xref: /titanic_41/usr/src/uts/common/vm/vm_page.c (revision d15447b6c777a1b2223924443bf36c9c8efb2ea4)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*	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  * pages_locked : How many pages are locked because of user specified
110  * locking through mlock or plock.
111  *
112  * pages_useclaim,pages_claimed : These two variables track the
113  * claim adjustments because of the protection changes on a segvn segment.
114  *
115  * All these globals are protected by the same lock which protects availrmem.
116  */
117 pgcnt_t pages_locked = 0;
118 pgcnt_t pages_useclaim = 0;
119 pgcnt_t pages_claimed = 0;
120 
121 
122 /*
123  * new_freemem_lock protects freemem, freemem_wait & freemem_cv.
124  */
125 static kmutex_t	new_freemem_lock;
126 static uint_t	freemem_wait;	/* someone waiting for freemem */
127 static kcondvar_t freemem_cv;
128 
129 /*
130  * The logical page free list is maintained as two lists, the 'free'
131  * and the 'cache' lists.
132  * The free list contains those pages that should be reused first.
133  *
134  * The implementation of the lists is machine dependent.
135  * page_get_freelist(), page_get_cachelist(),
136  * page_list_sub(), and page_list_add()
137  * form the interface to the machine dependent implementation.
138  *
139  * Pages with p_free set are on the cache list.
140  * Pages with p_free and p_age set are on the free list,
141  *
142  * A page may be locked while on either list.
143  */
144 
145 /*
146  * free list accounting stuff.
147  *
148  *
149  * Spread out the value for the number of pages on the
150  * page free and page cache lists.  If there is just one
151  * value, then it must be under just one lock.
152  * The lock contention and cache traffic are a real bother.
153  *
154  * When we acquire and then drop a single pcf lock
155  * we can start in the middle of the array of pcf structures.
156  * If we acquire more than one pcf lock at a time, we need to
157  * start at the front to avoid deadlocking.
158  *
159  * pcf_count holds the number of pages in each pool.
160  *
161  * pcf_block is set when page_create_get_something() has asked the
162  * PSM page freelist and page cachelist routines without specifying
163  * a color and nothing came back.  This is used to block anything
164  * else from moving pages from one list to the other while the
165  * lists are searched again.  If a page is freeed while pcf_block is
166  * set, then pcf_reserve is incremented.  pcgs_unblock() takes care
167  * of clearning pcf_block, doing the wakeups, etc.
168  */
169 
170 #define	MAX_PCF_FANOUT NCPU
171 static uint_t pcf_fanout = 1; /* Will get changed at boot time */
172 static uint_t pcf_fanout_mask = 0;
173 
174 struct pcf {
175 	kmutex_t	pcf_lock;	/* protects the structure */
176 	uint_t		pcf_count;	/* page count */
177 	uint_t		pcf_wait;	/* number of waiters */
178 	uint_t		pcf_block; 	/* pcgs flag to page_free() */
179 	uint_t		pcf_reserve; 	/* pages freed after pcf_block set */
180 	uint_t		pcf_fill[10];	/* to line up on the caches */
181 };
182 
183 /*
184  * PCF_INDEX hash needs to be dynamic (every so often the hash changes where
185  * it will hash the cpu to).  This is done to prevent a drain condition
186  * from happening.  This drain condition will occur when pcf_count decrement
187  * occurs on cpu A and the increment of pcf_count always occurs on cpu B.  An
188  * example of this shows up with device interrupts.  The dma buffer is allocated
189  * by the cpu requesting the IO thus the pcf_count is decremented based on that.
190  * When the memory is returned by the interrupt thread, the pcf_count will be
191  * incremented based on the cpu servicing the interrupt.
192  */
193 static struct pcf pcf[MAX_PCF_FANOUT];
194 #define	PCF_INDEX() ((int)(((long)CPU->cpu_seqid) + \
195 	(randtick() >> 24)) & (pcf_fanout_mask))
196 
197 static int pcf_decrement_bucket(pgcnt_t);
198 static int pcf_decrement_multiple(pgcnt_t *, pgcnt_t, int);
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[9];
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  * /etc/system tunable to control large page allocation hueristic.
332  *
333  * Setting to LPAP_LOCAL will heavily prefer the local lgroup over remote lgroup
334  * for large page allocation requests.  If a large page is not readily
335  * avaliable on the local freelists we will go through additional effort
336  * to create a large page, potentially moving smaller pages around to coalesce
337  * larger pages in the local lgroup.
338  * Default value of LPAP_DEFAULT will go to remote freelists if large pages
339  * are not readily available in the local lgroup.
340  */
341 enum lpap {
342 	LPAP_DEFAULT,	/* default large page allocation policy */
343 	LPAP_LOCAL	/* local large page allocation policy */
344 };
345 
346 enum lpap lpg_alloc_prefer = LPAP_DEFAULT;
347 
348 static void page_init_mem_config(void);
349 static int page_do_hashin(page_t *, vnode_t *, u_offset_t);
350 static void page_do_hashout(page_t *);
351 static void page_capture_init();
352 int page_capture_take_action(page_t *, uint_t, void *);
353 
354 static void page_demote_vp_pages(page_t *);
355 
356 
357 void
358 pcf_init(void)
359 
360 {
361 	if (boot_ncpus != -1) {
362 		pcf_fanout = boot_ncpus;
363 	} else {
364 		pcf_fanout = max_ncpus;
365 	}
366 #ifdef sun4v
367 	/*
368 	 * Force at least 4 buckets if possible for sun4v.
369 	 */
370 	pcf_fanout = MAX(pcf_fanout, 4);
371 #endif /* sun4v */
372 
373 	/*
374 	 * Round up to the nearest power of 2.
375 	 */
376 	pcf_fanout = MIN(pcf_fanout, MAX_PCF_FANOUT);
377 	if (!ISP2(pcf_fanout)) {
378 		pcf_fanout = 1 << highbit(pcf_fanout);
379 
380 		if (pcf_fanout > MAX_PCF_FANOUT) {
381 			pcf_fanout = 1 << (highbit(MAX_PCF_FANOUT) - 1);
382 		}
383 	}
384 	pcf_fanout_mask = pcf_fanout - 1;
385 }
386 
387 /*
388  * vm subsystem related initialization
389  */
390 void
391 vm_init(void)
392 {
393 	boolean_t callb_vm_cpr(void *, int);
394 
395 	(void) callb_add(callb_vm_cpr, 0, CB_CL_CPR_VM, "vm");
396 	page_init_mem_config();
397 	page_retire_init();
398 	vm_usage_init();
399 	page_capture_init();
400 }
401 
402 /*
403  * This function is called at startup and when memory is added or deleted.
404  */
405 void
406 init_pages_pp_maximum()
407 {
408 	static pgcnt_t p_min;
409 	static pgcnt_t pages_pp_maximum_startup;
410 	static pgcnt_t avrmem_delta;
411 	static int init_done;
412 	static int user_set;	/* true if set in /etc/system */
413 
414 	if (init_done == 0) {
415 
416 		/* If the user specified a value, save it */
417 		if (pages_pp_maximum != 0) {
418 			user_set = 1;
419 			pages_pp_maximum_startup = pages_pp_maximum;
420 		}
421 
422 		/*
423 		 * Setting of pages_pp_maximum is based first time
424 		 * on the value of availrmem just after the start-up
425 		 * allocations. To preserve this relationship at run
426 		 * time, use a delta from availrmem_initial.
427 		 */
428 		ASSERT(availrmem_initial >= availrmem);
429 		avrmem_delta = availrmem_initial - availrmem;
430 
431 		/* The allowable floor of pages_pp_maximum */
432 		p_min = tune.t_minarmem + 100;
433 
434 		/* Make sure we don't come through here again. */
435 		init_done = 1;
436 	}
437 	/*
438 	 * Determine pages_pp_maximum, the number of currently available
439 	 * pages (availrmem) that can't be `locked'. If not set by
440 	 * the user, we set it to 4% of the currently available memory
441 	 * plus 4MB.
442 	 * But we also insist that it be greater than tune.t_minarmem;
443 	 * otherwise a process could lock down a lot of memory, get swapped
444 	 * out, and never have enough to get swapped back in.
445 	 */
446 	if (user_set)
447 		pages_pp_maximum = pages_pp_maximum_startup;
448 	else
449 		pages_pp_maximum = ((availrmem_initial - avrmem_delta) / 25)
450 		    + btop(4 * 1024 * 1024);
451 
452 	if (pages_pp_maximum <= p_min) {
453 		pages_pp_maximum = p_min;
454 	}
455 }
456 
457 void
458 set_max_page_get(pgcnt_t target_total_pages)
459 {
460 	max_page_get = target_total_pages / 2;
461 }
462 
463 static pgcnt_t pending_delete;
464 
465 /*ARGSUSED*/
466 static void
467 page_mem_config_post_add(
468 	void *arg,
469 	pgcnt_t delta_pages)
470 {
471 	set_max_page_get(total_pages - pending_delete);
472 	init_pages_pp_maximum();
473 }
474 
475 /*ARGSUSED*/
476 static int
477 page_mem_config_pre_del(
478 	void *arg,
479 	pgcnt_t delta_pages)
480 {
481 	pgcnt_t nv;
482 
483 	nv = atomic_add_long_nv(&pending_delete, (spgcnt_t)delta_pages);
484 	set_max_page_get(total_pages - nv);
485 	return (0);
486 }
487 
488 /*ARGSUSED*/
489 static void
490 page_mem_config_post_del(
491 	void *arg,
492 	pgcnt_t delta_pages,
493 	int cancelled)
494 {
495 	pgcnt_t nv;
496 
497 	nv = atomic_add_long_nv(&pending_delete, -(spgcnt_t)delta_pages);
498 	set_max_page_get(total_pages - nv);
499 	if (!cancelled)
500 		init_pages_pp_maximum();
501 }
502 
503 static kphysm_setup_vector_t page_mem_config_vec = {
504 	KPHYSM_SETUP_VECTOR_VERSION,
505 	page_mem_config_post_add,
506 	page_mem_config_pre_del,
507 	page_mem_config_post_del,
508 };
509 
510 static void
511 page_init_mem_config(void)
512 {
513 	int ret;
514 
515 	ret = kphysm_setup_func_register(&page_mem_config_vec, (void *)NULL);
516 	ASSERT(ret == 0);
517 }
518 
519 /*
520  * Evenly spread out the PCF counters for large free pages
521  */
522 static void
523 page_free_large_ctr(pgcnt_t npages)
524 {
525 	static struct pcf	*p = pcf;
526 	pgcnt_t			lump;
527 
528 	freemem += npages;
529 
530 	lump = roundup(npages, pcf_fanout) / pcf_fanout;
531 
532 	while (npages > 0) {
533 
534 		ASSERT(!p->pcf_block);
535 
536 		if (lump < npages) {
537 			p->pcf_count += (uint_t)lump;
538 			npages -= lump;
539 		} else {
540 			p->pcf_count += (uint_t)npages;
541 			npages = 0;
542 		}
543 
544 		ASSERT(!p->pcf_wait);
545 
546 		if (++p > &pcf[pcf_fanout - 1])
547 			p = pcf;
548 	}
549 
550 	ASSERT(npages == 0);
551 }
552 
553 /*
554  * Add a physical chunk of memory to the system free lists during startup.
555  * Platform specific startup() allocates the memory for the page structs.
556  *
557  * num	- number of page structures
558  * base - page number (pfn) to be associated with the first page.
559  *
560  * Since we are doing this during startup (ie. single threaded), we will
561  * use shortcut routines to avoid any locking overhead while putting all
562  * these pages on the freelists.
563  *
564  * NOTE: Any changes performed to page_free(), must also be performed to
565  *	 add_physmem() since this is how we initialize all page_t's at
566  *	 boot time.
567  */
568 void
569 add_physmem(
570 	page_t	*pp,
571 	pgcnt_t	num,
572 	pfn_t	pnum)
573 {
574 	page_t	*root = NULL;
575 	uint_t	szc = page_num_pagesizes() - 1;
576 	pgcnt_t	large = page_get_pagecnt(szc);
577 	pgcnt_t	cnt = 0;
578 
579 	TRACE_2(TR_FAC_VM, TR_PAGE_INIT,
580 	    "add_physmem:pp %p num %lu", pp, num);
581 
582 	/*
583 	 * Arbitrarily limit the max page_get request
584 	 * to 1/2 of the page structs we have.
585 	 */
586 	total_pages += num;
587 	set_max_page_get(total_pages);
588 
589 	PLCNT_MODIFY_MAX(pnum, (long)num);
590 
591 	/*
592 	 * The physical space for the pages array
593 	 * representing ram pages has already been
594 	 * allocated.  Here we initialize each lock
595 	 * in the page structure, and put each on
596 	 * the free list
597 	 */
598 	for (; num; pp++, pnum++, num--) {
599 
600 		/*
601 		 * this needs to fill in the page number
602 		 * and do any other arch specific initialization
603 		 */
604 		add_physmem_cb(pp, pnum);
605 
606 		pp->p_lckcnt = 0;
607 		pp->p_cowcnt = 0;
608 		pp->p_slckcnt = 0;
609 
610 		/*
611 		 * Initialize the page lock as unlocked, since nobody
612 		 * can see or access this page yet.
613 		 */
614 		pp->p_selock = 0;
615 
616 		/*
617 		 * Initialize IO lock
618 		 */
619 		page_iolock_init(pp);
620 
621 		/*
622 		 * initialize other fields in the page_t
623 		 */
624 		PP_SETFREE(pp);
625 		page_clr_all_props(pp);
626 		PP_SETAGED(pp);
627 		pp->p_offset = (u_offset_t)-1;
628 		pp->p_next = pp;
629 		pp->p_prev = pp;
630 
631 		/*
632 		 * Simple case: System doesn't support large pages.
633 		 */
634 		if (szc == 0) {
635 			pp->p_szc = 0;
636 			page_free_at_startup(pp);
637 			continue;
638 		}
639 
640 		/*
641 		 * Handle unaligned pages, we collect them up onto
642 		 * the root page until we have a full large page.
643 		 */
644 		if (!IS_P2ALIGNED(pnum, large)) {
645 
646 			/*
647 			 * If not in a large page,
648 			 * just free as small page.
649 			 */
650 			if (root == NULL) {
651 				pp->p_szc = 0;
652 				page_free_at_startup(pp);
653 				continue;
654 			}
655 
656 			/*
657 			 * Link a constituent page into the large page.
658 			 */
659 			pp->p_szc = szc;
660 			page_list_concat(&root, &pp);
661 
662 			/*
663 			 * When large page is fully formed, free it.
664 			 */
665 			if (++cnt == large) {
666 				page_free_large_ctr(cnt);
667 				page_list_add_pages(root, PG_LIST_ISINIT);
668 				root = NULL;
669 				cnt = 0;
670 			}
671 			continue;
672 		}
673 
674 		/*
675 		 * At this point we have a page number which
676 		 * is aligned. We assert that we aren't already
677 		 * in a different large page.
678 		 */
679 		ASSERT(IS_P2ALIGNED(pnum, large));
680 		ASSERT(root == NULL && cnt == 0);
681 
682 		/*
683 		 * If insufficient number of pages left to form
684 		 * a large page, just free the small page.
685 		 */
686 		if (num < large) {
687 			pp->p_szc = 0;
688 			page_free_at_startup(pp);
689 			continue;
690 		}
691 
692 		/*
693 		 * Otherwise start a new large page.
694 		 */
695 		pp->p_szc = szc;
696 		cnt++;
697 		root = pp;
698 	}
699 	ASSERT(root == NULL && cnt == 0);
700 }
701 
702 /*
703  * Find a page representing the specified [vp, offset].
704  * If we find the page but it is intransit coming in,
705  * it will have an "exclusive" lock and we wait for
706  * the i/o to complete.  A page found on the free list
707  * is always reclaimed and then locked.  On success, the page
708  * is locked, its data is valid and it isn't on the free
709  * list, while a NULL is returned if the page doesn't exist.
710  */
711 page_t *
712 page_lookup(vnode_t *vp, u_offset_t off, se_t se)
713 {
714 	return (page_lookup_create(vp, off, se, NULL, NULL, 0));
715 }
716 
717 /*
718  * Find a page representing the specified [vp, offset].
719  * We either return the one we found or, if passed in,
720  * create one with identity of [vp, offset] of the
721  * pre-allocated page. If we find existing page but it is
722  * intransit coming in, it will have an "exclusive" lock
723  * and we wait for the i/o to complete.  A page found on
724  * the free list is always reclaimed and then locked.
725  * On success, the page is locked, its data is valid and
726  * it isn't on the free list, while a NULL is returned
727  * if the page doesn't exist and newpp is NULL;
728  */
729 page_t *
730 page_lookup_create(
731 	vnode_t *vp,
732 	u_offset_t off,
733 	se_t se,
734 	page_t *newpp,
735 	spgcnt_t *nrelocp,
736 	int flags)
737 {
738 	page_t		*pp;
739 	kmutex_t	*phm;
740 	ulong_t		index;
741 	uint_t		hash_locked;
742 	uint_t		es;
743 
744 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
745 	VM_STAT_ADD(page_lookup_cnt[0]);
746 	ASSERT(newpp ? PAGE_EXCL(newpp) : 1);
747 
748 	/*
749 	 * Acquire the appropriate page hash lock since
750 	 * we have to search the hash list.  Pages that
751 	 * hash to this list can't change identity while
752 	 * this lock is held.
753 	 */
754 	hash_locked = 0;
755 	index = PAGE_HASH_FUNC(vp, off);
756 	phm = NULL;
757 top:
758 	PAGE_HASH_SEARCH(index, pp, vp, off);
759 	if (pp != NULL) {
760 		VM_STAT_ADD(page_lookup_cnt[1]);
761 		es = (newpp != NULL) ? 1 : 0;
762 		es |= flags;
763 		if (!hash_locked) {
764 			VM_STAT_ADD(page_lookup_cnt[2]);
765 			if (!page_try_reclaim_lock(pp, se, es)) {
766 				/*
767 				 * On a miss, acquire the phm.  Then
768 				 * next time, page_lock() will be called,
769 				 * causing a wait if the page is busy.
770 				 * just looping with page_trylock() would
771 				 * get pretty boring.
772 				 */
773 				VM_STAT_ADD(page_lookup_cnt[3]);
774 				phm = PAGE_HASH_MUTEX(index);
775 				mutex_enter(phm);
776 				hash_locked = 1;
777 				goto top;
778 			}
779 		} else {
780 			VM_STAT_ADD(page_lookup_cnt[4]);
781 			if (!page_lock_es(pp, se, phm, P_RECLAIM, es)) {
782 				VM_STAT_ADD(page_lookup_cnt[5]);
783 				goto top;
784 			}
785 		}
786 
787 		/*
788 		 * Since `pp' is locked it can not change identity now.
789 		 * Reconfirm we locked the correct page.
790 		 *
791 		 * Both the p_vnode and p_offset *must* be cast volatile
792 		 * to force a reload of their values: The PAGE_HASH_SEARCH
793 		 * macro will have stuffed p_vnode and p_offset into
794 		 * registers before calling page_trylock(); another thread,
795 		 * actually holding the hash lock, could have changed the
796 		 * page's identity in memory, but our registers would not
797 		 * be changed, fooling the reconfirmation.  If the hash
798 		 * lock was held during the search, the casting would
799 		 * not be needed.
800 		 */
801 		VM_STAT_ADD(page_lookup_cnt[6]);
802 		if (((volatile struct vnode *)(pp->p_vnode) != vp) ||
803 		    ((volatile u_offset_t)(pp->p_offset) != off)) {
804 			VM_STAT_ADD(page_lookup_cnt[7]);
805 			if (hash_locked) {
806 				panic("page_lookup_create: lost page %p",
807 				    (void *)pp);
808 				/*NOTREACHED*/
809 			}
810 			page_unlock(pp);
811 			phm = PAGE_HASH_MUTEX(index);
812 			mutex_enter(phm);
813 			hash_locked = 1;
814 			goto top;
815 		}
816 
817 		/*
818 		 * If page_trylock() was called, then pp may still be on
819 		 * the cachelist (can't be on the free list, it would not
820 		 * have been found in the search).  If it is on the
821 		 * cachelist it must be pulled now. To pull the page from
822 		 * the cachelist, it must be exclusively locked.
823 		 *
824 		 * The other big difference between page_trylock() and
825 		 * page_lock(), is that page_lock() will pull the
826 		 * page from whatever free list (the cache list in this
827 		 * case) the page is on.  If page_trylock() was used
828 		 * above, then we have to do the reclaim ourselves.
829 		 */
830 		if ((!hash_locked) && (PP_ISFREE(pp))) {
831 			ASSERT(PP_ISAGED(pp) == 0);
832 			VM_STAT_ADD(page_lookup_cnt[8]);
833 
834 			/*
835 			 * page_relcaim will insure that we
836 			 * have this page exclusively
837 			 */
838 
839 			if (!page_reclaim(pp, NULL)) {
840 				/*
841 				 * Page_reclaim dropped whatever lock
842 				 * we held.
843 				 */
844 				VM_STAT_ADD(page_lookup_cnt[9]);
845 				phm = PAGE_HASH_MUTEX(index);
846 				mutex_enter(phm);
847 				hash_locked = 1;
848 				goto top;
849 			} else if (se == SE_SHARED && newpp == NULL) {
850 				VM_STAT_ADD(page_lookup_cnt[10]);
851 				page_downgrade(pp);
852 			}
853 		}
854 
855 		if (hash_locked) {
856 			mutex_exit(phm);
857 		}
858 
859 		if (newpp != NULL && pp->p_szc < newpp->p_szc &&
860 		    PAGE_EXCL(pp) && nrelocp != NULL) {
861 			ASSERT(nrelocp != NULL);
862 			(void) page_relocate(&pp, &newpp, 1, 1, nrelocp,
863 			    NULL);
864 			if (*nrelocp > 0) {
865 				VM_STAT_COND_ADD(*nrelocp == 1,
866 				    page_lookup_cnt[11]);
867 				VM_STAT_COND_ADD(*nrelocp > 1,
868 				    page_lookup_cnt[12]);
869 				pp = newpp;
870 				se = SE_EXCL;
871 			} else {
872 				if (se == SE_SHARED) {
873 					page_downgrade(pp);
874 				}
875 				VM_STAT_ADD(page_lookup_cnt[13]);
876 			}
877 		} else if (newpp != NULL && nrelocp != NULL) {
878 			if (PAGE_EXCL(pp) && se == SE_SHARED) {
879 				page_downgrade(pp);
880 			}
881 			VM_STAT_COND_ADD(pp->p_szc < newpp->p_szc,
882 			    page_lookup_cnt[14]);
883 			VM_STAT_COND_ADD(pp->p_szc == newpp->p_szc,
884 			    page_lookup_cnt[15]);
885 			VM_STAT_COND_ADD(pp->p_szc > newpp->p_szc,
886 			    page_lookup_cnt[16]);
887 		} else if (newpp != NULL && PAGE_EXCL(pp)) {
888 			se = SE_EXCL;
889 		}
890 	} else if (!hash_locked) {
891 		VM_STAT_ADD(page_lookup_cnt[17]);
892 		phm = PAGE_HASH_MUTEX(index);
893 		mutex_enter(phm);
894 		hash_locked = 1;
895 		goto top;
896 	} else if (newpp != NULL) {
897 		/*
898 		 * If we have a preallocated page then
899 		 * insert it now and basically behave like
900 		 * page_create.
901 		 */
902 		VM_STAT_ADD(page_lookup_cnt[18]);
903 		/*
904 		 * Since we hold the page hash mutex and
905 		 * just searched for this page, page_hashin
906 		 * had better not fail.  If it does, that
907 		 * means some thread did not follow the
908 		 * page hash mutex rules.  Panic now and
909 		 * get it over with.  As usual, go down
910 		 * holding all the locks.
911 		 */
912 		ASSERT(MUTEX_HELD(phm));
913 		if (!page_hashin(newpp, vp, off, phm)) {
914 			ASSERT(MUTEX_HELD(phm));
915 			panic("page_lookup_create: hashin failed %p %p %llx %p",
916 			    (void *)newpp, (void *)vp, off, (void *)phm);
917 			/*NOTREACHED*/
918 		}
919 		ASSERT(MUTEX_HELD(phm));
920 		mutex_exit(phm);
921 		phm = NULL;
922 		page_set_props(newpp, P_REF);
923 		page_io_lock(newpp);
924 		pp = newpp;
925 		se = SE_EXCL;
926 	} else {
927 		VM_STAT_ADD(page_lookup_cnt[19]);
928 		mutex_exit(phm);
929 	}
930 
931 	ASSERT(pp ? PAGE_LOCKED_SE(pp, se) : 1);
932 
933 	ASSERT(pp ? ((PP_ISFREE(pp) == 0) && (PP_ISAGED(pp) == 0)) : 1);
934 
935 	return (pp);
936 }
937 
938 /*
939  * Search the hash list for the page representing the
940  * specified [vp, offset] and return it locked.  Skip
941  * free pages and pages that cannot be locked as requested.
942  * Used while attempting to kluster pages.
943  */
944 page_t *
945 page_lookup_nowait(vnode_t *vp, u_offset_t off, se_t se)
946 {
947 	page_t		*pp;
948 	kmutex_t	*phm;
949 	ulong_t		index;
950 	uint_t		locked;
951 
952 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
953 	VM_STAT_ADD(page_lookup_nowait_cnt[0]);
954 
955 	index = PAGE_HASH_FUNC(vp, off);
956 	PAGE_HASH_SEARCH(index, pp, vp, off);
957 	locked = 0;
958 	if (pp == NULL) {
959 top:
960 		VM_STAT_ADD(page_lookup_nowait_cnt[1]);
961 		locked = 1;
962 		phm = PAGE_HASH_MUTEX(index);
963 		mutex_enter(phm);
964 		PAGE_HASH_SEARCH(index, pp, vp, off);
965 	}
966 
967 	if (pp == NULL || PP_ISFREE(pp)) {
968 		VM_STAT_ADD(page_lookup_nowait_cnt[2]);
969 		pp = NULL;
970 	} else {
971 		if (!page_trylock(pp, se)) {
972 			VM_STAT_ADD(page_lookup_nowait_cnt[3]);
973 			pp = NULL;
974 		} else {
975 			VM_STAT_ADD(page_lookup_nowait_cnt[4]);
976 			/*
977 			 * See the comment in page_lookup()
978 			 */
979 			if (((volatile struct vnode *)(pp->p_vnode) != vp) ||
980 			    ((u_offset_t)(pp->p_offset) != off)) {
981 				VM_STAT_ADD(page_lookup_nowait_cnt[5]);
982 				if (locked) {
983 					panic("page_lookup_nowait %p",
984 					    (void *)pp);
985 					/*NOTREACHED*/
986 				}
987 				page_unlock(pp);
988 				goto top;
989 			}
990 			if (PP_ISFREE(pp)) {
991 				VM_STAT_ADD(page_lookup_nowait_cnt[6]);
992 				page_unlock(pp);
993 				pp = NULL;
994 			}
995 		}
996 	}
997 	if (locked) {
998 		VM_STAT_ADD(page_lookup_nowait_cnt[7]);
999 		mutex_exit(phm);
1000 	}
1001 
1002 	ASSERT(pp ? PAGE_LOCKED_SE(pp, se) : 1);
1003 
1004 	return (pp);
1005 }
1006 
1007 /*
1008  * Search the hash list for a page with the specified [vp, off]
1009  * that is known to exist and is already locked.  This routine
1010  * is typically used by segment SOFTUNLOCK routines.
1011  */
1012 page_t *
1013 page_find(vnode_t *vp, u_offset_t off)
1014 {
1015 	page_t		*pp;
1016 	kmutex_t	*phm;
1017 	ulong_t		index;
1018 
1019 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
1020 	VM_STAT_ADD(page_find_cnt);
1021 
1022 	index = PAGE_HASH_FUNC(vp, off);
1023 	phm = PAGE_HASH_MUTEX(index);
1024 
1025 	mutex_enter(phm);
1026 	PAGE_HASH_SEARCH(index, pp, vp, off);
1027 	mutex_exit(phm);
1028 
1029 	ASSERT(pp == NULL || PAGE_LOCKED(pp) || panicstr);
1030 	return (pp);
1031 }
1032 
1033 /*
1034  * Determine whether a page with the specified [vp, off]
1035  * currently exists in the system.  Obviously this should
1036  * only be considered as a hint since nothing prevents the
1037  * page from disappearing or appearing immediately after
1038  * the return from this routine. Subsequently, we don't
1039  * even bother to lock the list.
1040  */
1041 page_t *
1042 page_exists(vnode_t *vp, u_offset_t off)
1043 {
1044 	page_t	*pp;
1045 	ulong_t		index;
1046 
1047 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
1048 	VM_STAT_ADD(page_exists_cnt);
1049 
1050 	index = PAGE_HASH_FUNC(vp, off);
1051 	PAGE_HASH_SEARCH(index, pp, vp, off);
1052 
1053 	return (pp);
1054 }
1055 
1056 /*
1057  * Determine if physically contiguous pages exist for [vp, off] - [vp, off +
1058  * page_size(szc)) range.  if they exist and ppa is not NULL fill ppa array
1059  * with these pages locked SHARED. If necessary reclaim pages from
1060  * freelist. Return 1 if contiguous pages exist and 0 otherwise.
1061  *
1062  * If we fail to lock pages still return 1 if pages exist and contiguous.
1063  * But in this case return value is just a hint. ppa array won't be filled.
1064  * Caller should initialize ppa[0] as NULL to distinguish return value.
1065  *
1066  * Returns 0 if pages don't exist or not physically contiguous.
1067  *
1068  * This routine doesn't work for anonymous(swapfs) pages.
1069  */
1070 int
1071 page_exists_physcontig(vnode_t *vp, u_offset_t off, uint_t szc, page_t *ppa[])
1072 {
1073 	pgcnt_t pages;
1074 	pfn_t pfn;
1075 	page_t *rootpp;
1076 	pgcnt_t i;
1077 	pgcnt_t j;
1078 	u_offset_t save_off = off;
1079 	ulong_t index;
1080 	kmutex_t *phm;
1081 	page_t *pp;
1082 	uint_t pszc;
1083 	int loopcnt = 0;
1084 
1085 	ASSERT(szc != 0);
1086 	ASSERT(vp != NULL);
1087 	ASSERT(!IS_SWAPFSVP(vp));
1088 	ASSERT(!VN_ISKAS(vp));
1089 
1090 again:
1091 	if (++loopcnt > 3) {
1092 		VM_STAT_ADD(page_exphcontg[0]);
1093 		return (0);
1094 	}
1095 
1096 	index = PAGE_HASH_FUNC(vp, off);
1097 	phm = PAGE_HASH_MUTEX(index);
1098 
1099 	mutex_enter(phm);
1100 	PAGE_HASH_SEARCH(index, pp, vp, off);
1101 	mutex_exit(phm);
1102 
1103 	VM_STAT_ADD(page_exphcontg[1]);
1104 
1105 	if (pp == NULL) {
1106 		VM_STAT_ADD(page_exphcontg[2]);
1107 		return (0);
1108 	}
1109 
1110 	pages = page_get_pagecnt(szc);
1111 	rootpp = pp;
1112 	pfn = rootpp->p_pagenum;
1113 
1114 	if ((pszc = pp->p_szc) >= szc && ppa != NULL) {
1115 		VM_STAT_ADD(page_exphcontg[3]);
1116 		if (!page_trylock(pp, SE_SHARED)) {
1117 			VM_STAT_ADD(page_exphcontg[4]);
1118 			return (1);
1119 		}
1120 		if (pp->p_szc != pszc || pp->p_vnode != vp ||
1121 		    pp->p_offset != off) {
1122 			VM_STAT_ADD(page_exphcontg[5]);
1123 			page_unlock(pp);
1124 			off = save_off;
1125 			goto again;
1126 		}
1127 		/*
1128 		 * szc was non zero and vnode and offset matched after we
1129 		 * locked the page it means it can't become free on us.
1130 		 */
1131 		ASSERT(!PP_ISFREE(pp));
1132 		if (!IS_P2ALIGNED(pfn, pages)) {
1133 			page_unlock(pp);
1134 			return (0);
1135 		}
1136 		ppa[0] = pp;
1137 		pp++;
1138 		off += PAGESIZE;
1139 		pfn++;
1140 		for (i = 1; i < pages; i++, pp++, off += PAGESIZE, pfn++) {
1141 			if (!page_trylock(pp, SE_SHARED)) {
1142 				VM_STAT_ADD(page_exphcontg[6]);
1143 				pp--;
1144 				while (i-- > 0) {
1145 					page_unlock(pp);
1146 					pp--;
1147 				}
1148 				ppa[0] = NULL;
1149 				return (1);
1150 			}
1151 			if (pp->p_szc != pszc) {
1152 				VM_STAT_ADD(page_exphcontg[7]);
1153 				page_unlock(pp);
1154 				pp--;
1155 				while (i-- > 0) {
1156 					page_unlock(pp);
1157 					pp--;
1158 				}
1159 				ppa[0] = NULL;
1160 				off = save_off;
1161 				goto again;
1162 			}
1163 			/*
1164 			 * szc the same as for previous already locked pages
1165 			 * with right identity. Since this page had correct
1166 			 * szc after we locked it can't get freed or destroyed
1167 			 * and therefore must have the expected identity.
1168 			 */
1169 			ASSERT(!PP_ISFREE(pp));
1170 			if (pp->p_vnode != vp ||
1171 			    pp->p_offset != off) {
1172 				panic("page_exists_physcontig: "
1173 				    "large page identity doesn't match");
1174 			}
1175 			ppa[i] = pp;
1176 			ASSERT(pp->p_pagenum == pfn);
1177 		}
1178 		VM_STAT_ADD(page_exphcontg[8]);
1179 		ppa[pages] = NULL;
1180 		return (1);
1181 	} else if (pszc >= szc) {
1182 		VM_STAT_ADD(page_exphcontg[9]);
1183 		if (!IS_P2ALIGNED(pfn, pages)) {
1184 			return (0);
1185 		}
1186 		return (1);
1187 	}
1188 
1189 	if (!IS_P2ALIGNED(pfn, pages)) {
1190 		VM_STAT_ADD(page_exphcontg[10]);
1191 		return (0);
1192 	}
1193 
1194 	if (page_numtomemseg_nolock(pfn) !=
1195 	    page_numtomemseg_nolock(pfn + pages - 1)) {
1196 		VM_STAT_ADD(page_exphcontg[11]);
1197 		return (0);
1198 	}
1199 
1200 	/*
1201 	 * We loop up 4 times across pages to promote page size.
1202 	 * We're extra cautious to promote page size atomically with respect
1203 	 * to everybody else.  But we can probably optimize into 1 loop if
1204 	 * this becomes an issue.
1205 	 */
1206 
1207 	for (i = 0; i < pages; i++, pp++, off += PAGESIZE, pfn++) {
1208 		ASSERT(pp->p_pagenum == pfn);
1209 		if (!page_trylock(pp, SE_EXCL)) {
1210 			VM_STAT_ADD(page_exphcontg[12]);
1211 			break;
1212 		}
1213 		if (pp->p_vnode != vp ||
1214 		    pp->p_offset != off) {
1215 			VM_STAT_ADD(page_exphcontg[13]);
1216 			page_unlock(pp);
1217 			break;
1218 		}
1219 		if (pp->p_szc >= szc) {
1220 			ASSERT(i == 0);
1221 			page_unlock(pp);
1222 			off = save_off;
1223 			goto again;
1224 		}
1225 	}
1226 
1227 	if (i != pages) {
1228 		VM_STAT_ADD(page_exphcontg[14]);
1229 		--pp;
1230 		while (i-- > 0) {
1231 			page_unlock(pp);
1232 			--pp;
1233 		}
1234 		return (0);
1235 	}
1236 
1237 	pp = rootpp;
1238 	for (i = 0; i < pages; i++, pp++) {
1239 		if (PP_ISFREE(pp)) {
1240 			VM_STAT_ADD(page_exphcontg[15]);
1241 			ASSERT(!PP_ISAGED(pp));
1242 			ASSERT(pp->p_szc == 0);
1243 			if (!page_reclaim(pp, NULL)) {
1244 				break;
1245 			}
1246 		} else {
1247 			ASSERT(pp->p_szc < szc);
1248 			VM_STAT_ADD(page_exphcontg[16]);
1249 			(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
1250 		}
1251 	}
1252 	if (i < pages) {
1253 		VM_STAT_ADD(page_exphcontg[17]);
1254 		/*
1255 		 * page_reclaim failed because we were out of memory.
1256 		 * drop the rest of the locks and return because this page
1257 		 * must be already reallocated anyway.
1258 		 */
1259 		pp = rootpp;
1260 		for (j = 0; j < pages; j++, pp++) {
1261 			if (j != i) {
1262 				page_unlock(pp);
1263 			}
1264 		}
1265 		return (0);
1266 	}
1267 
1268 	off = save_off;
1269 	pp = rootpp;
1270 	for (i = 0; i < pages; i++, pp++, off += PAGESIZE) {
1271 		ASSERT(PAGE_EXCL(pp));
1272 		ASSERT(!PP_ISFREE(pp));
1273 		ASSERT(!hat_page_is_mapped(pp));
1274 		ASSERT(pp->p_vnode == vp);
1275 		ASSERT(pp->p_offset == off);
1276 		pp->p_szc = szc;
1277 	}
1278 	pp = rootpp;
1279 	for (i = 0; i < pages; i++, pp++) {
1280 		if (ppa == NULL) {
1281 			page_unlock(pp);
1282 		} else {
1283 			ppa[i] = pp;
1284 			page_downgrade(ppa[i]);
1285 		}
1286 	}
1287 	if (ppa != NULL) {
1288 		ppa[pages] = NULL;
1289 	}
1290 	VM_STAT_ADD(page_exphcontg[18]);
1291 	ASSERT(vp->v_pages != NULL);
1292 	return (1);
1293 }
1294 
1295 /*
1296  * Determine whether a page with the specified [vp, off]
1297  * currently exists in the system and if so return its
1298  * size code. Obviously this should only be considered as
1299  * a hint since nothing prevents the page from disappearing
1300  * or appearing immediately after the return from this routine.
1301  */
1302 int
1303 page_exists_forreal(vnode_t *vp, u_offset_t off, uint_t *szc)
1304 {
1305 	page_t		*pp;
1306 	kmutex_t	*phm;
1307 	ulong_t		index;
1308 	int		rc = 0;
1309 
1310 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
1311 	ASSERT(szc != NULL);
1312 	VM_STAT_ADD(page_exists_forreal_cnt);
1313 
1314 	index = PAGE_HASH_FUNC(vp, off);
1315 	phm = PAGE_HASH_MUTEX(index);
1316 
1317 	mutex_enter(phm);
1318 	PAGE_HASH_SEARCH(index, pp, vp, off);
1319 	if (pp != NULL) {
1320 		*szc = pp->p_szc;
1321 		rc = 1;
1322 	}
1323 	mutex_exit(phm);
1324 	return (rc);
1325 }
1326 
1327 /* wakeup threads waiting for pages in page_create_get_something() */
1328 void
1329 wakeup_pcgs(void)
1330 {
1331 	if (!CV_HAS_WAITERS(&pcgs_cv))
1332 		return;
1333 	cv_broadcast(&pcgs_cv);
1334 }
1335 
1336 /*
1337  * 'freemem' is used all over the kernel as an indication of how many
1338  * pages are free (either on the cache list or on the free page list)
1339  * in the system.  In very few places is a really accurate 'freemem'
1340  * needed.  To avoid contention of the lock protecting a the
1341  * single freemem, it was spread out into NCPU buckets.  Set_freemem
1342  * sets freemem to the total of all NCPU buckets.  It is called from
1343  * clock() on each TICK.
1344  */
1345 void
1346 set_freemem()
1347 {
1348 	struct pcf	*p;
1349 	ulong_t		t;
1350 	uint_t		i;
1351 
1352 	t = 0;
1353 	p = pcf;
1354 	for (i = 0;  i < pcf_fanout; i++) {
1355 		t += p->pcf_count;
1356 		p++;
1357 	}
1358 	freemem = t;
1359 
1360 	/*
1361 	 * Don't worry about grabbing mutex.  It's not that
1362 	 * critical if we miss a tick or two.  This is
1363 	 * where we wakeup possible delayers in
1364 	 * page_create_get_something().
1365 	 */
1366 	wakeup_pcgs();
1367 }
1368 
1369 ulong_t
1370 get_freemem()
1371 {
1372 	struct pcf	*p;
1373 	ulong_t		t;
1374 	uint_t		i;
1375 
1376 	t = 0;
1377 	p = pcf;
1378 	for (i = 0; i < pcf_fanout; i++) {
1379 		t += p->pcf_count;
1380 		p++;
1381 	}
1382 	/*
1383 	 * We just calculated it, might as well set it.
1384 	 */
1385 	freemem = t;
1386 	return (t);
1387 }
1388 
1389 /*
1390  * Acquire all of the page cache & free (pcf) locks.
1391  */
1392 void
1393 pcf_acquire_all()
1394 {
1395 	struct pcf	*p;
1396 	uint_t		i;
1397 
1398 	p = pcf;
1399 	for (i = 0; i < pcf_fanout; i++) {
1400 		mutex_enter(&p->pcf_lock);
1401 		p++;
1402 	}
1403 }
1404 
1405 /*
1406  * Release all the pcf_locks.
1407  */
1408 void
1409 pcf_release_all()
1410 {
1411 	struct pcf	*p;
1412 	uint_t		i;
1413 
1414 	p = pcf;
1415 	for (i = 0; i < pcf_fanout; i++) {
1416 		mutex_exit(&p->pcf_lock);
1417 		p++;
1418 	}
1419 }
1420 
1421 /*
1422  * Inform the VM system that we need some pages freed up.
1423  * Calls must be symmetric, e.g.:
1424  *
1425  *	page_needfree(100);
1426  *	wait a bit;
1427  *	page_needfree(-100);
1428  */
1429 void
1430 page_needfree(spgcnt_t npages)
1431 {
1432 	mutex_enter(&new_freemem_lock);
1433 	needfree += npages;
1434 	mutex_exit(&new_freemem_lock);
1435 }
1436 
1437 /*
1438  * Throttle for page_create(): try to prevent freemem from dropping
1439  * below throttlefree.  We can't provide a 100% guarantee because
1440  * KM_NOSLEEP allocations, page_reclaim(), and various other things
1441  * nibble away at the freelist.  However, we can block all PG_WAIT
1442  * allocations until memory becomes available.  The motivation is
1443  * that several things can fall apart when there's no free memory:
1444  *
1445  * (1) If pageout() needs memory to push a page, the system deadlocks.
1446  *
1447  * (2) By (broken) specification, timeout(9F) can neither fail nor
1448  *     block, so it has no choice but to panic the system if it
1449  *     cannot allocate a callout structure.
1450  *
1451  * (3) Like timeout(), ddi_set_callback() cannot fail and cannot block;
1452  *     it panics if it cannot allocate a callback structure.
1453  *
1454  * (4) Untold numbers of third-party drivers have not yet been hardened
1455  *     against KM_NOSLEEP and/or allocb() failures; they simply assume
1456  *     success and panic the system with a data fault on failure.
1457  *     (The long-term solution to this particular problem is to ship
1458  *     hostile fault-injecting DEBUG kernels with the DDK.)
1459  *
1460  * It is theoretically impossible to guarantee success of non-blocking
1461  * allocations, but in practice, this throttle is very hard to break.
1462  */
1463 static int
1464 page_create_throttle(pgcnt_t npages, int flags)
1465 {
1466 	ulong_t	fm;
1467 	uint_t	i;
1468 	pgcnt_t tf;	/* effective value of throttlefree */
1469 
1470 	/*
1471 	 * Never deny pages when:
1472 	 * - it's a thread that cannot block [NOMEMWAIT()]
1473 	 * - the allocation cannot block and must not fail
1474 	 * - the allocation cannot block and is pageout dispensated
1475 	 */
1476 	if (NOMEMWAIT() ||
1477 	    ((flags & (PG_WAIT | PG_PANIC)) == PG_PANIC) ||
1478 	    ((flags & (PG_WAIT | PG_PUSHPAGE)) == PG_PUSHPAGE))
1479 		return (1);
1480 
1481 	/*
1482 	 * If the allocation can't block, we look favorably upon it
1483 	 * unless we're below pageout_reserve.  In that case we fail
1484 	 * the allocation because we want to make sure there are a few
1485 	 * pages available for pageout.
1486 	 */
1487 	if ((flags & PG_WAIT) == 0)
1488 		return (freemem >= npages + pageout_reserve);
1489 
1490 	/* Calculate the effective throttlefree value */
1491 	tf = throttlefree -
1492 	    ((flags & PG_PUSHPAGE) ? pageout_reserve : 0);
1493 
1494 	cv_signal(&proc_pageout->p_cv);
1495 
1496 	for (;;) {
1497 		fm = 0;
1498 		pcf_acquire_all();
1499 		mutex_enter(&new_freemem_lock);
1500 		for (i = 0; i < pcf_fanout; i++) {
1501 			fm += pcf[i].pcf_count;
1502 			pcf[i].pcf_wait++;
1503 			mutex_exit(&pcf[i].pcf_lock);
1504 		}
1505 		freemem = fm;
1506 		if (freemem >= npages + tf) {
1507 			mutex_exit(&new_freemem_lock);
1508 			break;
1509 		}
1510 		needfree += npages;
1511 		freemem_wait++;
1512 		cv_wait(&freemem_cv, &new_freemem_lock);
1513 		freemem_wait--;
1514 		needfree -= npages;
1515 		mutex_exit(&new_freemem_lock);
1516 	}
1517 	return (1);
1518 }
1519 
1520 /*
1521  * page_create_wait() is called to either coalesce pages from the
1522  * different pcf buckets or to wait because there simply are not
1523  * enough pages to satisfy the caller's request.
1524  *
1525  * Sadly, this is called from platform/vm/vm_machdep.c
1526  */
1527 int
1528 page_create_wait(pgcnt_t npages, uint_t flags)
1529 {
1530 	pgcnt_t		total;
1531 	uint_t		i;
1532 	struct pcf	*p;
1533 
1534 	/*
1535 	 * Wait until there are enough free pages to satisfy our
1536 	 * entire request.
1537 	 * We set needfree += npages before prodding pageout, to make sure
1538 	 * it does real work when npages > lotsfree > freemem.
1539 	 */
1540 	VM_STAT_ADD(page_create_not_enough);
1541 
1542 	ASSERT(!kcage_on ? !(flags & PG_NORELOC) : 1);
1543 checkagain:
1544 	if ((flags & PG_NORELOC) &&
1545 	    kcage_freemem < kcage_throttlefree + npages)
1546 		(void) kcage_create_throttle(npages, flags);
1547 
1548 	if (freemem < npages + throttlefree)
1549 		if (!page_create_throttle(npages, flags))
1550 			return (0);
1551 
1552 	if (pcf_decrement_bucket(npages) ||
1553 	    pcf_decrement_multiple(&total, npages, 0))
1554 		return (1);
1555 
1556 	/*
1557 	 * All of the pcf locks are held, there are not enough pages
1558 	 * to satisfy the request (npages < total).
1559 	 * Be sure to acquire the new_freemem_lock before dropping
1560 	 * the pcf locks.  This prevents dropping wakeups in page_free().
1561 	 * The order is always pcf_lock then new_freemem_lock.
1562 	 *
1563 	 * Since we hold all the pcf locks, it is a good time to set freemem.
1564 	 *
1565 	 * If the caller does not want to wait, return now.
1566 	 * Else turn the pageout daemon loose to find something
1567 	 * and wait till it does.
1568 	 *
1569 	 */
1570 	freemem = total;
1571 
1572 	if ((flags & PG_WAIT) == 0) {
1573 		pcf_release_all();
1574 
1575 		TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_NOMEM,
1576 		"page_create_nomem:npages %ld freemem %ld", npages, freemem);
1577 		return (0);
1578 	}
1579 
1580 	ASSERT(proc_pageout != NULL);
1581 	cv_signal(&proc_pageout->p_cv);
1582 
1583 	TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SLEEP_START,
1584 	    "page_create_sleep_start: freemem %ld needfree %ld",
1585 	    freemem, needfree);
1586 
1587 	/*
1588 	 * We are going to wait.
1589 	 * We currently hold all of the pcf_locks,
1590 	 * get the new_freemem_lock (it protects freemem_wait),
1591 	 * before dropping the pcf_locks.
1592 	 */
1593 	mutex_enter(&new_freemem_lock);
1594 
1595 	p = pcf;
1596 	for (i = 0; i < pcf_fanout; i++) {
1597 		p->pcf_wait++;
1598 		mutex_exit(&p->pcf_lock);
1599 		p++;
1600 	}
1601 
1602 	needfree += npages;
1603 	freemem_wait++;
1604 
1605 	cv_wait(&freemem_cv, &new_freemem_lock);
1606 
1607 	freemem_wait--;
1608 	needfree -= npages;
1609 
1610 	mutex_exit(&new_freemem_lock);
1611 
1612 	TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SLEEP_END,
1613 	    "page_create_sleep_end: freemem %ld needfree %ld",
1614 	    freemem, needfree);
1615 
1616 	VM_STAT_ADD(page_create_not_enough_again);
1617 	goto checkagain;
1618 }
1619 /*
1620  * A routine to do the opposite of page_create_wait().
1621  */
1622 void
1623 page_create_putback(spgcnt_t npages)
1624 {
1625 	struct pcf	*p;
1626 	pgcnt_t		lump;
1627 	uint_t		*which;
1628 
1629 	/*
1630 	 * When a contiguous lump is broken up, we have to
1631 	 * deal with lots of pages (min 64) so lets spread
1632 	 * the wealth around.
1633 	 */
1634 	lump = roundup(npages, pcf_fanout) / pcf_fanout;
1635 	freemem += npages;
1636 
1637 	for (p = pcf; (npages > 0) && (p < &pcf[pcf_fanout]); p++) {
1638 		which = &p->pcf_count;
1639 
1640 		mutex_enter(&p->pcf_lock);
1641 
1642 		if (p->pcf_block) {
1643 			which = &p->pcf_reserve;
1644 		}
1645 
1646 		if (lump < npages) {
1647 			*which += (uint_t)lump;
1648 			npages -= lump;
1649 		} else {
1650 			*which += (uint_t)npages;
1651 			npages = 0;
1652 		}
1653 
1654 		if (p->pcf_wait) {
1655 			mutex_enter(&new_freemem_lock);
1656 			/*
1657 			 * Check to see if some other thread
1658 			 * is actually waiting.  Another bucket
1659 			 * may have woken it up by now.  If there
1660 			 * are no waiters, then set our pcf_wait
1661 			 * count to zero to avoid coming in here
1662 			 * next time.
1663 			 */
1664 			if (freemem_wait) {
1665 				if (npages > 1) {
1666 					cv_broadcast(&freemem_cv);
1667 				} else {
1668 					cv_signal(&freemem_cv);
1669 				}
1670 				p->pcf_wait--;
1671 			} else {
1672 				p->pcf_wait = 0;
1673 			}
1674 			mutex_exit(&new_freemem_lock);
1675 		}
1676 		mutex_exit(&p->pcf_lock);
1677 	}
1678 	ASSERT(npages == 0);
1679 }
1680 
1681 /*
1682  * A helper routine for page_create_get_something.
1683  * The indenting got to deep down there.
1684  * Unblock the pcf counters.  Any pages freed after
1685  * pcf_block got set are moved to pcf_count and
1686  * wakeups (cv_broadcast() or cv_signal()) are done as needed.
1687  */
1688 static void
1689 pcgs_unblock(void)
1690 {
1691 	int		i;
1692 	struct pcf	*p;
1693 
1694 	/* Update freemem while we're here. */
1695 	freemem = 0;
1696 	p = pcf;
1697 	for (i = 0; i < pcf_fanout; i++) {
1698 		mutex_enter(&p->pcf_lock);
1699 		ASSERT(p->pcf_count == 0);
1700 		p->pcf_count = p->pcf_reserve;
1701 		p->pcf_block = 0;
1702 		freemem += p->pcf_count;
1703 		if (p->pcf_wait) {
1704 			mutex_enter(&new_freemem_lock);
1705 			if (freemem_wait) {
1706 				if (p->pcf_reserve > 1) {
1707 					cv_broadcast(&freemem_cv);
1708 					p->pcf_wait = 0;
1709 				} else {
1710 					cv_signal(&freemem_cv);
1711 					p->pcf_wait--;
1712 				}
1713 			} else {
1714 				p->pcf_wait = 0;
1715 			}
1716 			mutex_exit(&new_freemem_lock);
1717 		}
1718 		p->pcf_reserve = 0;
1719 		mutex_exit(&p->pcf_lock);
1720 		p++;
1721 	}
1722 }
1723 
1724 /*
1725  * Called from page_create_va() when both the cache and free lists
1726  * have been checked once.
1727  *
1728  * Either returns a page or panics since the accounting was done
1729  * way before we got here.
1730  *
1731  * We don't come here often, so leave the accounting on permanently.
1732  */
1733 
1734 #define	MAX_PCGS	100
1735 
1736 #ifdef	DEBUG
1737 #define	PCGS_TRIES	100
1738 #else	/* DEBUG */
1739 #define	PCGS_TRIES	10
1740 #endif	/* DEBUG */
1741 
1742 #ifdef	VM_STATS
1743 uint_t	pcgs_counts[PCGS_TRIES];
1744 uint_t	pcgs_too_many;
1745 uint_t	pcgs_entered;
1746 uint_t	pcgs_entered_noreloc;
1747 uint_t	pcgs_locked;
1748 uint_t	pcgs_cagelocked;
1749 #endif	/* VM_STATS */
1750 
1751 static page_t *
1752 page_create_get_something(vnode_t *vp, u_offset_t off, struct seg *seg,
1753     caddr_t vaddr, uint_t flags)
1754 {
1755 	uint_t		count;
1756 	page_t		*pp;
1757 	uint_t		locked, i;
1758 	struct	pcf	*p;
1759 	lgrp_t		*lgrp;
1760 	int		cagelocked = 0;
1761 
1762 	VM_STAT_ADD(pcgs_entered);
1763 
1764 	/*
1765 	 * Tap any reserve freelists: if we fail now, we'll die
1766 	 * since the page(s) we're looking for have already been
1767 	 * accounted for.
1768 	 */
1769 	flags |= PG_PANIC;
1770 
1771 	if ((flags & PG_NORELOC) != 0) {
1772 		VM_STAT_ADD(pcgs_entered_noreloc);
1773 		/*
1774 		 * Requests for free pages from critical threads
1775 		 * such as pageout still won't throttle here, but
1776 		 * we must try again, to give the cageout thread
1777 		 * another chance to catch up. Since we already
1778 		 * accounted for the pages, we had better get them
1779 		 * this time.
1780 		 *
1781 		 * N.B. All non-critical threads acquire the pcgs_cagelock
1782 		 * to serialize access to the freelists. This implements a
1783 		 * turnstile-type synchornization to avoid starvation of
1784 		 * critical requests for PG_NORELOC memory by non-critical
1785 		 * threads: all non-critical threads must acquire a 'ticket'
1786 		 * before passing through, which entails making sure
1787 		 * kcage_freemem won't fall below minfree prior to grabbing
1788 		 * pages from the freelists.
1789 		 */
1790 		if (kcage_create_throttle(1, flags) == KCT_NONCRIT) {
1791 			mutex_enter(&pcgs_cagelock);
1792 			cagelocked = 1;
1793 			VM_STAT_ADD(pcgs_cagelocked);
1794 		}
1795 	}
1796 
1797 	/*
1798 	 * Time to get serious.
1799 	 * We failed to get a `correctly colored' page from both the
1800 	 * free and cache lists.
1801 	 * We escalate in stage.
1802 	 *
1803 	 * First try both lists without worring about color.
1804 	 *
1805 	 * Then, grab all page accounting locks (ie. pcf[]) and
1806 	 * steal any pages that they have and set the pcf_block flag to
1807 	 * stop deletions from the lists.  This will help because
1808 	 * a page can get added to the free list while we are looking
1809 	 * at the cache list, then another page could be added to the cache
1810 	 * list allowing the page on the free list to be removed as we
1811 	 * move from looking at the cache list to the free list. This
1812 	 * could happen over and over. We would never find the page
1813 	 * we have accounted for.
1814 	 *
1815 	 * Noreloc pages are a subset of the global (relocatable) page pool.
1816 	 * They are not tracked separately in the pcf bins, so it is
1817 	 * impossible to know when doing pcf accounting if the available
1818 	 * page(s) are noreloc pages or not. When looking for a noreloc page
1819 	 * it is quite easy to end up here even if the global (relocatable)
1820 	 * page pool has plenty of free pages but the noreloc pool is empty.
1821 	 *
1822 	 * When the noreloc pool is empty (or low), additional noreloc pages
1823 	 * are created by converting pages from the global page pool. This
1824 	 * process will stall during pcf accounting if the pcf bins are
1825 	 * already locked. Such is the case when a noreloc allocation is
1826 	 * looping here in page_create_get_something waiting for more noreloc
1827 	 * pages to appear.
1828 	 *
1829 	 * Short of adding a new field to the pcf bins to accurately track
1830 	 * the number of free noreloc pages, we instead do not grab the
1831 	 * pcgs_lock, do not set the pcf blocks and do not timeout when
1832 	 * allocating a noreloc page. This allows noreloc allocations to
1833 	 * loop without blocking global page pool allocations.
1834 	 *
1835 	 * NOTE: the behaviour of page_create_get_something has not changed
1836 	 * for the case of global page pool allocations.
1837 	 */
1838 
1839 	flags &= ~PG_MATCH_COLOR;
1840 	locked = 0;
1841 #if defined(__i386) || defined(__amd64)
1842 	flags = page_create_update_flags_x86(flags);
1843 #endif
1844 
1845 	lgrp = lgrp_mem_choose(seg, vaddr, PAGESIZE);
1846 
1847 	for (count = 0; kcage_on || count < MAX_PCGS; count++) {
1848 		pp = page_get_freelist(vp, off, seg, vaddr, PAGESIZE,
1849 		    flags, lgrp);
1850 		if (pp == NULL) {
1851 			pp = page_get_cachelist(vp, off, seg, vaddr,
1852 			    flags, lgrp);
1853 		}
1854 		if (pp == NULL) {
1855 			/*
1856 			 * Serialize.  Don't fight with other pcgs().
1857 			 */
1858 			if (!locked && (!kcage_on || !(flags & PG_NORELOC))) {
1859 				mutex_enter(&pcgs_lock);
1860 				VM_STAT_ADD(pcgs_locked);
1861 				locked = 1;
1862 				p = pcf;
1863 				for (i = 0; i < pcf_fanout; i++) {
1864 					mutex_enter(&p->pcf_lock);
1865 					ASSERT(p->pcf_block == 0);
1866 					p->pcf_block = 1;
1867 					p->pcf_reserve = p->pcf_count;
1868 					p->pcf_count = 0;
1869 					mutex_exit(&p->pcf_lock);
1870 					p++;
1871 				}
1872 				freemem = 0;
1873 			}
1874 
1875 			if (count) {
1876 				/*
1877 				 * Since page_free() puts pages on
1878 				 * a list then accounts for it, we
1879 				 * just have to wait for page_free()
1880 				 * to unlock any page it was working
1881 				 * with. The page_lock()-page_reclaim()
1882 				 * path falls in the same boat.
1883 				 *
1884 				 * We don't need to check on the
1885 				 * PG_WAIT flag, we have already
1886 				 * accounted for the page we are
1887 				 * looking for in page_create_va().
1888 				 *
1889 				 * We just wait a moment to let any
1890 				 * locked pages on the lists free up,
1891 				 * then continue around and try again.
1892 				 *
1893 				 * Will be awakened by set_freemem().
1894 				 */
1895 				mutex_enter(&pcgs_wait_lock);
1896 				cv_wait(&pcgs_cv, &pcgs_wait_lock);
1897 				mutex_exit(&pcgs_wait_lock);
1898 			}
1899 		} else {
1900 #ifdef VM_STATS
1901 			if (count >= PCGS_TRIES) {
1902 				VM_STAT_ADD(pcgs_too_many);
1903 			} else {
1904 				VM_STAT_ADD(pcgs_counts[count]);
1905 			}
1906 #endif
1907 			if (locked) {
1908 				pcgs_unblock();
1909 				mutex_exit(&pcgs_lock);
1910 			}
1911 			if (cagelocked)
1912 				mutex_exit(&pcgs_cagelock);
1913 			return (pp);
1914 		}
1915 	}
1916 	/*
1917 	 * we go down holding the pcf locks.
1918 	 */
1919 	panic("no %spage found %d",
1920 	    ((flags & PG_NORELOC) ? "non-reloc " : ""), count);
1921 	/*NOTREACHED*/
1922 }
1923 
1924 /*
1925  * Create enough pages for "bytes" worth of data starting at
1926  * "off" in "vp".
1927  *
1928  *	Where flag must be one of:
1929  *
1930  *		PG_EXCL:	Exclusive create (fail if any page already
1931  *				exists in the page cache) which does not
1932  *				wait for memory to become available.
1933  *
1934  *		PG_WAIT:	Non-exclusive create which can wait for
1935  *				memory to become available.
1936  *
1937  *		PG_PHYSCONTIG:	Allocate physically contiguous pages.
1938  *				(Not Supported)
1939  *
1940  * A doubly linked list of pages is returned to the caller.  Each page
1941  * on the list has the "exclusive" (p_selock) lock and "iolock" (p_iolock)
1942  * lock.
1943  *
1944  * Unable to change the parameters to page_create() in a minor release,
1945  * we renamed page_create() to page_create_va(), changed all known calls
1946  * from page_create() to page_create_va(), and created this wrapper.
1947  *
1948  * Upon a major release, we should break compatibility by deleting this
1949  * wrapper, and replacing all the strings "page_create_va", with "page_create".
1950  *
1951  * NOTE: There is a copy of this interface as page_create_io() in
1952  *	 i86/vm/vm_machdep.c. Any bugs fixed here should be applied
1953  *	 there.
1954  */
1955 page_t *
1956 page_create(vnode_t *vp, u_offset_t off, size_t bytes, uint_t flags)
1957 {
1958 	caddr_t random_vaddr;
1959 	struct seg kseg;
1960 
1961 #ifdef DEBUG
1962 	cmn_err(CE_WARN, "Using deprecated interface page_create: caller %p",
1963 	    (void *)caller());
1964 #endif
1965 
1966 	random_vaddr = (caddr_t)(((uintptr_t)vp >> 7) ^
1967 	    (uintptr_t)(off >> PAGESHIFT));
1968 	kseg.s_as = &kas;
1969 
1970 	return (page_create_va(vp, off, bytes, flags, &kseg, random_vaddr));
1971 }
1972 
1973 #ifdef DEBUG
1974 uint32_t pg_alloc_pgs_mtbf = 0;
1975 #endif
1976 
1977 /*
1978  * Used for large page support. It will attempt to allocate
1979  * a large page(s) off the freelist.
1980  *
1981  * Returns non zero on failure.
1982  */
1983 int
1984 page_alloc_pages(struct vnode *vp, struct seg *seg, caddr_t addr,
1985     page_t **basepp, page_t *ppa[], uint_t szc, int anypgsz, int pgflags)
1986 {
1987 	pgcnt_t		npgs, curnpgs, totpgs;
1988 	size_t		pgsz;
1989 	page_t		*pplist = NULL, *pp;
1990 	int		err = 0;
1991 	lgrp_t		*lgrp;
1992 
1993 	ASSERT(szc != 0 && szc <= (page_num_pagesizes() - 1));
1994 	ASSERT(pgflags == 0 || pgflags == PG_LOCAL);
1995 
1996 	/*
1997 	 * Check if system heavily prefers local large pages over remote
1998 	 * on systems with multiple lgroups.
1999 	 */
2000 	if (lpg_alloc_prefer == LPAP_LOCAL && nlgrps > 1) {
2001 		pgflags = PG_LOCAL;
2002 	}
2003 
2004 	VM_STAT_ADD(alloc_pages[0]);
2005 
2006 #ifdef DEBUG
2007 	if (pg_alloc_pgs_mtbf && !(gethrtime() % pg_alloc_pgs_mtbf)) {
2008 		return (ENOMEM);
2009 	}
2010 #endif
2011 
2012 	/*
2013 	 * One must be NULL but not both.
2014 	 * And one must be non NULL but not both.
2015 	 */
2016 	ASSERT(basepp != NULL || ppa != NULL);
2017 	ASSERT(basepp == NULL || ppa == NULL);
2018 
2019 #if defined(__i386) || defined(__amd64)
2020 	while (page_chk_freelist(szc) == 0) {
2021 		VM_STAT_ADD(alloc_pages[8]);
2022 		if (anypgsz == 0 || --szc == 0)
2023 			return (ENOMEM);
2024 	}
2025 #endif
2026 
2027 	pgsz = page_get_pagesize(szc);
2028 	totpgs = curnpgs = npgs = pgsz >> PAGESHIFT;
2029 
2030 	ASSERT(((uintptr_t)addr & (pgsz - 1)) == 0);
2031 
2032 	(void) page_create_wait(npgs, PG_WAIT);
2033 
2034 	while (npgs && szc) {
2035 		lgrp = lgrp_mem_choose(seg, addr, pgsz);
2036 		if (pgflags == PG_LOCAL) {
2037 			pp = page_get_freelist(vp, 0, seg, addr, pgsz,
2038 			    pgflags, lgrp);
2039 			if (pp == NULL) {
2040 				pp = page_get_freelist(vp, 0, seg, addr, pgsz,
2041 				    0, lgrp);
2042 			}
2043 		} else {
2044 			pp = page_get_freelist(vp, 0, seg, addr, pgsz,
2045 			    0, lgrp);
2046 		}
2047 		if (pp != NULL) {
2048 			VM_STAT_ADD(alloc_pages[1]);
2049 			page_list_concat(&pplist, &pp);
2050 			ASSERT(npgs >= curnpgs);
2051 			npgs -= curnpgs;
2052 		} else if (anypgsz) {
2053 			VM_STAT_ADD(alloc_pages[2]);
2054 			szc--;
2055 			pgsz = page_get_pagesize(szc);
2056 			curnpgs = pgsz >> PAGESHIFT;
2057 		} else {
2058 			VM_STAT_ADD(alloc_pages[3]);
2059 			ASSERT(npgs == totpgs);
2060 			page_create_putback(npgs);
2061 			return (ENOMEM);
2062 		}
2063 	}
2064 	if (szc == 0) {
2065 		VM_STAT_ADD(alloc_pages[4]);
2066 		ASSERT(npgs != 0);
2067 		page_create_putback(npgs);
2068 		err = ENOMEM;
2069 	} else if (basepp != NULL) {
2070 		ASSERT(npgs == 0);
2071 		ASSERT(ppa == NULL);
2072 		*basepp = pplist;
2073 	}
2074 
2075 	npgs = totpgs - npgs;
2076 	pp = pplist;
2077 
2078 	/*
2079 	 * Clear the free and age bits. Also if we were passed in a ppa then
2080 	 * fill it in with all the constituent pages from the large page. But
2081 	 * if we failed to allocate all the pages just free what we got.
2082 	 */
2083 	while (npgs != 0) {
2084 		ASSERT(PP_ISFREE(pp));
2085 		ASSERT(PP_ISAGED(pp));
2086 		if (ppa != NULL || err != 0) {
2087 			if (err == 0) {
2088 				VM_STAT_ADD(alloc_pages[5]);
2089 				PP_CLRFREE(pp);
2090 				PP_CLRAGED(pp);
2091 				page_sub(&pplist, pp);
2092 				*ppa++ = pp;
2093 				npgs--;
2094 			} else {
2095 				VM_STAT_ADD(alloc_pages[6]);
2096 				ASSERT(pp->p_szc != 0);
2097 				curnpgs = page_get_pagecnt(pp->p_szc);
2098 				page_list_break(&pp, &pplist, curnpgs);
2099 				page_list_add_pages(pp, 0);
2100 				page_create_putback(curnpgs);
2101 				ASSERT(npgs >= curnpgs);
2102 				npgs -= curnpgs;
2103 			}
2104 			pp = pplist;
2105 		} else {
2106 			VM_STAT_ADD(alloc_pages[7]);
2107 			PP_CLRFREE(pp);
2108 			PP_CLRAGED(pp);
2109 			pp = pp->p_next;
2110 			npgs--;
2111 		}
2112 	}
2113 	return (err);
2114 }
2115 
2116 /*
2117  * Get a single large page off of the freelists, and set it up for use.
2118  * Number of bytes requested must be a supported page size.
2119  *
2120  * Note that this call may fail even if there is sufficient
2121  * memory available or PG_WAIT is set, so the caller must
2122  * be willing to fallback on page_create_va(), block and retry,
2123  * or fail the requester.
2124  */
2125 page_t *
2126 page_create_va_large(vnode_t *vp, u_offset_t off, size_t bytes, uint_t flags,
2127     struct seg *seg, caddr_t vaddr, void *arg)
2128 {
2129 	pgcnt_t		npages;
2130 	page_t		*pp;
2131 	page_t		*rootpp;
2132 	lgrp_t		*lgrp;
2133 	lgrp_id_t	*lgrpid = (lgrp_id_t *)arg;
2134 
2135 	ASSERT(vp != NULL);
2136 
2137 	ASSERT((flags & ~(PG_EXCL | PG_WAIT |
2138 	    PG_NORELOC | PG_PANIC | PG_PUSHPAGE)) == 0);
2139 	/* but no others */
2140 
2141 	ASSERT((flags & PG_EXCL) == PG_EXCL);
2142 
2143 	npages = btop(bytes);
2144 
2145 	if (!kcage_on || panicstr) {
2146 		/*
2147 		 * Cage is OFF, or we are single threaded in
2148 		 * panic, so make everything a RELOC request.
2149 		 */
2150 		flags &= ~PG_NORELOC;
2151 	}
2152 
2153 	/*
2154 	 * Make sure there's adequate physical memory available.
2155 	 * Note: PG_WAIT is ignored here.
2156 	 */
2157 	if (freemem <= throttlefree + npages) {
2158 		VM_STAT_ADD(page_create_large_cnt[1]);
2159 		return (NULL);
2160 	}
2161 
2162 	/*
2163 	 * If cage is on, dampen draw from cage when available
2164 	 * cage space is low.
2165 	 */
2166 	if ((flags & (PG_NORELOC | PG_WAIT)) ==  (PG_NORELOC | PG_WAIT) &&
2167 	    kcage_freemem < kcage_throttlefree + npages) {
2168 
2169 		/*
2170 		 * The cage is on, the caller wants PG_NORELOC
2171 		 * pages and available cage memory is very low.
2172 		 * Call kcage_create_throttle() to attempt to
2173 		 * control demand on the cage.
2174 		 */
2175 		if (kcage_create_throttle(npages, flags) == KCT_FAILURE) {
2176 			VM_STAT_ADD(page_create_large_cnt[2]);
2177 			return (NULL);
2178 		}
2179 	}
2180 
2181 	if (!pcf_decrement_bucket(npages) &&
2182 	    !pcf_decrement_multiple(NULL, npages, 1)) {
2183 		VM_STAT_ADD(page_create_large_cnt[4]);
2184 		return (NULL);
2185 	}
2186 
2187 	/*
2188 	 * This is where this function behaves fundamentally differently
2189 	 * than page_create_va(); since we're intending to map the page
2190 	 * with a single TTE, we have to get it as a physically contiguous
2191 	 * hardware pagesize chunk.  If we can't, we fail.
2192 	 */
2193 	if (lgrpid != NULL && *lgrpid >= 0 && *lgrpid <= lgrp_alloc_max &&
2194 	    LGRP_EXISTS(lgrp_table[*lgrpid]))
2195 		lgrp = lgrp_table[*lgrpid];
2196 	else
2197 		lgrp = lgrp_mem_choose(seg, vaddr, bytes);
2198 
2199 	if ((rootpp = page_get_freelist(&kvp, off, seg, vaddr,
2200 	    bytes, flags & ~PG_MATCH_COLOR, lgrp)) == NULL) {
2201 		page_create_putback(npages);
2202 		VM_STAT_ADD(page_create_large_cnt[5]);
2203 		return (NULL);
2204 	}
2205 
2206 	/*
2207 	 * if we got the page with the wrong mtype give it back this is a
2208 	 * workaround for CR 6249718. When CR 6249718 is fixed we never get
2209 	 * inside "if" and the workaround becomes just a nop
2210 	 */
2211 	if (kcage_on && (flags & PG_NORELOC) && !PP_ISNORELOC(rootpp)) {
2212 		page_list_add_pages(rootpp, 0);
2213 		page_create_putback(npages);
2214 		VM_STAT_ADD(page_create_large_cnt[6]);
2215 		return (NULL);
2216 	}
2217 
2218 	/*
2219 	 * If satisfying this request has left us with too little
2220 	 * memory, start the wheels turning to get some back.  The
2221 	 * first clause of the test prevents waking up the pageout
2222 	 * daemon in situations where it would decide that there's
2223 	 * nothing to do.
2224 	 */
2225 	if (nscan < desscan && freemem < minfree) {
2226 		TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL,
2227 		    "pageout_cv_signal:freemem %ld", freemem);
2228 		cv_signal(&proc_pageout->p_cv);
2229 	}
2230 
2231 	pp = rootpp;
2232 	while (npages--) {
2233 		ASSERT(PAGE_EXCL(pp));
2234 		ASSERT(pp->p_vnode == NULL);
2235 		ASSERT(!hat_page_is_mapped(pp));
2236 		PP_CLRFREE(pp);
2237 		PP_CLRAGED(pp);
2238 		if (!page_hashin(pp, vp, off, NULL))
2239 			panic("page_create_large: hashin failed: page %p",
2240 			    (void *)pp);
2241 		page_io_lock(pp);
2242 		off += PAGESIZE;
2243 		pp = pp->p_next;
2244 	}
2245 
2246 	VM_STAT_ADD(page_create_large_cnt[0]);
2247 	return (rootpp);
2248 }
2249 
2250 page_t *
2251 page_create_va(vnode_t *vp, u_offset_t off, size_t bytes, uint_t flags,
2252     struct seg *seg, caddr_t vaddr)
2253 {
2254 	page_t		*plist = NULL;
2255 	pgcnt_t		npages;
2256 	pgcnt_t		found_on_free = 0;
2257 	pgcnt_t		pages_req;
2258 	page_t		*npp = NULL;
2259 	struct pcf	*p;
2260 	lgrp_t		*lgrp;
2261 
2262 	TRACE_4(TR_FAC_VM, TR_PAGE_CREATE_START,
2263 	    "page_create_start:vp %p off %llx bytes %lu flags %x",
2264 	    vp, off, bytes, flags);
2265 
2266 	ASSERT(bytes != 0 && vp != NULL);
2267 
2268 	if ((flags & PG_EXCL) == 0 && (flags & PG_WAIT) == 0) {
2269 		panic("page_create: invalid flags");
2270 		/*NOTREACHED*/
2271 	}
2272 	ASSERT((flags & ~(PG_EXCL | PG_WAIT |
2273 	    PG_NORELOC | PG_PANIC | PG_PUSHPAGE)) == 0);
2274 	    /* but no others */
2275 
2276 	pages_req = npages = btopr(bytes);
2277 	/*
2278 	 * Try to see whether request is too large to *ever* be
2279 	 * satisfied, in order to prevent deadlock.  We arbitrarily
2280 	 * decide to limit maximum size requests to max_page_get.
2281 	 */
2282 	if (npages >= max_page_get) {
2283 		if ((flags & PG_WAIT) == 0) {
2284 			TRACE_4(TR_FAC_VM, TR_PAGE_CREATE_TOOBIG,
2285 			    "page_create_toobig:vp %p off %llx npages "
2286 			    "%lu max_page_get %lu",
2287 			    vp, off, npages, max_page_get);
2288 			return (NULL);
2289 		} else {
2290 			cmn_err(CE_WARN,
2291 			    "Request for too much kernel memory "
2292 			    "(%lu bytes), will hang forever", bytes);
2293 			for (;;)
2294 				delay(1000000000);
2295 		}
2296 	}
2297 
2298 	if (!kcage_on || panicstr) {
2299 		/*
2300 		 * Cage is OFF, or we are single threaded in
2301 		 * panic, so make everything a RELOC request.
2302 		 */
2303 		flags &= ~PG_NORELOC;
2304 	}
2305 
2306 	if (freemem <= throttlefree + npages)
2307 		if (!page_create_throttle(npages, flags))
2308 			return (NULL);
2309 
2310 	/*
2311 	 * If cage is on, dampen draw from cage when available
2312 	 * cage space is low.
2313 	 */
2314 	if ((flags & PG_NORELOC) &&
2315 	    kcage_freemem < kcage_throttlefree + npages) {
2316 
2317 		/*
2318 		 * The cage is on, the caller wants PG_NORELOC
2319 		 * pages and available cage memory is very low.
2320 		 * Call kcage_create_throttle() to attempt to
2321 		 * control demand on the cage.
2322 		 */
2323 		if (kcage_create_throttle(npages, flags) == KCT_FAILURE)
2324 			return (NULL);
2325 	}
2326 
2327 	VM_STAT_ADD(page_create_cnt[0]);
2328 
2329 	if (!pcf_decrement_bucket(npages)) {
2330 		/*
2331 		 * Have to look harder.  If npages is greater than
2332 		 * one, then we might have to coalesce the counters.
2333 		 *
2334 		 * Go wait.  We come back having accounted
2335 		 * for the memory.
2336 		 */
2337 		VM_STAT_ADD(page_create_cnt[1]);
2338 		if (!page_create_wait(npages, flags)) {
2339 			VM_STAT_ADD(page_create_cnt[2]);
2340 			return (NULL);
2341 		}
2342 	}
2343 
2344 	TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SUCCESS,
2345 	    "page_create_success:vp %p off %llx", vp, off);
2346 
2347 	/*
2348 	 * If satisfying this request has left us with too little
2349 	 * memory, start the wheels turning to get some back.  The
2350 	 * first clause of the test prevents waking up the pageout
2351 	 * daemon in situations where it would decide that there's
2352 	 * nothing to do.
2353 	 */
2354 	if (nscan < desscan && freemem < minfree) {
2355 		TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL,
2356 		    "pageout_cv_signal:freemem %ld", freemem);
2357 		cv_signal(&proc_pageout->p_cv);
2358 	}
2359 
2360 	/*
2361 	 * Loop around collecting the requested number of pages.
2362 	 * Most of the time, we have to `create' a new page. With
2363 	 * this in mind, pull the page off the free list before
2364 	 * getting the hash lock.  This will minimize the hash
2365 	 * lock hold time, nesting, and the like.  If it turns
2366 	 * out we don't need the page, we put it back at the end.
2367 	 */
2368 	while (npages--) {
2369 		page_t		*pp;
2370 		kmutex_t	*phm = NULL;
2371 		ulong_t		index;
2372 
2373 		index = PAGE_HASH_FUNC(vp, off);
2374 top:
2375 		ASSERT(phm == NULL);
2376 		ASSERT(index == PAGE_HASH_FUNC(vp, off));
2377 		ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
2378 
2379 		if (npp == NULL) {
2380 			/*
2381 			 * Try to get a page from the freelist (ie,
2382 			 * a page with no [vp, off] tag).  If that
2383 			 * fails, use the cachelist.
2384 			 *
2385 			 * During the first attempt at both the free
2386 			 * and cache lists we try for the correct color.
2387 			 */
2388 			/*
2389 			 * XXXX-how do we deal with virtual indexed
2390 			 * caches and and colors?
2391 			 */
2392 			VM_STAT_ADD(page_create_cnt[4]);
2393 			/*
2394 			 * Get lgroup to allocate next page of shared memory
2395 			 * from and use it to specify where to allocate
2396 			 * the physical memory
2397 			 */
2398 			lgrp = lgrp_mem_choose(seg, vaddr, PAGESIZE);
2399 			npp = page_get_freelist(vp, off, seg, vaddr, PAGESIZE,
2400 			    flags | PG_MATCH_COLOR, lgrp);
2401 			if (npp == NULL) {
2402 				npp = page_get_cachelist(vp, off, seg,
2403 				    vaddr, flags | PG_MATCH_COLOR, lgrp);
2404 				if (npp == NULL) {
2405 					npp = page_create_get_something(vp,
2406 					    off, seg, vaddr,
2407 					    flags & ~PG_MATCH_COLOR);
2408 				}
2409 
2410 				if (PP_ISAGED(npp) == 0) {
2411 					/*
2412 					 * Since this page came from the
2413 					 * cachelist, we must destroy the
2414 					 * old vnode association.
2415 					 */
2416 					page_hashout(npp, NULL);
2417 				}
2418 			}
2419 		}
2420 
2421 		/*
2422 		 * We own this page!
2423 		 */
2424 		ASSERT(PAGE_EXCL(npp));
2425 		ASSERT(npp->p_vnode == NULL);
2426 		ASSERT(!hat_page_is_mapped(npp));
2427 		PP_CLRFREE(npp);
2428 		PP_CLRAGED(npp);
2429 
2430 		/*
2431 		 * Here we have a page in our hot little mits and are
2432 		 * just waiting to stuff it on the appropriate lists.
2433 		 * Get the mutex and check to see if it really does
2434 		 * not exist.
2435 		 */
2436 		phm = PAGE_HASH_MUTEX(index);
2437 		mutex_enter(phm);
2438 		PAGE_HASH_SEARCH(index, pp, vp, off);
2439 		if (pp == NULL) {
2440 			VM_STAT_ADD(page_create_new);
2441 			pp = npp;
2442 			npp = NULL;
2443 			if (!page_hashin(pp, vp, off, phm)) {
2444 				/*
2445 				 * Since we hold the page hash mutex and
2446 				 * just searched for this page, page_hashin
2447 				 * had better not fail.  If it does, that
2448 				 * means somethread did not follow the
2449 				 * page hash mutex rules.  Panic now and
2450 				 * get it over with.  As usual, go down
2451 				 * holding all the locks.
2452 				 */
2453 				ASSERT(MUTEX_HELD(phm));
2454 				panic("page_create: "
2455 				    "hashin failed %p %p %llx %p",
2456 				    (void *)pp, (void *)vp, off, (void *)phm);
2457 				/*NOTREACHED*/
2458 			}
2459 			ASSERT(MUTEX_HELD(phm));
2460 			mutex_exit(phm);
2461 			phm = NULL;
2462 
2463 			/*
2464 			 * Hat layer locking need not be done to set
2465 			 * the following bits since the page is not hashed
2466 			 * and was on the free list (i.e., had no mappings).
2467 			 *
2468 			 * Set the reference bit to protect
2469 			 * against immediate pageout
2470 			 *
2471 			 * XXXmh modify freelist code to set reference
2472 			 * bit so we don't have to do it here.
2473 			 */
2474 			page_set_props(pp, P_REF);
2475 			found_on_free++;
2476 		} else {
2477 			VM_STAT_ADD(page_create_exists);
2478 			if (flags & PG_EXCL) {
2479 				/*
2480 				 * Found an existing page, and the caller
2481 				 * wanted all new pages.  Undo all of the work
2482 				 * we have done.
2483 				 */
2484 				mutex_exit(phm);
2485 				phm = NULL;
2486 				while (plist != NULL) {
2487 					pp = plist;
2488 					page_sub(&plist, pp);
2489 					page_io_unlock(pp);
2490 					/* large pages should not end up here */
2491 					ASSERT(pp->p_szc == 0);
2492 					/*LINTED: constant in conditional ctx*/
2493 					VN_DISPOSE(pp, B_INVAL, 0, kcred);
2494 				}
2495 				VM_STAT_ADD(page_create_found_one);
2496 				goto fail;
2497 			}
2498 			ASSERT(flags & PG_WAIT);
2499 			if (!page_lock(pp, SE_EXCL, phm, P_NO_RECLAIM)) {
2500 				/*
2501 				 * Start all over again if we blocked trying
2502 				 * to lock the page.
2503 				 */
2504 				mutex_exit(phm);
2505 				VM_STAT_ADD(page_create_page_lock_failed);
2506 				phm = NULL;
2507 				goto top;
2508 			}
2509 			mutex_exit(phm);
2510 			phm = NULL;
2511 
2512 			if (PP_ISFREE(pp)) {
2513 				ASSERT(PP_ISAGED(pp) == 0);
2514 				VM_STAT_ADD(pagecnt.pc_get_cache);
2515 				page_list_sub(pp, PG_CACHE_LIST);
2516 				PP_CLRFREE(pp);
2517 				found_on_free++;
2518 			}
2519 		}
2520 
2521 		/*
2522 		 * Got a page!  It is locked.  Acquire the i/o
2523 		 * lock since we are going to use the p_next and
2524 		 * p_prev fields to link the requested pages together.
2525 		 */
2526 		page_io_lock(pp);
2527 		page_add(&plist, pp);
2528 		plist = plist->p_next;
2529 		off += PAGESIZE;
2530 		vaddr += PAGESIZE;
2531 	}
2532 
2533 	ASSERT((flags & PG_EXCL) ? (found_on_free == pages_req) : 1);
2534 fail:
2535 	if (npp != NULL) {
2536 		/*
2537 		 * Did not need this page after all.
2538 		 * Put it back on the free list.
2539 		 */
2540 		VM_STAT_ADD(page_create_putbacks);
2541 		PP_SETFREE(npp);
2542 		PP_SETAGED(npp);
2543 		npp->p_offset = (u_offset_t)-1;
2544 		page_list_add(npp, PG_FREE_LIST | PG_LIST_TAIL);
2545 		page_unlock(npp);
2546 
2547 	}
2548 
2549 	ASSERT(pages_req >= found_on_free);
2550 
2551 	{
2552 		uint_t overshoot = (uint_t)(pages_req - found_on_free);
2553 
2554 		if (overshoot) {
2555 			VM_STAT_ADD(page_create_overshoot);
2556 			p = &pcf[PCF_INDEX()];
2557 			mutex_enter(&p->pcf_lock);
2558 			if (p->pcf_block) {
2559 				p->pcf_reserve += overshoot;
2560 			} else {
2561 				p->pcf_count += overshoot;
2562 				if (p->pcf_wait) {
2563 					mutex_enter(&new_freemem_lock);
2564 					if (freemem_wait) {
2565 						cv_signal(&freemem_cv);
2566 						p->pcf_wait--;
2567 					} else {
2568 						p->pcf_wait = 0;
2569 					}
2570 					mutex_exit(&new_freemem_lock);
2571 				}
2572 			}
2573 			mutex_exit(&p->pcf_lock);
2574 			/* freemem is approximate, so this test OK */
2575 			if (!p->pcf_block)
2576 				freemem += overshoot;
2577 		}
2578 	}
2579 
2580 	return (plist);
2581 }
2582 
2583 /*
2584  * One or more constituent pages of this large page has been marked
2585  * toxic. Simply demote the large page to PAGESIZE pages and let
2586  * page_free() handle it. This routine should only be called by
2587  * large page free routines (page_free_pages() and page_destroy_pages().
2588  * All pages are locked SE_EXCL and have already been marked free.
2589  */
2590 static void
2591 page_free_toxic_pages(page_t *rootpp)
2592 {
2593 	page_t	*tpp;
2594 	pgcnt_t	i, pgcnt = page_get_pagecnt(rootpp->p_szc);
2595 	uint_t	szc = rootpp->p_szc;
2596 
2597 	for (i = 0, tpp = rootpp; i < pgcnt; i++, tpp = tpp->p_next) {
2598 		ASSERT(tpp->p_szc == szc);
2599 		ASSERT((PAGE_EXCL(tpp) &&
2600 		    !page_iolock_assert(tpp)) || panicstr);
2601 		tpp->p_szc = 0;
2602 	}
2603 
2604 	while (rootpp != NULL) {
2605 		tpp = rootpp;
2606 		page_sub(&rootpp, tpp);
2607 		ASSERT(PP_ISFREE(tpp));
2608 		PP_CLRFREE(tpp);
2609 		page_free(tpp, 1);
2610 	}
2611 }
2612 
2613 /*
2614  * Put page on the "free" list.
2615  * The free list is really two lists maintained by
2616  * the PSM of whatever machine we happen to be on.
2617  */
2618 void
2619 page_free(page_t *pp, int dontneed)
2620 {
2621 	struct pcf	*p;
2622 	uint_t		pcf_index;
2623 
2624 	ASSERT((PAGE_EXCL(pp) &&
2625 	    !page_iolock_assert(pp)) || panicstr);
2626 
2627 	if (PP_ISFREE(pp)) {
2628 		panic("page_free: page %p is free", (void *)pp);
2629 	}
2630 
2631 	if (pp->p_szc != 0) {
2632 		if (pp->p_vnode == NULL || IS_SWAPFSVP(pp->p_vnode) ||
2633 		    PP_ISKAS(pp)) {
2634 			panic("page_free: anon or kernel "
2635 			    "or no vnode large page %p", (void *)pp);
2636 		}
2637 		page_demote_vp_pages(pp);
2638 		ASSERT(pp->p_szc == 0);
2639 	}
2640 
2641 	/*
2642 	 * The page_struct_lock need not be acquired to examine these
2643 	 * fields since the page has an "exclusive" lock.
2644 	 */
2645 	if (hat_page_is_mapped(pp) || pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
2646 	    pp->p_slckcnt != 0) {
2647 		panic("page_free pp=%p, pfn=%lx, lckcnt=%d, cowcnt=%d "
2648 		    "slckcnt = %d", pp, page_pptonum(pp), pp->p_lckcnt,
2649 		    pp->p_cowcnt, pp->p_slckcnt);
2650 		/*NOTREACHED*/
2651 	}
2652 
2653 	ASSERT(!hat_page_getshare(pp));
2654 
2655 	PP_SETFREE(pp);
2656 	ASSERT(pp->p_vnode == NULL || !IS_VMODSORT(pp->p_vnode) ||
2657 	    !hat_ismod(pp));
2658 	page_clr_all_props(pp);
2659 	ASSERT(!hat_page_getshare(pp));
2660 
2661 	/*
2662 	 * Now we add the page to the head of the free list.
2663 	 * But if this page is associated with a paged vnode
2664 	 * then we adjust the head forward so that the page is
2665 	 * effectively at the end of the list.
2666 	 */
2667 	if (pp->p_vnode == NULL) {
2668 		/*
2669 		 * Page has no identity, put it on the free list.
2670 		 */
2671 		PP_SETAGED(pp);
2672 		pp->p_offset = (u_offset_t)-1;
2673 		page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL);
2674 		VM_STAT_ADD(pagecnt.pc_free_free);
2675 		TRACE_1(TR_FAC_VM, TR_PAGE_FREE_FREE,
2676 		    "page_free_free:pp %p", pp);
2677 	} else {
2678 		PP_CLRAGED(pp);
2679 
2680 		if (!dontneed || nopageage) {
2681 			/* move it to the tail of the list */
2682 			page_list_add(pp, PG_CACHE_LIST | PG_LIST_TAIL);
2683 
2684 			VM_STAT_ADD(pagecnt.pc_free_cache);
2685 			TRACE_1(TR_FAC_VM, TR_PAGE_FREE_CACHE_TAIL,
2686 			    "page_free_cache_tail:pp %p", pp);
2687 		} else {
2688 			page_list_add(pp, PG_CACHE_LIST | PG_LIST_HEAD);
2689 
2690 			VM_STAT_ADD(pagecnt.pc_free_dontneed);
2691 			TRACE_1(TR_FAC_VM, TR_PAGE_FREE_CACHE_HEAD,
2692 			    "page_free_cache_head:pp %p", pp);
2693 		}
2694 	}
2695 	page_unlock(pp);
2696 
2697 	/*
2698 	 * Now do the `freemem' accounting.
2699 	 */
2700 	pcf_index = PCF_INDEX();
2701 	p = &pcf[pcf_index];
2702 
2703 	mutex_enter(&p->pcf_lock);
2704 	if (p->pcf_block) {
2705 		p->pcf_reserve += 1;
2706 	} else {
2707 		p->pcf_count += 1;
2708 		if (p->pcf_wait) {
2709 			mutex_enter(&new_freemem_lock);
2710 			/*
2711 			 * Check to see if some other thread
2712 			 * is actually waiting.  Another bucket
2713 			 * may have woken it up by now.  If there
2714 			 * are no waiters, then set our pcf_wait
2715 			 * count to zero to avoid coming in here
2716 			 * next time.  Also, since only one page
2717 			 * was put on the free list, just wake
2718 			 * up one waiter.
2719 			 */
2720 			if (freemem_wait) {
2721 				cv_signal(&freemem_cv);
2722 				p->pcf_wait--;
2723 			} else {
2724 				p->pcf_wait = 0;
2725 			}
2726 			mutex_exit(&new_freemem_lock);
2727 		}
2728 	}
2729 	mutex_exit(&p->pcf_lock);
2730 
2731 	/* freemem is approximate, so this test OK */
2732 	if (!p->pcf_block)
2733 		freemem += 1;
2734 }
2735 
2736 /*
2737  * Put page on the "free" list during intial startup.
2738  * This happens during initial single threaded execution.
2739  */
2740 void
2741 page_free_at_startup(page_t *pp)
2742 {
2743 	struct pcf	*p;
2744 	uint_t		pcf_index;
2745 
2746 	page_list_add(pp, PG_FREE_LIST | PG_LIST_HEAD | PG_LIST_ISINIT);
2747 	VM_STAT_ADD(pagecnt.pc_free_free);
2748 
2749 	/*
2750 	 * Now do the `freemem' accounting.
2751 	 */
2752 	pcf_index = PCF_INDEX();
2753 	p = &pcf[pcf_index];
2754 
2755 	ASSERT(p->pcf_block == 0);
2756 	ASSERT(p->pcf_wait == 0);
2757 	p->pcf_count += 1;
2758 
2759 	/* freemem is approximate, so this is OK */
2760 	freemem += 1;
2761 }
2762 
2763 void
2764 page_free_pages(page_t *pp)
2765 {
2766 	page_t	*tpp, *rootpp = NULL;
2767 	pgcnt_t	pgcnt = page_get_pagecnt(pp->p_szc);
2768 	pgcnt_t	i;
2769 	uint_t	szc = pp->p_szc;
2770 
2771 	VM_STAT_ADD(pagecnt.pc_free_pages);
2772 	TRACE_1(TR_FAC_VM, TR_PAGE_FREE_FREE,
2773 	    "page_free_free:pp %p", pp);
2774 
2775 	ASSERT(pp->p_szc != 0 && pp->p_szc < page_num_pagesizes());
2776 	if ((page_pptonum(pp) & (pgcnt - 1)) != 0) {
2777 		panic("page_free_pages: not root page %p", (void *)pp);
2778 		/*NOTREACHED*/
2779 	}
2780 
2781 	for (i = 0, tpp = pp; i < pgcnt; i++, tpp++) {
2782 		ASSERT((PAGE_EXCL(tpp) &&
2783 		    !page_iolock_assert(tpp)) || panicstr);
2784 		if (PP_ISFREE(tpp)) {
2785 			panic("page_free_pages: page %p is free", (void *)tpp);
2786 			/*NOTREACHED*/
2787 		}
2788 		if (hat_page_is_mapped(tpp) || tpp->p_lckcnt != 0 ||
2789 		    tpp->p_cowcnt != 0 || tpp->p_slckcnt != 0) {
2790 			panic("page_free_pages %p", (void *)tpp);
2791 			/*NOTREACHED*/
2792 		}
2793 
2794 		ASSERT(!hat_page_getshare(tpp));
2795 		ASSERT(tpp->p_vnode == NULL);
2796 		ASSERT(tpp->p_szc == szc);
2797 
2798 		PP_SETFREE(tpp);
2799 		page_clr_all_props(tpp);
2800 		PP_SETAGED(tpp);
2801 		tpp->p_offset = (u_offset_t)-1;
2802 		ASSERT(tpp->p_next == tpp);
2803 		ASSERT(tpp->p_prev == tpp);
2804 		page_list_concat(&rootpp, &tpp);
2805 	}
2806 	ASSERT(rootpp == pp);
2807 
2808 	page_list_add_pages(rootpp, 0);
2809 	page_create_putback(pgcnt);
2810 }
2811 
2812 int free_pages = 1;
2813 
2814 /*
2815  * This routine attempts to return pages to the cachelist via page_release().
2816  * It does not *have* to be successful in all cases, since the pageout scanner
2817  * will catch any pages it misses.  It does need to be fast and not introduce
2818  * too much overhead.
2819  *
2820  * If a page isn't found on the unlocked sweep of the page_hash bucket, we
2821  * don't lock and retry.  This is ok, since the page scanner will eventually
2822  * find any page we miss in free_vp_pages().
2823  */
2824 void
2825 free_vp_pages(vnode_t *vp, u_offset_t off, size_t len)
2826 {
2827 	page_t *pp;
2828 	u_offset_t eoff;
2829 	extern int swap_in_range(vnode_t *, u_offset_t, size_t);
2830 
2831 	eoff = off + len;
2832 
2833 	if (free_pages == 0)
2834 		return;
2835 	if (swap_in_range(vp, off, len))
2836 		return;
2837 
2838 	for (; off < eoff; off += PAGESIZE) {
2839 
2840 		/*
2841 		 * find the page using a fast, but inexact search. It'll be OK
2842 		 * if a few pages slip through the cracks here.
2843 		 */
2844 		pp = page_exists(vp, off);
2845 
2846 		/*
2847 		 * If we didn't find the page (it may not exist), the page
2848 		 * is free, looks still in use (shared), or we can't lock it,
2849 		 * just give up.
2850 		 */
2851 		if (pp == NULL ||
2852 		    PP_ISFREE(pp) ||
2853 		    page_share_cnt(pp) > 0 ||
2854 		    !page_trylock(pp, SE_EXCL))
2855 			continue;
2856 
2857 		/*
2858 		 * Once we have locked pp, verify that it's still the
2859 		 * correct page and not already free
2860 		 */
2861 		ASSERT(PAGE_LOCKED_SE(pp, SE_EXCL));
2862 		if (pp->p_vnode != vp || pp->p_offset != off || PP_ISFREE(pp)) {
2863 			page_unlock(pp);
2864 			continue;
2865 		}
2866 
2867 		/*
2868 		 * try to release the page...
2869 		 */
2870 		(void) page_release(pp, 1);
2871 	}
2872 }
2873 
2874 /*
2875  * Reclaim the given page from the free list.
2876  * If pp is part of a large pages, only the given constituent page is reclaimed
2877  * and the large page it belonged to will be demoted.  This can only happen
2878  * if the page is not on the cachelist.
2879  *
2880  * Returns 1 on success or 0 on failure.
2881  *
2882  * The page is unlocked if it can't be reclaimed (when freemem == 0).
2883  * If `lock' is non-null, it will be dropped and re-acquired if
2884  * the routine must wait while freemem is 0.
2885  *
2886  * As it turns out, boot_getpages() does this.  It picks a page,
2887  * based on where OBP mapped in some address, gets its pfn, searches
2888  * the memsegs, locks the page, then pulls it off the free list!
2889  */
2890 int
2891 page_reclaim(page_t *pp, kmutex_t *lock)
2892 {
2893 	struct pcf	*p;
2894 	struct cpu	*cpup;
2895 	int		enough;
2896 	uint_t		i;
2897 
2898 	ASSERT(lock != NULL ? MUTEX_HELD(lock) : 1);
2899 	ASSERT(PAGE_EXCL(pp) && PP_ISFREE(pp));
2900 
2901 	/*
2902 	 * If `freemem' is 0, we cannot reclaim this page from the
2903 	 * freelist, so release every lock we might hold: the page,
2904 	 * and the `lock' before blocking.
2905 	 *
2906 	 * The only way `freemem' can become 0 while there are pages
2907 	 * marked free (have their p->p_free bit set) is when the
2908 	 * system is low on memory and doing a page_create().  In
2909 	 * order to guarantee that once page_create() starts acquiring
2910 	 * pages it will be able to get all that it needs since `freemem'
2911 	 * was decreased by the requested amount.  So, we need to release
2912 	 * this page, and let page_create() have it.
2913 	 *
2914 	 * Since `freemem' being zero is not supposed to happen, just
2915 	 * use the usual hash stuff as a starting point.  If that bucket
2916 	 * is empty, then assume the worst, and start at the beginning
2917 	 * of the pcf array.  If we always start at the beginning
2918 	 * when acquiring more than one pcf lock, there won't be any
2919 	 * deadlock problems.
2920 	 */
2921 
2922 	/* TODO: Do we need to test kcage_freemem if PG_NORELOC(pp)? */
2923 
2924 	if (freemem <= throttlefree && !page_create_throttle(1l, 0)) {
2925 		pcf_acquire_all();
2926 		goto page_reclaim_nomem;
2927 	}
2928 
2929 	enough = pcf_decrement_bucket(1);
2930 
2931 	if (!enough) {
2932 		VM_STAT_ADD(page_reclaim_zero);
2933 		/*
2934 		 * Check again. Its possible that some other thread
2935 		 * could have been right behind us, and added one
2936 		 * to a list somewhere.  Acquire each of the pcf locks
2937 		 * until we find a page.
2938 		 */
2939 		p = pcf;
2940 		for (i = 0; i < pcf_fanout; i++) {
2941 			mutex_enter(&p->pcf_lock);
2942 			if (p->pcf_count >= 1) {
2943 				p->pcf_count -= 1;
2944 				enough = 1;
2945 				break;
2946 			}
2947 			p++;
2948 		}
2949 
2950 		if (!enough) {
2951 page_reclaim_nomem:
2952 			/*
2953 			 * We really can't have page `pp'.
2954 			 * Time for the no-memory dance with
2955 			 * page_free().  This is just like
2956 			 * page_create_wait().  Plus the added
2957 			 * attraction of releasing whatever mutex
2958 			 * we held when we were called with in `lock'.
2959 			 * Page_unlock() will wakeup any thread
2960 			 * waiting around for this page.
2961 			 */
2962 			if (lock) {
2963 				VM_STAT_ADD(page_reclaim_zero_locked);
2964 				mutex_exit(lock);
2965 			}
2966 			page_unlock(pp);
2967 
2968 			/*
2969 			 * get this before we drop all the pcf locks.
2970 			 */
2971 			mutex_enter(&new_freemem_lock);
2972 
2973 			p = pcf;
2974 			for (i = 0; i < pcf_fanout; i++) {
2975 				p->pcf_wait++;
2976 				mutex_exit(&p->pcf_lock);
2977 				p++;
2978 			}
2979 
2980 			freemem_wait++;
2981 			cv_wait(&freemem_cv, &new_freemem_lock);
2982 			freemem_wait--;
2983 
2984 			mutex_exit(&new_freemem_lock);
2985 
2986 			if (lock) {
2987 				mutex_enter(lock);
2988 			}
2989 			return (0);
2990 		}
2991 
2992 		/*
2993 		 * The pcf accounting has been done,
2994 		 * though none of the pcf_wait flags have been set,
2995 		 * drop the locks and continue on.
2996 		 */
2997 		while (p >= pcf) {
2998 			mutex_exit(&p->pcf_lock);
2999 			p--;
3000 		}
3001 	}
3002 
3003 	/*
3004 	 * freemem is not protected by any lock. Thus, we cannot
3005 	 * have any assertion containing freemem here.
3006 	 */
3007 	freemem -= 1;
3008 
3009 	VM_STAT_ADD(pagecnt.pc_reclaim);
3010 
3011 	/*
3012 	 * page_list_sub will handle the case where pp is a large page.
3013 	 * It's possible that the page was promoted while on the freelist
3014 	 */
3015 	if (PP_ISAGED(pp)) {
3016 		page_list_sub(pp, PG_FREE_LIST);
3017 		TRACE_1(TR_FAC_VM, TR_PAGE_UNFREE_FREE,
3018 		    "page_reclaim_free:pp %p", pp);
3019 	} else {
3020 		page_list_sub(pp, PG_CACHE_LIST);
3021 		TRACE_1(TR_FAC_VM, TR_PAGE_UNFREE_CACHE,
3022 		    "page_reclaim_cache:pp %p", pp);
3023 	}
3024 
3025 	/*
3026 	 * clear the p_free & p_age bits since this page is no longer
3027 	 * on the free list.  Notice that there was a brief time where
3028 	 * a page is marked as free, but is not on the list.
3029 	 *
3030 	 * Set the reference bit to protect against immediate pageout.
3031 	 */
3032 	PP_CLRFREE(pp);
3033 	PP_CLRAGED(pp);
3034 	page_set_props(pp, P_REF);
3035 
3036 	CPU_STATS_ENTER_K();
3037 	cpup = CPU;	/* get cpup now that CPU cannot change */
3038 	CPU_STATS_ADDQ(cpup, vm, pgrec, 1);
3039 	CPU_STATS_ADDQ(cpup, vm, pgfrec, 1);
3040 	CPU_STATS_EXIT_K();
3041 	ASSERT(pp->p_szc == 0);
3042 
3043 	return (1);
3044 }
3045 
3046 /*
3047  * Destroy identity of the page and put it back on
3048  * the page free list.  Assumes that the caller has
3049  * acquired the "exclusive" lock on the page.
3050  */
3051 void
3052 page_destroy(page_t *pp, int dontfree)
3053 {
3054 	ASSERT((PAGE_EXCL(pp) &&
3055 	    !page_iolock_assert(pp)) || panicstr);
3056 	ASSERT(pp->p_slckcnt == 0 || panicstr);
3057 
3058 	if (pp->p_szc != 0) {
3059 		if (pp->p_vnode == NULL || IS_SWAPFSVP(pp->p_vnode) ||
3060 		    PP_ISKAS(pp)) {
3061 			panic("page_destroy: anon or kernel or no vnode "
3062 			    "large page %p", (void *)pp);
3063 		}
3064 		page_demote_vp_pages(pp);
3065 		ASSERT(pp->p_szc == 0);
3066 	}
3067 
3068 	TRACE_1(TR_FAC_VM, TR_PAGE_DESTROY, "page_destroy:pp %p", pp);
3069 
3070 	/*
3071 	 * Unload translations, if any, then hash out the
3072 	 * page to erase its identity.
3073 	 */
3074 	(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
3075 	page_hashout(pp, NULL);
3076 
3077 	if (!dontfree) {
3078 		/*
3079 		 * Acquire the "freemem_lock" for availrmem.
3080 		 * The page_struct_lock need not be acquired for lckcnt
3081 		 * and cowcnt since the page has an "exclusive" lock.
3082 		 */
3083 		if ((pp->p_lckcnt != 0) || (pp->p_cowcnt != 0)) {
3084 			mutex_enter(&freemem_lock);
3085 			if (pp->p_lckcnt != 0) {
3086 				availrmem++;
3087 				pp->p_lckcnt = 0;
3088 			}
3089 			if (pp->p_cowcnt != 0) {
3090 				availrmem += pp->p_cowcnt;
3091 				pp->p_cowcnt = 0;
3092 			}
3093 			mutex_exit(&freemem_lock);
3094 		}
3095 		/*
3096 		 * Put the page on the "free" list.
3097 		 */
3098 		page_free(pp, 0);
3099 	}
3100 }
3101 
3102 void
3103 page_destroy_pages(page_t *pp)
3104 {
3105 
3106 	page_t	*tpp, *rootpp = NULL;
3107 	pgcnt_t	pgcnt = page_get_pagecnt(pp->p_szc);
3108 	pgcnt_t	i, pglcks = 0;
3109 	uint_t	szc = pp->p_szc;
3110 
3111 	ASSERT(pp->p_szc != 0 && pp->p_szc < page_num_pagesizes());
3112 
3113 	VM_STAT_ADD(pagecnt.pc_destroy_pages);
3114 
3115 	TRACE_1(TR_FAC_VM, TR_PAGE_DESTROY, "page_destroy_pages:pp %p", pp);
3116 
3117 	if ((page_pptonum(pp) & (pgcnt - 1)) != 0) {
3118 		panic("page_destroy_pages: not root page %p", (void *)pp);
3119 		/*NOTREACHED*/
3120 	}
3121 
3122 	for (i = 0, tpp = pp; i < pgcnt; i++, tpp++) {
3123 		ASSERT((PAGE_EXCL(tpp) &&
3124 		    !page_iolock_assert(tpp)) || panicstr);
3125 		ASSERT(tpp->p_slckcnt == 0 || panicstr);
3126 		(void) hat_pageunload(tpp, HAT_FORCE_PGUNLOAD);
3127 		page_hashout(tpp, NULL);
3128 		ASSERT(tpp->p_offset == (u_offset_t)-1);
3129 		if (tpp->p_lckcnt != 0) {
3130 			pglcks++;
3131 			tpp->p_lckcnt = 0;
3132 		} else if (tpp->p_cowcnt != 0) {
3133 			pglcks += tpp->p_cowcnt;
3134 			tpp->p_cowcnt = 0;
3135 		}
3136 		ASSERT(!hat_page_getshare(tpp));
3137 		ASSERT(tpp->p_vnode == NULL);
3138 		ASSERT(tpp->p_szc == szc);
3139 
3140 		PP_SETFREE(tpp);
3141 		page_clr_all_props(tpp);
3142 		PP_SETAGED(tpp);
3143 		ASSERT(tpp->p_next == tpp);
3144 		ASSERT(tpp->p_prev == tpp);
3145 		page_list_concat(&rootpp, &tpp);
3146 	}
3147 
3148 	ASSERT(rootpp == pp);
3149 	if (pglcks != 0) {
3150 		mutex_enter(&freemem_lock);
3151 		availrmem += pglcks;
3152 		mutex_exit(&freemem_lock);
3153 	}
3154 
3155 	page_list_add_pages(rootpp, 0);
3156 	page_create_putback(pgcnt);
3157 }
3158 
3159 /*
3160  * Similar to page_destroy(), but destroys pages which are
3161  * locked and known to be on the page free list.  Since
3162  * the page is known to be free and locked, no one can access
3163  * it.
3164  *
3165  * Also, the number of free pages does not change.
3166  */
3167 void
3168 page_destroy_free(page_t *pp)
3169 {
3170 	ASSERT(PAGE_EXCL(pp));
3171 	ASSERT(PP_ISFREE(pp));
3172 	ASSERT(pp->p_vnode);
3173 	ASSERT(hat_page_getattr(pp, P_MOD | P_REF | P_RO) == 0);
3174 	ASSERT(!hat_page_is_mapped(pp));
3175 	ASSERT(PP_ISAGED(pp) == 0);
3176 	ASSERT(pp->p_szc == 0);
3177 
3178 	VM_STAT_ADD(pagecnt.pc_destroy_free);
3179 	page_list_sub(pp, PG_CACHE_LIST);
3180 
3181 	page_hashout(pp, NULL);
3182 	ASSERT(pp->p_vnode == NULL);
3183 	ASSERT(pp->p_offset == (u_offset_t)-1);
3184 	ASSERT(pp->p_hash == NULL);
3185 
3186 	PP_SETAGED(pp);
3187 	page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL);
3188 	page_unlock(pp);
3189 
3190 	mutex_enter(&new_freemem_lock);
3191 	if (freemem_wait) {
3192 		cv_signal(&freemem_cv);
3193 	}
3194 	mutex_exit(&new_freemem_lock);
3195 }
3196 
3197 /*
3198  * Rename the page "opp" to have an identity specified
3199  * by [vp, off].  If a page already exists with this name
3200  * it is locked and destroyed.  Note that the page's
3201  * translations are not unloaded during the rename.
3202  *
3203  * This routine is used by the anon layer to "steal" the
3204  * original page and is not unlike destroying a page and
3205  * creating a new page using the same page frame.
3206  *
3207  * XXX -- Could deadlock if caller 1 tries to rename A to B while
3208  * caller 2 tries to rename B to A.
3209  */
3210 void
3211 page_rename(page_t *opp, vnode_t *vp, u_offset_t off)
3212 {
3213 	page_t		*pp;
3214 	int		olckcnt = 0;
3215 	int		ocowcnt = 0;
3216 	kmutex_t	*phm;
3217 	ulong_t		index;
3218 
3219 	ASSERT(PAGE_EXCL(opp) && !page_iolock_assert(opp));
3220 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
3221 	ASSERT(PP_ISFREE(opp) == 0);
3222 
3223 	VM_STAT_ADD(page_rename_count);
3224 
3225 	TRACE_3(TR_FAC_VM, TR_PAGE_RENAME,
3226 	    "page rename:pp %p vp %p off %llx", opp, vp, off);
3227 
3228 	/*
3229 	 * CacheFS may call page_rename for a large NFS page
3230 	 * when both CacheFS and NFS mount points are used
3231 	 * by applications. Demote this large page before
3232 	 * renaming it, to ensure that there are no "partial"
3233 	 * large pages left lying around.
3234 	 */
3235 	if (opp->p_szc != 0) {
3236 		vnode_t *ovp = opp->p_vnode;
3237 		ASSERT(ovp != NULL);
3238 		ASSERT(!IS_SWAPFSVP(ovp));
3239 		ASSERT(!VN_ISKAS(ovp));
3240 		page_demote_vp_pages(opp);
3241 		ASSERT(opp->p_szc == 0);
3242 	}
3243 
3244 	page_hashout(opp, NULL);
3245 	PP_CLRAGED(opp);
3246 
3247 	/*
3248 	 * Acquire the appropriate page hash lock, since
3249 	 * we're going to rename the page.
3250 	 */
3251 	index = PAGE_HASH_FUNC(vp, off);
3252 	phm = PAGE_HASH_MUTEX(index);
3253 	mutex_enter(phm);
3254 top:
3255 	/*
3256 	 * Look for an existing page with this name and destroy it if found.
3257 	 * By holding the page hash lock all the way to the page_hashin()
3258 	 * call, we are assured that no page can be created with this
3259 	 * identity.  In the case when the phm lock is dropped to undo any
3260 	 * hat layer mappings, the existing page is held with an "exclusive"
3261 	 * lock, again preventing another page from being created with
3262 	 * this identity.
3263 	 */
3264 	PAGE_HASH_SEARCH(index, pp, vp, off);
3265 	if (pp != NULL) {
3266 		VM_STAT_ADD(page_rename_exists);
3267 
3268 		/*
3269 		 * As it turns out, this is one of only two places where
3270 		 * page_lock() needs to hold the passed in lock in the
3271 		 * successful case.  In all of the others, the lock could
3272 		 * be dropped as soon as the attempt is made to lock
3273 		 * the page.  It is tempting to add yet another arguement,
3274 		 * PL_KEEP or PL_DROP, to let page_lock know what to do.
3275 		 */
3276 		if (!page_lock(pp, SE_EXCL, phm, P_RECLAIM)) {
3277 			/*
3278 			 * Went to sleep because the page could not
3279 			 * be locked.  We were woken up when the page
3280 			 * was unlocked, or when the page was destroyed.
3281 			 * In either case, `phm' was dropped while we
3282 			 * slept.  Hence we should not just roar through
3283 			 * this loop.
3284 			 */
3285 			goto top;
3286 		}
3287 
3288 		/*
3289 		 * If an existing page is a large page, then demote
3290 		 * it to ensure that no "partial" large pages are
3291 		 * "created" after page_rename. An existing page
3292 		 * can be a CacheFS page, and can't belong to swapfs.
3293 		 */
3294 		if (hat_page_is_mapped(pp)) {
3295 			/*
3296 			 * Unload translations.  Since we hold the
3297 			 * exclusive lock on this page, the page
3298 			 * can not be changed while we drop phm.
3299 			 * This is also not a lock protocol violation,
3300 			 * but rather the proper way to do things.
3301 			 */
3302 			mutex_exit(phm);
3303 			(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
3304 			if (pp->p_szc != 0) {
3305 				ASSERT(!IS_SWAPFSVP(vp));
3306 				ASSERT(!VN_ISKAS(vp));
3307 				page_demote_vp_pages(pp);
3308 				ASSERT(pp->p_szc == 0);
3309 			}
3310 			mutex_enter(phm);
3311 		} else if (pp->p_szc != 0) {
3312 			ASSERT(!IS_SWAPFSVP(vp));
3313 			ASSERT(!VN_ISKAS(vp));
3314 			mutex_exit(phm);
3315 			page_demote_vp_pages(pp);
3316 			ASSERT(pp->p_szc == 0);
3317 			mutex_enter(phm);
3318 		}
3319 		page_hashout(pp, phm);
3320 	}
3321 	/*
3322 	 * Hash in the page with the new identity.
3323 	 */
3324 	if (!page_hashin(opp, vp, off, phm)) {
3325 		/*
3326 		 * We were holding phm while we searched for [vp, off]
3327 		 * and only dropped phm if we found and locked a page.
3328 		 * If we can't create this page now, then some thing
3329 		 * is really broken.
3330 		 */
3331 		panic("page_rename: Can't hash in page: %p", (void *)pp);
3332 		/*NOTREACHED*/
3333 	}
3334 
3335 	ASSERT(MUTEX_HELD(phm));
3336 	mutex_exit(phm);
3337 
3338 	/*
3339 	 * Now that we have dropped phm, lets get around to finishing up
3340 	 * with pp.
3341 	 */
3342 	if (pp != NULL) {
3343 		ASSERT(!hat_page_is_mapped(pp));
3344 		/* for now large pages should not end up here */
3345 		ASSERT(pp->p_szc == 0);
3346 		/*
3347 		 * Save the locks for transfer to the new page and then
3348 		 * clear them so page_free doesn't think they're important.
3349 		 * The page_struct_lock need not be acquired for lckcnt and
3350 		 * cowcnt since the page has an "exclusive" lock.
3351 		 */
3352 		olckcnt = pp->p_lckcnt;
3353 		ocowcnt = pp->p_cowcnt;
3354 		pp->p_lckcnt = pp->p_cowcnt = 0;
3355 
3356 		/*
3357 		 * Put the page on the "free" list after we drop
3358 		 * the lock.  The less work under the lock the better.
3359 		 */
3360 		/*LINTED: constant in conditional context*/
3361 		VN_DISPOSE(pp, B_FREE, 0, kcred);
3362 	}
3363 
3364 	/*
3365 	 * Transfer the lock count from the old page (if any).
3366 	 * The page_struct_lock need not be acquired for lckcnt and
3367 	 * cowcnt since the page has an "exclusive" lock.
3368 	 */
3369 	opp->p_lckcnt += olckcnt;
3370 	opp->p_cowcnt += ocowcnt;
3371 }
3372 
3373 /*
3374  * low level routine to add page `pp' to the hash and vp chains for [vp, offset]
3375  *
3376  * Pages are normally inserted at the start of a vnode's v_pages list.
3377  * If the vnode is VMODSORT and the page is modified, it goes at the end.
3378  * This can happen when a modified page is relocated for DR.
3379  *
3380  * Returns 1 on success and 0 on failure.
3381  */
3382 static int
3383 page_do_hashin(page_t *pp, vnode_t *vp, u_offset_t offset)
3384 {
3385 	page_t		**listp;
3386 	page_t		*tp;
3387 	ulong_t		index;
3388 
3389 	ASSERT(PAGE_EXCL(pp));
3390 	ASSERT(vp != NULL);
3391 	ASSERT(MUTEX_HELD(page_vnode_mutex(vp)));
3392 
3393 	/*
3394 	 * Be sure to set these up before the page is inserted on the hash
3395 	 * list.  As soon as the page is placed on the list some other
3396 	 * thread might get confused and wonder how this page could
3397 	 * possibly hash to this list.
3398 	 */
3399 	pp->p_vnode = vp;
3400 	pp->p_offset = offset;
3401 
3402 	/*
3403 	 * record if this page is on a swap vnode
3404 	 */
3405 	if ((vp->v_flag & VISSWAP) != 0)
3406 		PP_SETSWAP(pp);
3407 
3408 	index = PAGE_HASH_FUNC(vp, offset);
3409 	ASSERT(MUTEX_HELD(PAGE_HASH_MUTEX(index)));
3410 	listp = &page_hash[index];
3411 
3412 	/*
3413 	 * If this page is already hashed in, fail this attempt to add it.
3414 	 */
3415 	for (tp = *listp; tp != NULL; tp = tp->p_hash) {
3416 		if (tp->p_vnode == vp && tp->p_offset == offset) {
3417 			pp->p_vnode = NULL;
3418 			pp->p_offset = (u_offset_t)(-1);
3419 			return (0);
3420 		}
3421 	}
3422 	pp->p_hash = *listp;
3423 	*listp = pp;
3424 
3425 	/*
3426 	 * Add the page to the vnode's list of pages
3427 	 */
3428 	if (vp->v_pages != NULL && IS_VMODSORT(vp) && hat_ismod(pp))
3429 		listp = &vp->v_pages->p_vpprev->p_vpnext;
3430 	else
3431 		listp = &vp->v_pages;
3432 
3433 	page_vpadd(listp, pp);
3434 
3435 	return (1);
3436 }
3437 
3438 /*
3439  * Add page `pp' to both the hash and vp chains for [vp, offset].
3440  *
3441  * Returns 1 on success and 0 on failure.
3442  * If hold is passed in, it is not dropped.
3443  */
3444 int
3445 page_hashin(page_t *pp, vnode_t *vp, u_offset_t offset, kmutex_t *hold)
3446 {
3447 	kmutex_t	*phm = NULL;
3448 	kmutex_t	*vphm;
3449 	int		rc;
3450 
3451 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
3452 
3453 	TRACE_3(TR_FAC_VM, TR_PAGE_HASHIN,
3454 	    "page_hashin:pp %p vp %p offset %llx",
3455 	    pp, vp, offset);
3456 
3457 	VM_STAT_ADD(hashin_count);
3458 
3459 	if (hold != NULL)
3460 		phm = hold;
3461 	else {
3462 		VM_STAT_ADD(hashin_not_held);
3463 		phm = PAGE_HASH_MUTEX(PAGE_HASH_FUNC(vp, offset));
3464 		mutex_enter(phm);
3465 	}
3466 
3467 	vphm = page_vnode_mutex(vp);
3468 	mutex_enter(vphm);
3469 	rc = page_do_hashin(pp, vp, offset);
3470 	mutex_exit(vphm);
3471 	if (hold == NULL)
3472 		mutex_exit(phm);
3473 	if (rc == 0)
3474 		VM_STAT_ADD(hashin_already);
3475 	return (rc);
3476 }
3477 
3478 /*
3479  * Remove page ``pp'' from the hash and vp chains and remove vp association.
3480  * All mutexes must be held
3481  */
3482 static void
3483 page_do_hashout(page_t *pp)
3484 {
3485 	page_t	**hpp;
3486 	page_t	*hp;
3487 	vnode_t	*vp = pp->p_vnode;
3488 
3489 	ASSERT(vp != NULL);
3490 	ASSERT(MUTEX_HELD(page_vnode_mutex(vp)));
3491 
3492 	/*
3493 	 * First, take pp off of its hash chain.
3494 	 */
3495 	hpp = &page_hash[PAGE_HASH_FUNC(vp, pp->p_offset)];
3496 
3497 	for (;;) {
3498 		hp = *hpp;
3499 		if (hp == pp)
3500 			break;
3501 		if (hp == NULL) {
3502 			panic("page_do_hashout");
3503 			/*NOTREACHED*/
3504 		}
3505 		hpp = &hp->p_hash;
3506 	}
3507 	*hpp = pp->p_hash;
3508 
3509 	/*
3510 	 * Now remove it from its associated vnode.
3511 	 */
3512 	if (vp->v_pages)
3513 		page_vpsub(&vp->v_pages, pp);
3514 
3515 	pp->p_hash = NULL;
3516 	page_clr_all_props(pp);
3517 	PP_CLRSWAP(pp);
3518 	pp->p_vnode = NULL;
3519 	pp->p_offset = (u_offset_t)-1;
3520 }
3521 
3522 /*
3523  * Remove page ``pp'' from the hash and vp chains and remove vp association.
3524  *
3525  * When `phm' is non-NULL it contains the address of the mutex protecting the
3526  * hash list pp is on.  It is not dropped.
3527  */
3528 void
3529 page_hashout(page_t *pp, kmutex_t *phm)
3530 {
3531 	vnode_t		*vp;
3532 	ulong_t		index;
3533 	kmutex_t	*nphm;
3534 	kmutex_t	*vphm;
3535 	kmutex_t	*sep;
3536 
3537 	ASSERT(phm != NULL ? MUTEX_HELD(phm) : 1);
3538 	ASSERT(pp->p_vnode != NULL);
3539 	ASSERT((PAGE_EXCL(pp) && !page_iolock_assert(pp)) || panicstr);
3540 	ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(pp->p_vnode)));
3541 
3542 	vp = pp->p_vnode;
3543 
3544 	TRACE_2(TR_FAC_VM, TR_PAGE_HASHOUT,
3545 	    "page_hashout:pp %p vp %p", pp, vp);
3546 
3547 	/* Kernel probe */
3548 	TNF_PROBE_2(page_unmap, "vm pagefault", /* CSTYLED */,
3549 	    tnf_opaque, vnode, vp,
3550 	    tnf_offset, offset, pp->p_offset);
3551 
3552 	/*
3553 	 *
3554 	 */
3555 	VM_STAT_ADD(hashout_count);
3556 	index = PAGE_HASH_FUNC(vp, pp->p_offset);
3557 	if (phm == NULL) {
3558 		VM_STAT_ADD(hashout_not_held);
3559 		nphm = PAGE_HASH_MUTEX(index);
3560 		mutex_enter(nphm);
3561 	}
3562 	ASSERT(phm ? phm == PAGE_HASH_MUTEX(index) : 1);
3563 
3564 
3565 	/*
3566 	 * grab page vnode mutex and remove it...
3567 	 */
3568 	vphm = page_vnode_mutex(vp);
3569 	mutex_enter(vphm);
3570 
3571 	page_do_hashout(pp);
3572 
3573 	mutex_exit(vphm);
3574 	if (phm == NULL)
3575 		mutex_exit(nphm);
3576 
3577 	/*
3578 	 * Wake up processes waiting for this page.  The page's
3579 	 * identity has been changed, and is probably not the
3580 	 * desired page any longer.
3581 	 */
3582 	sep = page_se_mutex(pp);
3583 	mutex_enter(sep);
3584 	pp->p_selock &= ~SE_EWANTED;
3585 	if (CV_HAS_WAITERS(&pp->p_cv))
3586 		cv_broadcast(&pp->p_cv);
3587 	mutex_exit(sep);
3588 }
3589 
3590 /*
3591  * Add the page to the front of a linked list of pages
3592  * using the p_next & p_prev pointers for the list.
3593  * The caller is responsible for protecting the list pointers.
3594  */
3595 void
3596 page_add(page_t **ppp, page_t *pp)
3597 {
3598 	ASSERT(PAGE_EXCL(pp) || (PAGE_SHARED(pp) && page_iolock_assert(pp)));
3599 
3600 	page_add_common(ppp, pp);
3601 }
3602 
3603 
3604 
3605 /*
3606  *  Common code for page_add() and mach_page_add()
3607  */
3608 void
3609 page_add_common(page_t **ppp, page_t *pp)
3610 {
3611 	if (*ppp == NULL) {
3612 		pp->p_next = pp->p_prev = pp;
3613 	} else {
3614 		pp->p_next = *ppp;
3615 		pp->p_prev = (*ppp)->p_prev;
3616 		(*ppp)->p_prev = pp;
3617 		pp->p_prev->p_next = pp;
3618 	}
3619 	*ppp = pp;
3620 }
3621 
3622 
3623 /*
3624  * Remove this page from a linked list of pages
3625  * using the p_next & p_prev pointers for the list.
3626  *
3627  * The caller is responsible for protecting the list pointers.
3628  */
3629 void
3630 page_sub(page_t **ppp, page_t *pp)
3631 {
3632 	ASSERT((PP_ISFREE(pp)) ? 1 :
3633 	    (PAGE_EXCL(pp)) || (PAGE_SHARED(pp) && page_iolock_assert(pp)));
3634 
3635 	if (*ppp == NULL || pp == NULL) {
3636 		panic("page_sub: bad arg(s): pp %p, *ppp %p",
3637 		    (void *)pp, (void *)(*ppp));
3638 		/*NOTREACHED*/
3639 	}
3640 
3641 	page_sub_common(ppp, pp);
3642 }
3643 
3644 
3645 /*
3646  *  Common code for page_sub() and mach_page_sub()
3647  */
3648 void
3649 page_sub_common(page_t **ppp, page_t *pp)
3650 {
3651 	if (*ppp == pp)
3652 		*ppp = pp->p_next;		/* go to next page */
3653 
3654 	if (*ppp == pp)
3655 		*ppp = NULL;			/* page list is gone */
3656 	else {
3657 		pp->p_prev->p_next = pp->p_next;
3658 		pp->p_next->p_prev = pp->p_prev;
3659 	}
3660 	pp->p_prev = pp->p_next = pp;		/* make pp a list of one */
3661 }
3662 
3663 
3664 /*
3665  * Break page list cppp into two lists with npages in the first list.
3666  * The tail is returned in nppp.
3667  */
3668 void
3669 page_list_break(page_t **oppp, page_t **nppp, pgcnt_t npages)
3670 {
3671 	page_t *s1pp = *oppp;
3672 	page_t *s2pp;
3673 	page_t *e1pp, *e2pp;
3674 	long n = 0;
3675 
3676 	if (s1pp == NULL) {
3677 		*nppp = NULL;
3678 		return;
3679 	}
3680 	if (npages == 0) {
3681 		*nppp = s1pp;
3682 		*oppp = NULL;
3683 		return;
3684 	}
3685 	for (n = 0, s2pp = *oppp; n < npages; n++) {
3686 		s2pp = s2pp->p_next;
3687 	}
3688 	/* Fix head and tail of new lists */
3689 	e1pp = s2pp->p_prev;
3690 	e2pp = s1pp->p_prev;
3691 	s1pp->p_prev = e1pp;
3692 	e1pp->p_next = s1pp;
3693 	s2pp->p_prev = e2pp;
3694 	e2pp->p_next = s2pp;
3695 
3696 	/* second list empty */
3697 	if (s2pp == s1pp) {
3698 		*oppp = s1pp;
3699 		*nppp = NULL;
3700 	} else {
3701 		*oppp = s1pp;
3702 		*nppp = s2pp;
3703 	}
3704 }
3705 
3706 /*
3707  * Concatenate page list nppp onto the end of list ppp.
3708  */
3709 void
3710 page_list_concat(page_t **ppp, page_t **nppp)
3711 {
3712 	page_t *s1pp, *s2pp, *e1pp, *e2pp;
3713 
3714 	if (*nppp == NULL) {
3715 		return;
3716 	}
3717 	if (*ppp == NULL) {
3718 		*ppp = *nppp;
3719 		return;
3720 	}
3721 	s1pp = *ppp;
3722 	e1pp =  s1pp->p_prev;
3723 	s2pp = *nppp;
3724 	e2pp = s2pp->p_prev;
3725 	s1pp->p_prev = e2pp;
3726 	e2pp->p_next = s1pp;
3727 	e1pp->p_next = s2pp;
3728 	s2pp->p_prev = e1pp;
3729 }
3730 
3731 /*
3732  * return the next page in the page list
3733  */
3734 page_t *
3735 page_list_next(page_t *pp)
3736 {
3737 	return (pp->p_next);
3738 }
3739 
3740 
3741 /*
3742  * Add the page to the front of the linked list of pages
3743  * using p_vpnext/p_vpprev pointers for the list.
3744  *
3745  * The caller is responsible for protecting the lists.
3746  */
3747 void
3748 page_vpadd(page_t **ppp, page_t *pp)
3749 {
3750 	if (*ppp == NULL) {
3751 		pp->p_vpnext = pp->p_vpprev = pp;
3752 	} else {
3753 		pp->p_vpnext = *ppp;
3754 		pp->p_vpprev = (*ppp)->p_vpprev;
3755 		(*ppp)->p_vpprev = pp;
3756 		pp->p_vpprev->p_vpnext = pp;
3757 	}
3758 	*ppp = pp;
3759 }
3760 
3761 /*
3762  * Remove this page from the linked list of pages
3763  * using p_vpnext/p_vpprev pointers for the list.
3764  *
3765  * The caller is responsible for protecting the lists.
3766  */
3767 void
3768 page_vpsub(page_t **ppp, page_t *pp)
3769 {
3770 	if (*ppp == NULL || pp == NULL) {
3771 		panic("page_vpsub: bad arg(s): pp %p, *ppp %p",
3772 		    (void *)pp, (void *)(*ppp));
3773 		/*NOTREACHED*/
3774 	}
3775 
3776 	if (*ppp == pp)
3777 		*ppp = pp->p_vpnext;		/* go to next page */
3778 
3779 	if (*ppp == pp)
3780 		*ppp = NULL;			/* page list is gone */
3781 	else {
3782 		pp->p_vpprev->p_vpnext = pp->p_vpnext;
3783 		pp->p_vpnext->p_vpprev = pp->p_vpprev;
3784 	}
3785 	pp->p_vpprev = pp->p_vpnext = pp;	/* make pp a list of one */
3786 }
3787 
3788 /*
3789  * Lock a physical page into memory "long term".  Used to support "lock
3790  * in memory" functions.  Accepts the page to be locked, and a cow variable
3791  * to indicate whether a the lock will travel to the new page during
3792  * a potential copy-on-write.
3793  */
3794 int
3795 page_pp_lock(
3796 	page_t *pp,			/* page to be locked */
3797 	int cow,			/* cow lock */
3798 	int kernel)			/* must succeed -- ignore checking */
3799 {
3800 	int r = 0;			/* result -- assume failure */
3801 
3802 	ASSERT(PAGE_LOCKED(pp));
3803 
3804 	page_struct_lock(pp);
3805 	/*
3806 	 * Acquire the "freemem_lock" for availrmem.
3807 	 */
3808 	if (cow) {
3809 		mutex_enter(&freemem_lock);
3810 		if ((availrmem > pages_pp_maximum) &&
3811 		    (pp->p_cowcnt < (ushort_t)PAGE_LOCK_MAXIMUM)) {
3812 			availrmem--;
3813 			pages_locked++;
3814 			mutex_exit(&freemem_lock);
3815 			r = 1;
3816 			if (++pp->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
3817 				cmn_err(CE_WARN,
3818 				    "COW lock limit reached on pfn 0x%lx",
3819 				    page_pptonum(pp));
3820 			}
3821 		} else
3822 			mutex_exit(&freemem_lock);
3823 	} else {
3824 		if (pp->p_lckcnt) {
3825 			if (pp->p_lckcnt < (ushort_t)PAGE_LOCK_MAXIMUM) {
3826 				r = 1;
3827 				if (++pp->p_lckcnt ==
3828 				    (ushort_t)PAGE_LOCK_MAXIMUM) {
3829 					cmn_err(CE_WARN, "Page lock limit "
3830 					    "reached on pfn 0x%lx",
3831 					    page_pptonum(pp));
3832 				}
3833 			}
3834 		} else {
3835 			if (kernel) {
3836 				/* availrmem accounting done by caller */
3837 				++pp->p_lckcnt;
3838 				r = 1;
3839 			} else {
3840 				mutex_enter(&freemem_lock);
3841 				if (availrmem > pages_pp_maximum) {
3842 					availrmem--;
3843 					pages_locked++;
3844 					++pp->p_lckcnt;
3845 					r = 1;
3846 				}
3847 				mutex_exit(&freemem_lock);
3848 			}
3849 		}
3850 	}
3851 	page_struct_unlock(pp);
3852 	return (r);
3853 }
3854 
3855 /*
3856  * Decommit a lock on a physical page frame.  Account for cow locks if
3857  * appropriate.
3858  */
3859 void
3860 page_pp_unlock(
3861 	page_t *pp,			/* page to be unlocked */
3862 	int cow,			/* expect cow lock */
3863 	int kernel)			/* this was a kernel lock */
3864 {
3865 	ASSERT(PAGE_LOCKED(pp));
3866 
3867 	page_struct_lock(pp);
3868 	/*
3869 	 * Acquire the "freemem_lock" for availrmem.
3870 	 * If cowcnt or lcknt is already 0 do nothing; i.e., we
3871 	 * could be called to unlock even if nothing is locked. This could
3872 	 * happen if locked file pages were truncated (removing the lock)
3873 	 * and the file was grown again and new pages faulted in; the new
3874 	 * pages are unlocked but the segment still thinks they're locked.
3875 	 */
3876 	if (cow) {
3877 		if (pp->p_cowcnt) {
3878 			mutex_enter(&freemem_lock);
3879 			pp->p_cowcnt--;
3880 			availrmem++;
3881 			pages_locked--;
3882 			mutex_exit(&freemem_lock);
3883 		}
3884 	} else {
3885 		if (pp->p_lckcnt && --pp->p_lckcnt == 0) {
3886 			if (!kernel) {
3887 				mutex_enter(&freemem_lock);
3888 				availrmem++;
3889 				pages_locked--;
3890 				mutex_exit(&freemem_lock);
3891 			}
3892 		}
3893 	}
3894 	page_struct_unlock(pp);
3895 }
3896 
3897 /*
3898  * This routine reserves availrmem for npages;
3899  * 	flags: KM_NOSLEEP or KM_SLEEP
3900  * 	returns 1 on success or 0 on failure
3901  */
3902 int
3903 page_resv(pgcnt_t npages, uint_t flags)
3904 {
3905 	mutex_enter(&freemem_lock);
3906 	while (availrmem < tune.t_minarmem + npages) {
3907 		if (flags & KM_NOSLEEP) {
3908 			mutex_exit(&freemem_lock);
3909 			return (0);
3910 		}
3911 		mutex_exit(&freemem_lock);
3912 		page_needfree(npages);
3913 		kmem_reap();
3914 		delay(hz >> 2);
3915 		page_needfree(-(spgcnt_t)npages);
3916 		mutex_enter(&freemem_lock);
3917 	}
3918 	availrmem -= npages;
3919 	mutex_exit(&freemem_lock);
3920 	return (1);
3921 }
3922 
3923 /*
3924  * This routine unreserves availrmem for npages;
3925  */
3926 void
3927 page_unresv(pgcnt_t npages)
3928 {
3929 	mutex_enter(&freemem_lock);
3930 	availrmem += npages;
3931 	mutex_exit(&freemem_lock);
3932 }
3933 
3934 /*
3935  * See Statement at the beginning of segvn_lockop() regarding
3936  * the way we handle cowcnts and lckcnts.
3937  *
3938  * Transfer cowcnt on 'opp' to cowcnt on 'npp' if the vpage
3939  * that breaks COW has PROT_WRITE.
3940  *
3941  * Note that, we may also break COW in case we are softlocking
3942  * on read access during physio;
3943  * in this softlock case, the vpage may not have PROT_WRITE.
3944  * So, we need to transfer lckcnt on 'opp' to lckcnt on 'npp'
3945  * if the vpage doesn't have PROT_WRITE.
3946  *
3947  * This routine is never called if we are stealing a page
3948  * in anon_private.
3949  *
3950  * The caller subtracted from availrmem for read only mapping.
3951  * if lckcnt is 1 increment availrmem.
3952  */
3953 void
3954 page_pp_useclaim(
3955 	page_t *opp,		/* original page frame losing lock */
3956 	page_t *npp,		/* new page frame gaining lock */
3957 	uint_t	write_perm) 	/* set if vpage has PROT_WRITE */
3958 {
3959 	int payback = 0;
3960 
3961 	ASSERT(PAGE_LOCKED(opp));
3962 	ASSERT(PAGE_LOCKED(npp));
3963 
3964 	page_struct_lock(opp);
3965 
3966 	ASSERT(npp->p_cowcnt == 0);
3967 	ASSERT(npp->p_lckcnt == 0);
3968 
3969 	/* Don't use claim if nothing is locked (see page_pp_unlock above) */
3970 	if ((write_perm && opp->p_cowcnt != 0) ||
3971 	    (!write_perm && opp->p_lckcnt != 0)) {
3972 
3973 		if (write_perm) {
3974 			npp->p_cowcnt++;
3975 			ASSERT(opp->p_cowcnt != 0);
3976 			opp->p_cowcnt--;
3977 		} else {
3978 
3979 			ASSERT(opp->p_lckcnt != 0);
3980 
3981 			/*
3982 			 * We didn't need availrmem decremented if p_lckcnt on
3983 			 * original page is 1. Here, we are unlocking
3984 			 * read-only copy belonging to original page and
3985 			 * are locking a copy belonging to new page.
3986 			 */
3987 			if (opp->p_lckcnt == 1)
3988 				payback = 1;
3989 
3990 			npp->p_lckcnt++;
3991 			opp->p_lckcnt--;
3992 		}
3993 	}
3994 	if (payback) {
3995 		mutex_enter(&freemem_lock);
3996 		availrmem++;
3997 		pages_useclaim--;
3998 		mutex_exit(&freemem_lock);
3999 	}
4000 	page_struct_unlock(opp);
4001 }
4002 
4003 /*
4004  * Simple claim adjust functions -- used to support changes in
4005  * claims due to changes in access permissions.  Used by segvn_setprot().
4006  */
4007 int
4008 page_addclaim(page_t *pp)
4009 {
4010 	int r = 0;			/* result */
4011 
4012 	ASSERT(PAGE_LOCKED(pp));
4013 
4014 	page_struct_lock(pp);
4015 	ASSERT(pp->p_lckcnt != 0);
4016 
4017 	if (pp->p_lckcnt == 1) {
4018 		if (pp->p_cowcnt < (ushort_t)PAGE_LOCK_MAXIMUM) {
4019 			--pp->p_lckcnt;
4020 			r = 1;
4021 			if (++pp->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
4022 				cmn_err(CE_WARN,
4023 				    "COW lock limit reached on pfn 0x%lx",
4024 				    page_pptonum(pp));
4025 			}
4026 		}
4027 	} else {
4028 		mutex_enter(&freemem_lock);
4029 		if ((availrmem > pages_pp_maximum) &&
4030 		    (pp->p_cowcnt < (ushort_t)PAGE_LOCK_MAXIMUM)) {
4031 			--availrmem;
4032 			++pages_claimed;
4033 			mutex_exit(&freemem_lock);
4034 			--pp->p_lckcnt;
4035 			r = 1;
4036 			if (++pp->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
4037 				cmn_err(CE_WARN,
4038 				    "COW lock limit reached on pfn 0x%lx",
4039 				    page_pptonum(pp));
4040 			}
4041 		} else
4042 			mutex_exit(&freemem_lock);
4043 	}
4044 	page_struct_unlock(pp);
4045 	return (r);
4046 }
4047 
4048 int
4049 page_subclaim(page_t *pp)
4050 {
4051 	int r = 0;
4052 
4053 	ASSERT(PAGE_LOCKED(pp));
4054 
4055 	page_struct_lock(pp);
4056 	ASSERT(pp->p_cowcnt != 0);
4057 
4058 	if (pp->p_lckcnt) {
4059 		if (pp->p_lckcnt < (ushort_t)PAGE_LOCK_MAXIMUM) {
4060 			r = 1;
4061 			/*
4062 			 * for availrmem
4063 			 */
4064 			mutex_enter(&freemem_lock);
4065 			availrmem++;
4066 			pages_claimed--;
4067 			mutex_exit(&freemem_lock);
4068 
4069 			pp->p_cowcnt--;
4070 
4071 			if (++pp->p_lckcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
4072 				cmn_err(CE_WARN,
4073 				    "Page lock limit reached on pfn 0x%lx",
4074 				    page_pptonum(pp));
4075 			}
4076 		}
4077 	} else {
4078 		r = 1;
4079 		pp->p_cowcnt--;
4080 		pp->p_lckcnt++;
4081 	}
4082 	page_struct_unlock(pp);
4083 	return (r);
4084 }
4085 
4086 int
4087 page_addclaim_pages(page_t  **ppa)
4088 {
4089 
4090 	pgcnt_t	lckpgs = 0, pg_idx;
4091 
4092 	VM_STAT_ADD(pagecnt.pc_addclaim_pages);
4093 
4094 	mutex_enter(&page_llock);
4095 	for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) {
4096 
4097 		ASSERT(PAGE_LOCKED(ppa[pg_idx]));
4098 		ASSERT(ppa[pg_idx]->p_lckcnt != 0);
4099 		if (ppa[pg_idx]->p_cowcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
4100 			mutex_exit(&page_llock);
4101 			return (0);
4102 		}
4103 		if (ppa[pg_idx]->p_lckcnt > 1)
4104 			lckpgs++;
4105 	}
4106 
4107 	if (lckpgs != 0) {
4108 		mutex_enter(&freemem_lock);
4109 		if (availrmem >= pages_pp_maximum + lckpgs) {
4110 			availrmem -= lckpgs;
4111 			pages_claimed += lckpgs;
4112 		} else {
4113 			mutex_exit(&freemem_lock);
4114 			mutex_exit(&page_llock);
4115 			return (0);
4116 		}
4117 		mutex_exit(&freemem_lock);
4118 	}
4119 
4120 	for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) {
4121 		ppa[pg_idx]->p_lckcnt--;
4122 		ppa[pg_idx]->p_cowcnt++;
4123 	}
4124 	mutex_exit(&page_llock);
4125 	return (1);
4126 }
4127 
4128 int
4129 page_subclaim_pages(page_t  **ppa)
4130 {
4131 	pgcnt_t	ulckpgs = 0, pg_idx;
4132 
4133 	VM_STAT_ADD(pagecnt.pc_subclaim_pages);
4134 
4135 	mutex_enter(&page_llock);
4136 	for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) {
4137 
4138 		ASSERT(PAGE_LOCKED(ppa[pg_idx]));
4139 		ASSERT(ppa[pg_idx]->p_cowcnt != 0);
4140 		if (ppa[pg_idx]->p_lckcnt == (ushort_t)PAGE_LOCK_MAXIMUM) {
4141 			mutex_exit(&page_llock);
4142 			return (0);
4143 		}
4144 		if (ppa[pg_idx]->p_lckcnt != 0)
4145 			ulckpgs++;
4146 	}
4147 
4148 	if (ulckpgs != 0) {
4149 		mutex_enter(&freemem_lock);
4150 		availrmem += ulckpgs;
4151 		pages_claimed -= ulckpgs;
4152 		mutex_exit(&freemem_lock);
4153 	}
4154 
4155 	for (pg_idx = 0; ppa[pg_idx] != NULL; pg_idx++) {
4156 		ppa[pg_idx]->p_cowcnt--;
4157 		ppa[pg_idx]->p_lckcnt++;
4158 
4159 	}
4160 	mutex_exit(&page_llock);
4161 	return (1);
4162 }
4163 
4164 page_t *
4165 page_numtopp(pfn_t pfnum, se_t se)
4166 {
4167 	page_t *pp;
4168 
4169 retry:
4170 	pp = page_numtopp_nolock(pfnum);
4171 	if (pp == NULL) {
4172 		return ((page_t *)NULL);
4173 	}
4174 
4175 	/*
4176 	 * Acquire the appropriate lock on the page.
4177 	 */
4178 	while (!page_lock(pp, se, (kmutex_t *)NULL, P_RECLAIM)) {
4179 		if (page_pptonum(pp) != pfnum)
4180 			goto retry;
4181 		continue;
4182 	}
4183 
4184 	if (page_pptonum(pp) != pfnum) {
4185 		page_unlock(pp);
4186 		goto retry;
4187 	}
4188 
4189 	return (pp);
4190 }
4191 
4192 page_t *
4193 page_numtopp_noreclaim(pfn_t pfnum, se_t se)
4194 {
4195 	page_t *pp;
4196 
4197 retry:
4198 	pp = page_numtopp_nolock(pfnum);
4199 	if (pp == NULL) {
4200 		return ((page_t *)NULL);
4201 	}
4202 
4203 	/*
4204 	 * Acquire the appropriate lock on the page.
4205 	 */
4206 	while (!page_lock(pp, se, (kmutex_t *)NULL, P_NO_RECLAIM)) {
4207 		if (page_pptonum(pp) != pfnum)
4208 			goto retry;
4209 		continue;
4210 	}
4211 
4212 	if (page_pptonum(pp) != pfnum) {
4213 		page_unlock(pp);
4214 		goto retry;
4215 	}
4216 
4217 	return (pp);
4218 }
4219 
4220 /*
4221  * This routine is like page_numtopp, but will only return page structs
4222  * for pages which are ok for loading into hardware using the page struct.
4223  */
4224 page_t *
4225 page_numtopp_nowait(pfn_t pfnum, se_t se)
4226 {
4227 	page_t *pp;
4228 
4229 retry:
4230 	pp = page_numtopp_nolock(pfnum);
4231 	if (pp == NULL) {
4232 		return ((page_t *)NULL);
4233 	}
4234 
4235 	/*
4236 	 * Try to acquire the appropriate lock on the page.
4237 	 */
4238 	if (PP_ISFREE(pp))
4239 		pp = NULL;
4240 	else {
4241 		if (!page_trylock(pp, se))
4242 			pp = NULL;
4243 		else {
4244 			if (page_pptonum(pp) != pfnum) {
4245 				page_unlock(pp);
4246 				goto retry;
4247 			}
4248 			if (PP_ISFREE(pp)) {
4249 				page_unlock(pp);
4250 				pp = NULL;
4251 			}
4252 		}
4253 	}
4254 	return (pp);
4255 }
4256 
4257 /*
4258  * Returns a count of dirty pages that are in the process
4259  * of being written out.  If 'cleanit' is set, try to push the page.
4260  */
4261 pgcnt_t
4262 page_busy(int cleanit)
4263 {
4264 	page_t *page0 = page_first();
4265 	page_t *pp = page0;
4266 	pgcnt_t nppbusy = 0;
4267 	u_offset_t off;
4268 
4269 	do {
4270 		vnode_t *vp = pp->p_vnode;
4271 
4272 		/*
4273 		 * A page is a candidate for syncing if it is:
4274 		 *
4275 		 * (a)	On neither the freelist nor the cachelist
4276 		 * (b)	Hashed onto a vnode
4277 		 * (c)	Not a kernel page
4278 		 * (d)	Dirty
4279 		 * (e)	Not part of a swapfile
4280 		 * (f)	a page which belongs to a real vnode; eg has a non-null
4281 		 *	v_vfsp pointer.
4282 		 * (g)	Backed by a filesystem which doesn't have a
4283 		 *	stubbed-out sync operation
4284 		 */
4285 		if (!PP_ISFREE(pp) && vp != NULL && !VN_ISKAS(vp) &&
4286 		    hat_ismod(pp) && !IS_SWAPVP(vp) && vp->v_vfsp != NULL &&
4287 		    vfs_can_sync(vp->v_vfsp)) {
4288 			nppbusy++;
4289 			vfs_syncprogress();
4290 
4291 			if (!cleanit)
4292 				continue;
4293 			if (!page_trylock(pp, SE_EXCL))
4294 				continue;
4295 
4296 			if (PP_ISFREE(pp) || vp == NULL || IS_SWAPVP(vp) ||
4297 			    pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
4298 			    !(hat_pagesync(pp,
4299 			    HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD) & P_MOD)) {
4300 				page_unlock(pp);
4301 				continue;
4302 			}
4303 			off = pp->p_offset;
4304 			VN_HOLD(vp);
4305 			page_unlock(pp);
4306 			(void) VOP_PUTPAGE(vp, off, PAGESIZE,
4307 			    B_ASYNC | B_FREE, kcred, NULL);
4308 			VN_RELE(vp);
4309 		}
4310 	} while ((pp = page_next(pp)) != page0);
4311 
4312 	return (nppbusy);
4313 }
4314 
4315 void page_invalidate_pages(void);
4316 
4317 /*
4318  * callback handler to vm sub-system
4319  *
4320  * callers make sure no recursive entries to this func.
4321  */
4322 /*ARGSUSED*/
4323 boolean_t
4324 callb_vm_cpr(void *arg, int code)
4325 {
4326 	if (code == CB_CODE_CPR_CHKPT)
4327 		page_invalidate_pages();
4328 	return (B_TRUE);
4329 }
4330 
4331 /*
4332  * Invalidate all pages of the system.
4333  * It shouldn't be called until all user page activities are all stopped.
4334  */
4335 void
4336 page_invalidate_pages()
4337 {
4338 	page_t *pp;
4339 	page_t *page0;
4340 	pgcnt_t nbusypages;
4341 	int retry = 0;
4342 	const int MAXRETRIES = 4;
4343 #if defined(__sparc)
4344 	extern struct vnode prom_ppages;
4345 #endif /* __sparc */
4346 
4347 top:
4348 	/*
4349 	 * Flush dirty pages and destroy the clean ones.
4350 	 */
4351 	nbusypages = 0;
4352 
4353 	pp = page0 = page_first();
4354 	do {
4355 		struct vnode	*vp;
4356 		u_offset_t	offset;
4357 		int		mod;
4358 
4359 		/*
4360 		 * skip the page if it has no vnode or the page associated
4361 		 * with the kernel vnode or prom allocated kernel mem.
4362 		 */
4363 #if defined(__sparc)
4364 		if ((vp = pp->p_vnode) == NULL || VN_ISKAS(vp) ||
4365 		    vp == &prom_ppages)
4366 #else /* x86 doesn't have prom or prom_ppage */
4367 		if ((vp = pp->p_vnode) == NULL || VN_ISKAS(vp))
4368 #endif /* __sparc */
4369 			continue;
4370 
4371 		/*
4372 		 * skip the page which is already free invalidated.
4373 		 */
4374 		if (PP_ISFREE(pp) && PP_ISAGED(pp))
4375 			continue;
4376 
4377 		/*
4378 		 * skip pages that are already locked or can't be "exclusively"
4379 		 * locked or are already free.  After we lock the page, check
4380 		 * the free and age bits again to be sure it's not destroied
4381 		 * yet.
4382 		 * To achieve max. parallelization, we use page_trylock instead
4383 		 * of page_lock so that we don't get block on individual pages
4384 		 * while we have thousands of other pages to process.
4385 		 */
4386 		if (!page_trylock(pp, SE_EXCL)) {
4387 			nbusypages++;
4388 			continue;
4389 		} else if (PP_ISFREE(pp)) {
4390 			if (!PP_ISAGED(pp)) {
4391 				page_destroy_free(pp);
4392 			} else {
4393 				page_unlock(pp);
4394 			}
4395 			continue;
4396 		}
4397 		/*
4398 		 * Is this page involved in some I/O? shared?
4399 		 *
4400 		 * The page_struct_lock need not be acquired to
4401 		 * examine these fields since the page has an
4402 		 * "exclusive" lock.
4403 		 */
4404 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
4405 			page_unlock(pp);
4406 			continue;
4407 		}
4408 
4409 		if (vp->v_type == VCHR) {
4410 			panic("vp->v_type == VCHR");
4411 			/*NOTREACHED*/
4412 		}
4413 
4414 		if (!page_try_demote_pages(pp)) {
4415 			page_unlock(pp);
4416 			continue;
4417 		}
4418 
4419 		/*
4420 		 * Check the modified bit. Leave the bits alone in hardware
4421 		 * (they will be modified if we do the putpage).
4422 		 */
4423 		mod = (hat_pagesync(pp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD)
4424 		    & P_MOD);
4425 		if (mod) {
4426 			offset = pp->p_offset;
4427 			/*
4428 			 * Hold the vnode before releasing the page lock
4429 			 * to prevent it from being freed and re-used by
4430 			 * some other thread.
4431 			 */
4432 			VN_HOLD(vp);
4433 			page_unlock(pp);
4434 			/*
4435 			 * No error return is checked here. Callers such as
4436 			 * cpr deals with the dirty pages at the dump time
4437 			 * if this putpage fails.
4438 			 */
4439 			(void) VOP_PUTPAGE(vp, offset, PAGESIZE, B_INVAL,
4440 			    kcred, NULL);
4441 			VN_RELE(vp);
4442 		} else {
4443 			page_destroy(pp, 0);
4444 		}
4445 	} while ((pp = page_next(pp)) != page0);
4446 	if (nbusypages && retry++ < MAXRETRIES) {
4447 		delay(1);
4448 		goto top;
4449 	}
4450 }
4451 
4452 /*
4453  * Replace the page "old" with the page "new" on the page hash and vnode lists
4454  *
4455  * the replacement must be done in place, ie the equivalent sequence:
4456  *
4457  *	vp = old->p_vnode;
4458  *	off = old->p_offset;
4459  *	page_do_hashout(old)
4460  *	page_do_hashin(new, vp, off)
4461  *
4462  * doesn't work, since
4463  *  1) if old is the only page on the vnode, the v_pages list has a window
4464  *     where it looks empty. This will break file system assumptions.
4465  * and
4466  *  2) pvn_vplist_dirty() can't deal with pages moving on the v_pages list.
4467  */
4468 static void
4469 page_do_relocate_hash(page_t *new, page_t *old)
4470 {
4471 	page_t	**hash_list;
4472 	vnode_t	*vp = old->p_vnode;
4473 	kmutex_t *sep;
4474 
4475 	ASSERT(PAGE_EXCL(old));
4476 	ASSERT(PAGE_EXCL(new));
4477 	ASSERT(vp != NULL);
4478 	ASSERT(MUTEX_HELD(page_vnode_mutex(vp)));
4479 	ASSERT(MUTEX_HELD(PAGE_HASH_MUTEX(PAGE_HASH_FUNC(vp, old->p_offset))));
4480 
4481 	/*
4482 	 * First find old page on the page hash list
4483 	 */
4484 	hash_list = &page_hash[PAGE_HASH_FUNC(vp, old->p_offset)];
4485 
4486 	for (;;) {
4487 		if (*hash_list == old)
4488 			break;
4489 		if (*hash_list == NULL) {
4490 			panic("page_do_hashout");
4491 			/*NOTREACHED*/
4492 		}
4493 		hash_list = &(*hash_list)->p_hash;
4494 	}
4495 
4496 	/*
4497 	 * update new and replace old with new on the page hash list
4498 	 */
4499 	new->p_vnode = old->p_vnode;
4500 	new->p_offset = old->p_offset;
4501 	new->p_hash = old->p_hash;
4502 	*hash_list = new;
4503 
4504 	if ((new->p_vnode->v_flag & VISSWAP) != 0)
4505 		PP_SETSWAP(new);
4506 
4507 	/*
4508 	 * replace old with new on the vnode's page list
4509 	 */
4510 	if (old->p_vpnext == old) {
4511 		new->p_vpnext = new;
4512 		new->p_vpprev = new;
4513 	} else {
4514 		new->p_vpnext = old->p_vpnext;
4515 		new->p_vpprev = old->p_vpprev;
4516 		new->p_vpnext->p_vpprev = new;
4517 		new->p_vpprev->p_vpnext = new;
4518 	}
4519 	if (vp->v_pages == old)
4520 		vp->v_pages = new;
4521 
4522 	/*
4523 	 * clear out the old page
4524 	 */
4525 	old->p_hash = NULL;
4526 	old->p_vpnext = NULL;
4527 	old->p_vpprev = NULL;
4528 	old->p_vnode = NULL;
4529 	PP_CLRSWAP(old);
4530 	old->p_offset = (u_offset_t)-1;
4531 	page_clr_all_props(old);
4532 
4533 	/*
4534 	 * Wake up processes waiting for this page.  The page's
4535 	 * identity has been changed, and is probably not the
4536 	 * desired page any longer.
4537 	 */
4538 	sep = page_se_mutex(old);
4539 	mutex_enter(sep);
4540 	old->p_selock &= ~SE_EWANTED;
4541 	if (CV_HAS_WAITERS(&old->p_cv))
4542 		cv_broadcast(&old->p_cv);
4543 	mutex_exit(sep);
4544 }
4545 
4546 /*
4547  * This function moves the identity of page "pp_old" to page "pp_new".
4548  * Both pages must be locked on entry.  "pp_new" is free, has no identity,
4549  * and need not be hashed out from anywhere.
4550  */
4551 void
4552 page_relocate_hash(page_t *pp_new, page_t *pp_old)
4553 {
4554 	vnode_t *vp = pp_old->p_vnode;
4555 	u_offset_t off = pp_old->p_offset;
4556 	kmutex_t *phm, *vphm;
4557 
4558 	/*
4559 	 * Rehash two pages
4560 	 */
4561 	ASSERT(PAGE_EXCL(pp_old));
4562 	ASSERT(PAGE_EXCL(pp_new));
4563 	ASSERT(vp != NULL);
4564 	ASSERT(pp_new->p_vnode == NULL);
4565 
4566 	/*
4567 	 * hashout then hashin while holding the mutexes
4568 	 */
4569 	phm = PAGE_HASH_MUTEX(PAGE_HASH_FUNC(vp, off));
4570 	mutex_enter(phm);
4571 	vphm = page_vnode_mutex(vp);
4572 	mutex_enter(vphm);
4573 
4574 	page_do_relocate_hash(pp_new, pp_old);
4575 
4576 	mutex_exit(vphm);
4577 	mutex_exit(phm);
4578 
4579 	/*
4580 	 * The page_struct_lock need not be acquired for lckcnt and
4581 	 * cowcnt since the page has an "exclusive" lock.
4582 	 */
4583 	ASSERT(pp_new->p_lckcnt == 0);
4584 	ASSERT(pp_new->p_cowcnt == 0);
4585 	pp_new->p_lckcnt = pp_old->p_lckcnt;
4586 	pp_new->p_cowcnt = pp_old->p_cowcnt;
4587 	pp_old->p_lckcnt = pp_old->p_cowcnt = 0;
4588 
4589 	/* The following comment preserved from page_flip(). */
4590 	/* XXX - Do we need to protect fsdata? */
4591 	pp_new->p_fsdata = pp_old->p_fsdata;
4592 }
4593 
4594 /*
4595  * Helper routine used to lock all remaining members of a
4596  * large page. The caller is responsible for passing in a locked
4597  * pp. If pp is a large page, then it succeeds in locking all the
4598  * remaining constituent pages or it returns with only the
4599  * original page locked.
4600  *
4601  * Returns 1 on success, 0 on failure.
4602  *
4603  * If success is returned this routine guarantees p_szc for all constituent
4604  * pages of a large page pp belongs to can't change. To achieve this we
4605  * recheck szc of pp after locking all constituent pages and retry if szc
4606  * changed (it could only decrease). Since hat_page_demote() needs an EXCL
4607  * lock on one of constituent pages it can't be running after all constituent
4608  * pages are locked.  hat_page_demote() with a lock on a constituent page
4609  * outside of this large page (i.e. pp belonged to a larger large page) is
4610  * already done with all constituent pages of pp since the root's p_szc is
4611  * changed last. Therefore no need to synchronize with hat_page_demote() that
4612  * locked a constituent page outside of pp's current large page.
4613  */
4614 #ifdef DEBUG
4615 uint32_t gpg_trylock_mtbf = 0;
4616 #endif
4617 
4618 int
4619 group_page_trylock(page_t *pp, se_t se)
4620 {
4621 	page_t  *tpp;
4622 	pgcnt_t	npgs, i, j;
4623 	uint_t pszc = pp->p_szc;
4624 
4625 #ifdef DEBUG
4626 	if (gpg_trylock_mtbf && !(gethrtime() % gpg_trylock_mtbf)) {
4627 		return (0);
4628 	}
4629 #endif
4630 
4631 	if (pp != PP_GROUPLEADER(pp, pszc)) {
4632 		return (0);
4633 	}
4634 
4635 retry:
4636 	ASSERT(PAGE_LOCKED_SE(pp, se));
4637 	ASSERT(!PP_ISFREE(pp));
4638 	if (pszc == 0) {
4639 		return (1);
4640 	}
4641 	npgs = page_get_pagecnt(pszc);
4642 	tpp = pp + 1;
4643 	for (i = 1; i < npgs; i++, tpp++) {
4644 		if (!page_trylock(tpp, se)) {
4645 			tpp = pp + 1;
4646 			for (j = 1; j < i; j++, tpp++) {
4647 				page_unlock(tpp);
4648 			}
4649 			return (0);
4650 		}
4651 	}
4652 	if (pp->p_szc != pszc) {
4653 		ASSERT(pp->p_szc < pszc);
4654 		ASSERT(pp->p_vnode != NULL && !PP_ISKAS(pp) &&
4655 		    !IS_SWAPFSVP(pp->p_vnode));
4656 		tpp = pp + 1;
4657 		for (i = 1; i < npgs; i++, tpp++) {
4658 			page_unlock(tpp);
4659 		}
4660 		pszc = pp->p_szc;
4661 		goto retry;
4662 	}
4663 	return (1);
4664 }
4665 
4666 void
4667 group_page_unlock(page_t *pp)
4668 {
4669 	page_t *tpp;
4670 	pgcnt_t	npgs, i;
4671 
4672 	ASSERT(PAGE_LOCKED(pp));
4673 	ASSERT(!PP_ISFREE(pp));
4674 	ASSERT(pp == PP_PAGEROOT(pp));
4675 	npgs = page_get_pagecnt(pp->p_szc);
4676 	for (i = 1, tpp = pp + 1; i < npgs; i++, tpp++) {
4677 		page_unlock(tpp);
4678 	}
4679 }
4680 
4681 /*
4682  * returns
4683  * 0 		: on success and *nrelocp is number of relocated PAGESIZE pages
4684  * ERANGE	: this is not a base page
4685  * EBUSY	: failure to get locks on the page/pages
4686  * ENOMEM	: failure to obtain replacement pages
4687  * EAGAIN	: OBP has not yet completed its boot-time handoff to the kernel
4688  * EIO		: An error occurred while trying to copy the page data
4689  *
4690  * Return with all constituent members of target and replacement
4691  * SE_EXCL locked. It is the callers responsibility to drop the
4692  * locks.
4693  */
4694 int
4695 do_page_relocate(
4696 	page_t **target,
4697 	page_t **replacement,
4698 	int grouplock,
4699 	spgcnt_t *nrelocp,
4700 	lgrp_t *lgrp)
4701 {
4702 	page_t *first_repl;
4703 	page_t *repl;
4704 	page_t *targ;
4705 	page_t *pl = NULL;
4706 	uint_t ppattr;
4707 	pfn_t   pfn, repl_pfn;
4708 	uint_t	szc;
4709 	spgcnt_t npgs, i;
4710 	int repl_contig = 0;
4711 	uint_t flags = 0;
4712 	spgcnt_t dofree = 0;
4713 
4714 	*nrelocp = 0;
4715 
4716 #if defined(__sparc)
4717 	/*
4718 	 * We need to wait till OBP has completed
4719 	 * its boot-time handoff of its resources to the kernel
4720 	 * before we allow page relocation
4721 	 */
4722 	if (page_relocate_ready == 0) {
4723 		return (EAGAIN);
4724 	}
4725 #endif
4726 
4727 	/*
4728 	 * If this is not a base page,
4729 	 * just return with 0x0 pages relocated.
4730 	 */
4731 	targ = *target;
4732 	ASSERT(PAGE_EXCL(targ));
4733 	ASSERT(!PP_ISFREE(targ));
4734 	szc = targ->p_szc;
4735 	ASSERT(szc < mmu_page_sizes);
4736 	VM_STAT_ADD(vmm_vmstats.ppr_reloc[szc]);
4737 	pfn = targ->p_pagenum;
4738 	if (pfn != PFN_BASE(pfn, szc)) {
4739 		VM_STAT_ADD(vmm_vmstats.ppr_relocnoroot[szc]);
4740 		return (ERANGE);
4741 	}
4742 
4743 	if ((repl = *replacement) != NULL && repl->p_szc >= szc) {
4744 		repl_pfn = repl->p_pagenum;
4745 		if (repl_pfn != PFN_BASE(repl_pfn, szc)) {
4746 			VM_STAT_ADD(vmm_vmstats.ppr_reloc_replnoroot[szc]);
4747 			return (ERANGE);
4748 		}
4749 		repl_contig = 1;
4750 	}
4751 
4752 	/*
4753 	 * We must lock all members of this large page or we cannot
4754 	 * relocate any part of it.
4755 	 */
4756 	if (grouplock != 0 && !group_page_trylock(targ, SE_EXCL)) {
4757 		VM_STAT_ADD(vmm_vmstats.ppr_relocnolock[targ->p_szc]);
4758 		return (EBUSY);
4759 	}
4760 
4761 	/*
4762 	 * reread szc it could have been decreased before
4763 	 * group_page_trylock() was done.
4764 	 */
4765 	szc = targ->p_szc;
4766 	ASSERT(szc < mmu_page_sizes);
4767 	VM_STAT_ADD(vmm_vmstats.ppr_reloc[szc]);
4768 	ASSERT(pfn == PFN_BASE(pfn, szc));
4769 
4770 	npgs = page_get_pagecnt(targ->p_szc);
4771 
4772 	if (repl == NULL) {
4773 		dofree = npgs;		/* Size of target page in MMU pages */
4774 		if (!page_create_wait(dofree, 0)) {
4775 			if (grouplock != 0) {
4776 				group_page_unlock(targ);
4777 			}
4778 			VM_STAT_ADD(vmm_vmstats.ppr_relocnomem[szc]);
4779 			return (ENOMEM);
4780 		}
4781 
4782 		/*
4783 		 * seg kmem pages require that the target and replacement
4784 		 * page be the same pagesize.
4785 		 */
4786 		flags = (VN_ISKAS(targ->p_vnode)) ? PGR_SAMESZC : 0;
4787 		repl = page_get_replacement_page(targ, lgrp, flags);
4788 		if (repl == NULL) {
4789 			if (grouplock != 0) {
4790 				group_page_unlock(targ);
4791 			}
4792 			page_create_putback(dofree);
4793 			VM_STAT_ADD(vmm_vmstats.ppr_relocnomem[szc]);
4794 			return (ENOMEM);
4795 		}
4796 	}
4797 #ifdef DEBUG
4798 	else {
4799 		ASSERT(PAGE_LOCKED(repl));
4800 	}
4801 #endif /* DEBUG */
4802 
4803 #if defined(__sparc)
4804 	/*
4805 	 * Let hat_page_relocate() complete the relocation if it's kernel page
4806 	 */
4807 	if (VN_ISKAS(targ->p_vnode)) {
4808 		*replacement = repl;
4809 		if (hat_page_relocate(target, replacement, nrelocp) != 0) {
4810 			if (grouplock != 0) {
4811 				group_page_unlock(targ);
4812 			}
4813 			if (dofree) {
4814 				*replacement = NULL;
4815 				page_free_replacement_page(repl);
4816 				page_create_putback(dofree);
4817 			}
4818 			VM_STAT_ADD(vmm_vmstats.ppr_krelocfail[szc]);
4819 			return (EAGAIN);
4820 		}
4821 		VM_STAT_ADD(vmm_vmstats.ppr_relocok[szc]);
4822 		return (0);
4823 	}
4824 #else
4825 #if defined(lint)
4826 	dofree = dofree;
4827 #endif
4828 #endif
4829 
4830 	first_repl = repl;
4831 
4832 	for (i = 0; i < npgs; i++) {
4833 		ASSERT(PAGE_EXCL(targ));
4834 		ASSERT(targ->p_slckcnt == 0);
4835 		ASSERT(repl->p_slckcnt == 0);
4836 
4837 		(void) hat_pageunload(targ, HAT_FORCE_PGUNLOAD);
4838 
4839 		ASSERT(hat_page_getshare(targ) == 0);
4840 		ASSERT(!PP_ISFREE(targ));
4841 		ASSERT(targ->p_pagenum == (pfn + i));
4842 		ASSERT(repl_contig == 0 ||
4843 		    repl->p_pagenum == (repl_pfn + i));
4844 
4845 		/*
4846 		 * Copy the page contents and attributes then
4847 		 * relocate the page in the page hash.
4848 		 */
4849 		if (ppcopy(targ, repl) == 0) {
4850 			targ = *target;
4851 			repl = first_repl;
4852 			VM_STAT_ADD(vmm_vmstats.ppr_copyfail);
4853 			if (grouplock != 0) {
4854 				group_page_unlock(targ);
4855 			}
4856 			if (dofree) {
4857 				*replacement = NULL;
4858 				page_free_replacement_page(repl);
4859 				page_create_putback(dofree);
4860 			}
4861 			return (EIO);
4862 		}
4863 
4864 		targ++;
4865 		if (repl_contig != 0) {
4866 			repl++;
4867 		} else {
4868 			repl = repl->p_next;
4869 		}
4870 	}
4871 
4872 	repl = first_repl;
4873 	targ = *target;
4874 
4875 	for (i = 0; i < npgs; i++) {
4876 		ppattr = hat_page_getattr(targ, (P_MOD | P_REF | P_RO));
4877 		page_clr_all_props(repl);
4878 		page_set_props(repl, ppattr);
4879 		page_relocate_hash(repl, targ);
4880 
4881 		ASSERT(hat_page_getshare(targ) == 0);
4882 		ASSERT(hat_page_getshare(repl) == 0);
4883 		/*
4884 		 * Now clear the props on targ, after the
4885 		 * page_relocate_hash(), they no longer
4886 		 * have any meaning.
4887 		 */
4888 		page_clr_all_props(targ);
4889 		ASSERT(targ->p_next == targ);
4890 		ASSERT(targ->p_prev == targ);
4891 		page_list_concat(&pl, &targ);
4892 
4893 		targ++;
4894 		if (repl_contig != 0) {
4895 			repl++;
4896 		} else {
4897 			repl = repl->p_next;
4898 		}
4899 	}
4900 	/* assert that we have come full circle with repl */
4901 	ASSERT(repl_contig == 1 || first_repl == repl);
4902 
4903 	*target = pl;
4904 	if (*replacement == NULL) {
4905 		ASSERT(first_repl == repl);
4906 		*replacement = repl;
4907 	}
4908 	VM_STAT_ADD(vmm_vmstats.ppr_relocok[szc]);
4909 	*nrelocp = npgs;
4910 	return (0);
4911 }
4912 /*
4913  * On success returns 0 and *nrelocp the number of PAGESIZE pages relocated.
4914  */
4915 int
4916 page_relocate(
4917 	page_t **target,
4918 	page_t **replacement,
4919 	int grouplock,
4920 	int freetarget,
4921 	spgcnt_t *nrelocp,
4922 	lgrp_t *lgrp)
4923 {
4924 	spgcnt_t ret;
4925 
4926 	/* do_page_relocate returns 0 on success or errno value */
4927 	ret = do_page_relocate(target, replacement, grouplock, nrelocp, lgrp);
4928 
4929 	if (ret != 0 || freetarget == 0) {
4930 		return (ret);
4931 	}
4932 	if (*nrelocp == 1) {
4933 		ASSERT(*target != NULL);
4934 		page_free(*target, 1);
4935 	} else {
4936 		page_t *tpp = *target;
4937 		uint_t szc = tpp->p_szc;
4938 		pgcnt_t npgs = page_get_pagecnt(szc);
4939 		ASSERT(npgs > 1);
4940 		ASSERT(szc != 0);
4941 		do {
4942 			ASSERT(PAGE_EXCL(tpp));
4943 			ASSERT(!hat_page_is_mapped(tpp));
4944 			ASSERT(tpp->p_szc == szc);
4945 			PP_SETFREE(tpp);
4946 			PP_SETAGED(tpp);
4947 			npgs--;
4948 		} while ((tpp = tpp->p_next) != *target);
4949 		ASSERT(npgs == 0);
4950 		page_list_add_pages(*target, 0);
4951 		npgs = page_get_pagecnt(szc);
4952 		page_create_putback(npgs);
4953 	}
4954 	return (ret);
4955 }
4956 
4957 /*
4958  * it is up to the caller to deal with pcf accounting.
4959  */
4960 void
4961 page_free_replacement_page(page_t *pplist)
4962 {
4963 	page_t *pp;
4964 
4965 	while (pplist != NULL) {
4966 		/*
4967 		 * pp_targ is a linked list.
4968 		 */
4969 		pp = pplist;
4970 		if (pp->p_szc == 0) {
4971 			page_sub(&pplist, pp);
4972 			page_clr_all_props(pp);
4973 			PP_SETFREE(pp);
4974 			PP_SETAGED(pp);
4975 			page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL);
4976 			page_unlock(pp);
4977 			VM_STAT_ADD(pagecnt.pc_free_replacement_page[0]);
4978 		} else {
4979 			spgcnt_t curnpgs = page_get_pagecnt(pp->p_szc);
4980 			page_t *tpp;
4981 			page_list_break(&pp, &pplist, curnpgs);
4982 			tpp = pp;
4983 			do {
4984 				ASSERT(PAGE_EXCL(tpp));
4985 				ASSERT(!hat_page_is_mapped(tpp));
4986 				page_clr_all_props(pp);
4987 				PP_SETFREE(tpp);
4988 				PP_SETAGED(tpp);
4989 			} while ((tpp = tpp->p_next) != pp);
4990 			page_list_add_pages(pp, 0);
4991 			VM_STAT_ADD(pagecnt.pc_free_replacement_page[1]);
4992 		}
4993 	}
4994 }
4995 
4996 /*
4997  * Relocate target to non-relocatable replacement page.
4998  */
4999 int
5000 page_relocate_cage(page_t **target, page_t **replacement)
5001 {
5002 	page_t *tpp, *rpp;
5003 	spgcnt_t pgcnt, npgs;
5004 	int result;
5005 
5006 	tpp = *target;
5007 
5008 	ASSERT(PAGE_EXCL(tpp));
5009 	ASSERT(tpp->p_szc == 0);
5010 
5011 	pgcnt = btop(page_get_pagesize(tpp->p_szc));
5012 
5013 	do {
5014 		(void) page_create_wait(pgcnt, PG_WAIT | PG_NORELOC);
5015 		rpp = page_get_replacement_page(tpp, NULL, PGR_NORELOC);
5016 		if (rpp == NULL) {
5017 			page_create_putback(pgcnt);
5018 			kcage_cageout_wakeup();
5019 		}
5020 	} while (rpp == NULL);
5021 
5022 	ASSERT(PP_ISNORELOC(rpp));
5023 
5024 	result = page_relocate(&tpp, &rpp, 0, 1, &npgs, NULL);
5025 
5026 	if (result == 0) {
5027 		*replacement = rpp;
5028 		if (pgcnt != npgs)
5029 			panic("page_relocate_cage: partial relocation");
5030 	}
5031 
5032 	return (result);
5033 }
5034 
5035 /*
5036  * Release the page lock on a page, place on cachelist
5037  * tail if no longer mapped. Caller can let us know if
5038  * the page is known to be clean.
5039  */
5040 int
5041 page_release(page_t *pp, int checkmod)
5042 {
5043 	int status;
5044 
5045 	ASSERT(PAGE_LOCKED(pp) && !PP_ISFREE(pp) &&
5046 	    (pp->p_vnode != NULL));
5047 
5048 	if (!hat_page_is_mapped(pp) && !IS_SWAPVP(pp->p_vnode) &&
5049 	    ((PAGE_SHARED(pp) && page_tryupgrade(pp)) || PAGE_EXCL(pp)) &&
5050 	    pp->p_lckcnt == 0 && pp->p_cowcnt == 0 &&
5051 	    !hat_page_is_mapped(pp)) {
5052 
5053 		/*
5054 		 * If page is modified, unlock it
5055 		 *
5056 		 * (p_nrm & P_MOD) bit has the latest stuff because:
5057 		 * (1) We found that this page doesn't have any mappings
5058 		 *	_after_ holding SE_EXCL and
5059 		 * (2) We didn't drop SE_EXCL lock after the check in (1)
5060 		 */
5061 		if (checkmod && hat_ismod(pp)) {
5062 			page_unlock(pp);
5063 			status = PGREL_MOD;
5064 		} else {
5065 			/*LINTED: constant in conditional context*/
5066 			VN_DISPOSE(pp, B_FREE, 0, kcred);
5067 			status = PGREL_CLEAN;
5068 		}
5069 	} else {
5070 		page_unlock(pp);
5071 		status = PGREL_NOTREL;
5072 	}
5073 	return (status);
5074 }
5075 
5076 /*
5077  * Given a constituent page, try to demote the large page on the freelist.
5078  *
5079  * Returns nonzero if the page could be demoted successfully. Returns with
5080  * the constituent page still locked.
5081  */
5082 int
5083 page_try_demote_free_pages(page_t *pp)
5084 {
5085 	page_t *rootpp = pp;
5086 	pfn_t	pfn = page_pptonum(pp);
5087 	spgcnt_t npgs;
5088 	uint_t	szc = pp->p_szc;
5089 
5090 	ASSERT(PP_ISFREE(pp));
5091 	ASSERT(PAGE_EXCL(pp));
5092 
5093 	/*
5094 	 * Adjust rootpp and lock it, if `pp' is not the base
5095 	 * constituent page.
5096 	 */
5097 	npgs = page_get_pagecnt(pp->p_szc);
5098 	if (npgs == 1) {
5099 		return (0);
5100 	}
5101 
5102 	if (!IS_P2ALIGNED(pfn, npgs)) {
5103 		pfn = P2ALIGN(pfn, npgs);
5104 		rootpp = page_numtopp_nolock(pfn);
5105 	}
5106 
5107 	if (pp != rootpp && !page_trylock(rootpp, SE_EXCL)) {
5108 		return (0);
5109 	}
5110 
5111 	if (rootpp->p_szc != szc) {
5112 		if (pp != rootpp)
5113 			page_unlock(rootpp);
5114 		return (0);
5115 	}
5116 
5117 	page_demote_free_pages(rootpp);
5118 
5119 	if (pp != rootpp)
5120 		page_unlock(rootpp);
5121 
5122 	ASSERT(PP_ISFREE(pp));
5123 	ASSERT(PAGE_EXCL(pp));
5124 	return (1);
5125 }
5126 
5127 /*
5128  * Given a constituent page, try to demote the large page.
5129  *
5130  * Returns nonzero if the page could be demoted successfully. Returns with
5131  * the constituent page still locked.
5132  */
5133 int
5134 page_try_demote_pages(page_t *pp)
5135 {
5136 	page_t *tpp, *rootpp = pp;
5137 	pfn_t	pfn = page_pptonum(pp);
5138 	spgcnt_t i, npgs;
5139 	uint_t	szc = pp->p_szc;
5140 	vnode_t *vp = pp->p_vnode;
5141 
5142 	ASSERT(PAGE_EXCL(pp));
5143 
5144 	VM_STAT_ADD(pagecnt.pc_try_demote_pages[0]);
5145 
5146 	if (pp->p_szc == 0) {
5147 		VM_STAT_ADD(pagecnt.pc_try_demote_pages[1]);
5148 		return (1);
5149 	}
5150 
5151 	if (vp != NULL && !IS_SWAPFSVP(vp) && !VN_ISKAS(vp)) {
5152 		VM_STAT_ADD(pagecnt.pc_try_demote_pages[2]);
5153 		page_demote_vp_pages(pp);
5154 		ASSERT(pp->p_szc == 0);
5155 		return (1);
5156 	}
5157 
5158 	/*
5159 	 * Adjust rootpp if passed in is not the base
5160 	 * constituent page.
5161 	 */
5162 	npgs = page_get_pagecnt(pp->p_szc);
5163 	ASSERT(npgs > 1);
5164 	if (!IS_P2ALIGNED(pfn, npgs)) {
5165 		pfn = P2ALIGN(pfn, npgs);
5166 		rootpp = page_numtopp_nolock(pfn);
5167 		VM_STAT_ADD(pagecnt.pc_try_demote_pages[3]);
5168 		ASSERT(rootpp->p_vnode != NULL);
5169 		ASSERT(rootpp->p_szc == szc);
5170 	}
5171 
5172 	/*
5173 	 * We can't demote kernel pages since we can't hat_unload()
5174 	 * the mappings.
5175 	 */
5176 	if (VN_ISKAS(rootpp->p_vnode))
5177 		return (0);
5178 
5179 	/*
5180 	 * Attempt to lock all constituent pages except the page passed
5181 	 * in since it's already locked.
5182 	 */
5183 	for (tpp = rootpp, i = 0; i < npgs; i++, tpp++) {
5184 		ASSERT(!PP_ISFREE(tpp));
5185 		ASSERT(tpp->p_vnode != NULL);
5186 
5187 		if (tpp != pp && !page_trylock(tpp, SE_EXCL))
5188 			break;
5189 		ASSERT(tpp->p_szc == rootpp->p_szc);
5190 		ASSERT(page_pptonum(tpp) == page_pptonum(rootpp) + i);
5191 	}
5192 
5193 	/*
5194 	 * If we failed to lock them all then unlock what we have
5195 	 * locked so far and bail.
5196 	 */
5197 	if (i < npgs) {
5198 		tpp = rootpp;
5199 		while (i-- > 0) {
5200 			if (tpp != pp)
5201 				page_unlock(tpp);
5202 			tpp++;
5203 		}
5204 		VM_STAT_ADD(pagecnt.pc_try_demote_pages[4]);
5205 		return (0);
5206 	}
5207 
5208 	for (tpp = rootpp, i = 0; i < npgs; i++, tpp++) {
5209 		ASSERT(PAGE_EXCL(tpp));
5210 		ASSERT(tpp->p_slckcnt == 0);
5211 		(void) hat_pageunload(tpp, HAT_FORCE_PGUNLOAD);
5212 		tpp->p_szc = 0;
5213 	}
5214 
5215 	/*
5216 	 * Unlock all pages except the page passed in.
5217 	 */
5218 	for (tpp = rootpp, i = 0; i < npgs; i++, tpp++) {
5219 		ASSERT(!hat_page_is_mapped(tpp));
5220 		if (tpp != pp)
5221 			page_unlock(tpp);
5222 	}
5223 
5224 	VM_STAT_ADD(pagecnt.pc_try_demote_pages[5]);
5225 	return (1);
5226 }
5227 
5228 /*
5229  * Called by page_free() and page_destroy() to demote the page size code
5230  * (p_szc) to 0 (since we can't just put a single PAGESIZE page with non zero
5231  * p_szc on free list, neither can we just clear p_szc of a single page_t
5232  * within a large page since it will break other code that relies on p_szc
5233  * being the same for all page_t's of a large page). Anonymous pages should
5234  * never end up here because anon_map_getpages() cannot deal with p_szc
5235  * changes after a single constituent page is locked.  While anonymous or
5236  * kernel large pages are demoted or freed the entire large page at a time
5237  * with all constituent pages locked EXCL for the file system pages we
5238  * have to be able to demote a large page (i.e. decrease all constituent pages
5239  * p_szc) with only just an EXCL lock on one of constituent pages. The reason
5240  * we can easily deal with anonymous page demotion the entire large page at a
5241  * time is that those operation originate at address space level and concern
5242  * the entire large page region with actual demotion only done when pages are
5243  * not shared with any other processes (therefore we can always get EXCL lock
5244  * on all anonymous constituent pages after clearing segment page
5245  * cache). However file system pages can be truncated or invalidated at a
5246  * PAGESIZE level from the file system side and end up in page_free() or
5247  * page_destroy() (we also allow only part of the large page to be SOFTLOCKed
5248  * and therefore pageout should be able to demote a large page by EXCL locking
5249  * any constituent page that is not under SOFTLOCK). In those cases we cannot
5250  * rely on being able to lock EXCL all constituent pages.
5251  *
5252  * To prevent szc changes on file system pages one has to lock all constituent
5253  * pages at least SHARED (or call page_szc_lock()). The only subsystem that
5254  * doesn't rely on locking all constituent pages (or using page_szc_lock()) to
5255  * prevent szc changes is hat layer that uses its own page level mlist
5256  * locks. hat assumes that szc doesn't change after mlist lock for a page is
5257  * taken. Therefore we need to change szc under hat level locks if we only
5258  * have an EXCL lock on a single constituent page and hat still references any
5259  * of constituent pages.  (Note we can't "ignore" hat layer by simply
5260  * hat_pageunload() all constituent pages without having EXCL locks on all of
5261  * constituent pages). We use hat_page_demote() call to safely demote szc of
5262  * all constituent pages under hat locks when we only have an EXCL lock on one
5263  * of constituent pages.
5264  *
5265  * This routine calls page_szc_lock() before calling hat_page_demote() to
5266  * allow segvn in one special case not to lock all constituent pages SHARED
5267  * before calling hat_memload_array() that relies on p_szc not changing even
5268  * before hat level mlist lock is taken.  In that case segvn uses
5269  * page_szc_lock() to prevent hat_page_demote() changing p_szc values.
5270  *
5271  * Anonymous or kernel page demotion still has to lock all pages exclusively
5272  * and do hat_pageunload() on all constituent pages before demoting the page
5273  * therefore there's no need for anonymous or kernel page demotion to use
5274  * hat_page_demote() mechanism.
5275  *
5276  * hat_page_demote() removes all large mappings that map pp and then decreases
5277  * p_szc starting from the last constituent page of the large page. By working
5278  * from the tail of a large page in pfn decreasing order allows one looking at
5279  * the root page to know that hat_page_demote() is done for root's szc area.
5280  * e.g. if a root page has szc 1 one knows it only has to lock all constituent
5281  * pages within szc 1 area to prevent szc changes because hat_page_demote()
5282  * that started on this page when it had szc > 1 is done for this szc 1 area.
5283  *
5284  * We are guaranteed that all constituent pages of pp's large page belong to
5285  * the same vnode with the consecutive offsets increasing in the direction of
5286  * the pfn i.e. the identity of constituent pages can't change until their
5287  * p_szc is decreased. Therefore it's safe for hat_page_demote() to remove
5288  * large mappings to pp even though we don't lock any constituent page except
5289  * pp (i.e. we won't unload e.g. kernel locked page).
5290  */
5291 static void
5292 page_demote_vp_pages(page_t *pp)
5293 {
5294 	kmutex_t *mtx;
5295 
5296 	ASSERT(PAGE_EXCL(pp));
5297 	ASSERT(!PP_ISFREE(pp));
5298 	ASSERT(pp->p_vnode != NULL);
5299 	ASSERT(!IS_SWAPFSVP(pp->p_vnode));
5300 	ASSERT(!PP_ISKAS(pp));
5301 
5302 	VM_STAT_ADD(pagecnt.pc_demote_pages[0]);
5303 
5304 	mtx = page_szc_lock(pp);
5305 	if (mtx != NULL) {
5306 		hat_page_demote(pp);
5307 		mutex_exit(mtx);
5308 	}
5309 	ASSERT(pp->p_szc == 0);
5310 }
5311 
5312 /*
5313  * Mark any existing pages for migration in the given range
5314  */
5315 void
5316 page_mark_migrate(struct seg *seg, caddr_t addr, size_t len,
5317     struct anon_map *amp, ulong_t anon_index, vnode_t *vp,
5318     u_offset_t vnoff, int rflag)
5319 {
5320 	struct anon	*ap;
5321 	vnode_t		*curvp;
5322 	lgrp_t		*from;
5323 	pgcnt_t		i;
5324 	pgcnt_t		nlocked;
5325 	u_offset_t	off;
5326 	pfn_t		pfn;
5327 	size_t		pgsz;
5328 	size_t		segpgsz;
5329 	pgcnt_t		pages;
5330 	uint_t		pszc;
5331 	page_t		**ppa;
5332 	pgcnt_t		ppa_nentries;
5333 	page_t		*pp;
5334 	caddr_t		va;
5335 	ulong_t		an_idx;
5336 	anon_sync_obj_t	cookie;
5337 
5338 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5339 
5340 	/*
5341 	 * Don't do anything if don't need to do lgroup optimizations
5342 	 * on this system
5343 	 */
5344 	if (!lgrp_optimizations())
5345 		return;
5346 
5347 	/*
5348 	 * Align address and length to (potentially large) page boundary
5349 	 */
5350 	segpgsz = page_get_pagesize(seg->s_szc);
5351 	addr = (caddr_t)P2ALIGN((uintptr_t)addr, segpgsz);
5352 	if (rflag)
5353 		len = P2ROUNDUP(len, segpgsz);
5354 
5355 	/*
5356 	 * Allocate page array to accommodate largest page size
5357 	 */
5358 	pgsz = page_get_pagesize(page_num_pagesizes() - 1);
5359 	ppa_nentries = btop(pgsz);
5360 	ppa = kmem_zalloc(ppa_nentries * sizeof (page_t *), KM_SLEEP);
5361 
5362 	/*
5363 	 * Do one (large) page at a time
5364 	 */
5365 	va = addr;
5366 	while (va < addr + len) {
5367 		/*
5368 		 * Lookup (root) page for vnode and offset corresponding to
5369 		 * this virtual address
5370 		 * Try anonmap first since there may be copy-on-write
5371 		 * pages, but initialize vnode pointer and offset using
5372 		 * vnode arguments just in case there isn't an amp.
5373 		 */
5374 		curvp = vp;
5375 		off = vnoff + va - seg->s_base;
5376 		if (amp) {
5377 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5378 			an_idx = anon_index + seg_page(seg, va);
5379 			anon_array_enter(amp, an_idx, &cookie);
5380 			ap = anon_get_ptr(amp->ahp, an_idx);
5381 			if (ap)
5382 				swap_xlate(ap, &curvp, &off);
5383 			anon_array_exit(&cookie);
5384 			ANON_LOCK_EXIT(&amp->a_rwlock);
5385 		}
5386 
5387 		pp = NULL;
5388 		if (curvp)
5389 			pp = page_lookup(curvp, off, SE_SHARED);
5390 
5391 		/*
5392 		 * If there isn't a page at this virtual address,
5393 		 * skip to next page
5394 		 */
5395 		if (pp == NULL) {
5396 			va += PAGESIZE;
5397 			continue;
5398 		}
5399 
5400 		/*
5401 		 * Figure out which lgroup this page is in for kstats
5402 		 */
5403 		pfn = page_pptonum(pp);
5404 		from = lgrp_pfn_to_lgrp(pfn);
5405 
5406 		/*
5407 		 * Get page size, and round up and skip to next page boundary
5408 		 * if unaligned address
5409 		 */
5410 		pszc = pp->p_szc;
5411 		pgsz = page_get_pagesize(pszc);
5412 		pages = btop(pgsz);
5413 		if (!IS_P2ALIGNED(va, pgsz) ||
5414 		    !IS_P2ALIGNED(pfn, pages) ||
5415 		    pgsz > segpgsz) {
5416 			pgsz = MIN(pgsz, segpgsz);
5417 			page_unlock(pp);
5418 			i = btop(P2END((uintptr_t)va, pgsz) -
5419 			    (uintptr_t)va);
5420 			va = (caddr_t)P2END((uintptr_t)va, pgsz);
5421 			lgrp_stat_add(from->lgrp_id, LGRP_PMM_FAIL_PGS, i);
5422 			continue;
5423 		}
5424 
5425 		/*
5426 		 * Upgrade to exclusive lock on page
5427 		 */
5428 		if (!page_tryupgrade(pp)) {
5429 			page_unlock(pp);
5430 			va += pgsz;
5431 			lgrp_stat_add(from->lgrp_id, LGRP_PMM_FAIL_PGS,
5432 			    btop(pgsz));
5433 			continue;
5434 		}
5435 
5436 		/*
5437 		 * Remember pages locked exclusively and how many
5438 		 */
5439 		ppa[0] = pp;
5440 		nlocked = 1;
5441 
5442 		/*
5443 		 * Lock constituent pages if this is large page
5444 		 */
5445 		if (pages > 1) {
5446 			/*
5447 			 * Lock all constituents except root page, since it
5448 			 * should be locked already.
5449 			 */
5450 			for (i = 1; i < pages; i++) {
5451 				pp++;
5452 				if (!page_trylock(pp, SE_EXCL)) {
5453 					break;
5454 				}
5455 				if (PP_ISFREE(pp) ||
5456 				    pp->p_szc != pszc) {
5457 					/*
5458 					 * hat_page_demote() raced in with us.
5459 					 */
5460 					ASSERT(!IS_SWAPFSVP(curvp));
5461 					page_unlock(pp);
5462 					break;
5463 				}
5464 				ppa[nlocked] = pp;
5465 				nlocked++;
5466 			}
5467 		}
5468 
5469 		/*
5470 		 * If all constituent pages couldn't be locked,
5471 		 * unlock pages locked so far and skip to next page.
5472 		 */
5473 		if (nlocked != pages) {
5474 			for (i = 0; i < nlocked; i++)
5475 				page_unlock(ppa[i]);
5476 			va += pgsz;
5477 			lgrp_stat_add(from->lgrp_id, LGRP_PMM_FAIL_PGS,
5478 			    btop(pgsz));
5479 			continue;
5480 		}
5481 
5482 		/*
5483 		 * hat_page_demote() can no longer happen
5484 		 * since last cons page had the right p_szc after
5485 		 * all cons pages were locked. all cons pages
5486 		 * should now have the same p_szc.
5487 		 */
5488 
5489 		/*
5490 		 * All constituent pages locked successfully, so mark
5491 		 * large page for migration and unload the mappings of
5492 		 * constituent pages, so a fault will occur on any part of the
5493 		 * large page
5494 		 */
5495 		PP_SETMIGRATE(ppa[0]);
5496 		for (i = 0; i < nlocked; i++) {
5497 			pp = ppa[i];
5498 			(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
5499 			ASSERT(hat_page_getshare(pp) == 0);
5500 			page_unlock(pp);
5501 		}
5502 		lgrp_stat_add(from->lgrp_id, LGRP_PMM_PGS, nlocked);
5503 
5504 		va += pgsz;
5505 	}
5506 	kmem_free(ppa, ppa_nentries * sizeof (page_t *));
5507 }
5508 
5509 /*
5510  * Migrate any pages that have been marked for migration in the given range
5511  */
5512 void
5513 page_migrate(
5514 	struct seg	*seg,
5515 	caddr_t		addr,
5516 	page_t		**ppa,
5517 	pgcnt_t		npages)
5518 {
5519 	lgrp_t		*from;
5520 	lgrp_t		*to;
5521 	page_t		*newpp;
5522 	page_t		*pp;
5523 	pfn_t		pfn;
5524 	size_t		pgsz;
5525 	spgcnt_t	page_cnt;
5526 	spgcnt_t	i;
5527 	uint_t		pszc;
5528 
5529 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5530 
5531 	while (npages > 0) {
5532 		pp = *ppa;
5533 		pszc = pp->p_szc;
5534 		pgsz = page_get_pagesize(pszc);
5535 		page_cnt = btop(pgsz);
5536 
5537 		/*
5538 		 * Check to see whether this page is marked for migration
5539 		 *
5540 		 * Assume that root page of large page is marked for
5541 		 * migration and none of the other constituent pages
5542 		 * are marked.  This really simplifies clearing the
5543 		 * migrate bit by not having to clear it from each
5544 		 * constituent page.
5545 		 *
5546 		 * note we don't want to relocate an entire large page if
5547 		 * someone is only using one subpage.
5548 		 */
5549 		if (npages < page_cnt)
5550 			break;
5551 
5552 		/*
5553 		 * Is it marked for migration?
5554 		 */
5555 		if (!PP_ISMIGRATE(pp))
5556 			goto next;
5557 
5558 		/*
5559 		 * Determine lgroups that page is being migrated between
5560 		 */
5561 		pfn = page_pptonum(pp);
5562 		if (!IS_P2ALIGNED(pfn, page_cnt)) {
5563 			break;
5564 		}
5565 		from = lgrp_pfn_to_lgrp(pfn);
5566 		to = lgrp_mem_choose(seg, addr, pgsz);
5567 
5568 		/*
5569 		 * Need to get exclusive lock's to migrate
5570 		 */
5571 		for (i = 0; i < page_cnt; i++) {
5572 			ASSERT(PAGE_LOCKED(ppa[i]));
5573 			if (page_pptonum(ppa[i]) != pfn + i ||
5574 			    ppa[i]->p_szc != pszc) {
5575 				break;
5576 			}
5577 			if (!page_tryupgrade(ppa[i])) {
5578 				lgrp_stat_add(from->lgrp_id,
5579 				    LGRP_PM_FAIL_LOCK_PGS,
5580 				    page_cnt);
5581 				break;
5582 			}
5583 
5584 			/*
5585 			 * Check to see whether we are trying to migrate
5586 			 * page to lgroup where it is allocated already.
5587 			 * If so, clear the migrate bit and skip to next
5588 			 * page.
5589 			 */
5590 			if (i == 0 && to == from) {
5591 				PP_CLRMIGRATE(ppa[0]);
5592 				page_downgrade(ppa[0]);
5593 				goto next;
5594 			}
5595 		}
5596 
5597 		/*
5598 		 * If all constituent pages couldn't be locked,
5599 		 * unlock pages locked so far and skip to next page.
5600 		 */
5601 		if (i != page_cnt) {
5602 			while (--i != -1) {
5603 				page_downgrade(ppa[i]);
5604 			}
5605 			goto next;
5606 		}
5607 
5608 		(void) page_create_wait(page_cnt, PG_WAIT);
5609 		newpp = page_get_replacement_page(pp, to, PGR_SAMESZC);
5610 		if (newpp == NULL) {
5611 			page_create_putback(page_cnt);
5612 			for (i = 0; i < page_cnt; i++) {
5613 				page_downgrade(ppa[i]);
5614 			}
5615 			lgrp_stat_add(to->lgrp_id, LGRP_PM_FAIL_ALLOC_PGS,
5616 			    page_cnt);
5617 			goto next;
5618 		}
5619 		ASSERT(newpp->p_szc == pszc);
5620 		/*
5621 		 * Clear migrate bit and relocate page
5622 		 */
5623 		PP_CLRMIGRATE(pp);
5624 		if (page_relocate(&pp, &newpp, 0, 1, &page_cnt, to)) {
5625 			panic("page_migrate: page_relocate failed");
5626 		}
5627 		ASSERT(page_cnt * PAGESIZE == pgsz);
5628 
5629 		/*
5630 		 * Keep stats for number of pages migrated from and to
5631 		 * each lgroup
5632 		 */
5633 		lgrp_stat_add(from->lgrp_id, LGRP_PM_SRC_PGS, page_cnt);
5634 		lgrp_stat_add(to->lgrp_id, LGRP_PM_DEST_PGS, page_cnt);
5635 		/*
5636 		 * update the page_t array we were passed in and
5637 		 * unlink constituent pages of a large page.
5638 		 */
5639 		for (i = 0; i < page_cnt; ++i, ++pp) {
5640 			ASSERT(PAGE_EXCL(newpp));
5641 			ASSERT(newpp->p_szc == pszc);
5642 			ppa[i] = newpp;
5643 			pp = newpp;
5644 			page_sub(&newpp, pp);
5645 			page_downgrade(pp);
5646 		}
5647 		ASSERT(newpp == NULL);
5648 next:
5649 		addr += pgsz;
5650 		ppa += page_cnt;
5651 		npages -= page_cnt;
5652 	}
5653 }
5654 
5655 ulong_t mem_waiters 	= 0;
5656 ulong_t	max_count 	= 20;
5657 #define	MAX_DELAY	0x1ff
5658 
5659 /*
5660  * Check if enough memory is available to proceed.
5661  * Depending on system configuration and how much memory is
5662  * reserved for swap we need to check against two variables.
5663  * e.g. on systems with little physical swap availrmem can be
5664  * more reliable indicator of how much memory is available.
5665  * On systems with large phys swap freemem can be better indicator.
5666  * If freemem drops below threshold level don't return an error
5667  * immediately but wake up pageout to free memory and block.
5668  * This is done number of times. If pageout is not able to free
5669  * memory within certain time return an error.
5670  * The same applies for availrmem but kmem_reap is used to
5671  * free memory.
5672  */
5673 int
5674 page_mem_avail(pgcnt_t npages)
5675 {
5676 	ulong_t count;
5677 
5678 #if defined(__i386)
5679 	if (freemem > desfree + npages &&
5680 	    availrmem > swapfs_reserve + npages &&
5681 	    btop(vmem_size(heap_arena, VMEM_FREE)) > tune.t_minarmem +
5682 	    npages)
5683 		return (1);
5684 #else
5685 	if (freemem > desfree + npages &&
5686 	    availrmem > swapfs_reserve + npages)
5687 		return (1);
5688 #endif
5689 
5690 	count = max_count;
5691 	atomic_add_long(&mem_waiters, 1);
5692 
5693 	while (freemem < desfree + npages && --count) {
5694 		cv_signal(&proc_pageout->p_cv);
5695 		if (delay_sig(hz + (mem_waiters & MAX_DELAY))) {
5696 			atomic_add_long(&mem_waiters, -1);
5697 			return (0);
5698 		}
5699 	}
5700 	if (count == 0) {
5701 		atomic_add_long(&mem_waiters, -1);
5702 		return (0);
5703 	}
5704 
5705 	count = max_count;
5706 	while (availrmem < swapfs_reserve + npages && --count) {
5707 		kmem_reap();
5708 		if (delay_sig(hz + (mem_waiters & MAX_DELAY))) {
5709 			atomic_add_long(&mem_waiters, -1);
5710 			return (0);
5711 		}
5712 	}
5713 	atomic_add_long(&mem_waiters, -1);
5714 	if (count == 0)
5715 		return (0);
5716 
5717 #if defined(__i386)
5718 	if (btop(vmem_size(heap_arena, VMEM_FREE)) <
5719 	    tune.t_minarmem + npages)
5720 		return (0);
5721 #endif
5722 	return (1);
5723 }
5724 
5725 #define	MAX_CNT	60	/* max num of iterations */
5726 /*
5727  * Reclaim/reserve availrmem for npages.
5728  * If there is not enough memory start reaping seg, kmem caches.
5729  * Start pageout scanner (via page_needfree()).
5730  * Exit after ~ MAX_CNT s regardless of how much memory has been released.
5731  * Note: There is no guarantee that any availrmem will be freed as
5732  * this memory typically is locked (kernel heap) or reserved for swap.
5733  * Also due to memory fragmentation kmem allocator may not be able
5734  * to free any memory (single user allocated buffer will prevent
5735  * freeing slab or a page).
5736  */
5737 int
5738 page_reclaim_mem(pgcnt_t npages, pgcnt_t epages, int adjust)
5739 {
5740 	int	i = 0;
5741 	int	ret = 0;
5742 	pgcnt_t	deficit;
5743 	pgcnt_t old_availrmem;
5744 
5745 	mutex_enter(&freemem_lock);
5746 	old_availrmem = availrmem - 1;
5747 	while ((availrmem < tune.t_minarmem + npages + epages) &&
5748 	    (old_availrmem < availrmem) && (i++ < MAX_CNT)) {
5749 		old_availrmem = availrmem;
5750 		deficit = tune.t_minarmem + npages + epages - availrmem;
5751 		mutex_exit(&freemem_lock);
5752 		page_needfree(deficit);
5753 		kmem_reap();
5754 		delay(hz);
5755 		page_needfree(-(spgcnt_t)deficit);
5756 		mutex_enter(&freemem_lock);
5757 	}
5758 
5759 	if (adjust && (availrmem >= tune.t_minarmem + npages + epages)) {
5760 		availrmem -= npages;
5761 		ret = 1;
5762 	}
5763 
5764 	mutex_exit(&freemem_lock);
5765 
5766 	return (ret);
5767 }
5768 
5769 /*
5770  * Search the memory segments to locate the desired page.  Within a
5771  * segment, pages increase linearly with one page structure per
5772  * physical page frame (size PAGESIZE).  The search begins
5773  * with the segment that was accessed last, to take advantage of locality.
5774  * If the hint misses, we start from the beginning of the sorted memseg list
5775  */
5776 
5777 
5778 /*
5779  * Some data structures for pfn to pp lookup.
5780  */
5781 ulong_t mhash_per_slot;
5782 struct memseg *memseg_hash[N_MEM_SLOTS];
5783 
5784 page_t *
5785 page_numtopp_nolock(pfn_t pfnum)
5786 {
5787 	struct memseg *seg;
5788 	page_t *pp;
5789 	vm_cpu_data_t *vc = CPU->cpu_vm_data;
5790 
5791 	ASSERT(vc != NULL);
5792 
5793 	MEMSEG_STAT_INCR(nsearch);
5794 
5795 	/* Try last winner first */
5796 	if (((seg = vc->vc_pnum_memseg) != NULL) &&
5797 	    (pfnum >= seg->pages_base) && (pfnum < seg->pages_end)) {
5798 		MEMSEG_STAT_INCR(nlastwon);
5799 		pp = seg->pages + (pfnum - seg->pages_base);
5800 		if (pp->p_pagenum == pfnum)
5801 			return ((page_t *)pp);
5802 	}
5803 
5804 	/* Else Try hash */
5805 	if (((seg = memseg_hash[MEMSEG_PFN_HASH(pfnum)]) != NULL) &&
5806 	    (pfnum >= seg->pages_base) && (pfnum < seg->pages_end)) {
5807 		MEMSEG_STAT_INCR(nhashwon);
5808 		vc->vc_pnum_memseg = seg;
5809 		pp = seg->pages + (pfnum - seg->pages_base);
5810 		if (pp->p_pagenum == pfnum)
5811 			return ((page_t *)pp);
5812 	}
5813 
5814 	/* Else Brute force */
5815 	for (seg = memsegs; seg != NULL; seg = seg->next) {
5816 		if (pfnum >= seg->pages_base && pfnum < seg->pages_end) {
5817 			vc->vc_pnum_memseg = seg;
5818 			pp = seg->pages + (pfnum - seg->pages_base);
5819 			return ((page_t *)pp);
5820 		}
5821 	}
5822 	vc->vc_pnum_memseg = NULL;
5823 	MEMSEG_STAT_INCR(nnotfound);
5824 	return ((page_t *)NULL);
5825 
5826 }
5827 
5828 struct memseg *
5829 page_numtomemseg_nolock(pfn_t pfnum)
5830 {
5831 	struct memseg *seg;
5832 	page_t *pp;
5833 
5834 	/* Try hash */
5835 	if (((seg = memseg_hash[MEMSEG_PFN_HASH(pfnum)]) != NULL) &&
5836 	    (pfnum >= seg->pages_base) && (pfnum < seg->pages_end)) {
5837 		pp = seg->pages + (pfnum - seg->pages_base);
5838 		if (pp->p_pagenum == pfnum)
5839 			return (seg);
5840 	}
5841 
5842 	/* Else Brute force */
5843 	for (seg = memsegs; seg != NULL; seg = seg->next) {
5844 		if (pfnum >= seg->pages_base && pfnum < seg->pages_end) {
5845 			return (seg);
5846 		}
5847 	}
5848 	return ((struct memseg *)NULL);
5849 }
5850 
5851 /*
5852  * Given a page and a count return the page struct that is
5853  * n structs away from the current one in the global page
5854  * list.
5855  *
5856  * This function wraps to the first page upon
5857  * reaching the end of the memseg list.
5858  */
5859 page_t *
5860 page_nextn(page_t *pp, ulong_t n)
5861 {
5862 	struct memseg *seg;
5863 	page_t *ppn;
5864 	vm_cpu_data_t *vc = (vm_cpu_data_t *)CPU->cpu_vm_data;
5865 
5866 	ASSERT(vc != NULL);
5867 
5868 	if (((seg = vc->vc_pnext_memseg) == NULL) ||
5869 	    (seg->pages_base == seg->pages_end) ||
5870 	    !(pp >= seg->pages && pp < seg->epages)) {
5871 
5872 		for (seg = memsegs; seg; seg = seg->next) {
5873 			if (pp >= seg->pages && pp < seg->epages)
5874 				break;
5875 		}
5876 
5877 		if (seg == NULL) {
5878 			/* Memory delete got in, return something valid. */
5879 			/* TODO: fix me. */
5880 			seg = memsegs;
5881 			pp = seg->pages;
5882 		}
5883 	}
5884 
5885 	/* check for wraparound - possible if n is large */
5886 	while ((ppn = (pp + n)) >= seg->epages || ppn < pp) {
5887 		n -= seg->epages - pp;
5888 		seg = seg->next;
5889 		if (seg == NULL)
5890 			seg = memsegs;
5891 		pp = seg->pages;
5892 	}
5893 	vc->vc_pnext_memseg = seg;
5894 	return (ppn);
5895 }
5896 
5897 /*
5898  * Initialize for a loop using page_next_scan_large().
5899  */
5900 page_t *
5901 page_next_scan_init(void **cookie)
5902 {
5903 	ASSERT(cookie != NULL);
5904 	*cookie = (void *)memsegs;
5905 	return ((page_t *)memsegs->pages);
5906 }
5907 
5908 /*
5909  * Return the next page in a scan of page_t's, assuming we want
5910  * to skip over sub-pages within larger page sizes.
5911  *
5912  * The cookie is used to keep track of the current memseg.
5913  */
5914 page_t *
5915 page_next_scan_large(
5916 	page_t		*pp,
5917 	ulong_t		*n,
5918 	void		**cookie)
5919 {
5920 	struct memseg	*seg = (struct memseg *)*cookie;
5921 	page_t		*new_pp;
5922 	ulong_t		cnt;
5923 	pfn_t		pfn;
5924 
5925 
5926 	/*
5927 	 * get the count of page_t's to skip based on the page size
5928 	 */
5929 	ASSERT(pp != NULL);
5930 	if (pp->p_szc == 0) {
5931 		cnt = 1;
5932 	} else {
5933 		pfn = page_pptonum(pp);
5934 		cnt = page_get_pagecnt(pp->p_szc);
5935 		cnt -= pfn & (cnt - 1);
5936 	}
5937 	*n += cnt;
5938 	new_pp = pp + cnt;
5939 
5940 	/*
5941 	 * Catch if we went past the end of the current memory segment. If so,
5942 	 * just move to the next segment with pages.
5943 	 */
5944 	if (new_pp >= seg->epages) {
5945 		do {
5946 			seg = seg->next;
5947 			if (seg == NULL)
5948 				seg = memsegs;
5949 		} while (seg->pages == seg->epages);
5950 		new_pp = seg->pages;
5951 		*cookie = (void *)seg;
5952 	}
5953 
5954 	return (new_pp);
5955 }
5956 
5957 
5958 /*
5959  * Returns next page in list. Note: this function wraps
5960  * to the first page in the list upon reaching the end
5961  * of the list. Callers should be aware of this fact.
5962  */
5963 
5964 /* We should change this be a #define */
5965 
5966 page_t *
5967 page_next(page_t *pp)
5968 {
5969 	return (page_nextn(pp, 1));
5970 }
5971 
5972 page_t *
5973 page_first()
5974 {
5975 	return ((page_t *)memsegs->pages);
5976 }
5977 
5978 
5979 /*
5980  * This routine is called at boot with the initial memory configuration
5981  * and when memory is added or removed.
5982  */
5983 void
5984 build_pfn_hash()
5985 {
5986 	pfn_t cur;
5987 	pgcnt_t index;
5988 	struct memseg *pseg;
5989 	int	i;
5990 
5991 	/*
5992 	 * Clear memseg_hash array.
5993 	 * Since memory add/delete is designed to operate concurrently
5994 	 * with normal operation, the hash rebuild must be able to run
5995 	 * concurrently with page_numtopp_nolock(). To support this
5996 	 * functionality, assignments to memseg_hash array members must
5997 	 * be done atomically.
5998 	 *
5999 	 * NOTE: bzero() does not currently guarantee this for kernel
6000 	 * threads, and cannot be used here.
6001 	 */
6002 	for (i = 0; i < N_MEM_SLOTS; i++)
6003 		memseg_hash[i] = NULL;
6004 
6005 	hat_kpm_mseghash_clear(N_MEM_SLOTS);
6006 
6007 	/*
6008 	 * Physmax is the last valid pfn.
6009 	 */
6010 	mhash_per_slot = (physmax + 1) >> MEM_HASH_SHIFT;
6011 	for (pseg = memsegs; pseg != NULL; pseg = pseg->next) {
6012 		index = MEMSEG_PFN_HASH(pseg->pages_base);
6013 		cur = pseg->pages_base;
6014 		do {
6015 			if (index >= N_MEM_SLOTS)
6016 				index = MEMSEG_PFN_HASH(cur);
6017 
6018 			if (memseg_hash[index] == NULL ||
6019 			    memseg_hash[index]->pages_base > pseg->pages_base) {
6020 				memseg_hash[index] = pseg;
6021 				hat_kpm_mseghash_update(index, pseg);
6022 			}
6023 			cur += mhash_per_slot;
6024 			index++;
6025 		} while (cur < pseg->pages_end);
6026 	}
6027 }
6028 
6029 /*
6030  * Return the pagenum for the pp
6031  */
6032 pfn_t
6033 page_pptonum(page_t *pp)
6034 {
6035 	return (pp->p_pagenum);
6036 }
6037 
6038 /*
6039  * interface to the referenced and modified etc bits
6040  * in the PSM part of the page struct
6041  * when no locking is desired.
6042  */
6043 void
6044 page_set_props(page_t *pp, uint_t flags)
6045 {
6046 	ASSERT((flags & ~(P_MOD | P_REF | P_RO)) == 0);
6047 	pp->p_nrm |= (uchar_t)flags;
6048 }
6049 
6050 void
6051 page_clr_all_props(page_t *pp)
6052 {
6053 	pp->p_nrm = 0;
6054 }
6055 
6056 /*
6057  * Clear p_lckcnt and p_cowcnt, adjusting freemem if required.
6058  */
6059 int
6060 page_clear_lck_cow(page_t *pp, int adjust)
6061 {
6062 	int	f_amount;
6063 
6064 	ASSERT(PAGE_EXCL(pp));
6065 
6066 	/*
6067 	 * The page_struct_lock need not be acquired here since
6068 	 * we require the caller hold the page exclusively locked.
6069 	 */
6070 	f_amount = 0;
6071 	if (pp->p_lckcnt) {
6072 		f_amount = 1;
6073 		pp->p_lckcnt = 0;
6074 	}
6075 	if (pp->p_cowcnt) {
6076 		f_amount += pp->p_cowcnt;
6077 		pp->p_cowcnt = 0;
6078 	}
6079 
6080 	if (adjust && f_amount) {
6081 		mutex_enter(&freemem_lock);
6082 		availrmem += f_amount;
6083 		mutex_exit(&freemem_lock);
6084 	}
6085 
6086 	return (f_amount);
6087 }
6088 
6089 /*
6090  * The following functions is called from free_vp_pages()
6091  * for an inexact estimate of a newly free'd page...
6092  */
6093 ulong_t
6094 page_share_cnt(page_t *pp)
6095 {
6096 	return (hat_page_getshare(pp));
6097 }
6098 
6099 int
6100 page_isshared(page_t *pp)
6101 {
6102 	return (hat_page_checkshare(pp, 1));
6103 }
6104 
6105 int
6106 page_isfree(page_t *pp)
6107 {
6108 	return (PP_ISFREE(pp));
6109 }
6110 
6111 int
6112 page_isref(page_t *pp)
6113 {
6114 	return (hat_page_getattr(pp, P_REF));
6115 }
6116 
6117 int
6118 page_ismod(page_t *pp)
6119 {
6120 	return (hat_page_getattr(pp, P_MOD));
6121 }
6122 
6123 /*
6124  * The following code all currently relates to the page capture logic:
6125  *
6126  * This logic is used for cases where there is a desire to claim a certain
6127  * physical page in the system for the caller.  As it may not be possible
6128  * to capture the page immediately, the p_toxic bits are used in the page
6129  * structure to indicate that someone wants to capture this page.  When the
6130  * page gets unlocked, the toxic flag will be noted and an attempt to capture
6131  * the page will be made.  If it is successful, the original callers callback
6132  * will be called with the page to do with it what they please.
6133  *
6134  * There is also an async thread which wakes up to attempt to capture
6135  * pages occasionally which have the capture bit set.  All of the pages which
6136  * need to be captured asynchronously have been inserted into the
6137  * page_capture_hash and thus this thread walks that hash list.  Items in the
6138  * hash have an expiration time so this thread handles that as well by removing
6139  * the item from the hash if it has expired.
6140  *
6141  * Some important things to note are:
6142  * - if the PR_CAPTURE bit is set on a page, then the page is in the
6143  *   page_capture_hash.  The page_capture_hash_head.pchh_mutex is needed
6144  *   to set and clear this bit, and while the lock is held is the only time
6145  *   you can add or remove an entry from the hash.
6146  * - the PR_CAPTURE bit can only be set and cleared while holding the
6147  *   page_capture_hash_head.pchh_mutex
6148  * - the t_flag field of the thread struct is used with the T_CAPTURING
6149  *   flag to prevent recursion while dealing with large pages.
6150  * - pages which need to be retired never expire on the page_capture_hash.
6151  */
6152 
6153 static void page_capture_thread(void);
6154 static kthread_t *pc_thread_id;
6155 kcondvar_t pc_cv;
6156 static kmutex_t pc_thread_mutex;
6157 static clock_t pc_thread_shortwait;
6158 static clock_t pc_thread_longwait;
6159 static int pc_thread_retry;
6160 
6161 struct page_capture_callback pc_cb[PC_NUM_CALLBACKS];
6162 
6163 /* Note that this is a circular linked list */
6164 typedef struct page_capture_hash_bucket {
6165 	page_t *pp;
6166 	uint_t szc;
6167 	uint_t flags;
6168 	clock_t expires;	/* lbolt at which this request expires. */
6169 	void *datap;		/* Cached data passed in for callback */
6170 	struct page_capture_hash_bucket *next;
6171 	struct page_capture_hash_bucket *prev;
6172 } page_capture_hash_bucket_t;
6173 
6174 /*
6175  * Each hash bucket will have it's own mutex and two lists which are:
6176  * active (0):	represents requests which have not been processed by
6177  *		the page_capture async thread yet.
6178  * walked (1):	represents requests which have been processed by the
6179  *		page_capture async thread within it's given walk of this bucket.
6180  *
6181  * These are all needed so that we can synchronize all async page_capture
6182  * events.  When the async thread moves to a new bucket, it will append the
6183  * walked list to the active list and walk each item one at a time, moving it
6184  * from the active list to the walked list.  Thus if there is an async request
6185  * outstanding for a given page, it will always be in one of the two lists.
6186  * New requests will always be added to the active list.
6187  * If we were not able to capture a page before the request expired, we'd free
6188  * up the request structure which would indicate to page_capture that there is
6189  * no longer a need for the given page, and clear the PR_CAPTURE flag if
6190  * possible.
6191  */
6192 typedef struct page_capture_hash_head {
6193 	kmutex_t pchh_mutex;
6194 	uint_t num_pages;
6195 	page_capture_hash_bucket_t lists[2]; /* sentinel nodes */
6196 } page_capture_hash_head_t;
6197 
6198 #ifdef DEBUG
6199 #define	NUM_PAGE_CAPTURE_BUCKETS 4
6200 #else
6201 #define	NUM_PAGE_CAPTURE_BUCKETS 64
6202 #endif
6203 
6204 page_capture_hash_head_t page_capture_hash[NUM_PAGE_CAPTURE_BUCKETS];
6205 
6206 /* for now use a very simple hash based upon the size of a page struct */
6207 #define	PAGE_CAPTURE_HASH(pp)	\
6208 	((int)(((uintptr_t)pp >> 7) & (NUM_PAGE_CAPTURE_BUCKETS - 1)))
6209 
6210 extern pgcnt_t swapfs_minfree;
6211 
6212 int page_trycapture(page_t *pp, uint_t szc, uint_t flags, void *datap);
6213 
6214 /*
6215  * a callback function is required for page capture requests.
6216  */
6217 void
6218 page_capture_register_callback(uint_t index, clock_t duration,
6219     int (*cb_func)(page_t *, void *, uint_t))
6220 {
6221 	ASSERT(pc_cb[index].cb_active == 0);
6222 	ASSERT(cb_func != NULL);
6223 	rw_enter(&pc_cb[index].cb_rwlock, RW_WRITER);
6224 	pc_cb[index].duration = duration;
6225 	pc_cb[index].cb_func = cb_func;
6226 	pc_cb[index].cb_active = 1;
6227 	rw_exit(&pc_cb[index].cb_rwlock);
6228 }
6229 
6230 void
6231 page_capture_unregister_callback(uint_t index)
6232 {
6233 	int i, j;
6234 	struct page_capture_hash_bucket *bp1;
6235 	struct page_capture_hash_bucket *bp2;
6236 	struct page_capture_hash_bucket *head = NULL;
6237 	uint_t flags = (1 << index);
6238 
6239 	rw_enter(&pc_cb[index].cb_rwlock, RW_WRITER);
6240 	ASSERT(pc_cb[index].cb_active == 1);
6241 	pc_cb[index].duration = 0;	/* Paranoia */
6242 	pc_cb[index].cb_func = NULL;	/* Paranoia */
6243 	pc_cb[index].cb_active = 0;
6244 	rw_exit(&pc_cb[index].cb_rwlock);
6245 
6246 	/*
6247 	 * Just move all the entries to a private list which we can walk
6248 	 * through without the need to hold any locks.
6249 	 * No more requests can get added to the hash lists for this consumer
6250 	 * as the cb_active field for the callback has been cleared.
6251 	 */
6252 	for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) {
6253 		mutex_enter(&page_capture_hash[i].pchh_mutex);
6254 		for (j = 0; j < 2; j++) {
6255 			bp1 = page_capture_hash[i].lists[j].next;
6256 			/* walk through all but first (sentinel) element */
6257 			while (bp1 != &page_capture_hash[i].lists[j]) {
6258 				bp2 = bp1;
6259 				if (bp2->flags & flags) {
6260 					bp1 = bp2->next;
6261 					bp1->prev = bp2->prev;
6262 					bp2->prev->next = bp1;
6263 					bp2->next = head;
6264 					head = bp2;
6265 					/*
6266 					 * Clear the PR_CAPTURE bit as we
6267 					 * hold appropriate locks here.
6268 					 */
6269 					page_clrtoxic(head->pp, PR_CAPTURE);
6270 					page_capture_hash[i].num_pages--;
6271 					continue;
6272 				}
6273 				bp1 = bp1->next;
6274 			}
6275 		}
6276 		mutex_exit(&page_capture_hash[i].pchh_mutex);
6277 	}
6278 
6279 	while (head != NULL) {
6280 		bp1 = head;
6281 		head = head->next;
6282 		kmem_free(bp1, sizeof (*bp1));
6283 	}
6284 }
6285 
6286 
6287 /*
6288  * Find pp in the active list and move it to the walked list if it
6289  * exists.
6290  * Note that most often pp should be at the front of the active list
6291  * as it is currently used and thus there is no other sort of optimization
6292  * being done here as this is a linked list data structure.
6293  * Returns 1 on successful move or 0 if page could not be found.
6294  */
6295 static int
6296 page_capture_move_to_walked(page_t *pp)
6297 {
6298 	page_capture_hash_bucket_t *bp;
6299 	int index;
6300 
6301 	index = PAGE_CAPTURE_HASH(pp);
6302 
6303 	mutex_enter(&page_capture_hash[index].pchh_mutex);
6304 	bp = page_capture_hash[index].lists[0].next;
6305 	while (bp != &page_capture_hash[index].lists[0]) {
6306 		if (bp->pp == pp) {
6307 			/* Remove from old list */
6308 			bp->next->prev = bp->prev;
6309 			bp->prev->next = bp->next;
6310 
6311 			/* Add to new list */
6312 			bp->next = page_capture_hash[index].lists[1].next;
6313 			bp->prev = &page_capture_hash[index].lists[1];
6314 			page_capture_hash[index].lists[1].next = bp;
6315 			bp->next->prev = bp;
6316 			mutex_exit(&page_capture_hash[index].pchh_mutex);
6317 
6318 			return (1);
6319 		}
6320 		bp = bp->next;
6321 	}
6322 	mutex_exit(&page_capture_hash[index].pchh_mutex);
6323 	return (0);
6324 }
6325 
6326 /*
6327  * Add a new entry to the page capture hash.  The only case where a new
6328  * entry is not added is when the page capture consumer is no longer registered.
6329  * In this case, we'll silently not add the page to the hash.  We know that
6330  * page retire will always be registered for the case where we are currently
6331  * unretiring a page and thus there are no conflicts.
6332  */
6333 static void
6334 page_capture_add_hash(page_t *pp, uint_t szc, uint_t flags, void *datap)
6335 {
6336 	page_capture_hash_bucket_t *bp1;
6337 	page_capture_hash_bucket_t *bp2;
6338 	int index;
6339 	int cb_index;
6340 	int i;
6341 #ifdef DEBUG
6342 	page_capture_hash_bucket_t *tp1;
6343 	int l;
6344 #endif
6345 
6346 	ASSERT(!(flags & CAPTURE_ASYNC));
6347 
6348 	bp1 = kmem_alloc(sizeof (struct page_capture_hash_bucket), KM_SLEEP);
6349 
6350 	bp1->pp = pp;
6351 	bp1->szc = szc;
6352 	bp1->flags = flags;
6353 	bp1->datap = datap;
6354 
6355 	for (cb_index = 0; cb_index < PC_NUM_CALLBACKS; cb_index++) {
6356 		if ((flags >> cb_index) & 1) {
6357 			break;
6358 		}
6359 	}
6360 
6361 	ASSERT(cb_index != PC_NUM_CALLBACKS);
6362 
6363 	rw_enter(&pc_cb[cb_index].cb_rwlock, RW_READER);
6364 	if (pc_cb[cb_index].cb_active) {
6365 		if (pc_cb[cb_index].duration == -1) {
6366 			bp1->expires = (clock_t)-1;
6367 		} else {
6368 			bp1->expires = lbolt + pc_cb[cb_index].duration;
6369 		}
6370 	} else {
6371 		/* There's no callback registered so don't add to the hash */
6372 		rw_exit(&pc_cb[cb_index].cb_rwlock);
6373 		kmem_free(bp1, sizeof (*bp1));
6374 		return;
6375 	}
6376 
6377 	index = PAGE_CAPTURE_HASH(pp);
6378 
6379 	/*
6380 	 * Only allow capture flag to be modified under this mutex.
6381 	 * Prevents multiple entries for same page getting added.
6382 	 */
6383 	mutex_enter(&page_capture_hash[index].pchh_mutex);
6384 
6385 	/*
6386 	 * if not already on the hash, set capture bit and add to the hash
6387 	 */
6388 	if (!(pp->p_toxic & PR_CAPTURE)) {
6389 #ifdef DEBUG
6390 		/* Check for duplicate entries */
6391 		for (l = 0; l < 2; l++) {
6392 			tp1 = page_capture_hash[index].lists[l].next;
6393 			while (tp1 != &page_capture_hash[index].lists[l]) {
6394 				if (tp1->pp == pp) {
6395 					panic("page pp 0x%p already on hash "
6396 					    "at 0x%p\n", pp, tp1);
6397 				}
6398 				tp1 = tp1->next;
6399 			}
6400 		}
6401 
6402 #endif
6403 		page_settoxic(pp, PR_CAPTURE);
6404 		bp1->next = page_capture_hash[index].lists[0].next;
6405 		bp1->prev = &page_capture_hash[index].lists[0];
6406 		bp1->next->prev = bp1;
6407 		page_capture_hash[index].lists[0].next = bp1;
6408 		page_capture_hash[index].num_pages++;
6409 		if (flags & CAPTURE_RETIRE) {
6410 			page_retire_incr_pend_count();
6411 		}
6412 		mutex_exit(&page_capture_hash[index].pchh_mutex);
6413 		rw_exit(&pc_cb[cb_index].cb_rwlock);
6414 		cv_signal(&pc_cv);
6415 		return;
6416 	}
6417 
6418 	/*
6419 	 * A page retire request will replace any other request.
6420 	 * A second physmem request which is for a different process than
6421 	 * the currently registered one will be dropped as there is
6422 	 * no way to hold the private data for both calls.
6423 	 * In the future, once there are more callers, this will have to
6424 	 * be worked out better as there needs to be private storage for
6425 	 * at least each type of caller (maybe have datap be an array of
6426 	 * *void's so that we can index based upon callers index).
6427 	 */
6428 
6429 	/* walk hash list to update expire time */
6430 	for (i = 0; i < 2; i++) {
6431 		bp2 = page_capture_hash[index].lists[i].next;
6432 		while (bp2 != &page_capture_hash[index].lists[i]) {
6433 			if (bp2->pp == pp) {
6434 				if (flags & CAPTURE_RETIRE) {
6435 					if (!(bp2->flags & CAPTURE_RETIRE)) {
6436 						page_retire_incr_pend_count();
6437 						bp2->flags = flags;
6438 						bp2->expires = bp1->expires;
6439 						bp2->datap = datap;
6440 					}
6441 				} else {
6442 					ASSERT(flags & CAPTURE_PHYSMEM);
6443 					if (!(bp2->flags & CAPTURE_RETIRE) &&
6444 					    (datap == bp2->datap)) {
6445 						bp2->expires = bp1->expires;
6446 					}
6447 				}
6448 				mutex_exit(&page_capture_hash[index].
6449 				    pchh_mutex);
6450 				rw_exit(&pc_cb[cb_index].cb_rwlock);
6451 				kmem_free(bp1, sizeof (*bp1));
6452 				return;
6453 			}
6454 			bp2 = bp2->next;
6455 		}
6456 	}
6457 
6458 	/*
6459 	 * the PR_CAPTURE flag is protected by the page_capture_hash mutexes
6460 	 * and thus it either has to be set or not set and can't change
6461 	 * while holding the mutex above.
6462 	 */
6463 	panic("page_capture_add_hash, PR_CAPTURE flag set on pp %p\n", pp);
6464 }
6465 
6466 /*
6467  * We have a page in our hands, lets try and make it ours by turning
6468  * it into a clean page like it had just come off the freelists.
6469  *
6470  * Returns 0 on success, with the page still EXCL locked.
6471  * On failure, the page will be unlocked, and returns EAGAIN
6472  */
6473 static int
6474 page_capture_clean_page(page_t *pp)
6475 {
6476 	page_t *newpp;
6477 	int skip_unlock = 0;
6478 	spgcnt_t count;
6479 	page_t *tpp;
6480 	int ret = 0;
6481 	int extra;
6482 
6483 	ASSERT(PAGE_EXCL(pp));
6484 	ASSERT(!PP_RETIRED(pp));
6485 	ASSERT(curthread->t_flag & T_CAPTURING);
6486 
6487 	if (PP_ISFREE(pp)) {
6488 		if (!page_reclaim(pp, NULL)) {
6489 			skip_unlock = 1;
6490 			ret = EAGAIN;
6491 			goto cleanup;
6492 		}
6493 		ASSERT(pp->p_szc == 0);
6494 		if (pp->p_vnode != NULL) {
6495 			/*
6496 			 * Since this page came from the
6497 			 * cachelist, we must destroy the
6498 			 * old vnode association.
6499 			 */
6500 			page_hashout(pp, NULL);
6501 		}
6502 		goto cleanup;
6503 	}
6504 
6505 	/*
6506 	 * If we know page_relocate will fail, skip it
6507 	 * It could still fail due to a UE on another page but we
6508 	 * can't do anything about that.
6509 	 */
6510 	if (pp->p_toxic & PR_UE) {
6511 		goto skip_relocate;
6512 	}
6513 
6514 	/*
6515 	 * It's possible that pages can not have a vnode as fsflush comes
6516 	 * through and cleans up these pages.  It's ugly but that's how it is.
6517 	 */
6518 	if (pp->p_vnode == NULL) {
6519 		goto skip_relocate;
6520 	}
6521 
6522 	/*
6523 	 * Page was not free, so lets try to relocate it.
6524 	 * page_relocate only works with root pages, so if this is not a root
6525 	 * page, we need to demote it to try and relocate it.
6526 	 * Unfortunately this is the best we can do right now.
6527 	 */
6528 	newpp = NULL;
6529 	if ((pp->p_szc > 0) && (pp != PP_PAGEROOT(pp))) {
6530 		if (page_try_demote_pages(pp) == 0) {
6531 			ret = EAGAIN;
6532 			goto cleanup;
6533 		}
6534 	}
6535 	ret = page_relocate(&pp, &newpp, 1, 0, &count, NULL);
6536 	if (ret == 0) {
6537 		page_t *npp;
6538 		/* unlock the new page(s) */
6539 		while (count-- > 0) {
6540 			ASSERT(newpp != NULL);
6541 			npp = newpp;
6542 			page_sub(&newpp, npp);
6543 			page_unlock(npp);
6544 		}
6545 		ASSERT(newpp == NULL);
6546 		/*
6547 		 * Check to see if the page we have is too large.
6548 		 * If so, demote it freeing up the extra pages.
6549 		 */
6550 		if (pp->p_szc > 0) {
6551 			/* For now demote extra pages to szc == 0 */
6552 			extra = page_get_pagecnt(pp->p_szc) - 1;
6553 			while (extra > 0) {
6554 				tpp = pp->p_next;
6555 				page_sub(&pp, tpp);
6556 				tpp->p_szc = 0;
6557 				page_free(tpp, 1);
6558 				extra--;
6559 			}
6560 			/* Make sure to set our page to szc 0 as well */
6561 			ASSERT(pp->p_next == pp && pp->p_prev == pp);
6562 			pp->p_szc = 0;
6563 		}
6564 		goto cleanup;
6565 	} else if (ret == EIO) {
6566 		ret = EAGAIN;
6567 		goto cleanup;
6568 	} else {
6569 		/*
6570 		 * Need to reset return type as we failed to relocate the page
6571 		 * but that does not mean that some of the next steps will not
6572 		 * work.
6573 		 */
6574 		ret = 0;
6575 	}
6576 
6577 skip_relocate:
6578 
6579 	if (pp->p_szc > 0) {
6580 		if (page_try_demote_pages(pp) == 0) {
6581 			ret = EAGAIN;
6582 			goto cleanup;
6583 		}
6584 	}
6585 
6586 	ASSERT(pp->p_szc == 0);
6587 
6588 	if (hat_ismod(pp)) {
6589 		ret = EAGAIN;
6590 		goto cleanup;
6591 	}
6592 	if (PP_ISKAS(pp)) {
6593 		ret = EAGAIN;
6594 		goto cleanup;
6595 	}
6596 	if (pp->p_lckcnt || pp->p_cowcnt) {
6597 		ret = EAGAIN;
6598 		goto cleanup;
6599 	}
6600 
6601 	(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
6602 	ASSERT(!hat_page_is_mapped(pp));
6603 
6604 	if (hat_ismod(pp)) {
6605 		/*
6606 		 * This is a semi-odd case as the page is now modified but not
6607 		 * mapped as we just unloaded the mappings above.
6608 		 */
6609 		ret = EAGAIN;
6610 		goto cleanup;
6611 	}
6612 	if (pp->p_vnode != NULL) {
6613 		page_hashout(pp, NULL);
6614 	}
6615 
6616 	/*
6617 	 * At this point, the page should be in a clean state and
6618 	 * we can do whatever we want with it.
6619 	 */
6620 
6621 cleanup:
6622 	if (ret != 0) {
6623 		if (!skip_unlock) {
6624 			page_unlock(pp);
6625 		}
6626 	} else {
6627 		ASSERT(pp->p_szc == 0);
6628 		ASSERT(PAGE_EXCL(pp));
6629 
6630 		pp->p_next = pp;
6631 		pp->p_prev = pp;
6632 	}
6633 	return (ret);
6634 }
6635 
6636 /*
6637  * Various callers of page_trycapture() can have different restrictions upon
6638  * what memory they have access to.
6639  * Returns 0 on success, with the following error codes on failure:
6640  *      EPERM - The requested page is long term locked, and thus repeated
6641  *              requests to capture this page will likely fail.
6642  *      ENOMEM - There was not enough free memory in the system to safely
6643  *              map the requested page.
6644  *      ENOENT - The requested page was inside the kernel cage, and the
6645  *              PHYSMEM_CAGE flag was not set.
6646  */
6647 int
6648 page_capture_pre_checks(page_t *pp, uint_t flags)
6649 {
6650 #if defined(__sparc)
6651 	extern struct vnode prom_ppages;
6652 #endif /* __sparc */
6653 
6654 	ASSERT(pp != NULL);
6655 
6656 #if defined(__sparc)
6657 	if (pp->p_vnode == &prom_ppages) {
6658 		return (EPERM);
6659 	}
6660 
6661 	if (PP_ISNORELOC(pp) && !(flags & CAPTURE_GET_CAGE) &&
6662 	    (flags & CAPTURE_PHYSMEM)) {
6663 		return (ENOENT);
6664 	}
6665 
6666 	if (PP_ISNORELOCKERNEL(pp)) {
6667 		return (EPERM);
6668 	}
6669 #else
6670 	if (PP_ISKAS(pp)) {
6671 		return (EPERM);
6672 	}
6673 #endif /* __sparc */
6674 
6675 	/* only physmem currently has the restrictions checked below */
6676 	if (!(flags & CAPTURE_PHYSMEM)) {
6677 		return (0);
6678 	}
6679 
6680 	if (availrmem < swapfs_minfree) {
6681 		/*
6682 		 * We won't try to capture this page as we are
6683 		 * running low on memory.
6684 		 */
6685 		return (ENOMEM);
6686 	}
6687 	return (0);
6688 }
6689 
6690 /*
6691  * Once we have a page in our mits, go ahead and complete the capture
6692  * operation.
6693  * Returns 1 on failure where page is no longer needed
6694  * Returns 0 on success
6695  * Returns -1 if there was a transient failure.
6696  * Failure cases must release the SE_EXCL lock on pp (usually via page_free).
6697  */
6698 int
6699 page_capture_take_action(page_t *pp, uint_t flags, void *datap)
6700 {
6701 	int cb_index;
6702 	int ret = 0;
6703 	page_capture_hash_bucket_t *bp1;
6704 	page_capture_hash_bucket_t *bp2;
6705 	int index;
6706 	int found = 0;
6707 	int i;
6708 
6709 	ASSERT(PAGE_EXCL(pp));
6710 	ASSERT(curthread->t_flag & T_CAPTURING);
6711 
6712 	for (cb_index = 0; cb_index < PC_NUM_CALLBACKS; cb_index++) {
6713 		if ((flags >> cb_index) & 1) {
6714 			break;
6715 		}
6716 	}
6717 	ASSERT(cb_index < PC_NUM_CALLBACKS);
6718 
6719 	/*
6720 	 * Remove the entry from the page_capture hash, but don't free it yet
6721 	 * as we may need to put it back.
6722 	 * Since we own the page at this point in time, we should find it
6723 	 * in the hash if this is an ASYNC call.  If we don't it's likely
6724 	 * that the page_capture_async() thread decided that this request
6725 	 * had expired, in which case we just continue on.
6726 	 */
6727 	if (flags & CAPTURE_ASYNC) {
6728 
6729 		index = PAGE_CAPTURE_HASH(pp);
6730 
6731 		mutex_enter(&page_capture_hash[index].pchh_mutex);
6732 		for (i = 0; i < 2 && !found; i++) {
6733 			bp1 = page_capture_hash[index].lists[i].next;
6734 			while (bp1 != &page_capture_hash[index].lists[i]) {
6735 				if (bp1->pp == pp) {
6736 					bp1->next->prev = bp1->prev;
6737 					bp1->prev->next = bp1->next;
6738 					page_capture_hash[index].num_pages--;
6739 					page_clrtoxic(pp, PR_CAPTURE);
6740 					found = 1;
6741 					break;
6742 				}
6743 				bp1 = bp1->next;
6744 			}
6745 		}
6746 		mutex_exit(&page_capture_hash[index].pchh_mutex);
6747 	}
6748 
6749 	/* Synchronize with the unregister func. */
6750 	rw_enter(&pc_cb[cb_index].cb_rwlock, RW_READER);
6751 	if (!pc_cb[cb_index].cb_active) {
6752 		page_free(pp, 1);
6753 		rw_exit(&pc_cb[cb_index].cb_rwlock);
6754 		if (found) {
6755 			kmem_free(bp1, sizeof (*bp1));
6756 		}
6757 		return (1);
6758 	}
6759 
6760 	/*
6761 	 * We need to remove the entry from the page capture hash and turn off
6762 	 * the PR_CAPTURE bit before calling the callback.  We'll need to cache
6763 	 * the entry here, and then based upon the return value, cleanup
6764 	 * appropriately or re-add it to the hash, making sure that someone else
6765 	 * hasn't already done so.
6766 	 * It should be rare for the callback to fail and thus it's ok for
6767 	 * the failure path to be a bit complicated as the success path is
6768 	 * cleaner and the locking rules are easier to follow.
6769 	 */
6770 
6771 	ret = pc_cb[cb_index].cb_func(pp, datap, flags);
6772 
6773 	rw_exit(&pc_cb[cb_index].cb_rwlock);
6774 
6775 	/*
6776 	 * If this was an ASYNC request, we need to cleanup the hash if the
6777 	 * callback was successful or if the request was no longer valid.
6778 	 * For non-ASYNC requests, we return failure to map and the caller
6779 	 * will take care of adding the request to the hash.
6780 	 * Note also that the callback itself is responsible for the page
6781 	 * at this point in time in terms of locking ...  The most common
6782 	 * case for the failure path should just be a page_free.
6783 	 */
6784 	if (ret >= 0) {
6785 		if (found) {
6786 			if (bp1->flags & CAPTURE_RETIRE) {
6787 				page_retire_decr_pend_count();
6788 			}
6789 			kmem_free(bp1, sizeof (*bp1));
6790 		}
6791 		return (ret);
6792 	}
6793 	if (!found) {
6794 		return (ret);
6795 	}
6796 
6797 	ASSERT(flags & CAPTURE_ASYNC);
6798 
6799 	/*
6800 	 * Check for expiration time first as we can just free it up if it's
6801 	 * expired.
6802 	 */
6803 	if (lbolt > bp1->expires && bp1->expires != -1) {
6804 		kmem_free(bp1, sizeof (*bp1));
6805 		return (ret);
6806 	}
6807 
6808 	/*
6809 	 * The callback failed and there used to be an entry in the hash for
6810 	 * this page, so we need to add it back to the hash.
6811 	 */
6812 	mutex_enter(&page_capture_hash[index].pchh_mutex);
6813 	if (!(pp->p_toxic & PR_CAPTURE)) {
6814 		/* just add bp1 back to head of walked list */
6815 		page_settoxic(pp, PR_CAPTURE);
6816 		bp1->next = page_capture_hash[index].lists[1].next;
6817 		bp1->prev = &page_capture_hash[index].lists[1];
6818 		bp1->next->prev = bp1;
6819 		page_capture_hash[index].lists[1].next = bp1;
6820 		page_capture_hash[index].num_pages++;
6821 		mutex_exit(&page_capture_hash[index].pchh_mutex);
6822 		return (ret);
6823 	}
6824 
6825 	/*
6826 	 * Otherwise there was a new capture request added to list
6827 	 * Need to make sure that our original data is represented if
6828 	 * appropriate.
6829 	 */
6830 	for (i = 0; i < 2; i++) {
6831 		bp2 = page_capture_hash[index].lists[i].next;
6832 		while (bp2 != &page_capture_hash[index].lists[i]) {
6833 			if (bp2->pp == pp) {
6834 				if (bp1->flags & CAPTURE_RETIRE) {
6835 					if (!(bp2->flags & CAPTURE_RETIRE)) {
6836 						bp2->szc = bp1->szc;
6837 						bp2->flags = bp1->flags;
6838 						bp2->expires = bp1->expires;
6839 						bp2->datap = bp1->datap;
6840 					}
6841 				} else {
6842 					ASSERT(bp1->flags & CAPTURE_PHYSMEM);
6843 					if (!(bp2->flags & CAPTURE_RETIRE)) {
6844 						bp2->szc = bp1->szc;
6845 						bp2->flags = bp1->flags;
6846 						bp2->expires = bp1->expires;
6847 						bp2->datap = bp1->datap;
6848 					}
6849 				}
6850 				mutex_exit(&page_capture_hash[index].
6851 				    pchh_mutex);
6852 				kmem_free(bp1, sizeof (*bp1));
6853 				return (ret);
6854 			}
6855 			bp2 = bp2->next;
6856 		}
6857 	}
6858 	panic("PR_CAPTURE set but not on hash for pp 0x%p\n", pp);
6859 	/*NOTREACHED*/
6860 }
6861 
6862 /*
6863  * Try to capture the given page for the caller specified in the flags
6864  * parameter.  The page will either be captured and handed over to the
6865  * appropriate callback, or will be queued up in the page capture hash
6866  * to be captured asynchronously.
6867  * If the current request is due to an async capture, the page must be
6868  * exclusively locked before calling this function.
6869  * Currently szc must be 0 but in the future this should be expandable to
6870  * other page sizes.
6871  * Returns 0 on success, with the following error codes on failure:
6872  *      EPERM - The requested page is long term locked, and thus repeated
6873  *              requests to capture this page will likely fail.
6874  *      ENOMEM - There was not enough free memory in the system to safely
6875  *              map the requested page.
6876  *      ENOENT - The requested page was inside the kernel cage, and the
6877  *              CAPTURE_GET_CAGE flag was not set.
6878  *	EAGAIN - The requested page could not be capturead at this point in
6879  *		time but future requests will likely work.
6880  *	EBUSY - The requested page is retired and the CAPTURE_GET_RETIRED flag
6881  *		was not set.
6882  */
6883 int
6884 page_itrycapture(page_t *pp, uint_t szc, uint_t flags, void *datap)
6885 {
6886 	int ret;
6887 	int cb_index;
6888 
6889 	if (flags & CAPTURE_ASYNC) {
6890 		ASSERT(PAGE_EXCL(pp));
6891 		goto async;
6892 	}
6893 
6894 	/* Make sure there's enough availrmem ... */
6895 	ret = page_capture_pre_checks(pp, flags);
6896 	if (ret != 0) {
6897 		return (ret);
6898 	}
6899 
6900 	if (!page_trylock(pp, SE_EXCL)) {
6901 		for (cb_index = 0; cb_index < PC_NUM_CALLBACKS; cb_index++) {
6902 			if ((flags >> cb_index) & 1) {
6903 				break;
6904 			}
6905 		}
6906 		ASSERT(cb_index < PC_NUM_CALLBACKS);
6907 		ret = EAGAIN;
6908 		/* Special case for retired pages */
6909 		if (PP_RETIRED(pp)) {
6910 			if (flags & CAPTURE_GET_RETIRED) {
6911 				if (!page_unretire_pp(pp, PR_UNR_TEMP)) {
6912 					/*
6913 					 * Need to set capture bit and add to
6914 					 * hash so that the page will be
6915 					 * retired when freed.
6916 					 */
6917 					page_capture_add_hash(pp, szc,
6918 					    CAPTURE_RETIRE, NULL);
6919 					ret = 0;
6920 					goto own_page;
6921 				}
6922 			} else {
6923 				return (EBUSY);
6924 			}
6925 		}
6926 		page_capture_add_hash(pp, szc, flags, datap);
6927 		return (ret);
6928 	}
6929 
6930 async:
6931 	ASSERT(PAGE_EXCL(pp));
6932 
6933 	/* Need to check for physmem async requests that availrmem is sane */
6934 	if ((flags & (CAPTURE_ASYNC | CAPTURE_PHYSMEM)) ==
6935 	    (CAPTURE_ASYNC | CAPTURE_PHYSMEM) &&
6936 	    (availrmem < swapfs_minfree)) {
6937 		page_unlock(pp);
6938 		return (ENOMEM);
6939 	}
6940 
6941 	ret = page_capture_clean_page(pp);
6942 
6943 	if (ret != 0) {
6944 		/* We failed to get the page, so lets add it to the hash */
6945 		if (!(flags & CAPTURE_ASYNC)) {
6946 			page_capture_add_hash(pp, szc, flags, datap);
6947 		}
6948 		return (ret);
6949 	}
6950 
6951 own_page:
6952 	ASSERT(PAGE_EXCL(pp));
6953 	ASSERT(pp->p_szc == 0);
6954 
6955 	/* Call the callback */
6956 	ret = page_capture_take_action(pp, flags, datap);
6957 
6958 	if (ret == 0) {
6959 		return (0);
6960 	}
6961 
6962 	/*
6963 	 * Note that in the failure cases from page_capture_take_action, the
6964 	 * EXCL lock will have already been dropped.
6965 	 */
6966 	if ((ret == -1) && (!(flags & CAPTURE_ASYNC))) {
6967 		page_capture_add_hash(pp, szc, flags, datap);
6968 	}
6969 	return (EAGAIN);
6970 }
6971 
6972 int
6973 page_trycapture(page_t *pp, uint_t szc, uint_t flags, void *datap)
6974 {
6975 	int ret;
6976 
6977 	curthread->t_flag |= T_CAPTURING;
6978 	ret = page_itrycapture(pp, szc, flags, datap);
6979 	curthread->t_flag &= ~T_CAPTURING; /* xor works as we know its set */
6980 	return (ret);
6981 }
6982 
6983 /*
6984  * When unlocking a page which has the PR_CAPTURE bit set, this routine
6985  * gets called to try and capture the page.
6986  */
6987 void
6988 page_unlock_capture(page_t *pp)
6989 {
6990 	page_capture_hash_bucket_t *bp;
6991 	int index;
6992 	int i;
6993 	uint_t szc;
6994 	uint_t flags = 0;
6995 	void *datap;
6996 	kmutex_t *mp;
6997 	extern vnode_t retired_pages;
6998 
6999 	/*
7000 	 * We need to protect against a possible deadlock here where we own
7001 	 * the vnode page hash mutex and want to acquire it again as there
7002 	 * are locations in the code, where we unlock a page while holding
7003 	 * the mutex which can lead to the page being captured and eventually
7004 	 * end up here.  As we may be hashing out the old page and hashing into
7005 	 * the retire vnode, we need to make sure we don't own them.
7006 	 * Other callbacks who do hash operations also need to make sure that
7007 	 * before they hashin to a vnode that they do not currently own the
7008 	 * vphm mutex otherwise there will be a panic.
7009 	 */
7010 	if (mutex_owned(page_vnode_mutex(&retired_pages))) {
7011 		page_unlock_nocapture(pp);
7012 		return;
7013 	}
7014 	if (pp->p_vnode != NULL && mutex_owned(page_vnode_mutex(pp->p_vnode))) {
7015 		page_unlock_nocapture(pp);
7016 		return;
7017 	}
7018 
7019 	index = PAGE_CAPTURE_HASH(pp);
7020 
7021 	mp = &page_capture_hash[index].pchh_mutex;
7022 	mutex_enter(mp);
7023 	for (i = 0; i < 2; i++) {
7024 		bp = page_capture_hash[index].lists[i].next;
7025 		while (bp != &page_capture_hash[index].lists[i]) {
7026 			if (bp->pp == pp) {
7027 				szc = bp->szc;
7028 				flags = bp->flags | CAPTURE_ASYNC;
7029 				datap = bp->datap;
7030 				mutex_exit(mp);
7031 				(void) page_trycapture(pp, szc, flags, datap);
7032 				return;
7033 			}
7034 			bp = bp->next;
7035 		}
7036 	}
7037 
7038 	/* Failed to find page in hash so clear flags and unlock it. */
7039 	page_clrtoxic(pp, PR_CAPTURE);
7040 	page_unlock(pp);
7041 
7042 	mutex_exit(mp);
7043 }
7044 
7045 void
7046 page_capture_init()
7047 {
7048 	int i;
7049 	for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) {
7050 		page_capture_hash[i].lists[0].next =
7051 		    &page_capture_hash[i].lists[0];
7052 		page_capture_hash[i].lists[0].prev =
7053 		    &page_capture_hash[i].lists[0];
7054 		page_capture_hash[i].lists[1].next =
7055 		    &page_capture_hash[i].lists[1];
7056 		page_capture_hash[i].lists[1].prev =
7057 		    &page_capture_hash[i].lists[1];
7058 	}
7059 
7060 	pc_thread_shortwait = 23 * hz;
7061 	pc_thread_longwait = 1201 * hz;
7062 	pc_thread_retry = 3;
7063 	mutex_init(&pc_thread_mutex, NULL, MUTEX_DEFAULT, NULL);
7064 	cv_init(&pc_cv, NULL, CV_DEFAULT, NULL);
7065 	pc_thread_id = thread_create(NULL, 0, page_capture_thread, NULL, 0, &p0,
7066 	    TS_RUN, minclsyspri);
7067 }
7068 
7069 /*
7070  * It is necessary to scrub any failing pages prior to reboot in order to
7071  * prevent a latent error trap from occurring on the next boot.
7072  */
7073 void
7074 page_retire_mdboot()
7075 {
7076 	page_t *pp;
7077 	int i, j;
7078 	page_capture_hash_bucket_t *bp;
7079 
7080 	/* walk lists looking for pages to scrub */
7081 	for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) {
7082 		if (page_capture_hash[i].num_pages == 0)
7083 			continue;
7084 
7085 		mutex_enter(&page_capture_hash[i].pchh_mutex);
7086 
7087 		for (j = 0; j < 2; j++) {
7088 			bp = page_capture_hash[i].lists[j].next;
7089 			while (bp != &page_capture_hash[i].lists[j]) {
7090 				pp = bp->pp;
7091 				if (!PP_ISKAS(pp) && PP_TOXIC(pp)) {
7092 					pp->p_selock = -1;  /* pacify ASSERTs */
7093 					PP_CLRFREE(pp);
7094 					pagescrub(pp, 0, PAGESIZE);
7095 					pp->p_selock = 0;
7096 				}
7097 				bp = bp->next;
7098 			}
7099 		}
7100 		mutex_exit(&page_capture_hash[i].pchh_mutex);
7101 	}
7102 }
7103 
7104 /*
7105  * Walk the page_capture_hash trying to capture pages and also cleanup old
7106  * entries which have expired.
7107  */
7108 void
7109 page_capture_async()
7110 {
7111 	page_t *pp;
7112 	int i;
7113 	int ret;
7114 	page_capture_hash_bucket_t *bp1, *bp2;
7115 	uint_t szc;
7116 	uint_t flags;
7117 	void *datap;
7118 
7119 	/* If there are outstanding pages to be captured, get to work */
7120 	for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++) {
7121 		if (page_capture_hash[i].num_pages == 0)
7122 			continue;
7123 		/* Append list 1 to list 0 and then walk through list 0 */
7124 		mutex_enter(&page_capture_hash[i].pchh_mutex);
7125 		bp1 = &page_capture_hash[i].lists[1];
7126 		bp2 = bp1->next;
7127 		if (bp1 != bp2) {
7128 			bp1->prev->next = page_capture_hash[i].lists[0].next;
7129 			bp2->prev = &page_capture_hash[i].lists[0];
7130 			page_capture_hash[i].lists[0].next->prev = bp1->prev;
7131 			page_capture_hash[i].lists[0].next = bp2;
7132 			bp1->next = bp1;
7133 			bp1->prev = bp1;
7134 		}
7135 
7136 		/* list[1] will be empty now */
7137 
7138 		bp1 = page_capture_hash[i].lists[0].next;
7139 		while (bp1 != &page_capture_hash[i].lists[0]) {
7140 			/* Check expiration time */
7141 			if ((lbolt > bp1->expires && bp1->expires != -1) ||
7142 			    page_deleted(bp1->pp)) {
7143 				page_capture_hash[i].lists[0].next = bp1->next;
7144 				bp1->next->prev =
7145 				    &page_capture_hash[i].lists[0];
7146 				page_capture_hash[i].num_pages--;
7147 
7148 				/*
7149 				 * We can safely remove the PR_CAPTURE bit
7150 				 * without holding the EXCL lock on the page
7151 				 * as the PR_CAPTURE bit requres that the
7152 				 * page_capture_hash[].pchh_mutex be held
7153 				 * to modify it.
7154 				 */
7155 				page_clrtoxic(bp1->pp, PR_CAPTURE);
7156 				mutex_exit(&page_capture_hash[i].pchh_mutex);
7157 				kmem_free(bp1, sizeof (*bp1));
7158 				mutex_enter(&page_capture_hash[i].pchh_mutex);
7159 				bp1 = page_capture_hash[i].lists[0].next;
7160 				continue;
7161 			}
7162 			pp = bp1->pp;
7163 			szc = bp1->szc;
7164 			flags = bp1->flags;
7165 			datap = bp1->datap;
7166 			mutex_exit(&page_capture_hash[i].pchh_mutex);
7167 			if (page_trylock(pp, SE_EXCL)) {
7168 				ret = page_trycapture(pp, szc,
7169 				    flags | CAPTURE_ASYNC, datap);
7170 			} else {
7171 				ret = 1;	/* move to walked hash */
7172 			}
7173 
7174 			if (ret != 0) {
7175 				/* Move to walked hash */
7176 				(void) page_capture_move_to_walked(pp);
7177 			}
7178 			mutex_enter(&page_capture_hash[i].pchh_mutex);
7179 			bp1 = page_capture_hash[i].lists[0].next;
7180 		}
7181 
7182 		mutex_exit(&page_capture_hash[i].pchh_mutex);
7183 	}
7184 }
7185 
7186 /*
7187  * This function is called by the page_capture_thread, and is needed in
7188  * in order to initiate aio cleanup, so that pages used in aio
7189  * will be unlocked and subsequently retired by page_capture_thread.
7190  */
7191 static int
7192 do_aio_cleanup(void)
7193 {
7194 	proc_t *procp;
7195 	int (*aio_cleanup_dr_delete_memory)(proc_t *);
7196 	int cleaned = 0;
7197 
7198 	if (modload("sys", "kaio") == -1) {
7199 		cmn_err(CE_WARN, "do_aio_cleanup: cannot load kaio");
7200 		return (0);
7201 	}
7202 	/*
7203 	 * We use the aio_cleanup_dr_delete_memory function to
7204 	 * initiate the actual clean up; this function will wake
7205 	 * up the per-process aio_cleanup_thread.
7206 	 */
7207 	aio_cleanup_dr_delete_memory = (int (*)(proc_t *))
7208 	    modgetsymvalue("aio_cleanup_dr_delete_memory", 0);
7209 	if (aio_cleanup_dr_delete_memory == NULL) {
7210 		cmn_err(CE_WARN,
7211 	    "aio_cleanup_dr_delete_memory not found in kaio");
7212 		return (0);
7213 	}
7214 	mutex_enter(&pidlock);
7215 	for (procp = practive; (procp != NULL); procp = procp->p_next) {
7216 		mutex_enter(&procp->p_lock);
7217 		if (procp->p_aio != NULL) {
7218 			/* cleanup proc's outstanding kaio */
7219 			cleaned += (*aio_cleanup_dr_delete_memory)(procp);
7220 		}
7221 		mutex_exit(&procp->p_lock);
7222 	}
7223 	mutex_exit(&pidlock);
7224 	return (cleaned);
7225 }
7226 
7227 /*
7228  * helper function for page_capture_thread
7229  */
7230 static void
7231 page_capture_handle_outstanding(void)
7232 {
7233 	int ntry;
7234 
7235 	if (!page_retire_pend_count()) {
7236 		/*
7237 		 * Do we really want to be this aggressive
7238 		 * for things other than page_retire?
7239 		 * Maybe have a counter for each callback
7240 		 * type to guide how aggressive we should
7241 		 * be here.  Thus if there's at least one
7242 		 * page for page_retire we go ahead and reap
7243 		 * like this.
7244 		 */
7245 		kmem_reap();
7246 		seg_preap();
7247 		page_capture_async();
7248 	} else {
7249 		/*
7250 		 * There are pages pending retirement, so
7251 		 * we reap prior to attempting to capture.
7252 		 */
7253 		kmem_reap();
7254 
7255 		/* disable and purge seg_pcache */
7256 		(void) seg_p_disable();
7257 		for (ntry = 0; ntry < pc_thread_retry; ntry++) {
7258 			if (!page_retire_pend_count())
7259 				break;
7260 			if (do_aio_cleanup()) {
7261 				/*
7262 				 * allow the apps cleanup threads
7263 				 * to run
7264 				 */
7265 				delay(pc_thread_shortwait);
7266 			}
7267 			page_capture_async();
7268 		}
7269 		/* reenable seg_pcache */
7270 		seg_p_enable();
7271 	}
7272 }
7273 
7274 /*
7275  * The page_capture_thread loops forever, looking to see if there are
7276  * pages still waiting to be captured.
7277  */
7278 static void
7279 page_capture_thread(void)
7280 {
7281 	callb_cpr_t c;
7282 	int outstanding;
7283 	int i;
7284 
7285 	CALLB_CPR_INIT(&c, &pc_thread_mutex, callb_generic_cpr, "page_capture");
7286 
7287 	mutex_enter(&pc_thread_mutex);
7288 	for (;;) {
7289 		outstanding = 0;
7290 		for (i = 0; i < NUM_PAGE_CAPTURE_BUCKETS; i++)
7291 			outstanding += page_capture_hash[i].num_pages;
7292 		if (outstanding) {
7293 			page_capture_handle_outstanding();
7294 			CALLB_CPR_SAFE_BEGIN(&c);
7295 			(void) cv_timedwait(&pc_cv, &pc_thread_mutex,
7296 			    lbolt + pc_thread_shortwait);
7297 			CALLB_CPR_SAFE_END(&c, &pc_thread_mutex);
7298 		} else {
7299 			CALLB_CPR_SAFE_BEGIN(&c);
7300 			(void) cv_timedwait(&pc_cv, &pc_thread_mutex,
7301 			    lbolt + pc_thread_longwait);
7302 			CALLB_CPR_SAFE_END(&c, &pc_thread_mutex);
7303 		}
7304 	}
7305 	/*NOTREACHED*/
7306 }
7307 /*
7308  * Attempt to locate a bucket that has enough pages to satisfy the request.
7309  * The initial check is done without the lock to avoid unneeded contention.
7310  * The function returns 1 if enough pages were found, else 0 if it could not
7311  * find enough pages in a bucket.
7312  */
7313 static int
7314 pcf_decrement_bucket(pgcnt_t npages)
7315 {
7316 	struct pcf	*p;
7317 	struct pcf	*q;
7318 	int i;
7319 
7320 	p = &pcf[PCF_INDEX()];
7321 	q = &pcf[pcf_fanout];
7322 	for (i = 0; i < pcf_fanout; i++) {
7323 		if (p->pcf_count > npages) {
7324 			/*
7325 			 * a good one to try.
7326 			 */
7327 			mutex_enter(&p->pcf_lock);
7328 			if (p->pcf_count > npages) {
7329 				p->pcf_count -= (uint_t)npages;
7330 				/*
7331 				 * freemem is not protected by any lock.
7332 				 * Thus, we cannot have any assertion
7333 				 * containing freemem here.
7334 				 */
7335 				freemem -= npages;
7336 				mutex_exit(&p->pcf_lock);
7337 				return (1);
7338 			}
7339 			mutex_exit(&p->pcf_lock);
7340 		}
7341 		p++;
7342 		if (p >= q) {
7343 			p = pcf;
7344 		}
7345 	}
7346 	return (0);
7347 }
7348 
7349 /*
7350  * Arguments:
7351  *	pcftotal_ret:	If the value is not NULL and we have walked all the
7352  *			buckets but did not find enough pages then it will
7353  *			be set to the total number of pages in all the pcf
7354  *			buckets.
7355  *	npages:		Is the number of pages we have been requested to
7356  *			find.
7357  *	unlock:		If set to 0 we will leave the buckets locked if the
7358  *			requested number of pages are not found.
7359  *
7360  * Go and try to satisfy the page request  from any number of buckets.
7361  * This can be a very expensive operation as we have to lock the buckets
7362  * we are checking (and keep them locked), starting at bucket 0.
7363  *
7364  * The function returns 1 if enough pages were found, else 0 if it could not
7365  * find enough pages in the buckets.
7366  *
7367  */
7368 static int
7369 pcf_decrement_multiple(pgcnt_t *pcftotal_ret, pgcnt_t npages, int unlock)
7370 {
7371 	struct pcf	*p;
7372 	pgcnt_t pcftotal;
7373 	int i;
7374 
7375 	p = pcf;
7376 	/* try to collect pages from several pcf bins */
7377 	for (pcftotal = 0, i = 0; i < pcf_fanout; i++) {
7378 		mutex_enter(&p->pcf_lock);
7379 		pcftotal += p->pcf_count;
7380 		if (pcftotal >= npages) {
7381 			/*
7382 			 * Wow!  There are enough pages laying around
7383 			 * to satisfy the request.  Do the accounting,
7384 			 * drop the locks we acquired, and go back.
7385 			 *
7386 			 * freemem is not protected by any lock. So,
7387 			 * we cannot have any assertion containing
7388 			 * freemem.
7389 			 */
7390 			freemem -= npages;
7391 			while (p >= pcf) {
7392 				if (p->pcf_count <= npages) {
7393 					npages -= p->pcf_count;
7394 					p->pcf_count = 0;
7395 				} else {
7396 					p->pcf_count -= (uint_t)npages;
7397 					npages = 0;
7398 				}
7399 				mutex_exit(&p->pcf_lock);
7400 				p--;
7401 			}
7402 			ASSERT(npages == 0);
7403 			return (1);
7404 		}
7405 		p++;
7406 	}
7407 	if (unlock) {
7408 		/* failed to collect pages - release the locks */
7409 		while (--p >= pcf) {
7410 			mutex_exit(&p->pcf_lock);
7411 		}
7412 	}
7413 	if (pcftotal_ret != NULL)
7414 		*pcftotal_ret = pcftotal;
7415 	return (0);
7416 }
7417