * anon_index + seg_pgs <= anon hdr size.
293 		 */
294 		ahp->size = P2ROUNDUP(npages, ANON_CHUNK_SIZE);
295 		nchunks = ahp->size >> ANON_CHUNK_SHIFT;
296 
297 		ahp->array_chunk = kmem_zalloc(nchunks * sizeof (ulong_t *),
298 		    kmemflags);
299 
300 		if (ahp->array_chunk == NULL) {
301 			kmem_free(ahp, sizeof (struct anon_hdr));
302 			return (NULL);
303 		}
304 	}
305 	return (ahp);
306 }
307 
308 /*
309  * Free the array of pointers
310  */
311 void
312 anon_release(struct anon_hdr *ahp, pgcnt_t npages)
313 {
314 	ulong_t i;
315 	void **ppp;
316 	ulong_t nchunks;
317 
318 	ASSERT(npages <= ahp->size);
319 
320 	/*
321 	 * Single level case.
322 	 */
323 	if (npages <= ANON_CHUNK_SIZE || (ahp->flags & ANON_ALLOC_FORCE)) {
324 		kmem_free(ahp->array_chunk, ahp->size * sizeof (struct anon *));
325 	} else {
326 		/*
327 		 * 2 level case.
328 		 */
329 		nchunks = ahp->size >> ANON_CHUNK_SHIFT;
330 		for (i = 0; i < nchunks; i++) {
331 			ppp = &ahp->array_chunk[i];
332 			if (*ppp != NULL)
333 				kmem_free(*ppp, PAGESIZE);
334 		}
335 		kmem_free(ahp->array_chunk, nchunks * sizeof (ulong_t *));
336 	}
337 	mutex_destroy(&ahp->serial_lock);
338 	kmem_free(ahp, sizeof (struct anon_hdr));
339 }
340 
341 /*
342  * Return the pointer from the list for a
343  * specified anon index.
344  */
345 struct anon *
346 anon_get_ptr(struct anon_hdr *ahp, ulong_t an_idx)
347 {
348 	struct anon **app;
349 
350 	ASSERT(an_idx < ahp->size);
351 
352 	/*
353 	 * Single level case.
354 	 */
355 	if ((ahp->size <= ANON_CHUNK_SIZE) || (ahp->flags & ANON_ALLOC_FORCE)) {
356 		return ((struct anon *)
357 		    ((uintptr_t)ahp->array_chunk[an_idx] & ANON_PTRMASK));
358 	} else {
359 
360 		/*
361 		 * 2 level case.
362 		 */
363 		app = ahp->array_chunk[an_idx >> ANON_CHUNK_SHIFT];
364 		if (app) {
365 			return ((struct anon *)
366 			    ((uintptr_t)app[an_idx & ANON_CHUNK_OFF] &
367 			    ANON_PTRMASK));
368 		} else {
369 			return (NULL);
370 		}
371 	}
372 }
373 
374 /*
375  * Return the anon pointer for the first valid entry in the anon list,
376  * starting from the given index.
377  */
378 struct anon *
379 anon_get_next_ptr(struct anon_hdr *ahp, ulong_t *index)
380 {
381 	struct anon *ap;
382 	struct anon **app;
383 	ulong_t chunkoff;
384 	ulong_t i;
385 	ulong_t j;
386 	pgcnt_t size;
387 
388 	i = *index;
389 	size = ahp->size;
390 
391 	ASSERT(i < size);
392 
393 	if ((size <= ANON_CHUNK_SIZE) || (ahp->flags & ANON_ALLOC_FORCE)) {
394 		/*
395 		 * 1 level case
396 		 */
397 		while (i < size) {
398 			ap = (struct anon *)
399 			    ((uintptr_t)ahp->array_chunk[i] & ANON_PTRMASK);
400 			if (ap) {
401 				*index = i;
402 				return (ap);
403 			}
404 			i++;
405 		}
406 	} else {
407 		/*
408 		 * 2 level case
409 		 */
410 		chunkoff = i & ANON_CHUNK_OFF;
411 		while (i < size) {
412 			app = ahp->array_chunk[i >> ANON_CHUNK_SHIFT];
413 			if (app)
414 				for (j = chunkoff; j < ANON_CHUNK_SIZE; j++) {
415 					ap = (struct anon *)
416 					    ((uintptr_t)app[j] & ANON_PTRMASK);
417 					if (ap) {
418 						*index = i + (j - chunkoff);
419 						return (ap);
420 					}
421 				}
422 			chunkoff = 0;
423 			i = (i + ANON_CHUNK_SIZE) & ~ANON_CHUNK_OFF;
424 		}
425 	}
426 	*index = size;
427 	return (NULL);
428 }
429 
430 /*
431  * Set list entry with a given pointer for a specified offset
432  */
433 int
434 anon_set_ptr(struct anon_hdr *ahp, ulong_t an_idx, struct anon *ap, int flags)
435 {
436 	void		**ppp;
437 	struct anon	**app;
438 	int kmemflags = (flags & ANON_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
439 	uintptr_t	*ap_addr;
440 
441 	ASSERT(an_idx < ahp->size);
442 
443 	/*
444 	 * Single level case.
445 	 */
446 	if (ahp->size <= ANON_CHUNK_SIZE || (ahp->flags & ANON_ALLOC_FORCE)) {
447 		ap_addr = (uintptr_t *)&ahp->array_chunk[an_idx];
448 	} else {
449 
450 		/*
451 		 * 2 level case.
452 		 */
453 		ppp = &ahp->array_chunk[an_idx >> ANON_CHUNK_SHIFT];
454 
455 		ASSERT(ppp != NULL);
456 		if (*ppp == NULL) {
457 			mutex_enter(&ahp->serial_lock);
458 			ppp = &ahp->array_chunk[an_idx >> ANON_CHUNK_SHIFT];
459 			if (*ppp == NULL) {
460 				*ppp = kmem_zalloc(PAGESIZE, kmemflags);
461 				if (*ppp == NULL) {
462 					mutex_exit(&ahp->serial_lock);
463 					return (ENOMEM);
464 				}
465 			}
466 			mutex_exit(&ahp->serial_lock);
467 		}
468 		app = *ppp;
469 		ap_addr = (uintptr_t *)&app[an_idx & ANON_CHUNK_OFF];
470 	}
471 	*ap_addr = (*ap_addr & ~ANON_PTRMASK) | (uintptr_t)ap;
472 	return (0);
473 }
474 
475 /*
476  * Copy anon array into a given new anon array
477  */
478 int
479 anon_copy_ptr(struct anon_hdr *sahp, ulong_t s_idx,
480 	struct anon_hdr *dahp, ulong_t d_idx,
481 	pgcnt_t npages, int flags)
482 {
483 	void **sapp, **dapp;
484 	void *ap;
485 	int kmemflags = (flags & ANON_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
486 
487 	ASSERT((s_idx < sahp->size) && (d_idx < dahp->size));
488 	ASSERT((npages <= sahp->size) && (npages <= dahp->size));
489 
490 	/*
491 	 * Both arrays are 1 level.
492 	 */
493 	if (((sahp->size <= ANON_CHUNK_SIZE) &&
494 	    (dahp->size <= ANON_CHUNK_SIZE)) ||
495 	    ((sahp->flags & ANON_ALLOC_FORCE) &&
496 	    (dahp->flags & ANON_ALLOC_FORCE))) {
497 
498 		bcopy(&sahp->array_chunk[s_idx], &dahp->array_chunk[d_idx],
499 		    npages * sizeof (struct anon *));
500 		return (0);
501 	}
502 
503 	/*
504 	 * Both arrays are 2 levels.
505 	 */
506 	if (sahp->size > ANON_CHUNK_SIZE &&
507 	    dahp->size > ANON_CHUNK_SIZE &&
508 	    ((sahp->flags & ANON_ALLOC_FORCE) == 0) &&
509 	    ((dahp->flags & ANON_ALLOC_FORCE) == 0)) {
510 
511 		ulong_t sapidx, dapidx;
512 		ulong_t *sap, *dap;
513 		ulong_t chknp;
514 
515 		while (npages != 0) {
516 
517 			sapidx = s_idx & ANON_CHUNK_OFF;
518 			dapidx = d_idx & ANON_CHUNK_OFF;
519 			chknp = ANON_CHUNK_SIZE - MAX(sapidx, dapidx);
520 			if (chknp > npages)
521 				chknp = npages;
522 
523 			sapp = &sahp->array_chunk[s_idx >> ANON_CHUNK_SHIFT];
524 			if ((sap = *sapp) != NULL) {
525 				dapp = &dahp->array_chunk[d_idx
526 				    >> ANON_CHUNK_SHIFT];
527 				if ((dap = *dapp) == NULL) {
528 					*dapp = kmem_zalloc(PAGESIZE,
529 					    kmemflags);
530 					if ((dap = *dapp) == NULL)
531 						return (ENOMEM);
532 				}
533 				bcopy((sap + sapidx), (dap + dapidx),
534 				    chknp << ANON_PTRSHIFT);
535 			}
536 			s_idx += chknp;
537 			d_idx += chknp;
538 			npages -= chknp;
539 		}
540 		return (0);
541 	}
542 
543 	/*
544 	 * At least one of the arrays is 2 level.
545 	 */
546 	while (npages--) {
547 		if ((ap = anon_get_ptr(sahp, s_idx)) != NULL) {
548 			ASSERT(!ANON_ISBUSY(anon_get_slot(sahp, s_idx)));
549 			if (anon_set_ptr(dahp, d_idx, ap, flags) == ENOMEM)
550 					return (ENOMEM);
551 		}
552 		s_idx++;
553 		d_idx++;
554 	}
555 	return (0);
556 }
557 
558 
559 /*
560  * ANON_INITBUF is a convenience macro for anon_grow() below. It
561  * takes a buffer dst, which is at least as large as buffer src. It
562  * does a bcopy from src into dst, and then bzeros the extra bytes
563  * of dst. If tail is set, the data in src is tail aligned within
564  * dst instead of head aligned.
565  */
566 
567 #define	ANON_INITBUF(src, srclen, dst, dstsize, tail)			      \
568 	if (tail) {							      \
569 		bzero((dst), (dstsize) - (srclen));			      \
570 		bcopy((src), (char *)(dst) + (dstsize) - (srclen), (srclen)); \
571 	} else {							      \
572 		bcopy((src), (dst), (srclen));				      \
573 		bzero((char *)(dst) + (srclen), (dstsize) - (srclen));	      \
574 	}
575 
576 #define	ANON_1_LEVEL_INC	(ANON_CHUNK_SIZE / 8)
577 #define	ANON_2_LEVEL_INC	(ANON_1_LEVEL_INC * ANON_CHUNK_SIZE)
578 
579 /*
580  * anon_grow() is used to efficiently extend an existing anon array.
581  * startidx_p points to the index into the anon array of the first page
582  * that is in use. oldseg_pgs is the number of pages in use, starting at
583  * *startidx_p. newpages is the number of additional pages desired.
584  *
585  * If startidx_p == NULL, startidx is taken to be 0 and cannot be changed.
586  *
587  * The growth is done by creating a new top level of the anon array,
588  * and (if the array is 2-level) reusing the existing second level arrays.
589  *
590  * flags can be used to specify ANON_NOSLEEP and ANON_GROWDOWN.
591  *
592  * Returns the new number of pages in the anon array.
593  */
594 pgcnt_t
595 anon_grow(struct anon_hdr *ahp, ulong_t *startidx_p, pgcnt_t oldseg_pgs,
596     pgcnt_t newseg_pgs, int flags)
597 {
598 	ulong_t startidx = startidx_p ? *startidx_p : 0;
599 	pgcnt_t oldamp_pgs = ahp->size, newamp_pgs;
600 	pgcnt_t oelems, nelems, totpages;
601 	void **level1;
602 	int kmemflags = (flags & ANON_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
603 	int growdown = (flags & ANON_GROWDOWN);
604 	size_t newarrsz, oldarrsz;
605 	void *level2;
606 
607 	ASSERT(!(startidx_p == NULL && growdown));
608 	ASSERT(startidx + oldseg_pgs <= ahp->size);
609 
610 	/*
611 	 * Determine the total number of pages needed in the new
612 	 * anon array. If growing down, totpages is all pages from
613 	 * startidx through the end of the array, plus <newseg_pgs>
614 	 * pages. If growing up, keep all pages from page 0 through
615 	 * the last page currently in use, plus <newseg_pgs> pages.
616 	 */
617 	if (growdown)
618 		totpages = oldamp_pgs - startidx + newseg_pgs;
619 	else
620 		totpages = startidx + oldseg_pgs + newseg_pgs;
621 
622 	/* If the array is already large enough, just return. */
623 
624 	if (oldamp_pgs >= totpages) {
625 		if (growdown)
626 			*startidx_p = oldamp_pgs - totpages;
627 		return (oldamp_pgs);
628 	}
629 
630 	/*
631 	 * oldamp_pgs/newamp_pgs are the total numbers of pages represented
632 	 * by the corresponding arrays.
633 	 * oelems/nelems are the number of pointers in the top level arrays
634 	 * which may be either level 1 or level 2.
635 	 * Will the new anon array be one level or two levels?
636 	 */
637 	if (totpages <= ANON_CHUNK_SIZE || (ahp->flags & ANON_ALLOC_FORCE)) {
638 		newamp_pgs = P2ROUNDUP(totpages, ANON_1_LEVEL_INC);
639 		oelems = oldamp_pgs;
640 		nelems = newamp_pgs;
641 	} else {
642 		newamp_pgs = P2ROUNDUP(totpages, ANON_2_LEVEL_INC);
643 		oelems = (oldamp_pgs + ANON_CHUNK_OFF) >> ANON_CHUNK_SHIFT;
644 		nelems = newamp_pgs >> ANON_CHUNK_SHIFT;
645 	}
646 
647 	newarrsz = nelems * sizeof (void *);
648 	level1 = kmem_alloc(newarrsz, kmemflags);
649 	if (level1 == NULL)
650 		return (0);
651 
652 	/* Are we converting from a one level to a two level anon array? */
653 
654 	if (newamp_pgs > ANON_CHUNK_SIZE && oldamp_pgs <= ANON_CHUNK_SIZE &&
655 	    !(ahp->flags & ANON_ALLOC_FORCE)) {
656 
657 		/*
658 		 * Yes, we're converting to a two level. Reuse old level 1
659 		 * as new level 2 if it is exactly PAGESIZE. Otherwise
660 		 * alloc a new level 2 and copy the old level 1 data into it.
661 		 */
662 		if (oldamp_pgs == ANON_CHUNK_SIZE) {
663 			level2 = (void *)ahp->array_chunk;
664 		} else {
665 			level2 = kmem_alloc(PAGESIZE, kmemflags);
666 			if (level2 == NULL) {
667 				kmem_free(level1, newarrsz);
668 				return (0);
669 			}
670 			oldarrsz = oldamp_pgs * sizeof (void *);
671 
672 			ANON_INITBUF(ahp->array_chunk, oldarrsz,
673 			    level2, PAGESIZE, growdown);
674 			kmem_free(ahp->array_chunk, oldarrsz);
675 		}
676 		bzero(level1, newarrsz);
677 		if (growdown)
678 			level1[nelems - 1] = level2;
679 		else
680 			level1[0] = level2;
681 	} else {
682 		oldarrsz = oelems * sizeof (void *);
683 
684 		ANON_INITBUF(ahp->array_chunk, oldarrsz,
685 		    level1, newarrsz, growdown);
686 		kmem_free(ahp->array_chunk, oldarrsz);
687 	}
688 
689 	ahp->array_chunk = level1;
690 	ahp->size = newamp_pgs;
691 	if (growdown)
692 		*startidx_p = newamp_pgs - totpages;
693 
694 	return (newamp_pgs);
695 }
696 
697 
698 /*
699  * Called from clock handler to sync ani_free value.
700  */
701 
702 void
703 set_anoninfo(void)
704 {
705 	int	ix;
706 	pgcnt_t	total = 0;
707 
708 	for (ix = 0; ix < ANI_MAX_POOL; ix++) {
709 		total += ani_free_pool[ix].ani_count;
710 	}
711 	k_anoninfo.ani_free = total;
712 }
713 
714 /*
715  * Reserve anon space.
716  *
717  * It's no longer simply a matter of incrementing ani_resv to
718  * reserve swap space, we need to check memory-based as well
719  * as disk-backed (physical) swap.  The following algorithm
720  * is used:
721  * 	Check the space on physical swap
722  * 		i.e. amount needed < ani_max - ani_phys_resv
723  * 	If we are swapping on swapfs check
724  *		amount needed < (availrmem - swapfs_minfree)
725  * Since the algorithm to check for the quantity of swap space is
726  * almost the same as that for reserving it, we'll just use anon_resvmem
727  * with a flag to decrement availrmem.
728  *
729  * Return non-zero on success.
730  */
731 int
732 anon_resvmem(size_t size, boolean_t takemem, zone_t *zone, int tryhard)
733 {
734 	pgcnt_t npages = btopr(size);
735 	pgcnt_t mswap_pages = 0;
736 	pgcnt_t pswap_pages = 0;
737 	proc_t *p = curproc;
738 
739 	if (zone != NULL && takemem) {
740 		/* test zone.max-swap resource control */
741 		mutex_enter(&p->p_lock);
742 		if (rctl_incr_swap(p, zone, ptob(npages)) != 0) {
743 			mutex_exit(&p->p_lock);
744 			return (0);
745 		}
746 		mutex_exit(&p->p_lock);
747 	}
748 	mutex_enter(&anoninfo_lock);
749 
750 	/*
751 	 * pswap_pages is the number of pages we can take from
752 	 * physical (i.e. disk-backed) swap.
753 	 */
754 	ASSERT(k_anoninfo.ani_max >= k_anoninfo.ani_phys_resv);
755 	pswap_pages = k_anoninfo.ani_max - k_anoninfo.ani_phys_resv;
756 
757 	ANON_PRINT(A_RESV,
758 	    ("anon_resvmem: npages %lu takemem %u pswap %lu caller %p\n",
759 	    npages, takemem, pswap_pages, (void *)caller()));
760 
761 	if (npages <= pswap_pages) {
762 		/*
763 		 * we have enough space on a physical swap
764 		 */
765 		if (takemem)
766 			k_anoninfo.ani_phys_resv += npages;
767 		mutex_exit(&anoninfo_lock);
768 		return (1);
769 	} else if (pswap_pages != 0) {
770 		/*
771 		 * we have some space on a physical swap
772 		 */
773 		if (takemem) {
774 			/*
775 			 * use up remainder of phys swap
776 			 */
777 			k_anoninfo.ani_phys_resv += pswap_pages;
778 			ASSERT(k_anoninfo.ani_phys_resv == k_anoninfo.ani_max);
779 		}
780 	}
781 	/*
782 	 * since (npages > pswap_pages) we need mem swap
783 	 * mswap_pages is the number of pages needed from availrmem
784 	 */
785 	ASSERT(npages > pswap_pages);
786 	mswap_pages = npages - pswap_pages;
787 
788 	ANON_PRINT(A_RESV, ("anon_resvmem: need %ld pages from memory\n",
789 	    mswap_pages));
790 
791 	/*
792 	 * priv processes can reserve memory as swap as long as availrmem
793 	 * remains greater than swapfs_minfree; in the case of non-priv
794 	 * processes, memory can be reserved as swap only if availrmem
795 	 * doesn't fall below (swapfs_minfree + swapfs_reserve). Thus,
796 	 * swapfs_reserve amount of memswap is not available to non-priv
797 	 * processes. This protects daemons such as automounter dying
798 	 * as a result of application processes eating away almost entire
799 	 * membased swap. This safeguard becomes useless if apps are run
800 	 * with root access.
801 	 *
802 	 * swapfs_reserve is minimum of 4Mb or 1/16 of physmem.
803 	 *
804 	 */
805 	if (tryhard) {
806 		mutex_exit(&anoninfo_lock);
807 		(void) page_reclaim_mem(mswap_pages,
808 		    swapfs_minfree + swapfs_reserve, 0);
809 		mutex_enter(&anoninfo_lock);
810 	}
811 
812 	mutex_enter(&freemem_lock);
813 	if (availrmem > (swapfs_minfree + swapfs_reserve + mswap_pages) ||
814 	    (availrmem > (swapfs_minfree + mswap_pages) &&
815 	    secpolicy_resource(CRED()) == 0)) {
816 
817 		if (takemem) {
818 			/*
819 			 * Take the memory from the rest of the system.
820 			 */
821 			availrmem -= mswap_pages;
822 			mutex_exit(&freemem_lock);
823 			k_anoninfo.ani_mem_resv += mswap_pages;
824 			ANI_ADD(mswap_pages);
825 			ANON_PRINT((A_RESV | A_MRESV),
826 			    ("anon_resvmem: took %ld pages of availrmem\n",
827 			    mswap_pages));
828 		} else {
829 			mutex_exit(&freemem_lock);
830 		}
831 
832 		ASSERT(k_anoninfo.ani_max >= k_anoninfo.ani_phys_resv);
833 		mutex_exit(&anoninfo_lock);
834 		return (1);
835 
836 	} else {
837 		/*
838 		 * Fail if not enough memory
839 		 */
840 
841 		if (takemem) {
842 			k_anoninfo.ani_phys_resv -= pswap_pages;
843 		}
844 
845 		mutex_exit(&freemem_lock);
846 		mutex_exit(&anoninfo_lock);
847 		ANON_PRINT(A_RESV,
848 		    ("anon_resvmem: not enough space from swapfs\n"));
849 		if (zone != NULL && takemem)
850 			rctl_decr_swap(zone, ptob(npages));
851 		return (0);
852 	}
853 }
854 
855 /*
856  * Give back an anon reservation.
857  */
858 void
859 anon_unresvmem(size_t size, zone_t *zone)
860 {
861 	pgcnt_t npages = btopr(size);
862 	spgcnt_t mem_free_pages = 0;
863 	pgcnt_t phys_free_slots;
864 #ifdef	ANON_DEBUG
865 	pgcnt_t mem_resv;
866 #endif
867 	if (zone != NULL)
868 		rctl_decr_swap(zone, ptob(npages));
869 
870 	mutex_enter(&anoninfo_lock);
871 
872 	ASSERT(k_anoninfo.ani_mem_resv >= k_anoninfo.ani_locked_swap);
873 	/*
874 	 * If some of this reservation belonged to swapfs
875 	 * give it back to availrmem.
876 	 * ani_mem_resv is the amount of availrmem swapfs has reserved.
877 	 * but some of that memory could be locked by segspt so we can only
878 	 * return non locked ani_mem_resv back to availrmem
879 	 */
880 	if (k_anoninfo.ani_mem_resv > k_anoninfo.ani_locked_swap) {
881 		ANON_PRINT((A_RESV | A_MRESV),
882 		    ("anon_unresv: growing availrmem by %ld pages\n",
883 		    MIN(k_anoninfo.ani_mem_resv, npages)));
884 
885 		mem_free_pages = MIN((spgcnt_t)(k_anoninfo.ani_mem_resv -
886 		    k_anoninfo.ani_locked_swap), npages);
887 		mutex_enter(&freemem_lock);
888 		availrmem += mem_free_pages;
889 		mutex_exit(&freemem_lock);
890 		k_anoninfo.ani_mem_resv -= mem_free_pages;
891 
892 		ANI_ADD(-mem_free_pages);
893 	}
894 	/*
895 	 * The remainder of the pages is returned to phys swap
896 	 */
897 	ASSERT(npages >= mem_free_pages);
898 	phys_free_slots = npages - mem_free_pages;
899 
900 	if (phys_free_slots) {
901 		k_anoninfo.ani_phys_resv -= phys_free_slots;
902 	}
903 
904 #ifdef	ANON_DEBUG
905 	mem_resv = k_anoninfo.ani_mem_resv;
906 #endif
907 
908 	ASSERT(k_anoninfo.ani_mem_resv >= k_anoninfo.ani_locked_swap);
909 	ASSERT(k_anoninfo.ani_max >= k_anoninfo.ani_phys_resv);
910 
911 	mutex_exit(&anoninfo_lock);
912 
913 	ANON_PRINT(A_RESV, ("anon_unresv: %lu, tot %lu, caller %p\n",
914 	    npages, mem_resv, (void *)caller()));
915 }
916 
917 /*
918  * Allocate an anon slot and return it with the lock held.
919  */
920 struct anon *
921 anon_alloc(struct vnode *vp, anoff_t off)
922 {
923 	struct anon	*ap;
924 	kmutex_t	*ahm;
925 
926 	ap = kmem_cache_alloc(anon_cache, KM_SLEEP);
927 	if (vp == NULL) {
928 		swap_alloc(ap);
929 	} else {
930 		ap->an_vp = vp;
931 		ap->an_off = off;
932 	}
933 	ap->an_refcnt = 1;
934 	ap->an_pvp = NULL;
935 	ap->an_poff = 0;
936 	ahm = &anonhash_lock[AH_LOCK(ap->an_vp, ap->an_off)];
937 	mutex_enter(ahm);
938 	anon_addhash(ap);
939 	mutex_exit(ahm);
940 	ANI_ADD(-1);
941 	ANON_PRINT(A_ANON, ("anon_alloc: returning ap %p, vp %p\n",
942 	    (void *)ap, (ap ? (void *)ap->an_vp : NULL)));
943 	return (ap);
944 }
945 
946 /*
947  * Decrement the reference count of an anon page.
948  * If reference count goes to zero, free it and
949  * its associated page (if any).
950  */
951 void
952 anon_decref(struct anon *ap)
953 {
954 	page_t *pp;
955 	struct vnode *vp;
956 	anoff_t off;
957 	kmutex_t *ahm;
958 
959 	ahm = &anonhash_lock[AH_LOCK(ap->an_vp, ap->an_off)];
960 	mutex_enter(ahm);
961 	ASSERT(ap->an_refcnt != 0);
962 	if (ap->an_refcnt == 0)
963 		panic("anon_decref: slot count 0");
964 	if (--ap->an_refcnt == 0) {
965 		swap_xlate(ap, &vp, &off);
966 		anon_rmhash(ap);
967 		if (ap->an_pvp != NULL)
968 			swap_phys_free(ap->an_pvp, ap->an_poff, PAGESIZE);
969 		mutex_exit(ahm);
970 
971 		/*
972 		 * If there is a page for this anon slot we will need to
973 		 * call VN_DISPOSE to get rid of the vp association and
974 		 * put the page back on the free list as really free.
975 		 * Acquire the "exclusive" lock to ensure that any
976 		 * pending i/o always completes before the swap slot
977 		 * is freed.
978 		 */
979 		pp = page_lookup(vp, (u_offset_t)off, SE_EXCL);
980 		if (pp != NULL) {
981 			/*LINTED: constant in conditional context */
982 			VN_DISPOSE(pp, B_INVAL, 0, kcred);
983 		}
984 		ANON_PRINT(A_ANON, ("anon_decref: free ap %p, vp %p\n",
985 		    (void *)ap, (void *)ap->an_vp));
986 
987 		kmem_cache_free(anon_cache, ap);
988 
989 		ANI_ADD(1);
990 	} else {
991 		mutex_exit(ahm);
992 	}
993 }
994 
995 
996 /*
997  * check an_refcnt of the root anon slot (anon_index argument is aligned at
998  * seg->s_szc level) to determine whether COW processing is required.
999  * anonpages_hash_lock[] held on the root ap ensures that if root's
1000  * refcnt is 1 all other refcnt's are 1 as well (and they can't increase
1001  * later since this process can't fork while its AS lock is held).
1002  *
1003  * returns 1 if the root anon slot has a refcnt > 1 otherwise returns 0.
1004  */
1005 int
1006 anon_szcshare(struct anon_hdr *ahp, ulong_t anon_index)
1007 {
1008 	struct anon	*ap;
1009 	kmutex_t	*ahmpages = NULL;
1010 
1011 	ap = anon_get_ptr(ahp, anon_index);
1012 	if (ap == NULL)
1013 		return (0);
1014 
1015 	ahmpages = &anonpages_hash_lock[AH_LOCK(ap->an_vp, ap->an_off)];
1016 	mutex_enter(ahmpages);
1017 	ASSERT(ap->an_refcnt >= 1);
1018 	if (ap->an_refcnt == 1) {
1019 		mutex_exit(ahmpages);
1020 		return (0);
1021 	}
1022 	mutex_exit(ahmpages);
1023 	return (1);
1024 }
1025 /*
1026  * Check 'nslots' anon slots for refcnt > 1.
1027  *
1028  * returns 1 if any of the 'nslots' anon slots has a refcnt > 1 otherwise
1029  * returns 0.
1030  */
1031 static int
1032 anon_share(struct anon_hdr *ahp, ulong_t anon_index, pgcnt_t nslots)
1033 {
1034 	struct anon *ap;
1035 
1036 	while (nslots-- > 0) {
1037 		if ((ap = anon_get_ptr(ahp, anon_index)) != NULL &&
1038 		    ap->an_refcnt > 1)
1039 			return (1);
1040 		anon_index++;
1041 	}
1042 
1043 	return (0);
1044 }
1045 
1046 static void
1047 anon_decref_pages(
1048 	struct anon_hdr *ahp,
1049 	ulong_t an_idx,
1050 	uint_t szc)
1051 {
1052 	struct anon *ap = anon_get_ptr(ahp, an_idx);
1053 	kmutex_t *ahmpages = NULL;
1054 	page_t *pp;
1055 	pgcnt_t pgcnt = page_get_pagecnt(szc);
1056 	pgcnt_t i;
1057 	struct vnode *vp;
1058 	anoff_t   off;
1059 	kmutex_t *ahm;
1060 #ifdef DEBUG
1061 	int refcnt = 1;
1062 #endif
1063 
1064 	ASSERT(szc != 0);
1065 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
1066 	ASSERT(IS_P2ALIGNED(an_idx, pgcnt));
1067 	ASSERT(an_idx < ahp->size);
1068 
1069 	if (ahp->size - an_idx < pgcnt) {
1070 		/*
1071 		 * In case of shared mappings total anon map size may not be
1072 		 * the largest page size aligned.
1073 		 */
1074 		pgcnt = ahp->size - an_idx;
1075 	}
1076 
1077 	VM_STAT_ADD(anonvmstats.decrefpages[0]);
1078 
1079 	if (ap != NULL) {
1080 		ahmpages = &anonpages_hash_lock[AH_LOCK(ap->an_vp, ap->an_off)];
1081 		mutex_enter(ahmpages);
1082 		ASSERT((refcnt = ap->an_refcnt) != 0);
1083 		VM_STAT_ADD(anonvmstats.decrefpages[1]);
1084 		if (ap->an_refcnt == 1) {
1085 			VM_STAT_ADD(anonvmstats.decrefpages[2]);
1086 			ASSERT(!anon_share(ahp, an_idx, pgcnt));
1087 			mutex_exit(ahmpages);
1088 			ahmpages = NULL;
1089 		}
1090 	}
1091 
1092 	i = 0;
1093 	while (i < pgcnt) {
1094 		if ((ap = anon_get_ptr(ahp, an_idx + i)) == NULL) {
1095 			ASSERT(refcnt == 1 && ahmpages == NULL);
1096 			i++;
1097 			continue;
1098 		}
1099 		ASSERT(ap->an_refcnt == refcnt);
1100 		ASSERT(ahmpages != NULL || ap->an_refcnt == 1);
1101 		ASSERT(ahmpages == NULL || ap->an_refcnt > 1);
1102 
1103 		if (ahmpages == NULL) {
1104 			swap_xlate(ap, &vp, &off);
1105 			pp = page_lookup(vp, (u_offset_t)off, SE_EXCL);
1106 			if (pp == NULL || pp->p_szc == 0) {
1107 				VM_STAT_ADD(anonvmstats.decrefpages[3]);
1108 				ahm = &anonhash_lock[AH_LOCK(ap->an_vp,
1109 				    ap->an_off)];
1110 				(void) anon_set_ptr(ahp, an_idx + i, NULL,
1111 				    ANON_SLEEP);
1112 				mutex_enter(ahm);
1113 				ap->an_refcnt--;
1114 				ASSERT(ap->an_refcnt == 0);
1115 				anon_rmhash(ap);
1116 				if (ap->an_pvp)
1117 					swap_phys_free(ap->an_pvp, ap->an_poff,
1118 					    PAGESIZE);
1119 				mutex_exit(ahm);
1120 				if (pp == NULL) {
1121 					pp = page_lookup(vp, (u_offset_t)off,
1122 					    SE_EXCL);
1123 					ASSERT(pp == NULL || pp->p_szc == 0);
1124 				}
1125 				if (pp != NULL) {
1126 					VM_STAT_ADD(anonvmstats.decrefpages[4]);
1127 					/*LINTED*/
1128 					VN_DISPOSE(pp, B_INVAL, 0, kcred);
1129 				}
1130 				kmem_cache_free(anon_cache, ap);
1131 				ANI_ADD(1);
1132 				i++;
1133 			} else {
1134 				pgcnt_t j;
1135 				pgcnt_t curpgcnt =
1136 				    page_get_pagecnt(pp->p_szc);
1137 				size_t ppasize = curpgcnt * sizeof (page_t *);
1138 				page_t **ppa = kmem_alloc(ppasize, KM_SLEEP);
1139 				int dispose = 0;
1140 
1141 				VM_STAT_ADD(anonvmstats.decrefpages[5]);
1142 
1143 				ASSERT(pp->p_szc <= szc);
1144 				ASSERT(IS_P2ALIGNED(curpgcnt, curpgcnt));
1145 				ASSERT(IS_P2ALIGNED(i, curpgcnt));
1146 				ASSERT(i + curpgcnt <= pgcnt);
1147 				ASSERT(!(page_pptonum(pp) & (curpgcnt - 1)));
1148 				ppa[0] = pp;
1149 				for (j = i + 1; j < i + curpgcnt; j++) {
1150 					ap = anon_get_ptr(ahp, an_idx + j);
1151 					ASSERT(ap != NULL &&
1152 					    ap->an_refcnt == 1);
1153 					swap_xlate(ap, &vp, &off);
1154 					pp = page_lookup(vp, (u_offset_t)off,
1155 					    SE_EXCL);
1156 					if (pp == NULL)
1157 						panic("anon_decref_pages: "
1158 						    "no page");
1159 
1160 					(void) hat_pageunload(pp,
1161 					    HAT_FORCE_PGUNLOAD);
1162 					ASSERT(pp->p_szc == ppa[0]->p_szc);
1163 					ASSERT(page_pptonum(pp) - 1 ==
1164 					    page_pptonum(ppa[j - i - 1]));
1165 					ppa[j - i] = pp;
1166 					if (ap->an_pvp != NULL &&
1167 					    !vn_matchopval(ap->an_pvp,
1168 					    VOPNAME_DISPOSE,
1169 					    (fs_generic_func_p)fs_dispose))
1170 						dispose = 1;
1171 				}
1172 				for (j = i; j < i + curpgcnt; j++) {
1173 					ap = anon_get_ptr(ahp, an_idx + j);
1174 					ASSERT(ap != NULL &&
1175 					    ap->an_refcnt == 1);
1176 					ahm = &anonhash_lock[AH_LOCK(ap->an_vp,
1177 					    ap->an_off)];
1178 					(void) anon_set_ptr(ahp, an_idx + j,
1179 					    NULL, ANON_SLEEP);
1180 					mutex_enter(ahm);
1181 					ap->an_refcnt--;
1182 					ASSERT(ap->an_refcnt == 0);
1183 					anon_rmhash(ap);
1184 					if (ap->an_pvp)
1185 						swap_phys_free(ap->an_pvp,
1186 						    ap->an_poff, PAGESIZE);
1187 					mutex_exit(ahm);
1188 					kmem_cache_free(anon_cache, ap);
1189 					ANI_ADD(1);
1190 				}
1191 				if (!dispose) {
1192 					VM_STAT_ADD(anonvmstats.decrefpages[6]);
1193 					page_destroy_pages(ppa[0]);
1194 				} else {
1195 					VM_STAT_ADD(anonvmstats.decrefpages[7]);
1196 					for (j = 0; j < curpgcnt; j++) {
1197 						ASSERT(PAGE_EXCL(ppa[j]));
1198 						ppa[j]->p_szc = 0;
1199 					}
1200 					for (j = 0; j < curpgcnt; j++) {
1201 						ASSERT(!hat_page_is_mapped(
1202 						    ppa[j]));
1203 						/*LINTED*/
1204 						VN_DISPOSE(ppa[j], B_INVAL, 0,
1205 						    kcred);
1206 					}
1207 				}
1208 				kmem_free(ppa, ppasize);
1209 				i += curpgcnt;
1210 			}
1211 		} else {
1212 			VM_STAT_ADD(anonvmstats.decrefpages[8]);
1213 			(void) anon_set_ptr(ahp, an_idx + i, NULL, ANON_SLEEP);
1214 			ahm = &anonhash_lock[AH_LOCK(ap->an_vp, ap->an_off)];
1215 			mutex_enter(ahm);
1216 			ap->an_refcnt--;
1217 			mutex_exit(ahm);
1218 			i++;
1219 		}
1220 	}
1221 
1222 	if (ahmpages != NULL) {
1223 		mutex_exit(ahmpages);
1224 	}
1225 }
1226 
1227 /*
1228  * Duplicate references to size bytes worth of anon pages.
1229  * Used when duplicating a segment that contains private anon pages.
1230  * This code assumes that procedure calling this one has already used
1231  * hat_chgprot() to disable write access to the range of addresses that
1232  * that *old actually refers to.
1233  */
1234 void
1235 anon_dup(struct anon_hdr *old, ulong_t old_idx, struct anon_hdr *new,
1236 			ulong_t new_idx, size_t size)
1237 {
1238 	spgcnt_t npages;
1239 	kmutex_t *ahm;
1240 	struct anon *ap;
1241 	ulong_t off;
1242 	ulong_t index;
1243 
1244 	npages = btopr(size);
1245 	while (npages > 0) {
1246 		index = old_idx;
1247 		if ((ap = anon_get_next_ptr(old, &index)) == NULL)
1248 			break;
1249 
1250 		ASSERT(!ANON_ISBUSY(anon_get_slot(old, index)));
1251 		off = index - old_idx;
1252 		npages -= off;
1253 		if (npages <= 0)
1254 			break;
1255 
1256 		(void) anon_set_ptr(new, new_idx + off, ap, ANON_SLEEP);
1257 		ahm = &anonhash_lock[AH_LOCK(ap->an_vp, ap->an_off)];
1258 
1259 		mutex_enter(ahm);
1260 		ap->an_refcnt++;
1261 		mutex_exit(ahm);
1262 
1263 		off++;
1264 		new_idx += off;
1265 		old_idx += off;
1266 		npages--;
1267 	}
1268 }
1269 
1270 /*
1271  * Just like anon_dup but also guarantees there are no holes (unallocated anon
1272  * slots) within any large page region. That means if a large page region is
1273  * empty in the old array it will skip it. If there are 1 or more valid slots
1274  * in the large page region of the old array it will make sure to fill in any
1275  * unallocated ones and also copy them to the new array. If noalloc is 1 large
1276  * page region should either have no valid anon slots or all slots should be
1277  * valid.
1278  */
1279 void
1280 anon_dup_fill_holes(
1281 	struct anon_hdr *old,
1282 	ulong_t old_idx,
1283 	struct anon_hdr *new,
1284 	ulong_t new_idx,
1285 	size_t size,
1286 	uint_t szc,
1287 	int noalloc)
1288 {
1289 	struct anon	*ap;
1290 	spgcnt_t	npages;
1291 	kmutex_t	*ahm, *ahmpages = NULL;
1292 	pgcnt_t		pgcnt, i;
1293 	ulong_t		index, off;
1294 #ifdef DEBUG
1295 	int		refcnt;
1296 #endif
1297 
1298 	ASSERT(szc != 0);
1299 	pgcnt = page_get_pagecnt(szc);
1300 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
1301 	npages = btopr(size);
1302 	ASSERT(IS_P2ALIGNED(npages, pgcnt));
1303 	ASSERT(IS_P2ALIGNED(old_idx, pgcnt));
1304 
1305 	VM_STAT_ADD(anonvmstats.dupfillholes[0]);
1306 
1307 	while (npages > 0) {
1308 		index = old_idx;
1309 
1310 		/*
1311 		 * Find the next valid slot.
1312 		 */
1313 		if (anon_get_next_ptr(old, &index) == NULL)
1314 			break;
1315 
1316 		ASSERT(!ANON_ISBUSY(anon_get_slot(old, index)));
1317 		/*
1318 		 * Now backup index to the beginning of the
1319 		 * current large page region of the old array.
1320 		 */
1321 		index = P2ALIGN(index, pgcnt);
1322 		off = index - old_idx;
1323 		ASSERT(IS_P2ALIGNED(off, pgcnt));
1324 		npages -= off;
1325 		if (npages <= 0)
1326 			break;
1327 
1328 		/*
1329 		 * Fill and copy a large page regions worth
1330 		 * of anon slots.
1331 		 */
1332 		for (i = 0; i < pgcnt; i++) {
1333 			if ((ap = anon_get_ptr(old, index + i)) == NULL) {
1334 				if (noalloc) {
1335 					panic("anon_dup_fill_holes: "
1336 					    "empty anon slot\n");
1337 				}
1338 				VM_STAT_ADD(anonvmstats.dupfillholes[1]);
1339 				ap = anon_alloc(NULL, 0);
1340 				(void) anon_set_ptr(old, index + i, ap,
1341 				    ANON_SLEEP);
1342 			} else if (i == 0) {
1343 				/*
1344 				 * make the increment of all refcnts of all
1345 				 * anon slots of a large page appear atomic by
1346 				 * getting an anonpages_hash_lock for the
1347 				 * first anon slot of a large page.
1348 				 */
1349 				int hash = AH_LOCK(ap->an_vp, ap->an_off);
1350 
1351 				VM_STAT_ADD(anonvmstats.dupfillholes[2]);
1352 
1353 				ahmpages = &anonpages_hash_lock[hash];
1354 				mutex_enter(ahmpages);
1355 				/*LINTED*/
1356 				ASSERT(refcnt = ap->an_refcnt);
1357 
1358 				VM_STAT_COND_ADD(ap->an_refcnt > 1,
1359 				    anonvmstats.dupfillholes[3]);
1360 			}
1361 			(void) anon_set_ptr(new, new_idx + off + i, ap,
1362 			    ANON_SLEEP);
1363 			ahm = &anonhash_lock[AH_LOCK(ap->an_vp, ap->an_off)];
1364 			mutex_enter(ahm);
1365 			ASSERT(ahmpages != NULL || ap->an_refcnt == 1);
1366 			ASSERT(i == 0 || ahmpages == NULL ||
1367 			    refcnt == ap->an_refcnt);
1368 			ap->an_refcnt++;
1369 			mutex_exit(ahm);
1370 		}
1371 		if (ahmpages != NULL) {
1372 			mutex_exit(ahmpages);
1373 			ahmpages = NULL;
1374 		}
1375 		off += pgcnt;
1376 		new_idx += off;
1377 		old_idx += off;
1378 		npages -= pgcnt;
1379 	}
1380 }
1381 
1382 /*
1383  * Used when a segment with a vnode changes szc. similarly to
1384  * anon_dup_fill_holes() makes sure each large page region either has no anon
1385  * slots or all of them. but new slots are created by COWing the file
1386  * pages. on entrance no anon slots should be shared.
1387  */
1388 int
1389 anon_fill_cow_holes(
1390 	struct seg *seg,
1391 	caddr_t addr,
1392 	struct anon_hdr *ahp,
1393 	ulong_t an_idx,
1394 	struct vnode *vp,
1395 	u_offset_t vp_off,
1396 	size_t size,
1397 	uint_t szc,
1398 	uint_t prot,
1399 	struct vpage vpage[],
1400 	struct cred *cred)
1401 {
1402 	struct anon	*ap;
1403 	spgcnt_t	npages;
1404 	pgcnt_t		pgcnt, i;
1405 	ulong_t		index, off;
1406 	int		err = 0;
1407 	int		pageflags = 0;
1408 
1409 	ASSERT(szc != 0);
1410 	pgcnt = page_get_pagecnt(szc);
1411 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
1412 	npages = btopr(size);
1413 	ASSERT(IS_P2ALIGNED(npages, pgcnt));
1414 	ASSERT(IS_P2ALIGNED(an_idx, pgcnt));
1415 
1416 	while (npages > 0) {
1417 		index = an_idx;
1418 
1419 		/*
1420 		 * Find the next valid slot.
1421 		 */
1422 		if (anon_get_next_ptr(ahp, &index) == NULL) {
1423 			break;
1424 		}
1425 
1426 		ASSERT(!ANON_ISBUSY(anon_get_slot(ahp, index)));
1427 		/*
1428 		 * Now backup index to the beginning of the
1429 		 * current large page region of the anon array.
1430 		 */
1431 		index = P2ALIGN(index, pgcnt);
1432 		off = index - an_idx;
1433 		ASSERT(IS_P2ALIGNED(off, pgcnt));
1434 		npages -= off;
1435 		if (npages <= 0)
1436 			break;
1437 		an_idx += off;
1438 		vp_off += ptob(off);
1439 		addr += ptob(off);
1440 		if (vpage != NULL) {
1441 			vpage += off;
1442 		}
1443 
1444 		for (i = 0; i < pgcnt; i++, an_idx++, vp_off += PAGESIZE) {
1445 			if ((ap = anon_get_ptr(ahp, an_idx)) == NULL) {
1446 				page_t *pl[1 + 1];
1447 				page_t *pp;
1448 
1449 				err = VOP_GETPAGE(vp, vp_off, PAGESIZE, NULL,
1450 				    pl, PAGESIZE, seg, addr, S_READ, cred,
1451 				    NULL);
1452 				if (err) {
1453 					break;
1454 				}
1455 				if (vpage != NULL) {
1456 					prot = VPP_PROT(vpage);
1457 					pageflags = VPP_ISPPLOCK(vpage) ?
1458 					    LOCK_PAGE : 0;
1459 				}
1460 				pp = anon_private(&ap, seg, addr, prot, pl[0],
1461 				    pageflags, cred);
1462 				if (pp == NULL) {
1463 					err = ENOMEM;
1464 					break;
1465 				}
1466 				(void) anon_set_ptr(ahp, an_idx, ap,
1467 				    ANON_SLEEP);
1468 				page_unlock(pp);
1469 			}
1470 			ASSERT(ap->an_refcnt == 1);
1471 			addr += PAGESIZE;
1472 			if (vpage != NULL) {
1473 				vpage++;
1474 			}
1475 		}
1476 		npages -= pgcnt;
1477 	}
1478 
1479 	return (err);
1480 }
1481 
1482 /*
1483  * Free a group of "size" anon pages, size in bytes,
1484  * and clear out the pointers to the anon entries.
1485  */
1486 void
1487 anon_free(struct anon_hdr *ahp, ulong_t index, size_t size)
1488 {
1489 	spgcnt_t npages;
1490 	struct anon *ap;
1491 	ulong_t old;
1492 
1493 	npages = btopr(size);
1494 
1495 	while (npages > 0) {
1496 		old = index;
1497 		if ((ap = anon_get_next_ptr(ahp, &index)) == NULL)
1498 			break;
1499 
1500 		ASSERT(!ANON_ISBUSY(anon_get_slot(ahp, index)));
1501 		npages -= index - old;
1502 		if (npages <= 0)
1503 			break;
1504 
1505 		(void) anon_set_ptr(ahp, index, NULL, ANON_SLEEP);
1506 		anon_decref(ap);
1507 		/*
1508 		 * Bump index and decrement page count
1509 		 */
1510 		index++;
1511 		npages--;
1512 	}
1513 }
1514 
1515 void
1516 anon_free_pages(
1517 	struct anon_hdr *ahp,
1518 	ulong_t an_idx,
1519 	size_t size,
1520 	uint_t szc)
1521 {
1522 	spgcnt_t	npages;
1523 	pgcnt_t		pgcnt;
1524 	ulong_t		index, off;
1525 
1526 	ASSERT(szc != 0);
1527 	pgcnt = page_get_pagecnt(szc);
1528 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
1529 	npages = btopr(size);
1530 	ASSERT(IS_P2ALIGNED(npages, pgcnt));
1531 	ASSERT(IS_P2ALIGNED(an_idx, pgcnt));
1532 	ASSERT(an_idx < ahp->size);
1533 
1534 	VM_STAT_ADD(anonvmstats.freepages[0]);
1535 
1536 	while (npages > 0) {
1537 		index = an_idx;
1538 
1539 		/*
1540 		 * Find the next valid slot.
1541 		 */
1542 		if (anon_get_next_ptr(ahp, &index) == NULL)
1543 			break;
1544 
1545 		ASSERT(!ANON_ISBUSY(anon_get_slot(ahp, index)));
1546 		/*
1547 		 * Now backup index to the beginning of the
1548 		 * current large page region of the old array.
1549 		 */
1550 		index = P2ALIGN(index, pgcnt);
1551 		off = index - an_idx;
1552 		ASSERT(IS_P2ALIGNED(off, pgcnt));
1553 		npages -= off;
1554 		if (npages <= 0)
1555 			break;
1556 
1557 		anon_decref_pages(ahp, index, szc);
1558 
1559 		off += pgcnt;
1560 		an_idx += off;
1561 		npages -= pgcnt;
1562 	}
1563 }
1564 
1565 /*
1566  * Make anonymous pages discardable
1567  */
1568 void
1569 anon_disclaim(struct anon_map *amp, ulong_t index, size_t size)
1570 {
1571 	spgcnt_t npages = btopr(size);
1572 	struct anon *ap;
1573 	struct vnode *vp;
1574 	anoff_t off;
1575 	page_t *pp, *root_pp;
1576 	kmutex_t *ahm;
1577 	pgcnt_t pgcnt;
1578 	ulong_t old_idx, idx, i;
1579 	struct anon_hdr *ahp = amp->ahp;
1580 	anon_sync_obj_t cookie;
1581 
1582 	ASSERT(RW_READ_HELD(&amp->a_rwlock));
1583 	pgcnt = 1;
1584 	for (; npages > 0; index = (pgcnt == 1) ? index + 1 :
1585 	    P2ROUNDUP(index + 1, pgcnt), npages -= pgcnt) {
1586 
1587 		/*
1588 		 * get anon pointer and index for the first valid entry
1589 		 * in the anon list, starting from "index"
1590 		 */
1591 		old_idx = index;
1592 		if ((ap = anon_get_next_ptr(ahp, &index)) == NULL)
1593 			break;
1594 
1595 		/*
1596 		 * decrement npages by number of NULL anon slots we skipped
1597 		 */
1598 		npages -= index - old_idx;
1599 		if (npages <= 0)
1600 			break;
1601 
1602 		anon_array_enter(amp, index, &cookie);
1603 		ap = anon_get_ptr(ahp, index);
1604 		ASSERT(ap != NULL);
1605 
1606 		/*
1607 		 * Get anonymous page and try to lock it SE_EXCL;
1608 		 * if we couldn't grab the lock we skip to next page.
1609 		 */
1610 		swap_xlate(ap, &vp, &off);
1611 		pp = page_lookup_nowait(vp, (u_offset_t)off, SE_EXCL);
1612 		if (pp == NULL) {
1613 			segadvstat.MADV_FREE_miss.value.ul++;
1614 			pgcnt = 1;
1615 			anon_array_exit(&cookie);
1616 			continue;
1617 		}
1618 		pgcnt = page_get_pagecnt(pp->p_szc);
1619 
1620 		/*
1621 		 * we cannot free a page which is permanently locked.
1622 		 * The page_struct_lock need not be acquired to examine
1623 		 * these fields since the page has an "exclusive" lock.
1624 		 */
1625 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
1626 			page_unlock(pp);
1627 			segadvstat.MADV_FREE_miss.value.ul++;
1628 			anon_array_exit(&cookie);
1629 			continue;
1630 		}
1631 
1632 		ahm = &anonhash_lock[AH_LOCK(vp, off)];
1633 		mutex_enter(ahm);
1634 		ASSERT(ap->an_refcnt != 0);
1635 		/*
1636 		 * skip this one if copy-on-write is not yet broken.
1637 		 */
1638 		if (ap->an_refcnt > 1) {
1639 			mutex_exit(ahm);
1640 			page_unlock(pp);
1641 			segadvstat.MADV_FREE_miss.value.ul++;
1642 			anon_array_exit(&cookie);
1643 			continue;
1644 		}
1645 
1646 		if (pp->p_szc == 0) {
1647 			pgcnt = 1;
1648 
1649 			/*
1650 			 * free swap slot;
1651 			 */
1652 			if (ap->an_pvp) {
1653 				swap_phys_free(ap->an_pvp, ap->an_poff,
1654 				    PAGESIZE);
1655 				ap->an_pvp = NULL;
1656 				ap->an_poff = 0;
1657 			}
1658 			mutex_exit(ahm);
1659 			segadvstat.MADV_FREE_hit.value.ul++;
1660 
1661 			/*
1662 			 * while we are at it, unload all the translations
1663 			 * and attempt to free the page.
1664 			 */
1665 			(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
1666 			/*LINTED: constant in conditional context */
1667 			VN_DISPOSE(pp, B_FREE, 0, kcred);
1668 			anon_array_exit(&cookie);
1669 			continue;
1670 		}
1671 
1672 		pgcnt = page_get_pagecnt(pp->p_szc);
1673 		if (!IS_P2ALIGNED(index, pgcnt) || npages < pgcnt) {
1674 			if (!page_try_demote_pages(pp)) {
1675 				mutex_exit(ahm);
1676 				page_unlock(pp);
1677 				segadvstat.MADV_FREE_miss.value.ul++;
1678 				anon_array_exit(&cookie);
1679 				continue;
1680 			} else {
1681 				pgcnt = 1;
1682 				if (ap->an_pvp) {
1683 					swap_phys_free(ap->an_pvp,
1684 					    ap->an_poff, PAGESIZE);
1685 					ap->an_pvp = NULL;
1686 					ap->an_poff = 0;
1687 				}
1688 				mutex_exit(ahm);
1689 				(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
1690 				/*LINTED*/
1691 				VN_DISPOSE(pp, B_FREE, 0, kcred);
1692 				segadvstat.MADV_FREE_hit.value.ul++;
1693 				anon_array_exit(&cookie);
1694 				continue;
1695 			}
1696 		}
1697 		mutex_exit(ahm);
1698 		root_pp = pp;
1699 
1700 		/*
1701 		 * try to lock remaining pages
1702 		 */
1703 		for (idx = 1; idx < pgcnt; idx++) {
1704 			pp++;
1705 			if (!page_trylock(pp, SE_EXCL))
1706 				break;
1707 			if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
1708 				page_unlock(pp);
1709 				break;
1710 			}
1711 		}
1712 
1713 		if (idx == pgcnt) {
1714 			for (i = 0; i < pgcnt; i++) {
1715 				ap = anon_get_ptr(ahp, index + i);
1716 				if (ap == NULL)
1717 					break;
1718 				swap_xlate(ap, &vp, &off);
1719 				ahm = &anonhash_lock[AH_LOCK(vp, off)];
1720 				mutex_enter(ahm);
1721 				ASSERT(ap->an_refcnt != 0);
1722 
1723 				/*
1724 				 * skip this one if copy-on-write
1725 				 * is not yet broken.
1726 				 */
1727 				if (ap->an_refcnt > 1) {
1728 					mutex_exit(ahm);
1729 					goto skiplp;
1730 				}
1731 				if (ap->an_pvp) {
1732 					swap_phys_free(ap->an_pvp,
1733 					    ap->an_poff, PAGESIZE);
1734 					ap->an_pvp = NULL;
1735 					ap->an_poff = 0;
1736 				}
1737 				mutex_exit(ahm);
1738 			}
1739 			page_destroy_pages(root_pp);
1740 			segadvstat.MADV_FREE_hit.value.ul += pgcnt;
1741 			anon_array_exit(&cookie);
1742 			continue;
1743 		}
1744 skiplp:
1745 		segadvstat.MADV_FREE_miss.value.ul += pgcnt;
1746 		for (i = 0, pp = root_pp; i < idx; pp++, i++)
1747 			page_unlock(pp);
1748 		anon_array_exit(&cookie);
1749 	}
1750 }
1751 
1752 /*
1753  * Return the kept page(s) and protections back to the segment driver.
1754  */
1755 int
1756 anon_getpage(
1757 	struct anon **app,
1758 	uint_t *protp,
1759 	page_t *pl[],
1760 	size_t plsz,
1761 	struct seg *seg,
1762 	caddr_t addr,
1763 	enum seg_rw rw,
1764 	struct cred *cred)
1765 {
1766 	page_t *pp;
1767 	struct anon *ap = *app;
1768 	struct vnode *vp;
1769 	anoff_t off;
1770 	int err;
1771 	kmutex_t *ahm;
1772 
1773 	swap_xlate(ap, &vp, &off);
1774 
1775 	/*
1776 	 * Lookup the page. If page is being paged in,
1777 	 * wait for it to finish as we must return a list of
1778 	 * pages since this routine acts like the VOP_GETPAGE
1779 	 * routine does.
1780 	 */
1781 	if (pl != NULL && (pp = page_lookup(vp, (u_offset_t)off, SE_SHARED))) {
1782 		ahm = &anonhash_lock[AH_LOCK(ap->an_vp, ap->an_off)];
1783 		mutex_enter(ahm);
1784 		if (ap->an_refcnt == 1)
1785 			*protp = PROT_ALL;
1786 		else
1787 			*protp = PROT_ALL & ~PROT_WRITE;
1788 		mutex_exit(ahm);
1789 		pl[0] = pp;
1790 		pl[1] = NULL;
1791 		return (0);
1792 	}
1793 
1794 	/*
1795 	 * Simply treat it as a vnode fault on the anon vp.
1796 	 */
1797 
1798 	TRACE_3(TR_FAC_VM, TR_ANON_GETPAGE,
1799 	    "anon_getpage:seg %x addr %x vp %x",
1800 	    seg, addr, vp);
1801 
1802 	err = VOP_GETPAGE(vp, (u_offset_t)off, PAGESIZE, protp, pl, plsz,
1803 	    seg, addr, rw, cred, NULL);
1804 
1805 	if (err == 0 && pl != NULL) {
1806 		ahm = &anonhash_lock[AH_LOCK(ap->an_vp, ap->an_off)];
1807 		mutex_enter(ahm);
1808 		if (ap->an_refcnt != 1)
1809 			*protp &= ~PROT_WRITE;	/* make read-only */
1810 		mutex_exit(ahm);
1811 	}
1812 	return (err);
1813 }
1814 
1815 /*
1816  * Creates or returns kept pages to the segment driver.  returns -1 if a large
1817  * page cannot be allocated. returns -2 if some other process has allocated a
1818  * larger page.
1819  *
1820  * For cowfault it will allocate any size pages to fill the requested area to
1821  * avoid partially overwriting anon slots (i.e. sharing only some of the anon
1822  * slots within a large page with other processes). This policy greatly
1823  * simplifies large page freeing (which is only freed when all anon slot
1824  * refcnts are 0).
1825  */
1826 int
1827 anon_map_getpages(
1828 	struct anon_map *amp,
1829 	ulong_t	start_idx,
1830 	uint_t	szc,
1831 	struct seg *seg,
1832 	caddr_t	addr,
1833 	uint_t prot,
1834 	uint_t *protp,
1835 	page_t	*ppa[],
1836 	uint_t	*ppa_szc,
1837 	struct vpage vpage[],
1838 	enum seg_rw rw,
1839 	int brkcow,
1840 	int anypgsz,
1841 	int pgflags,
1842 	struct cred *cred)
1843 {
1844 	pgcnt_t		pgcnt;
1845 	struct anon	*ap;
1846 	struct vnode	*vp;
1847 	anoff_t		off;
1848 	page_t		*pp, *pl[2], *conpp = NULL;
1849 	caddr_t		vaddr;
1850 	ulong_t		pg_idx, an_idx, i;
1851 	spgcnt_t	nreloc = 0;
1852 	int		prealloc = 1;
1853 	int		err, slotcreate;
1854 	uint_t		vpprot;
1855 	int		upsize = (szc < seg->s_szc);
1856 
1857 #if !defined(__i386) && !defined(__amd64)
1858 	ASSERT(seg->s_szc != 0);
1859 #endif
1860 	ASSERT(szc <= seg->s_szc);
1861 	ASSERT(ppa_szc != NULL);
1862 	ASSERT(rw != S_CREATE);
1863 
1864 	*protp = PROT_ALL;
1865 
1866 	VM_STAT_ADD(anonvmstats.getpages[0]);
1867 
1868 	if (szc == 0) {
1869 		VM_STAT_ADD(anonvmstats.getpages[1]);
1870 		if ((ap = anon_get_ptr(amp->ahp, start_idx)) != NULL) {
1871 			err = anon_getpage(&ap, protp, pl, PAGESIZE, seg,
1872 			    addr, rw, cred);
1873 			if (err)
1874 				return (err);
1875 			ppa[0] = pl[0];
1876 			if (brkcow == 0 || (*protp & PROT_WRITE)) {
1877 				VM_STAT_ADD(anonvmstats.getpages[2]);
1878 				if (ppa[0]->p_szc != 0 && upsize) {
1879 					VM_STAT_ADD(anonvmstats.getpages[3]);
1880 					*ppa_szc = MIN(ppa[0]->p_szc,
1881 					    seg->s_szc);
1882 					page_unlock(ppa[0]);
1883 					return (-2);
1884 				}
1885 				return (0);
1886 			}
1887 			panic("anon_map_getpages: cowfault for szc 0");
1888 		} else {
1889 			VM_STAT_ADD(anonvmstats.getpages[4]);
1890 			ppa[0] = anon_zero(seg, addr, &ap, cred);
1891 			if (ppa[0] == NULL)
1892 				return (ENOMEM);
1893 			(void) anon_set_ptr(amp->ahp, start_idx, ap,
1894 			    ANON_SLEEP);
1895 			return (0);
1896 		}
1897 	}
1898 
1899 	pgcnt = page_get_pagecnt(szc);
1900 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
1901 	ASSERT(IS_P2ALIGNED(start_idx, pgcnt));
1902 
1903 	/*
1904 	 * First we check for the case that the requtested large
1905 	 * page or larger page already exists in the system.
1906 	 * Actually we only check if the first constituent page
1907 	 * exists and only preallocate if it's not found.
1908 	 */
1909 	ap = anon_get_ptr(amp->ahp, start_idx);
1910 	if (ap) {
1911 		uint_t pszc;
1912 		swap_xlate(ap, &vp, &off);
1913 		if (page_exists_forreal(vp, (u_offset_t)off, &pszc)) {
1914 			if (pszc > szc && upsize) {
1915 				*ppa_szc = MIN(pszc, seg->s_szc);
1916 				return (-2);
1917 			}
1918 			if (pszc >= szc) {
1919 				prealloc = 0;
1920 			}
1921 		}
1922 	}
1923 
1924 	VM_STAT_COND_ADD(prealloc == 0, anonvmstats.getpages[5]);
1925 	VM_STAT_COND_ADD(prealloc != 0, anonvmstats.getpages[6]);
1926 
1927 top:
1928 	/*
1929 	 * If a smaller page or no page at all was found,
1930 	 * grab a large page off the freelist.
1931 	 */
1932 	if (prealloc) {
1933 		ASSERT(conpp == NULL);
1934 		if (page_alloc_pages(anon_vp, seg, addr, NULL, ppa,
1935 		    szc, 0, pgflags) != 0) {
1936 			VM_STAT_ADD(anonvmstats.getpages[7]);
1937 			if (brkcow == 0 || szc < seg->s_szc ||
1938 			    !anon_szcshare(amp->ahp, start_idx)) {
1939 				/*
1940 				 * If the refcnt's of all anon slots are <= 1
1941 				 * they can't increase since we are holding
1942 				 * the address space's lock. So segvn can
1943 				 * safely decrease szc without risking to
1944 				 * generate a cow fault for the region smaller
1945 				 * than the segment's largest page size.
1946 				 */
1947 				VM_STAT_ADD(anonvmstats.getpages[8]);
1948 				return (-1);
1949 			}
1950 		docow:
1951 			/*
1952 			 * This is a cow fault. Copy away the entire 1 large
1953 			 * page region of this segment.
1954 			 */
1955 			if (szc != seg->s_szc)
1956 				panic("anon_map_getpages: cowfault for szc %d",
1957 				    szc);
1958 			vaddr = addr;
1959 			for (pg_idx = 0, an_idx = start_idx; pg_idx < pgcnt;
1960 			    pg_idx++, an_idx++, vaddr += PAGESIZE) {
1961 				if ((ap = anon_get_ptr(amp->ahp, an_idx)) !=
1962 				    NULL) {
1963 					err = anon_getpage(&ap, &vpprot, pl,
1964 					    PAGESIZE, seg, vaddr, rw, cred);
1965 					if (err) {
1966 						for (i = 0; i < pg_idx; i++) {
1967 							if ((pp = ppa[i]) !=
1968 							    NULL)
1969 								page_unlock(pp);
1970 						}
1971 						return (err);
1972 					}
1973 					ppa[pg_idx] = pl[0];
1974 				} else {
1975 					/*
1976 					 * Since this is a cowfault we know
1977 					 * that this address space has a
1978 					 * parent or children which means
1979 					 * anon_dup_fill_holes() has initialized
1980 					 * all anon slots within a large page
1981 					 * region that had at least one anon
1982 					 * slot at the time of fork().
1983 					 */
1984 					panic("anon_map_getpages: "
1985 					    "cowfault but anon slot is empty");
1986 				}
1987 			}
1988 			VM_STAT_ADD(anonvmstats.getpages[9]);
1989 			*protp = PROT_ALL;
1990 			return (anon_map_privatepages(amp, start_idx, szc, seg,
1991 			    addr, prot, ppa, vpage, anypgsz, pgflags, cred));
1992 		}
1993 	}
1994 
1995 	VM_STAT_ADD(anonvmstats.getpages[10]);
1996 
1997 	an_idx = start_idx;
1998 	pg_idx = 0;
1999 	vaddr = addr;
2000 	while (pg_idx < pgcnt) {
2001 		slotcreate = 0;
2002 		if ((ap = anon_get_ptr(amp->ahp, an_idx)) == NULL) {
2003 			VM_STAT_ADD(anonvmstats.getpages[11]);
2004 			/*
2005 			 * For us to have decided not to preallocate
2006 			 * would have meant that a large page
2007 			 * was found. Which also means that all of the
2008 			 * anon slots for that page would have been
2009 			 * already created for us.
2010 			 */
2011 			if (prealloc == 0)
2012 				panic("anon_map_getpages: prealloc = 0");
2013 
2014 			slotcreate = 1;
2015 			ap = anon_alloc(NULL, 0);
2016 		}
2017 		swap_xlate(ap, &vp, &off);
2018 
2019 		/*
2020 		 * Now setup our preallocated page to pass down
2021 		 * to swap_getpage().
2022 		 */
2023 		if (prealloc) {
2024 			ASSERT(ppa[pg_idx]->p_szc == szc);
2025 			conpp = ppa[pg_idx];
2026 		}
2027 		ASSERT(prealloc || conpp == NULL);
2028 
2029 		/*
2030 		 * If we just created this anon slot then call
2031 		 * with S_CREATE to prevent doing IO on the page.
2032 		 * Similar to the anon_zero case.
2033 		 */
2034 		err = swap_getconpage(vp, (u_offset_t)off, PAGESIZE,
2035 		    NULL, pl, PAGESIZE, conpp, ppa_szc, &nreloc, seg, vaddr,
2036 		    slotcreate == 1 ? S_CREATE : rw, cred);
2037 
2038 		if (err) {
2039 			ASSERT(err != -2 || upsize);
2040 			VM_STAT_ADD(anonvmstats.getpages[12]);
2041 			ASSERT(slotcreate == 0);
2042 			goto io_err;
2043 		}
2044 
2045 		pp = pl[0];
2046 
2047 		if (pp->p_szc < szc || (pp->p_szc > szc && upsize)) {
2048 			VM_STAT_ADD(anonvmstats.getpages[13]);
2049 			ASSERT(slotcreate == 0);
2050 			ASSERT(prealloc == 0);
2051 			ASSERT(pg_idx == 0);
2052 			if (pp->p_szc > szc) {
2053 				ASSERT(upsize);
2054 				*ppa_szc = MIN(pp->p_szc, seg->s_szc);
2055 				page_unlock(pp);
2056 				VM_STAT_ADD(anonvmstats.getpages[14]);
2057 				return (-2);
2058 			}
2059 			page_unlock(pp);
2060 			prealloc = 1;
2061 			goto top;
2062 		}
2063 
2064 		/*
2065 		 * If we decided to preallocate but VOP_GETPAGE
2066 		 * found a page in the system that satisfies our
2067 		 * request then free up our preallocated large page
2068 		 * and continue looping accross the existing large
2069 		 * page via VOP_GETPAGE.
2070 		 */
2071 		if (prealloc && pp != ppa[pg_idx]) {
2072 			VM_STAT_ADD(anonvmstats.getpages[15]);
2073 			ASSERT(slotcreate == 0);
2074 			ASSERT(pg_idx == 0);
2075 			conpp = NULL;
2076 			prealloc = 0;
2077 			page_free_pages(ppa[0]);
2078 		}
2079 
2080 		if (prealloc && nreloc > 1) {
2081 			/*
2082 			 * we have relocated out of a smaller large page.
2083 			 * skip npgs - 1 iterations and continue which will
2084 			 * increment by one the loop indices.
2085 			 */
2086 			spgcnt_t npgs = nreloc;
2087 
2088 			VM_STAT_ADD(anonvmstats.getpages[16]);
2089 
2090 			ASSERT(pp == ppa[pg_idx]);
2091 			ASSERT(slotcreate == 0);
2092 			ASSERT(pg_idx + npgs <= pgcnt);
2093 			if ((*protp & PROT_WRITE) &&
2094 			    anon_share(amp->ahp, an_idx, npgs)) {
2095 				*protp &= ~PROT_WRITE;
2096 			}
2097 			pg_idx += npgs;
2098 			an_idx += npgs;
2099 			vaddr += PAGESIZE * npgs;
2100 			continue;
2101 		}
2102 
2103 		VM_STAT_ADD(anonvmstats.getpages[17]);
2104 
2105 		/*
2106 		 * Anon_zero case.
2107 		 */
2108 		if (slotcreate) {
2109 			ASSERT(prealloc);
2110 			pagezero(pp, 0, PAGESIZE);
2111 			CPU_STATS_ADD_K(vm, zfod, 1);
2112 			hat_setrefmod(pp);
2113 		}
2114 
2115 		ASSERT(prealloc == 0 || ppa[pg_idx] == pp);
2116 		ASSERT(prealloc != 0 || PAGE_SHARED(pp));
2117 		ASSERT(prealloc == 0 || PAGE_EXCL(pp));
2118 
2119 		if (pg_idx > 0 &&
2120 		    ((page_pptonum(pp) != page_pptonum(ppa[pg_idx - 1]) + 1) ||
2121 		    (pp->p_szc != ppa[pg_idx - 1]->p_szc))) {
2122 			panic("anon_map_getpages: unexpected page");
2123 		} else if (pg_idx == 0 && (page_pptonum(pp) & (pgcnt - 1))) {
2124 			panic("anon_map_getpages: unaligned page");
2125 		}
2126 
2127 		if (prealloc == 0) {
2128 			ppa[pg_idx] = pp;
2129 		}
2130 
2131 		if (ap->an_refcnt > 1) {
2132 			VM_STAT_ADD(anonvmstats.getpages[18]);
2133 			*protp &= ~PROT_WRITE;
2134 		}
2135 
2136 		/*
2137 		 * If this is a new anon slot then initialize
2138 		 * the anon array entry.
2139 		 */
2140 		if (slotcreate) {
2141 			(void) anon_set_ptr(amp->ahp, an_idx, ap, ANON_SLEEP);
2142 		}
2143 		pg_idx++;
2144 		an_idx++;
2145 		vaddr += PAGESIZE;
2146 	}
2147 
2148 	/*
2149 	 * Since preallocated pages come off the freelist
2150 	 * they are locked SE_EXCL. Simply downgrade and return.
2151 	 */
2152 	if (prealloc) {
2153 		VM_STAT_ADD(anonvmstats.getpages[19]);
2154 		conpp = NULL;
2155 		for (pg_idx = 0; pg_idx < pgcnt; pg_idx++) {
2156 			page_downgrade(ppa[pg_idx]);
2157 		}
2158 	}
2159 	ASSERT(conpp == NULL);
2160 
2161 	if (brkcow == 0 || (*protp & PROT_WRITE)) {
2162 		VM_STAT_ADD(anonvmstats.getpages[20]);
2163 		return (0);
2164 	}
2165 
2166 	if (szc < seg->s_szc)
2167 		panic("anon_map_getpages: cowfault for szc %d", szc);
2168 
2169 	VM_STAT_ADD(anonvmstats.getpages[21]);
2170 
2171 	*protp = PROT_ALL;
2172 	return (anon_map_privatepages(amp, start_idx, szc, seg, addr, prot,
2173 	    ppa, vpage, anypgsz, pgflags, cred));
2174 io_err:
2175 	/*
2176 	 * We got an IO error somewhere in our large page.
2177 	 * If we were using a preallocated page then just demote
2178 	 * all the constituent pages that we've succeeded with sofar
2179 	 * to PAGESIZE pages and leave them in the system
2180 	 * unlocked.
2181 	 */
2182 
2183 	ASSERT(err != -2 || ((pg_idx == 0) && upsize));
2184 
2185 	VM_STAT_COND_ADD(err > 0, anonvmstats.getpages[22]);
2186 	VM_STAT_COND_ADD(err == -1, anonvmstats.getpages[23]);
2187 	VM_STAT_COND_ADD(err == -2, anonvmstats.getpages[24]);
2188 
2189 	if (prealloc) {
2190 		conpp = NULL;
2191 		if (pg_idx > 0) {
2192 			VM_STAT_ADD(anonvmstats.getpages[25]);
2193 			for (i = 0; i < pgcnt; i++) {
2194 				pp = ppa[i];
2195 				ASSERT(PAGE_EXCL(pp));
2196 				ASSERT(pp->p_szc == szc);
2197 				pp->p_szc = 0;
2198 			}
2199 			for (i = 0; i < pg_idx; i++) {
2200 				ASSERT(!hat_page_is_mapped(ppa[i]));
2201 				page_unlock(ppa[i]);
2202 			}
2203 			/*
2204 			 * Now free up the remaining unused constituent
2205 			 * pages.
2206 			 */
2207 			while (pg_idx < pgcnt) {
2208 				ASSERT(!hat_page_is_mapped(ppa[pg_idx]));
2209 				page_free(ppa[pg_idx], 0);
2210 				pg_idx++;
2211 			}
2212 		} else {
2213 			VM_STAT_ADD(anonvmstats.getpages[26]);
2214 			page_free_pages(ppa[0]);
2215 		}
2216 	} else {
2217 		VM_STAT_ADD(anonvmstats.getpages[27]);
2218 		ASSERT(err > 0);
2219 		for (i = 0; i < pg_idx; i++)
2220 			page_unlock(ppa[i]);
2221 	}
2222 	ASSERT(conpp == NULL);
2223 	if (err != -1)
2224 		return (err);
2225 	/*
2226 	 * we are here because we failed to relocate.
2227 	 */
2228 	ASSERT(prealloc);
2229 	if (brkcow == 0 || szc < seg->s_szc ||
2230 	    !anon_szcshare(amp->ahp, start_idx)) {
2231 		VM_STAT_ADD(anonvmstats.getpages[28]);
2232 		return (-1);
2233 	}
2234 	VM_STAT_ADD(anonvmstats.getpages[29]);
2235 	goto docow;
2236 }
2237 
2238 
2239 /*
2240  * Turn a reference to an object or shared anon page
2241  * into a private page with a copy of the data from the
2242  * original page which is always locked by the caller.
2243  * This routine unloads the translation and unlocks the
2244  * original page, if it isn't being stolen, before returning
2245  * to the caller.
2246  *
2247  * NOTE:  The original anon slot is not freed by this routine
2248  *	  It must be freed by the caller while holding the
2249  *	  "anon_map" lock to prevent races which can occur if
2250  *	  a process has multiple lwps in its address space.
2251  */
2252 page_t *
2253 anon_private(
2254 	struct anon **app,
2255 	struct seg *seg,
2256 	caddr_t addr,
2257 	uint_t	prot,
2258 	page_t *opp,
2259 	int oppflags,
2260 	struct cred *cred)
2261 {
2262 	struct anon *old = *app;
2263 	struct anon *new;
2264 	page_t *pp = NULL;
2265 	struct vnode *vp;
2266 	anoff_t off;
2267 	page_t *anon_pl[1 + 1];
2268 	int err;
2269 
2270 	if (oppflags & STEAL_PAGE)
2271 		ASSERT(PAGE_EXCL(opp));
2272 	else
2273 		ASSERT(PAGE_LOCKED(opp));
2274 
2275 	CPU_STATS_ADD_K(vm, cow_fault, 1);
2276 
2277 	/* Kernel probe */
2278 	TNF_PROBE_1(anon_private, "vm pagefault", /* CSTYLED */,
2279 		tnf_opaque,	address,	addr);
2280 
2281 	*app = new = anon_alloc(NULL, 0);
2282 	swap_xlate(new, &vp, &off);
2283 
2284 	if (oppflags & STEAL_PAGE) {
2285 		page_rename(opp, vp, (u_offset_t)off);
2286 		pp = opp;
2287 		TRACE_5(TR_FAC_VM, TR_ANON_PRIVATE,
2288 		    "anon_private:seg %p addr %x pp %p vp %p off %lx",
2289 		    seg, addr, pp, vp, off);
2290 		hat_setmod(pp);
2291 
2292 		/* bug 4026339 */
2293 		page_downgrade(pp);
2294 		return (pp);
2295 	}
2296 
2297 	/*
2298 	 * Call the VOP_GETPAGE routine to create the page, thereby
2299 	 * enabling the vnode driver to allocate any filesystem
2300 	 * space (e.g., disk block allocation for UFS).  This also
2301 	 * prevents more than one page from being added to the
2302 	 * vnode at the same time.
2303 	 */
2304 	err = VOP_GETPAGE(vp, (u_offset_t)off, PAGESIZE, NULL,
2305 	    anon_pl, PAGESIZE, seg, addr, S_CREATE, cred, NULL);
2306 	if (err)
2307 		goto out;
2308 
2309 	pp = anon_pl[0];
2310 
2311 	/*
2312 	 * If the original page was locked, we need to move the lock
2313 	 * to the new page by transfering 'cowcnt/lckcnt' of the original
2314 	 * page to 'cowcnt/lckcnt' of the new page.
2315 	 *
2316 	 * See Statement at the beginning of segvn_lockop() and
2317 	 * comments in page_pp_useclaim() regarding the way
2318 	 * cowcnts/lckcnts are handled.
2319 	 *
2320 	 * Also availrmem must be decremented up front for read only mapping
2321 	 * before calling page_pp_useclaim. page_pp_useclaim will bump it back
2322 	 * if availrmem did not need to be decremented after all.
2323 	 */
2324 	if (oppflags & LOCK_PAGE) {
2325 		if ((prot & PROT_WRITE) == 0) {
2326 			mutex_enter(&freemem_lock);
2327 			if (availrmem > pages_pp_maximum) {
2328 				availrmem--;
2329 				pages_useclaim++;
2330 			} else {
2331 				mutex_exit(&freemem_lock);
2332 				goto out;
2333 			}
2334 			mutex_exit(&freemem_lock);
2335 		}
2336 		page_pp_useclaim(opp, pp, prot & PROT_WRITE);
2337 	}
2338 
2339 	/*
2340 	 * Now copy the contents from the original page,
2341 	 * which is locked and loaded in the MMU by
2342 	 * the caller to prevent yet another page fault.
2343 	 */
2344 	/* XXX - should set mod bit in here */
2345 	if (ppcopy(opp, pp) == 0) {
2346 		/*
2347 		 * Before ppcopy could hanlde UE or other faults, we
2348 		 * would have panicked here, and still have no option
2349 		 * but to do so now.
2350 		 */
2351 		panic("anon_private, ppcopy failed, opp = 0x%p, pp = 0x%p",
2352 		    opp, pp);
2353 	}
2354 
2355 	hat_setrefmod(pp);		/* mark as modified */
2356 
2357 	/*
2358 	 * Unload the old translation.
2359 	 */
2360 	hat_unload(seg->s_as->a_hat, addr, PAGESIZE, HAT_UNLOAD);
2361 
2362 	/*
2363 	 * Free unmapped, unmodified original page.
2364 	 * or release the lock on the original page,
2365 	 * otherwise the process will sleep forever in
2366 	 * anon_decref() waiting for the "exclusive" lock
2367 	 * on the page.
2368 	 */
2369 	(void) page_release(opp, 1);
2370 
2371 	/*
2372 	 * we are done with page creation so downgrade the new
2373 	 * page's selock to shared, this helps when multiple
2374 	 * as_fault(...SOFTLOCK...) are done to the same
2375 	 * page(aio)
2376 	 */
2377 	page_downgrade(pp);
2378 
2379 	/*
2380 	 * NOTE:  The original anon slot must be freed by the
2381 	 * caller while holding the "anon_map" lock, if we
2382 	 * copied away from an anonymous page.
2383 	 */
2384 	return (pp);
2385 
2386 out:
2387 	*app = old;
2388 	if (pp)
2389 		page_unlock(pp);
2390 	anon_decref(new);
2391 	page_unlock(opp);
2392 	return ((page_t *)NULL);
2393 }
2394 
2395 int
2396 anon_map_privatepages(
2397 	struct anon_map *amp,
2398 	ulong_t	start_idx,
2399 	uint_t	szc,
2400 	struct seg *seg,
2401 	caddr_t addr,
2402 	uint_t	prot,
2403 	page_t	*ppa[],
2404 	struct vpage vpage[],
2405 	int anypgsz,
2406 	int pgflags,
2407 	struct cred *cred)
2408 {
2409 	pgcnt_t		pgcnt;
2410 	struct vnode	*vp;
2411 	anoff_t		off;
2412 	page_t		*pl[2], *conpp = NULL;
2413 	int		err;
2414 	int		prealloc = 1;
2415 	struct anon	*ap, *oldap;
2416 	caddr_t		vaddr;
2417 	page_t		*pplist, *pp;
2418 	ulong_t		pg_idx, an_idx;
2419 	spgcnt_t	nreloc = 0;
2420 	int		pagelock = 0;
2421 	kmutex_t	*ahmpages = NULL;
2422 #ifdef DEBUG
2423 	int		refcnt;
2424 #endif
2425 
2426 	ASSERT(szc != 0);
2427 	ASSERT(szc == seg->s_szc);
2428 
2429 	VM_STAT_ADD(anonvmstats.privatepages[0]);
2430 
2431 	pgcnt = page_get_pagecnt(szc);
2432 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
2433 	ASSERT(IS_P2ALIGNED(start_idx, pgcnt));
2434 
2435 	ASSERT(amp != NULL);
2436 	ap = anon_get_ptr(amp->ahp, start_idx);
2437 	ASSERT(ap == NULL || ap->an_refcnt >= 1);
2438 
2439 	VM_STAT_COND_ADD(ap == NULL, anonvmstats.privatepages[1]);
2440 
2441 	/*
2442 	 * Now try and allocate the large page. If we fail then just
2443 	 * let VOP_GETPAGE give us PAGESIZE pages. Normally we let
2444 	 * the caller make this decision but to avoid added complexity
2445 	 * it's simplier to handle that case here.
2446 	 */
2447 	if (anypgsz == -1) {
2448 		VM_STAT_ADD(anonvmstats.privatepages[2]);
2449 		prealloc = 0;
2450 	} else if (page_alloc_pages(anon_vp, seg, addr, &pplist, NULL, szc,
2451 	    anypgsz, pgflags) != 0) {
2452 		VM_STAT_ADD(anonvmstats.privatepages[3]);
2453 		prealloc = 0;
2454 	}
2455 
2456 	/*
2457 	 * make the decrement of all refcnts of all
2458 	 * anon slots of a large page appear atomic by
2459 	 * getting an anonpages_hash_lock for the
2460 	 * first anon slot of a large page.
2461 	 */
2462 	if (ap != NULL) {
2463 		ahmpages = &anonpages_hash_lock[AH_LOCK(ap->an_vp,
2464 		    ap->an_off)];
2465 		mutex_enter(ahmpages);
2466 		if (ap->an_refcnt == 1) {
2467 			VM_STAT_ADD(anonvmstats.privatepages[4]);
2468 			ASSERT(!anon_share(amp->ahp, start_idx, pgcnt));
2469 			mutex_exit(ahmpages);
2470 
2471 			if (prealloc) {
2472 				page_free_replacement_page(pplist);
2473 				page_create_putback(pgcnt);
2474 			}
2475 			ASSERT(ppa[0]->p_szc <= szc);
2476 			if (ppa[0]->p_szc == szc) {
2477 				VM_STAT_ADD(anonvmstats.privatepages[5]);
2478 				return (0);
2479 			}
2480 			for (pg_idx = 0; pg_idx < pgcnt; pg_idx++) {
2481 				ASSERT(ppa[pg_idx] != NULL);
2482 				page_unlock(ppa[pg_idx]);
2483 			}
2484 			return (-1);
2485 		}
2486 	}
2487 
2488 	/*
2489 	 * If we are passed in the vpage array and this is
2490 	 * not PROT_WRITE then we need to decrement availrmem
2491 	 * up front before we try anything. If we need to and
2492 	 * can't decrement availrmem then its better to fail now
2493 	 * than in the middle of processing the new large page.
2494 	 * page_pp_usclaim() on behalf of each constituent page
2495 	 * below will adjust availrmem back for the cases not needed.
2496 	 */
2497 	if (vpage != NULL && (prot & PROT_WRITE) == 0) {
2498 		for (pg_idx = 0; pg_idx < pgcnt; pg_idx++) {
2499 			if (VPP_ISPPLOCK(&vpage[pg_idx])) {
2500 				pagelock = 1;
2501 				break;
2502 			}
2503 		}
2504 		if (pagelock) {
2505 			VM_STAT_ADD(anonvmstats.privatepages[6]);
2506 			mutex_enter(&freemem_lock);
2507 			if (availrmem >= pages_pp_maximum + pgcnt) {
2508 				availrmem -= pgcnt;
2509 				pages_useclaim += pgcnt;
2510 			} else {
2511 				VM_STAT_ADD(anonvmstats.privatepages[7]);
2512 				mutex_exit(&freemem_lock);
2513 				if (ahmpages != NULL) {
2514 					mutex_exit(ahmpages);
2515 				}
2516 				if (prealloc) {
2517 					page_free_replacement_page(pplist);
2518 					page_create_putback(pgcnt);
2519 				}
2520 				for (pg_idx = 0; pg_idx < pgcnt; pg_idx++)
2521 					if (ppa[pg_idx] != NULL)
2522 						page_unlock(ppa[pg_idx]);
2523 				return (ENOMEM);
2524 			}
2525 			mutex_exit(&freemem_lock);
2526 		}
2527 	}
2528 
2529 	CPU_STATS_ADD_K(vm, cow_fault, pgcnt);
2530 
2531 	VM_STAT_ADD(anonvmstats.privatepages[8]);
2532 
2533 	an_idx = start_idx;
2534 	pg_idx = 0;
2535 	vaddr = addr;
2536 	for (; pg_idx < pgcnt; pg_idx++, an_idx++, vaddr += PAGESIZE) {
2537 		ASSERT(ppa[pg_idx] != NULL);
2538 		oldap = anon_get_ptr(amp->ahp, an_idx);
2539 		ASSERT(ahmpages != NULL || oldap == NULL);
2540 		ASSERT(ahmpages == NULL || oldap != NULL);
2541 		ASSERT(ahmpages == NULL || oldap->an_refcnt > 1);
2542 		ASSERT(ahmpages == NULL || pg_idx != 0 ||
2543 		    (refcnt = oldap->an_refcnt));
2544 		ASSERT(ahmpages == NULL || pg_idx == 0 ||
2545 		    refcnt == oldap->an_refcnt);
2546 
2547 		ap = anon_alloc(NULL, 0);
2548 
2549 		swap_xlate(ap, &vp, &off);
2550 
2551 		/*
2552 		 * Now setup our preallocated page to pass down to
2553 		 * swap_getpage().
2554 		 */
2555 		if (prealloc) {
2556 			pp = pplist;
2557 			page_sub(&pplist, pp);
2558 			conpp = pp;
2559 		}
2560 
2561 		err = swap_getconpage(vp, (u_offset_t)off, PAGESIZE, NULL, pl,
2562 		    PAGESIZE, conpp, NULL, &nreloc, seg, vaddr,
2563 		    S_CREATE, cred);
2564 
2565 		/*
2566 		 * Impossible to fail this is S_CREATE.
2567 		 */
2568 		if (err)
2569 			panic("anon_map_privatepages: VOP_GETPAGE failed");
2570 
2571 		ASSERT(prealloc ? pp == pl[0] : pl[0]->p_szc == 0);
2572 		ASSERT(prealloc == 0 || nreloc == 1);
2573 
2574 		pp = pl[0];
2575 
2576 		/*
2577 		 * If the original page was locked, we need to move
2578 		 * the lock to the new page by transfering
2579 		 * 'cowcnt/lckcnt' of the original page to 'cowcnt/lckcnt'
2580 		 * of the new page. pg_idx can be used to index
2581 		 * into the vpage array since the caller will guarentee
2582 		 * that vpage struct passed in corresponds to addr
2583 		 * and forward.
2584 		 */
2585 		if (vpage != NULL && VPP_ISPPLOCK(&vpage[pg_idx])) {
2586 			page_pp_useclaim(ppa[pg_idx], pp, prot & PROT_WRITE);
2587 		} else if (pagelock) {
2588 			mutex_enter(&freemem_lock);
2589 			availrmem++;
2590 			pages_useclaim--;
2591 			mutex_exit(&freemem_lock);
2592 		}
2593 
2594 		/*
2595 		 * Now copy the contents from the original page.
2596 		 */
2597 		if (ppcopy(ppa[pg_idx], pp) == 0) {
2598 			/*
2599 			 * Before ppcopy could hanlde UE or other faults, we
2600 			 * would have panicked here, and still have no option
2601 			 * but to do so now.
2602 			 */
2603 			panic("anon_map_privatepages, ppcopy failed");
2604 		}
2605 
2606 		hat_setrefmod(pp);		/* mark as modified */
2607 
2608 		/*
2609 		 * Release the lock on the original page,
2610 		 * derement the old slot, and down grade the lock
2611 		 * on the new copy.
2612 		 */
2613 		page_unlock(ppa[pg_idx]);
2614 
2615 		if (!prealloc)
2616 			page_downgrade(pp);
2617 
2618 		ppa[pg_idx] = pp;
2619 
2620 		/*
2621 		 * Now reflect the copy in the new anon array.
2622 		 */
2623 		ASSERT(ahmpages == NULL || oldap->an_refcnt > 1);
2624 		if (oldap != NULL)
2625 			anon_decref(oldap);
2626 		(void) anon_set_ptr(amp->ahp, an_idx, ap, ANON_SLEEP);
2627 	}
2628 	if (ahmpages != NULL) {
2629 		mutex_exit(ahmpages);
2630 	}
2631 	ASSERT(prealloc == 0 || pplist == NULL);
2632 	if (prealloc) {
2633 		VM_STAT_ADD(anonvmstats.privatepages[9]);
2634 		for (pg_idx = 0; pg_idx < pgcnt; pg_idx++) {
2635 			page_downgrade(ppa[pg_idx]);
2636 		}
2637 	}
2638 
2639 	/*
2640 	 * Unload the old large page translation.
2641 	 */
2642 	hat_unload(seg->s_as->a_hat, addr, pgcnt << PAGESHIFT, HAT_UNLOAD);
2643 	return (0);
2644 }
2645 
2646 /*
2647  * Allocate a private zero-filled anon page.
2648  */
2649 page_t *
2650 anon_zero(struct seg *seg, caddr_t addr, struct anon **app, struct cred *cred)
2651 {
2652 	struct anon *ap;
2653 	page_t *pp;
2654 	struct vnode *vp;
2655 	anoff_t off;
2656 	page_t *anon_pl[1 + 1];
2657 	int err;
2658 
2659 	/* Kernel probe */
2660 	TNF_PROBE_1(anon_zero, "vm pagefault", /* CSTYLED */,
2661 		tnf_opaque,	address,	addr);
2662 
2663 	*app = ap = anon_alloc(NULL, 0);
2664 	swap_xlate(ap, &vp, &off);
2665 
2666 	/*
2667 	 * Call the VOP_GETPAGE routine to create the page, thereby
2668 	 * enabling the vnode driver to allocate any filesystem
2669 	 * dependent structures (e.g., disk block allocation for UFS).
2670 	 * This also prevents more than on page from being added to
2671 	 * the vnode at the same time since it is locked.
2672 	 */
2673 	err = VOP_GETPAGE(vp, off, PAGESIZE, NULL,
2674 	    anon_pl, PAGESIZE, seg, addr, S_CREATE, cred, NULL);
2675 	if (err) {
2676 		*app = NULL;
2677 		anon_decref(ap);
2678 		return (NULL);
2679 	}
2680 	pp = anon_pl[0];
2681 
2682 	pagezero(pp, 0, PAGESIZE);	/* XXX - should set mod bit */
2683 	page_downgrade(pp);
2684 	CPU_STATS_ADD_K(vm, zfod, 1);
2685 	hat_setrefmod(pp);	/* mark as modified so pageout writes back */
2686 	return (pp);
2687 }
2688 
2689 
2690 /*
2691  * Allocate array of private zero-filled anon pages for empty slots
2692  * and kept pages for non empty slots within given range.
2693  *
2694  * NOTE: This rontine will try and use large pages
2695  *	if available and supported by underlying platform.
2696  */
2697 int
2698 anon_map_createpages(
2699 	struct anon_map *amp,
2700 	ulong_t start_index,
2701 	size_t len,
2702 	page_t *ppa[],
2703 	struct seg *seg,
2704 	caddr_t addr,
2705 	enum seg_rw rw,
2706 	struct cred *cred)
2707 {
2708 
2709 	struct anon	*ap;
2710 	struct vnode	*ap_vp;
2711 	page_t		*pp, *pplist, *anon_pl[1 + 1], *conpp = NULL;
2712 	int		err = 0;
2713 	ulong_t		p_index, index;
2714 	pgcnt_t		npgs, pg_cnt;
2715 	spgcnt_t	nreloc = 0;
2716 	uint_t		l_szc, szc, prot;
2717 	anoff_t		ap_off;
2718 	size_t		pgsz;
2719 	lgrp_t		*lgrp;
2720 	kmutex_t	*ahm;
2721 
2722 	/*
2723 	 * XXX For now only handle S_CREATE.
2724 	 */
2725 	ASSERT(rw == S_CREATE);
2726 
2727 	index	= start_index;
2728 	p_index	= 0;
2729 	npgs = btopr(len);
2730 
2731 	/*
2732 	 * If this platform supports multiple page sizes
2733 	 * then try and allocate directly from the free
2734 	 * list for pages larger than PAGESIZE.
2735 	 *
2736 	 * NOTE:When we have page_create_ru we can stop
2737 	 *	directly allocating from the freelist.
2738 	 */
2739 	l_szc  = seg->s_szc;
2740 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2741 	while (npgs) {
2742 
2743 		/*
2744 		 * if anon slot already exists
2745 		 *   (means page has been created)
2746 		 * so 1) look up the page
2747 		 *    2) if the page is still in memory, get it.
2748 		 *    3) if not, create a page and
2749 		 *	  page in from physical swap device.
2750 		 * These are done in anon_getpage().
2751 		 */
2752 		ap = anon_get_ptr(amp->ahp, index);
2753 		if (ap) {
2754 			err = anon_getpage(&ap, &prot, anon_pl, PAGESIZE,
2755 			    seg, addr, S_READ, cred);
2756 			if (err) {
2757 				ANON_LOCK_EXIT(&amp->a_rwlock);
2758 				panic("anon_map_createpages: anon_getpage");
2759 			}
2760 			pp = anon_pl[0];
2761 			ppa[p_index++] = pp;
2762 
2763 			/*
2764 			 * an_pvp can become non-NULL after SysV's page was
2765 			 * paged out before ISM was attached to this SysV
2766 			 * shared memory segment. So free swap slot if needed.
2767 			 */
2768 			if (ap->an_pvp != NULL) {
2769 				page_io_lock(pp);
2770 				ahm = &anonhash_lock[AH_LOCK(ap->an_vp,
2771 				    ap->an_off)];
2772 				mutex_enter(ahm);
2773 				if (ap->an_pvp != NULL) {
2774 					swap_phys_free(ap->an_pvp,
2775 					    ap->an_poff, PAGESIZE);
2776 					ap->an_pvp = NULL;
2777 					ap->an_poff = 0;
2778 					mutex_exit(ahm);
2779 					hat_setmod(pp);
2780 				} else {
2781 					mutex_exit(ahm);
2782 				}
2783 				page_io_unlock(pp);
2784 			}
2785 
2786 			addr += PAGESIZE;
2787 			index++;
2788 			npgs--;
2789 			continue;
2790 		}
2791 		/*
2792 		 * Now try and allocate the largest page possible
2793 		 * for the current address and range.
2794 		 * Keep dropping down in page size until:
2795 		 *
2796 		 *	1) Properly aligned
2797 		 *	2) Does not overlap existing anon pages
2798 		 *	3) Fits in remaining range.
2799 		 *	4) able to allocate one.
2800 		 *
2801 		 * NOTE: XXX When page_create_ru is completed this code
2802 		 *	 will change.
2803 		 */
2804 		szc    = l_szc;
2805 		pplist = NULL;
2806 		pg_cnt = 0;
2807 		while (szc) {
2808 			pgsz	= page_get_pagesize(szc);
2809 			pg_cnt	= pgsz >> PAGESHIFT;
2810 			if (IS_P2ALIGNED(addr, pgsz) && pg_cnt <= npgs &&
2811 			    anon_pages(amp->ahp, index, pg_cnt) == 0) {
2812 				/*
2813 				 * XXX
2814 				 * Since we are faking page_create()
2815 				 * we also need to do the freemem and
2816 				 * pcf accounting.
2817 				 */
2818 				(void) page_create_wait(pg_cnt, PG_WAIT);
2819 
2820 				/*
2821 				 * Get lgroup to allocate next page of shared
2822 				 * memory from and use it to specify where to
2823 				 * allocate the physical memory
2824 				 */
2825 				lgrp = lgrp_mem_choose(seg, addr, pgsz);
2826 
2827 				pplist = page_get_freelist(
2828 				    anon_vp, (u_offset_t)0, seg,
2829 				    addr, pgsz, 0, lgrp);
2830 
2831 				if (pplist == NULL) {
2832 					page_create_putback(pg_cnt);
2833 				}
2834 
2835 				/*
2836 				 * If a request for a page of size
2837 				 * larger than PAGESIZE failed
2838 				 * then don't try that size anymore.
2839 				 */
2840 				if (pplist == NULL) {
2841 					l_szc = szc - 1;
2842 				} else {
2843 					break;
2844 				}
2845 			}
2846 			szc--;
2847 		}
2848 
2849 		/*
2850 		 * If just using PAGESIZE pages then don't
2851 		 * directly allocate from the free list.
2852 		 */
2853 		if (pplist == NULL) {
2854 			ASSERT(szc == 0);
2855 			pp = anon_zero(seg, addr, &ap, cred);
2856 			if (pp == NULL) {
2857 				ANON_LOCK_EXIT(&amp->a_rwlock);
2858 				panic("anon_map_createpages: anon_zero");
2859 			}
2860 			ppa[p_index++] = pp;
2861 
2862 			ASSERT(anon_get_ptr(amp->ahp, index) == NULL);
2863 			(void) anon_set_ptr(amp->ahp, index, ap, ANON_SLEEP);
2864 
2865 			addr += PAGESIZE;
2866 			index++;
2867 			npgs--;
2868 			continue;
2869 		}
2870 
2871 		/*
2872 		 * pplist is a list of pg_cnt PAGESIZE pages.
2873 		 * These pages are locked SE_EXCL since they
2874 		 * came directly off the free list.
2875 		 */
2876 		ASSERT(IS_P2ALIGNED(pg_cnt, pg_cnt));
2877 		ASSERT(IS_P2ALIGNED(index, pg_cnt));
2878 		ASSERT(conpp == NULL);
2879 		while (pg_cnt--) {
2880 
2881 			ap = anon_alloc(NULL, 0);
2882 			swap_xlate(ap, &ap_vp, &ap_off);
2883 
2884 			ASSERT(pplist != NULL);
2885 			pp = pplist;
2886 			page_sub(&pplist, pp);
2887 			PP_CLRFREE(pp);
2888 			PP_CLRAGED(pp);
2889 			conpp = pp;
2890 
2891 			err = swap_getconpage(ap_vp, ap_off, PAGESIZE,
2892 			    (uint_t *)NULL, anon_pl, PAGESIZE, conpp, NULL,
2893 			    &nreloc, seg, addr, S_CREATE, cred);
2894 
2895 			if (err) {
2896 				ANON_LOCK_EXIT(&amp->a_rwlock);
2897 				panic("anon_map_createpages: S_CREATE");
2898 			}
2899 
2900 			ASSERT(anon_pl[0] == pp);
2901 			ASSERT(nreloc == 1);
2902 			pagezero(pp, 0, PAGESIZE);
2903 			CPU_STATS_ADD_K(vm, zfod, 1);
2904 			hat_setrefmod(pp);
2905 
2906 			ASSERT(anon_get_ptr(amp->ahp, index) == NULL);
2907 			(void) anon_set_ptr(amp->ahp, index, ap, ANON_SLEEP);
2908 
2909 			ppa[p_index++] = pp;
2910 
2911 			addr += PAGESIZE;
2912 			index++;
2913 			npgs--;
2914 		}
2915 		conpp = NULL;
2916 		pg_cnt	= pgsz >> PAGESHIFT;
2917 		p_index = p_index - pg_cnt;
2918 		while (pg_cnt--) {
2919 			page_downgrade(ppa[p_index++]);
2920 		}
2921 	}
2922 	ANON_LOCK_EXIT(&amp->a_rwlock);
2923 	return (0);
2924 }
2925 
2926 static int
2927 anon_try_demote_pages(
2928 	struct anon_hdr *ahp,
2929 	ulong_t sidx,
2930 	uint_t szc,
2931 	page_t **ppa,
2932 	int private)
2933 {
2934 	struct anon	*ap;
2935 	pgcnt_t		pgcnt = page_get_pagecnt(szc);
2936 	page_t		*pp;
2937 	pgcnt_t		i;
2938 	kmutex_t	*ahmpages = NULL;
2939 	int		root = 0;
2940 	pgcnt_t		npgs;
2941 	pgcnt_t		curnpgs = 0;
2942 	size_t		ppasize = 0;
2943 
2944 	ASSERT(szc != 0);
2945 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
2946 	ASSERT(IS_P2ALIGNED(sidx, pgcnt));
2947 	ASSERT(sidx < ahp->size);
2948 
2949 	if (ppa == NULL) {
2950 		ppasize = pgcnt * sizeof (page_t *);
2951 		ppa = kmem_alloc(ppasize, KM_SLEEP);
2952 	}
2953 
2954 	ap = anon_get_ptr(ahp, sidx);
2955 	if (ap != NULL && private) {
2956 		VM_STAT_ADD(anonvmstats.demotepages[1]);
2957 		ahmpages = &anonpages_hash_lock[AH_LOCK(ap->an_vp, ap->an_off)];
2958 		mutex_enter(ahmpages);
2959 	}
2960 
2961 	if (ap != NULL && ap->an_refcnt > 1) {
2962 		if (ahmpages != NULL) {
2963 			VM_STAT_ADD(anonvmstats.demotepages[2]);
2964 			mutex_exit(ahmpages);
2965 		}
2966 		if (ppasize != 0) {
2967 			kmem_free(ppa, ppasize);
2968 		}
2969 		return (0);
2970 	}
2971 	if (ahmpages != NULL) {
2972 		mutex_exit(ahmpages);
2973 	}
2974 	if (ahp->size - sidx < pgcnt) {
2975 		ASSERT(private == 0);
2976 		pgcnt = ahp->size - sidx;
2977 	}
2978 	for (i = 0; i < pgcnt; i++, sidx++) {
2979 		ap = anon_get_ptr(ahp, sidx);
2980 		if (ap != NULL) {
2981 			if (ap->an_refcnt != 1) {
2982 				panic("anon_try_demote_pages: an_refcnt != 1");
2983 			}
2984 			pp = ppa[i] = page_lookup(ap->an_vp, ap->an_off,
2985 			    SE_EXCL);
2986 			if (pp != NULL) {
2987 				(void) hat_pageunload(pp,
2988 				    HAT_FORCE_PGUNLOAD);
2989 			}
2990 		} else {
2991 			ppa[i] = NULL;
2992 		}
2993 	}
2994 	for (i = 0; i < pgcnt; i++) {
2995 		if ((pp = ppa[i]) != NULL && pp->p_szc != 0) {
2996 			ASSERT(pp->p_szc <= szc);
2997 			if (!root) {
2998 				VM_STAT_ADD(anonvmstats.demotepages[3]);
2999 				if (curnpgs != 0)
3000 					panic("anon_try_demote_pages: "
3001 					    "bad large page");
3002 
3003 				root = 1;
3004 				curnpgs = npgs =
3005 				    page_get_pagecnt(pp->p_szc);
3006 
3007 				ASSERT(npgs <= pgcnt);
3008 				ASSERT(IS_P2ALIGNED(npgs, npgs));
3009 				ASSERT(!(page_pptonum(pp) & (npgs - 1)));
3010 			} else {
3011 				ASSERT(i > 0);
3012 				ASSERT(page_pptonum(pp) - 1 ==
3013 				    page_pptonum(ppa[i - 1]));
3014 				if ((page_pptonum(pp) & (npgs - 1)) ==
3015 				    npgs - 1)
3016 					root = 0;
3017 			}
3018 			ASSERT(PAGE_EXCL(pp));
3019 			pp->p_szc = 0;
3020 			ASSERT(curnpgs > 0);
3021 			curnpgs--;
3022 		}
3023 	}
3024 	if (root != 0 || curnpgs != 0)
3025 		panic("anon_try_demote_pages: bad large page");
3026 
3027 	for (i = 0; i < pgcnt; i++) {
3028 		if ((pp = ppa[i]) != NULL) {
3029 			ASSERT(!hat_page_is_mapped(pp));
3030 			ASSERT(pp->p_szc == 0);
3031 			page_unlock(pp);
3032 		}
3033 	}
3034 	if (ppasize != 0) {
3035 		kmem_free(ppa, ppasize);
3036 	}
3037 	return (1);
3038 }
3039 
3040 /*
3041  * anon_map_demotepages() can only be called by MAP_PRIVATE segments.
3042  */
3043 int
3044 anon_map_demotepages(
3045 	struct anon_map *amp,
3046 	ulong_t	start_idx,
3047 	struct seg *seg,
3048 	caddr_t addr,
3049 	uint_t prot,
3050 	struct vpage vpage[],
3051 	struct cred *cred)
3052 {
3053 	struct anon	*ap;
3054 	uint_t		szc = seg->s_szc;
3055 	pgcnt_t		pgcnt = page_get_pagecnt(szc);
3056 	size_t		ppasize = pgcnt * sizeof (page_t *);
3057 	page_t		**ppa = kmem_alloc(ppasize, KM_SLEEP);
3058 	page_t		*pp;
3059 	page_t		*pl[2];
3060 	pgcnt_t		i, pg_idx;
3061 	ulong_t		an_idx;
3062 	caddr_t		vaddr;
3063 	int 		err;
3064 	int		retry = 0;
3065 	uint_t		vpprot;
3066 
3067 	ASSERT(RW_WRITE_HELD(&amp->a_rwlock));
3068 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
3069 	ASSERT(IS_P2ALIGNED(start_idx, pgcnt));
3070 	ASSERT(ppa != NULL);
3071 	ASSERT(szc != 0);
3072 	ASSERT(szc == amp->a_szc);
3073 
3074 	VM_STAT_ADD(anonvmstats.demotepages[0]);
3075 
3076 top:
3077 	if (anon_try_demote_pages(amp->ahp, start_idx, szc, ppa, 1)) {
3078 		kmem_free(ppa, ppasize);
3079 		return (0);
3080 	}
3081 
3082 	VM_STAT_ADD(anonvmstats.demotepages[4]);
3083 
3084 	ASSERT(retry == 0); /* we can be here only once */
3085 
3086 	vaddr = addr;
3087 	for (pg_idx = 0, an_idx = start_idx; pg_idx < pgcnt;
3088 	    pg_idx++, an_idx++, vaddr += PAGESIZE) {
3089 		ap = anon_get_ptr(amp->ahp, an_idx);
3090 		if (ap == NULL)
3091 			panic("anon_map_demotepages: no anon slot");
3092 		err = anon_getpage(&ap, &vpprot, pl, PAGESIZE, seg, vaddr,
3093 		    S_READ, cred);
3094 		if (err) {
3095 			for (i = 0; i < pg_idx; i++) {
3096 				if ((pp = ppa[i]) != NULL)
3097 					page_unlock(pp);
3098 			}
3099 			kmem_free(ppa, ppasize);
3100 			return (err);
3101 		}
3102 		ppa[pg_idx] = pl[0];
3103 	}
3104 
3105 	err = anon_map_privatepages(amp, start_idx, szc, seg, addr, prot, ppa,
3106 	    vpage, -1, 0, cred);
3107 	if (err > 0) {
3108 		VM_STAT_ADD(anonvmstats.demotepages[5]);
3109 		kmem_free(ppa, ppasize);
3110 		return (err);
3111 	}
3112 	ASSERT(err == 0 || err == -1);
3113 	if (err == -1) {
3114 		VM_STAT_ADD(anonvmstats.demotepages[6]);
3115 		retry = 1;
3116 		goto top;
3117 	}
3118 	for (i = 0; i < pgcnt; i++) {
3119 		ASSERT(ppa[i] != NULL);
3120 		if (ppa[i]->p_szc != 0)
3121 			retry = 1;
3122 		page_unlock(ppa[i]);
3123 	}
3124 	if (retry) {
3125 		VM_STAT_ADD(anonvmstats.demotepages[7]);
3126 		goto top;
3127 	}
3128 
3129 	VM_STAT_ADD(anonvmstats.demotepages[8]);
3130 
3131 	kmem_free(ppa, ppasize);
3132 
3133 	return (0);
3134 }
3135 
3136 /*
3137  * Free pages of shared anon map. It's assumed that anon maps don't share anon
3138  * structures with private anon maps. Therefore all anon structures should
3139  * have at most one reference at this point. This means underlying pages can
3140  * be exclusively locked and demoted or freed.  If not freeing the entire
3141  * large pages demote the ends of the region we free to be able to free
3142  * subpages. Page roots correspond to aligned index positions in anon map.
3143  */
3144 void
3145 anon_shmap_free_pages(struct anon_map *amp, ulong_t sidx, size_t len)
3146 {
3147 	ulong_t eidx = sidx + btopr(len);
3148 	pgcnt_t pages = page_get_pagecnt(amp->a_szc);
3149 	struct anon_hdr *ahp = amp->ahp;
3150 	ulong_t tidx;
3151 	size_t size;
3152 	ulong_t sidx_aligned;
3153 	ulong_t eidx_aligned;
3154 
3155 	ASSERT(RW_WRITE_HELD(&amp->a_rwlock));
3156 	ASSERT(amp->refcnt <= 1);
3157 	ASSERT(amp->a_szc > 0);
3158 	ASSERT(eidx <= ahp->size);
3159 	ASSERT(!anon_share(ahp, sidx, btopr(len)));
3160 
3161 	if (len == 0) {	/* XXX */
3162 		return;
3163 	}
3164 
3165 	sidx_aligned = P2ALIGN(sidx, pages);
3166 	if (sidx_aligned != sidx ||
3167 	    (eidx < sidx_aligned + pages && eidx < ahp->size)) {
3168 		if (!anon_try_demote_pages(ahp, sidx_aligned,
3169 		    amp->a_szc, NULL, 0)) {
3170 			panic("anon_shmap_free_pages: demote failed");
3171 		}
3172 		size = (eidx <= sidx_aligned + pages) ? (eidx - sidx) :
3173 		    P2NPHASE(sidx, pages);
3174 		size <<= PAGESHIFT;
3175 		anon_free(ahp, sidx, size);
3176 		sidx = sidx_aligned + pages;
3177 		if (eidx <= sidx) {
3178 			return;
3179 		}
3180 	}
3181 	eidx_aligned = P2ALIGN(eidx, pages);
3182 	if (sidx < eidx_aligned) {
3183 		anon_free_pages(ahp, sidx,
3184 		    (eidx_aligned - sidx) << PAGESHIFT,
3185 		    amp->a_szc);
3186 		sidx = eidx_aligned;
3187 	}
3188 	ASSERT(sidx == eidx_aligned);
3189 	if (eidx == eidx_aligned) {
3190 		return;
3191 	}
3192 	tidx = eidx;
3193 	if (eidx != ahp->size && anon_get_next_ptr(ahp, &tidx) != NULL &&
3194 	    tidx - sidx < pages) {
3195 		if (!anon_try_demote_pages(ahp, sidx, amp->a_szc, NULL, 0)) {
3196 			panic("anon_shmap_free_pages: demote failed");
3197 		}
3198 		size = (eidx - sidx) << PAGESHIFT;
3199 		anon_free(ahp, sidx, size);
3200 	} else {
3201 		anon_free_pages(ahp, sidx, pages << PAGESHIFT, amp->a_szc);
3202 	}
3203 }
3204 
3205 /*
3206  * Allocate and initialize an anon_map structure for seg
3207  * associating the given swap reservation with the new anon_map.
3208  */
3209 struct anon_map *
3210 anonmap_alloc(size_t size, size_t swresv, int flags)
3211 {
3212 	struct anon_map *amp;
3213 	int kmflags = (flags & ANON_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
3214 
3215 	amp = kmem_cache_alloc(anonmap_cache, kmflags);
3216 	if (amp == NULL) {
3217 		ASSERT(kmflags == KM_NOSLEEP);
3218 		return (NULL);
3219 	}
3220 
3221 	amp->ahp = anon_create(btopr(size), flags);
3222 	if (amp->ahp == NULL) {
3223 		ASSERT(flags == ANON_NOSLEEP);
3224 		kmem_cache_free(anonmap_cache, amp);
3225 		return (NULL);
3226 	}
3227 	amp->refcnt = 1;
3228 	amp->size = size;
3229 	amp->swresv = swresv;
3230 	amp->locality = 0;
3231 	amp->a_szc = 0;
3232 	amp->a_sp = NULL;
3233 	return (amp);
3234 }
3235 
3236 void
3237 anonmap_free(struct anon_map *amp)
3238 {
3239 	ASSERT(amp->ahp);
3240 	ASSERT(amp->refcnt == 0);
3241 
3242 	lgrp_shm_policy_fini(amp, NULL);
3243 	anon_release(amp->ahp, btopr(amp->size));
3244 	kmem_cache_free(anonmap_cache, amp);
3245 }
3246 
3247 /*
3248  * Returns true if the app array has some empty slots.
3249  * The offp and lenp parameters are in/out parameters.  On entry
3250  * these values represent the starting offset and length of the
3251  * mapping.  When true is returned, these values may be modified
3252  * to be the largest range which includes empty slots.
3253  */
3254 int
3255 non_anon(struct anon_hdr *ahp, ulong_t anon_idx, u_offset_t *offp,
3256 				size_t *lenp)
3257 {
3258 	ulong_t i, el;
3259 	ssize_t low, high;
3260 	struct anon *ap;
3261 
3262 	low = -1;
3263 	for (i = 0, el = *lenp; i < el; i += PAGESIZE, anon_idx++) {
3264 		ap = anon_get_ptr(ahp, anon_idx);
3265 		if (ap == NULL) {
3266 			if (low == -1)
3267 				low = i;
3268 			high = i;
3269 		}
3270 	}
3271 	if (low != -1) {
3272 		/*
3273 		 * Found at least one non-anon page.
3274 		 * Set up the off and len return values.
3275 		 */
3276 		if (low != 0)
3277 			*offp += low;
3278 		*lenp = high - low + PAGESIZE;
3279 		return (1);
3280 	}
3281 	return (0);
3282 }
3283 
3284 /*
3285  * Return a count of the number of existing anon pages in the anon array
3286  * app in the range (off, off+len). The array and slots must be guaranteed
3287  * stable by the caller.
3288  */
3289 pgcnt_t
3290 anon_pages(struct anon_hdr *ahp, ulong_t anon_index, pgcnt_t nslots)
3291 {
3292 	pgcnt_t cnt = 0;
3293 
3294 	while (nslots-- > 0) {
3295 		if ((anon_get_ptr(ahp, anon_index)) != NULL)
3296 			cnt++;
3297 		anon_index++;
3298 	}
3299 	return (cnt);
3300 }
3301 
3302 /*
3303  * Move reserved phys swap into memory swap (unreserve phys swap
3304  * and reserve mem swap by the same amount).
3305  * Used by segspt when it needs to lock reserved swap npages in memory
3306  */
3307 int
3308 anon_swap_adjust(pgcnt_t npages)
3309 {
3310 	pgcnt_t unlocked_mem_swap;
3311 
3312 	mutex_enter(&anoninfo_lock);
3313 
3314 	ASSERT(k_anoninfo.ani_mem_resv >= k_anoninfo.ani_locked_swap);
3315 	ASSERT(k_anoninfo.ani_max >= k_anoninfo.ani_phys_resv);
3316 
3317 	unlocked_mem_swap = k_anoninfo.ani_mem_resv
3318 	    - k_anoninfo.ani_locked_swap;
3319 	if (npages > unlocked_mem_swap) {
3320 		spgcnt_t adjusted_swap = npages - unlocked_mem_swap;
3321 
3322 		/*
3323 		 * if there is not enough unlocked mem swap we take missing
3324 		 * amount from phys swap and give it to mem swap
3325 		 */
3326 		if (!page_reclaim_mem(adjusted_swap, segspt_minfree, 1)) {
3327 			mutex_exit(&anoninfo_lock);
3328 			return (ENOMEM);
3329 		}
3330 
3331 		k_anoninfo.ani_mem_resv += adjusted_swap;
3332 		ASSERT(k_anoninfo.ani_phys_resv >= adjusted_swap);
3333 		k_anoninfo.ani_phys_resv -= adjusted_swap;
3334 
3335 		ANI_ADD(adjusted_swap);
3336 	}
3337 	k_anoninfo.ani_locked_swap += npages;
3338 
3339 	ASSERT(k_anoninfo.ani_mem_resv >= k_anoninfo.ani_locked_swap);
3340 	ASSERT(k_anoninfo.ani_max >= k_anoninfo.ani_phys_resv);
3341 
3342 	mutex_exit(&anoninfo_lock);
3343 
3344 	return (0);
3345 }
3346 
3347 /*
3348  * 'unlocked' reserved mem swap so when it is unreserved it
3349  * can be moved back phys (disk) swap
3350  */
3351 void
3352 anon_swap_restore(pgcnt_t npages)
3353 {
3354 	mutex_enter(&anoninfo_lock);
3355 
3356 	ASSERT(k_anoninfo.ani_locked_swap <= k_anoninfo.ani_mem_resv);
3357 
3358 	ASSERT(k_anoninfo.ani_locked_swap >= npages);
3359 	k_anoninfo.ani_locked_swap -= npages;
3360 
3361 	ASSERT(k_anoninfo.ani_locked_swap <= k_anoninfo.ani_mem_resv);
3362 
3363 	mutex_exit(&anoninfo_lock);
3364 }
3365 
3366 /*
3367  * Return the pointer from the list for a
3368  * specified anon index.
3369  */
3370 ulong_t *
3371 anon_get_slot(struct anon_hdr *ahp, ulong_t an_idx)
3372 {
3373 	struct anon	**app;
3374 	void 		**ppp;
3375 
3376 	ASSERT(an_idx < ahp->size);
3377 
3378 	/*
3379 	 * Single level case.
3380 	 */
3381 	if ((ahp->size <= ANON_CHUNK_SIZE) || (ahp->flags & ANON_ALLOC_FORCE)) {
3382 		return ((ulong_t *)&ahp->array_chunk[an_idx]);
3383 	} else {
3384 
3385 		/*
3386 		 * 2 level case.
3387 		 */
3388 		ppp = &ahp->array_chunk[an_idx >> ANON_CHUNK_SHIFT];
3389 		if (*ppp == NULL) {
3390 			mutex_enter(&ahp->serial_lock);
3391 			ppp = &ahp->array_chunk[an_idx >> ANON_CHUNK_SHIFT];
3392 			if (*ppp == NULL)
3393 				*ppp = kmem_zalloc(PAGESIZE, KM_SLEEP);
3394 			mutex_exit(&ahp->serial_lock);
3395 		}
3396 		app = *ppp;
3397 		return ((ulong_t *)&app[an_idx & ANON_CHUNK_OFF]);
3398 	}
3399 }
3400 
3401 void
3402 anon_array_enter(struct anon_map *amp, ulong_t an_idx, anon_sync_obj_t *sobj)
3403 {
3404 	ulong_t		*ap_slot;
3405 	kmutex_t	*mtx;
3406 	kcondvar_t	*cv;
3407 	int		hash;
3408 
3409 	/*
3410 	 * Use szc to determine anon slot(s) to appear atomic.
3411 	 * If szc = 0, then lock the anon slot and mark it busy.
3412 	 * If szc > 0, then lock the range of slots by getting the
3413 	 * anon_array_lock for the first anon slot, and mark only the
3414 	 * first anon slot busy to represent whole range being busy.
3415 	 */
3416 
3417 	ASSERT(RW_READ_HELD(&amp->a_rwlock));
3418 	an_idx = P2ALIGN(an_idx, page_get_pagecnt(amp->a_szc));
3419 	hash = ANON_ARRAY_HASH(amp, an_idx);
3420 	sobj->sync_mutex = mtx = &anon_array_lock[hash].pad_mutex;
3421 	sobj->sync_cv = cv = &anon_array_cv[hash];
3422 	mutex_enter(mtx);
3423 	ap_slot = anon_get_slot(amp->ahp, an_idx);
3424 	while (ANON_ISBUSY(ap_slot))
3425 		cv_wait(cv, mtx);
3426 	ANON_SETBUSY(ap_slot);
3427 	sobj->sync_data = ap_slot;
3428 	mutex_exit(mtx);
3429 }
3430 
3431 int
3432 anon_array_try_enter(struct anon_map *amp, ulong_t an_idx,
3433 			anon_sync_obj_t *sobj)
3434 {
3435 	ulong_t		*ap_slot;
3436 	kmutex_t	*mtx;
3437 	int		hash;
3438 
3439 	/*
3440 	 * Try to lock a range of anon slots.
3441 	 * Use szc to determine anon slot(s) to appear atomic.
3442 	 * If szc = 0, then lock the anon slot and mark it busy.
3443 	 * If szc > 0, then lock the range of slots by getting the
3444 	 * anon_array_lock for the first anon slot, and mark only the
3445 	 * first anon slot busy to represent whole range being busy.
3446 	 * Fail if the mutex or the anon_array are busy.
3447 	 */
3448 
3449 	ASSERT(RW_READ_HELD(&amp->a_rwlock));
3450 	an_idx = P2ALIGN(an_idx, page_get_pagecnt(amp->a_szc));
3451 	hash = ANON_ARRAY_HASH(amp, an_idx);
3452 	sobj->sync_mutex = mtx = &anon_array_lock[hash].pad_mutex;
3453 	sobj->sync_cv = &anon_array_cv[hash];
3454 	if (!mutex_tryenter(mtx)) {
3455 		return (EWOULDBLOCK);
3456 	}
3457 	ap_slot = anon_get_slot(amp->ahp, an_idx);
3458 	if (ANON_ISBUSY(ap_slot)) {
3459 		mutex_exit(mtx);
3460 		return (EWOULDBLOCK);
3461 	}
3462 	ANON_SETBUSY(ap_slot);
3463 	sobj->sync_data = ap_slot;
3464 	mutex_exit(mtx);
3465 	return (0);
3466 }
3467 
3468 void
3469 anon_array_exit(anon_sync_obj_t *sobj)
3470 {
3471 	mutex_enter(sobj->sync_mutex);
3472 	ASSERT(ANON_ISBUSY(sobj->sync_data));
3473 	ANON_CLRBUSY(sobj->sync_data);
3474 	if (CV_HAS_WAITERS(sobj->sync_cv))
3475 		cv_broadcast(sobj->sync_cv);
3476 	mutex_exit(sobj->sync_mutex);
3477 }
3478