xref: /illumos-gate/usr/src/uts/common/vm/vm_page.c (revision bb57d1f5164aca913cbd286ae1b61c896167cfa7)
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) 1983, 1984, 1985, 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 - physical page management.
43  */
44 
45 #include <sys/types.h>
46 #include <sys/t_lock.h>
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/errno.h>
50 #include <sys/time.h>
51 #include <sys/vnode.h>
52 #include <sys/vm.h>
53 #include <sys/vtrace.h>
54 #include <sys/swap.h>
55 #include <sys/cmn_err.h>
56 #include <sys/tuneable.h>
57 #include <sys/sysmacros.h>
58 #include <sys/cpuvar.h>
59 #include <sys/callb.h>
60 #include <sys/debug.h>
61 #include <sys/tnf_probe.h>
62 #include <sys/condvar_impl.h>
63 #include <sys/mem_config.h>
64 #include <sys/mem_cage.h>
65 #include <sys/kmem.h>
66 #include <sys/atomic.h>
67 #include <sys/strlog.h>
68 #include <sys/mman.h>
69 #include <sys/ontrap.h>
70 #include <sys/lgrp.h>
71 #include <sys/vfs.h>
72 
73 #include <vm/hat.h>
74 #include <vm/anon.h>
75 #include <vm/page.h>
76 #include <vm/seg.h>
77 #include <vm/pvn.h>
78 #include <vm/seg_kmem.h>
79 #include <vm/vm_dep.h>
80 #include <sys/vm_usage.h>
81 #include <fs/fs_subr.h>
82 #include <sys/ddi.h>
83 #include <sys/modctl.h>
84 
85 static int nopageage = 0;
86 
87 static pgcnt_t max_page_get;	/* max page_get request size in pages */
88 pgcnt_t total_pages = 0;	/* total number of pages (used by /proc) */
89 
90 /*
91  * freemem_lock protects all freemem variables:
92  * availrmem. Also this lock protects the globals which track the
93  * availrmem changes for accurate kernel footprint calculation.
94  * See below for an explanation of these
95  * globals.
96  */
97 kmutex_t freemem_lock;
98 pgcnt_t availrmem;
99 pgcnt_t availrmem_initial;
100 
101 /*
102  * These globals track availrmem changes to get a more accurate
103  * estimate of tke kernel size. Historically pp_kernel is used for
104  * kernel size and is based on availrmem. But availrmem is adjusted for
105  * locked pages in the system not just for kernel locked pages.
106  * These new counters will track the pages locked through segvn and
107  * by explicit user locking.
108  *
109  * segvn_pages_locked : This keeps track on a global basis how many pages
110  * are currently locked because of I/O.
111  *
112  * pages_locked : How many pages are locked because of user specified
113  * locking through mlock or plock.
114  *
115  * pages_useclaim,pages_claimed : These two variables track the
116  * claim adjustments because of the protection changes on a segvn segment.
117  *
118  * All these globals are protected by the same lock which protects availrmem.
119  */
120 pgcnt_t segvn_pages_locked;
121 pgcnt_t pages_locked;
122 pgcnt_t pages_useclaim;
123 pgcnt_t pages_claimed;
124 
125 
126 /*
127  * new_freemem_lock protects freemem, freemem_wait & freemem_cv.
128  */
129 static kmutex_t	new_freemem_lock;
130 static uint_t	freemem_wait;	/* someone waiting for freemem */
131 static kcondvar_t freemem_cv;
132 
133 /*
134  * The logical page free list is maintained as two lists, the 'free'
135  * and the 'cache' lists.
136  * The free list contains those pages that should be reused first.
137  *
138  * The implementation of the lists is machine dependent.
139  * page_get_freelist(), page_get_cachelist(),
140  * page_list_sub(), and page_list_add()
141  * form the interface to the machine dependent implementation.
142  *
143  * Pages with p_free set are on the cache list.
144  * Pages with p_free and p_age set are on the free list,
145  *
146  * A page may be locked while on either list.
147  */
148 
149 /*
150  * free list accounting stuff.
151  *
152  *
153  * Spread out the value for the number of pages on the
154  * page free and page cache lists.  If there is just one
155  * value, then it must be under just one lock.
156  * The lock contention and cache traffic are a real bother.
157  *
158  * When we acquire and then drop a single pcf lock
159  * we can start in the middle of the array of pcf structures.
160  * If we acquire more than one pcf lock at a time, we need to
161  * start at the front to avoid deadlocking.
162  *
163  * pcf_count holds the number of pages in each pool.
164  *
165  * pcf_block is set when page_create_get_something() has asked the
166  * PSM page freelist and page cachelist routines without specifying
167  * a color and nothing came back.  This is used to block anything
168  * else from moving pages from one list to the other while the
169  * lists are searched again.  If a page is freeed while pcf_block is
170  * set, then pcf_reserve is incremented.  pcgs_unblock() takes care
171  * of clearning pcf_block, doing the wakeups, etc.
172  */
173 
174 #if NCPU <= 4
175 #define	PAD	2
176 #define	PCF_FANOUT	4
177 static	uint_t	pcf_mask = PCF_FANOUT - 1;
178 #else
179 #define	PAD	10
180 #ifdef sun4v
181 #define	PCF_FANOUT	32
182 #else
183 #define	PCF_FANOUT	128
184 #endif
185 static	uint_t	pcf_mask = PCF_FANOUT - 1;
186 #endif
187 
188 struct pcf {
189 	kmutex_t	pcf_lock;	/* protects the structure */
190 	uint_t		pcf_count;	/* page count */
191 	uint_t		pcf_wait;	/* number of waiters */
192 	uint_t		pcf_block; 	/* pcgs flag to page_free() */
193 	uint_t		pcf_reserve; 	/* pages freed after pcf_block set */
194 	uint_t		pcf_fill[PAD];	/* to line up on the caches */
195 };
196 
197 static struct	pcf	pcf[PCF_FANOUT];
198 #define	PCF_INDEX()	((CPU->cpu_id) & (pcf_mask))
199 
200 kmutex_t	pcgs_lock;		/* serializes page_create_get_ */
201 kmutex_t	pcgs_cagelock;		/* serializes NOSLEEP cage allocs */
202 kmutex_t	pcgs_wait_lock;		/* used for delay in pcgs */
203 static kcondvar_t	pcgs_cv;	/* cv for delay in pcgs */
204 
205 #ifdef VM_STATS
206 
207 /*
208  * No locks, but so what, they are only statistics.
209  */
210 
211 static struct page_tcnt {
212 	int	pc_free_cache;		/* free's into cache list */
213 	int	pc_free_dontneed;	/* free's with dontneed */
214 	int	pc_free_pageout;	/* free's from pageout */
215 	int	pc_free_free;		/* free's into free list */
216 	int	pc_free_pages;		/* free's into large page free list */
217 	int	pc_destroy_pages;	/* large page destroy's */
218 	int	pc_get_cache;		/* get's from cache list */
219 	int	pc_get_free;		/* get's from free list */
220 	int	pc_reclaim;		/* reclaim's */
221 	int	pc_abortfree;		/* abort's of free pages */
222 	int	pc_find_hit;		/* find's that find page */
223 	int	pc_find_miss;		/* find's that don't find page */
224 	int	pc_destroy_free;	/* # of free pages destroyed */
225 #define	PC_HASH_CNT	(4*PAGE_HASHAVELEN)
226 	int	pc_find_hashlen[PC_HASH_CNT+1];
227 	int	pc_addclaim_pages;
228 	int	pc_subclaim_pages;
229 	int	pc_free_replacement_page[2];
230 	int	pc_try_demote_pages[6];
231 	int	pc_demote_pages[2];
232 } pagecnt;
233 
234 uint_t	hashin_count;
235 uint_t	hashin_not_held;
236 uint_t	hashin_already;
237 
238 uint_t	hashout_count;
239 uint_t	hashout_not_held;
240 
241 uint_t	page_create_count;
242 uint_t	page_create_not_enough;
243 uint_t	page_create_not_enough_again;
244 uint_t	page_create_zero;
245 uint_t	page_create_hashout;
246 uint_t	page_create_page_lock_failed;
247 uint_t	page_create_trylock_failed;
248 uint_t	page_create_found_one;
249 uint_t	page_create_hashin_failed;
250 uint_t	page_create_dropped_phm;
251 
252 uint_t	page_create_new;
253 uint_t	page_create_exists;
254 uint_t	page_create_putbacks;
255 uint_t	page_create_overshoot;
256 
257 uint_t	page_reclaim_zero;
258 uint_t	page_reclaim_zero_locked;
259 
260 uint_t	page_rename_exists;
261 uint_t	page_rename_count;
262 
263 uint_t	page_lookup_cnt[20];
264 uint_t	page_lookup_nowait_cnt[10];
265 uint_t	page_find_cnt;
266 uint_t	page_exists_cnt;
267 uint_t	page_exists_forreal_cnt;
268 uint_t	page_lookup_dev_cnt;
269 uint_t	get_cachelist_cnt;
270 uint_t	page_create_cnt[10];
271 uint_t	alloc_pages[8];
272 uint_t	page_exphcontg[19];
273 uint_t  page_create_large_cnt[10];
274 
275 /*
276  * Collects statistics.
277  */
278 #define	PAGE_HASH_SEARCH(index, pp, vp, off) { \
279 	uint_t	mylen = 0; \
280 			\
281 	for ((pp) = page_hash[(index)]; (pp); (pp) = (pp)->p_hash, mylen++) { \
282 		if ((pp)->p_vnode == (vp) && (pp)->p_offset == (off)) \
283 			break; \
284 	} \
285 	if ((pp) != NULL) \
286 		pagecnt.pc_find_hit++; \
287 	else \
288 		pagecnt.pc_find_miss++; \
289 	if (mylen > PC_HASH_CNT) \
290 		mylen = PC_HASH_CNT; \
291 	pagecnt.pc_find_hashlen[mylen]++; \
292 }
293 
294 #else	/* VM_STATS */
295 
296 /*
297  * Don't collect statistics
298  */
299 #define	PAGE_HASH_SEARCH(index, pp, vp, off) { \
300 	for ((pp) = page_hash[(index)]; (pp); (pp) = (pp)->p_hash) { \
301 		if ((pp)->p_vnode == (vp) && (pp)->p_offset == (off)) \
302 			break; \
303 	} \
304 }
305 
306 #endif	/* VM_STATS */
307 
308 
309 
310 #ifdef DEBUG
311 #define	MEMSEG_SEARCH_STATS
312 #endif
313 
314 #ifdef MEMSEG_SEARCH_STATS
315 struct memseg_stats {
316     uint_t nsearch;
317     uint_t nlastwon;
318     uint_t nhashwon;
319     uint_t nnotfound;
320 } memseg_stats;
321 
322 #define	MEMSEG_STAT_INCR(v) \
323 	atomic_add_32(&memseg_stats.v, 1)
324 #else
325 #define	MEMSEG_STAT_INCR(x)
326 #endif
327 
328 struct memseg *memsegs;		/* list of memory segments */
329 
330 
331 static void page_init_mem_config(void);
332 static int page_do_hashin(page_t *, vnode_t *, u_offset_t);
333 static void page_do_hashout(page_t *);
334 static void page_capture_init();
335 int page_capture_take_action(page_t *, uint_t, void *);
336 
337 static void page_demote_vp_pages(page_t *);
338 
339 /*
340  * vm subsystem related initialization
341  */
342 void
343 vm_init(void)
344 {
345 	boolean_t callb_vm_cpr(void *, int);
346 
347 	(void) callb_add(callb_vm_cpr, 0, CB_CL_CPR_VM, "vm");
348 	page_init_mem_config();
349 	page_retire_init();
350 	vm_usage_init();
351 	page_capture_init();
352 }
353 
354 /*
355  * This function is called at startup and when memory is added or deleted.
356  */
357 void
358 init_pages_pp_maximum()
359 {
360 	static pgcnt_t p_min;
361 	static pgcnt_t pages_pp_maximum_startup;
362 	static pgcnt_t avrmem_delta;
363 	static int init_done;
364 	static int user_set;	/* true if set in /etc/system */
365 
366 	if (init_done == 0) {
367 
368 		/* If the user specified a value, save it */
369 		if (pages_pp_maximum != 0) {
370 			user_set = 1;
371 			pages_pp_maximum_startup = pages_pp_maximum;
372 		}
373 
374 		/*
375 		 * Setting of pages_pp_maximum is based first time
376 		 * on the value of availrmem just after the start-up
377 		 * allocations. To preserve this relationship at run
378 		 * time, use a delta from availrmem_initial.
379 		 */
380 		ASSERT(availrmem_initial >= availrmem);
381 		avrmem_delta = availrmem_initial - availrmem;
382 
383 		/* The allowable floor of pages_pp_maximum */
384 		p_min = tune.t_minarmem + 100;
385 
386 		/* Make sure we don't come through here again. */
387 		init_done = 1;
388 	}
389 	/*
390 	 * Determine pages_pp_maximum, the number of currently available
391 	 * pages (availrmem) that can't be `locked'. If not set by
392 	 * the user, we set it to 4% of the currently available memory
393 	 * plus 4MB.
394 	 * But we also insist that it be greater than tune.t_minarmem;
395 	 * otherwise a process could lock down a lot of memory, get swapped
396 	 * out, and never have enough to get swapped back in.
397 	 */
398 	if (user_set)
399 		pages_pp_maximum = pages_pp_maximum_startup;
400 	else
401 		pages_pp_maximum = ((availrmem_initial - avrmem_delta) / 25)
402 		    + btop(4 * 1024 * 1024);
403 
404 	if (pages_pp_maximum <= p_min) {
405 		pages_pp_maximum = p_min;
406 	}
407 }
408 
409 void
410 set_max_page_get(pgcnt_t target_total_pages)
411 {
412 	max_page_get = target_total_pages / 2;
413 }
414 
415 static pgcnt_t pending_delete;
416 
417 /*ARGSUSED*/
418 static void
419 page_mem_config_post_add(
420 	void *arg,
421 	pgcnt_t delta_pages)
422 {
423 	set_max_page_get(total_pages - pending_delete);
424 	init_pages_pp_maximum();
425 }
426 
427 /*ARGSUSED*/
428 static int
429 page_mem_config_pre_del(
430 	void *arg,
431 	pgcnt_t delta_pages)
432 {
433 	pgcnt_t nv;
434 
435 	nv = atomic_add_long_nv(&pending_delete, (spgcnt_t)delta_pages);
436 	set_max_page_get(total_pages - nv);
437 	return (0);
438 }
439 
440 /*ARGSUSED*/
441 static void
442 page_mem_config_post_del(
443 	void *arg,
444 	pgcnt_t delta_pages,
445 	int cancelled)
446 {
447 	pgcnt_t nv;
448 
449 	nv = atomic_add_long_nv(&pending_delete, -(spgcnt_t)delta_pages);
450 	set_max_page_get(total_pages - nv);
451 	if (!cancelled)
452 		init_pages_pp_maximum();
453 }
454 
455 static kphysm_setup_vector_t page_mem_config_vec = {
456 	KPHYSM_SETUP_VECTOR_VERSION,
457 	page_mem_config_post_add,
458 	page_mem_config_pre_del,
459 	page_mem_config_post_del,
460 };
461 
462 static void
463 page_init_mem_config(void)
464 {
465 	int ret;
466 
467 	ret = kphysm_setup_func_register(&page_mem_config_vec, (void *)NULL);
468 	ASSERT(ret == 0);
469 }
470 
471 /*
472  * Evenly spread out the PCF counters for large free pages
473  */
474 static void
475 page_free_large_ctr(pgcnt_t npages)
476 {
477 	static struct pcf	*p = pcf;
478 	pgcnt_t			lump;
479 
480 	freemem += npages;
481 
482 	lump = roundup(npages, PCF_FANOUT) / PCF_FANOUT;
483 
484 	while (npages > 0) {
485 
486 		ASSERT(!p->pcf_block);
487 
488 		if (lump < npages) {
489 			p->pcf_count += (uint_t)lump;
490 			npages -= lump;
491 		} else {
492 			p->pcf_count += (uint_t)npages;
493 			npages = 0;
494 		}
495 
496 		ASSERT(!p->pcf_wait);
497 
498 		if (++p > &pcf[PCF_FANOUT - 1])
499 			p = pcf;
500 	}
501 
502 	ASSERT(npages == 0);
503 }
504 
505 /*
506  * Add a physical chunk of memory to the system free lists during startup.
507  * Platform specific startup() allocates the memory for the page structs.
508  *
509  * num	- number of page structures
510  * base - page number (pfn) to be associated with the first page.
511  *
512  * Since we are doing this during startup (ie. single threaded), we will
513  * use shortcut routines to avoid any locking overhead while putting all
514  * these pages on the freelists.
515  *
516  * NOTE: Any changes performed to page_free(), must also be performed to
517  *	 add_physmem() since this is how we initialize all page_t's at
518  *	 boot time.
519  */
520 void
521 add_physmem(
522 	page_t	*pp,
523 	pgcnt_t	num,
524 	pfn_t	pnum)
525 {
526 	page_t	*root = NULL;
527 	uint_t	szc = page_num_pagesizes() - 1;
528 	pgcnt_t	large = page_get_pagecnt(szc);
529 	pgcnt_t	cnt = 0;
530 
531 	TRACE_2(TR_FAC_VM, TR_PAGE_INIT,
532 	    "add_physmem:pp %p num %lu", pp, num);
533 
534 	/*
535 	 * Arbitrarily limit the max page_get request
536 	 * to 1/2 of the page structs we have.
537 	 */
538 	total_pages += num;
539 	set_max_page_get(total_pages);
540 
541 	PLCNT_MODIFY_MAX(pnum, (long)num);
542 
543 	/*
544 	 * The physical space for the pages array
545 	 * representing ram pages has already been
546 	 * allocated.  Here we initialize each lock
547 	 * in the page structure, and put each on
548 	 * the free list
549 	 */
550 	for (; num; pp++, pnum++, num--) {
551 
552 		/*
553 		 * this needs to fill in the page number
554 		 * and do any other arch specific initialization
555 		 */
556 		add_physmem_cb(pp, pnum);
557 
558 		pp->p_lckcnt = 0;
559 		pp->p_cowcnt = 0;
560 		pp->p_slckcnt = 0;
561 
562 		/*
563 		 * Initialize the page lock as unlocked, since nobody
564 		 * can see or access this page yet.
565 		 */
566 		pp->p_selock = 0;
567 
568 		/*
569 		 * Initialize IO lock
570 		 */
571 		page_iolock_init(pp);
572 
573 		/*
574 		 * initialize other fields in the page_t
575 		 */
576 		PP_SETFREE(pp);
577 		page_clr_all_props(pp);
578 		PP_SETAGED(pp);
579 		pp->p_offset = (u_offset_t)-1;
580 		pp->p_next = pp;
581 		pp->p_prev = pp;
582 
583 		/*
584 		 * Simple case: System doesn't support large pages.
585 		 */
586 		if (szc == 0) {
587 			pp->p_szc = 0;
588 			page_free_at_startup(pp);
589 			continue;
590 		}
591 
592 		/*
593 		 * Handle unaligned pages, we collect them up onto
594 		 * the root page until we have a full large page.
595 		 */
596 		if (!IS_P2ALIGNED(pnum, large)) {
597 
598 			/*
599 			 * If not in a large page,
600 			 * just free as small page.
601 			 */
602 			if (root == NULL) {
603 				pp->p_szc = 0;
604 				page_free_at_startup(pp);
605 				continue;
606 			}
607 
608 			/*
609 			 * Link a constituent page into the large page.
610 			 */
611 			pp->p_szc = szc;
612 			page_list_concat(&root, &pp);
613 
614 			/*
615 			 * When large page is fully formed, free it.
616 			 */
617 			if (++cnt == large) {
618 				page_free_large_ctr(cnt);
619 				page_list_add_pages(root, PG_LIST_ISINIT);
620 				root = NULL;
621 				cnt = 0;
622 			}
623 			continue;
624 		}
625 
626 		/*
627 		 * At this point we have a page number which
628 		 * is aligned. We assert that we aren't already
629 		 * in a different large page.
630 		 */
631 		ASSERT(IS_P2ALIGNED(pnum, large));
632 		ASSERT(root == NULL && cnt == 0);
633 
634 		/*
635 		 * If insufficient number of pages left to form
636 		 * a large page, just free the small page.
637 		 */
638 		if (num < large) {
639 			pp->p_szc = 0;
640 			page_free_at_startup(pp);
641 			continue;
642 		}
643 
644 		/*
645 		 * Otherwise start a new large page.
646 		 */
647 		pp->p_szc = szc;
648 		cnt++;
649 		root = pp;
650 	}
651 	ASSERT(root == NULL && cnt == 0);
652 }
653 
654 /*
655  * Find a page representing the specified [vp, offset].
656  * If we find the page but it is intransit coming in,
657  * it will have an "exclusive" lock and we wait for
658  * the i/o to complete.  A page found on the free list
659  * is always reclaimed and then locked.  On success, the page
660  * is locked, its data is valid and it isn't on the free
661  * list, while a NULL is returned if the page doesn't exist.
662  */
663 page_t *
664 page_lookup(vnode_t *vp, u_offset_t off, se_t se)
665 {
666 	return (page_lookup_create(vp, off, se, NULL, NULL, 0));
667 }
668 
669 /*
670  * Find a page representing the specified [vp, offset].
671  * We either return the one we found or, if passed in,
672  * create one with identity of [vp, offset] of the
673  * pre-allocated page. If we find existing page but it is
674  * intransit coming in, it will have an "exclusive" lock
675  * and we wait for the i/o to complete.  A page found on
676  * the free list is always reclaimed and then locked.
677  * On success, the page is locked, its data is valid and
678  * it isn't on the free list, while a NULL is returned
679  * if the page doesn't exist and newpp is NULL;
680  */
681 page_t *
682 page_lookup_create(
683 	vnode_t *vp,
684 	u_offset_t off,
685 	se_t se,
686 	page_t *newpp,
687 	spgcnt_t *nrelocp,
688 	int flags)
689 {
690 	page_t		*pp;
691 	kmutex_t	*phm;
692 	ulong_t		index;
693 	uint_t		hash_locked;
694 	uint_t		es;
695 
696 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
697 	VM_STAT_ADD(page_lookup_cnt[0]);
698 	ASSERT(newpp ? PAGE_EXCL(newpp) : 1);
699 
700 	/*
701 	 * Acquire the appropriate page hash lock since
702 	 * we have to search the hash list.  Pages that
703 	 * hash to this list can't change identity while
704 	 * this lock is held.
705 	 */
706 	hash_locked = 0;
707 	index = PAGE_HASH_FUNC(vp, off);
708 	phm = NULL;
709 top:
710 	PAGE_HASH_SEARCH(index, pp, vp, off);
711 	if (pp != NULL) {
712 		VM_STAT_ADD(page_lookup_cnt[1]);
713 		es = (newpp != NULL) ? 1 : 0;
714 		es |= flags;
715 		if (!hash_locked) {
716 			VM_STAT_ADD(page_lookup_cnt[2]);
717 			if (!page_try_reclaim_lock(pp, se, es)) {
718 				/*
719 				 * On a miss, acquire the phm.  Then
720 				 * next time, page_lock() will be called,
721 				 * causing a wait if the page is busy.
722 				 * just looping with page_trylock() would
723 				 * get pretty boring.
724 				 */
725 				VM_STAT_ADD(page_lookup_cnt[3]);
726 				phm = PAGE_HASH_MUTEX(index);
727 				mutex_enter(phm);
728 				hash_locked = 1;
729 				goto top;
730 			}
731 		} else {
732 			VM_STAT_ADD(page_lookup_cnt[4]);
733 			if (!page_lock_es(pp, se, phm, P_RECLAIM, es)) {
734 				VM_STAT_ADD(page_lookup_cnt[5]);
735 				goto top;
736 			}
737 		}
738 
739 		/*
740 		 * Since `pp' is locked it can not change identity now.
741 		 * Reconfirm we locked the correct page.
742 		 *
743 		 * Both the p_vnode and p_offset *must* be cast volatile
744 		 * to force a reload of their values: The PAGE_HASH_SEARCH
745 		 * macro will have stuffed p_vnode and p_offset into
746 		 * registers before calling page_trylock(); another thread,
747 		 * actually holding the hash lock, could have changed the
748 		 * page's identity in memory, but our registers would not
749 		 * be changed, fooling the reconfirmation.  If the hash
750 		 * lock was held during the search, the casting would
751 		 * not be needed.
752 		 */
753 		VM_STAT_ADD(page_lookup_cnt[6]);
754 		if (((volatile struct vnode *)(pp->p_vnode) != vp) ||
755 		    ((volatile u_offset_t)(pp->p_offset) != off)) {
756 			VM_STAT_ADD(page_lookup_cnt[7]);
757 			if (hash_locked) {
758 				panic("page_lookup_create: lost page %p",
759 				    (void *)pp);
760 				/*NOTREACHED*/
761 			}
762 			page_unlock(pp);
763 			phm = PAGE_HASH_MUTEX(index);
764 			mutex_enter(phm);
765 			hash_locked = 1;
766 			goto top;
767 		}
768 
769 		/*
770 		 * If page_trylock() was called, then pp may still be on
771 		 * the cachelist (can't be on the free list, it would not
772 		 * have been found in the search).  If it is on the
773 		 * cachelist it must be pulled now. To pull the page from
774 		 * the cachelist, it must be exclusively locked.
775 		 *
776 		 * The other big difference between page_trylock() and
777 		 * page_lock(), is that page_lock() will pull the
778 		 * page from whatever free list (the cache list in this
779 		 * case) the page is on.  If page_trylock() was used
780 		 * above, then we have to do the reclaim ourselves.
781 		 */
782 		if ((!hash_locked) && (PP_ISFREE(pp))) {
783 			ASSERT(PP_ISAGED(pp) == 0);
784 			VM_STAT_ADD(page_lookup_cnt[8]);
785 
786 			/*
787 			 * page_relcaim will insure that we
788 			 * have this page exclusively
789 			 */
790 
791 			if (!page_reclaim(pp, NULL)) {
792 				/*
793 				 * Page_reclaim dropped whatever lock
794 				 * we held.
795 				 */
796 				VM_STAT_ADD(page_lookup_cnt[9]);
797 				phm = PAGE_HASH_MUTEX(index);
798 				mutex_enter(phm);
799 				hash_locked = 1;
800 				goto top;
801 			} else if (se == SE_SHARED && newpp == NULL) {
802 				VM_STAT_ADD(page_lookup_cnt[10]);
803 				page_downgrade(pp);
804 			}
805 		}
806 
807 		if (hash_locked) {
808 			mutex_exit(phm);
809 		}
810 
811 		if (newpp != NULL && pp->p_szc < newpp->p_szc &&
812 		    PAGE_EXCL(pp) && nrelocp != NULL) {
813 			ASSERT(nrelocp != NULL);
814 			(void) page_relocate(&pp, &newpp, 1, 1, nrelocp,
815 			    NULL);
816 			if (*nrelocp > 0) {
817 				VM_STAT_COND_ADD(*nrelocp == 1,
818 				    page_lookup_cnt[11]);
819 				VM_STAT_COND_ADD(*nrelocp > 1,
820 				    page_lookup_cnt[12]);
821 				pp = newpp;
822 				se = SE_EXCL;
823 			} else {
824 				if (se == SE_SHARED) {
825 					page_downgrade(pp);
826 				}
827 				VM_STAT_ADD(page_lookup_cnt[13]);
828 			}
829 		} else if (newpp != NULL && nrelocp != NULL) {
830 			if (PAGE_EXCL(pp) && se == SE_SHARED) {
831 				page_downgrade(pp);
832 			}
833 			VM_STAT_COND_ADD(pp->p_szc < newpp->p_szc,
834 			    page_lookup_cnt[14]);
835 			VM_STAT_COND_ADD(pp->p_szc == newpp->p_szc,
836 			    page_lookup_cnt[15]);
837 			VM_STAT_COND_ADD(pp->p_szc > newpp->p_szc,
838 			    page_lookup_cnt[16]);
839 		} else if (newpp != NULL && PAGE_EXCL(pp)) {
840 			se = SE_EXCL;
841 		}
842 	} else if (!hash_locked) {
843 		VM_STAT_ADD(page_lookup_cnt[17]);
844 		phm = PAGE_HASH_MUTEX(index);
845 		mutex_enter(phm);
846 		hash_locked = 1;
847 		goto top;
848 	} else if (newpp != NULL) {
849 		/*
850 		 * If we have a preallocated page then
851 		 * insert it now and basically behave like
852 		 * page_create.
853 		 */
854 		VM_STAT_ADD(page_lookup_cnt[18]);
855 		/*
856 		 * Since we hold the page hash mutex and
857 		 * just searched for this page, page_hashin
858 		 * had better not fail.  If it does, that
859 		 * means some thread did not follow the
860 		 * page hash mutex rules.  Panic now and
861 		 * get it over with.  As usual, go down
862 		 * holding all the locks.
863 		 */
864 		ASSERT(MUTEX_HELD(phm));
865 		if (!page_hashin(newpp, vp, off, phm)) {
866 			ASSERT(MUTEX_HELD(phm));
867 			panic("page_lookup_create: hashin failed %p %p %llx %p",
868 			    (void *)newpp, (void *)vp, off, (void *)phm);
869 			/*NOTREACHED*/
870 		}
871 		ASSERT(MUTEX_HELD(phm));
872 		mutex_exit(phm);
873 		phm = NULL;
874 		page_set_props(newpp, P_REF);
875 		page_io_lock(newpp);
876 		pp = newpp;
877 		se = SE_EXCL;
878 	} else {
879 		VM_STAT_ADD(page_lookup_cnt[19]);
880 		mutex_exit(phm);
881 	}
882 
883 	ASSERT(pp ? PAGE_LOCKED_SE(pp, se) : 1);
884 
885 	ASSERT(pp ? ((PP_ISFREE(pp) == 0) && (PP_ISAGED(pp) == 0)) : 1);
886 
887 	return (pp);
888 }
889 
890 /*
891  * Search the hash list for the page representing the
892  * specified [vp, offset] and return it locked.  Skip
893  * free pages and pages that cannot be locked as requested.
894  * Used while attempting to kluster pages.
895  */
896 page_t *
897 page_lookup_nowait(vnode_t *vp, u_offset_t off, se_t se)
898 {
899 	page_t		*pp;
900 	kmutex_t	*phm;
901 	ulong_t		index;
902 	uint_t		locked;
903 
904 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
905 	VM_STAT_ADD(page_lookup_nowait_cnt[0]);
906 
907 	index = PAGE_HASH_FUNC(vp, off);
908 	PAGE_HASH_SEARCH(index, pp, vp, off);
909 	locked = 0;
910 	if (pp == NULL) {
911 top:
912 		VM_STAT_ADD(page_lookup_nowait_cnt[1]);
913 		locked = 1;
914 		phm = PAGE_HASH_MUTEX(index);
915 		mutex_enter(phm);
916 		PAGE_HASH_SEARCH(index, pp, vp, off);
917 	}
918 
919 	if (pp == NULL || PP_ISFREE(pp)) {
920 		VM_STAT_ADD(page_lookup_nowait_cnt[2]);
921 		pp = NULL;
922 	} else {
923 		if (!page_trylock(pp, se)) {
924 			VM_STAT_ADD(page_lookup_nowait_cnt[3]);
925 			pp = NULL;
926 		} else {
927 			VM_STAT_ADD(page_lookup_nowait_cnt[4]);
928 			/*
929 			 * See the comment in page_lookup()
930 			 */
931 			if (((volatile struct vnode *)(pp->p_vnode) != vp) ||
932 			    ((u_offset_t)(pp->p_offset) != off)) {
933 				VM_STAT_ADD(page_lookup_nowait_cnt[5]);
934 				if (locked) {
935 					panic("page_lookup_nowait %p",
936 					    (void *)pp);
937 					/*NOTREACHED*/
938 				}
939 				page_unlock(pp);
940 				goto top;
941 			}
942 			if (PP_ISFREE(pp)) {
943 				VM_STAT_ADD(page_lookup_nowait_cnt[6]);
944 				page_unlock(pp);
945 				pp = NULL;
946 			}
947 		}
948 	}
949 	if (locked) {
950 		VM_STAT_ADD(page_lookup_nowait_cnt[7]);
951 		mutex_exit(phm);
952 	}
953 
954 	ASSERT(pp ? PAGE_LOCKED_SE(pp, se) : 1);
955 
956 	return (pp);
957 }
958 
959 /*
960  * Search the hash list for a page with the specified [vp, off]
961  * that is known to exist and is already locked.  This routine
962  * is typically used by segment SOFTUNLOCK routines.
963  */
964 page_t *
965 page_find(vnode_t *vp, u_offset_t off)
966 {
967 	page_t		*pp;
968 	kmutex_t	*phm;
969 	ulong_t		index;
970 
971 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
972 	VM_STAT_ADD(page_find_cnt);
973 
974 	index = PAGE_HASH_FUNC(vp, off);
975 	phm = PAGE_HASH_MUTEX(index);
976 
977 	mutex_enter(phm);
978 	PAGE_HASH_SEARCH(index, pp, vp, off);
979 	mutex_exit(phm);
980 
981 	ASSERT(pp == NULL || PAGE_LOCKED(pp) || panicstr);
982 	return (pp);
983 }
984 
985 /*
986  * Determine whether a page with the specified [vp, off]
987  * currently exists in the system.  Obviously this should
988  * only be considered as a hint since nothing prevents the
989  * page from disappearing or appearing immediately after
990  * the return from this routine. Subsequently, we don't
991  * even bother to lock the list.
992  */
993 page_t *
994 page_exists(vnode_t *vp, u_offset_t off)
995 {
996 	page_t	*pp;
997 	ulong_t		index;
998 
999 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
1000 	VM_STAT_ADD(page_exists_cnt);
1001 
1002 	index = PAGE_HASH_FUNC(vp, off);
1003 	PAGE_HASH_SEARCH(index, pp, vp, off);
1004 
1005 	return (pp);
1006 }
1007 
1008 /*
1009  * Determine if physically contiguous pages exist for [vp, off] - [vp, off +
1010  * page_size(szc)) range.  if they exist and ppa is not NULL fill ppa array
1011  * with these pages locked SHARED. If necessary reclaim pages from
1012  * freelist. Return 1 if contiguous pages exist and 0 otherwise.
1013  *
1014  * If we fail to lock pages still return 1 if pages exist and contiguous.
1015  * But in this case return value is just a hint. ppa array won't be filled.
1016  * Caller should initialize ppa[0] as NULL to distinguish return value.
1017  *
1018  * Returns 0 if pages don't exist or not physically contiguous.
1019  *
1020  * This routine doesn't work for anonymous(swapfs) pages.
1021  */
1022 int
1023 page_exists_physcontig(vnode_t *vp, u_offset_t off, uint_t szc, page_t *ppa[])
1024 {
1025 	pgcnt_t pages;
1026 	pfn_t pfn;
1027 	page_t *rootpp;
1028 	pgcnt_t i;
1029 	pgcnt_t j;
1030 	u_offset_t save_off = off;
1031 	ulong_t index;
1032 	kmutex_t *phm;
1033 	page_t *pp;
1034 	uint_t pszc;
1035 	int loopcnt = 0;
1036 
1037 	ASSERT(szc != 0);
1038 	ASSERT(vp != NULL);
1039 	ASSERT(!IS_SWAPFSVP(vp));
1040 	ASSERT(!VN_ISKAS(vp));
1041 
1042 again:
1043 	if (++loopcnt > 3) {
1044 		VM_STAT_ADD(page_exphcontg[0]);
1045 		return (0);
1046 	}
1047 
1048 	index = PAGE_HASH_FUNC(vp, off);
1049 	phm = PAGE_HASH_MUTEX(index);
1050 
1051 	mutex_enter(phm);
1052 	PAGE_HASH_SEARCH(index, pp, vp, off);
1053 	mutex_exit(phm);
1054 
1055 	VM_STAT_ADD(page_exphcontg[1]);
1056 
1057 	if (pp == NULL) {
1058 		VM_STAT_ADD(page_exphcontg[2]);
1059 		return (0);
1060 	}
1061 
1062 	pages = page_get_pagecnt(szc);
1063 	rootpp = pp;
1064 	pfn = rootpp->p_pagenum;
1065 
1066 	if ((pszc = pp->p_szc) >= szc && ppa != NULL) {
1067 		VM_STAT_ADD(page_exphcontg[3]);
1068 		if (!page_trylock(pp, SE_SHARED)) {
1069 			VM_STAT_ADD(page_exphcontg[4]);
1070 			return (1);
1071 		}
1072 		if (pp->p_szc != pszc || pp->p_vnode != vp ||
1073 		    pp->p_offset != off) {
1074 			VM_STAT_ADD(page_exphcontg[5]);
1075 			page_unlock(pp);
1076 			off = save_off;
1077 			goto again;
1078 		}
1079 		/*
1080 		 * szc was non zero and vnode and offset matched after we
1081 		 * locked the page it means it can't become free on us.
1082 		 */
1083 		ASSERT(!PP_ISFREE(pp));
1084 		if (!IS_P2ALIGNED(pfn, pages)) {
1085 			page_unlock(pp);
1086 			return (0);
1087 		}
1088 		ppa[0] = pp;
1089 		pp++;
1090 		off += PAGESIZE;
1091 		pfn++;
1092 		for (i = 1; i < pages; i++, pp++, off += PAGESIZE, pfn++) {
1093 			if (!page_trylock(pp, SE_SHARED)) {
1094 				VM_STAT_ADD(page_exphcontg[6]);
1095 				pp--;
1096 				while (i-- > 0) {
1097 					page_unlock(pp);
1098 					pp--;
1099 				}
1100 				ppa[0] = NULL;
1101 				return (1);
1102 			}
1103 			if (pp->p_szc != pszc) {
1104 				VM_STAT_ADD(page_exphcontg[7]);
1105 				page_unlock(pp);
1106 				pp--;
1107 				while (i-- > 0) {
1108 					page_unlock(pp);
1109 					pp--;
1110 				}
1111 				ppa[0] = NULL;
1112 				off = save_off;
1113 				goto again;
1114 			}
1115 			/*
1116 			 * szc the same as for previous already locked pages
1117 			 * with right identity. Since this page had correct
1118 			 * szc after we locked it can't get freed or destroyed
1119 			 * and therefore must have the expected identity.
1120 			 */
1121 			ASSERT(!PP_ISFREE(pp));
1122 			if (pp->p_vnode != vp ||
1123 			    pp->p_offset != off) {
1124 				panic("page_exists_physcontig: "
1125 				    "large page identity doesn't match");
1126 			}
1127 			ppa[i] = pp;
1128 			ASSERT(pp->p_pagenum == pfn);
1129 		}
1130 		VM_STAT_ADD(page_exphcontg[8]);
1131 		ppa[pages] = NULL;
1132 		return (1);
1133 	} else if (pszc >= szc) {
1134 		VM_STAT_ADD(page_exphcontg[9]);
1135 		if (!IS_P2ALIGNED(pfn, pages)) {
1136 			return (0);
1137 		}
1138 		return (1);
1139 	}
1140 
1141 	if (!IS_P2ALIGNED(pfn, pages)) {
1142 		VM_STAT_ADD(page_exphcontg[10]);
1143 		return (0);
1144 	}
1145 
1146 	if (page_numtomemseg_nolock(pfn) !=
1147 	    page_numtomemseg_nolock(pfn + pages - 1)) {
1148 		VM_STAT_ADD(page_exphcontg[11]);
1149 		return (0);
1150 	}
1151 
1152 	/*
1153 	 * We loop up 4 times across pages to promote page size.
1154 	 * We're extra cautious to promote page size atomically with respect
1155 	 * to everybody else.  But we can probably optimize into 1 loop if
1156 	 * this becomes an issue.
1157 	 */
1158 
1159 	for (i = 0; i < pages; i++, pp++, off += PAGESIZE, pfn++) {
1160 		ASSERT(pp->p_pagenum == pfn);
1161 		if (!page_trylock(pp, SE_EXCL)) {
1162 			VM_STAT_ADD(page_exphcontg[12]);
1163 			break;
1164 		}
1165 		if (pp->p_vnode != vp ||
1166 		    pp->p_offset != off) {
1167 			VM_STAT_ADD(page_exphcontg[13]);
1168 			page_unlock(pp);
1169 			break;
1170 		}
1171 		if (pp->p_szc >= szc) {
1172 			ASSERT(i == 0);
1173 			page_unlock(pp);
1174 			off = save_off;
1175 			goto again;
1176 		}
1177 	}
1178 
1179 	if (i != pages) {
1180 		VM_STAT_ADD(page_exphcontg[14]);
1181 		--pp;
1182 		while (i-- > 0) {
1183 			page_unlock(pp);
1184 			--pp;
1185 		}
1186 		return (0);
1187 	}
1188 
1189 	pp = rootpp;
1190 	for (i = 0; i < pages; i++, pp++) {
1191 		if (PP_ISFREE(pp)) {
1192 			VM_STAT_ADD(page_exphcontg[15]);
1193 			ASSERT(!PP_ISAGED(pp));
1194 			ASSERT(pp->p_szc == 0);
1195 			if (!page_reclaim(pp, NULL)) {
1196 				break;
1197 			}
1198 		} else {
1199 			ASSERT(pp->p_szc < szc);
1200 			VM_STAT_ADD(page_exphcontg[16]);
1201 			(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
1202 		}
1203 	}
1204 	if (i < pages) {
1205 		VM_STAT_ADD(page_exphcontg[17]);
1206 		/*
1207 		 * page_reclaim failed because we were out of memory.
1208 		 * drop the rest of the locks and return because this page
1209 		 * must be already reallocated anyway.
1210 		 */
1211 		pp = rootpp;
1212 		for (j = 0; j < pages; j++, pp++) {
1213 			if (j != i) {
1214 				page_unlock(pp);
1215 			}
1216 		}
1217 		return (0);
1218 	}
1219 
1220 	off = save_off;
1221 	pp = rootpp;
1222 	for (i = 0; i < pages; i++, pp++, off += PAGESIZE) {
1223 		ASSERT(PAGE_EXCL(pp));
1224 		ASSERT(!PP_ISFREE(pp));
1225 		ASSERT(!hat_page_is_mapped(pp));
1226 		ASSERT(pp->p_vnode == vp);
1227 		ASSERT(pp->p_offset == off);
1228 		pp->p_szc = szc;
1229 	}
1230 	pp = rootpp;
1231 	for (i = 0; i < pages; i++, pp++) {
1232 		if (ppa == NULL) {
1233 			page_unlock(pp);
1234 		} else {
1235 			ppa[i] = pp;
1236 			page_downgrade(ppa[i]);
1237 		}
1238 	}
1239 	if (ppa != NULL) {
1240 		ppa[pages] = NULL;
1241 	}
1242 	VM_STAT_ADD(page_exphcontg[18]);
1243 	ASSERT(vp->v_pages != NULL);
1244 	return (1);
1245 }
1246 
1247 /*
1248  * Determine whether a page with the specified [vp, off]
1249  * currently exists in the system and if so return its
1250  * size code. Obviously this should only be considered as
1251  * a hint since nothing prevents the page from disappearing
1252  * or appearing immediately after the return from this routine.
1253  */
1254 int
1255 page_exists_forreal(vnode_t *vp, u_offset_t off, uint_t *szc)
1256 {
1257 	page_t		*pp;
1258 	kmutex_t	*phm;
1259 	ulong_t		index;
1260 	int		rc = 0;
1261 
1262 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
1263 	ASSERT(szc != NULL);
1264 	VM_STAT_ADD(page_exists_forreal_cnt);
1265 
1266 	index = PAGE_HASH_FUNC(vp, off);
1267 	phm = PAGE_HASH_MUTEX(index);
1268 
1269 	mutex_enter(phm);
1270 	PAGE_HASH_SEARCH(index, pp, vp, off);
1271 	if (pp != NULL) {
1272 		*szc = pp->p_szc;
1273 		rc = 1;
1274 	}
1275 	mutex_exit(phm);
1276 	return (rc);
1277 }
1278 
1279 /* wakeup threads waiting for pages in page_create_get_something() */
1280 void
1281 wakeup_pcgs(void)
1282 {
1283 	if (!CV_HAS_WAITERS(&pcgs_cv))
1284 		return;
1285 	cv_broadcast(&pcgs_cv);
1286 }
1287 
1288 /*
1289  * 'freemem' is used all over the kernel as an indication of how many
1290  * pages are free (either on the cache list or on the free page list)
1291  * in the system.  In very few places is a really accurate 'freemem'
1292  * needed.  To avoid contention of the lock protecting a the
1293  * single freemem, it was spread out into NCPU buckets.  Set_freemem
1294  * sets freemem to the total of all NCPU buckets.  It is called from
1295  * clock() on each TICK.
1296  */
1297 void
1298 set_freemem()
1299 {
1300 	struct pcf	*p;
1301 	ulong_t		t;
1302 	uint_t		i;
1303 
1304 	t = 0;
1305 	p = pcf;
1306 	for (i = 0;  i < PCF_FANOUT; i++) {
1307 		t += p->pcf_count;
1308 		p++;
1309 	}
1310 	freemem = t;
1311 
1312 	/*
1313 	 * Don't worry about grabbing mutex.  It's not that
1314 	 * critical if we miss a tick or two.  This is
1315 	 * where we wakeup possible delayers in
1316 	 * page_create_get_something().
1317 	 */
1318 	wakeup_pcgs();
1319 }
1320 
1321 ulong_t
1322 get_freemem()
1323 {
1324 	struct pcf	*p;
1325 	ulong_t		t;
1326 	uint_t		i;
1327 
1328 	t = 0;
1329 	p = pcf;
1330 	for (i = 0; i < PCF_FANOUT; i++) {
1331 		t += p->pcf_count;
1332 		p++;
1333 	}
1334 	/*
1335 	 * We just calculated it, might as well set it.
1336 	 */
1337 	freemem = t;
1338 	return (t);
1339 }
1340 
1341 /*
1342  * Acquire all of the page cache & free (pcf) locks.
1343  */
1344 void
1345 pcf_acquire_all()
1346 {
1347 	struct pcf	*p;
1348 	uint_t		i;
1349 
1350 	p = pcf;
1351 	for (i = 0; i < PCF_FANOUT; i++) {
1352 		mutex_enter(&p->pcf_lock);
1353 		p++;
1354 	}
1355 }
1356 
1357 /*
1358  * Release all the pcf_locks.
1359  */
1360 void
1361 pcf_release_all()
1362 {
1363 	struct pcf	*p;
1364 	uint_t		i;
1365 
1366 	p = pcf;
1367 	for (i = 0; i < PCF_FANOUT; i++) {
1368 		mutex_exit(&p->pcf_lock);
1369 		p++;
1370 	}
1371 }
1372 
1373 /*
1374  * Inform the VM system that we need some pages freed up.
1375  * Calls must be symmetric, e.g.:
1376  *
1377  *	page_needfree(100);
1378  *	wait a bit;
1379  *	page_needfree(-100);
1380  */
1381 void
1382 page_needfree(spgcnt_t npages)
1383 {
1384 	mutex_enter(&new_freemem_lock);
1385 	needfree += npages;
1386 	mutex_exit(&new_freemem_lock);
1387 }
1388 
1389 /*
1390  * Throttle for page_create(): try to prevent freemem from dropping
1391  * below throttlefree.  We can't provide a 100% guarantee because
1392  * KM_NOSLEEP allocations, page_reclaim(), and various other things
1393  * nibble away at the freelist.  However, we can block all PG_WAIT
1394  * allocations until memory becomes available.  The motivation is
1395  * that several things can fall apart when there's no free memory:
1396  *
1397  * (1) If pageout() needs memory to push a page, the system deadlocks.
1398  *
1399  * (2) By (broken) specification, timeout(9F) can neither fail nor
1400  *     block, so it has no choice but to panic the system if it
1401  *     cannot allocate a callout structure.
1402  *
1403  * (3) Like timeout(), ddi_set_callback() cannot fail and cannot block;
1404  *     it panics if it cannot allocate a callback structure.
1405  *
1406  * (4) Untold numbers of third-party drivers have not yet been hardened
1407  *     against KM_NOSLEEP and/or allocb() failures; they simply assume
1408  *     success and panic the system with a data fault on failure.
1409  *     (The long-term solution to this particular problem is to ship
1410  *     hostile fault-injecting DEBUG kernels with the DDK.)
1411  *
1412  * It is theoretically impossible to guarantee success of non-blocking
1413  * allocations, but in practice, this throttle is very hard to break.
1414  */
1415 static int
1416 page_create_throttle(pgcnt_t npages, int flags)
1417 {
1418 	ulong_t	fm;
1419 	uint_t	i;
1420 	pgcnt_t tf;	/* effective value of throttlefree */
1421 
1422 	/*
1423 	 * Never deny pages when:
1424 	 * - it's a thread that cannot block [NOMEMWAIT()]
1425 	 * - the allocation cannot block and must not fail
1426 	 * - the allocation cannot block and is pageout dispensated
1427 	 */
1428 	if (NOMEMWAIT() ||
1429 	    ((flags & (PG_WAIT | PG_PANIC)) == PG_PANIC) ||
1430 	    ((flags & (PG_WAIT | PG_PUSHPAGE)) == PG_PUSHPAGE))
1431 		return (1);
1432 
1433 	/*
1434 	 * If the allocation can't block, we look favorably upon it
1435 	 * unless we're below pageout_reserve.  In that case we fail
1436 	 * the allocation because we want to make sure there are a few
1437 	 * pages available for pageout.
1438 	 */
1439 	if ((flags & PG_WAIT) == 0)
1440 		return (freemem >= npages + pageout_reserve);
1441 
1442 	/* Calculate the effective throttlefree value */
1443 	tf = throttlefree -
1444 	    ((flags & PG_PUSHPAGE) ? pageout_reserve : 0);
1445 
1446 	cv_signal(&proc_pageout->p_cv);
1447 
1448 	while (freemem < npages + tf) {
1449 		pcf_acquire_all();
1450 		mutex_enter(&new_freemem_lock);
1451 		fm = 0;
1452 		for (i = 0; i < PCF_FANOUT; i++) {
1453 			fm += pcf[i].pcf_count;
1454 			pcf[i].pcf_wait++;
1455 			mutex_exit(&pcf[i].pcf_lock);
1456 		}
1457 		freemem = fm;
1458 		needfree += npages;
1459 		freemem_wait++;
1460 		cv_wait(&freemem_cv, &new_freemem_lock);
1461 		freemem_wait--;
1462 		needfree -= npages;
1463 		mutex_exit(&new_freemem_lock);
1464 	}
1465 	return (1);
1466 }
1467 
1468 /*
1469  * page_create_wait() is called to either coalesce pages from the
1470  * different pcf buckets or to wait because there simply are not
1471  * enough pages to satisfy the caller's request.
1472  *
1473  * Sadly, this is called from platform/vm/vm_machdep.c
1474  */
1475 int
1476 page_create_wait(size_t npages, uint_t flags)
1477 {
1478 	pgcnt_t		total;
1479 	uint_t		i;
1480 	struct pcf	*p;
1481 
1482 	/*
1483 	 * Wait until there are enough free pages to satisfy our
1484 	 * entire request.
1485 	 * We set needfree += npages before prodding pageout, to make sure
1486 	 * it does real work when npages > lotsfree > freemem.
1487 	 */
1488 	VM_STAT_ADD(page_create_not_enough);
1489 
1490 	ASSERT(!kcage_on ? !(flags & PG_NORELOC) : 1);
1491 checkagain:
1492 	if ((flags & PG_NORELOC) &&
1493 	    kcage_freemem < kcage_throttlefree + npages)
1494 		(void) kcage_create_throttle(npages, flags);
1495 
1496 	if (freemem < npages + throttlefree)
1497 		if (!page_create_throttle(npages, flags))
1498 			return (0);
1499 
1500 	/*
1501 	 * Since page_create_va() looked at every
1502 	 * bucket, assume we are going to have to wait.
1503 	 * Get all of the pcf locks.
1504 	 */
1505 	total = 0;
1506 	p = pcf;
1507 	for (i = 0; i < PCF_FANOUT; i++) {
1508 		mutex_enter(&p->pcf_lock);
1509 		total += p->pcf_count;
1510 		if (total >= npages) {
1511 			/*
1512 			 * Wow!  There are enough pages laying around
1513 			 * to satisfy the request.  Do the accounting,
1514 			 * drop the locks we acquired, and go back.
1515 			 *
1516 			 * freemem is not protected by any lock. So,
1517 			 * we cannot have any assertion containing
1518 			 * freemem.
1519 			 */
1520 			freemem -= npages;
1521 
1522 			while (p >= pcf) {
1523 				if (p->pcf_count <= npages) {
1524 					npages -= p->pcf_count;
1525 					p->pcf_count = 0;
1526 				} else {
1527 					p->pcf_count -= (uint_t)npages;
1528 					npages = 0;
1529 				}
1530 				mutex_exit(&p->pcf_lock);
1531 				p--;
1532 			}
1533 			ASSERT(npages == 0);
1534 			return (1);
1535 		}
1536 		p++;
1537 	}
1538 
1539 	/*
1540 	 * All of the pcf locks are held, there are not enough pages
1541 	 * to satisfy the request (npages < total).
1542 	 * Be sure to acquire the new_freemem_lock before dropping
1543 	 * the pcf locks.  This prevents dropping wakeups in page_free().
1544 	 * The order is always pcf_lock then new_freemem_lock.
1545 	 *
1546 	 * Since we hold all the pcf locks, it is a good time to set freemem.
1547 	 *
1548 	 * If the caller does not want to wait, return now.
1549 	 * Else turn the pageout daemon loose to find something
1550 	 * and wait till it does.
1551 	 *
1552 	 */
1553 	freemem = total;
1554 
1555 	if ((flags & PG_WAIT) == 0) {
1556 		pcf_release_all();
1557 
1558 		TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_NOMEM,
1559 		"page_create_nomem:npages %ld freemem %ld", npages, freemem);
1560 		return (0);
1561 	}
1562 
1563 	ASSERT(proc_pageout != NULL);
1564 	cv_signal(&proc_pageout->p_cv);
1565 
1566 	TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SLEEP_START,
1567 	    "page_create_sleep_start: freemem %ld needfree %ld",
1568 	    freemem, needfree);
1569 
1570 	/*
1571 	 * We are going to wait.
1572 	 * We currently hold all of the pcf_locks,
1573 	 * get the new_freemem_lock (it protects freemem_wait),
1574 	 * before dropping the pcf_locks.
1575 	 */
1576 	mutex_enter(&new_freemem_lock);
1577 
1578 	p = pcf;
1579 	for (i = 0; i < PCF_FANOUT; i++) {
1580 		p->pcf_wait++;
1581 		mutex_exit(&p->pcf_lock);
1582 		p++;
1583 	}
1584 
1585 	needfree += npages;
1586 	freemem_wait++;
1587 
1588 	cv_wait(&freemem_cv, &new_freemem_lock);
1589 
1590 	freemem_wait--;
1591 	needfree -= npages;
1592 
1593 	mutex_exit(&new_freemem_lock);
1594 
1595 	TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SLEEP_END,
1596 	    "page_create_sleep_end: freemem %ld needfree %ld",
1597 	    freemem, needfree);
1598 
1599 	VM_STAT_ADD(page_create_not_enough_again);
1600 	goto checkagain;
1601 }
1602 
1603 /*
1604  * A routine to do the opposite of page_create_wait().
1605  */
1606 void
1607 page_create_putback(spgcnt_t npages)
1608 {
1609 	struct pcf	*p;
1610 	pgcnt_t		lump;
1611 	uint_t		*which;
1612 
1613 	/*
1614 	 * When a contiguous lump is broken up, we have to
1615 	 * deal with lots of pages (min 64) so lets spread
1616 	 * the wealth around.
1617 	 */
1618 	lump = roundup(npages, PCF_FANOUT) / PCF_FANOUT;
1619 	freemem += npages;
1620 
1621 	for (p = pcf; (npages > 0) && (p < &pcf[PCF_FANOUT]); p++) {
1622 		which = &p->pcf_count;
1623 
1624 		mutex_enter(&p->pcf_lock);
1625 
1626 		if (p->pcf_block) {
1627 			which = &p->pcf_reserve;
1628 		}
1629 
1630 		if (lump < npages) {
1631 			*which += (uint_t)lump;
1632 			npages -= lump;
1633 		} else {
1634 			*which += (uint_t)npages;
1635 			npages = 0;
1636 		}
1637 
1638 		if (p->pcf_wait) {
1639 			mutex_enter(&new_freemem_lock);
1640 			/*
1641 			 * Check to see if some other thread
1642 			 * is actually waiting.  Another bucket
1643 			 * may have woken it up by now.  If there
1644 			 * are no waiters, then set our pcf_wait
1645 			 * count to zero to avoid coming in here
1646 			 * next time.
1647 			 */
1648 			if (freemem_wait) {
1649 				if (npages > 1) {
1650 					cv_broadcast(&freemem_cv);
1651 				} else {
1652 					cv_signal(&freemem_cv);
1653 				}
1654 				p->pcf_wait--;
1655 			} else {
1656 				p->pcf_wait = 0;
1657 			}
1658 			mutex_exit(&new_freemem_lock);
1659 		}
1660 		mutex_exit(&p->pcf_lock);
1661 	}
1662 	ASSERT(npages == 0);
1663 }
1664 
1665 /*
1666  * A helper routine for page_create_get_something.
1667  * The indenting got to deep down there.
1668  * Unblock the pcf counters.  Any pages freed after
1669  * pcf_block got set are moved to pcf_count and
1670  * wakeups (cv_broadcast() or cv_signal()) are done as needed.
1671  */
1672 static void
1673 pcgs_unblock(void)
1674 {
1675 	int		i;
1676 	struct pcf	*p;
1677 
1678 	/* Update freemem while we're here. */
1679 	freemem = 0;
1680 	p = pcf;
1681 	for (i = 0; i < PCF_FANOUT; i++) {
1682 		mutex_enter(&p->pcf_lock);
1683 		ASSERT(p->pcf_count == 0);
1684 		p->pcf_count = p->pcf_reserve;
1685 		p->pcf_block = 0;
1686 		freemem += p->pcf_count;
1687 		if (p->pcf_wait) {
1688 			mutex_enter(&new_freemem_lock);
1689 			if (freemem_wait) {
1690 				if (p->pcf_reserve > 1) {
1691 					cv_broadcast(&freemem_cv);
1692 					p->pcf_wait = 0;
1693 				} else {
1694 					cv_signal(&freemem_cv);
1695 					p->pcf_wait--;
1696 				}
1697 			} else {
1698 				p->pcf_wait = 0;
1699 			}
1700 			mutex_exit(&new_freemem_lock);
1701 		}
1702 		p->pcf_reserve = 0;
1703 		mutex_exit(&p->pcf_lock);
1704 		p++;
1705 	}
1706 }
1707 
1708 /*
1709  * Called from page_create_va() when both the cache and free lists
1710  * have been checked once.
1711  *
1712  * Either returns a page or panics since the accounting was done
1713  * way before we got here.
1714  *
1715  * We don't come here often, so leave the accounting on permanently.
1716  */
1717 
1718 #define	MAX_PCGS	100
1719 
1720 #ifdef	DEBUG
1721 #define	PCGS_TRIES	100
1722 #else	/* DEBUG */
1723 #define	PCGS_TRIES	10
1724 #endif	/* DEBUG */
1725 
1726 #ifdef	VM_STATS
1727 uint_t	pcgs_counts[PCGS_TRIES];
1728 uint_t	pcgs_too_many;
1729 uint_t	pcgs_entered;
1730 uint_t	pcgs_entered_noreloc;
1731 uint_t	pcgs_locked;
1732 uint_t	pcgs_cagelocked;
1733 #endif	/* VM_STATS */
1734 
1735 static page_t *
1736 page_create_get_something(vnode_t *vp, u_offset_t off, struct seg *seg,
1737     caddr_t vaddr, uint_t flags)
1738 {
1739 	uint_t		count;
1740 	page_t		*pp;
1741 	uint_t		locked, i;
1742 	struct	pcf	*p;
1743 	lgrp_t		*lgrp;
1744 	int		cagelocked = 0;
1745 
1746 	VM_STAT_ADD(pcgs_entered);
1747 
1748 	/*
1749 	 * Tap any reserve freelists: if we fail now, we'll die
1750 	 * since the page(s) we're looking for have already been
1751 	 * accounted for.
1752 	 */
1753 	flags |= PG_PANIC;
1754 
1755 	if ((flags & PG_NORELOC) != 0) {
1756 		VM_STAT_ADD(pcgs_entered_noreloc);
1757 		/*
1758 		 * Requests for free pages from critical threads
1759 		 * such as pageout still won't throttle here, but
1760 		 * we must try again, to give the cageout thread
1761 		 * another chance to catch up. Since we already
1762 		 * accounted for the pages, we had better get them
1763 		 * this time.
1764 		 *
1765 		 * N.B. All non-critical threads acquire the pcgs_cagelock
1766 		 * to serialize access to the freelists. This implements a
1767 		 * turnstile-type synchornization to avoid starvation of
1768 		 * critical requests for PG_NORELOC memory by non-critical
1769 		 * threads: all non-critical threads must acquire a 'ticket'
1770 		 * before passing through, which entails making sure
1771 		 * kcage_freemem won't fall below minfree prior to grabbing
1772 		 * pages from the freelists.
1773 		 */
1774 		if (kcage_create_throttle(1, flags) == KCT_NONCRIT) {
1775 			mutex_enter(&pcgs_cagelock);
1776 			cagelocked = 1;
1777 			VM_STAT_ADD(pcgs_cagelocked);
1778 		}
1779 	}
1780 
1781 	/*
1782 	 * Time to get serious.
1783 	 * We failed to get a `correctly colored' page from both the
1784 	 * free and cache lists.
1785 	 * We escalate in stage.
1786 	 *
1787 	 * First try both lists without worring about color.
1788 	 *
1789 	 * Then, grab all page accounting locks (ie. pcf[]) and
1790 	 * steal any pages that they have and set the pcf_block flag to
1791 	 * stop deletions from the lists.  This will help because
1792 	 * a page can get added to the free list while we are looking
1793 	 * at the cache list, then another page could be added to the cache
1794 	 * list allowing the page on the free list to be removed as we
1795 	 * move from looking at the cache list to the free list. This
1796 	 * could happen over and over. We would never find the page
1797 	 * we have accounted for.
1798 	 *
1799 	 * Noreloc pages are a subset of the global (relocatable) page pool.
1800 	 * They are not tracked separately in the pcf bins, so it is
1801 	 * impossible to know when doing pcf accounting if the available
1802 	 * page(s) are noreloc pages or not. When looking for a noreloc page
1803 	 * it is quite easy to end up here even if the global (relocatable)
1804 	 * page pool has plenty of free pages but the noreloc pool is empty.
1805 	 *
1806 	 * When the noreloc pool is empty (or low), additional noreloc pages
1807 	 * are created by converting pages from the global page pool. This
1808 	 * process will stall during pcf accounting if the pcf bins are
1809 	 * already locked. Such is the case when a noreloc allocation is
1810 	 * looping here in page_create_get_something waiting for more noreloc
1811 	 * pages to appear.
1812 	 *
1813 	 * Short of adding a new field to the pcf bins to accurately track
1814 	 * the number of free noreloc pages, we instead do not grab the
1815 	 * pcgs_lock, do not set the pcf blocks and do not timeout when
1816 	 * allocating a noreloc page. This allows noreloc allocations to
1817 	 * loop without blocking global page pool allocations.
1818 	 *
1819 	 * NOTE: the behaviour of page_create_get_something has not changed
1820 	 * for the case of global page pool allocations.
1821 	 */
1822 
1823 	flags &= ~PG_MATCH_COLOR;
1824 	locked = 0;
1825 #if defined(__i386) || defined(__amd64)
1826 	flags = page_create_update_flags_x86(flags);
1827 #endif
1828 
1829 	lgrp = lgrp_mem_choose(seg, vaddr, PAGESIZE);
1830 
1831 	for (count = 0; kcage_on || count < MAX_PCGS; count++) {
1832 		pp = page_get_freelist(vp, off, seg, vaddr, PAGESIZE,
1833 		    flags, lgrp);
1834 		if (pp == NULL) {
1835 			pp = page_get_cachelist(vp, off, seg, vaddr,
1836 			    flags, lgrp);
1837 		}
1838 		if (pp == NULL) {
1839 			/*
1840 			 * Serialize.  Don't fight with other pcgs().
1841 			 */
1842 			if (!locked && (!kcage_on || !(flags & PG_NORELOC))) {
1843 				mutex_enter(&pcgs_lock);
1844 				VM_STAT_ADD(pcgs_locked);
1845 				locked = 1;
1846 				p = pcf;
1847 				for (i = 0; i < PCF_FANOUT; i++) {
1848 					mutex_enter(&p->pcf_lock);
1849 					ASSERT(p->pcf_block == 0);
1850 					p->pcf_block = 1;
1851 					p->pcf_reserve = p->pcf_count;
1852 					p->pcf_count = 0;
1853 					mutex_exit(&p->pcf_lock);
1854 					p++;
1855 				}
1856 				freemem = 0;
1857 			}
1858 
1859 			if (count) {
1860 				/*
1861 				 * Since page_free() puts pages on
1862 				 * a list then accounts for it, we
1863 				 * just have to wait for page_free()
1864 				 * to unlock any page it was working
1865 				 * with. The page_lock()-page_reclaim()
1866 				 * path falls in the same boat.
1867 				 *
1868 				 * We don't need to check on the
1869 				 * PG_WAIT flag, we have already
1870 				 * accounted for the page we are
1871 				 * looking for in page_create_va().
1872 				 *
1873 				 * We just wait a moment to let any
1874 				 * locked pages on the lists free up,
1875 				 * then continue around and try again.
1876 				 *
1877 				 * Will be awakened by set_freemem().
1878 				 */
1879 				mutex_enter(&pcgs_wait_lock);
1880 				cv_wait(&pcgs_cv, &pcgs_wait_lock);
1881 				mutex_exit(&pcgs_wait_lock);
1882 			}
1883 		} else {
1884 #ifdef VM_STATS
1885 			if (count >= PCGS_TRIES) {
1886 				VM_STAT_ADD(pcgs_too_many);
1887 			} else {
1888 				VM_STAT_ADD(pcgs_counts[count]);
1889 			}
1890 #endif
1891 			if (locked) {
1892 				pcgs_unblock();
1893 				mutex_exit(&pcgs_lock);
1894 			}
1895 			if (cagelocked)
1896 				mutex_exit(&pcgs_cagelock);
1897 			return (pp);
1898 		}
1899 	}
1900 	/*
1901 	 * we go down holding the pcf locks.
1902 	 */
1903 	panic("no %spage found %d",
1904 	    ((flags & PG_NORELOC) ? "non-reloc " : ""), count);
1905 	/*NOTREACHED*/
1906 }
1907 
1908 /*
1909  * Create enough pages for "bytes" worth of data starting at
1910  * "off" in "vp".
1911  *
1912  *	Where flag must be one of:
1913  *
1914  *		PG_EXCL:	Exclusive create (fail if any page already
1915  *				exists in the page cache) which does not
1916  *				wait for memory to become available.
1917  *
1918  *		PG_WAIT:	Non-exclusive create which can wait for
1919  *				memory to become available.
1920  *
1921  *		PG_PHYSCONTIG:	Allocate physically contiguous pages.
1922  *				(Not Supported)
1923  *
1924  * A doubly linked list of pages is returned to the caller.  Each page
1925  * on the list has the "exclusive" (p_selock) lock and "iolock" (p_iolock)
1926  * lock.
1927  *
1928  * Unable to change the parameters to page_create() in a minor release,
1929  * we renamed page_create() to page_create_va(), changed all known calls
1930  * from page_create() to page_create_va(), and created this wrapper.
1931  *
1932  * Upon a major release, we should break compatibility by deleting this
1933  * wrapper, and replacing all the strings "page_create_va", with "page_create".
1934  *
1935  * NOTE: There is a copy of this interface as page_create_io() in
1936  *	 i86/vm/vm_machdep.c. Any bugs fixed here should be applied
1937  *	 there.
1938  */
1939 page_t *
1940 page_create(vnode_t *vp, u_offset_t off, size_t bytes, uint_t flags)
1941 {
1942 	caddr_t random_vaddr;
1943 	struct seg kseg;
1944 
1945 #ifdef DEBUG
1946 	cmn_err(CE_WARN, "Using deprecated interface page_create: caller %p",
1947 	    (void *)caller());
1948 #endif
1949 
1950 	random_vaddr = (caddr_t)(((uintptr_t)vp >> 7) ^
1951 	    (uintptr_t)(off >> PAGESHIFT));
1952 	kseg.s_as = &kas;
1953 
1954 	return (page_create_va(vp, off, bytes, flags, &kseg, random_vaddr));
1955 }
1956 
1957 #ifdef DEBUG
1958 uint32_t pg_alloc_pgs_mtbf = 0;
1959 #endif
1960 
1961 /*
1962  * Used for large page support. It will attempt to allocate
1963  * a large page(s) off the freelist.
1964  *
1965  * Returns non zero on failure.
1966  */
1967 int
1968 page_alloc_pages(struct vnode *vp, struct seg *seg, caddr_t addr,
1969     page_t **basepp, page_t *ppa[], uint_t szc, int anypgsz, int pgflags)
1970 {
1971 	pgcnt_t		npgs, curnpgs, totpgs;
1972 	size_t		pgsz;
1973 	page_t		*pplist = NULL, *pp;
1974 	int		err = 0;
1975 	lgrp_t		*lgrp;
1976 
1977 	ASSERT(szc != 0 && szc <= (page_num_pagesizes() - 1));
1978 	ASSERT(pgflags == 0 || pgflags == PG_LOCAL);
1979 
1980 	VM_STAT_ADD(alloc_pages[0]);
1981 
1982 #ifdef DEBUG
1983 	if (pg_alloc_pgs_mtbf && !(gethrtime() % pg_alloc_pgs_mtbf)) {
1984 		return (ENOMEM);
1985 	}
1986 #endif
1987 
1988 	pgsz = page_get_pagesize(szc);
1989 	totpgs = curnpgs = npgs = pgsz >> PAGESHIFT;
1990 
1991 	ASSERT(((uintptr_t)addr & (pgsz - 1)) == 0);
1992 	/*
1993 	 * One must be NULL but not both.
1994 	 * And one must be non NULL but not both.
1995 	 */
1996 	ASSERT(basepp != NULL || ppa != NULL);
1997 	ASSERT(basepp == NULL || ppa == NULL);
1998 
1999 	(void) page_create_wait(npgs, PG_WAIT);
2000 
2001 	while (npgs && szc) {
2002 		lgrp = lgrp_mem_choose(seg, addr, pgsz);
2003 		if (pgflags == PG_LOCAL) {
2004 			pp = page_get_freelist(vp, 0, seg, addr, pgsz,
2005 			    pgflags, lgrp);
2006 			if (pp == NULL) {
2007 				pp = page_get_freelist(vp, 0, seg, addr, pgsz,
2008 				    0, lgrp);
2009 			}
2010 		} else {
2011 			pp = page_get_freelist(vp, 0, seg, addr, pgsz,
2012 			    0, lgrp);
2013 		}
2014 		if (pp != NULL) {
2015 			VM_STAT_ADD(alloc_pages[1]);
2016 			page_list_concat(&pplist, &pp);
2017 			ASSERT(npgs >= curnpgs);
2018 			npgs -= curnpgs;
2019 		} else if (anypgsz) {
2020 			VM_STAT_ADD(alloc_pages[2]);
2021 			szc--;
2022 			pgsz = page_get_pagesize(szc);
2023 			curnpgs = pgsz >> PAGESHIFT;
2024 		} else {
2025 			VM_STAT_ADD(alloc_pages[3]);
2026 			ASSERT(npgs == totpgs);
2027 			page_create_putback(npgs);
2028 			return (ENOMEM);
2029 		}
2030 	}
2031 	if (szc == 0) {
2032 		VM_STAT_ADD(alloc_pages[4]);
2033 		ASSERT(npgs != 0);
2034 		page_create_putback(npgs);
2035 		err = ENOMEM;
2036 	} else if (basepp != NULL) {
2037 		ASSERT(npgs == 0);
2038 		ASSERT(ppa == NULL);
2039 		*basepp = pplist;
2040 	}
2041 
2042 	npgs = totpgs - npgs;
2043 	pp = pplist;
2044 
2045 	/*
2046 	 * Clear the free and age bits. Also if we were passed in a ppa then
2047 	 * fill it in with all the constituent pages from the large page. But
2048 	 * if we failed to allocate all the pages just free what we got.
2049 	 */
2050 	while (npgs != 0) {
2051 		ASSERT(PP_ISFREE(pp));
2052 		ASSERT(PP_ISAGED(pp));
2053 		if (ppa != NULL || err != 0) {
2054 			if (err == 0) {
2055 				VM_STAT_ADD(alloc_pages[5]);
2056 				PP_CLRFREE(pp);
2057 				PP_CLRAGED(pp);
2058 				page_sub(&pplist, pp);
2059 				*ppa++ = pp;
2060 				npgs--;
2061 			} else {
2062 				VM_STAT_ADD(alloc_pages[6]);
2063 				ASSERT(pp->p_szc != 0);
2064 				curnpgs = page_get_pagecnt(pp->p_szc);
2065 				page_list_break(&pp, &pplist, curnpgs);
2066 				page_list_add_pages(pp, 0);
2067 				page_create_putback(curnpgs);
2068 				ASSERT(npgs >= curnpgs);
2069 				npgs -= curnpgs;
2070 			}
2071 			pp = pplist;
2072 		} else {
2073 			VM_STAT_ADD(alloc_pages[7]);
2074 			PP_CLRFREE(pp);
2075 			PP_CLRAGED(pp);
2076 			pp = pp->p_next;
2077 			npgs--;
2078 		}
2079 	}
2080 	return (err);
2081 }
2082 
2083 /*
2084  * Get a single large page off of the freelists, and set it up for use.
2085  * Number of bytes requested must be a supported page size.
2086  *
2087  * Note that this call may fail even if there is sufficient
2088  * memory available or PG_WAIT is set, so the caller must
2089  * be willing to fallback on page_create_va(), block and retry,
2090  * or fail the requester.
2091  */
2092 page_t *
2093 page_create_va_large(vnode_t *vp, u_offset_t off, size_t bytes, uint_t flags,
2094     struct seg *seg, caddr_t vaddr, void *arg)
2095 {
2096 	pgcnt_t		npages, pcftotal;
2097 	page_t		*pp;
2098 	page_t		*rootpp;
2099 	lgrp_t		*lgrp;
2100 	uint_t		enough;
2101 	uint_t		pcf_index;
2102 	uint_t		i;
2103 	struct pcf	*p;
2104 	struct pcf	*q;
2105 	lgrp_id_t	*lgrpid = (lgrp_id_t *)arg;
2106 
2107 	ASSERT(vp != NULL);
2108 
2109 	ASSERT((flags & ~(PG_EXCL | PG_WAIT |
2110 	    PG_NORELOC | PG_PANIC | PG_PUSHPAGE)) == 0);
2111 	/* but no others */
2112 
2113 	ASSERT((flags & PG_EXCL) == PG_EXCL);
2114 
2115 	npages = btop(bytes);
2116 
2117 	if (!kcage_on || panicstr) {
2118 		/*
2119 		 * Cage is OFF, or we are single threaded in
2120 		 * panic, so make everything a RELOC request.
2121 		 */
2122 		flags &= ~PG_NORELOC;
2123 	}
2124 
2125 	/*
2126 	 * Make sure there's adequate physical memory available.
2127 	 * Note: PG_WAIT is ignored here.
2128 	 */
2129 	if (freemem <= throttlefree + npages) {
2130 		VM_STAT_ADD(page_create_large_cnt[1]);
2131 		return (NULL);
2132 	}
2133 
2134 	/*
2135 	 * If cage is on, dampen draw from cage when available
2136 	 * cage space is low.
2137 	 */
2138 	if ((flags & (PG_NORELOC | PG_WAIT)) ==  (PG_NORELOC | PG_WAIT) &&
2139 	    kcage_freemem < kcage_throttlefree + npages) {
2140 
2141 		/*
2142 		 * The cage is on, the caller wants PG_NORELOC
2143 		 * pages and available cage memory is very low.
2144 		 * Call kcage_create_throttle() to attempt to
2145 		 * control demand on the cage.
2146 		 */
2147 		if (kcage_create_throttle(npages, flags) == KCT_FAILURE) {
2148 			VM_STAT_ADD(page_create_large_cnt[2]);
2149 			return (NULL);
2150 		}
2151 	}
2152 
2153 	enough = 0;
2154 	pcf_index = PCF_INDEX();
2155 	p = &pcf[pcf_index];
2156 	q = &pcf[PCF_FANOUT];
2157 	for (pcftotal = 0, i = 0; i < PCF_FANOUT; i++) {
2158 		if (p->pcf_count > npages) {
2159 			/*
2160 			 * a good one to try.
2161 			 */
2162 			mutex_enter(&p->pcf_lock);
2163 			if (p->pcf_count > npages) {
2164 				p->pcf_count -= (uint_t)npages;
2165 				/*
2166 				 * freemem is not protected by any lock.
2167 				 * Thus, we cannot have any assertion
2168 				 * containing freemem here.
2169 				 */
2170 				freemem -= npages;
2171 				enough = 1;
2172 				mutex_exit(&p->pcf_lock);
2173 				break;
2174 			}
2175 			mutex_exit(&p->pcf_lock);
2176 		}
2177 		pcftotal += p->pcf_count;
2178 		p++;
2179 		if (p >= q) {
2180 			p = pcf;
2181 		}
2182 	}
2183 
2184 	if (!enough) {
2185 		/* If there isn't enough memory available, give up. */
2186 		if (pcftotal < npages) {
2187 			VM_STAT_ADD(page_create_large_cnt[3]);
2188 			return (NULL);
2189 		}
2190 
2191 		/* try to collect pages from several pcf bins */
2192 		for (p = pcf, pcftotal = 0, i = 0; i < PCF_FANOUT; i++) {
2193 			mutex_enter(&p->pcf_lock);
2194 			pcftotal += p->pcf_count;
2195 			if (pcftotal >= npages) {
2196 				/*
2197 				 * Wow!  There are enough pages laying around
2198 				 * to satisfy the request.  Do the accounting,
2199 				 * drop the locks we acquired, and go back.
2200 				 *
2201 				 * freemem is not protected by any lock. So,
2202 				 * we cannot have any assertion containing
2203 				 * freemem.
2204 				 */
2205 				pgcnt_t	tpages = npages;
2206 				freemem -= npages;
2207 				while (p >= pcf) {
2208 					if (p->pcf_count <= tpages) {
2209 						tpages -= p->pcf_count;
2210 						p->pcf_count = 0;
2211 					} else {
2212 						p->pcf_count -= (uint_t)tpages;
2213 						tpages = 0;
2214 					}
2215 					mutex_exit(&p->pcf_lock);
2216 					p--;
2217 				}
2218 				ASSERT(tpages == 0);
2219 				break;
2220 			}
2221 			p++;
2222 		}
2223 		if (i == PCF_FANOUT) {
2224 			/* failed to collect pages - release the locks */
2225 			while (--p >= pcf) {
2226 				mutex_exit(&p->pcf_lock);
2227 			}
2228 			VM_STAT_ADD(page_create_large_cnt[4]);
2229 			return (NULL);
2230 		}
2231 	}
2232 
2233 	/*
2234 	 * This is where this function behaves fundamentally differently
2235 	 * than page_create_va(); since we're intending to map the page
2236 	 * with a single TTE, we have to get it as a physically contiguous
2237 	 * hardware pagesize chunk.  If we can't, we fail.
2238 	 */
2239 	if (lgrpid != NULL && *lgrpid >= 0 && *lgrpid <= lgrp_alloc_max &&
2240 	    LGRP_EXISTS(lgrp_table[*lgrpid]))
2241 		lgrp = lgrp_table[*lgrpid];
2242 	else
2243 		lgrp = lgrp_mem_choose(seg, vaddr, bytes);
2244 
2245 	if ((rootpp = page_get_freelist(&kvp, off, seg, vaddr,
2246 	    bytes, flags & ~PG_MATCH_COLOR, lgrp)) == NULL) {
2247 		page_create_putback(npages);
2248 		VM_STAT_ADD(page_create_large_cnt[5]);
2249 		return (NULL);
2250 	}
2251 
2252 	/*
2253 	 * if we got the page with the wrong mtype give it back this is a
2254 	 * workaround for CR 6249718. When CR 6249718 is fixed we never get
2255 	 * inside "if" and the workaround becomes just a nop
2256 	 */
2257 	if (kcage_on && (flags & PG_NORELOC) && !PP_ISNORELOC(rootpp)) {
2258 		page_list_add_pages(rootpp, 0);
2259 		page_create_putback(npages);
2260 		VM_STAT_ADD(page_create_large_cnt[6]);
2261 		return (NULL);
2262 	}
2263 
2264 	/*
2265 	 * If satisfying this request has left us with too little
2266 	 * memory, start the wheels turning to get some back.  The
2267 	 * first clause of the test prevents waking up the pageout
2268 	 * daemon in situations where it would decide that there's
2269 	 * nothing to do.
2270 	 */
2271 	if (nscan < desscan && freemem < minfree) {
2272 		TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL,
2273 		    "pageout_cv_signal:freemem %ld", freemem);
2274 		cv_signal(&proc_pageout->p_cv);
2275 	}
2276 
2277 	pp = rootpp;
2278 	while (npages--) {
2279 		ASSERT(PAGE_EXCL(pp));
2280 		ASSERT(pp->p_vnode == NULL);
2281 		ASSERT(!hat_page_is_mapped(pp));
2282 		PP_CLRFREE(pp);
2283 		PP_CLRAGED(pp);
2284 		if (!page_hashin(pp, vp, off, NULL))
2285 			panic("page_create_large: hashin failed: page %p",
2286 			    (void *)pp);
2287 		page_io_lock(pp);
2288 		off += PAGESIZE;
2289 		pp = pp->p_next;
2290 	}
2291 
2292 	VM_STAT_ADD(page_create_large_cnt[0]);
2293 	return (rootpp);
2294 }
2295 
2296 page_t *
2297 page_create_va(vnode_t *vp, u_offset_t off, size_t bytes, uint_t flags,
2298     struct seg *seg, caddr_t vaddr)
2299 {
2300 	page_t		*plist = NULL;
2301 	pgcnt_t		npages;
2302 	pgcnt_t		found_on_free = 0;
2303 	pgcnt_t		pages_req;
2304 	page_t		*npp = NULL;
2305 	uint_t		enough;
2306 	uint_t		i;
2307 	uint_t		pcf_index;
2308 	struct pcf	*p;
2309 	struct pcf	*q;
2310 	lgrp_t		*lgrp;
2311 
2312 	TRACE_4(TR_FAC_VM, TR_PAGE_CREATE_START,
2313 	    "page_create_start:vp %p off %llx bytes %lu flags %x",
2314 	    vp, off, bytes, flags);
2315 
2316 	ASSERT(bytes != 0 && vp != NULL);
2317 
2318 	if ((flags & PG_EXCL) == 0 && (flags & PG_WAIT) == 0) {
2319 		panic("page_create: invalid flags");
2320 		/*NOTREACHED*/
2321 	}
2322 	ASSERT((flags & ~(PG_EXCL | PG_WAIT |
2323 	    PG_NORELOC | PG_PANIC | PG_PUSHPAGE)) == 0);
2324 	    /* but no others */
2325 
2326 	pages_req = npages = btopr(bytes);
2327 	/*
2328 	 * Try to see whether request is too large to *ever* be
2329 	 * satisfied, in order to prevent deadlock.  We arbitrarily
2330 	 * decide to limit maximum size requests to max_page_get.
2331 	 */
2332 	if (npages >= max_page_get) {
2333 		if ((flags & PG_WAIT) == 0) {
2334 			TRACE_4(TR_FAC_VM, TR_PAGE_CREATE_TOOBIG,
2335 			    "page_create_toobig:vp %p off %llx npages "
2336 			    "%lu max_page_get %lu",
2337 			    vp, off, npages, max_page_get);
2338 			return (NULL);
2339 		} else {
2340 			cmn_err(CE_WARN,
2341 			    "Request for too much kernel memory "
2342 			    "(%lu bytes), will hang forever", bytes);
2343 			for (;;)
2344 				delay(1000000000);
2345 		}
2346 	}
2347 
2348 	if (!kcage_on || panicstr) {
2349 		/*
2350 		 * Cage is OFF, or we are single threaded in
2351 		 * panic, so make everything a RELOC request.
2352 		 */
2353 		flags &= ~PG_NORELOC;
2354 	}
2355 
2356 	if (freemem <= throttlefree + npages)
2357 		if (!page_create_throttle(npages, flags))
2358 			return (NULL);
2359 
2360 	/*
2361 	 * If cage is on, dampen draw from cage when available
2362 	 * cage space is low.
2363 	 */
2364 	if ((flags & PG_NORELOC) &&
2365 	    kcage_freemem < kcage_throttlefree + npages) {
2366 
2367 		/*
2368 		 * The cage is on, the caller wants PG_NORELOC
2369 		 * pages and available cage memory is very low.
2370 		 * Call kcage_create_throttle() to attempt to
2371 		 * control demand on the cage.
2372 		 */
2373 		if (kcage_create_throttle(npages, flags) == KCT_FAILURE)
2374 			return (NULL);
2375 	}
2376 
2377 	VM_STAT_ADD(page_create_cnt[0]);
2378 
2379 	enough = 0;
2380 	pcf_index = PCF_INDEX();
2381 
2382 	p = &pcf[pcf_index];
2383 	q = &pcf[PCF_FANOUT];
2384 	for (i = 0; i < PCF_FANOUT; i++) {
2385 		if (p->pcf_count > npages) {
2386 			/*
2387 			 * a good one to try.
2388 			 */
2389 			mutex_enter(&p->pcf_lock);
2390 			if (p->pcf_count > npages) {
2391 				p->pcf_count -= (uint_t)npages;
2392 				/*
2393 				 * freemem is not protected by any lock.
2394 				 * Thus, we cannot have any assertion
2395 				 * containing freemem here.
2396 				 */
2397 				freemem -= npages;
2398 				enough = 1;
2399 				mutex_exit(&p->pcf_lock);
2400 				break;
2401 			}
2402 			mutex_exit(&p->pcf_lock);
2403 		}
2404 		p++;
2405 		if (p >= q) {
2406 			p = pcf;
2407 		}
2408 	}
2409 
2410 	if (!enough) {
2411 		/*
2412 		 * Have to look harder.  If npages is greater than
2413 		 * one, then we might have to coalesce the counters.
2414 		 *
2415 		 * Go wait.  We come back having accounted
2416 		 * for the memory.
2417 		 */
2418 		VM_STAT_ADD(page_create_cnt[1]);
2419 		if (!page_create_wait(npages, flags)) {
2420 			VM_STAT_ADD(page_create_cnt[2]);
2421 			return (NULL);
2422 		}
2423 	}
2424 
2425 	TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SUCCESS,
2426 	    "page_create_success:vp %p off %llx", vp, off);
2427 
2428 	/*
2429 	 * If satisfying this request has left us with too little
2430 	 * memory, start the wheels turning to get some back.  The
2431 	 * first clause of the test prevents waking up the pageout
2432 	 * daemon in situations where it would decide that there's
2433 	 * nothing to do.
2434 	 */
2435 	if (nscan < desscan && freemem < minfree) {
2436 		TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL,
2437 		    "pageout_cv_signal:freemem %ld", freemem);
2438 		cv_signal(&proc_pageout->p_cv);
2439 	}
2440 
2441 	/*
2442 	 * Loop around collecting the requested number of pages.
2443 	 * Most of the time, we have to `create' a new page. With
2444 	 * this in mind, pull the page off the free list before
2445 	 * getting the hash lock.  This will minimize the hash
2446 	 * lock hold time, nesting, and the like.  If it turns
2447 	 * out we don't need the page, we put it back at the end.
2448 	 */
2449 	while (npages--) {
2450 		page_t		*pp;
2451 		kmutex_t	*phm = NULL;
2452 		ulong_t		index;
2453 
2454 		index = PAGE_HASH_FUNC(vp, off);
2455 top:
2456 		ASSERT(phm == NULL);
2457 		ASSERT(index == PAGE_HASH_FUNC(vp, off));
2458 		ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
2459 
2460 		if (npp == NULL) {
2461 			/*
2462 			 * Try to get a page from the freelist (ie,
2463 			 * a page with no [vp, off] tag).  If that
2464 			 * fails, use the cachelist.
2465 			 *
2466 			 * During the first attempt at both the free
2467 			 * and cache lists we try for the correct color.
2468 			 */
2469 			/*
2470 			 * XXXX-how do we deal with virtual indexed
2471 			 * caches and and colors?
2472 			 */
2473 			VM_STAT_ADD(page_create_cnt[4]);
2474 			/*
2475 			 * Get lgroup to allocate next page of shared memory
2476 			 * from and use it to specify where to allocate
2477 			 * the physical memory
2478 			 */
2479 			lgrp = lgrp_mem_choose(seg, vaddr, PAGESIZE);
2480 			npp = page_get_freelist(vp, off, seg, vaddr, PAGESIZE,
2481 			    flags | PG_MATCH_COLOR, lgrp);
2482 			if (npp == NULL) {
2483 				npp = page_get_cachelist(vp, off, seg,
2484 				    vaddr, flags | PG_MATCH_COLOR, lgrp);
2485 				if (npp == NULL) {
2486 					npp = page_create_get_something(vp,
2487 					    off, seg, vaddr,
2488 					    flags & ~PG_MATCH_COLOR);
2489 				}
2490 
2491 				if (PP_ISAGED(npp) == 0) {
2492 					/*
2493 					 * Since this page came from the
2494 					 * cachelist, we must destroy the
2495 					 * old vnode association.
2496 					 */
2497 					page_hashout(npp, NULL);
2498 				}
2499 			}
2500 		}
2501 
2502 		/*
2503 		 * We own this page!
2504 		 */
2505 		ASSERT(PAGE_EXCL(npp));
2506 		ASSERT(npp->p_vnode == NULL);
2507 		ASSERT(!hat_page_is_mapped(npp));
2508 		PP_CLRFREE(npp);
2509 		PP_CLRAGED(npp);
2510 
2511 		/*
2512 		 * Here we have a page in our hot little mits and are
2513 		 * just waiting to stuff it on the appropriate lists.
2514 		 * Get the mutex and check to see if it really does
2515 		 * not exist.
2516 		 */
2517 		phm = PAGE_HASH_MUTEX(index);
2518 		mutex_enter(phm);
2519 		PAGE_HASH_SEARCH(index, pp, vp, off);
2520 		if (pp == NULL) {
2521 			VM_STAT_ADD(page_create_new);
2522 			pp = npp;
2523 			npp = NULL;
2524 			if (!page_hashin(pp, vp, off, phm)) {
2525 				/*
2526 				 * Since we hold the page hash mutex and
2527 				 * just searched for this page, page_hashin
2528 				 * had better not fail.  If it does, that
2529 				 * means somethread did not follow the
2530 				 * page hash mutex rules.  Panic now and
2531 				 * get it over with.  As usual, go down
2532 				 * holding all the locks.
2533 				 */
2534 				ASSERT(MUTEX_HELD(phm));
2535 				panic("page_create: "
2536 				    "hashin failed %p %p %llx %p",
2537 				    (void *)pp, (void *)vp, off, (void *)phm);
2538 				/*NOTREACHED*/
2539 			}
2540 			ASSERT(MUTEX_HELD(phm));
2541 			mutex_exit(phm);
2542 			phm = NULL;
2543 
2544 			/*
2545 			 * Hat layer locking need not be done to set
2546 			 * the following bits since the page is not hashed
2547 			 * and was on the free list (i.e., had no mappings).
2548 			 *
2549 			 * Set the reference bit to protect
2550 			 * against immediate pageout
2551 			 *
2552 			 * XXXmh modify freelist code to set reference
2553 			 * bit so we don't have to do it here.
2554 			 */
2555 			page_set_props(pp, P_REF);
2556 			found_on_free++;
2557 		} else {
2558 			VM_STAT_ADD(page_create_exists);
2559 			if (flags & PG_EXCL) {
2560 				/*
2561 				 * Found an existing page, and the caller
2562 				 * wanted all new pages.  Undo all of the work
2563 				 * we have done.
2564 				 */
2565 				mutex_exit(phm);
2566 				phm = NULL;
2567 				while (plist != NULL) {
2568 					pp = plist;
2569 					page_sub(&plist, pp);
2570 					page_io_unlock(pp);
2571 					/* large pages should not end up here */
2572 					ASSERT(pp->p_szc == 0);
2573 					/*LINTED: constant in conditional ctx*/
2574 					VN_DISPOSE(pp, B_INVAL, 0, kcred);
2575 				}
2576 				VM_STAT_ADD(page_create_found_one);
2577 				goto fail;
2578 			}
2579 			ASSERT(flags & PG_WAIT);
2580 			if (!page_lock(pp, SE_EXCL, phm, P_NO_RECLAIM)) {
2581 				/*
2582 				 * Start all over again if we blocked trying
2583 				 * to lock the page.
2584 				 */
2585 				mutex_exit(phm);
2586 				VM_STAT_ADD(page_create_page_lock_failed);
2587 				phm = NULL;
2588 				goto top;
2589 			}
2590 			mutex_exit(phm);
2591 			phm = NULL;
2592 
2593 			if (PP_ISFREE(pp)) {
2594 				ASSERT(PP_ISAGED(pp) == 0);
2595 				VM_STAT_ADD(pagecnt.pc_get_cache);
2596 				page_list_sub(pp, PG_CACHE_LIST);
2597 				PP_CLRFREE(pp);
2598 				found_on_free++;
2599 			}
2600 		}
2601 
2602 		/*
2603 		 * Got a page!  It is locked.  Acquire the i/o
2604 		 * lock since we are going to use the p_next and
2605 		 * p_prev fields to link the requested pages together.
2606 		 */
2607 		page_io_lock(pp);
2608 		page_add(&plist, pp);
2609 		plist = plist->p_next;
2610 		off += PAGESIZE;
2611 		vaddr += PAGESIZE;
2612 	}
2613 
2614 	ASSERT((flags & PG_EXCL) ? (found_on_free == pages_req) : 1);
2615 fail:
2616 	if (npp != NULL) {
2617 		/*
2618 		 * Did not need this page after all.
2619 		 * Put it back on the free list.
2620 		 */
2621 		VM_STAT_ADD(page_create_putbacks);
2622 		PP_SETFREE(npp);
2623 		PP_SETAGED(npp);
2624 		npp->p_offset = (u_offset_t)-1;
2625 		page_list_add(npp, PG_FREE_LIST | PG_LIST_TAIL);
2626 		page_unlock(npp);
2627 
2628 	}
2629 
2630 	ASSERT(pages_req >= found_on_free);
2631 
2632 	{
2633 		uint_t overshoot = (uint_t)(pages_req - found_on_free);
2634 
2635 		if (overshoot) {
2636 			VM_STAT_ADD(page_create_overshoot);
2637 			p = &pcf[pcf_index];
2638 			mutex_enter(&p->pcf_lock);
2639 			if (p->pcf_block) {
2640 				p->pcf_reserve += overshoot;
2641 			} else {
2642 				p->pcf_count += overshoot;
2643 				if (p->pcf_wait) {
2644 					mutex_enter(&new_freemem_lock);
2645 					if (freemem_wait) {
2646 						cv_signal(&freemem_cv);
2647 						p->pcf_wait--;
2648 					} else {
2649 						p->pcf_wait = 0;
2650 					}
2651 					mutex_exit(&new_freemem_lock);
2652 				}
2653 			}
2654 			mutex_exit(&p->pcf_lock);
2655 			/* freemem is approximate, so this test OK */
2656 			if (!p->pcf_block)
2657 				freemem += overshoot;
2658 		}
2659 	}
2660 
2661 	return (plist);
2662 }
2663 
2664 /*
2665  * One or more constituent pages of this large page has been marked
2666  * toxic. Simply demote the large page to PAGESIZE pages and let
2667  * page_free() handle it. This routine should only be called by
2668  * large page free routines (page_free_pages() and page_destroy_pages().
2669  * All pages are locked SE_EXCL and have already been marked free.
2670  */
2671 static void
2672 page_free_toxic_pages(page_t *rootpp)
2673 {
2674 	page_t	*tpp;
2675 	pgcnt_t	i, pgcnt = page_get_pagecnt(rootpp->p_szc);
2676 	uint_t	szc = rootpp->p_szc;
2677 
2678 	for (i = 0, tpp = rootpp; i < pgcnt; i++, tpp = tpp->p_next) {
2679 		ASSERT(tpp->p_szc == szc);
2680 		ASSERT((PAGE_EXCL(tpp) &&
2681 		    !page_iolock_assert(tpp)) || panicstr);
2682 		tpp->p_szc = 0;
2683 	}
2684 
2685 	while (rootpp != NULL) {
2686 		tpp = rootpp;
2687 		page_sub(&rootpp, tpp);
2688 		ASSERT(PP_ISFREE(tpp));
2689 		PP_CLRFREE(tpp);
2690 		page_free(tpp, 1);
2691 	}
2692 }
2693 
2694 /*
2695  * Put page on the "free" list.
2696  * The free list is really two lists maintained by
2697  * the PSM of whatever machine we happen to be on.
2698  */
2699 void
2700 page_free(page_t *pp, int dontneed)
2701 {
2702 	struct pcf	*p;
2703 	uint_t		pcf_index;
2704 
2705 	ASSERT((PAGE_EXCL(pp) &&
2706 	    !page_iolock_assert(pp)) || panicstr);
2707 
2708 	if (PP_ISFREE(pp)) {
2709 		panic("page_free: page %p is free", (void *)pp);
2710 	}
2711 
2712 	if (pp->p_szc != 0) {
2713 		if (pp->p_vnode == NULL || IS_SWAPFSVP(pp->p_vnode) ||
2714 		    PP_ISKAS(pp)) {
2715 			panic("page_free: anon or kernel "
2716 			    "or no vnode large page %p", (void *)pp);
2717 		}
2718 		page_demote_vp_pages(pp);
2719 		ASSERT(pp->p_szc == 0);
2720 	}
2721 
2722 	/*
2723 	 * The page_struct_lock need not be acquired to examine these
2724 	 * fields since the page has an "exclusive" lock.
2725 	 */
2726 	if (hat_page_is_mapped(pp) || pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
2727 	    pp->p_slckcnt != 0) {
2728 		panic("page_free pp=%p, pfn=%lx, lckcnt=%d, cowcnt=%d "
2729 		    "slckcnt = %d", pp, page_pptonum(pp), pp->p_lckcnt,
2730 		    pp->p_cowcnt, pp->p_slckcnt);
2731 		/*NOTREACHED*/
2732 	}
2733 
2734 	ASSERT(!hat_page_getshare(pp));
2735 
2736 	PP_SETFREE(pp);
2737 	ASSERT(pp->p_vnode == NULL || !IS_VMODSORT(pp->p_vnode) ||
2738 	    !hat_ismod(pp));
2739 	page_clr_all_props(pp);
2740 	ASSERT(!hat_page_getshare(pp));
2741 
2742 	/*
2743 	 * Now we add the page to the head of the free list.
2744 	 * But if this page is associated with a paged vnode
2745 	 * then we adjust the head forward so that the page is
2746 	 * effectively at the end of the list.
2747 	 */
2748 	if (pp->p_vnode == NULL) {
2749 		/*
2750 		 * Page has no identity, put it on the free list.
2751 		 */
2752 		PP_SETAGED(pp);
2753 		pp->p_offset = (u_offset_t)-1;
2754 		page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL);
2755 		VM_STAT_ADD(pagecnt.pc_free_free);
2756 		TRACE_1(TR_FAC_VM, TR_PAGE_FREE_FREE,
2757 		    "page_free_free:pp %p", pp);
2758 	} else {
2759 		PP_CLRAGED(pp);
2760 
2761 		if (!dontneed || nopageage) {
2762 			/* move it to the tail of the list */
2763 			page_list_add(pp, PG_CACHE_LIST | PG_LIST_TAIL);
2764 
2765 			VM_STAT_ADD(pagecnt.pc_free_cache);
2766 			TRACE_1(TR_FAC_VM, TR_PAGE_FREE_CACHE_TAIL,
2767 			    "page_free_cache_tail:pp %p", pp);
2768 		} else {
2769 			page_list_add(pp, PG_CACHE_LIST | PG_LIST_HEAD);
2770 
2771 			VM_STAT_ADD(pagecnt.pc_free_dontneed);
2772 			TRACE_1(TR_FAC_VM, TR_PAGE_FREE_CACHE_HEAD,
2773 			    "page_free_cache_head:pp %p", pp);
2774 		}
2775 	}
2776 	page_unlock(pp);
2777 
2778 	/*
2779 	 * Now do the `freemem' accounting.
2780 	 */
2781 	pcf_index = PCF_INDEX();
2782 	p = &pcf[pcf_index];
2783 
2784 	mutex_enter(&p->pcf_lock);
2785 	if (p->pcf_block) {
2786 		p->pcf_reserve += 1;
2787 	} else {
2788 		p->pcf_count += 1;
2789 		if (p->pcf_wait) {
2790 			mutex_enter(&new_freemem_lock);
2791 			/*
2792 			 * Check to see if some other thread
2793 			 * is actually waiting.  Another bucket
2794 			 * may have woken it up by now.  If there
2795 			 * are no waiters, then set our pcf_wait
2796 			 * count to zero to avoid coming in here
2797 			 * next time.  Also, since only one page
2798 			 * was put on the free list, just wake
2799 			 * up one waiter.
2800 			 */
2801 			if (freemem_wait) {
2802 				cv_signal(&freemem_cv);
2803 				p->pcf_wait--;
2804 			} else {
2805 				p->pcf_wait = 0;
2806 			}
2807 			mutex_exit(&new_freemem_lock);
2808 		}
2809 	}
2810 	mutex_exit(&p->pcf_lock);
2811 
2812 	/* freemem is approximate, so this test OK */
2813 	if (!p->pcf_block)
2814 		freemem += 1;
2815 }
2816 
2817 /*
2818  * Put page on the "free" list during intial startup.
2819  * This happens during initial single threaded execution.
2820  */
2821 void
2822 page_free_at_startup(page_t *pp)
2823 {
2824 	struct pcf	*p;
2825 	uint_t		pcf_index;
2826 
2827 	page_list_add(pp, PG_FREE_LIST | PG_LIST_HEAD | PG_LIST_ISINIT);
2828 	VM_STAT_ADD(pagecnt.pc_free_free);
2829 
2830 	/*
2831 	 * Now do the `freemem' accounting.
2832 	 */
2833 	pcf_index = PCF_INDEX();
2834 	p = &pcf[pcf_index];
2835 
2836 	ASSERT(p->pcf_block == 0);
2837 	ASSERT(p->pcf_wait == 0);
2838 	p->pcf_count += 1;
2839 
2840 	/* freemem is approximate, so this is OK */
2841 	freemem += 1;
2842 }
2843 
2844 void
2845 page_free_pages(page_t *pp)
2846 {
2847 	page_t	*tpp, *rootpp = NULL;
2848 	pgcnt_t	pgcnt = page_get_pagecnt(pp->p_szc);
2849 	pgcnt_t	i;
2850 	uint_t	szc = pp->p_szc;
2851 
2852 	VM_STAT_ADD(pagecnt.pc_free_pages);
2853 	TRACE_1(TR_FAC_VM, TR_PAGE_FREE_FREE,
2854 	    "page_free_free:pp %p", pp);
2855 
2856 	ASSERT(pp->p_szc != 0 && pp->p_szc < page_num_pagesizes());
2857 	if ((page_pptonum(pp) & (pgcnt - 1)) != 0) {
2858 		panic("page_free_pages: not root page %p", (void *)pp);
2859 		/*NOTREACHED*/
2860 	}
2861 
2862 	for (i = 0, tpp = pp; i < pgcnt; i++, tpp++) {
2863 		ASSERT((PAGE_EXCL(tpp) &&
2864 		    !page_iolock_assert(tpp)) || panicstr);
2865 		if (PP_ISFREE(tpp)) {
2866 			panic("page_free_pages: page %p is free", (void *)tpp);
2867 			/*NOTREACHED*/
2868 		}
2869 		if (hat_page_is_mapped(tpp) || tpp->p_lckcnt != 0 ||
2870 		    tpp->p_cowcnt != 0 || tpp->p_slckcnt != 0) {
2871 			panic("page_free_pages %p", (void *)tpp);
2872 			/*NOTREACHED*/
2873 		}
2874 
2875 		ASSERT(!hat_page_getshare(tpp));
2876 		ASSERT(tpp->p_vnode == NULL);
2877 		ASSERT(tpp->p_szc == szc);
2878 
2879 		PP_SETFREE(tpp);
2880 		page_clr_all_props(tpp);
2881 		PP_SETAGED(tpp);
2882 		tpp->p_offset = (u_offset_t)-1;
2883 		ASSERT(tpp->p_next == tpp);
2884 		ASSERT(tpp->p_prev == tpp);
2885 		page_list_concat(&rootpp, &tpp);
2886 	}
2887 	ASSERT(rootpp == pp);
2888 
2889 	page_list_add_pages(rootpp, 0);
2890 	page_create_putback(pgcnt);
2891 }
2892 
2893 int free_pages = 1;
2894 
2895 /*
2896  * This routine attempts to return pages to the cachelist via page_release().
2897  * It does not *have* to be successful in all cases, since the pageout scanner
2898  * will catch any pages it misses.  It does need to be fast and not introduce
2899  * too much overhead.
2900  *
2901  * If a page isn't found on the unlocked sweep of the page_hash bucket, we
2902  * don't lock and retry.  This is ok, since the page scanner will eventually
2903  * find any page we miss in free_vp_pages().
2904  */
2905 void
2906 free_vp_pages(vnode_t *vp, u_offset_t off, size_t len)
2907 {
2908 	page_t *pp;
2909 	u_offset_t eoff;
2910 	extern int swap_in_range(vnode_t *, u_offset_t, size_t);
2911 
2912 	eoff = off + len;
2913 
2914 	if (free_pages == 0)
2915 		return;
2916 	if (swap_in_range(vp, off, len))
2917 		return;
2918 
2919 	for (; off < eoff; off += PAGESIZE) {
2920 
2921 		/*
2922 		 * find the page using a fast, but inexact search. It'll be OK
2923 		 * if a few pages slip through the cracks here.
2924 		 */
2925 		pp = page_exists(vp, off);
2926 
2927 		/*
2928 		 * If we didn't find the page (it may not exist), the page
2929 		 * is free, looks still in use (shared), or we can't lock it,
2930 		 * just give up.
2931 		 */
2932 		if (pp == NULL ||
2933 		    PP_ISFREE(pp) ||
2934 		    page_share_cnt(pp) > 0 ||
2935 		    !page_trylock(pp, SE_EXCL))
2936 			continue;
2937 
2938 		/*
2939 		 * Once we have locked pp, verify that it's still the
2940 		 * correct page and not already free
2941 		 */
2942 		ASSERT(PAGE_LOCKED_SE(pp, SE_EXCL));
2943 		if (pp->p_vnode != vp || pp->p_offset != off || PP_ISFREE(pp)) {
2944 			page_unlock(pp);
2945 			continue;
2946 		}
2947 
2948 		/*
2949 		 * try to release the page...
2950 		 */
2951 		(void) page_release(pp, 1);
2952 	}
2953 }
2954 
2955 /*
2956  * Reclaim the given page from the free list.
2957  * If pp is part of a large pages, only the given constituent page is reclaimed
2958  * and the large page it belonged to will be demoted.  This can only happen
2959  * if the page is not on the cachelist.
2960  *
2961  * Returns 1 on success or 0 on failure.
2962  *
2963  * The page is unlocked if it can't be reclaimed (when freemem == 0).
2964  * If `lock' is non-null, it will be dropped and re-acquired if
2965  * the routine must wait while freemem is 0.
2966  *
2967  * As it turns out, boot_getpages() does this.  It picks a page,
2968  * based on where OBP mapped in some address, gets its pfn, searches
2969  * the memsegs, locks the page, then pulls it off the free list!
2970  */
2971 int
2972 page_reclaim(page_t *pp, kmutex_t *lock)
2973 {
2974 	struct pcf	*p;
2975 	uint_t		pcf_index;
2976 	struct cpu	*cpup;
2977 	int		enough;
2978 	uint_t		i;
2979 
2980 	ASSERT(lock != NULL ? MUTEX_HELD(lock) : 1);
2981 	ASSERT(PAGE_EXCL(pp) && PP_ISFREE(pp));
2982 
2983 	/*
2984 	 * If `freemem' is 0, we cannot reclaim this page from the
2985 	 * freelist, so release every lock we might hold: the page,
2986 	 * and the `lock' before blocking.
2987 	 *
2988 	 * The only way `freemem' can become 0 while there are pages
2989 	 * marked free (have their p->p_free bit set) is when the
2990 	 * system is low on memory and doing a page_create().  In
2991 	 * order to guarantee that once page_create() starts acquiring
2992 	 * pages it will be able to get all that it needs since `freemem'
2993 	 * was decreased by the requested amount.  So, we need to release
2994 	 * this page, and let page_create() have it.
2995 	 *
2996 	 * Since `freemem' being zero is not supposed to happen, just
2997 	 * use the usual hash stuff as a starting point.  If that bucket
2998 	 * is empty, then assume the worst, and start at the beginning
2999 	 * of the pcf array.  If we always start at the beginning
3000 	 * when acquiring more than one pcf lock, there won't be any
3001 	 * deadlock problems.
3002 	 */
3003 
3004 	/* TODO: Do we need to test kcage_freemem if PG_NORELOC(pp)? */
3005 
3006 	if (freemem <= throttlefree && !page_create_throttle(1l, 0)) {
3007 		pcf_acquire_all();
3008 		goto page_reclaim_nomem;
3009 	}
3010 
3011 	enough = 0;
3012 	pcf_index = PCF_INDEX();
3013 	p = &pcf[pcf_index];
3014 	mutex_enter(&p->pcf_lock);
3015 	if (p->pcf_count >= 1) {
3016 		enough = 1;
3017 		p->pcf_count--;
3018 	}
3019 	mutex_exit(&p->pcf_lock);
3020 
3021 	if (!enough) {
3022 		VM_STAT_ADD(page_reclaim_zero);
3023 		/*
3024 		 * Check again. Its possible that some other thread
3025 		 * could have been right behind us, and added one
3026 		 * to a list somewhere.  Acquire each of the pcf locks
3027 		 * until we find a page.
3028 		 */
3029 		p = pcf;
3030 		for (i = 0; i < PCF_FANOUT; i++) {
3031 			mutex_enter(&p->pcf_lock);
3032 			if (p->pcf_count >= 1) {
3033 				p->pcf_count -= 1;
3034 				enough = 1;
3035 				break;
3036 			}
3037 			p++;
3038 		}
3039 
3040 		if (!enough) {
3041 page_reclaim_nomem:
3042 			/*
3043 			 * We really can't have page `pp'.
3044 			 * Time for the no-memory dance with
3045 			 * page_free().  This is just like
3046 			 * page_create_wait().  Plus the added
3047 			 * attraction of releasing whatever mutex
3048 			 * we held when we were called with in `lock'.
3049 			 * Page_unlock() will wakeup any thread
3050 			 * waiting around for this page.
3051 			 */
3052 			if (lock) {
3053 				VM_STAT_ADD(page_reclaim_zero_locked);
3054 				mutex_exit(lock);
3055 			}
3056 			page_unlock(pp);
3057 
3058 			/*
3059 			 * get this before we drop all the pcf locks.
3060 			 */
3061 			mutex_enter(&new_freemem_lock);
3062 
3063 			p = pcf;
3064 			for (i = 0; i < PCF_FANOUT; i++) {
3065 				p->pcf_wait++;
3066 				mutex_exit(&p->pcf_lock);
3067 				p++;
3068 			}
3069 
3070 			freemem_wait++;
3071 			cv_wait(&freemem_cv, &new_freemem_lock);
3072 			freemem_wait--;
3073 
3074 			mutex_exit(&new_freemem_lock);
3075 
3076 			if (lock) {
3077 				mutex_enter(lock);
3078 			}
3079 			return (0);
3080 		}
3081 
3082 		/*
3083 		 * The pcf accounting has been done,
3084 		 * though none of the pcf_wait flags have been set,
3085 		 * drop the locks and continue on.
3086 		 */
3087 		while (p >= pcf) {
3088 			mutex_exit(&p->pcf_lock);
3089 			p--;
3090 		}
3091 	}
3092 
3093 	/*
3094 	 * freemem is not protected by any lock. Thus, we cannot
3095 	 * have any assertion containing freemem here.
3096 	 */
3097 	freemem -= 1;
3098 
3099 	VM_STAT_ADD(pagecnt.pc_reclaim);
3100 
3101 	/*
3102 	 * page_list_sub will handle the case where pp is a large page.
3103 	 * It's possible that the page was promoted while on the freelist
3104 	 */
3105 	if (PP_ISAGED(pp)) {
3106 		page_list_sub(pp, PG_FREE_LIST);
3107 		TRACE_1(TR_FAC_VM, TR_PAGE_UNFREE_FREE,
3108 		    "page_reclaim_free:pp %p", pp);
3109 	} else {
3110 		page_list_sub(pp, PG_CACHE_LIST);
3111 		TRACE_1(TR_FAC_VM, TR_PAGE_UNFREE_CACHE,
3112 		    "page_reclaim_cache:pp %p", pp);
3113 	}
3114 
3115 	/*
3116 	 * clear the p_free & p_age bits since this page is no longer
3117 	 * on the free list.  Notice that there was a brief time where
3118 	 * a page is marked as free, but is not on the list.
3119 	 *
3120 	 * Set the reference bit to protect against immediate pageout.
3121 	 */
3122 	PP_CLRFREE(pp);
3123 	PP_CLRAGED(pp);
3124 	page_set_props(pp, P_REF);
3125 
3126 	CPU_STATS_ENTER_K();
3127 	cpup = CPU;	/* get cpup now that CPU cannot change */
3128 	CPU_STATS_ADDQ(cpup, vm, pgrec, 1);
3129 	CPU_STATS_ADDQ(cpup, vm, pgfrec, 1);
3130 	CPU_STATS_EXIT_K();
3131 	ASSERT(pp->p_szc == 0);
3132 
3133 	return (1);
3134 }
3135 
3136 /*
3137  * Destroy identity of the page and put it back on
3138  * the page free list.  Assumes that the caller has
3139  * acquired the "exclusive" lock on the page.
3140  */
3141 void
3142 page_destroy(page_t *pp, int dontfree)
3143 {
3144 	ASSERT((PAGE_EXCL(pp) &&
3145 	    !page_iolock_assert(pp)) || panicstr);
3146 	ASSERT(pp->p_slckcnt == 0 || panicstr);
3147 
3148 	if (pp->p_szc != 0) {
3149 		if (pp->p_vnode == NULL || IS_SWAPFSVP(pp->p_vnode) ||
3150 		    PP_ISKAS(pp)) {
3151 			panic("page_destroy: anon or kernel or no vnode "
3152 			    "large page %p", (void *)pp);
3153 		}
3154 		page_demote_vp_pages(pp);
3155 		ASSERT(pp->p_szc == 0);
3156 	}
3157 
3158 	TRACE_1(TR_FAC_VM, TR_PAGE_DESTROY, "page_destroy:pp %p", pp);
3159 
3160 	/*
3161 	 * Unload translations, if any, then hash out the
3162 	 * page to erase its identity.
3163 	 */
3164 	(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
3165 	page_hashout(pp, NULL);
3166 
3167 	if (!dontfree) {
3168 		/*
3169 		 * Acquire the "freemem_lock" for availrmem.
3170 		 * The page_struct_lock need not be acquired for lckcnt
3171 		 * and cowcnt since the page has an "exclusive" lock.
3172 		 */
3173 		if ((pp->p_lckcnt != 0) || (pp->p_cowcnt != 0)) {
3174 			mutex_enter(&freemem_lock);
3175 			if (pp->p_lckcnt != 0) {
3176 				availrmem++;
3177 				pp->p_lckcnt = 0;
3178 			}
3179 			if (pp->p_cowcnt != 0) {
3180 				availrmem += pp->p_cowcnt;
3181 				pp->p_cowcnt = 0;
3182 			}
3183 			mutex_exit(&freemem_lock);
3184 		}
3185 		/*
3186 		 * Put the page on the "free" list.
3187 		 */
3188 		page_free(pp, 0);
3189 	}
3190 }
3191 
3192 void
3193 page_destroy_pages(page_t *pp)
3194 {
3195 
3196 	page_t	*tpp, *rootpp = NULL;
3197 	pgcnt_t	pgcnt = page_get_pagecnt(pp->p_szc);
3198 	pgcnt_t	i, pglcks = 0;
3199 	uint_t	szc = pp->p_szc;
3200 
3201 	ASSERT(pp->p_szc != 0 && pp->p_szc < page_num_pagesizes());
3202 
3203 	VM_STAT_ADD(pagecnt.pc_destroy_pages);
3204 
3205 	TRACE_1(TR_FAC_VM, TR_PAGE_DESTROY, "page_destroy_pages:pp %p", pp);
3206 
3207 	if ((page_pptonum(pp) & (pgcnt - 1)) != 0) {
3208 		panic("page_destroy_pages: not root page %p", (void *)pp);
3209 		/*NOTREACHED*/
3210 	}
3211 
3212 	for (i = 0, tpp = pp; i < pgcnt; i++, tpp++) {
3213 		ASSERT((PAGE_EXCL(tpp) &&
3214 		    !page_iolock_assert(tpp)) || panicstr);
3215 		ASSERT(tpp->p_slckcnt == 0 || panicstr);
3216 		(void) hat_pageunload(tpp, HAT_FORCE_PGUNLOAD);
3217 		page_hashout(tpp, NULL);
3218 		ASSERT(tpp->p_offset == (u_offset_t)-1);
3219 		if (tpp->p_lckcnt != 0) {
3220 			pglcks++;
3221 			tpp->p_lckcnt = 0;
3222 		} else if (tpp->p_cowcnt != 0) {
3223 			pglcks += tpp->p_cowcnt;
3224 			tpp->p_cowcnt = 0;
3225 		}
3226 		ASSERT(!hat_page_getshare(tpp));
3227 		ASSERT(tpp->p_vnode == NULL);
3228 		ASSERT(tpp->p_szc == szc);
3229 
3230 		PP_SETFREE(tpp);
3231 		page_clr_all_props(tpp);
3232 		PP_SETAGED(tpp);
3233 		ASSERT(tpp->p_next == tpp);
3234 		ASSERT(tpp->p_prev == tpp);
3235 		page_list_concat(&rootpp, &tpp);
3236 	}
3237 
3238 	ASSERT(rootpp == pp);
3239 	if (pglcks != 0) {
3240 		mutex_enter(&freemem_lock);
3241 		availrmem += pglcks;
3242 		mutex_exit(&freemem_lock);
3243 	}
3244 
3245 	page_list_add_pages(rootpp, 0);
3246 	page_create_putback(pgcnt);
3247 }
3248 
3249 /*
3250  * Similar to page_destroy(), but destroys pages which are
3251  * locked and known to be on the page free list.  Since
3252  * the page is known to be free and locked, no one can access
3253  * it.
3254  *
3255  * Also, the number of free pages does not change.
3256  */
3257 void
3258 page_destroy_free(page_t *pp)
3259 {
3260 	ASSERT(PAGE_EXCL(pp));
3261 	ASSERT(PP_ISFREE(pp));
3262 	ASSERT(pp->p_vnode);
3263 	ASSERT(hat_page_getattr(pp, P_MOD | P_REF | P_RO) == 0);
3264 	ASSERT(!hat_page_is_mapped(pp));
3265 	ASSERT(PP_ISAGED(pp) == 0);
3266 	ASSERT(pp->p_szc == 0);
3267 
3268 	VM_STAT_ADD(pagecnt.pc_destroy_free);
3269 	page_list_sub(pp, PG_CACHE_LIST);
3270 
3271 	page_hashout(pp, NULL);
3272 	ASSERT(pp->p_vnode == NULL);
3273 	ASSERT(pp->p_offset == (u_offset_t)-1);
3274 	ASSERT(pp->p_hash == NULL);
3275 
3276 	PP_SETAGED(pp);
3277 	page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL);
3278 	page_unlock(pp);
3279 
3280 	mutex_enter(&new_freemem_lock);
3281 	if (freemem_wait) {
3282 		cv_signal(&freemem_cv);
3283 	}
3284 	mutex_exit(&new_freemem_lock);
3285 }
3286 
3287 /*
3288  * Rename the page "opp" to have an identity specified
3289  * by [vp, off].  If a page already exists with this name
3290  * it is locked and destroyed.  Note that the page's
3291  * translations are not unloaded during the rename.
3292  *
3293  * This routine is used by the anon layer to "steal" the
3294  * original page and is not unlike destroying a page and
3295  * creating a new page using the same page frame.
3296  *
3297  * XXX -- Could deadlock if caller 1 tries to rename A to B while
3298  * caller 2 tries to rename B to A.
3299  */
3300 void
3301 page_rename(page_t *opp, vnode_t *vp, u_offset_t off)
3302 {
3303 	page_t		*pp;
3304 	int		olckcnt = 0;
3305 	int		ocowcnt = 0;
3306 	kmutex_t	*phm;
3307 	ulong_t		index;
3308 
3309 	ASSERT(PAGE_EXCL(opp) && !page_iolock_assert(opp));
3310 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
3311 	ASSERT(PP_ISFREE(opp) == 0);
3312 
3313 	VM_STAT_ADD(page_rename_count);
3314 
3315 	TRACE_3(TR_FAC_VM, TR_PAGE_RENAME,
3316 	    "page rename:pp %p vp %p off %llx", opp, vp, off);
3317 
3318 	/*
3319 	 * CacheFS may call page_rename for a large NFS page
3320 	 * when both CacheFS and NFS mount points are used
3321 	 * by applications. Demote this large page before
3322 	 * renaming it, to ensure that there are no "partial"
3323 	 * large pages left lying around.
3324 	 */
3325 	if (opp->p_szc != 0) {
3326 		vnode_t *ovp = opp->p_vnode;
3327 		ASSERT(ovp != NULL);
3328 		ASSERT(!IS_SWAPFSVP(ovp));
3329 		ASSERT(!VN_ISKAS(ovp));
3330 		page_demote_vp_pages(opp);
3331 		ASSERT(opp->p_szc == 0);
3332 	}
3333 
3334 	page_hashout(opp, NULL);
3335 	PP_CLRAGED(opp);
3336 
3337 	/*
3338 	 * Acquire the appropriate page hash lock, since
3339 	 * we're going to rename the page.
3340 	 */
3341 	index = PAGE_HASH_FUNC(vp, off);
3342 	phm = PAGE_HASH_MUTEX(index);
3343 	mutex_enter(phm);
3344 top:
3345 	/*
3346 	 * Look for an existing page with this name and destroy it if found.
3347 	 * By holding the page hash lock all the way to the page_hashin()
3348 	 * call, we are assured that no page can be created with this
3349 	 * identity.  In the case when the phm lock is dropped to undo any
3350 	 * hat layer mappings, the existing page is held with an "exclusive"
3351 	 * lock, again preventing another page from being created with
3352 	 * this identity.
3353 	 */
3354 	PAGE_HASH_SEARCH(index, pp, vp, off);
3355 	if (pp != NULL) {
3356 		VM_STAT_ADD(page_rename_exists);
3357 
3358 		/*
3359 		 * As it turns out, this is one of only two places where
3360 		 * page_lock() needs to hold the passed in lock in the
3361 		 * successful case.  In all of the others, the lock could
3362 		 * be dropped as soon as the attempt is made to lock
3363 		 * the page.  It is tempting to add yet another arguement,
3364 		 * PL_KEEP or PL_DROP, to let page_lock know what to do.
3365 		 */
3366 		if (!page_lock(pp, SE_EXCL, phm, P_RECLAIM)) {
3367 			/*
3368 			 * Went to sleep because the page could not
3369 			 * be locked.  We were woken up when the page
3370 			 * was unlocked, or when the page was destroyed.
3371 			 * In either case, `phm' was dropped while we
3372 			 * slept.  Hence we should not just roar through
3373 			 * this loop.
3374 			 */
3375 			goto top;
3376 		}
3377 
3378 		/*
3379 		 * If an existing page is a large page, then demote
3380 		 * it to ensure that no "partial" large pages are
3381 		 * "created" after page_rename. An existing page
3382 		 * can be a CacheFS page, and can't belong to swapfs.
3383 		 */
3384 		if (hat_page_is_mapped(pp)) {
3385 			/*
3386 			 * Unload translations.  Since we hold the
3387 			 * exclusive lock on this page, the page
3388 			 * can not be changed while we drop phm.
3389 			 * This is also not a lock protocol violation,
3390 			 * but rather the proper way to do things.
3391 			 */
3392 			mutex_exit(phm);
3393 			(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
3394 			if (pp->p_szc != 0) {
3395 				ASSERT(!IS_SWAPFSVP(vp));
3396 				ASSERT(!VN_ISKAS(vp));
3397 				page_demote_vp_pages(pp);
3398 				ASSERT(pp->p_szc == 0);
3399 			}
3400 			mutex_enter(phm);
3401 		} else if (pp->p_szc != 0) {
3402 			ASSERT(!IS_SWAPFSVP(vp));
3403 			ASSERT(!VN_ISKAS(vp));
3404 			mutex_exit(phm);
3405 			page_demote_vp_pages(pp);
3406 			ASSERT(pp->p_szc == 0);
3407 			mutex_enter(phm);
3408 		}
3409 		page_hashout(pp, phm);
3410 	}
3411 	/*
3412 	 * Hash in the page with the new identity.
3413 	 */
3414 	if (!page_hashin(opp, vp, off, phm)) {
3415 		/*
3416 		 * We were holding phm while we searched for [vp, off]
3417 		 * and only dropped phm if we found and locked a page.
3418 		 * If we can't create this page now, then some thing
3419 		 * is really broken.
3420 		 */
3421 		panic("page_rename: Can't hash in page: %p", (void *)pp);
3422 		/*NOTREACHED*/
3423 	}
3424 
3425 	ASSERT(MUTEX_HELD(phm));
3426 	mutex_exit(phm);
3427 
3428 	/*
3429 	 * Now that we have dropped phm, lets get around to finishing up
3430 	 * with pp.
3431 	 */
3432 	if (pp != NULL) {
3433 		ASSERT(!hat_page_is_mapped(pp));
3434 		/* for now large pages should not end up here */
3435 		ASSERT(pp->p_szc == 0);
3436 		/*
3437 		 * Save the locks for transfer to the new page and then
3438 		 * clear them so page_free doesn't think they're important.
3439 		 * The page_struct_lock need not be acquired for lckcnt and
3440 		 * cowcnt since the page has an "exclusive" lock.
3441 		 */
3442 		olckcnt = pp->p_lckcnt;
3443 		ocowcnt = pp->p_cowcnt;
3444 		pp->p_lckcnt = pp->p_cowcnt = 0;
3445 
3446 		/*
3447 		 * Put the page on the "free" list after we drop
3448 		 * the lock.  The less work under the lock the better.
3449 		 */
3450 		/*LINTED: constant in conditional context*/
3451 		VN_DISPOSE(pp, B_FREE, 0, kcred);
3452 	}
3453 
3454 	/*
3455 	 * Transfer the lock count from the old page (if any).
3456 	 * The page_struct_lock need not be acquired for lckcnt and
3457 	 * cowcnt since the page has an "exclusive" lock.
3458 	 */
3459 	opp->p_lckcnt += olckcnt;
3460 	opp->p_cowcnt += ocowcnt;
3461 }
3462 
3463 /*
3464  * low level routine to add page `pp' to the hash and vp chains for [vp, offset]
3465  *
3466  * Pages are normally inserted at the start of a vnode's v_pages list.
3467  * If the vnode is VMODSORT and the page is modified, it goes at the end.
3468  * This can happen when a modified page is relocated for DR.
3469  *
3470  * Returns 1 on success and 0 on failure.
3471  */
3472 static int
3473 page_do_hashin(page_t *pp, vnode_t *vp, u_offset_t offset)
3474 {
3475 	page_t		**listp;
3476 	page_t		*tp;
3477 	ulong_t		index;
3478 
3479 	ASSERT(PAGE_EXCL(pp));
3480 	ASSERT(vp != NULL);
3481 	ASSERT(MUTEX_HELD(page_vnode_mutex(vp)));
3482 
3483 	/*
3484 	 * Be sure to set these up before the page is inserted on the hash
3485 	 * list.  As soon as the page is placed on the list some other
3486 	 * thread might get confused and wonder how this page could
3487 	 * possibly hash to this list.
3488 	 */
3489 	pp->p_vnode = vp;
3490 	pp->p_offset = offset;
3491 
3492 	/*
3493 	 * record if this page is on a swap vnode
3494 	 */
3495 	if ((vp->v_flag & VISSWAP) != 0)
3496 		PP_SETSWAP(pp);
3497 
3498 	index = PAGE_HASH_FUNC(vp, offset);
3499 	ASSERT(MUTEX_HELD(PAGE_HASH_MUTEX(index)));
3500 	listp = &page_hash[index];
3501 
3502 	/*
3503 	 * If this page is already hashed in, fail this attempt to add it.
3504 	 */
3505 	for (tp = *listp; tp != NULL; tp = tp->p_hash) {
3506 		if (tp->p_vnode == vp && tp->p_offset == offset) {
3507 			pp->p_vnode = NULL;
3508 			pp->p_offset = (u_offset_t)(-1);
3509 			return (0);
3510 		}
3511 	}
3512 	pp->p_hash = *listp;
3513 	*listp = pp;
3514 
3515 	/*
3516 	 * Add the page to the vnode's list of pages
3517 	 */
3518 	if (vp->v_pages != NULL && IS_VMODSORT(vp) && hat_ismod(pp))
3519 		listp = &vp->v_pages->p_vpprev->p_vpnext;
3520 	else
3521 		listp = &vp->v_pages;
3522 
3523 	page_vpadd(listp, pp);
3524 
3525 	return (1);
3526 }
3527 
3528 /*
3529  * Add page `pp' to both the hash and vp chains for [vp, offset].
3530  *
3531  * Returns 1 on success and 0 on failure.
3532  * If hold is passed in, it is not dropped.
3533  */
3534 int
3535 page_hashin(page_t *pp, vnode_t *vp, u_offset_t offset, kmutex_t *hold)
3536 {
3537 	kmutex_t	*phm = NULL;
3538 	kmutex_t	*vphm;
3539 	int		rc;
3540 
3541 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
3542 
3543 	TRACE_3(TR_FAC_VM, TR_PAGE_HASHIN,
3544 	    "page_hashin:pp %p vp %p offset %llx",
3545 	    pp, vp, offset);
3546 
3547 	VM_STAT_ADD(hashin_count);
3548 
3549 	if (hold != NULL)
3550 		phm = hold;
3551 	else {
3552 		VM_STAT_ADD(hashin_not_held);
3553 		phm = PAGE_HASH_MUTEX(PAGE_HASH_FUNC(vp, offset));
3554 		mutex_enter(phm);
3555 	}
3556 
3557 	vphm = page_vnode_mutex(vp);
3558 	mutex_enter(vphm);
3559 	rc = page_do_hashin(pp, vp, offset);
3560 	mutex_exit(vphm);
3561 	if (hold == NULL)
3562 		mutex_exit(phm);
3563 	if (rc == 0)
3564 		VM_STAT_ADD(hashin_already);
3565 	return (rc);
3566 }
3567 
3568 /*
3569  * Remove page ``pp'' from the hash and vp chains and remove vp association.
3570  * All mutexes must be held
3571  */
3572 static void
3573 page_do_hashout(page_t *pp)
3574 {
3575 	page_t	**hpp;
3576 	page_t	*hp;
3577 	vnode_t	*vp = pp->p_vnode;
3578 
3579 	ASSERT(vp != NULL);
3580 	ASSERT(MUTEX_HELD(page_vnode_mutex(vp)));
3581 
3582 	/*
3583 	 * First, take pp off of its hash chain.
3584 	 */
3585 	hpp = &page_hash[PAGE_HASH_FUNC(vp, pp->p_offset)];
3586 
3587 	for (;;) {
3588 		hp = *hpp;
3589 		if (hp == pp)
3590 			break;
3591 		if (hp == NULL) {
3592 			panic("page_do_hashout");
3593 			/*NOTREACHED*/
3594 		}
3595 		hpp = &hp->p_hash;
3596 	}
3597 	*hpp = pp->p_hash;
3598 
3599 	/*
3600 	 * Now remove it from its associated vnode.
3601 	 */
3602 	if (vp->v_pages)
3603 		page_vpsub(&vp->v_pages, pp);
3604 
3605 	pp->p_hash = NULL;
3606 	page_clr_all_props(pp);
3607 	PP_CLRSWAP(pp);
3608 	pp->p_vnode = NULL;
3609 	pp->p_offset = (u_offset_t)-1;
3610 }
3611 
3612 /*
3613  * Remove page ``pp'' from the hash and vp chains and remove vp association.
3614  *
3615  * When `phm' is non-NULL it contains the address of the mutex protecting the
3616  * hash list pp is on.  It is not dropped.
3617  */
3618 void
3619 page_hashout(page_t *pp, kmutex_t *phm)
3620 {
3621 	vnode_t		*vp;
3622 	ulong_t		index;
3623 	kmutex_t	*nphm;
3624 	kmutex_t	*vphm;
3625 	kmutex_t	*sep;
3626 
3627 	ASSERT(phm != NULL ? MUTEX_HELD(phm) : 1);
3628 	ASSERT(pp->p_vnode != NULL);
3629 	ASSERT((PAGE_EXCL(pp) && !page_iolock_assert(pp)) || panicstr);
3630 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(pp->p_vnode)));
3631 
3632 	vp = pp->p_vnode;
3633 
3634 	TRACE_2(TR_FAC_VM, TR_PAGE_HASHOUT,
3635 	    "page_hashout:pp %p vp %p", pp, vp);
3636 
3637 	/* Kernel probe */
3638 	TNF_PROBE_2(page_unmap, "vm pagefault", /* CSTYLED */,
3639 	    tnf_opaque, vnode, vp,
3640 	    tnf_offset, offset, pp->p_offset);
3641 
3642 	/*
3643 	 *
3644 	 */
3645 	VM_STAT_ADD(hashout_count);
3646 	index = PAGE_HASH_FUNC(vp, pp->p_offset);
3647 	if (phm == NULL) {
3648 		VM_STAT_ADD(hashout_not_held);
3649 		nphm = PAGE_HASH_MUTEX(index);
3650 		mutex_enter(nphm);
3651 	}
3652 	ASSERT(phm ? phm == PAGE_HASH_MUTEX(index) : 1);
3653 
3654 
3655 	/*
3656 	 * grab page vnode mutex and remove it...
3657 	 */
3658 	vphm = page_vnode_mutex(vp);
3659 	mutex_enter(vphm);
3660 
3661 	page_do_hashout(pp);
3662 
3663 	mutex_exit(vphm);
3664 	if (phm == NULL)
3665 		mutex_exit(nphm);
3666 
3667 	/*
3668 	 * Wake up processes waiting for this page.  The page's
3669 	 * identity has been changed, and is probably not the
3670 	 * desired page any longer.
3671 	 */
3672 	sep = page_se_mutex(pp);
3673 	mutex_enter(sep);
3674 	pp->p_selock &= ~SE_EWANTED;
3675 	if (CV_HAS_WAITERS(&pp->p_cv))
3676 		cv_broadcast(&pp->p_cv);
3677 	mutex_exit(sep);
3678 }
3679 
3680 /*
3681  * Add the page to the front of a linked list of pages
3682  * using the p_next & p_prev pointers for the list.
3683  * The caller is responsible for protecting the list pointers.
3684  */
3685 void
3686 page_add(page_t **ppp, page_t *pp)
3687 {
3688 	ASSERT(PAGE_EXCL(pp) || (PAGE_SHARED(pp) && page_iolock_assert(pp)));
3689 
3690 	page_add_common(ppp, pp);
3691 }
3692 
3693 
3694 
3695 /*
3696  *  Common code for page_add() and mach_page_add()
3697  */
3698 void
3699 page_add_common(page_t **ppp, page_t *pp)
3700 {
3701 	if (*ppp == NULL) {
3702 		pp->p_next = pp->p_prev = pp;
3703 	} else {
3704 		pp->p_next = *ppp;
3705 		pp->p_prev = (*ppp)->p_prev;
3706 		(*ppp)->p_prev = pp;
3707 		pp->p_prev->p_next = pp;
3708 	}
3709 	*ppp = pp;
3710 }
3711 
3712 
3713 /*
3714  * Remove this page from a linked list of pages
3715  * using the p_next & p_prev pointers for the list.
3716  *
3717  * The caller is responsible for protecting the list pointers.
3718  */
3719 void
3720 page_sub(page_t **ppp, page_t *pp)
3721 {
3722 	ASSERT((PP_ISFREE(pp)) ? 1 :
3723 	    (PAGE_EXCL(pp)) || (PAGE_SHARED(pp) && page_iolock_assert(pp)));
3724 
3725 	if (*ppp == NULL || pp == NULL) {
3726 		panic("page_sub: bad arg(s): pp %p, *ppp %p",
3727 		    (void *)pp, (void *)(*ppp));
3728 		/*NOTREACHED*/
3729 	}
3730 
3731 	page_sub_common(ppp, pp);
3732 }
3733 
3734 
3735 /*
3736  *  Common code for page_sub() and mach_page_sub()
3737  */
3738 void
3739 page_sub_common(page_t **ppp, page_t *pp)
3740 {
3741 	if (*ppp == pp)
3742 		*ppp = pp->p_next;		/* go to next page */
3743 
3744 	if (*ppp == pp)
3745 		*ppp = NULL;			/* page list is gone */
3746 	else {
3747 		pp->p_prev->p_next = pp->p_next;
3748 		pp->p_next->p_prev = pp->p_prev;
3749 	}
3750 	pp->p_prev = pp->p_next = pp;		/* make pp a list of one */
3751 }
3752 
3753 
3754 /*
3755  * Break page list cppp into two lists with npages in the first list.
3756  * The tail is returned in nppp.
3757  */
3758 void
3759 page_list_break(page_t **oppp, page_t **nppp, pgcnt_t npages)
3760 {
3761 	page_t *s1pp = *oppp;
3762 	page_t *s2pp;
3763 	page_t *e1pp, *e2pp;
3764 	long n = 0;
3765 
3766 	if (s1pp == NULL) {
3767 		*nppp = NULL;
3768 		return;
3769 	}
3770 	if (npages == 0) {
3771 		*nppp = s1pp;
3772 		*oppp = NULL;
3773 		return;
3774 	}
3775 	for (n = 0, s2pp = *oppp; n < npages; n++) {
3776 		s2pp = s2pp->p_next;
3777 	}
3778 	/* Fix head and tail of new lists */
3779 	e1pp = s2pp->p_prev;
3780 	e2pp = s1pp->p_prev;
3781 	s1pp->p_prev = e1pp;
3782 	e1pp->p_next = s1pp;
3783 	s2pp->p_prev = e2pp;
3784 	e2pp->p_next = s2pp;
3785 
3786 	/* second list empty */
3787 	if (s2pp == s1pp) {
3788 		*oppp = s1pp;
3789 		*nppp = NULL;
3790 	} else {
3791 		*oppp = s1pp;
3792 		*nppp = s2pp;
3793 	}
3794 }
3795 
3796 /*
3797  * Concatenate page list nppp onto the end of list ppp.
3798  */
3799 void
3800 page_list_concat(page_t **ppp, page_t **nppp)
3801 {
3802 	page_t *s1pp, *s2pp, *e1pp, *e2pp;
3803 
3804 	if (*nppp == NULL) {
3805 		return;
3806 	}
3807 	if (*ppp == NULL) {
3808 		*ppp = *nppp;
3809 		return;
3810 	}
3811 	s1pp = *ppp;
3812 	e1pp =  s1pp->p_prev;
3813 	s2pp = *nppp;
3814 	e2pp = s2pp->p_prev;
3815 	s1pp->p_prev = e2pp;
3816 	e2pp->p_next = s1pp;
3817 	e1pp->p_next = s2pp;
3818 	s2pp->p_prev = e1pp;
3819 }
3820 
3821 /*
3822  * return the next page in the page list
3823  */
3824 page_t *
3825 page_list_next(page_t *pp)
3826 {
3827 	return (pp->p_next);
3828 }
3829 
3830 
3831 /*
3832  * Add the page to the front of the linked list of pages
3833  * using p_vpnext/p_vpprev pointers for the list.
3834  *
3835  * The caller is responsible for protecting the lists.
3836  */
3837 void
3838 page_vpadd(page_t **ppp, page_t *pp)
3839 {
3840 	if (*ppp == NULL) {
3841 		pp->p_vpnext = pp->p_vpprev = pp;
3842 	} else {
3843 		pp->p_vpnext = *ppp;
3844 		pp->p_vpprev = (*ppp)->p_vpprev;
3845 		(*ppp)->p_vpprev = pp;
3846 		pp->p_vpprev->p_vpnext = pp;
3847 	}
3848 	*ppp = pp;
3849 }
3850 
3851 /*
3852  * Remove this page from the linked list of pages
3853  * using p_vpnext/p_vpprev pointers for the list.
3854  *
3855  * The caller is responsible for protecting the lists.
3856  */
3857 void
3858 page_vpsub(page_t **ppp, page_t *pp)
3859 {
3860 	if (*ppp == NULL || pp == NULL) {
3861 		panic("page_vpsub: bad arg(s): pp %p, *ppp %p",
3862 		    (void *)pp, (void *)(*ppp));
3863 		/*NOTREACHED*/
3864 	}
3865 
3866 	if (*ppp == pp)
3867 		*ppp = pp->p_vpnext;		/* go to next page */
3868 
3869 	if (*ppp == pp)
3870 		*ppp = NULL;			/* page list is gone */
3871 	else {
3872 		pp->p_vpprev->p_vpnext = pp->p_vpnext;
3873 		pp->p_vpnext->p_vpprev = pp->p_vpprev;
3874 	}
3875 	pp->p_vpprev = pp->p_vpnext = pp;	/* make pp a list of one */
3876 }
3877 
3878 /*
3879  * Lock a physical page into memory "long term".  Used to support "lock
3880  * in memory" functions.  Accepts the page to be locked, and a cow variable
3881  * to indicate whether a the lock will travel to the new page during
3882  * a potential copy-on-write.
3883  */
3884 int
3885 page_pp_lock(
3886 	page_t *pp,			/* page to be locked */
3887 	int cow,			/* cow lock */
3888 	int kernel)			/* must succeed -- ignore checking */
3889 {
3890 	int r = 0;			/* result -- assume failure */
3891 
3892 	ASSERT(PAGE_LOCKED(pp));
3893 
3894 	page_struct_lock(pp);
3895 	/*
3896 	 * Acquire the "freemem_lock" for availrmem.
3897 	 */
3898 	if (cow) {
3899 		mutex_enter(&freemem_lock);
3900 		if ((availrmem > pages_pp_maximum) &&
3901 		    (pp->p_cowcnt < (ushort_t)PAGE_LOCK_MAXIMUM)) {
3902 			availrmem--;
3903 			pages_locked++;
3904 			mutex_exit(&freemem_lock);
3905 			r = 1;
3906 			if (++pp->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
3907 				cmn_err(CE_WARN,
3908 				    "COW lock limit reached on pfn 0x%lx",
3909 				    page_pptonum(pp));
3910 			}
3911 		} else
3912 			mutex_exit(&freemem_lock);
3913 	} else {
3914 		if (pp->p_lckcnt) {
3915 			if (pp->p_lckcnt < (ushort_t)PAGE_LOCK_MAXIMUM) {
3916 				r = 1;
3917 				if (++pp->p_lckcnt ==
3918 				    (ushort_t)PAGE_LOCK_MAXIMUM) {
3919 					cmn_err(CE_WARN, "Page lock limit "
3920 					    "reached on pfn 0x%lx",
3921 					    page_pptonum(pp));
3922 				}
3923 			}
3924 		} else {
3925 			if (kernel) {
3926 				/* availrmem accounting done by caller */
3927 				++pp->p_lckcnt;
3928 				r = 1;
3929 			} else {
3930 				mutex_enter(&freemem_lock);
3931 				if (availrmem > pages_pp_maximum) {
3932 					availrmem--;
3933 					pages_locked++;
3934 					++pp->p_lckcnt;
3935 					r = 1;
3936 				}
3937 				mutex_exit(&freemem_lock);
3938 			}
3939 		}
3940 	}
3941 	page_struct_unlock(pp);
3942 	return (r);
3943 }
3944 
3945 /*
3946  * Decommit a lock on a physical page frame.  Account for cow locks if
3947  * appropriate.
3948  */
3949 void
3950 page_pp_unlock(
3951 	page_t *pp,			/* page to be unlocked */
3952 	int cow,			/* expect cow lock */
3953 	int kernel)			/* this was a kernel lock */
3954 {
3955 	ASSERT(PAGE_LOCKED(pp));
3956 
3957 	page_struct_lock(pp);
3958 	/*
3959 	 * Acquire the "freemem_lock" for availrmem.
3960 	 * If cowcnt or lcknt is already 0 do nothing; i.e., we
3961 	 * could be called to unlock even if nothing is locked. This could
3962 	 * happen if locked file pages were truncated (removing the lock)
3963 	 * and the file was grown again and new pages faulted in; the new
3964 	 * pages are unlocked but the segment still thinks they're locked.
3965 	 */
3966 	if (cow) {
3967 		if (pp->p_cowcnt) {
3968 			mutex_enter(&freemem_lock);
3969 			pp->p_cowcnt--;
3970 			availrmem++;
3971 			pages_locked--;
3972 			mutex_exit(&freemem_lock);
3973 		}
3974 	} else {
3975 		if (pp->p_lckcnt && --pp->p_lckcnt == 0) {
3976 			if (!kernel) {
3977 				mutex_enter(&freemem_lock);
3978 				availrmem++;
3979 				pages_locked--;
3980 				mutex_exit(&freemem_lock);
3981 			}
3982 		}
3983 	}
3984 	page_struct_unlock(pp);
3985 }
3986 
3987 /*
3988  * This routine reserves availrmem for npages;
3989  * 	flags: KM_NOSLEEP or KM_SLEEP
3990  * 	returns 1 on success or 0 on failure
3991  */
3992 int
3993 page_resv(pgcnt_t npages, uint_t flags)
3994 {
3995 	mutex_enter(&freemem_lock);
3996 	while (availrmem < tune.t_minarmem + npages) {
3997 		if (flags & KM_NOSLEEP) {
3998 			mutex_exit(&freemem_lock);
3999 			return (0);
4000 		}
4001 		mutex_exit(&freemem_lock);
4002 		page_needfree(npages);
4003 		kmem_reap();
4004 		delay(hz >> 2);
4005 		page_needfree(-(spgcnt_t)npages);
4006 		mutex_enter(&freemem_lock);
4007 	}
4008 	availrmem -= npages;
4009 	mutex_exit(&freemem_lock);
4010 	return (1);
4011 }
4012 
4013 /*
4014  * This routine unreserves availrmem for npages;
4015  */
4016 void
4017 page_unresv(pgcnt_t npages)
4018 {
4019 	mutex_enter(&freemem_lock);
4020 	availrmem += npages;
4021 	mutex_exit(&freemem_lock);
4022 }
4023 
4024 /*
4025  * See Statement at the beginning of segvn_lockop() regarding
4026  * the way we handle cowcnts and lckcnts.
4027  *
4028  * Transfer cowcnt on 'opp' to cowcnt on 'npp' if the vpage
4029  * that breaks COW has PROT_WRITE.
4030  *
4031  * Note that, we may also break COW in case we are softlocking
4032  * on read access during physio;
4033  * in this softlock case, the vpage may not have PROT_WRITE.
4034  * So, we need to transfer lckcnt on 'opp' to lckcnt on 'npp'
4035  * if the vpage doesn't have PROT_WRITE.
4036  *
4037  * This routine is never called if we are stealing a page
4038  * in anon_private.
4039  *
4040  * The caller subtracted from availrmem for read only mapping.
4041  * if lckcnt is 1 increment availrmem.
4042  */
4043 void
4044 page_pp_useclaim(
4045 	page_t *opp,		/* original page frame losing lock */
4046 	page_t *npp,		/* new page frame gaining lock */
4047 	uint_t	write_perm) 	/* set if vpage has PROT_WRITE */
4048 {
4049 	int payback = 0;
4050 
4051 	ASSERT(PAGE_LOCKED(opp));
4052 	ASSERT(PAGE_LOCKED(npp));
4053 
4054 	page_struct_lock(opp);
4055 
4056 	ASSERT(npp->p_cowcnt == 0);
4057 	ASSERT(npp->p_lckcnt == 0);
4058 
4059 	/* Don't use claim if nothing is locked (see page_pp_unlock above) */
4060 	if ((write_perm && opp->p_cowcnt != 0) ||
4061 	    (!write_perm && opp->p_lckcnt != 0)) {
4062 
4063 		if (write_perm) {
4064 			npp->p_cowcnt++;
4065 			ASSERT(opp->p_cowcnt != 0);
4066 			opp->p_cowcnt--;
4067 		} else {
4068 
4069 			ASSERT(opp->p_lckcnt != 0);
4070 
4071 			/*
4072 			 * We didn't need availrmem decremented if p_lckcnt on
4073 			 * original page is 1. Here, we are unlocking
4074 			 * read-only copy belonging to original page and
4075 			 * are locking a copy belonging to new page.
4076 			 */
4077 			if (opp->p_lckcnt == 1)
4078 				payback = 1;
4079 
4080 			npp->p_lckcnt++;
4081 			opp->p_lckcnt--;
4082 		}
4083 	}
4084 	if (payback) {
4085 		mutex_enter(&freemem_lock);
4086 		availrmem++;
4087 		pages_useclaim--;
4088 		mutex_exit(&freemem_lock);
4089 	}
4090 	page_struct_unlock(opp);
4091 }
4092 
4093 /*
4094  * Simple claim adjust functions -- used to support changes in
4095  * claims due to changes in access permissions.  Used by segvn_setprot().
4096  */
4097 int
4098 page_addclaim(page_t *pp)
4099 {
4100 	int r = 0;			/* result */
4101 
4102 	ASSERT(PAGE_LOCKED(pp));
4103 
4104 	page_struct_lock(pp);
4105 	ASSERT(pp->p_lckcnt != 0);
4106 
4107 	if (pp->p_lckcnt == 1) {
4108 		if (pp->p_cowcnt < (ushort_t)PAGE_LOCK_MAXIMUM) {
4109 			--pp->p_lckcnt;
4110 			r = 1;
4111 			if (++pp->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
4112 				cmn_err(CE_WARN,
4113 				    "COW lock limit reached on pfn 0x%lx",
4114 				    page_pptonum(pp));
4115 			}
4116 		}
4117 	} else {
4118 		mutex_enter(&freemem_lock);
4119 		if ((availrmem > pages_pp_maximum) &&
4120 		    (pp->p_cowcnt < (ushort_t)PAGE_LOCK_MAXIMUM)) {
4121 			--availrmem;
4122 			++pages_claimed;
4123 			mutex_exit(&freemem_lock);
4124 			--pp->p_lckcnt;
4125 			r = 1;
4126 			if (++pp->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
4127 				cmn_err(CE_WARN,
4128 				    "COW lock limit reached on pfn 0x%lx",
4129 				    page_pptonum(pp));
4130 			}
4131 		} else
4132 			mutex_exit(&freemem_lock);
4133 	}
4134 	page_struct_unlock(pp);
4135 	return (r);
4136 }
4137 
4138 int
4139 page_subclaim(page_t *pp)
4140 {
4141 	int r = 0;
4142 
4143 	ASSERT(PAGE_LOCKED(pp));
4144 
4145 	page_struct_lock(pp);
4146 	ASSERT(pp->p_cowcnt != 0);
4147 
4148 	if (pp->p_lckcnt) {
4149 		if (pp->p_lckcnt < (ushort_t)PAGE_LOCK_MAXIMUM) {
4150 			r = 1;
4151 			/*
4152 			 * for availrmem
4153 			 */
4154 			mutex_enter(&freemem_lock);
4155 			availrmem++;
4156 			pages_claimed--;
4157 			mutex_exit(&freemem_lock);
4158 
4159 			pp->p_cowcnt--;
4160 
4161 			if (++pp->p_lckcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
4162 				cmn_err(CE_WARN,
4163 				    "Page lock limit reached on pfn 0x%lx",
4164 				    page_pptonum(pp));
4165 			}
4166 		}
4167 	} else {
4168 		r = 1;
4169 		pp->p_cowcnt--;
4170 		pp->p_lckcnt++;
4171 	}
4172 	page_struct_unlock(pp);
4173 	return (r);
4174 }
4175 
4176 int
4177 page_addclaim_pages(page_t  **ppa)
4178 {
4179 
4180 	pgcnt_t	lckpgs = 0, pg_idx;
4181 
4182 	VM_STAT_ADD(pagecnt.pc_addclaim_pages);
4183 
4184 	mutex_enter(&page_llock);
4185 	for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) {
4186 
4187 		ASSERT(PAGE_LOCKED(ppa[pg_idx]));
4188 		ASSERT(ppa[pg_idx]->p_lckcnt != 0);
4189 		if (ppa[pg_idx]->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
4190 			mutex_exit(&page_llock);
4191 			return (0);
4192 		}
4193 		if (ppa[pg_idx]->p_lckcnt > 1)
4194 			lckpgs++;
4195 	}
4196 
4197 	if (lckpgs != 0) {
4198 		mutex_enter(&freemem_lock);
4199 		if (availrmem >= pages_pp_maximum + lckpgs) {
4200 			availrmem -= lckpgs;
4201 			pages_claimed += lckpgs;
4202 		} else {
4203 			mutex_exit(&freemem_lock);
4204 			mutex_exit(&page_llock);
4205 			return (0);
4206 		}
4207 		mutex_exit(&freemem_lock);
4208 	}
4209 
4210 	for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) {
4211 		ppa[pg_idx]->p_lckcnt--;
4212 		ppa[pg_idx]->p_cowcnt++;
4213 	}
4214 	mutex_exit(&page_llock);
4215 	return (1);
4216 }
4217 
4218 int
4219 page_subclaim_pages(page_t  **ppa)
4220 {
4221 	pgcnt_t	ulckpgs = 0, pg_idx;
4222 
4223 	VM_STAT_ADD(pagecnt.pc_subclaim_pages);
4224 
4225 	mutex_enter(&page_llock);
4226 	for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) {
4227 
4228 		ASSERT(PAGE_LOCKED(ppa[pg_idx]));
4229 		ASSERT(ppa[pg_idx]->p_cowcnt != 0);
4230 		if (ppa[pg_idx]->p_lckcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
4231 			mutex_exit(&page_llock);
4232 			return (0);
4233 		}
4234 		if (ppa[pg_idx]->p_lckcnt != 0)
4235 			ulckpgs++;
4236 	}
4237 
4238 	if (ulckpgs != 0) {
4239 		mutex_enter(&freemem_lock);
4240 		availrmem += ulckpgs;
4241 		pages_claimed -= ulckpgs;
4242 		mutex_exit(&freemem_lock);
4243 	}
4244 
4245 	for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) {
4246 		ppa[pg_idx]->p_cowcnt--;
4247 		ppa[pg_idx]->p_lckcnt++;
4248 
4249 	}
4250 	mutex_exit(&page_llock);
4251 	return (1);
4252 }
4253 
4254 page_t *
4255 page_numtopp(pfn_t pfnum, se_t se)
4256 {
4257 	page_t *pp;
4258 
4259 retry:
4260 	pp = page_numtopp_nolock(pfnum);
4261 	if (pp == NULL) {
4262 		return ((page_t *)NULL);
4263 	}
4264 
4265 	/*
4266 	 * Acquire the appropriate lock on the page.
4267 	 */
4268 	while (!page_lock(pp, se, (kmutex_t *)NULL, P_RECLAIM)) {
4269 		if (page_pptonum(pp) != pfnum)
4270 			goto retry;
4271 		continue;
4272 	}
4273 
4274 	if (page_pptonum(pp) != pfnum) {
4275 		page_unlock(pp);
4276 		goto retry;
4277 	}
4278 
4279 	return (pp);
4280 }
4281 
4282 page_t *
4283 page_numtopp_noreclaim(pfn_t pfnum, se_t se)
4284 {
4285 	page_t *pp;
4286 
4287 retry:
4288 	pp = page_numtopp_nolock(pfnum);
4289 	if (pp == NULL) {
4290 		return ((page_t *)NULL);
4291 	}
4292 
4293 	/*
4294 	 * Acquire the appropriate lock on the page.
4295 	 */
4296 	while (!page_lock(pp, se, (kmutex_t *)NULL, P_NO_RECLAIM)) {
4297 		if (page_pptonum(pp) != pfnum)
4298 			goto retry;
4299 		continue;
4300 	}
4301 
4302 	if (page_pptonum(pp) != pfnum) {
4303 		page_unlock(pp);
4304 		goto retry;
4305 	}
4306 
4307 	return (pp);
4308 }
4309 
4310 /*
4311  * This routine is like page_numtopp, but will only return page structs
4312  * for pages which are ok for loading into hardware using the page struct.
4313  */
4314 page_t *
4315 page_numtopp_nowait(pfn_t pfnum, se_t se)
4316 {
4317 	page_t *pp;
4318 
4319 retry:
4320 	pp = page_numtopp_nolock(pfnum);
4321 	if (pp == NULL) {
4322 		return ((page_t *)NULL);
4323 	}
4324 
4325 	/*
4326 	 * Try to acquire the appropriate lock on the page.
4327 	 */
4328 	if (PP_ISFREE(pp))
4329 		pp = NULL;
4330 	else {
4331 		if (!page_trylock(pp, se))
4332 			pp = NULL;
4333 		else {
4334 			if (page_pptonum(pp) != pfnum) {
4335 				page_unlock(pp);
4336 				goto retry;
4337 			}
4338 			if (PP_ISFREE(pp)) {
4339 				page_unlock(pp);
4340 				pp = NULL;
4341 			}
4342 		}
4343 	}
4344 	return (pp);
4345 }
4346 
4347 /*
4348  * Returns a count of dirty pages that are in the process
4349  * of being written out.  If 'cleanit' is set, try to push the page.
4350  */
4351 pgcnt_t
4352 page_busy(int cleanit)
4353 {
4354 	page_t *page0 = page_first();
4355 	page_t *pp = page0;
4356 	pgcnt_t nppbusy = 0;
4357 	u_offset_t off;
4358 
4359 	do {
4360 		vnode_t *vp = pp->p_vnode;
4361 
4362 		/*
4363 		 * A page is a candidate for syncing if it is:
4364 		 *
4365 		 * (a)	On neither the freelist nor the cachelist
4366 		 * (b)	Hashed onto a vnode
4367 		 * (c)	Not a kernel page
4368 		 * (d)	Dirty
4369 		 * (e)	Not part of a swapfile
4370 		 * (f)	a page which belongs to a real vnode; eg has a non-null
4371 		 *	v_vfsp pointer.
4372 		 * (g)	Backed by a filesystem which doesn't have a
4373 		 *	stubbed-out sync operation
4374 		 */
4375 		if (!PP_ISFREE(pp) && vp != NULL && !VN_ISKAS(vp) &&
4376 		    hat_ismod(pp) && !IS_SWAPVP(vp) && vp->v_vfsp != NULL &&
4377 		    vfs_can_sync(vp->v_vfsp)) {
4378 			nppbusy++;
4379 			vfs_syncprogress();
4380 
4381 			if (!cleanit)
4382 				continue;
4383 			if (!page_trylock(pp, SE_EXCL))
4384 				continue;
4385 
4386 			if (PP_ISFREE(pp) || vp == NULL || IS_SWAPVP(vp) ||
4387 			    pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
4388 			    !(hat_pagesync(pp,
4389 			    HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD) & P_MOD)) {
4390 				page_unlock(pp);
4391 				continue;
4392 			}
4393 			off = pp->p_offset;
4394 			VN_HOLD(vp);
4395 			page_unlock(pp);
4396 			(void) VOP_PUTPAGE(vp, off, PAGESIZE,
4397 			    B_ASYNC | B_FREE, kcred, NULL);
4398 			VN_RELE(vp);
4399 		}
4400 	} while ((pp = page_next(pp)) != page0);
4401 
4402 	return (nppbusy);
4403 }
4404 
4405 void page_invalidate_pages(void);
4406 
4407 /*
4408  * callback handler to vm sub-system
4409  *
4410  * callers make sure no recursive entries to this func.
4411  */
4412 /*ARGSUSED*/
4413 boolean_t
4414 callb_vm_cpr(void *arg, int code)
4415 {
4416 	if (code == CB_CODE_CPR_CHKPT)
4417 		page_invalidate_pages();
4418 	return (B_TRUE);
4419 }
4420 
4421 /*
4422  * Invalidate all pages of the system.
4423  * It shouldn't be called until all user page activities are all stopped.
4424  */
4425 void
4426 page_invalidate_pages()
4427 {
4428 	page_t *pp;
4429 	page_t *page0;
4430 	pgcnt_t nbusypages;
4431 	int retry = 0;
4432 	const int MAXRETRIES = 4;
4433 #if defined(__sparc)
4434 	extern struct vnode prom_ppages;
4435 #endif /* __sparc */
4436 
4437 top:
4438 	/*
4439 	 * Flush dirty pages and destroy the clean ones.
4440 	 */
4441 	nbusypages = 0;
4442 
4443 	pp = page0 = page_first();
4444 	do {
4445 		struct vnode	*vp;
4446 		u_offset_t	offset;
4447 		int		mod;
4448 
4449 		/*
4450 		 * skip the page if it has no vnode or the page associated
4451 		 * with the kernel vnode or prom allocated kernel mem.
4452 		 */
4453 #if defined(__sparc)
4454 		if ((vp = pp->p_vnode) == NULL || VN_ISKAS(vp) ||
4455 		    vp == &prom_ppages)
4456 #else /* x86 doesn't have prom or prom_ppage */
4457 		if ((vp = pp->p_vnode) == NULL || VN_ISKAS(vp))
4458 #endif /* __sparc */
4459 			continue;
4460 
4461 		/*
4462 		 * skip the page which is already free invalidated.
4463 		 */
4464 		if (PP_ISFREE(pp) && PP_ISAGED(pp))
4465 			continue;
4466 
4467 		/*
4468 		 * skip pages that are already locked or can't be "exclusively"
4469 		 * locked or are already free.  After we lock the page, check
4470 		 * the free and age bits again to be sure it's not destroied
4471 		 * yet.
4472 		 * To achieve max. parallelization, we use page_trylock instead
4473 		 * of page_lock so that we don't get block on individual pages
4474 		 * while we have thousands of other pages to process.
4475 		 */
4476 		if (!page_trylock(pp, SE_EXCL)) {
4477 			nbusypages++;
4478 			continue;
4479 		} else if (PP_ISFREE(pp)) {
4480 			if (!PP_ISAGED(pp)) {
4481 				page_destroy_free(pp);
4482 			} else {
4483 				page_unlock(pp);
4484 			}
4485 			continue;
4486 		}
4487 		/*
4488 		 * Is this page involved in some I/O? shared?
4489 		 *
4490 		 * The page_struct_lock need not be acquired to
4491 		 * examine these fields since the page has an
4492 		 * "exclusive" lock.
4493 		 */
4494 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
4495 			page_unlock(pp);
4496 			continue;
4497 		}
4498 
4499 		if (vp->v_type == VCHR) {
4500 			panic("vp->v_type == VCHR");
4501 			/*NOTREACHED*/
4502 		}
4503 
4504 		if (!page_try_demote_pages(pp)) {
4505 			page_unlock(pp);
4506 			continue;
4507 		}
4508 
4509 		/*
4510 		 * Check the modified bit. Leave the bits alone in hardware
4511 		 * (they will be modified if we do the putpage).
4512 		 */
4513 		mod = (hat_pagesync(pp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD)
4514 		    & P_MOD);
4515 		if (mod) {
4516 			offset = pp->p_offset;
4517 			/*
4518 			 * Hold the vnode before releasing the page lock
4519 			 * to prevent it from being freed and re-used by
4520 			 * some other thread.
4521 			 */
4522 			VN_HOLD(vp);
4523 			page_unlock(pp);
4524 			/*
4525 			 * No error return is checked here. Callers such as
4526 			 * cpr deals with the dirty pages at the dump time
4527 			 * if this putpage fails.
4528 			 */
4529 			(void) VOP_PUTPAGE(vp, offset, PAGESIZE, B_INVAL,
4530 			    kcred, NULL);
4531 			VN_RELE(vp);
4532 		} else {
4533 			page_destroy(pp, 0);
4534 		}
4535 	} while ((pp = page_next(pp)) != page0);
4536 	if (nbusypages && retry++ < MAXRETRIES) {
4537 		delay(1);
4538 		goto top;
4539 	}
4540 }
4541 
4542 /*
4543  * Replace the page "old" with the page "new" on the page hash and vnode lists
4544  *
4545  * the replacement must be done in place, ie the equivalent sequence:
4546  *
4547  *	vp = old->p_vnode;
4548  *	off = old->p_offset;
4549  *	page_do_hashout(old)
4550  *	page_do_hashin(new, vp, off)
4551  *
4552  * doesn't work, since
4553  *  1) if old is the only page on the vnode, the v_pages list has a window
4554  *     where it looks empty. This will break file system assumptions.
4555  * and
4556  *  2) pvn_vplist_dirty() can't deal with pages moving on the v_pages list.
4557  */
4558 static void
4559 page_do_relocate_hash(page_t *new, page_t *old)
4560 {
4561 	page_t	**hash_list;
4562 	vnode_t	*vp = old->p_vnode;
4563 	kmutex_t *sep;
4564 
4565 	ASSERT(PAGE_EXCL(old));
4566 	ASSERT(PAGE_EXCL(new));
4567 	ASSERT(vp != NULL);
4568 	ASSERT(MUTEX_HELD(page_vnode_mutex(vp)));
4569 	ASSERT(MUTEX_HELD(PAGE_HASH_MUTEX(PAGE_HASH_FUNC(vp, old->p_offset))));
4570 
4571 	/*
4572 	 * First find old page on the page hash list
4573 	 */
4574 	hash_list = &page_hash[PAGE_HASH_FUNC(vp, old->p_offset)];
4575 
4576 	for (;;) {
4577 		if (*hash_list == old)
4578 			break;
4579 		if (*hash_list == NULL) {
4580 			panic("page_do_hashout");
4581 			/*NOTREACHED*/
4582 		}
4583 		hash_list = &(*hash_list)->p_hash;
4584 	}
4585 
4586 	/*
4587 	 * update new and replace old with new on the page hash list
4588 	 */
4589 	new->p_vnode = old->p_vnode;
4590 	new->p_offset = old->p_offset;
4591 	new->p_hash = old->p_hash;
4592 	*hash_list = new;
4593 
4594 	if ((new->p_vnode->v_flag & VISSWAP) != 0)
4595 		PP_SETSWAP(new);
4596 
4597 	/*
4598 	 * replace old with new on the vnode's page list
4599 	 */
4600 	if (old->p_vpnext == old) {
4601 		new->p_vpnext = new;
4602 		new->p_vpprev = new;
4603 	} else {
4604 		new->p_vpnext = old->p_vpnext;
4605 		new->p_vpprev = old->p_vpprev;
4606 		new->p_vpnext->p_vpprev = new;
4607 		new->p_vpprev->p_vpnext = new;
4608 	}
4609 	if (vp->v_pages == old)
4610 		vp->v_pages = new;
4611 
4612 	/*
4613 	 * clear out the old page
4614 	 */
4615 	old->p_hash = NULL;
4616 	old->p_vpnext = NULL;
4617 	old->p_vpprev = NULL;
4618 	old->p_vnode = NULL;
4619 	PP_CLRSWAP(old);
4620 	old->p_offset = (u_offset_t)-1;
4621 	page_clr_all_props(old);
4622 
4623 	/*
4624 	 * Wake up processes waiting for this page.  The page's
4625 	 * identity has been changed, and is probably not the
4626 	 * desired page any longer.
4627 	 */
4628 	sep = page_se_mutex(old);
4629 	mutex_enter(sep);
4630 	old->p_selock &= ~SE_EWANTED;
4631 	if (CV_HAS_WAITERS(&old->p_cv))
4632 		cv_broadcast(&old->p_cv);
4633 	mutex_exit(sep);
4634 }
4635 
4636 /*
4637  * This function moves the identity of page "pp_old" to page "pp_new".
4638  * Both pages must be locked on entry.  "pp_new" is free, has no identity,
4639  * and need not be hashed out from anywhere.
4640  */
4641 void
4642 page_relocate_hash(page_t *pp_new, page_t *pp_old)
4643 {
4644 	vnode_t *vp = pp_old->p_vnode;
4645 	u_offset_t off = pp_old->p_offset;
4646 	kmutex_t *phm, *vphm;
4647 
4648 	/*
4649 	 * Rehash two pages
4650 	 */
4651 	ASSERT(PAGE_EXCL(pp_old));
4652 	ASSERT(PAGE_EXCL(pp_new));
4653 	ASSERT(vp != NULL);
4654 	ASSERT(pp_new->p_vnode == NULL);
4655 
4656 	/*
4657 	 * hashout then hashin while holding the mutexes
4658 	 */
4659 	phm = PAGE_HASH_MUTEX(PAGE_HASH_FUNC(vp, off));
4660 	mutex_enter(phm);
4661 	vphm = page_vnode_mutex(vp);
4662 	mutex_enter(vphm);
4663 
4664 	page_do_relocate_hash(pp_new, pp_old);
4665 
4666 	mutex_exit(vphm);
4667 	mutex_exit(phm);
4668 
4669 	/*
4670 	 * The page_struct_lock need not be acquired for lckcnt and
4671 	 * cowcnt since the page has an "exclusive" lock.
4672 	 */
4673 	ASSERT(pp_new->p_lckcnt == 0);
4674 	ASSERT(pp_new->p_cowcnt == 0);
4675 	pp_new->p_lckcnt = pp_old->p_lckcnt;
4676 	pp_new->p_cowcnt = pp_old->p_cowcnt;
4677 	pp_old->p_lckcnt = pp_old->p_cowcnt = 0;
4678 
4679 	/* The following comment preserved from page_flip(). */
4680 	/* XXX - Do we need to protect fsdata? */
4681 	pp_new->p_fsdata = pp_old->p_fsdata;
4682 }
4683 
4684 /*
4685  * Helper routine used to lock all remaining members of a
4686  * large page. The caller is responsible for passing in a locked
4687  * pp. If pp is a large page, then it succeeds in locking all the
4688  * remaining constituent pages or it returns with only the
4689  * original page locked.
4690  *
4691  * Returns 1 on success, 0 on failure.
4692  *
4693  * If success is returned this routine guarantees p_szc for all constituent
4694  * pages of a large page pp belongs to can't change. To achieve this we
4695  * recheck szc of pp after locking all constituent pages and retry if szc
4696  * changed (it could only decrease). Since hat_page_demote() needs an EXCL
4697  * lock on one of constituent pages it can't be running after all constituent
4698  * pages are locked.  hat_page_demote() with a lock on a constituent page
4699  * outside of this large page (i.e. pp belonged to a larger large page) is
4700  * already done with all constituent pages of pp since the root's p_szc is
4701  * changed last. Therefore no need to synchronize with hat_page_demote() that
4702  * locked a constituent page outside of pp's current large page.
4703  */
4704 #ifdef DEBUG
4705 uint32_t gpg_trylock_mtbf = 0;
4706 #endif
4707 
4708 int
4709 group_page_trylock(page_t *pp, se_t se)
4710 {
4711 	page_t  *tpp;
4712 	pgcnt_t	npgs, i, j;
4713 	uint_t pszc = pp->p_szc;
4714 
4715 #ifdef DEBUG
4716 	if (gpg_trylock_mtbf && !(gethrtime() % gpg_trylock_mtbf)) {
4717 		return (0);
4718 	}
4719 #endif
4720 
4721 	if (pp != PP_GROUPLEADER(pp, pszc)) {
4722 		return (0);
4723 	}
4724 
4725 retry:
4726 	ASSERT(PAGE_LOCKED_SE(pp, se));
4727 	ASSERT(!PP_ISFREE(pp));
4728 	if (pszc == 0) {
4729 		return (1);
4730 	}
4731 	npgs = page_get_pagecnt(pszc);
4732 	tpp = pp + 1;
4733 	for (i = 1; i < npgs; i++, tpp++) {
4734 		if (!page_trylock(tpp, se)) {
4735 			tpp = pp + 1;
4736 			for (j = 1; j < i; j++, tpp++) {
4737 				page_unlock(tpp);
4738 			}
4739 			return (0);
4740 		}
4741 	}
4742 	if (pp->p_szc != pszc) {
4743 		ASSERT(pp->p_szc < pszc);
4744 		ASSERT(pp->p_vnode != NULL && !PP_ISKAS(pp) &&
4745 		    !IS_SWAPFSVP(pp->p_vnode));
4746 		tpp = pp + 1;
4747 		for (i = 1; i < npgs; i++, tpp++) {
4748 			page_unlock(tpp);
4749 		}
4750 		pszc = pp->p_szc;
4751 		goto retry;
4752 	}
4753 	return (1);
4754 }
4755 
4756 void
4757 group_page_unlock(page_t *pp)
4758 {
4759 	page_t *tpp;
4760 	pgcnt_t	npgs, i;
4761 
4762 	ASSERT(PAGE_LOCKED(pp));
4763 	ASSERT(!PP_ISFREE(pp));
4764 	ASSERT(pp == PP_PAGEROOT(pp));
4765 	npgs = page_get_pagecnt(pp->p_szc);
4766 	for (i = 1, tpp = pp + 1; i < npgs; i++, tpp++) {
4767 		page_unlock(tpp);
4768 	}
4769 }
4770 
4771 /*
4772  * returns
4773  * 0 		: on success and *nrelocp is number of relocated PAGESIZE pages
4774  * ERANGE	: this is not a base page
4775  * EBUSY	: failure to get locks on the page/pages
4776  * ENOMEM	: failure to obtain replacement pages
4777  * EAGAIN	: OBP has not yet completed its boot-time handoff to the kernel
4778  * EIO		: An error occurred while trying to copy the page data
4779  *
4780  * Return with all constituent members of target and replacement
4781  * SE_EXCL locked. It is the callers responsibility to drop the
4782  * locks.
4783  */
4784 int
4785 do_page_relocate(
4786 	page_t **target,
4787 	page_t **replacement,
4788 	int grouplock,
4789 	spgcnt_t *nrelocp,
4790 	lgrp_t *lgrp)
4791 {
4792 	page_t *first_repl;
4793 	page_t *repl;
4794 	page_t *targ;
4795 	page_t *pl = NULL;
4796 	uint_t ppattr;
4797 	pfn_t   pfn, repl_pfn;
4798 	uint_t	szc;
4799 	spgcnt_t npgs, i;
4800 	int repl_contig = 0;
4801 	uint_t flags = 0;
4802 	spgcnt_t dofree = 0;
4803 
4804 	*nrelocp = 0;
4805 
4806 #if defined(__sparc)
4807 	/*
4808 	 * We need to wait till OBP has completed
4809 	 * its boot-time handoff of its resources to the kernel
4810 	 * before we allow page relocation
4811 	 */
4812 	if (page_relocate_ready == 0) {
4813 		return (EAGAIN);
4814 	}
4815 #endif
4816 
4817 	/*
4818 	 * If this is not a base page,
4819 	 * just return with 0x0 pages relocated.
4820 	 */
4821 	targ = *target;
4822 	ASSERT(PAGE_EXCL(targ));
4823 	ASSERT(!PP_ISFREE(targ));
4824 	szc = targ->p_szc;
4825 	ASSERT(szc < mmu_page_sizes);
4826 	VM_STAT_ADD(vmm_vmstats.ppr_reloc[szc]);
4827 	pfn = targ->p_pagenum;
4828 	if (pfn != PFN_BASE(pfn, szc)) {
4829 		VM_STAT_ADD(vmm_vmstats.ppr_relocnoroot[szc]);
4830 		return (ERANGE);
4831 	}
4832 
4833 	if ((repl = *replacement) != NULL && repl->p_szc >= szc) {
4834 		repl_pfn = repl->p_pagenum;
4835 		if (repl_pfn != PFN_BASE(repl_pfn, szc)) {
4836 			VM_STAT_ADD(vmm_vmstats.ppr_reloc_replnoroot[szc]);
4837 			return (ERANGE);
4838 		}
4839 		repl_contig = 1;
4840 	}
4841 
4842 	/*
4843 	 * We must lock all members of this large page or we cannot
4844 	 * relocate any part of it.
4845 	 */
4846 	if (grouplock != 0 && !group_page_trylock(targ, SE_EXCL)) {
4847 		VM_STAT_ADD(vmm_vmstats.ppr_relocnolock[targ->p_szc]);
4848 		return (EBUSY);
4849 	}
4850 
4851 	/*
4852 	 * reread szc it could have been decreased before
4853 	 * group_page_trylock() was done.
4854 	 */
4855 	szc = targ->p_szc;
4856 	ASSERT(szc < mmu_page_sizes);
4857 	VM_STAT_ADD(vmm_vmstats.ppr_reloc[szc]);
4858 	ASSERT(pfn == PFN_BASE(pfn, szc));
4859 
4860 	npgs = page_get_pagecnt(targ->p_szc);
4861 
4862 	if (repl == NULL) {
4863 		dofree = npgs;		/* Size of target page in MMU pages */
4864 		if (!page_create_wait(dofree, 0)) {
4865 			if (grouplock != 0) {
4866 				group_page_unlock(targ);
4867 			}
4868 			VM_STAT_ADD(vmm_vmstats.ppr_relocnomem[szc]);
4869 			return (ENOMEM);
4870 		}
4871 
4872 		/*
4873 		 * seg kmem pages require that the target and replacement
4874 		 * page be the same pagesize.
4875 		 */
4876 		flags = (VN_ISKAS(targ->p_vnode)) ? PGR_SAMESZC : 0;
4877 		repl = page_get_replacement_page(targ, lgrp, flags);
4878 		if (repl == NULL) {
4879 			if (grouplock != 0) {
4880 				group_page_unlock(targ);
4881 			}
4882 			page_create_putback(dofree);
4883 			VM_STAT_ADD(vmm_vmstats.ppr_relocnomem[szc]);
4884 			return (ENOMEM);
4885 		}
4886 	}
4887 #ifdef DEBUG
4888 	else {
4889 		ASSERT(PAGE_LOCKED(repl));
4890 	}
4891 #endif /* DEBUG */
4892 
4893 #if defined(__sparc)
4894 	/*
4895 	 * Let hat_page_relocate() complete the relocation if it's kernel page
4896 	 */
4897 	if (VN_ISKAS(targ->p_vnode)) {
4898 		*replacement = repl;
4899 		if (hat_page_relocate(target, replacement, nrelocp) != 0) {
4900 			if (grouplock != 0) {
4901 				group_page_unlock(targ);
4902 			}
4903 			if (dofree) {
4904 				*replacement = NULL;
4905 				page_free_replacement_page(repl);
4906 				page_create_putback(dofree);
4907 			}
4908 			VM_STAT_ADD(vmm_vmstats.ppr_krelocfail[szc]);
4909 			return (EAGAIN);
4910 		}
4911 		VM_STAT_ADD(vmm_vmstats.ppr_relocok[szc]);
4912 		return (0);
4913 	}
4914 #else
4915 #if defined(lint)
4916 	dofree = dofree;
4917 #endif
4918 #endif
4919 
4920 	first_repl = repl;
4921 
4922 	for (i = 0; i < npgs; i++) {
4923 		ASSERT(PAGE_EXCL(targ));
4924 		ASSERT(targ->p_slckcnt == 0);
4925 		ASSERT(repl->p_slckcnt == 0);
4926 
4927 		(void) hat_pageunload(targ, HAT_FORCE_PGUNLOAD);
4928 
4929 		ASSERT(hat_page_getshare(targ) == 0);
4930 		ASSERT(!PP_ISFREE(targ));
4931 		ASSERT(targ->p_pagenum == (pfn + i));
4932 		ASSERT(repl_contig == 0 ||
4933 		    repl->p_pagenum == (repl_pfn + i));
4934 
4935 		/*
4936 		 * Copy the page contents and attributes then
4937 		 * relocate the page in the page hash.
4938 		 */
4939 		if (ppcopy(targ, repl) == 0) {
4940 			targ = *target;
4941 			repl = first_repl;
4942 			VM_STAT_ADD(vmm_vmstats.ppr_copyfail);
4943 			if (grouplock != 0) {
4944 				group_page_unlock(targ);
4945 			}
4946 			if (dofree) {
4947 				*replacement = NULL;
4948 				page_free_replacement_page(repl);
4949 				page_create_putback(dofree);
4950 			}
4951 			return (EIO);
4952 		}
4953 
4954 		targ++;
4955 		if (repl_contig != 0) {
4956 			repl++;
4957 		} else {
4958 			repl = repl->p_next;
4959 		}
4960 	}
4961 
4962 	repl = first_repl;
4963 	targ = *target;
4964 
4965 	for (i = 0; i < npgs; i++) {
4966 		ppattr = hat_page_getattr(targ, (P_MOD | P_REF | P_RO));
4967 		page_clr_all_props(repl);
4968 		page_set_props(repl, ppattr);
4969 		page_relocate_hash(repl, targ);
4970 
4971 		ASSERT(hat_page_getshare(targ) == 0);
4972 		ASSERT(hat_page_getshare(repl) == 0);
4973 		/*
4974 		 * Now clear the props on targ, after the
4975 		 * page_relocate_hash(), they no longer
4976 		 * have any meaning.
4977 		 */
4978 		page_clr_all_props(targ);
4979 		ASSERT(targ->p_next == targ);
4980 		ASSERT(targ->p_prev == targ);
4981 		page_list_concat(&pl, &targ);
4982 
4983 		targ++;
4984 		if (repl_contig != 0) {
4985 			repl++;
4986 		} else {
4987 			repl = repl->p_next;
4988 		}
4989 	}
4990 	/* assert that we have come full circle with repl */
4991 	ASSERT(repl_contig == 1 || first_repl == repl);
4992 
4993 	*target = pl;
4994 	if (*replacement == NULL) {
4995 		ASSERT(first_repl == repl);
4996 		*replacement = repl;
4997 	}
4998 	VM_STAT_ADD(vmm_vmstats.ppr_relocok[szc]);
4999 	*nrelocp = npgs;
5000 	return (0);
5001 }
5002 /*
5003  * On success returns 0 and *nrelocp the number of PAGESIZE pages relocated.
5004  */
5005 int
5006 page_relocate(
5007 	page_t **target,
5008 	page_t **replacement,
5009 	int grouplock,
5010 	int freetarget,
5011 	spgcnt_t *nrelocp,
5012 	lgrp_t *lgrp)
5013 {
5014 	spgcnt_t ret;
5015 
5016 	/* do_page_relocate returns 0 on success or errno value */
5017 	ret = do_page_relocate(target, replacement, grouplock, nrelocp, lgrp);
5018 
5019 	if (ret != 0 || freetarget == 0) {
5020 		return (ret);
5021 	}
5022 	if (*nrelocp == 1) {
5023 		ASSERT(*target != NULL);
5024 		page_free(*target, 1);
5025 	} else {
5026 		page_t *tpp = *target;
5027 		uint_t szc = tpp->p_szc;
5028 		pgcnt_t npgs = page_get_pagecnt(szc);
5029 		ASSERT(npgs > 1);
5030 		ASSERT(szc != 0);
5031 		do {
5032 			ASSERT(PAGE_EXCL(tpp));
5033 			ASSERT(!hat_page_is_mapped(tpp));
5034 			ASSERT(tpp->p_szc == szc);
5035 			PP_SETFREE(tpp);
5036 			PP_SETAGED(tpp);
5037 			npgs--;
5038 		} while ((tpp = tpp->p_next) != *target);
5039 		ASSERT(npgs == 0);
5040 		page_list_add_pages(*target, 0);
5041 		npgs = page_get_pagecnt(szc);
5042 		page_create_putback(npgs);
5043 	}
5044 	return (ret);
5045 }
5046 
5047 /*
5048  * it is up to the caller to deal with pcf accounting.
5049  */
5050 void
5051 page_free_replacement_page(page_t *pplist)
5052 {
5053 	page_t *pp;
5054 
5055 	while (pplist != NULL) {
5056 		/*
5057 		 * pp_targ is a linked list.
5058 		 */
5059 		pp = pplist;
5060 		if (pp->p_szc == 0) {
5061 			page_sub(&pplist, pp);
5062 			page_clr_all_props(pp);
5063 			PP_SETFREE(pp);
5064 			PP_SETAGED(pp);
5065 			page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL);
5066 			page_unlock(pp);
5067 			VM_STAT_ADD(pagecnt.pc_free_replacement_page[0]);
5068 		} else {
5069 			spgcnt_t curnpgs = page_get_pagecnt(pp->p_szc);
5070 			page_t *tpp;
5071 			page_list_break(&pp, &pplist, curnpgs);
5072 			tpp = pp;
5073 			do {
5074 				ASSERT(PAGE_EXCL(tpp));
5075 				ASSERT(!hat_page_is_mapped(tpp));
5076 				page_clr_all_props(pp);
5077 				PP_SETFREE(tpp);
5078 				PP_SETAGED(tpp);
5079 			} while ((tpp = tpp->p_next) != pp);
5080 			page_list_add_pages(pp, 0);
5081 			VM_STAT_ADD(pagecnt.pc_free_replacement_page[1]);
5082 		}
5083 	}
5084 }
5085 
5086 /*
5087  * Relocate target to non-relocatable replacement page.
5088  */
5089 int
5090 page_relocate_cage(page_t **target, page_t **replacement)
5091 {
5092 	page_t *tpp, *rpp;
5093 	spgcnt_t pgcnt, npgs;
5094 	int result;
5095 
5096 	tpp = *target;
5097 
5098 	ASSERT(PAGE_EXCL(tpp));
5099 	ASSERT(tpp->p_szc == 0);
5100 
5101 	pgcnt = btop(page_get_pagesize(tpp->p_szc));
5102 
5103 	do {
5104 		(void) page_create_wait(pgcnt, PG_WAIT | PG_NORELOC);
5105 		rpp = page_get_replacement_page(tpp, NULL, PGR_NORELOC);
5106 		if (rpp == NULL) {
5107 			page_create_putback(pgcnt);
5108 			kcage_cageout_wakeup();
5109 		}
5110 	} while (rpp == NULL);
5111 
5112 	ASSERT(PP_ISNORELOC(rpp));
5113 
5114 	result = page_relocate(&tpp, &rpp, 0, 1, &npgs, NULL);
5115 
5116 	if (result == 0) {
5117 		*replacement = rpp;
5118 		if (pgcnt != npgs)
5119 			panic("page_relocate_cage: partial relocation");
5120 	}
5121 
5122 	return (result);
5123 }
5124 
5125 /*
5126  * Release the page lock on a page, place on cachelist
5127  * tail if no longer mapped. Caller can let us know if
5128  * the page is known to be clean.
5129  */
5130 int
5131 page_release(page_t *pp, int checkmod)
5132 {
5133 	int status;
5134 
5135 	ASSERT(PAGE_LOCKED(pp) && !PP_ISFREE(pp) &&
5136 	    (pp->p_vnode != NULL));
5137 
5138 	if (!hat_page_is_mapped(pp) && !IS_SWAPVP(pp->p_vnode) &&
5139 	    ((PAGE_SHARED(pp) && page_tryupgrade(pp)) || PAGE_EXCL(pp)) &&
5140 	    pp->p_lckcnt == 0 && pp->p_cowcnt == 0 &&
5141 	    !hat_page_is_mapped(pp)) {
5142 
5143 		/*
5144 		 * If page is modified, unlock it
5145 		 *
5146 		 * (p_nrm & P_MOD) bit has the latest stuff because:
5147 		 * (1) We found that this page doesn't have any mappings
5148 		 *	_after_ holding SE_EXCL and
5149 		 * (2) We didn't drop SE_EXCL lock after the check in (1)
5150 		 */
5151 		if (checkmod && hat_ismod(pp)) {
5152 			page_unlock(pp);
5153 			status = PGREL_MOD;
5154 		} else {
5155 			/*LINTED: constant in conditional context*/
5156 			VN_DISPOSE(pp, B_FREE, 0, kcred);
5157 			status = PGREL_CLEAN;
5158 		}
5159 	} else {
5160 		page_unlock(pp);
5161 		status = PGREL_NOTREL;
5162 	}
5163 	return (status);
5164 }
5165 
5166 /*
5167  * Given a constituent page, try to demote the large page on the freelist.
5168  *
5169  * Returns nonzero if the page could be demoted successfully. Returns with
5170  * the constituent page still locked.
5171  */
5172 int
5173 page_try_demote_free_pages(page_t *pp)
5174 {
5175 	page_t *rootpp = pp;
5176 	pfn_t	pfn = page_pptonum(pp);
5177 	spgcnt_t npgs;
5178 	uint_t	szc = pp->p_szc;
5179 
5180 	ASSERT(PP_ISFREE(pp));
5181 	ASSERT(PAGE_EXCL(pp));
5182 
5183 	/*
5184 	 * Adjust rootpp and lock it, if `pp' is not the base
5185 	 * constituent page.
5186 	 */
5187 	npgs = page_get_pagecnt(pp->p_szc);
5188 	if (npgs == 1) {
5189 		return (0);
5190 	}
5191 
5192 	if (!IS_P2ALIGNED(pfn, npgs)) {
5193 		pfn = P2ALIGN(pfn, npgs);
5194 		rootpp = page_numtopp_nolock(pfn);
5195 	}
5196 
5197 	if (pp != rootpp && !page_trylock(rootpp, SE_EXCL)) {
5198 		return (0);
5199 	}
5200 
5201 	if (rootpp->p_szc != szc) {
5202 		if (pp != rootpp)
5203 			page_unlock(rootpp);
5204 		return (0);
5205 	}
5206 
5207 	page_demote_free_pages(rootpp);
5208 
5209 	if (pp != rootpp)
5210 		page_unlock(rootpp);
5211 
5212 	ASSERT(PP_ISFREE(pp));
5213 	ASSERT(PAGE_EXCL(pp));
5214 	return (1);
5215 }
5216 
5217 /*
5218  * Given a constituent page, try to demote the large page.
5219  *
5220  * Returns nonzero if the page could be demoted successfully. Returns with
5221  * the constituent page still locked.
5222  */
5223 int
5224 page_try_demote_pages(page_t *pp)
5225 {
5226 	page_t *tpp, *rootpp = pp;
5227 	pfn_t	pfn = page_pptonum(pp);
5228 	spgcnt_t i, npgs;
5229 	uint_t	szc = pp->p_szc;
5230 	vnode_t *vp = pp->p_vnode;
5231 
5232 	ASSERT(PAGE_EXCL(pp));
5233 
5234 	VM_STAT_ADD(pagecnt.pc_try_demote_pages[0]);
5235 
5236 	if (pp->p_szc == 0) {
5237 		VM_STAT_ADD(pagecnt.pc_try_demote_pages[1]);
5238 		return (1);
5239 	}
5240 
5241 	if (vp != NULL && !IS_SWAPFSVP(vp) && !VN_ISKAS(vp)) {
5242 		VM_STAT_ADD(pagecnt.pc_try_demote_pages[2]);
5243 		page_demote_vp_pages(pp);
5244 		ASSERT(pp->p_szc == 0);
5245 		return (1);
5246 	}
5247 
5248 	/*
5249 	 * Adjust rootpp if passed in is not the base
5250 	 * constituent page.
5251 	 */
5252 	npgs = page_get_pagecnt(pp->p_szc);
5253 	ASSERT(npgs > 1);
5254 	if (!IS_P2ALIGNED(pfn, npgs)) {
5255 		pfn = P2ALIGN(pfn, npgs);
5256 		rootpp = page_numtopp_nolock(pfn);
5257 		VM_STAT_ADD(pagecnt.pc_try_demote_pages[3]);
5258 		ASSERT(rootpp->p_vnode != NULL);
5259 		ASSERT(rootpp->p_szc == szc);
5260 	}
5261 
5262 	/*
5263 	 * We can't demote kernel pages since we can't hat_unload()
5264 	 * the mappings.
5265 	 */
5266 	if (VN_ISKAS(rootpp->p_vnode))
5267 		return (0);
5268 
5269 	/*
5270 	 * Attempt to lock all constituent pages except the page passed
5271 	 * in since it's already locked.
5272 	 */
5273 	for (tpp = rootpp, i = 0; i < npgs; i++, tpp++) {
5274 		ASSERT(!PP_ISFREE(tpp));
5275 		ASSERT(tpp->p_vnode != NULL);
5276 
5277 		if (tpp != pp && !page_trylock(tpp, SE_EXCL))
5278 			break;
5279 		ASSERT(tpp->p_szc == rootpp->p_szc);
5280 		ASSERT(page_pptonum(tpp) == page_pptonum(rootpp) + i);
5281 	}
5282 
5283 	/*
5284 	 * If we failed to lock them all then unlock what we have
5285 	 * locked so far and bail.
5286 	 */
5287 	if (i < npgs) {
5288 		tpp = rootpp;
5289 		while (i-- > 0) {
5290 			if (tpp != pp)
5291 				page_unlock(tpp);
5292 			tpp++;
5293 		}
5294 		VM_STAT_ADD(pagecnt.pc_try_demote_pages[4]);
5295 		return (0);
5296 	}
5297 
5298 	for (tpp = rootpp, i = 0; i < npgs; i++, tpp++) {
5299 		ASSERT(PAGE_EXCL(tpp));
5300 		ASSERT(tpp->p_slckcnt == 0);
5301 		(void) hat_pageunload(tpp, HAT_FORCE_PGUNLOAD);
5302 		tpp->p_szc = 0;
5303 	}
5304 
5305 	/*
5306 	 * Unlock all pages except the page passed in.
5307 	 */
5308 	for (tpp = rootpp, i = 0; i < npgs; i++, tpp++) {
5309 		ASSERT(!hat_page_is_mapped(tpp));
5310 		if (tpp != pp)
5311 			page_unlock(tpp);
5312 	}
5313 
5314 	VM_STAT_ADD(pagecnt.pc_try_demote_pages[5]);
5315 	return (1);
5316 }
5317 
5318 /*
5319  * Called by page_free() and page_destroy() to demote the page size code
5320  * (p_szc) to 0 (since we can't just put a single PAGESIZE page with non zero
5321  * p_szc on free list, neither can we just clear p_szc of a single page_t
5322  * within a large page since it will break other code that relies on p_szc
5323  * being the same for all page_t's of a large page). Anonymous pages should
5324  * never end up here because anon_map_getpages() cannot deal with p_szc
5325  * changes after a single constituent page is locked.  While anonymous or
5326  * kernel large pages are demoted or freed the entire large page at a time
5327  * with all constituent pages locked EXCL for the file system pages we
5328  * have to be able to demote a large page (i.e. decrease all constituent pages
5329  * p_szc) with only just an EXCL lock on one of constituent pages. The reason
5330  * we can easily deal with anonymous page demotion the entire large page at a
5331  * time is that those operation originate at address space level and concern
5332  * the entire large page region with actual demotion only done when pages are
5333  * not shared with any other processes (therefore we can always get EXCL lock
5334  * on all anonymous constituent pages after clearing segment page
5335  * cache). However file system pages can be truncated or invalidated at a
5336  * PAGESIZE level from the file system side and end up in page_free() or
5337  * page_destroy() (we also allow only part of the large page to be SOFTLOCKed
5338  * and therefore pageout should be able to demote a large page by EXCL locking
5339  * any constituent page that is not under SOFTLOCK). In those cases we cannot
5340  * rely on being able to lock EXCL all constituent pages.
5341  *
5342  * To prevent szc changes on file system pages one has to lock all constituent
5343  * pages at least SHARED (or call page_szc_lock()). The only subsystem that
5344  * doesn't rely on locking all constituent pages (or using page_szc_lock()) to
5345  * prevent szc changes is hat layer that uses its own page level mlist
5346  * locks. hat assumes that szc doesn't change after mlist lock for a page is
5347  * taken. Therefore we need to change szc under hat level locks if we only
5348  * have an EXCL lock on a single constituent page and hat still references any
5349  * of constituent pages.  (Note we can't "ignore" hat layer by simply
5350  * hat_pageunload() all constituent pages without having EXCL locks on all of
5351  * constituent pages). We use hat_page_demote() call to safely demote szc of
5352  * all constituent pages under hat locks when we only have an EXCL lock on one
5353  * of constituent pages.
5354  *
5355  * This routine calls page_szc_lock() before calling hat_page_demote() to
5356  * allow segvn in one special case not to lock all constituent pages SHARED
5357  * before calling hat_memload_array() that relies on p_szc not changing even
5358  * before hat level mlist lock is taken.  In that case segvn uses
5359  * page_szc_lock() to prevent hat_page_demote() changing p_szc values.
5360  *
5361  * Anonymous or kernel page demotion still has to lock all pages exclusively
5362  * and do hat_pageunload() on all constituent pages before demoting the page
5363  * therefore there's no need for anonymous or kernel page demotion to use
5364  * hat_page_demote() mechanism.
5365  *
5366  * hat_page_demote() removes all large mappings that map pp and then decreases
5367  * p_szc starting from the last constituent page of the large page. By working
5368  * from the tail of a large page in pfn decreasing order allows one looking at
5369  * the root page to know that hat_page_demote() is done for root's szc area.
5370  * e.g. if a root page has szc 1 one knows it only has to lock all constituent
5371  * pages within szc 1 area to prevent szc changes because hat_page_demote()
5372  * that started on this page when it had szc > 1 is done for this szc 1 area.
5373  *
5374  * We are guaranteed that all constituent pages of pp's large page belong to
5375  * the same vnode with the consecutive offsets increasing in the direction of
5376  * the pfn i.e. the identity of constituent pages can't change until their
5377  * p_szc is decreased. Therefore it's safe for hat_page_demote() to remove
5378  * large mappings to pp even though we don't lock any constituent page except
5379  * pp (i.e. we won't unload e.g. kernel locked page).
5380  */
5381 static void
5382 page_demote_vp_pages(page_t *pp)
5383 {
5384 	kmutex_t *mtx;
5385 
5386 	ASSERT(PAGE_EXCL(pp));
5387 	ASSERT(!PP_ISFREE(pp));
5388 	ASSERT(pp->p_vnode != NULL);
5389 	ASSERT(!IS_SWAPFSVP(pp->p_vnode));
5390 	ASSERT(!PP_ISKAS(pp));
5391 
5392 	VM_STAT_ADD(pagecnt.pc_demote_pages[0]);
5393 
5394 	mtx = page_szc_lock(pp);
5395 	if (mtx != NULL) {
5396 		hat_page_demote(pp);
5397 		mutex_exit(mtx);
5398 	}
5399 	ASSERT(pp->p_szc == 0);
5400 }
5401 
5402 /*
5403  * Mark any existing pages for migration in the given range
5404  */
5405 void
5406 page_mark_migrate(struct seg *seg, caddr_t addr, size_t len,
5407     struct anon_map *amp, ulong_t anon_index, vnode_t *vp,
5408     u_offset_t vnoff, int rflag)
5409 {
5410 	struct anon	*ap;
5411 	vnode_t		*curvp;
5412 	lgrp_t		*from;
5413 	pgcnt_t		i;
5414 	pgcnt_t		nlocked;
5415 	u_offset_t	off;
5416 	pfn_t		pfn;
5417 	size_t		pgsz;
5418 	size_t		segpgsz;
5419 	pgcnt_t		pages;
5420 	uint_t		pszc;
5421 	page_t		**ppa;
5422 	pgcnt_t		ppa_nentries;
5423 	page_t		*pp;
5424 	caddr_t		va;
5425 	ulong_t		an_idx;
5426 	anon_sync_obj_t	cookie;
5427 
5428 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5429 
5430 	/*
5431 	 * Don't do anything if don't need to do lgroup optimizations
5432 	 * on this system
5433 	 */
5434 	if (!lgrp_optimizations())
5435 		return;
5436 
5437 	/*
5438 	 * Align address and length to (potentially large) page boundary
5439 	 */
5440 	segpgsz = page_get_pagesize(seg->s_szc);
5441 	addr = (caddr_t)P2ALIGN((uintptr_t)addr, segpgsz);
5442 	if (rflag)
5443 		len = P2ROUNDUP(len, segpgsz);
5444 
5445 	/*
5446 	 * Allocate page array to accommodate largest page size
5447 	 */
5448 	pgsz = page_get_pagesize(page_num_pagesizes() - 1);
5449 	ppa_nentries = btop(pgsz);
5450 	ppa = kmem_zalloc(ppa_nentries * sizeof (page_t *), KM_SLEEP);
5451 
5452 	/*
5453 	 * Do one (large) page at a time
5454 	 */
5455 	va = addr;
5456 	while (va < addr + len) {
5457 		/*
5458 		 * Lookup (root) page for vnode and offset corresponding to
5459 		 * this virtual address
5460 		 * Try anonmap first since there may be copy-on-write
5461 		 * pages, but initialize vnode pointer and offset using
5462 		 * vnode arguments just in case there isn't an amp.
5463 		 */
5464 		curvp = vp;
5465 		off = vnoff + va - seg->s_base;
5466 		if (amp) {
5467 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5468 			an_idx = anon_index + seg_page(seg, va);
5469 			anon_array_enter(amp, an_idx, &cookie);
5470 			ap = anon_get_ptr(amp->ahp, an_idx);
5471 			if (ap)
5472 				swap_xlate(ap, &curvp, &off);
5473 			anon_array_exit(&cookie);
5474 			ANON_LOCK_EXIT(&amp->a_rwlock);
5475 		}
5476 
5477 		pp = NULL;
5478 		if (curvp)
5479 			pp = page_lookup(curvp, off, SE_SHARED);
5480 
5481 		/*
5482 		 * If there isn't a page at this virtual address,
5483 		 * skip to next page
5484 		 */
5485 		if (pp == NULL) {
5486 			va += PAGESIZE;
5487 			continue;
5488 		}
5489 
5490 		/*
5491 		 * Figure out which lgroup this page is in for kstats
5492 		 */
5493 		pfn = page_pptonum(pp);
5494 		from = lgrp_pfn_to_lgrp(pfn);
5495 
5496 		/*
5497 		 * Get page size, and round up and skip to next page boundary
5498 		 * if unaligned address
5499 		 */
5500 		pszc = pp->p_szc;
5501 		pgsz = page_get_pagesize(pszc);
5502 		pages = btop(pgsz);
5503 		if (!IS_P2ALIGNED(va, pgsz) ||
5504 		    !IS_P2ALIGNED(pfn, pages) ||
5505 		    pgsz > segpgsz) {
5506 			pgsz = MIN(pgsz, segpgsz);
5507 			page_unlock(pp);
5508 			i = btop(P2END((uintptr_t)va, pgsz) -
5509 			    (uintptr_t)va);
5510 			va = (caddr_t)P2END((uintptr_t)va, pgsz);
5511 			lgrp_stat_add(from->lgrp_id, LGRP_PMM_FAIL_PGS, i);
5512 			continue;
5513 		}
5514 
5515 		/*
5516 		 * Upgrade to exclusive lock on page
5517 		 */
5518 		if (!page_tryupgrade(pp)) {
5519 			page_unlock(pp);
5520 			va += pgsz;
5521 			lgrp_stat_add(from->lgrp_id, LGRP_PMM_FAIL_PGS,
5522 			    btop(pgsz));
5523 			continue;
5524 		}
5525 
5526 		/*
5527 		 * Remember pages locked exclusively and how many
5528 		 */
5529 		ppa[0] = pp;
5530 		nlocked = 1;
5531 
5532 		/*
5533 		 * Lock constituent pages if this is large page
5534 		 */
5535 		if (pages > 1) {
5536 			/*
5537 			 * Lock all constituents except root page, since it
5538 			 * should be locked already.
5539 			 */
5540 			for (i = 1; i < pages; i++) {
5541 				pp++;
5542 				if (!page_trylock(pp, SE_EXCL)) {
5543 					break;
5544 				}
5545 				if (PP_ISFREE(pp) ||
5546 				    pp->p_szc != pszc) {
5547 					/*
5548 					 * hat_page_demote() raced in with us.
5549 					 */
5550 					ASSERT(!IS_SWAPFSVP(curvp));
5551 					page_unlock(pp);
5552 					break;
5553 				}
5554 				ppa[nlocked] = pp;
5555 				nlocked++;
5556 			}
5557 		}
5558 
5559 		/*
5560 		 * If all constituent pages couldn't be locked,
5561 		 * unlock pages locked so far and skip to next page.
5562 		 */
5563 		if (nlocked != pages) {
5564 			for (i = 0; i < nlocked; i++)
5565 				page_unlock(ppa[i]);
5566 			va += pgsz;
5567 			lgrp_stat_add(from->lgrp_id, LGRP_PMM_FAIL_PGS,
5568 			    btop(pgsz));
5569 			continue;
5570 		}
5571 
5572 		/*
5573 		 * hat_page_demote() can no longer happen
5574 		 * since last cons page had the right p_szc after
5575 		 * all cons pages were locked. all cons pages
5576 		 * should now have the same p_szc.
5577 		 */
5578 
5579 		/*
5580 		 * All constituent pages locked successfully, so mark
5581 		 * large page for migration and unload the mappings of
5582 		 * constituent pages, so a fault will occur on any part of the
5583 		 * large page
5584 		 */
5585 		PP_SETMIGRATE(ppa[0]);
5586 		for (i = 0; i < nlocked; i++) {
5587 			pp = ppa[i];
5588 			(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
5589 			ASSERT(hat_page_getshare(pp) == 0);
5590 			page_unlock(pp);
5591 		}
5592 		lgrp_stat_add(from->lgrp_id, LGRP_PMM_PGS, nlocked);
5593 
5594 		va += pgsz;
5595 	}
5596 	kmem_free(ppa, ppa_nentries * sizeof (page_t *));
5597 }
5598 
5599 /*
5600  * Migrate any pages that have been marked for migration in the given range
5601  */
5602 void
5603 page_migrate(
5604 	struct seg	*seg,
5605 	caddr_t		addr,
5606 	page_t		**ppa,
5607 	pgcnt_t		npages)
5608 {
5609 	lgrp_t		*from;
5610 	lgrp_t		*to;
5611 	page_t		*newpp;
5612 	page_t		*pp;
5613 	pfn_t		pfn;
5614 	size_t		pgsz;
5615 	spgcnt_t	page_cnt;
5616 	spgcnt_t	i;
5617 	uint_t		pszc;
5618 
5619 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5620 
5621 	while (npages > 0) {
5622 		pp = *ppa;
5623 		pszc = pp->p_szc;
5624 		pgsz = page_get_pagesize(pszc);
5625 		page_cnt = btop(pgsz);
5626 
5627 		/*
5628 		 * Check to see whether this page is marked for migration
5629 		 *
5630 		 * Assume that root page of large page is marked for
5631 		 * migration and none of the other constituent pages
5632 		 * are marked.  This really simplifies clearing the
5633 		 * migrate bit by not having to clear it from each
5634 		 * constituent page.
5635 		 *
5636 		 * note we don't want to relocate an entire large page if
5637 		 * someone is only using one subpage.
5638 		 */
5639 		if (npages < page_cnt)
5640 			break;
5641 
5642 		/*
5643 		 * Is it marked for migration?
5644 		 */
5645 		if (!PP_ISMIGRATE(pp))
5646 			goto next;
5647 
5648 		/*
5649 		 * Determine lgroups that page is being migrated between
5650 		 */
5651 		pfn = page_pptonum(pp);
5652 		if (!IS_P2ALIGNED(pfn, page_cnt)) {
5653 			break;
5654 		}
5655 		from = lgrp_pfn_to_lgrp(pfn);
5656 		to = lgrp_mem_choose(seg, addr, pgsz);
5657 
5658 		/*
5659 		 * Check to see whether we are trying to migrate page to lgroup
5660 		 * where it is allocated already
5661 		 */
5662 		if (to == from) {
5663 			PP_CLRMIGRATE(pp);
5664 			goto next;
5665 		}
5666 
5667 		/*
5668 		 * Need to get exclusive lock's to migrate
5669 		 */
5670 		for (i = 0; i < page_cnt; i++) {
5671 			ASSERT(PAGE_LOCKED(ppa[i]));
5672 			if (page_pptonum(ppa[i]) != pfn + i ||
5673 			    ppa[i]->p_szc != pszc) {
5674 				break;
5675 			}
5676 			if (!page_tryupgrade(ppa[i])) {
5677 				lgrp_stat_add(from->lgrp_id,
5678 				    LGRP_PM_FAIL_LOCK_PGS,
5679 				    page_cnt);
5680 				break;
5681 			}
5682 		}
5683 		if (i != page_cnt) {
5684 			while (--i != -1) {
5685 				page_downgrade(ppa[i]);
5686 			}
5687 			goto next;
5688 		}
5689 
5690 		(void) page_create_wait(page_cnt, PG_WAIT);
5691 		newpp = page_get_replacement_page(pp, to, PGR_SAMESZC);
5692 		if (newpp == NULL) {
5693 			page_create_putback(page_cnt);
5694 			for (i = 0; i < page_cnt; i++) {
5695 				page_downgrade(ppa[i]);
5696 			}
5697 			lgrp_stat_add(to->lgrp_id, LGRP_PM_FAIL_ALLOC_PGS,
5698 			    page_cnt);
5699 			goto next;
5700 		}
5701 		ASSERT(newpp->p_szc == pszc);
5702 		/*
5703 		 * Clear migrate bit and relocate page
5704 		 */
5705 		PP_CLRMIGRATE(pp);
5706 		if (page_relocate(&pp, &newpp, 0, 1, &page_cnt, to)) {
5707 			panic("page_migrate: page_relocate failed");
5708 		}
5709 		ASSERT(page_cnt * PAGESIZE == pgsz);
5710 
5711 		/*
5712 		 * Keep stats for number of pages migrated from and to
5713 		 * each lgroup
5714 		 */
5715 		lgrp_stat_add(from->lgrp_id, LGRP_PM_SRC_PGS, page_cnt);
5716 		lgrp_stat_add(to->lgrp_id, LGRP_PM_DEST_PGS, page_cnt);
5717 		/*
5718 		 * update the page_t array we were passed in and
5719 		 * unlink constituent pages of a large page.
5720 		 */
5721 		for (i = 0; i < page_cnt; ++i, ++pp) {
5722 			ASSERT(PAGE_EXCL(newpp));
5723 			ASSERT(newpp->p_szc == pszc);
5724 			ppa[i] = newpp;
5725 			pp = newpp;
5726 			page_sub(&newpp, pp);
5727 			page_downgrade(pp);
5728 		}
5729 		ASSERT(newpp == NULL);
5730 next:
5731 		addr += pgsz;
5732 		ppa += page_cnt;
5733 		npages -= page_cnt;
5734 	}
5735 }
5736 
5737 ulong_t mem_waiters 	= 0;
5738 ulong_t	max_count 	= 20;
5739 #define	MAX_DELAY	0x1ff
5740 
5741 /*
5742  * Check if enough memory is available to proceed.
5743  * Depending on system configuration and how much memory is
5744  * reserved for swap we need to check against two variables.
5745  * e.g. on systems with little physical swap availrmem can be
5746  * more reliable indicator of how much memory is available.
5747  * On systems with large phys swap freemem can be better indicator.
5748  * If freemem drops below threshold level don't return an error
5749  * immediately but wake up pageout to free memory and block.
5750  * This is done number of times. If pageout is not able to free
5751  * memory within certain time return an error.
5752  * The same applies for availrmem but kmem_reap is used to
5753  * free memory.
5754  */
5755 int
5756 page_mem_avail(pgcnt_t npages)
5757 {
5758 	ulong_t count;
5759 
5760 #if defined(__i386)
5761 	if (freemem > desfree + npages &&
5762 	    availrmem > swapfs_reserve + npages &&
5763 	    btop(vmem_size(heap_arena, VMEM_FREE)) > tune.t_minarmem +
5764 	    npages)
5765 		return (1);
5766 #else
5767 	if (freemem > desfree + npages &&
5768 	    availrmem > swapfs_reserve + npages)
5769 		return (1);
5770 #endif
5771 
5772 	count = max_count;
5773 	atomic_add_long(&mem_waiters, 1);
5774 
5775 	while (freemem < desfree + npages && --count) {
5776 		cv_signal(&proc_pageout->p_cv);
5777 		if (delay_sig(hz + (mem_waiters & MAX_DELAY))) {
5778 			atomic_add_long(&mem_waiters, -1);
5779 			return (0);
5780 		}
5781 	}
5782 	if (count == 0) {
5783 		atomic_add_long(&mem_waiters, -1);
5784 		return (0);
5785 	}
5786 
5787 	count = max_count;
5788 	while (availrmem < swapfs_reserve + npages && --count) {
5789 		kmem_reap();
5790 		if (delay_sig(hz + (mem_waiters & MAX_DELAY))) {
5791 			atomic_add_long(&mem_waiters, -1);
5792 			return (0);
5793 		}
5794 	}
5795 	atomic_add_long(&mem_waiters, -1);
5796 	if (count == 0)
5797 		return (0);
5798 
5799 #if defined(__i386)
5800 	if (btop(vmem_size(heap_arena, VMEM_FREE)) <
5801 	    tune.t_minarmem + npages)
5802 		return (0);
5803 #endif
5804 	return (1);
5805 }
5806 
5807 #define	MAX_CNT	60	/* max num of iterations */
5808 /*
5809  * Reclaim/reserve availrmem for npages.
5810  * If there is not enough memory start reaping seg, kmem caches.
5811  * Start pageout scanner (via page_needfree()).
5812  * Exit after ~ MAX_CNT s regardless of how much memory has been released.
5813  * Note: There is no guarantee that any availrmem will be freed as
5814  * this memory typically is locked (kernel heap) or reserved for swap.
5815  * Also due to memory fragmentation kmem allocator may not be able
5816  * to free any memory (single user allocated buffer will prevent
5817  * freeing slab or a page).
5818  */
5819 int
5820 page_reclaim_mem(pgcnt_t npages, pgcnt_t epages, int adjust)
5821 {
5822 	int	i = 0;
5823 	int	ret = 0;
5824 	pgcnt_t	deficit;
5825 	pgcnt_t old_availrmem;
5826 
5827 	mutex_enter(&freemem_lock);
5828 	old_availrmem = availrmem - 1;
5829 	while ((availrmem < tune.t_minarmem + npages + epages) &&
5830 	    (old_availrmem < availrmem) && (i++ < MAX_CNT)) {
5831 		old_availrmem = availrmem;
5832 		deficit = tune.t_minarmem + npages + epages - availrmem;
5833 		mutex_exit(&freemem_lock);
5834 		page_needfree(deficit);
5835 		seg_preap();
5836 		kmem_reap();
5837 		delay(hz);
5838 		page_needfree(-(spgcnt_t)deficit);
5839 		mutex_enter(&freemem_lock);
5840 	}
5841 
5842 	if (adjust && (availrmem >= tune.t_minarmem + npages + epages)) {
5843 		availrmem -= npages;
5844 		ret = 1;
5845 	}
5846 
5847 	mutex_exit(&freemem_lock);
5848 
5849 	return (ret);
5850 }
5851 
5852 /*
5853  * Search the memory segments to locate the desired page.  Within a
5854  * segment, pages increase linearly with one page structure per
5855  * physical page frame (size PAGESIZE).  The search begins
5856  * with the segment that was accessed last, to take advantage of locality.
5857  * If the hint misses, we start from the beginning of the sorted memseg list
5858  */
5859 
5860 
5861 /*
5862  * Some data structures for pfn to pp lookup.
5863  */
5864 ulong_t mhash_per_slot;
5865 struct memseg *memseg_hash[N_MEM_SLOTS];
5866 
5867 page_t *
5868 page_numtopp_nolock(pfn_t pfnum)
5869 {
5870 	struct memseg *seg;
5871 	page_t *pp;
5872 	vm_cpu_data_t *vc = CPU->cpu_vm_data;
5873 
5874 	ASSERT(vc != NULL);
5875 
5876 	MEMSEG_STAT_INCR(nsearch);
5877 
5878 	/* Try last winner first */
5879 	if (((seg = vc->vc_pnum_memseg) != NULL) &&
5880 	    (pfnum >= seg->pages_base) && (pfnum < seg->pages_end)) {
5881 		MEMSEG_STAT_INCR(nlastwon);
5882 		pp = seg->pages + (pfnum - seg->pages_base);
5883 		if (pp->p_pagenum == pfnum)
5884 			return ((page_t *)pp);
5885 	}
5886 
5887 	/* Else Try hash */
5888 	if (((seg = memseg_hash[MEMSEG_PFN_HASH(pfnum)]) != NULL) &&
5889 	    (pfnum >= seg->pages_base) && (pfnum < seg->pages_end)) {
5890 		MEMSEG_STAT_INCR(nhashwon);
5891 		vc->vc_pnum_memseg = seg;
5892 		pp = seg->pages + (pfnum - seg->pages_base);
5893 		if (pp->p_pagenum == pfnum)
5894 			return ((page_t *)pp);
5895 	}
5896 
5897 	/* Else Brute force */
5898 	for (seg = memsegs; seg != NULL; seg = seg->next) {
5899 		if (pfnum >= seg->pages_base && pfnum < seg->pages_end) {
5900 			vc->vc_pnum_memseg = seg;
5901 			pp = seg->pages + (pfnum - seg->pages_base);
5902 			return ((page_t *)pp);
5903 		}
5904 	}
5905 	vc->vc_pnum_memseg = NULL;
5906 	MEMSEG_STAT_INCR(nnotfound);
5907 	return ((page_t *)NULL);
5908 
5909 }
5910 
5911 struct memseg *
5912 page_numtomemseg_nolock(pfn_t pfnum)
5913 {
5914 	struct memseg *seg;
5915 	page_t *pp;
5916 
5917 	/* Try hash */
5918 	if (((seg = memseg_hash[MEMSEG_PFN_HASH(pfnum)]) != NULL) &&
5919 	    (pfnum >= seg->pages_base) && (pfnum < seg->pages_end)) {
5920 		pp = seg->pages + (pfnum - seg->pages_base);
5921 		if (pp->p_pagenum == pfnum)
5922 			return (seg);
5923 	}
5924 
5925 	/* Else Brute force */
5926 	for (seg = memsegs; seg != NULL; seg = seg->next) {
5927 		if (pfnum >= seg->pages_base && pfnum < seg->pages_end) {
5928 			return (seg);
5929 		}
5930 	}
5931 	return ((struct memseg *)NULL);
5932 }
5933 
5934 /*
5935  * Given a page and a count return the page struct that is
5936  * n structs away from the current one in the global page
5937  * list.
5938  *
5939  * This function wraps to the first page upon
5940  * reaching the end of the memseg list.
5941  */
5942 page_t *
5943 page_nextn(page_t *pp, ulong_t n)
5944 {
5945 	struct memseg *seg;
5946 	page_t *ppn;
5947 	vm_cpu_data_t *vc = (vm_cpu_data_t *)CPU->cpu_vm_data;
5948 
5949 	ASSERT(vc != NULL);
5950 
5951 	if (((seg = vc->vc_pnext_memseg) == NULL) ||
5952 	    (seg->pages_base == seg->pages_end) ||
5953 	    !(pp >= seg->pages && pp < seg->epages)) {
5954 
5955 		for (seg = memsegs; seg; seg = seg->next) {
5956 			if (pp >= seg->pages && pp < seg->epages)
5957 				break;
5958 		}
5959 
5960 		if (seg == NULL) {
5961 			/* Memory delete got in, return something valid. */
5962 			/* TODO: fix me. */
5963 			seg = memsegs;
5964 			pp = seg->pages;
5965 		}
5966 	}
5967 
5968 	/* check for wraparound - possible if n is large */
5969 	while ((ppn = (pp + n)) >= seg->epages || ppn < pp) {
5970 		n -= seg->epages - pp;
5971 		seg = seg->next;
5972 		if (seg == NULL)
5973 			seg = memsegs;
5974 		pp = seg->pages;
5975 	}
5976 	vc->vc_pnext_memseg = seg;
5977 	return (ppn);
5978 }
5979 
5980 /*
5981  * Initialize for a loop using page_next_scan_large().
5982  */
5983 page_t *
5984 page_next_scan_init(void **cookie)
5985 {
5986 	ASSERT(cookie != NULL);
5987 	*cookie = (void *)memsegs;
5988 	return ((page_t *)memsegs->pages);
5989 }
5990 
5991 /*
5992  * Return the next page in a scan of page_t's, assuming we want
5993  * to skip over sub-pages within larger page sizes.
5994  *
5995  * The cookie is used to keep track of the current memseg.
5996  */
5997 page_t *
5998 page_next_scan_large(
5999 	page_t		*pp,
6000 	ulong_t		*n,
6001 	void		**cookie)
6002 {
6003 	struct memseg	*seg = (struct memseg *)*cookie;
6004 	page_t		*new_pp;
6005 	ulong_t		cnt;
6006 	pfn_t		pfn;
6007 
6008 
6009 	/*
6010 	 * get the count of page_t's to skip based on the page size
6011 	 */
6012 	ASSERT(pp != NULL);
6013 	if (pp->p_szc == 0) {
6014 		cnt = 1;
6015 	} else {
6016 		pfn = page_pptonum(pp);
6017 		cnt = page_get_pagecnt(pp->p_szc);
6018 		cnt -= pfn & (cnt - 1);
6019 	}
6020 	*n += cnt;
6021 	new_pp = pp + cnt;
6022 
6023 	/*
6024 	 * Catch if we went past the end of the current memory segment. If so,
6025 	 * just move to the next segment with pages.
6026 	 */
6027 	if (new_pp >= seg->epages) {
6028 		do {
6029 			seg = seg->next;
6030 			if (seg == NULL)
6031 				seg = memsegs;
6032 		} while (seg->pages == seg->epages);
6033 		new_pp = seg->pages;
6034 		*cookie = (void *)seg;
6035 	}
6036 
6037 	return (new_pp);
6038 }
6039 
6040 
6041 /*
6042  * Returns next page in list. Note: this function wraps
6043  * to the first page in the list upon reaching the end
6044  * of the list. Callers should be aware of this fact.
6045  */
6046 
6047 /* We should change this be a #define */
6048 
6049 page_t *
6050 page_next(page_t *pp)
6051 {
6052 	return (page_nextn(pp, 1));
6053 }
6054 
6055 page_t *
6056 page_first()
6057 {
6058 	return ((page_t *)memsegs->pages);
6059 }
6060 
6061 
6062 /*
6063  * This routine is called at boot with the initial memory configuration
6064  * and when memory is added or removed.
6065  */
6066 void
6067 build_pfn_hash()
6068 {
6069 	pfn_t cur;
6070 	pgcnt_t index;
6071 	struct memseg *pseg;
6072 	int	i;
6073 
6074 	/*
6075 	 * Clear memseg_hash array.
6076 	 * Since memory add/delete is designed to operate concurrently
6077 	 * with normal operation, the hash rebuild must be able to run
6078 	 * concurrently with page_numtopp_nolock(). To support this
6079 	 * functionality, assignments to memseg_hash array members must
6080 	 * be done atomically.
6081 	 *
6082 	 * NOTE: bzero() does not currently guarantee this for kernel
6083 	 * threads, and cannot be used here.
6084 	 */
6085 	for (i = 0; i < N_MEM_SLOTS; i++)
6086 		memseg_hash[i] = NULL;
6087 
6088 	hat_kpm_mseghash_clear(N_MEM_SLOTS);
6089 
6090 	/*
6091 	 * Physmax is the last valid pfn.
6092 	 */
6093 	mhash_per_slot = (physmax + 1) >> MEM_HASH_SHIFT;
6094 	for (pseg = memsegs; pseg != NULL; pseg = pseg->next) {
6095 		index = MEMSEG_PFN_HASH(pseg->pages_base);
6096 		cur = pseg->pages_base;
6097 		do {
6098 			if (index >= N_MEM_SLOTS)
6099 				index = MEMSEG_PFN_HASH(cur);
6100 
6101 			if (memseg_hash[index] == NULL ||
6102 			    memseg_hash[index]->pages_base > pseg->pages_base) {
6103 				memseg_hash[index] = pseg;
6104 				hat_kpm_mseghash_update(index, pseg);
6105 			}
6106 			cur += mhash_per_slot;
6107 			index++;
6108 		} while (cur < pseg->pages_end);
6109 	}
6110 }
6111 
6112 /*
6113  * Return the pagenum for the pp
6114  */
6115 pfn_t
6116 page_pptonum(page_t *pp)
6117 {
6118 	return (pp->p_pagenum);
6119 }
6120 
6121 /*
6122  * interface to the referenced and modified etc bits
6123  * in the PSM part of the page struct
6124  * when no locking is desired.
6125  */
6126 void
6127 page_set_props(page_t *pp, uint_t flags)
6128 {
6129 	ASSERT((flags & ~(P_MOD | P_REF | P_RO)) == 0);
6130 	pp->p_nrm |= (uchar_t)flags;
6131 }
6132 
6133 void
6134 page_clr_all_props(page_t *pp)
6135 {
6136 	pp->p_nrm = 0;
6137 }
6138 
6139 /*
6140  * Clear p_lckcnt and p_cowcnt, adjusting freemem if required.
6141  */
6142 int
6143 page_clear_lck_cow(page_t *pp, int adjust)
6144 {
6145 	int	f_amount;
6146 
6147 	ASSERT(PAGE_EXCL(pp));
6148 
6149 	/*
6150 	 * The page_struct_lock need not be acquired here since
6151 	 * we require the caller hold the page exclusively locked.
6152 	 */
6153 	f_amount = 0;
6154 	if (pp->p_lckcnt) {
6155 		f_amount = 1;
6156 		pp->p_lckcnt = 0;
6157 	}
6158 	if (pp->p_cowcnt) {
6159 		f_amount += pp->p_cowcnt;
6160 		pp->p_cowcnt = 0;
6161 	}
6162 
6163 	if (adjust && f_amount) {
6164 		mutex_enter(&freemem_lock);
6165 		availrmem += f_amount;
6166 		mutex_exit(&freemem_lock);
6167 	}
6168 
6169 	return (f_amount);
6170 }
6171 
6172 /*
6173  * The following functions is called from free_vp_pages()
6174  * for an inexact estimate of a newly free'd page...
6175  */
6176 ulong_t
6177 page_share_cnt(page_t *pp)
6178 {
6179 	return (hat_page_getshare(pp));
6180 }
6181 
6182 int
6183 page_isshared(page_t *pp)
6184 {
6185 	return (hat_page_checkshare(pp, 1));
6186 }
6187 
6188 int
6189 page_isfree(page_t *pp)
6190 {
6191 	return (PP_ISFREE(pp));
6192 }
6193 
6194 int
6195 page_isref(page_t *pp)
6196 {
6197 	return (hat_page_getattr(pp, P_REF));
6198 }
6199 
6200 int
6201 page_ismod(page_t *pp)
6202 {
6203 	return (hat_page_getattr(pp, P_MOD));
6204 }
6205 
6206 /*
6207  * The following code all currently relates to the page capture logic:
6208  *
6209  * This logic is used for cases where there is a desire to claim a certain
6210  * physical page in the system for the caller.  As it may not be possible
6211  * to capture the page immediately, the p_toxic bits are used in the page
6212  * structure to indicate that someone wants to capture this page.  When the
6213  * page gets unlocked, the toxic flag will be noted and an attempt to capture
6214  * the page will be made.  If it is successful, the original callers callback
6215  * will be called with the page to do with it what they please.
6216  *
6217  * There is also an async thread which wakes up to attempt to capture
6218  * pages occasionally which have the capture bit set.  All of the pages which
6219  * need to be captured asynchronously have been inserted into the
6220  * page_capture_hash and thus this thread walks that hash list.  Items in the
6221  * hash have an expiration time so this thread handles that as well by removing
6222  * the item from the hash if it has expired.
6223  *
6224  * Some important things to note are:
6225  * - if the PR_CAPTURE bit is set on a page, then the page is in the
6226  *   page_capture_hash.  The page_capture_hash_head.pchh_mutex is needed
6227  *   to set and clear this bit, and while the lock is held is the only time
6228  *   you can add or remove an entry from the hash.
6229  * - the PR_CAPTURE bit can only be set and cleared while holding the
6230  *   page_capture_hash_head.pchh_mutex
6231  * - the t_flag field of the thread struct is used with the T_CAPTURING
6232  *   flag to prevent recursion while dealing with large pages.
6233  * - pages which need to be retired never expire on the page_capture_hash.
6234  */
6235 
6236 static void page_capture_thread(void);
6237 static kthread_t *pc_thread_id;
6238 kcondvar_t pc_cv;
6239 static kmutex_t pc_thread_mutex;
6240 static clock_t pc_thread_shortwait;
6241 static clock_t pc_thread_longwait;
6242 static int pc_thread_ism_retry;
6243 
6244 struct page_capture_callback pc_cb[PC_NUM_CALLBACKS];
6245 
6246 /* Note that this is a circular linked list */
6247 typedef struct page_capture_hash_bucket {
6248 	page_t *pp;
6249 	uint_t szc;
6250 	uint_t flags;
6251 	clock_t expires;	/* lbolt at which this request expires. */
6252 	void *datap;		/* Cached data passed in for callback */
6253 	struct page_capture_hash_bucket *next;
6254 	struct page_capture_hash_bucket *prev;
6255 } page_capture_hash_bucket_t;
6256 
6257 /*
6258  * Each hash bucket will have it's own mutex and two lists which are:
6259  * active (0):	represents requests which have not been processed by
6260  *		the page_capture async thread yet.
6261  * walked (1):	represents requests which have been processed by the
6262  *		page_capture async thread within it's given walk of this bucket.
6263  *
6264  * These are all needed so that we can synchronize all async page_capture
6265  * events.  When the async thread moves to a new bucket, it will append the
6266  * walked list to the active list and walk each item one at a time, moving it
6267  * from the active list to the walked list.  Thus if there is an async request
6268  * outstanding for a given page, it will always be in one of the two lists.
6269  * New requests will always be added to the active list.
6270  * If we were not able to capture a page before the request expired, we'd free
6271  * up the request structure which would indicate to page_capture that there is
6272  * no longer a need for the given page, and clear the PR_CAPTURE flag if
6273  * possible.
6274  */
6275 typedef struct page_capture_hash_head {
6276 	kmutex_t pchh_mutex;
6277 	uint_t num_pages;
6278 	page_capture_hash_bucket_t lists[2]; /* sentinel nodes */
6279 } page_capture_hash_head_t;
6280 
6281 #ifdef DEBUG
6282 #define	NUM_PAGE_CAPTURE_BUCKETS 4
6283 #else
6284 #define	NUM_PAGE_CAPTURE_BUCKETS 64
6285 #endif
6286 
6287 page_capture_hash_head_t page_capture_hash[NUM_PAGE_CAPTURE_BUCKETS];
6288 
6289 /* for now use a very simple hash based upon the size of a page struct */
6290 #define	PAGE_CAPTURE_HASH(pp)	\
6291 	((int)(((uintptr_t)pp >> 7) & (NUM_PAGE_CAPTURE_BUCKETS - 1)))
6292 
6293 extern pgcnt_t swapfs_minfree;
6294 
6295 int page_trycapture(page_t *pp, uint_t szc, uint_t flags, void *datap);
6296 
6297 /*
6298  * a callback function is required for page capture requests.
6299  */
6300 void
6301 page_capture_register_callback(uint_t index, clock_t duration,
6302     int (*cb_func)(page_t *, void *, uint_t))
6303 {
6304 	ASSERT(pc_cb[index].cb_active == 0);
6305 	ASSERT(cb_func != NULL);
6306 	rw_enter(&pc_cb[index].cb_rwlock, RW_WRITER);
6307 	pc_cb[index].duration = duration;
6308 	pc_cb[index].cb_func = cb_func;
6309 	pc_cb[index].cb_active = 1;
6310 	rw_exit(&pc_cb[index].cb_rwlock);
6311 }
6312 
6313 void
6314 page_capture_unregister_callback(uint_t index)
6315 {
6316 	int i, j;
6317 	struct page_capture_hash_bucket *bp1;
6318 	struct page_capture_hash_bucket *bp2;
6319 	struct page_capture_hash_bucket *head = NULL;
6320 	uint_t flags = (1 << index);
6321 
6322 	rw_enter(&pc_cb[index].cb_rwlock, RW_WRITER);
6323 	ASSERT(pc_cb[index].cb_active == 1);
6324 	pc_cb[index].duration = 0;	/* Paranoia */
6325 	pc_cb[index].cb_func = NULL;	/* Paranoia */
6326 	pc_cb[index].cb_active = 0;
6327 	rw_exit(&pc_cb[index].cb_rwlock);
6328 
6329 	/*
6330 	 * Just move all the entries to a private list which we can walk
6331 	 * through without the need to hold any locks.
6332 	 * No more requests can get added to the hash lists for this consumer
6333 	 * as the cb_active field for the callback has been cleared.
6334 	 */
6335 	for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) {
6336 		mutex_enter(&page_capture_hash[i].pchh_mutex);
6337 		for (j = 0; j < 2; j++) {
6338 			bp1 = page_capture_hash[i].lists[j].next;
6339 			/* walk through all but first (sentinel) element */
6340 			while (bp1 != &page_capture_hash[i].lists[j]) {
6341 				bp2 = bp1;
6342 				if (bp2->flags & flags) {
6343 					bp1 = bp2->next;
6344 					bp1->prev = bp2->prev;
6345 					bp2->prev->next = bp1;
6346 					bp2->next = head;
6347 					head = bp2;
6348 					/*
6349 					 * Clear the PR_CAPTURE bit as we
6350 					 * hold appropriate locks here.
6351 					 */
6352 					page_clrtoxic(head->pp, PR_CAPTURE);
6353 					page_capture_hash[i].num_pages--;
6354 					continue;
6355 				}
6356 				bp1 = bp1->next;
6357 			}
6358 		}
6359 		mutex_exit(&page_capture_hash[i].pchh_mutex);
6360 	}
6361 
6362 	while (head != NULL) {
6363 		bp1 = head;
6364 		head = head->next;
6365 		kmem_free(bp1, sizeof (*bp1));
6366 	}
6367 }
6368 
6369 
6370 /*
6371  * Find pp in the active list and move it to the walked list if it
6372  * exists.
6373  * Note that most often pp should be at the front of the active list
6374  * as it is currently used and thus there is no other sort of optimization
6375  * being done here as this is a linked list data structure.
6376  * Returns 1 on successful move or 0 if page could not be found.
6377  */
6378 static int
6379 page_capture_move_to_walked(page_t *pp)
6380 {
6381 	page_capture_hash_bucket_t *bp;
6382 	int index;
6383 
6384 	index = PAGE_CAPTURE_HASH(pp);
6385 
6386 	mutex_enter(&page_capture_hash[index].pchh_mutex);
6387 	bp = page_capture_hash[index].lists[0].next;
6388 	while (bp != &page_capture_hash[index].lists[0]) {
6389 		if (bp->pp == pp) {
6390 			/* Remove from old list */
6391 			bp->next->prev = bp->prev;
6392 			bp->prev->next = bp->next;
6393 
6394 			/* Add to new list */
6395 			bp->next = page_capture_hash[index].lists[1].next;
6396 			bp->prev = &page_capture_hash[index].lists[1];
6397 			page_capture_hash[index].lists[1].next = bp;
6398 			bp->next->prev = bp;
6399 			mutex_exit(&page_capture_hash[index].pchh_mutex);
6400 
6401 			return (1);
6402 		}
6403 		bp = bp->next;
6404 	}
6405 	mutex_exit(&page_capture_hash[index].pchh_mutex);
6406 	return (0);
6407 }
6408 
6409 /*
6410  * Add a new entry to the page capture hash.  The only case where a new
6411  * entry is not added is when the page capture consumer is no longer registered.
6412  * In this case, we'll silently not add the page to the hash.  We know that
6413  * page retire will always be registered for the case where we are currently
6414  * unretiring a page and thus there are no conflicts.
6415  */
6416 static void
6417 page_capture_add_hash(page_t *pp, uint_t szc, uint_t flags, void *datap)
6418 {
6419 	page_capture_hash_bucket_t *bp1;
6420 	page_capture_hash_bucket_t *bp2;
6421 	int index;
6422 	int cb_index;
6423 	int i;
6424 #ifdef DEBUG
6425 	page_capture_hash_bucket_t *tp1;
6426 	int l;
6427 #endif
6428 
6429 	ASSERT(!(flags & CAPTURE_ASYNC));
6430 
6431 	bp1 = kmem_alloc(sizeof (struct page_capture_hash_bucket), KM_SLEEP);
6432 
6433 	bp1->pp = pp;
6434 	bp1->szc = szc;
6435 	bp1->flags = flags;
6436 	bp1->datap = datap;
6437 
6438 	for (cb_index = 0; cb_index < PC_NUM_CALLBACKS; cb_index++) {
6439 		if ((flags >> cb_index) & 1) {
6440 			break;
6441 		}
6442 	}
6443 
6444 	ASSERT(cb_index != PC_NUM_CALLBACKS);
6445 
6446 	rw_enter(&pc_cb[cb_index].cb_rwlock, RW_READER);
6447 	if (pc_cb[cb_index].cb_active) {
6448 		if (pc_cb[cb_index].duration == -1) {
6449 			bp1->expires = (clock_t)-1;
6450 		} else {
6451 			bp1->expires = lbolt + pc_cb[cb_index].duration;
6452 		}
6453 	} else {
6454 		/* There's no callback registered so don't add to the hash */
6455 		rw_exit(&pc_cb[cb_index].cb_rwlock);
6456 		kmem_free(bp1, sizeof (*bp1));
6457 		return;
6458 	}
6459 
6460 	index = PAGE_CAPTURE_HASH(pp);
6461 
6462 	/*
6463 	 * Only allow capture flag to be modified under this mutex.
6464 	 * Prevents multiple entries for same page getting added.
6465 	 */
6466 	mutex_enter(&page_capture_hash[index].pchh_mutex);
6467 
6468 	/*
6469 	 * if not already on the hash, set capture bit and add to the hash
6470 	 */
6471 	if (!(pp->p_toxic & PR_CAPTURE)) {
6472 #ifdef DEBUG
6473 		/* Check for duplicate entries */
6474 		for (l = 0; l < 2; l++) {
6475 			tp1 = page_capture_hash[index].lists[l].next;
6476 			while (tp1 != &page_capture_hash[index].lists[l]) {
6477 				if (tp1->pp == pp) {
6478 					panic("page pp 0x%p already on hash "
6479 					    "at 0x%p\n", pp, tp1);
6480 				}
6481 				tp1 = tp1->next;
6482 			}
6483 		}
6484 
6485 #endif
6486 		page_settoxic(pp, PR_CAPTURE);
6487 		bp1->next = page_capture_hash[index].lists[0].next;
6488 		bp1->prev = &page_capture_hash[index].lists[0];
6489 		bp1->next->prev = bp1;
6490 		page_capture_hash[index].lists[0].next = bp1;
6491 		page_capture_hash[index].num_pages++;
6492 		if (flags & CAPTURE_RETIRE) {
6493 			page_retire_incr_pend_count();
6494 		}
6495 		mutex_exit(&page_capture_hash[index].pchh_mutex);
6496 		rw_exit(&pc_cb[cb_index].cb_rwlock);
6497 		cv_signal(&pc_cv);
6498 		return;
6499 	}
6500 
6501 	/*
6502 	 * A page retire request will replace any other request.
6503 	 * A second physmem request which is for a different process than
6504 	 * the currently registered one will be dropped as there is
6505 	 * no way to hold the private data for both calls.
6506 	 * In the future, once there are more callers, this will have to
6507 	 * be worked out better as there needs to be private storage for
6508 	 * at least each type of caller (maybe have datap be an array of
6509 	 * *void's so that we can index based upon callers index).
6510 	 */
6511 
6512 	/* walk hash list to update expire time */
6513 	for (i = 0; i < 2; i++) {
6514 		bp2 = page_capture_hash[index].lists[i].next;
6515 		while (bp2 != &page_capture_hash[index].lists[i]) {
6516 			if (bp2->pp == pp) {
6517 				if (flags & CAPTURE_RETIRE) {
6518 					if (!(bp2->flags & CAPTURE_RETIRE)) {
6519 						page_retire_incr_pend_count();
6520 						bp2->flags = flags;
6521 						bp2->expires = bp1->expires;
6522 						bp2->datap = datap;
6523 					}
6524 				} else {
6525 					ASSERT(flags & CAPTURE_PHYSMEM);
6526 					if (!(bp2->flags & CAPTURE_RETIRE) &&
6527 					    (datap == bp2->datap)) {
6528 						bp2->expires = bp1->expires;
6529 					}
6530 				}
6531 				mutex_exit(&page_capture_hash[index].
6532 				    pchh_mutex);
6533 				rw_exit(&pc_cb[cb_index].cb_rwlock);
6534 				kmem_free(bp1, sizeof (*bp1));
6535 				return;
6536 			}
6537 			bp2 = bp2->next;
6538 		}
6539 	}
6540 
6541 	/*
6542 	 * the PR_CAPTURE flag is protected by the page_capture_hash mutexes
6543 	 * and thus it either has to be set or not set and can't change
6544 	 * while holding the mutex above.
6545 	 */
6546 	panic("page_capture_add_hash, PR_CAPTURE flag set on pp %p\n", pp);
6547 }
6548 
6549 /*
6550  * We have a page in our hands, lets try and make it ours by turning
6551  * it into a clean page like it had just come off the freelists.
6552  *
6553  * Returns 0 on success, with the page still EXCL locked.
6554  * On failure, the page will be unlocked, and returns EAGAIN
6555  */
6556 static int
6557 page_capture_clean_page(page_t *pp)
6558 {
6559 	page_t *newpp;
6560 	int skip_unlock = 0;
6561 	spgcnt_t count;
6562 	page_t *tpp;
6563 	int ret = 0;
6564 	int extra;
6565 
6566 	ASSERT(PAGE_EXCL(pp));
6567 	ASSERT(!PP_RETIRED(pp));
6568 	ASSERT(curthread->t_flag & T_CAPTURING);
6569 
6570 	if (PP_ISFREE(pp)) {
6571 		if (!page_reclaim(pp, NULL)) {
6572 			skip_unlock = 1;
6573 			ret = EAGAIN;
6574 			goto cleanup;
6575 		}
6576 		ASSERT(pp->p_szc == 0);
6577 		if (pp->p_vnode != NULL) {
6578 			/*
6579 			 * Since this page came from the
6580 			 * cachelist, we must destroy the
6581 			 * old vnode association.
6582 			 */
6583 			page_hashout(pp, NULL);
6584 		}
6585 		goto cleanup;
6586 	}
6587 
6588 	/*
6589 	 * If we know page_relocate will fail, skip it
6590 	 * It could still fail due to a UE on another page but we
6591 	 * can't do anything about that.
6592 	 */
6593 	if (pp->p_toxic & PR_UE) {
6594 		goto skip_relocate;
6595 	}
6596 
6597 	/*
6598 	 * It's possible that pages can not have a vnode as fsflush comes
6599 	 * through and cleans up these pages.  It's ugly but that's how it is.
6600 	 */
6601 	if (pp->p_vnode == NULL) {
6602 		goto skip_relocate;
6603 	}
6604 
6605 	/*
6606 	 * Page was not free, so lets try to relocate it.
6607 	 * page_relocate only works with root pages, so if this is not a root
6608 	 * page, we need to demote it to try and relocate it.
6609 	 * Unfortunately this is the best we can do right now.
6610 	 */
6611 	newpp = NULL;
6612 	if ((pp->p_szc > 0) && (pp != PP_PAGEROOT(pp))) {
6613 		if (page_try_demote_pages(pp) == 0) {
6614 			ret = EAGAIN;
6615 			goto cleanup;
6616 		}
6617 	}
6618 	ret = page_relocate(&pp, &newpp, 1, 0, &count, NULL);
6619 	if (ret == 0) {
6620 		page_t *npp;
6621 		/* unlock the new page(s) */
6622 		while (count-- > 0) {
6623 			ASSERT(newpp != NULL);
6624 			npp = newpp;
6625 			page_sub(&newpp, npp);
6626 			page_unlock(npp);
6627 		}
6628 		ASSERT(newpp == NULL);
6629 		/*
6630 		 * Check to see if the page we have is too large.
6631 		 * If so, demote it freeing up the extra pages.
6632 		 */
6633 		if (pp->p_szc > 0) {
6634 			/* For now demote extra pages to szc == 0 */
6635 			extra = page_get_pagecnt(pp->p_szc) - 1;
6636 			while (extra > 0) {
6637 				tpp = pp->p_next;
6638 				page_sub(&pp, tpp);
6639 				tpp->p_szc = 0;
6640 				page_free(tpp, 1);
6641 				extra--;
6642 			}
6643 			/* Make sure to set our page to szc 0 as well */
6644 			ASSERT(pp->p_next == pp && pp->p_prev == pp);
6645 			pp->p_szc = 0;
6646 		}
6647 		goto cleanup;
6648 	} else if (ret == EIO) {
6649 		ret = EAGAIN;
6650 		goto cleanup;
6651 	} else {
6652 		/*
6653 		 * Need to reset return type as we failed to relocate the page
6654 		 * but that does not mean that some of the next steps will not
6655 		 * work.
6656 		 */
6657 		ret = 0;
6658 	}
6659 
6660 skip_relocate:
6661 
6662 	if (pp->p_szc > 0) {
6663 		if (page_try_demote_pages(pp) == 0) {
6664 			ret = EAGAIN;
6665 			goto cleanup;
6666 		}
6667 	}
6668 
6669 	ASSERT(pp->p_szc == 0);
6670 
6671 	if (hat_ismod(pp)) {
6672 		ret = EAGAIN;
6673 		goto cleanup;
6674 	}
6675 	if (PP_ISKAS(pp)) {
6676 		ret = EAGAIN;
6677 		goto cleanup;
6678 	}
6679 	if (pp->p_lckcnt || pp->p_cowcnt) {
6680 		ret = EAGAIN;
6681 		goto cleanup;
6682 	}
6683 
6684 	(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
6685 	ASSERT(!hat_page_is_mapped(pp));
6686 
6687 	if (hat_ismod(pp)) {
6688 		/*
6689 		 * This is a semi-odd case as the page is now modified but not
6690 		 * mapped as we just unloaded the mappings above.
6691 		 */
6692 		ret = EAGAIN;
6693 		goto cleanup;
6694 	}
6695 	if (pp->p_vnode != NULL) {
6696 		page_hashout(pp, NULL);
6697 	}
6698 
6699 	/*
6700 	 * At this point, the page should be in a clean state and
6701 	 * we can do whatever we want with it.
6702 	 */
6703 
6704 cleanup:
6705 	if (ret != 0) {
6706 		if (!skip_unlock) {
6707 			page_unlock(pp);
6708 		}
6709 	} else {
6710 		ASSERT(pp->p_szc == 0);
6711 		ASSERT(PAGE_EXCL(pp));
6712 
6713 		pp->p_next = pp;
6714 		pp->p_prev = pp;
6715 	}
6716 	return (ret);
6717 }
6718 
6719 /*
6720  * Various callers of page_trycapture() can have different restrictions upon
6721  * what memory they have access to.
6722  * Returns 0 on success, with the following error codes on failure:
6723  *      EPERM - The requested page is long term locked, and thus repeated
6724  *              requests to capture this page will likely fail.
6725  *      ENOMEM - There was not enough free memory in the system to safely
6726  *              map the requested page.
6727  *      ENOENT - The requested page was inside the kernel cage, and the
6728  *              PHYSMEM_CAGE flag was not set.
6729  */
6730 int
6731 page_capture_pre_checks(page_t *pp, uint_t flags)
6732 {
6733 #if defined(__sparc)
6734 	extern struct vnode prom_ppages;
6735 #endif /* __sparc */
6736 
6737 	ASSERT(pp != NULL);
6738 
6739 	/* only physmem currently has restrictions */
6740 	if (!(flags & CAPTURE_PHYSMEM)) {
6741 		return (0);
6742 	}
6743 
6744 #if defined(__sparc)
6745 	if (pp->p_vnode == &prom_ppages) {
6746 		return (EPERM);
6747 	}
6748 
6749 	if (PP_ISNORELOC(pp) && !(flags & CAPTURE_GET_CAGE)) {
6750 		return (ENOENT);
6751 	}
6752 
6753 	if (PP_ISNORELOCKERNEL(pp)) {
6754 		return (EPERM);
6755 	}
6756 #else
6757 	if (PP_ISKAS(pp)) {
6758 		return (EPERM);
6759 	}
6760 #endif /* __sparc */
6761 
6762 	if (availrmem < swapfs_minfree) {
6763 		/*
6764 		 * We won't try to capture this page as we are
6765 		 * running low on memory.
6766 		 */
6767 		return (ENOMEM);
6768 	}
6769 	return (0);
6770 }
6771 
6772 /*
6773  * Once we have a page in our mits, go ahead and complete the capture
6774  * operation.
6775  * Returns 1 on failure where page is no longer needed
6776  * Returns 0 on success
6777  * Returns -1 if there was a transient failure.
6778  * Failure cases must release the SE_EXCL lock on pp (usually via page_free).
6779  */
6780 int
6781 page_capture_take_action(page_t *pp, uint_t flags, void *datap)
6782 {
6783 	int cb_index;
6784 	int ret = 0;
6785 	page_capture_hash_bucket_t *bp1;
6786 	page_capture_hash_bucket_t *bp2;
6787 	int index;
6788 	int found = 0;
6789 	int i;
6790 
6791 	ASSERT(PAGE_EXCL(pp));
6792 	ASSERT(curthread->t_flag & T_CAPTURING);
6793 
6794 	for (cb_index = 0; cb_index < PC_NUM_CALLBACKS; cb_index++) {
6795 		if ((flags >> cb_index) & 1) {
6796 			break;
6797 		}
6798 	}
6799 	ASSERT(cb_index < PC_NUM_CALLBACKS);
6800 
6801 	/*
6802 	 * Remove the entry from the page_capture hash, but don't free it yet
6803 	 * as we may need to put it back.
6804 	 * Since we own the page at this point in time, we should find it
6805 	 * in the hash if this is an ASYNC call.  If we don't it's likely
6806 	 * that the page_capture_async() thread decided that this request
6807 	 * had expired, in which case we just continue on.
6808 	 */
6809 	if (flags & CAPTURE_ASYNC) {
6810 
6811 		index = PAGE_CAPTURE_HASH(pp);
6812 
6813 		mutex_enter(&page_capture_hash[index].pchh_mutex);
6814 		for (i = 0; i < 2 && !found; i++) {
6815 			bp1 = page_capture_hash[index].lists[i].next;
6816 			while (bp1 != &page_capture_hash[index].lists[i]) {
6817 				if (bp1->pp == pp) {
6818 					bp1->next->prev = bp1->prev;
6819 					bp1->prev->next = bp1->next;
6820 					page_capture_hash[index].num_pages--;
6821 					page_clrtoxic(pp, PR_CAPTURE);
6822 					found = 1;
6823 					break;
6824 				}
6825 				bp1 = bp1->next;
6826 			}
6827 		}
6828 		mutex_exit(&page_capture_hash[index].pchh_mutex);
6829 	}
6830 
6831 	/* Synchronize with the unregister func. */
6832 	rw_enter(&pc_cb[cb_index].cb_rwlock, RW_READER);
6833 	if (!pc_cb[cb_index].cb_active) {
6834 		page_free(pp, 1);
6835 		rw_exit(&pc_cb[cb_index].cb_rwlock);
6836 		if (found) {
6837 			kmem_free(bp1, sizeof (*bp1));
6838 		}
6839 		return (1);
6840 	}
6841 
6842 	/*
6843 	 * We need to remove the entry from the page capture hash and turn off
6844 	 * the PR_CAPTURE bit before calling the callback.  We'll need to cache
6845 	 * the entry here, and then based upon the return value, cleanup
6846 	 * appropriately or re-add it to the hash, making sure that someone else
6847 	 * hasn't already done so.
6848 	 * It should be rare for the callback to fail and thus it's ok for
6849 	 * the failure path to be a bit complicated as the success path is
6850 	 * cleaner and the locking rules are easier to follow.
6851 	 */
6852 
6853 	ret = pc_cb[cb_index].cb_func(pp, datap, flags);
6854 
6855 	rw_exit(&pc_cb[cb_index].cb_rwlock);
6856 
6857 	/*
6858 	 * If this was an ASYNC request, we need to cleanup the hash if the
6859 	 * callback was successful or if the request was no longer valid.
6860 	 * For non-ASYNC requests, we return failure to map and the caller
6861 	 * will take care of adding the request to the hash.
6862 	 * Note also that the callback itself is responsible for the page
6863 	 * at this point in time in terms of locking ...  The most common
6864 	 * case for the failure path should just be a page_free.
6865 	 */
6866 	if (ret >= 0) {
6867 		if (found) {
6868 			if (bp1->flags & CAPTURE_RETIRE) {
6869 				page_retire_decr_pend_count();
6870 			}
6871 			kmem_free(bp1, sizeof (*bp1));
6872 		}
6873 		return (ret);
6874 	}
6875 	if (!found) {
6876 		return (ret);
6877 	}
6878 
6879 	ASSERT(flags & CAPTURE_ASYNC);
6880 
6881 	/*
6882 	 * Check for expiration time first as we can just free it up if it's
6883 	 * expired.
6884 	 */
6885 	if (lbolt > bp1->expires && bp1->expires != -1) {
6886 		kmem_free(bp1, sizeof (*bp1));
6887 		return (ret);
6888 	}
6889 
6890 	/*
6891 	 * The callback failed and there used to be an entry in the hash for
6892 	 * this page, so we need to add it back to the hash.
6893 	 */
6894 	mutex_enter(&page_capture_hash[index].pchh_mutex);
6895 	if (!(pp->p_toxic & PR_CAPTURE)) {
6896 		/* just add bp1 back to head of walked list */
6897 		page_settoxic(pp, PR_CAPTURE);
6898 		bp1->next = page_capture_hash[index].lists[1].next;
6899 		bp1->prev = &page_capture_hash[index].lists[1];
6900 		bp1->next->prev = bp1;
6901 		page_capture_hash[index].lists[1].next = bp1;
6902 		page_capture_hash[index].num_pages++;
6903 		mutex_exit(&page_capture_hash[index].pchh_mutex);
6904 		return (ret);
6905 	}
6906 
6907 	/*
6908 	 * Otherwise there was a new capture request added to list
6909 	 * Need to make sure that our original data is represented if
6910 	 * appropriate.
6911 	 */
6912 	for (i = 0; i < 2; i++) {
6913 		bp2 = page_capture_hash[index].lists[i].next;
6914 		while (bp2 != &page_capture_hash[index].lists[i]) {
6915 			if (bp2->pp == pp) {
6916 				if (bp1->flags & CAPTURE_RETIRE) {
6917 					if (!(bp2->flags & CAPTURE_RETIRE)) {
6918 						bp2->szc = bp1->szc;
6919 						bp2->flags = bp1->flags;
6920 						bp2->expires = bp1->expires;
6921 						bp2->datap = bp1->datap;
6922 					}
6923 				} else {
6924 					ASSERT(bp1->flags & CAPTURE_PHYSMEM);
6925 					if (!(bp2->flags & CAPTURE_RETIRE)) {
6926 						bp2->szc = bp1->szc;
6927 						bp2->flags = bp1->flags;
6928 						bp2->expires = bp1->expires;
6929 						bp2->datap = bp1->datap;
6930 					}
6931 				}
6932 				mutex_exit(&page_capture_hash[index].
6933 				    pchh_mutex);
6934 				kmem_free(bp1, sizeof (*bp1));
6935 				return (ret);
6936 			}
6937 			bp2 = bp2->next;
6938 		}
6939 	}
6940 	panic("PR_CAPTURE set but not on hash for pp 0x%p\n", pp);
6941 	/*NOTREACHED*/
6942 }
6943 
6944 /*
6945  * Try to capture the given page for the caller specified in the flags
6946  * parameter.  The page will either be captured and handed over to the
6947  * appropriate callback, or will be queued up in the page capture hash
6948  * to be captured asynchronously.
6949  * If the current request is due to an async capture, the page must be
6950  * exclusively locked before calling this function.
6951  * Currently szc must be 0 but in the future this should be expandable to
6952  * other page sizes.
6953  * Returns 0 on success, with the following error codes on failure:
6954  *      EPERM - The requested page is long term locked, and thus repeated
6955  *              requests to capture this page will likely fail.
6956  *      ENOMEM - There was not enough free memory in the system to safely
6957  *              map the requested page.
6958  *      ENOENT - The requested page was inside the kernel cage, and the
6959  *              CAPTURE_GET_CAGE flag was not set.
6960  *	EAGAIN - The requested page could not be capturead at this point in
6961  *		time but future requests will likely work.
6962  *	EBUSY - The requested page is retired and the CAPTURE_GET_RETIRED flag
6963  *		was not set.
6964  */
6965 int
6966 page_itrycapture(page_t *pp, uint_t szc, uint_t flags, void *datap)
6967 {
6968 	int ret;
6969 	int cb_index;
6970 
6971 	if (flags & CAPTURE_ASYNC) {
6972 		ASSERT(PAGE_EXCL(pp));
6973 		goto async;
6974 	}
6975 
6976 	/* Make sure there's enough availrmem ... */
6977 	ret = page_capture_pre_checks(pp, flags);
6978 	if (ret != 0) {
6979 		return (ret);
6980 	}
6981 
6982 	if (!page_trylock(pp, SE_EXCL)) {
6983 		for (cb_index = 0; cb_index < PC_NUM_CALLBACKS; cb_index++) {
6984 			if ((flags >> cb_index) & 1) {
6985 				break;
6986 			}
6987 		}
6988 		ASSERT(cb_index < PC_NUM_CALLBACKS);
6989 		ret = EAGAIN;
6990 		/* Special case for retired pages */
6991 		if (PP_RETIRED(pp)) {
6992 			if (flags & CAPTURE_GET_RETIRED) {
6993 				if (!page_unretire_pp(pp, PR_UNR_TEMP)) {
6994 					/*
6995 					 * Need to set capture bit and add to
6996 					 * hash so that the page will be
6997 					 * retired when freed.
6998 					 */
6999 					page_capture_add_hash(pp, szc,
7000 					    CAPTURE_RETIRE, NULL);
7001 					ret = 0;
7002 					goto own_page;
7003 				}
7004 			} else {
7005 				return (EBUSY);
7006 			}
7007 		}
7008 		page_capture_add_hash(pp, szc, flags, datap);
7009 		return (ret);
7010 	}
7011 
7012 async:
7013 	ASSERT(PAGE_EXCL(pp));
7014 
7015 	/* Need to check for physmem async requests that availrmem is sane */
7016 	if ((flags & (CAPTURE_ASYNC | CAPTURE_PHYSMEM)) ==
7017 	    (CAPTURE_ASYNC | CAPTURE_PHYSMEM) &&
7018 	    (availrmem < swapfs_minfree)) {
7019 		page_unlock(pp);
7020 		return (ENOMEM);
7021 	}
7022 
7023 	ret = page_capture_clean_page(pp);
7024 
7025 	if (ret != 0) {
7026 		/* We failed to get the page, so lets add it to the hash */
7027 		if (!(flags & CAPTURE_ASYNC)) {
7028 			page_capture_add_hash(pp, szc, flags, datap);
7029 		}
7030 		return (ret);
7031 	}
7032 
7033 own_page:
7034 	ASSERT(PAGE_EXCL(pp));
7035 	ASSERT(pp->p_szc == 0);
7036 
7037 	/* Call the callback */
7038 	ret = page_capture_take_action(pp, flags, datap);
7039 
7040 	if (ret == 0) {
7041 		return (0);
7042 	}
7043 
7044 	/*
7045 	 * Note that in the failure cases from page_capture_take_action, the
7046 	 * EXCL lock will have already been dropped.
7047 	 */
7048 	if ((ret == -1) && (!(flags & CAPTURE_ASYNC))) {
7049 		page_capture_add_hash(pp, szc, flags, datap);
7050 	}
7051 	return (EAGAIN);
7052 }
7053 
7054 int
7055 page_trycapture(page_t *pp, uint_t szc, uint_t flags, void *datap)
7056 {
7057 	int ret;
7058 
7059 	curthread->t_flag |= T_CAPTURING;
7060 	ret = page_itrycapture(pp, szc, flags, datap);
7061 	curthread->t_flag &= ~T_CAPTURING; /* xor works as we know its set */
7062 	return (ret);
7063 }
7064 
7065 /*
7066  * When unlocking a page which has the PR_CAPTURE bit set, this routine
7067  * gets called to try and capture the page.
7068  */
7069 void
7070 page_unlock_capture(page_t *pp)
7071 {
7072 	page_capture_hash_bucket_t *bp;
7073 	int index;
7074 	int i;
7075 	uint_t szc;
7076 	uint_t flags = 0;
7077 	void *datap;
7078 	kmutex_t *mp;
7079 	extern vnode_t retired_pages;
7080 
7081 	/*
7082 	 * We need to protect against a possible deadlock here where we own
7083 	 * the vnode page hash mutex and want to acquire it again as there
7084 	 * are locations in the code, where we unlock a page while holding
7085 	 * the mutex which can lead to the page being captured and eventually
7086 	 * end up here.  As we may be hashing out the old page and hashing into
7087 	 * the retire vnode, we need to make sure we don't own them.
7088 	 * Other callbacks who do hash operations also need to make sure that
7089 	 * before they hashin to a vnode that they do not currently own the
7090 	 * vphm mutex otherwise there will be a panic.
7091 	 */
7092 	if (mutex_owned(page_vnode_mutex(&retired_pages))) {
7093 		page_unlock_nocapture(pp);
7094 		return;
7095 	}
7096 	if (pp->p_vnode != NULL && mutex_owned(page_vnode_mutex(pp->p_vnode))) {
7097 		page_unlock_nocapture(pp);
7098 		return;
7099 	}
7100 
7101 	index = PAGE_CAPTURE_HASH(pp);
7102 
7103 	mp = &page_capture_hash[index].pchh_mutex;
7104 	mutex_enter(mp);
7105 	for (i = 0; i < 2; i++) {
7106 		bp = page_capture_hash[index].lists[i].next;
7107 		while (bp != &page_capture_hash[index].lists[i]) {
7108 			if (bp->pp == pp) {
7109 				szc = bp->szc;
7110 				flags = bp->flags | CAPTURE_ASYNC;
7111 				datap = bp->datap;
7112 				mutex_exit(mp);
7113 				(void) page_trycapture(pp, szc, flags, datap);
7114 				return;
7115 			}
7116 			bp = bp->next;
7117 		}
7118 	}
7119 
7120 	/* Failed to find page in hash so clear flags and unlock it. */
7121 	page_clrtoxic(pp, PR_CAPTURE);
7122 	page_unlock(pp);
7123 
7124 	mutex_exit(mp);
7125 }
7126 
7127 void
7128 page_capture_init()
7129 {
7130 	int i;
7131 	for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) {
7132 		page_capture_hash[i].lists[0].next =
7133 		    &page_capture_hash[i].lists[0];
7134 		page_capture_hash[i].lists[0].prev =
7135 		    &page_capture_hash[i].lists[0];
7136 		page_capture_hash[i].lists[1].next =
7137 		    &page_capture_hash[i].lists[1];
7138 		page_capture_hash[i].lists[1].prev =
7139 		    &page_capture_hash[i].lists[1];
7140 	}
7141 
7142 	pc_thread_shortwait = 23 * hz;
7143 	pc_thread_longwait = 1201 * hz;
7144 	pc_thread_ism_retry = 3;
7145 	mutex_init(&pc_thread_mutex, NULL, MUTEX_DEFAULT, NULL);
7146 	cv_init(&pc_cv, NULL, CV_DEFAULT, NULL);
7147 	pc_thread_id = thread_create(NULL, 0, page_capture_thread, NULL, 0, &p0,
7148 	    TS_RUN, minclsyspri);
7149 }
7150 
7151 /*
7152  * It is necessary to scrub any failing pages prior to reboot in order to
7153  * prevent a latent error trap from occurring on the next boot.
7154  */
7155 void
7156 page_retire_mdboot()
7157 {
7158 	page_t *pp;
7159 	int i, j;
7160 	page_capture_hash_bucket_t *bp;
7161 
7162 	/* walk lists looking for pages to scrub */
7163 	for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) {
7164 		if (page_capture_hash[i].num_pages == 0)
7165 			continue;
7166 
7167 		mutex_enter(&page_capture_hash[i].pchh_mutex);
7168 
7169 		for (j = 0; j < 2; j++) {
7170 			bp = page_capture_hash[i].lists[j].next;
7171 			while (bp != &page_capture_hash[i].lists[j]) {
7172 				pp = bp->pp;
7173 				if (!PP_ISKAS(pp) && PP_TOXIC(pp)) {
7174 					pp->p_selock = -1;  /* pacify ASSERTs */
7175 					PP_CLRFREE(pp);
7176 					pagescrub(pp, 0, PAGESIZE);
7177 					pp->p_selock = 0;
7178 				}
7179 				bp = bp->next;
7180 			}
7181 		}
7182 		mutex_exit(&page_capture_hash[i].pchh_mutex);
7183 	}
7184 }
7185 
7186 /*
7187  * Walk the page_capture_hash trying to capture pages and also cleanup old
7188  * entries which have expired.
7189  */
7190 void
7191 page_capture_async()
7192 {
7193 	page_t *pp;
7194 	int i;
7195 	int ret;
7196 	page_capture_hash_bucket_t *bp1, *bp2;
7197 	uint_t szc;
7198 	uint_t flags;
7199 	void *datap;
7200 
7201 	/* If there are outstanding pages to be captured, get to work */
7202 	for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) {
7203 		if (page_capture_hash[i].num_pages == 0)
7204 			continue;
7205 		/* Append list 1 to list 0 and then walk through list 0 */
7206 		mutex_enter(&page_capture_hash[i].pchh_mutex);
7207 		bp1 = &page_capture_hash[i].lists[1];
7208 		bp2 = bp1->next;
7209 		if (bp1 != bp2) {
7210 			bp1->prev->next = page_capture_hash[i].lists[0].next;
7211 			bp2->prev = &page_capture_hash[i].lists[0];
7212 			page_capture_hash[i].lists[0].next->prev = bp1->prev;
7213 			page_capture_hash[i].lists[0].next = bp2;
7214 			bp1->next = bp1;
7215 			bp1->prev = bp1;
7216 		}
7217 
7218 		/* list[1] will be empty now */
7219 
7220 		bp1 = page_capture_hash[i].lists[0].next;
7221 		while (bp1 != &page_capture_hash[i].lists[0]) {
7222 			/* Check expiration time */
7223 			if ((lbolt > bp1->expires && bp1->expires != -1) ||
7224 			    page_deleted(bp1->pp)) {
7225 				page_capture_hash[i].lists[0].next = bp1->next;
7226 				bp1->next->prev =
7227 				    &page_capture_hash[i].lists[0];
7228 				page_capture_hash[i].num_pages--;
7229 
7230 				/*
7231 				 * We can safely remove the PR_CAPTURE bit
7232 				 * without holding the EXCL lock on the page
7233 				 * as the PR_CAPTURE bit requres that the
7234 				 * page_capture_hash[].pchh_mutex be held
7235 				 * to modify it.
7236 				 */
7237 				page_clrtoxic(bp1->pp, PR_CAPTURE);
7238 				mutex_exit(&page_capture_hash[i].pchh_mutex);
7239 				kmem_free(bp1, sizeof (*bp1));
7240 				mutex_enter(&page_capture_hash[i].pchh_mutex);
7241 				bp1 = page_capture_hash[i].lists[0].next;
7242 				continue;
7243 			}
7244 			pp = bp1->pp;
7245 			szc = bp1->szc;
7246 			flags = bp1->flags;
7247 			datap = bp1->datap;
7248 			mutex_exit(&page_capture_hash[i].pchh_mutex);
7249 			if (page_trylock(pp, SE_EXCL)) {
7250 				ret = page_trycapture(pp, szc,
7251 				    flags | CAPTURE_ASYNC, datap);
7252 			} else {
7253 				ret = 1;	/* move to walked hash */
7254 			}
7255 
7256 			if (ret != 0) {
7257 				/* Move to walked hash */
7258 				(void) page_capture_move_to_walked(pp);
7259 			}
7260 			mutex_enter(&page_capture_hash[i].pchh_mutex);
7261 			bp1 = page_capture_hash[i].lists[0].next;
7262 		}
7263 
7264 		mutex_exit(&page_capture_hash[i].pchh_mutex);
7265 	}
7266 }
7267 
7268 /*
7269  * This function is called by the page_capture_thread, and is needed in
7270  * in order to initiate aio cleanup, so that pages used in aio
7271  * will be unlocked and subsequently retired by page_capture_thread.
7272  */
7273 static int
7274 do_aio_cleanup(void)
7275 {
7276 	proc_t *procp;
7277 	int (*aio_cleanup_dr_delete_memory)(proc_t *);
7278 	int cleaned = 0;
7279 
7280 	if (modload("sys", "kaio") == -1) {
7281 		cmn_err(CE_WARN, "do_aio_cleanup: cannot load kaio");
7282 		return (0);
7283 	}
7284 	/*
7285 	 * We use the aio_cleanup_dr_delete_memory function to
7286 	 * initiate the actual clean up; this function will wake
7287 	 * up the per-process aio_cleanup_thread.
7288 	 */
7289 	aio_cleanup_dr_delete_memory = (int (*)(proc_t *))
7290 	    modgetsymvalue("aio_cleanup_dr_delete_memory", 0);
7291 	if (aio_cleanup_dr_delete_memory == NULL) {
7292 		cmn_err(CE_WARN,
7293 	    "aio_cleanup_dr_delete_memory not found in kaio");
7294 		return (0);
7295 	}
7296 	mutex_enter(&pidlock);
7297 	for (procp = practive; (procp != NULL); procp = procp->p_next) {
7298 		mutex_enter(&procp->p_lock);
7299 		if (procp->p_aio != NULL) {
7300 			/* cleanup proc's outstanding kaio */
7301 			cleaned += (*aio_cleanup_dr_delete_memory)(procp);
7302 		}
7303 		mutex_exit(&procp->p_lock);
7304 	}
7305 	mutex_exit(&pidlock);
7306 	return (cleaned);
7307 }
7308 
7309 /*
7310  * helper function for page_capture_thread
7311  */
7312 static void
7313 page_capture_handle_outstanding(void)
7314 {
7315 	extern size_t spt_used;
7316 	int ntry;
7317 
7318 	if (!page_retire_pend_count()) {
7319 		/*
7320 		 * Do we really want to be this aggressive
7321 		 * for things other than page_retire?
7322 		 * Maybe have a counter for each callback
7323 		 * type to guide how aggressive we should
7324 		 * be here.  Thus if there's at least one
7325 		 * page for page_retire we go ahead and reap
7326 		 * like this.
7327 		 */
7328 		kmem_reap();
7329 		seg_preap();
7330 		page_capture_async();
7331 	} else {
7332 		/*
7333 		 * There are pages pending retirement, so
7334 		 * we reap prior to attempting to capture.
7335 		 */
7336 		kmem_reap();
7337 		/*
7338 		 * When ISM is in use, we need to disable and
7339 		 * purge the seg_pcache, and initiate aio
7340 		 * cleanup in order to release page locks and
7341 		 * subsquently retire pages in need of
7342 		 * retirement.
7343 		 */
7344 		if (spt_used) {
7345 			/* disable and purge seg_pcache */
7346 			(void) seg_p_disable();
7347 			for (ntry = 0; ntry < pc_thread_ism_retry; ntry++) {
7348 				if (!page_retire_pend_count())
7349 					break;
7350 				if (do_aio_cleanup()) {
7351 					/*
7352 					 * allow the apps cleanup threads
7353 					 * to run
7354 					 */
7355 					delay(pc_thread_shortwait);
7356 				}
7357 				page_capture_async();
7358 			}
7359 			/* reenable seg_pcache */
7360 			seg_p_enable();
7361 		} else {
7362 			seg_preap();
7363 			page_capture_async();
7364 		}
7365 	}
7366 }
7367 
7368 /*
7369  * The page_capture_thread loops forever, looking to see if there are
7370  * pages still waiting to be captured.
7371  */
7372 static void
7373 page_capture_thread(void)
7374 {
7375 	callb_cpr_t c;
7376 	int outstanding;
7377 	int i;
7378 
7379 	CALLB_CPR_INIT(&c, &pc_thread_mutex, callb_generic_cpr, "page_capture");
7380 
7381 	mutex_enter(&pc_thread_mutex);
7382 	for (;;) {
7383 		outstanding = 0;
7384 		for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++)
7385 			outstanding += page_capture_hash[i].num_pages;
7386 		if (outstanding) {
7387 			page_capture_handle_outstanding();
7388 			CALLB_CPR_SAFE_BEGIN(&c);
7389 			(void) cv_timedwait(&pc_cv, &pc_thread_mutex,
7390 			    lbolt + pc_thread_shortwait);
7391 			CALLB_CPR_SAFE_END(&c, &pc_thread_mutex);
7392 		} else {
7393 			CALLB_CPR_SAFE_BEGIN(&c);
7394 			(void) cv_timedwait(&pc_cv, &pc_thread_mutex,
7395 			    lbolt + pc_thread_longwait);
7396 			CALLB_CPR_SAFE_END(&c, &pc_thread_mutex);
7397 		}
7398 	}
7399 	/*NOTREACHED*/
7400 }
7401