xref: /freebsd/sys/vm/vm_page.c (revision 83823d063ab57db8d3954c1530d036f1ccdceb41)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
36  */
37 
38 /*-
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  *	Resident memory management module.
67  */
68 
69 #include <sys/cdefs.h>
70 #include "opt_vm.h"
71 
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/counter.h>
75 #include <sys/domainset.h>
76 #include <sys/kernel.h>
77 #include <sys/limits.h>
78 #include <sys/linker.h>
79 #include <sys/lock.h>
80 #include <sys/malloc.h>
81 #include <sys/mman.h>
82 #include <sys/msgbuf.h>
83 #include <sys/mutex.h>
84 #include <sys/proc.h>
85 #include <sys/rwlock.h>
86 #include <sys/sleepqueue.h>
87 #include <sys/sbuf.h>
88 #include <sys/sched.h>
89 #include <sys/smp.h>
90 #include <sys/sysctl.h>
91 #include <sys/vmmeter.h>
92 #include <sys/vnode.h>
93 
94 #include <vm/vm.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_param.h>
97 #include <vm/vm_domainset.h>
98 #include <vm/vm_kern.h>
99 #include <vm/vm_map.h>
100 #include <vm/vm_object.h>
101 #include <vm/vm_page.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_phys.h>
104 #include <vm/vm_pagequeue.h>
105 #include <vm/vm_pager.h>
106 #include <vm/vm_radix.h>
107 #include <vm/vm_reserv.h>
108 #include <vm/vm_extern.h>
109 #include <vm/vm_dumpset.h>
110 #include <vm/uma.h>
111 #include <vm/uma_int.h>
112 
113 #include <machine/md_var.h>
114 
115 struct vm_domain vm_dom[MAXMEMDOM];
116 
117 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
118 
119 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
120 
121 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
122 /* The following fields are protected by the domainset lock. */
123 domainset_t __exclusive_cache_line vm_min_domains;
124 domainset_t __exclusive_cache_line vm_severe_domains;
125 static int vm_min_waiters;
126 static int vm_severe_waiters;
127 static int vm_pageproc_waiters;
128 
129 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
130     "VM page statistics");
131 
132 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
133 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
134     CTLFLAG_RD, &pqstate_commit_retries,
135     "Number of failed per-page atomic queue state updates");
136 
137 static COUNTER_U64_DEFINE_EARLY(queue_ops);
138 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
139     CTLFLAG_RD, &queue_ops,
140     "Number of batched queue operations");
141 
142 static COUNTER_U64_DEFINE_EARLY(queue_nops);
143 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
144     CTLFLAG_RD, &queue_nops,
145     "Number of batched queue operations with no effects");
146 
147 /*
148  * bogus page -- for I/O to/from partially complete buffers,
149  * or for paging into sparsely invalid regions.
150  */
151 vm_page_t bogus_page;
152 
153 vm_page_t vm_page_array;
154 long vm_page_array_size;
155 long first_page;
156 
157 struct bitset *vm_page_dump;
158 long vm_page_dump_pages;
159 
160 static TAILQ_HEAD(, vm_page) blacklist_head;
161 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
162 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
163     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
164 
165 static uma_zone_t fakepg_zone;
166 
167 static void vm_page_alloc_check(vm_page_t m);
168 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
169     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
170 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
171 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
172 static bool vm_page_free_prep(vm_page_t m);
173 static void vm_page_free_toq(vm_page_t m);
174 static void vm_page_init(void *dummy);
175 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
176     vm_pindex_t pindex, vm_page_t mpred);
177 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
178     vm_page_t mpred);
179 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
180     const uint16_t nflag);
181 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
182     vm_page_t m_run, vm_paddr_t high);
183 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
184 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
185     int req);
186 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
187     int flags);
188 static void vm_page_zone_release(void *arg, void **store, int cnt);
189 
190 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
191 
192 static void
193 vm_page_init(void *dummy)
194 {
195 
196 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
197 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
198 	bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED);
199 }
200 
201 static int pgcache_zone_max_pcpu;
202 SYSCTL_INT(_vm, OID_AUTO, pgcache_zone_max_pcpu,
203     CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pgcache_zone_max_pcpu, 0,
204     "Per-CPU page cache size");
205 
206 /*
207  * The cache page zone is initialized later since we need to be able to allocate
208  * pages before UMA is fully initialized.
209  */
210 static void
211 vm_page_init_cache_zones(void *dummy __unused)
212 {
213 	struct vm_domain *vmd;
214 	struct vm_pgcache *pgcache;
215 	int cache, domain, maxcache, pool;
216 
217 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &pgcache_zone_max_pcpu);
218 	maxcache = pgcache_zone_max_pcpu * mp_ncpus;
219 	for (domain = 0; domain < vm_ndomains; domain++) {
220 		vmd = VM_DOMAIN(domain);
221 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
222 			pgcache = &vmd->vmd_pgcache[pool];
223 			pgcache->domain = domain;
224 			pgcache->pool = pool;
225 			pgcache->zone = uma_zcache_create("vm pgcache",
226 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
227 			    vm_page_zone_import, vm_page_zone_release, pgcache,
228 			    UMA_ZONE_VM);
229 
230 			/*
231 			 * Limit each pool's zone to 0.1% of the pages in the
232 			 * domain.
233 			 */
234 			cache = maxcache != 0 ? maxcache :
235 			    vmd->vmd_page_count / 1000;
236 			uma_zone_set_maxcache(pgcache->zone, cache);
237 		}
238 	}
239 }
240 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
241 
242 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
243 #if PAGE_SIZE == 32768
244 #ifdef CTASSERT
245 CTASSERT(sizeof(u_long) >= 8);
246 #endif
247 #endif
248 
249 /*
250  *	vm_set_page_size:
251  *
252  *	Sets the page size, perhaps based upon the memory
253  *	size.  Must be called before any use of page-size
254  *	dependent functions.
255  */
256 void
257 vm_set_page_size(void)
258 {
259 	if (vm_cnt.v_page_size == 0)
260 		vm_cnt.v_page_size = PAGE_SIZE;
261 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
262 		panic("vm_set_page_size: page size not a power of two");
263 }
264 
265 /*
266  *	vm_page_blacklist_next:
267  *
268  *	Find the next entry in the provided string of blacklist
269  *	addresses.  Entries are separated by space, comma, or newline.
270  *	If an invalid integer is encountered then the rest of the
271  *	string is skipped.  Updates the list pointer to the next
272  *	character, or NULL if the string is exhausted or invalid.
273  */
274 static vm_paddr_t
275 vm_page_blacklist_next(char **list, char *end)
276 {
277 	vm_paddr_t bad;
278 	char *cp, *pos;
279 
280 	if (list == NULL || *list == NULL)
281 		return (0);
282 	if (**list =='\0') {
283 		*list = NULL;
284 		return (0);
285 	}
286 
287 	/*
288 	 * If there's no end pointer then the buffer is coming from
289 	 * the kenv and we know it's null-terminated.
290 	 */
291 	if (end == NULL)
292 		end = *list + strlen(*list);
293 
294 	/* Ensure that strtoq() won't walk off the end */
295 	if (*end != '\0') {
296 		if (*end == '\n' || *end == ' ' || *end  == ',')
297 			*end = '\0';
298 		else {
299 			printf("Blacklist not terminated, skipping\n");
300 			*list = NULL;
301 			return (0);
302 		}
303 	}
304 
305 	for (pos = *list; *pos != '\0'; pos = cp) {
306 		bad = strtoq(pos, &cp, 0);
307 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
308 			if (bad == 0) {
309 				if (++cp < end)
310 					continue;
311 				else
312 					break;
313 			}
314 		} else
315 			break;
316 		if (*cp == '\0' || ++cp >= end)
317 			*list = NULL;
318 		else
319 			*list = cp;
320 		return (trunc_page(bad));
321 	}
322 	printf("Garbage in RAM blacklist, skipping\n");
323 	*list = NULL;
324 	return (0);
325 }
326 
327 bool
328 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
329 {
330 	struct vm_domain *vmd;
331 	vm_page_t m;
332 	bool found;
333 
334 	m = vm_phys_paddr_to_vm_page(pa);
335 	if (m == NULL)
336 		return (true); /* page does not exist, no failure */
337 
338 	vmd = vm_pagequeue_domain(m);
339 	vm_domain_free_lock(vmd);
340 	found = vm_phys_unfree_page(m);
341 	vm_domain_free_unlock(vmd);
342 	if (found) {
343 		vm_domain_freecnt_inc(vmd, -1);
344 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
345 		if (verbose)
346 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
347 	}
348 	return (found);
349 }
350 
351 /*
352  *	vm_page_blacklist_check:
353  *
354  *	Iterate through the provided string of blacklist addresses, pulling
355  *	each entry out of the physical allocator free list and putting it
356  *	onto a list for reporting via the vm.page_blacklist sysctl.
357  */
358 static void
359 vm_page_blacklist_check(char *list, char *end)
360 {
361 	vm_paddr_t pa;
362 	char *next;
363 
364 	next = list;
365 	while (next != NULL) {
366 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
367 			continue;
368 		vm_page_blacklist_add(pa, bootverbose);
369 	}
370 }
371 
372 /*
373  *	vm_page_blacklist_load:
374  *
375  *	Search for a special module named "ram_blacklist".  It'll be a
376  *	plain text file provided by the user via the loader directive
377  *	of the same name.
378  */
379 static void
380 vm_page_blacklist_load(char **list, char **end)
381 {
382 	void *mod;
383 	u_char *ptr;
384 	u_int len;
385 
386 	mod = NULL;
387 	ptr = NULL;
388 
389 	mod = preload_search_by_type("ram_blacklist");
390 	if (mod != NULL) {
391 		ptr = preload_fetch_addr(mod);
392 		len = preload_fetch_size(mod);
393         }
394 	*list = ptr;
395 	if (ptr != NULL)
396 		*end = ptr + len;
397 	else
398 		*end = NULL;
399 	return;
400 }
401 
402 static int
403 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
404 {
405 	vm_page_t m;
406 	struct sbuf sbuf;
407 	int error, first;
408 
409 	first = 1;
410 	error = sysctl_wire_old_buffer(req, 0);
411 	if (error != 0)
412 		return (error);
413 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
414 	TAILQ_FOREACH(m, &blacklist_head, listq) {
415 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
416 		    (uintmax_t)m->phys_addr);
417 		first = 0;
418 	}
419 	error = sbuf_finish(&sbuf);
420 	sbuf_delete(&sbuf);
421 	return (error);
422 }
423 
424 /*
425  * Initialize a dummy page for use in scans of the specified paging queue.
426  * In principle, this function only needs to set the flag PG_MARKER.
427  * Nonetheless, it write busies the page as a safety precaution.
428  */
429 void
430 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
431 {
432 
433 	bzero(marker, sizeof(*marker));
434 	marker->flags = PG_MARKER;
435 	marker->a.flags = aflags;
436 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
437 	marker->a.queue = queue;
438 }
439 
440 static void
441 vm_page_domain_init(int domain)
442 {
443 	struct vm_domain *vmd;
444 	struct vm_pagequeue *pq;
445 	int i;
446 
447 	vmd = VM_DOMAIN(domain);
448 	bzero(vmd, sizeof(*vmd));
449 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
450 	    "vm inactive pagequeue";
451 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
452 	    "vm active pagequeue";
453 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
454 	    "vm laundry pagequeue";
455 	*__DECONST(const char **,
456 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
457 	    "vm unswappable pagequeue";
458 	vmd->vmd_domain = domain;
459 	vmd->vmd_page_count = 0;
460 	vmd->vmd_free_count = 0;
461 	vmd->vmd_segs = 0;
462 	vmd->vmd_oom = FALSE;
463 	for (i = 0; i < PQ_COUNT; i++) {
464 		pq = &vmd->vmd_pagequeues[i];
465 		TAILQ_INIT(&pq->pq_pl);
466 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
467 		    MTX_DEF | MTX_DUPOK);
468 		pq->pq_pdpages = 0;
469 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
470 	}
471 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
472 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
473 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
474 
475 	/*
476 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
477 	 * insertions.
478 	 */
479 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
480 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
481 	    &vmd->vmd_inacthead, plinks.q);
482 
483 	/*
484 	 * The clock pages are used to implement active queue scanning without
485 	 * requeues.  Scans start at clock[0], which is advanced after the scan
486 	 * ends.  When the two clock hands meet, they are reset and scanning
487 	 * resumes from the head of the queue.
488 	 */
489 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
490 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
491 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
492 	    &vmd->vmd_clock[0], plinks.q);
493 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
494 	    &vmd->vmd_clock[1], plinks.q);
495 }
496 
497 /*
498  * Initialize a physical page in preparation for adding it to the free
499  * lists.
500  */
501 void
502 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
503 {
504 
505 	m->object = NULL;
506 	m->ref_count = 0;
507 	m->busy_lock = VPB_FREED;
508 	m->flags = m->a.flags = 0;
509 	m->phys_addr = pa;
510 	m->a.queue = PQ_NONE;
511 	m->psind = 0;
512 	m->segind = segind;
513 	m->order = VM_NFREEORDER;
514 	m->pool = VM_FREEPOOL_DEFAULT;
515 	m->valid = m->dirty = 0;
516 	pmap_page_init(m);
517 }
518 
519 #ifndef PMAP_HAS_PAGE_ARRAY
520 static vm_paddr_t
521 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
522 {
523 	vm_paddr_t new_end;
524 
525 	/*
526 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
527 	 * However, because this page is allocated from KVM, out-of-bounds
528 	 * accesses using the direct map will not be trapped.
529 	 */
530 	*vaddr += PAGE_SIZE;
531 
532 	/*
533 	 * Allocate physical memory for the page structures, and map it.
534 	 */
535 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
536 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
537 	    VM_PROT_READ | VM_PROT_WRITE);
538 	vm_page_array_size = page_range;
539 
540 	return (new_end);
541 }
542 #endif
543 
544 /*
545  *	vm_page_startup:
546  *
547  *	Initializes the resident memory module.  Allocates physical memory for
548  *	bootstrapping UMA and some data structures that are used to manage
549  *	physical pages.  Initializes these structures, and populates the free
550  *	page queues.
551  */
552 vm_offset_t
553 vm_page_startup(vm_offset_t vaddr)
554 {
555 	struct vm_phys_seg *seg;
556 	struct vm_domain *vmd;
557 	vm_page_t m;
558 	char *list, *listend;
559 	vm_paddr_t end, high_avail, low_avail, new_end, size;
560 	vm_paddr_t page_range __unused;
561 	vm_paddr_t last_pa, pa, startp, endp;
562 	u_long pagecount;
563 #if MINIDUMP_PAGE_TRACKING
564 	u_long vm_page_dump_size;
565 #endif
566 	int biggestone, i, segind;
567 #ifdef WITNESS
568 	vm_offset_t mapped;
569 	int witness_size;
570 #endif
571 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
572 	long ii;
573 #endif
574 
575 	vaddr = round_page(vaddr);
576 
577 	vm_phys_early_startup();
578 	biggestone = vm_phys_avail_largest();
579 	end = phys_avail[biggestone+1];
580 
581 	/*
582 	 * Initialize the page and queue locks.
583 	 */
584 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
585 	for (i = 0; i < PA_LOCK_COUNT; i++)
586 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
587 	for (i = 0; i < vm_ndomains; i++)
588 		vm_page_domain_init(i);
589 
590 	new_end = end;
591 #ifdef WITNESS
592 	witness_size = round_page(witness_startup_count());
593 	new_end -= witness_size;
594 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
595 	    VM_PROT_READ | VM_PROT_WRITE);
596 	bzero((void *)mapped, witness_size);
597 	witness_startup((void *)mapped);
598 #endif
599 
600 #if MINIDUMP_PAGE_TRACKING
601 	/*
602 	 * Allocate a bitmap to indicate that a random physical page
603 	 * needs to be included in a minidump.
604 	 *
605 	 * The amd64 port needs this to indicate which direct map pages
606 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
607 	 *
608 	 * However, i386 still needs this workspace internally within the
609 	 * minidump code.  In theory, they are not needed on i386, but are
610 	 * included should the sf_buf code decide to use them.
611 	 */
612 	last_pa = 0;
613 	vm_page_dump_pages = 0;
614 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
615 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
616 		    dump_avail[i] / PAGE_SIZE;
617 		if (dump_avail[i + 1] > last_pa)
618 			last_pa = dump_avail[i + 1];
619 	}
620 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
621 	new_end -= vm_page_dump_size;
622 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
623 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
624 	bzero((void *)vm_page_dump, vm_page_dump_size);
625 #else
626 	(void)last_pa;
627 #endif
628 #if defined(__aarch64__) || defined(__amd64__) || \
629     defined(__riscv) || defined(__powerpc64__)
630 	/*
631 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
632 	 * in a crash dump.  When pmap_map() uses the direct map, they are
633 	 * not automatically included.
634 	 */
635 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
636 		dump_add_page(pa);
637 #endif
638 	phys_avail[biggestone + 1] = new_end;
639 #ifdef __amd64__
640 	/*
641 	 * Request that the physical pages underlying the message buffer be
642 	 * included in a crash dump.  Since the message buffer is accessed
643 	 * through the direct map, they are not automatically included.
644 	 */
645 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
646 	last_pa = pa + round_page(msgbufsize);
647 	while (pa < last_pa) {
648 		dump_add_page(pa);
649 		pa += PAGE_SIZE;
650 	}
651 #endif
652 	/*
653 	 * Compute the number of pages of memory that will be available for
654 	 * use, taking into account the overhead of a page structure per page.
655 	 * In other words, solve
656 	 *	"available physical memory" - round_page(page_range *
657 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
658 	 * for page_range.
659 	 */
660 	low_avail = phys_avail[0];
661 	high_avail = phys_avail[1];
662 	for (i = 0; i < vm_phys_nsegs; i++) {
663 		if (vm_phys_segs[i].start < low_avail)
664 			low_avail = vm_phys_segs[i].start;
665 		if (vm_phys_segs[i].end > high_avail)
666 			high_avail = vm_phys_segs[i].end;
667 	}
668 	/* Skip the first chunk.  It is already accounted for. */
669 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
670 		if (phys_avail[i] < low_avail)
671 			low_avail = phys_avail[i];
672 		if (phys_avail[i + 1] > high_avail)
673 			high_avail = phys_avail[i + 1];
674 	}
675 	first_page = low_avail / PAGE_SIZE;
676 #ifdef VM_PHYSSEG_SPARSE
677 	size = 0;
678 	for (i = 0; i < vm_phys_nsegs; i++)
679 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
680 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
681 		size += phys_avail[i + 1] - phys_avail[i];
682 #elif defined(VM_PHYSSEG_DENSE)
683 	size = high_avail - low_avail;
684 #else
685 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
686 #endif
687 
688 #ifdef PMAP_HAS_PAGE_ARRAY
689 	pmap_page_array_startup(size / PAGE_SIZE);
690 	biggestone = vm_phys_avail_largest();
691 	end = new_end = phys_avail[biggestone + 1];
692 #else
693 #ifdef VM_PHYSSEG_DENSE
694 	/*
695 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
696 	 * the overhead of a page structure per page only if vm_page_array is
697 	 * allocated from the last physical memory chunk.  Otherwise, we must
698 	 * allocate page structures representing the physical memory
699 	 * underlying vm_page_array, even though they will not be used.
700 	 */
701 	if (new_end != high_avail)
702 		page_range = size / PAGE_SIZE;
703 	else
704 #endif
705 	{
706 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
707 
708 		/*
709 		 * If the partial bytes remaining are large enough for
710 		 * a page (PAGE_SIZE) without a corresponding
711 		 * 'struct vm_page', then new_end will contain an
712 		 * extra page after subtracting the length of the VM
713 		 * page array.  Compensate by subtracting an extra
714 		 * page from new_end.
715 		 */
716 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
717 			if (new_end == high_avail)
718 				high_avail -= PAGE_SIZE;
719 			new_end -= PAGE_SIZE;
720 		}
721 	}
722 	end = new_end;
723 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
724 #endif
725 
726 #if VM_NRESERVLEVEL > 0
727 	/*
728 	 * Allocate physical memory for the reservation management system's
729 	 * data structures, and map it.
730 	 */
731 	new_end = vm_reserv_startup(&vaddr, new_end);
732 #endif
733 #if defined(__aarch64__) || defined(__amd64__) || \
734     defined(__riscv) || defined(__powerpc64__)
735 	/*
736 	 * Include vm_page_array and vm_reserv_array in a crash dump.
737 	 */
738 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
739 		dump_add_page(pa);
740 #endif
741 	phys_avail[biggestone + 1] = new_end;
742 
743 	/*
744 	 * Add physical memory segments corresponding to the available
745 	 * physical pages.
746 	 */
747 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
748 		if (vm_phys_avail_size(i) != 0)
749 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
750 
751 	/*
752 	 * Initialize the physical memory allocator.
753 	 */
754 	vm_phys_init();
755 
756 	/*
757 	 * Initialize the page structures and add every available page to the
758 	 * physical memory allocator's free lists.
759 	 */
760 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
761 	for (ii = 0; ii < vm_page_array_size; ii++) {
762 		m = &vm_page_array[ii];
763 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
764 		m->flags = PG_FICTITIOUS;
765 	}
766 #endif
767 	vm_cnt.v_page_count = 0;
768 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
769 		seg = &vm_phys_segs[segind];
770 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
771 		    m++, pa += PAGE_SIZE)
772 			vm_page_init_page(m, pa, segind);
773 
774 		/*
775 		 * Add the segment's pages that are covered by one of
776 		 * phys_avail's ranges to the free lists.
777 		 */
778 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
779 			if (seg->end <= phys_avail[i] ||
780 			    seg->start >= phys_avail[i + 1])
781 				continue;
782 
783 			startp = MAX(seg->start, phys_avail[i]);
784 			endp = MIN(seg->end, phys_avail[i + 1]);
785 			pagecount = (u_long)atop(endp - startp);
786 			if (pagecount == 0)
787 				continue;
788 
789 			m = seg->first_page + atop(startp - seg->start);
790 			vmd = VM_DOMAIN(seg->domain);
791 			vm_domain_free_lock(vmd);
792 			vm_phys_enqueue_contig(m, pagecount);
793 			vm_domain_free_unlock(vmd);
794 			vm_domain_freecnt_inc(vmd, pagecount);
795 			vm_cnt.v_page_count += (u_int)pagecount;
796 			vmd->vmd_page_count += (u_int)pagecount;
797 			vmd->vmd_segs |= 1UL << segind;
798 		}
799 	}
800 
801 	/*
802 	 * Remove blacklisted pages from the physical memory allocator.
803 	 */
804 	TAILQ_INIT(&blacklist_head);
805 	vm_page_blacklist_load(&list, &listend);
806 	vm_page_blacklist_check(list, listend);
807 
808 	list = kern_getenv("vm.blacklist");
809 	vm_page_blacklist_check(list, NULL);
810 
811 	freeenv(list);
812 #if VM_NRESERVLEVEL > 0
813 	/*
814 	 * Initialize the reservation management system.
815 	 */
816 	vm_reserv_init();
817 #endif
818 
819 	return (vaddr);
820 }
821 
822 void
823 vm_page_reference(vm_page_t m)
824 {
825 
826 	vm_page_aflag_set(m, PGA_REFERENCED);
827 }
828 
829 /*
830  *	vm_page_trybusy
831  *
832  *	Helper routine for grab functions to trylock busy.
833  *
834  *	Returns true on success and false on failure.
835  */
836 static bool
837 vm_page_trybusy(vm_page_t m, int allocflags)
838 {
839 
840 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
841 		return (vm_page_trysbusy(m));
842 	else
843 		return (vm_page_tryxbusy(m));
844 }
845 
846 /*
847  *	vm_page_tryacquire
848  *
849  *	Helper routine for grab functions to trylock busy and wire.
850  *
851  *	Returns true on success and false on failure.
852  */
853 static inline bool
854 vm_page_tryacquire(vm_page_t m, int allocflags)
855 {
856 	bool locked;
857 
858 	locked = vm_page_trybusy(m, allocflags);
859 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
860 		vm_page_wire(m);
861 	return (locked);
862 }
863 
864 /*
865  *	vm_page_busy_acquire:
866  *
867  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
868  *	and drop the object lock if necessary.
869  */
870 bool
871 vm_page_busy_acquire(vm_page_t m, int allocflags)
872 {
873 	vm_object_t obj;
874 	bool locked;
875 
876 	/*
877 	 * The page-specific object must be cached because page
878 	 * identity can change during the sleep, causing the
879 	 * re-lock of a different object.
880 	 * It is assumed that a reference to the object is already
881 	 * held by the callers.
882 	 */
883 	obj = atomic_load_ptr(&m->object);
884 	for (;;) {
885 		if (vm_page_tryacquire(m, allocflags))
886 			return (true);
887 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
888 			return (false);
889 		if (obj != NULL)
890 			locked = VM_OBJECT_WOWNED(obj);
891 		else
892 			locked = false;
893 		MPASS(locked || vm_page_wired(m));
894 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
895 		    locked) && locked)
896 			VM_OBJECT_WLOCK(obj);
897 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
898 			return (false);
899 		KASSERT(m->object == obj || m->object == NULL,
900 		    ("vm_page_busy_acquire: page %p does not belong to %p",
901 		    m, obj));
902 	}
903 }
904 
905 /*
906  *	vm_page_busy_downgrade:
907  *
908  *	Downgrade an exclusive busy page into a single shared busy page.
909  */
910 void
911 vm_page_busy_downgrade(vm_page_t m)
912 {
913 	u_int x;
914 
915 	vm_page_assert_xbusied(m);
916 
917 	x = vm_page_busy_fetch(m);
918 	for (;;) {
919 		if (atomic_fcmpset_rel_int(&m->busy_lock,
920 		    &x, VPB_SHARERS_WORD(1)))
921 			break;
922 	}
923 	if ((x & VPB_BIT_WAITERS) != 0)
924 		wakeup(m);
925 }
926 
927 /*
928  *
929  *	vm_page_busy_tryupgrade:
930  *
931  *	Attempt to upgrade a single shared busy into an exclusive busy.
932  */
933 int
934 vm_page_busy_tryupgrade(vm_page_t m)
935 {
936 	u_int ce, x;
937 
938 	vm_page_assert_sbusied(m);
939 
940 	x = vm_page_busy_fetch(m);
941 	ce = VPB_CURTHREAD_EXCLUSIVE;
942 	for (;;) {
943 		if (VPB_SHARERS(x) > 1)
944 			return (0);
945 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
946 		    ("vm_page_busy_tryupgrade: invalid lock state"));
947 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
948 		    ce | (x & VPB_BIT_WAITERS)))
949 			continue;
950 		return (1);
951 	}
952 }
953 
954 /*
955  *	vm_page_sbusied:
956  *
957  *	Return a positive value if the page is shared busied, 0 otherwise.
958  */
959 int
960 vm_page_sbusied(vm_page_t m)
961 {
962 	u_int x;
963 
964 	x = vm_page_busy_fetch(m);
965 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
966 }
967 
968 /*
969  *	vm_page_sunbusy:
970  *
971  *	Shared unbusy a page.
972  */
973 void
974 vm_page_sunbusy(vm_page_t m)
975 {
976 	u_int x;
977 
978 	vm_page_assert_sbusied(m);
979 
980 	x = vm_page_busy_fetch(m);
981 	for (;;) {
982 		KASSERT(x != VPB_FREED,
983 		    ("vm_page_sunbusy: Unlocking freed page."));
984 		if (VPB_SHARERS(x) > 1) {
985 			if (atomic_fcmpset_int(&m->busy_lock, &x,
986 			    x - VPB_ONE_SHARER))
987 				break;
988 			continue;
989 		}
990 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
991 		    ("vm_page_sunbusy: invalid lock state"));
992 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
993 			continue;
994 		if ((x & VPB_BIT_WAITERS) == 0)
995 			break;
996 		wakeup(m);
997 		break;
998 	}
999 }
1000 
1001 /*
1002  *	vm_page_busy_sleep:
1003  *
1004  *	Sleep if the page is busy, using the page pointer as wchan.
1005  *	This is used to implement the hard-path of the busying mechanism.
1006  *
1007  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1008  *	will not sleep if the page is shared-busy.
1009  *
1010  *	The object lock must be held on entry.
1011  *
1012  *	Returns true if it slept and dropped the object lock, or false
1013  *	if there was no sleep and the lock is still held.
1014  */
1015 bool
1016 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1017 {
1018 	vm_object_t obj;
1019 
1020 	obj = m->object;
1021 	VM_OBJECT_ASSERT_LOCKED(obj);
1022 
1023 	return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1024 	    true));
1025 }
1026 
1027 /*
1028  *	vm_page_busy_sleep_unlocked:
1029  *
1030  *	Sleep if the page is busy, using the page pointer as wchan.
1031  *	This is used to implement the hard-path of busying mechanism.
1032  *
1033  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1034  *	will not sleep if the page is shared-busy.
1035  *
1036  *	The object lock must not be held on entry.  The operation will
1037  *	return if the page changes identity.
1038  */
1039 void
1040 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1041     const char *wmesg, int allocflags)
1042 {
1043 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1044 
1045 	(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1046 }
1047 
1048 /*
1049  *	_vm_page_busy_sleep:
1050  *
1051  *	Internal busy sleep function.  Verifies the page identity and
1052  *	lockstate against parameters.  Returns true if it sleeps and
1053  *	false otherwise.
1054  *
1055  *	allocflags uses VM_ALLOC_* flags to specify the lock required.
1056  *
1057  *	If locked is true the lock will be dropped for any true returns
1058  *	and held for any false returns.
1059  */
1060 static bool
1061 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1062     const char *wmesg, int allocflags, bool locked)
1063 {
1064 	bool xsleep;
1065 	u_int x;
1066 
1067 	/*
1068 	 * If the object is busy we must wait for that to drain to zero
1069 	 * before trying the page again.
1070 	 */
1071 	if (obj != NULL && vm_object_busied(obj)) {
1072 		if (locked)
1073 			VM_OBJECT_DROP(obj);
1074 		vm_object_busy_wait(obj, wmesg);
1075 		return (true);
1076 	}
1077 
1078 	if (!vm_page_busied(m))
1079 		return (false);
1080 
1081 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1082 	sleepq_lock(m);
1083 	x = vm_page_busy_fetch(m);
1084 	do {
1085 		/*
1086 		 * If the page changes objects or becomes unlocked we can
1087 		 * simply return.
1088 		 */
1089 		if (x == VPB_UNBUSIED ||
1090 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1091 		    m->object != obj || m->pindex != pindex) {
1092 			sleepq_release(m);
1093 			return (false);
1094 		}
1095 		if ((x & VPB_BIT_WAITERS) != 0)
1096 			break;
1097 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1098 	if (locked)
1099 		VM_OBJECT_DROP(obj);
1100 	DROP_GIANT();
1101 	sleepq_add(m, NULL, wmesg, 0, 0);
1102 	sleepq_wait(m, PVM);
1103 	PICKUP_GIANT();
1104 	return (true);
1105 }
1106 
1107 /*
1108  *	vm_page_trysbusy:
1109  *
1110  *	Try to shared busy a page.
1111  *	If the operation succeeds 1 is returned otherwise 0.
1112  *	The operation never sleeps.
1113  */
1114 int
1115 vm_page_trysbusy(vm_page_t m)
1116 {
1117 	vm_object_t obj;
1118 	u_int x;
1119 
1120 	obj = m->object;
1121 	x = vm_page_busy_fetch(m);
1122 	for (;;) {
1123 		if ((x & VPB_BIT_SHARED) == 0)
1124 			return (0);
1125 		/*
1126 		 * Reduce the window for transient busies that will trigger
1127 		 * false negatives in vm_page_ps_test().
1128 		 */
1129 		if (obj != NULL && vm_object_busied(obj))
1130 			return (0);
1131 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1132 		    x + VPB_ONE_SHARER))
1133 			break;
1134 	}
1135 
1136 	/* Refetch the object now that we're guaranteed that it is stable. */
1137 	obj = m->object;
1138 	if (obj != NULL && vm_object_busied(obj)) {
1139 		vm_page_sunbusy(m);
1140 		return (0);
1141 	}
1142 	return (1);
1143 }
1144 
1145 /*
1146  *	vm_page_tryxbusy:
1147  *
1148  *	Try to exclusive busy a page.
1149  *	If the operation succeeds 1 is returned otherwise 0.
1150  *	The operation never sleeps.
1151  */
1152 int
1153 vm_page_tryxbusy(vm_page_t m)
1154 {
1155 	vm_object_t obj;
1156 
1157         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1158             VPB_CURTHREAD_EXCLUSIVE) == 0)
1159 		return (0);
1160 
1161 	obj = m->object;
1162 	if (obj != NULL && vm_object_busied(obj)) {
1163 		vm_page_xunbusy(m);
1164 		return (0);
1165 	}
1166 	return (1);
1167 }
1168 
1169 static void
1170 vm_page_xunbusy_hard_tail(vm_page_t m)
1171 {
1172 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1173 	/* Wake the waiter. */
1174 	wakeup(m);
1175 }
1176 
1177 /*
1178  *	vm_page_xunbusy_hard:
1179  *
1180  *	Called when unbusy has failed because there is a waiter.
1181  */
1182 void
1183 vm_page_xunbusy_hard(vm_page_t m)
1184 {
1185 	vm_page_assert_xbusied(m);
1186 	vm_page_xunbusy_hard_tail(m);
1187 }
1188 
1189 void
1190 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1191 {
1192 	vm_page_assert_xbusied_unchecked(m);
1193 	vm_page_xunbusy_hard_tail(m);
1194 }
1195 
1196 static void
1197 vm_page_busy_free(vm_page_t m)
1198 {
1199 	u_int x;
1200 
1201 	atomic_thread_fence_rel();
1202 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1203 	if ((x & VPB_BIT_WAITERS) != 0)
1204 		wakeup(m);
1205 }
1206 
1207 /*
1208  *	vm_page_unhold_pages:
1209  *
1210  *	Unhold each of the pages that is referenced by the given array.
1211  */
1212 void
1213 vm_page_unhold_pages(vm_page_t *ma, int count)
1214 {
1215 
1216 	for (; count != 0; count--) {
1217 		vm_page_unwire(*ma, PQ_ACTIVE);
1218 		ma++;
1219 	}
1220 }
1221 
1222 vm_page_t
1223 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1224 {
1225 	vm_page_t m;
1226 
1227 #ifdef VM_PHYSSEG_SPARSE
1228 	m = vm_phys_paddr_to_vm_page(pa);
1229 	if (m == NULL)
1230 		m = vm_phys_fictitious_to_vm_page(pa);
1231 	return (m);
1232 #elif defined(VM_PHYSSEG_DENSE)
1233 	long pi;
1234 
1235 	pi = atop(pa);
1236 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1237 		m = &vm_page_array[pi - first_page];
1238 		return (m);
1239 	}
1240 	return (vm_phys_fictitious_to_vm_page(pa));
1241 #else
1242 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1243 #endif
1244 }
1245 
1246 /*
1247  *	vm_page_getfake:
1248  *
1249  *	Create a fictitious page with the specified physical address and
1250  *	memory attribute.  The memory attribute is the only the machine-
1251  *	dependent aspect of a fictitious page that must be initialized.
1252  */
1253 vm_page_t
1254 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1255 {
1256 	vm_page_t m;
1257 
1258 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1259 	vm_page_initfake(m, paddr, memattr);
1260 	return (m);
1261 }
1262 
1263 void
1264 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1265 {
1266 
1267 	if ((m->flags & PG_FICTITIOUS) != 0) {
1268 		/*
1269 		 * The page's memattr might have changed since the
1270 		 * previous initialization.  Update the pmap to the
1271 		 * new memattr.
1272 		 */
1273 		goto memattr;
1274 	}
1275 	m->phys_addr = paddr;
1276 	m->a.queue = PQ_NONE;
1277 	/* Fictitious pages don't use "segind". */
1278 	m->flags = PG_FICTITIOUS;
1279 	/* Fictitious pages don't use "order" or "pool". */
1280 	m->oflags = VPO_UNMANAGED;
1281 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1282 	/* Fictitious pages are unevictable. */
1283 	m->ref_count = 1;
1284 	pmap_page_init(m);
1285 memattr:
1286 	pmap_page_set_memattr(m, memattr);
1287 }
1288 
1289 /*
1290  *	vm_page_putfake:
1291  *
1292  *	Release a fictitious page.
1293  */
1294 void
1295 vm_page_putfake(vm_page_t m)
1296 {
1297 
1298 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1299 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1300 	    ("vm_page_putfake: bad page %p", m));
1301 	vm_page_assert_xbusied(m);
1302 	vm_page_busy_free(m);
1303 	uma_zfree(fakepg_zone, m);
1304 }
1305 
1306 /*
1307  *	vm_page_updatefake:
1308  *
1309  *	Update the given fictitious page to the specified physical address and
1310  *	memory attribute.
1311  */
1312 void
1313 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1314 {
1315 
1316 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1317 	    ("vm_page_updatefake: bad page %p", m));
1318 	m->phys_addr = paddr;
1319 	pmap_page_set_memattr(m, memattr);
1320 }
1321 
1322 /*
1323  *	vm_page_free:
1324  *
1325  *	Free a page.
1326  */
1327 void
1328 vm_page_free(vm_page_t m)
1329 {
1330 
1331 	m->flags &= ~PG_ZERO;
1332 	vm_page_free_toq(m);
1333 }
1334 
1335 /*
1336  *	vm_page_free_zero:
1337  *
1338  *	Free a page to the zerod-pages queue
1339  */
1340 void
1341 vm_page_free_zero(vm_page_t m)
1342 {
1343 
1344 	m->flags |= PG_ZERO;
1345 	vm_page_free_toq(m);
1346 }
1347 
1348 /*
1349  * Unbusy and handle the page queueing for a page from a getpages request that
1350  * was optionally read ahead or behind.
1351  */
1352 void
1353 vm_page_readahead_finish(vm_page_t m)
1354 {
1355 
1356 	/* We shouldn't put invalid pages on queues. */
1357 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1358 
1359 	/*
1360 	 * Since the page is not the actually needed one, whether it should
1361 	 * be activated or deactivated is not obvious.  Empirical results
1362 	 * have shown that deactivating the page is usually the best choice,
1363 	 * unless the page is wanted by another thread.
1364 	 */
1365 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1366 		vm_page_activate(m);
1367 	else
1368 		vm_page_deactivate(m);
1369 	vm_page_xunbusy_unchecked(m);
1370 }
1371 
1372 /*
1373  * Destroy the identity of an invalid page and free it if possible.
1374  * This is intended to be used when reading a page from backing store fails.
1375  */
1376 void
1377 vm_page_free_invalid(vm_page_t m)
1378 {
1379 
1380 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1381 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1382 	KASSERT(m->object != NULL, ("page %p has no object", m));
1383 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1384 
1385 	/*
1386 	 * We may be attempting to free the page as part of the handling for an
1387 	 * I/O error, in which case the page was xbusied by a different thread.
1388 	 */
1389 	vm_page_xbusy_claim(m);
1390 
1391 	/*
1392 	 * If someone has wired this page while the object lock
1393 	 * was not held, then the thread that unwires is responsible
1394 	 * for freeing the page.  Otherwise just free the page now.
1395 	 * The wire count of this unmapped page cannot change while
1396 	 * we have the page xbusy and the page's object wlocked.
1397 	 */
1398 	if (vm_page_remove(m))
1399 		vm_page_free(m);
1400 }
1401 
1402 /*
1403  *	vm_page_dirty_KBI:		[ internal use only ]
1404  *
1405  *	Set all bits in the page's dirty field.
1406  *
1407  *	The object containing the specified page must be locked if the
1408  *	call is made from the machine-independent layer.
1409  *
1410  *	See vm_page_clear_dirty_mask().
1411  *
1412  *	This function should only be called by vm_page_dirty().
1413  */
1414 void
1415 vm_page_dirty_KBI(vm_page_t m)
1416 {
1417 
1418 	/* Refer to this operation by its public name. */
1419 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1420 	m->dirty = VM_PAGE_BITS_ALL;
1421 }
1422 
1423 /*
1424  *	vm_page_insert:		[ internal use only ]
1425  *
1426  *	Inserts the given mem entry into the object and object list.
1427  *
1428  *	The object must be locked.
1429  */
1430 int
1431 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1432 {
1433 	vm_page_t mpred;
1434 
1435 	VM_OBJECT_ASSERT_WLOCKED(object);
1436 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1437 	return (vm_page_insert_after(m, object, pindex, mpred));
1438 }
1439 
1440 /*
1441  *	vm_page_insert_after:
1442  *
1443  *	Inserts the page "m" into the specified object at offset "pindex".
1444  *
1445  *	The page "mpred" must immediately precede the offset "pindex" within
1446  *	the specified object.
1447  *
1448  *	The object must be locked.
1449  */
1450 static int
1451 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1452     vm_page_t mpred)
1453 {
1454 	vm_page_t msucc;
1455 
1456 	VM_OBJECT_ASSERT_WLOCKED(object);
1457 	KASSERT(m->object == NULL,
1458 	    ("vm_page_insert_after: page already inserted"));
1459 	if (mpred != NULL) {
1460 		KASSERT(mpred->object == object,
1461 		    ("vm_page_insert_after: object doesn't contain mpred"));
1462 		KASSERT(mpred->pindex < pindex,
1463 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1464 		msucc = TAILQ_NEXT(mpred, listq);
1465 	} else
1466 		msucc = TAILQ_FIRST(&object->memq);
1467 	if (msucc != NULL)
1468 		KASSERT(msucc->pindex > pindex,
1469 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1470 
1471 	/*
1472 	 * Record the object/offset pair in this page.
1473 	 */
1474 	m->object = object;
1475 	m->pindex = pindex;
1476 	m->ref_count |= VPRC_OBJREF;
1477 
1478 	/*
1479 	 * Now link into the object's ordered list of backed pages.
1480 	 */
1481 	if (vm_radix_insert(&object->rtree, m)) {
1482 		m->object = NULL;
1483 		m->pindex = 0;
1484 		m->ref_count &= ~VPRC_OBJREF;
1485 		return (1);
1486 	}
1487 	vm_page_insert_radixdone(m, object, mpred);
1488 	vm_pager_page_inserted(object, m);
1489 	return (0);
1490 }
1491 
1492 /*
1493  *	vm_page_insert_radixdone:
1494  *
1495  *	Complete page "m" insertion into the specified object after the
1496  *	radix trie hooking.
1497  *
1498  *	The page "mpred" must precede the offset "m->pindex" within the
1499  *	specified object.
1500  *
1501  *	The object must be locked.
1502  */
1503 static void
1504 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1505 {
1506 
1507 	VM_OBJECT_ASSERT_WLOCKED(object);
1508 	KASSERT(object != NULL && m->object == object,
1509 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1510 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1511 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1512 	if (mpred != NULL) {
1513 		KASSERT(mpred->object == object,
1514 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1515 		KASSERT(mpred->pindex < m->pindex,
1516 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1517 	}
1518 
1519 	if (mpred != NULL)
1520 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1521 	else
1522 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1523 
1524 	/*
1525 	 * Show that the object has one more resident page.
1526 	 */
1527 	object->resident_page_count++;
1528 
1529 	/*
1530 	 * Hold the vnode until the last page is released.
1531 	 */
1532 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1533 		vhold(object->handle);
1534 
1535 	/*
1536 	 * Since we are inserting a new and possibly dirty page,
1537 	 * update the object's generation count.
1538 	 */
1539 	if (pmap_page_is_write_mapped(m))
1540 		vm_object_set_writeable_dirty(object);
1541 }
1542 
1543 /*
1544  * Do the work to remove a page from its object.  The caller is responsible for
1545  * updating the page's fields to reflect this removal.
1546  */
1547 static void
1548 vm_page_object_remove(vm_page_t m)
1549 {
1550 	vm_object_t object;
1551 	vm_page_t mrem __diagused;
1552 
1553 	vm_page_assert_xbusied(m);
1554 	object = m->object;
1555 	VM_OBJECT_ASSERT_WLOCKED(object);
1556 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1557 	    ("page %p is missing its object ref", m));
1558 
1559 	/* Deferred free of swap space. */
1560 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1561 		vm_pager_page_unswapped(m);
1562 
1563 	vm_pager_page_removed(object, m);
1564 
1565 	m->object = NULL;
1566 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1567 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1568 
1569 	/*
1570 	 * Now remove from the object's list of backed pages.
1571 	 */
1572 	TAILQ_REMOVE(&object->memq, m, listq);
1573 
1574 	/*
1575 	 * And show that the object has one fewer resident page.
1576 	 */
1577 	object->resident_page_count--;
1578 
1579 	/*
1580 	 * The vnode may now be recycled.
1581 	 */
1582 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1583 		vdrop(object->handle);
1584 }
1585 
1586 /*
1587  *	vm_page_remove:
1588  *
1589  *	Removes the specified page from its containing object, but does not
1590  *	invalidate any backing storage.  Returns true if the object's reference
1591  *	was the last reference to the page, and false otherwise.
1592  *
1593  *	The object must be locked and the page must be exclusively busied.
1594  *	The exclusive busy will be released on return.  If this is not the
1595  *	final ref and the caller does not hold a wire reference it may not
1596  *	continue to access the page.
1597  */
1598 bool
1599 vm_page_remove(vm_page_t m)
1600 {
1601 	bool dropped;
1602 
1603 	dropped = vm_page_remove_xbusy(m);
1604 	vm_page_xunbusy(m);
1605 
1606 	return (dropped);
1607 }
1608 
1609 /*
1610  *	vm_page_remove_xbusy
1611  *
1612  *	Removes the page but leaves the xbusy held.  Returns true if this
1613  *	removed the final ref and false otherwise.
1614  */
1615 bool
1616 vm_page_remove_xbusy(vm_page_t m)
1617 {
1618 
1619 	vm_page_object_remove(m);
1620 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1621 }
1622 
1623 /*
1624  *	vm_page_lookup:
1625  *
1626  *	Returns the page associated with the object/offset
1627  *	pair specified; if none is found, NULL is returned.
1628  *
1629  *	The object must be locked.
1630  */
1631 vm_page_t
1632 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1633 {
1634 
1635 	VM_OBJECT_ASSERT_LOCKED(object);
1636 	return (vm_radix_lookup(&object->rtree, pindex));
1637 }
1638 
1639 /*
1640  *	vm_page_lookup_unlocked:
1641  *
1642  *	Returns the page associated with the object/offset pair specified;
1643  *	if none is found, NULL is returned.  The page may be no longer be
1644  *	present in the object at the time that this function returns.  Only
1645  *	useful for opportunistic checks such as inmem().
1646  */
1647 vm_page_t
1648 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1649 {
1650 
1651 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1652 }
1653 
1654 /*
1655  *	vm_page_relookup:
1656  *
1657  *	Returns a page that must already have been busied by
1658  *	the caller.  Used for bogus page replacement.
1659  */
1660 vm_page_t
1661 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1662 {
1663 	vm_page_t m;
1664 
1665 	m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1666 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1667 	    m->object == object && m->pindex == pindex,
1668 	    ("vm_page_relookup: Invalid page %p", m));
1669 	return (m);
1670 }
1671 
1672 /*
1673  * This should only be used by lockless functions for releasing transient
1674  * incorrect acquires.  The page may have been freed after we acquired a
1675  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1676  * further to do.
1677  */
1678 static void
1679 vm_page_busy_release(vm_page_t m)
1680 {
1681 	u_int x;
1682 
1683 	x = vm_page_busy_fetch(m);
1684 	for (;;) {
1685 		if (x == VPB_FREED)
1686 			break;
1687 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1688 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1689 			    x - VPB_ONE_SHARER))
1690 				break;
1691 			continue;
1692 		}
1693 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1694 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1695 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1696 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1697 			continue;
1698 		if ((x & VPB_BIT_WAITERS) != 0)
1699 			wakeup(m);
1700 		break;
1701 	}
1702 }
1703 
1704 /*
1705  *	vm_page_find_least:
1706  *
1707  *	Returns the page associated with the object with least pindex
1708  *	greater than or equal to the parameter pindex, or NULL.
1709  *
1710  *	The object must be locked.
1711  */
1712 vm_page_t
1713 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1714 {
1715 	vm_page_t m;
1716 
1717 	VM_OBJECT_ASSERT_LOCKED(object);
1718 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1719 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1720 	return (m);
1721 }
1722 
1723 /*
1724  * Returns the given page's successor (by pindex) within the object if it is
1725  * resident; if none is found, NULL is returned.
1726  *
1727  * The object must be locked.
1728  */
1729 vm_page_t
1730 vm_page_next(vm_page_t m)
1731 {
1732 	vm_page_t next;
1733 
1734 	VM_OBJECT_ASSERT_LOCKED(m->object);
1735 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1736 		MPASS(next->object == m->object);
1737 		if (next->pindex != m->pindex + 1)
1738 			next = NULL;
1739 	}
1740 	return (next);
1741 }
1742 
1743 /*
1744  * Returns the given page's predecessor (by pindex) within the object if it is
1745  * resident; if none is found, NULL is returned.
1746  *
1747  * The object must be locked.
1748  */
1749 vm_page_t
1750 vm_page_prev(vm_page_t m)
1751 {
1752 	vm_page_t prev;
1753 
1754 	VM_OBJECT_ASSERT_LOCKED(m->object);
1755 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1756 		MPASS(prev->object == m->object);
1757 		if (prev->pindex != m->pindex - 1)
1758 			prev = NULL;
1759 	}
1760 	return (prev);
1761 }
1762 
1763 /*
1764  * Uses the page mnew as a replacement for an existing page at index
1765  * pindex which must be already present in the object.
1766  *
1767  * Both pages must be exclusively busied on enter.  The old page is
1768  * unbusied on exit.
1769  *
1770  * A return value of true means mold is now free.  If this is not the
1771  * final ref and the caller does not hold a wire reference it may not
1772  * continue to access the page.
1773  */
1774 static bool
1775 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1776     vm_page_t mold)
1777 {
1778 	vm_page_t mret __diagused;
1779 	bool dropped;
1780 
1781 	VM_OBJECT_ASSERT_WLOCKED(object);
1782 	vm_page_assert_xbusied(mold);
1783 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1784 	    ("vm_page_replace: page %p already in object", mnew));
1785 
1786 	/*
1787 	 * This function mostly follows vm_page_insert() and
1788 	 * vm_page_remove() without the radix, object count and vnode
1789 	 * dance.  Double check such functions for more comments.
1790 	 */
1791 
1792 	mnew->object = object;
1793 	mnew->pindex = pindex;
1794 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1795 	mret = vm_radix_replace(&object->rtree, mnew);
1796 	KASSERT(mret == mold,
1797 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1798 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1799 	    (mnew->oflags & VPO_UNMANAGED),
1800 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1801 
1802 	/* Keep the resident page list in sorted order. */
1803 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1804 	TAILQ_REMOVE(&object->memq, mold, listq);
1805 	mold->object = NULL;
1806 
1807 	/*
1808 	 * The object's resident_page_count does not change because we have
1809 	 * swapped one page for another, but the generation count should
1810 	 * change if the page is dirty.
1811 	 */
1812 	if (pmap_page_is_write_mapped(mnew))
1813 		vm_object_set_writeable_dirty(object);
1814 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1815 	vm_page_xunbusy(mold);
1816 
1817 	return (dropped);
1818 }
1819 
1820 void
1821 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1822     vm_page_t mold)
1823 {
1824 
1825 	vm_page_assert_xbusied(mnew);
1826 
1827 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1828 		vm_page_free(mold);
1829 }
1830 
1831 /*
1832  *	vm_page_rename:
1833  *
1834  *	Move the given memory entry from its
1835  *	current object to the specified target object/offset.
1836  *
1837  *	Note: swap associated with the page must be invalidated by the move.  We
1838  *	      have to do this for several reasons:  (1) we aren't freeing the
1839  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1840  *	      moving the page from object A to B, and will then later move
1841  *	      the backing store from A to B and we can't have a conflict.
1842  *
1843  *	Note: we *always* dirty the page.  It is necessary both for the
1844  *	      fact that we moved it, and because we may be invalidating
1845  *	      swap.
1846  *
1847  *	The objects must be locked.
1848  */
1849 int
1850 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1851 {
1852 	vm_page_t mpred;
1853 	vm_pindex_t opidx;
1854 
1855 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1856 
1857 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1858 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1859 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1860 	    ("vm_page_rename: pindex already renamed"));
1861 
1862 	/*
1863 	 * Create a custom version of vm_page_insert() which does not depend
1864 	 * by m_prev and can cheat on the implementation aspects of the
1865 	 * function.
1866 	 */
1867 	opidx = m->pindex;
1868 	m->pindex = new_pindex;
1869 	if (vm_radix_insert(&new_object->rtree, m)) {
1870 		m->pindex = opidx;
1871 		return (1);
1872 	}
1873 
1874 	/*
1875 	 * The operation cannot fail anymore.  The removal must happen before
1876 	 * the listq iterator is tainted.
1877 	 */
1878 	m->pindex = opidx;
1879 	vm_page_object_remove(m);
1880 
1881 	/* Return back to the new pindex to complete vm_page_insert(). */
1882 	m->pindex = new_pindex;
1883 	m->object = new_object;
1884 
1885 	vm_page_insert_radixdone(m, new_object, mpred);
1886 	vm_page_dirty(m);
1887 	vm_pager_page_inserted(new_object, m);
1888 	return (0);
1889 }
1890 
1891 /*
1892  *	vm_page_alloc:
1893  *
1894  *	Allocate and return a page that is associated with the specified
1895  *	object and offset pair.  By default, this page is exclusive busied.
1896  *
1897  *	The caller must always specify an allocation class.
1898  *
1899  *	allocation classes:
1900  *	VM_ALLOC_NORMAL		normal process request
1901  *	VM_ALLOC_SYSTEM		system *really* needs a page
1902  *	VM_ALLOC_INTERRUPT	interrupt time request
1903  *
1904  *	optional allocation flags:
1905  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1906  *				intends to allocate
1907  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1908  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1909  *	VM_ALLOC_SBUSY		shared busy the allocated page
1910  *	VM_ALLOC_WIRED		wire the allocated page
1911  *	VM_ALLOC_ZERO		prefer a zeroed page
1912  */
1913 vm_page_t
1914 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1915 {
1916 
1917 	return (vm_page_alloc_after(object, pindex, req,
1918 	    vm_radix_lookup_le(&object->rtree, pindex)));
1919 }
1920 
1921 vm_page_t
1922 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1923     int req)
1924 {
1925 
1926 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1927 	    vm_radix_lookup_le(&object->rtree, pindex)));
1928 }
1929 
1930 /*
1931  * Allocate a page in the specified object with the given page index.  To
1932  * optimize insertion of the page into the object, the caller must also specifiy
1933  * the resident page in the object with largest index smaller than the given
1934  * page index, or NULL if no such page exists.
1935  */
1936 vm_page_t
1937 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1938     int req, vm_page_t mpred)
1939 {
1940 	struct vm_domainset_iter di;
1941 	vm_page_t m;
1942 	int domain;
1943 
1944 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1945 	do {
1946 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1947 		    mpred);
1948 		if (m != NULL)
1949 			break;
1950 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1951 
1952 	return (m);
1953 }
1954 
1955 /*
1956  * Returns true if the number of free pages exceeds the minimum
1957  * for the request class and false otherwise.
1958  */
1959 static int
1960 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1961 {
1962 	u_int limit, old, new;
1963 
1964 	if (req_class == VM_ALLOC_INTERRUPT)
1965 		limit = 0;
1966 	else if (req_class == VM_ALLOC_SYSTEM)
1967 		limit = vmd->vmd_interrupt_free_min;
1968 	else
1969 		limit = vmd->vmd_free_reserved;
1970 
1971 	/*
1972 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1973 	 */
1974 	limit += npages;
1975 	old = vmd->vmd_free_count;
1976 	do {
1977 		if (old < limit)
1978 			return (0);
1979 		new = old - npages;
1980 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1981 
1982 	/* Wake the page daemon if we've crossed the threshold. */
1983 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1984 		pagedaemon_wakeup(vmd->vmd_domain);
1985 
1986 	/* Only update bitsets on transitions. */
1987 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1988 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1989 		vm_domain_set(vmd);
1990 
1991 	return (1);
1992 }
1993 
1994 int
1995 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1996 {
1997 	int req_class;
1998 
1999 	/*
2000 	 * The page daemon is allowed to dig deeper into the free page list.
2001 	 */
2002 	req_class = req & VM_ALLOC_CLASS_MASK;
2003 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2004 		req_class = VM_ALLOC_SYSTEM;
2005 	return (_vm_domain_allocate(vmd, req_class, npages));
2006 }
2007 
2008 vm_page_t
2009 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2010     int req, vm_page_t mpred)
2011 {
2012 	struct vm_domain *vmd;
2013 	vm_page_t m;
2014 	int flags;
2015 
2016 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2017 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2018 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2019 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
2020 	KASSERT((req & ~VPA_FLAGS) == 0,
2021 	    ("invalid request %#x", req));
2022 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2023 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2024 	    ("invalid request %#x", req));
2025 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2026 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2027 	    (uintmax_t)pindex));
2028 	VM_OBJECT_ASSERT_WLOCKED(object);
2029 
2030 	flags = 0;
2031 	m = NULL;
2032 	if (!vm_pager_can_alloc_page(object, pindex))
2033 		return (NULL);
2034 again:
2035 #if VM_NRESERVLEVEL > 0
2036 	/*
2037 	 * Can we allocate the page from a reservation?
2038 	 */
2039 	if (vm_object_reserv(object) &&
2040 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2041 	    NULL) {
2042 		goto found;
2043 	}
2044 #endif
2045 	vmd = VM_DOMAIN(domain);
2046 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2047 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2048 		    M_NOWAIT | M_NOVM);
2049 		if (m != NULL) {
2050 			flags |= PG_PCPU_CACHE;
2051 			goto found;
2052 		}
2053 	}
2054 	if (vm_domain_allocate(vmd, req, 1)) {
2055 		/*
2056 		 * If not, allocate it from the free page queues.
2057 		 */
2058 		vm_domain_free_lock(vmd);
2059 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2060 		vm_domain_free_unlock(vmd);
2061 		if (m == NULL) {
2062 			vm_domain_freecnt_inc(vmd, 1);
2063 #if VM_NRESERVLEVEL > 0
2064 			if (vm_reserv_reclaim_inactive(domain))
2065 				goto again;
2066 #endif
2067 		}
2068 	}
2069 	if (m == NULL) {
2070 		/*
2071 		 * Not allocatable, give up.
2072 		 */
2073 		if (vm_domain_alloc_fail(vmd, object, req))
2074 			goto again;
2075 		return (NULL);
2076 	}
2077 
2078 	/*
2079 	 * At this point we had better have found a good page.
2080 	 */
2081 found:
2082 	vm_page_dequeue(m);
2083 	vm_page_alloc_check(m);
2084 
2085 	/*
2086 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2087 	 */
2088 	flags |= m->flags & PG_ZERO;
2089 	if ((req & VM_ALLOC_NODUMP) != 0)
2090 		flags |= PG_NODUMP;
2091 	m->flags = flags;
2092 	m->a.flags = 0;
2093 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2094 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2095 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2096 	else if ((req & VM_ALLOC_SBUSY) != 0)
2097 		m->busy_lock = VPB_SHARERS_WORD(1);
2098 	else
2099 		m->busy_lock = VPB_UNBUSIED;
2100 	if (req & VM_ALLOC_WIRED) {
2101 		vm_wire_add(1);
2102 		m->ref_count = 1;
2103 	}
2104 	m->a.act_count = 0;
2105 
2106 	if (vm_page_insert_after(m, object, pindex, mpred)) {
2107 		if (req & VM_ALLOC_WIRED) {
2108 			vm_wire_sub(1);
2109 			m->ref_count = 0;
2110 		}
2111 		KASSERT(m->object == NULL, ("page %p has object", m));
2112 		m->oflags = VPO_UNMANAGED;
2113 		m->busy_lock = VPB_UNBUSIED;
2114 		/* Don't change PG_ZERO. */
2115 		vm_page_free_toq(m);
2116 		if (req & VM_ALLOC_WAITFAIL) {
2117 			VM_OBJECT_WUNLOCK(object);
2118 			vm_radix_wait();
2119 			VM_OBJECT_WLOCK(object);
2120 		}
2121 		return (NULL);
2122 	}
2123 
2124 	/* Ignore device objects; the pager sets "memattr" for them. */
2125 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2126 	    (object->flags & OBJ_FICTITIOUS) == 0)
2127 		pmap_page_set_memattr(m, object->memattr);
2128 
2129 	return (m);
2130 }
2131 
2132 /*
2133  *	vm_page_alloc_contig:
2134  *
2135  *	Allocate a contiguous set of physical pages of the given size "npages"
2136  *	from the free lists.  All of the physical pages must be at or above
2137  *	the given physical address "low" and below the given physical address
2138  *	"high".  The given value "alignment" determines the alignment of the
2139  *	first physical page in the set.  If the given value "boundary" is
2140  *	non-zero, then the set of physical pages cannot cross any physical
2141  *	address boundary that is a multiple of that value.  Both "alignment"
2142  *	and "boundary" must be a power of two.
2143  *
2144  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2145  *	then the memory attribute setting for the physical pages is configured
2146  *	to the object's memory attribute setting.  Otherwise, the memory
2147  *	attribute setting for the physical pages is configured to "memattr",
2148  *	overriding the object's memory attribute setting.  However, if the
2149  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2150  *	memory attribute setting for the physical pages cannot be configured
2151  *	to VM_MEMATTR_DEFAULT.
2152  *
2153  *	The specified object may not contain fictitious pages.
2154  *
2155  *	The caller must always specify an allocation class.
2156  *
2157  *	allocation classes:
2158  *	VM_ALLOC_NORMAL		normal process request
2159  *	VM_ALLOC_SYSTEM		system *really* needs a page
2160  *	VM_ALLOC_INTERRUPT	interrupt time request
2161  *
2162  *	optional allocation flags:
2163  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2164  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2165  *	VM_ALLOC_SBUSY		shared busy the allocated page
2166  *	VM_ALLOC_WIRED		wire the allocated page
2167  *	VM_ALLOC_ZERO		prefer a zeroed page
2168  */
2169 vm_page_t
2170 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2171     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2172     vm_paddr_t boundary, vm_memattr_t memattr)
2173 {
2174 	struct vm_domainset_iter di;
2175 	vm_page_t m;
2176 	int domain;
2177 
2178 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2179 	do {
2180 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2181 		    npages, low, high, alignment, boundary, memattr);
2182 		if (m != NULL)
2183 			break;
2184 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2185 
2186 	return (m);
2187 }
2188 
2189 static vm_page_t
2190 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2191     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2192 {
2193 	struct vm_domain *vmd;
2194 	vm_page_t m_ret;
2195 
2196 	/*
2197 	 * Can we allocate the pages without the number of free pages falling
2198 	 * below the lower bound for the allocation class?
2199 	 */
2200 	vmd = VM_DOMAIN(domain);
2201 	if (!vm_domain_allocate(vmd, req, npages))
2202 		return (NULL);
2203 	/*
2204 	 * Try to allocate the pages from the free page queues.
2205 	 */
2206 	vm_domain_free_lock(vmd);
2207 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2208 	    alignment, boundary);
2209 	vm_domain_free_unlock(vmd);
2210 	if (m_ret != NULL)
2211 		return (m_ret);
2212 #if VM_NRESERVLEVEL > 0
2213 	/*
2214 	 * Try to break a reservation to allocate the pages.
2215 	 */
2216 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2217 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2218 	            high, alignment, boundary);
2219 		if (m_ret != NULL)
2220 			return (m_ret);
2221 	}
2222 #endif
2223 	vm_domain_freecnt_inc(vmd, npages);
2224 	return (NULL);
2225 }
2226 
2227 vm_page_t
2228 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2229     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2230     vm_paddr_t boundary, vm_memattr_t memattr)
2231 {
2232 	vm_page_t m, m_ret, mpred;
2233 	u_int busy_lock, flags, oflags;
2234 
2235 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2236 	KASSERT((req & ~VPAC_FLAGS) == 0,
2237 	    ("invalid request %#x", req));
2238 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2239 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2240 	    ("invalid request %#x", req));
2241 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2242 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2243 	    ("invalid request %#x", req));
2244 	VM_OBJECT_ASSERT_WLOCKED(object);
2245 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2246 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2247 	    object));
2248 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2249 
2250 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
2251 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2252 	    ("vm_page_alloc_contig: pindex already allocated"));
2253 	for (;;) {
2254 #if VM_NRESERVLEVEL > 0
2255 		/*
2256 		 * Can we allocate the pages from a reservation?
2257 		 */
2258 		if (vm_object_reserv(object) &&
2259 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2260 		    mpred, npages, low, high, alignment, boundary)) != NULL) {
2261 			break;
2262 		}
2263 #endif
2264 		if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2265 		    low, high, alignment, boundary)) != NULL)
2266 			break;
2267 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2268 			return (NULL);
2269 	}
2270 	for (m = m_ret; m < &m_ret[npages]; m++) {
2271 		vm_page_dequeue(m);
2272 		vm_page_alloc_check(m);
2273 	}
2274 
2275 	/*
2276 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2277 	 */
2278 	flags = PG_ZERO;
2279 	if ((req & VM_ALLOC_NODUMP) != 0)
2280 		flags |= PG_NODUMP;
2281 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2282 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2283 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2284 	else if ((req & VM_ALLOC_SBUSY) != 0)
2285 		busy_lock = VPB_SHARERS_WORD(1);
2286 	else
2287 		busy_lock = VPB_UNBUSIED;
2288 	if ((req & VM_ALLOC_WIRED) != 0)
2289 		vm_wire_add(npages);
2290 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2291 	    memattr == VM_MEMATTR_DEFAULT)
2292 		memattr = object->memattr;
2293 	for (m = m_ret; m < &m_ret[npages]; m++) {
2294 		m->a.flags = 0;
2295 		m->flags = (m->flags | PG_NODUMP) & flags;
2296 		m->busy_lock = busy_lock;
2297 		if ((req & VM_ALLOC_WIRED) != 0)
2298 			m->ref_count = 1;
2299 		m->a.act_count = 0;
2300 		m->oflags = oflags;
2301 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2302 			if ((req & VM_ALLOC_WIRED) != 0)
2303 				vm_wire_sub(npages);
2304 			KASSERT(m->object == NULL,
2305 			    ("page %p has object", m));
2306 			mpred = m;
2307 			for (m = m_ret; m < &m_ret[npages]; m++) {
2308 				if (m <= mpred &&
2309 				    (req & VM_ALLOC_WIRED) != 0)
2310 					m->ref_count = 0;
2311 				m->oflags = VPO_UNMANAGED;
2312 				m->busy_lock = VPB_UNBUSIED;
2313 				/* Don't change PG_ZERO. */
2314 				vm_page_free_toq(m);
2315 			}
2316 			if (req & VM_ALLOC_WAITFAIL) {
2317 				VM_OBJECT_WUNLOCK(object);
2318 				vm_radix_wait();
2319 				VM_OBJECT_WLOCK(object);
2320 			}
2321 			return (NULL);
2322 		}
2323 		mpred = m;
2324 		if (memattr != VM_MEMATTR_DEFAULT)
2325 			pmap_page_set_memattr(m, memattr);
2326 		pindex++;
2327 	}
2328 	return (m_ret);
2329 }
2330 
2331 /*
2332  * Allocate a physical page that is not intended to be inserted into a VM
2333  * object.  If the "freelist" parameter is not equal to VM_NFREELIST, then only
2334  * pages from the specified vm_phys freelist will be returned.
2335  */
2336 static __always_inline vm_page_t
2337 _vm_page_alloc_noobj_domain(int domain, const int freelist, int req)
2338 {
2339 	struct vm_domain *vmd;
2340 	vm_page_t m;
2341 	int flags;
2342 
2343 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2344 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2345 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2346 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
2347 	KASSERT((req & ~VPAN_FLAGS) == 0,
2348 	    ("invalid request %#x", req));
2349 
2350 	flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0;
2351 	vmd = VM_DOMAIN(domain);
2352 again:
2353 	if (freelist == VM_NFREELIST &&
2354 	    vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2355 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2356 		    M_NOWAIT | M_NOVM);
2357 		if (m != NULL) {
2358 			flags |= PG_PCPU_CACHE;
2359 			goto found;
2360 		}
2361 	}
2362 
2363 	if (vm_domain_allocate(vmd, req, 1)) {
2364 		vm_domain_free_lock(vmd);
2365 		if (freelist == VM_NFREELIST)
2366 			m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2367 		else
2368 			m = vm_phys_alloc_freelist_pages(domain, freelist,
2369 			    VM_FREEPOOL_DIRECT, 0);
2370 		vm_domain_free_unlock(vmd);
2371 		if (m == NULL) {
2372 			vm_domain_freecnt_inc(vmd, 1);
2373 #if VM_NRESERVLEVEL > 0
2374 			if (freelist == VM_NFREELIST &&
2375 			    vm_reserv_reclaim_inactive(domain))
2376 				goto again;
2377 #endif
2378 		}
2379 	}
2380 	if (m == NULL) {
2381 		if (vm_domain_alloc_fail(vmd, NULL, req))
2382 			goto again;
2383 		return (NULL);
2384 	}
2385 
2386 found:
2387 	vm_page_dequeue(m);
2388 	vm_page_alloc_check(m);
2389 
2390 	/*
2391 	 * Consumers should not rely on a useful default pindex value.
2392 	 */
2393 	m->pindex = 0xdeadc0dedeadc0de;
2394 	m->flags = (m->flags & PG_ZERO) | flags;
2395 	m->a.flags = 0;
2396 	m->oflags = VPO_UNMANAGED;
2397 	m->busy_lock = VPB_UNBUSIED;
2398 	if ((req & VM_ALLOC_WIRED) != 0) {
2399 		vm_wire_add(1);
2400 		m->ref_count = 1;
2401 	}
2402 
2403 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2404 		pmap_zero_page(m);
2405 
2406 	return (m);
2407 }
2408 
2409 vm_page_t
2410 vm_page_alloc_freelist(int freelist, int req)
2411 {
2412 	struct vm_domainset_iter di;
2413 	vm_page_t m;
2414 	int domain;
2415 
2416 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2417 	do {
2418 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2419 		if (m != NULL)
2420 			break;
2421 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2422 
2423 	return (m);
2424 }
2425 
2426 vm_page_t
2427 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2428 {
2429 	KASSERT(freelist >= 0 && freelist < VM_NFREELIST,
2430 	    ("%s: invalid freelist %d", __func__, freelist));
2431 
2432 	return (_vm_page_alloc_noobj_domain(domain, freelist, req));
2433 }
2434 
2435 vm_page_t
2436 vm_page_alloc_noobj(int req)
2437 {
2438 	struct vm_domainset_iter di;
2439 	vm_page_t m;
2440 	int domain;
2441 
2442 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2443 	do {
2444 		m = vm_page_alloc_noobj_domain(domain, req);
2445 		if (m != NULL)
2446 			break;
2447 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2448 
2449 	return (m);
2450 }
2451 
2452 vm_page_t
2453 vm_page_alloc_noobj_domain(int domain, int req)
2454 {
2455 	return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req));
2456 }
2457 
2458 vm_page_t
2459 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2460     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2461     vm_memattr_t memattr)
2462 {
2463 	struct vm_domainset_iter di;
2464 	vm_page_t m;
2465 	int domain;
2466 
2467 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2468 	do {
2469 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2470 		    high, alignment, boundary, memattr);
2471 		if (m != NULL)
2472 			break;
2473 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2474 
2475 	return (m);
2476 }
2477 
2478 vm_page_t
2479 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2480     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2481     vm_memattr_t memattr)
2482 {
2483 	vm_page_t m, m_ret;
2484 	u_int flags;
2485 
2486 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2487 	KASSERT((req & ~VPANC_FLAGS) == 0,
2488 	    ("invalid request %#x", req));
2489 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2490 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2491 	    ("invalid request %#x", req));
2492 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2493 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2494 	    ("invalid request %#x", req));
2495 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2496 
2497 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2498 	    low, high, alignment, boundary)) == NULL) {
2499 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2500 			return (NULL);
2501 	}
2502 
2503 	/*
2504 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2505 	 */
2506 	flags = PG_ZERO;
2507 	if ((req & VM_ALLOC_NODUMP) != 0)
2508 		flags |= PG_NODUMP;
2509 	if ((req & VM_ALLOC_WIRED) != 0)
2510 		vm_wire_add(npages);
2511 	for (m = m_ret; m < &m_ret[npages]; m++) {
2512 		vm_page_dequeue(m);
2513 		vm_page_alloc_check(m);
2514 
2515 		/*
2516 		 * Consumers should not rely on a useful default pindex value.
2517 		 */
2518 		m->pindex = 0xdeadc0dedeadc0de;
2519 		m->a.flags = 0;
2520 		m->flags = (m->flags | PG_NODUMP) & flags;
2521 		m->busy_lock = VPB_UNBUSIED;
2522 		if ((req & VM_ALLOC_WIRED) != 0)
2523 			m->ref_count = 1;
2524 		m->a.act_count = 0;
2525 		m->oflags = VPO_UNMANAGED;
2526 
2527 		/*
2528 		 * Zero the page before updating any mappings since the page is
2529 		 * not yet shared with any devices which might require the
2530 		 * non-default memory attribute.  pmap_page_set_memattr()
2531 		 * flushes data caches before returning.
2532 		 */
2533 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2534 			pmap_zero_page(m);
2535 		if (memattr != VM_MEMATTR_DEFAULT)
2536 			pmap_page_set_memattr(m, memattr);
2537 	}
2538 	return (m_ret);
2539 }
2540 
2541 /*
2542  * Check a page that has been freshly dequeued from a freelist.
2543  */
2544 static void
2545 vm_page_alloc_check(vm_page_t m)
2546 {
2547 
2548 	KASSERT(m->object == NULL, ("page %p has object", m));
2549 	KASSERT(m->a.queue == PQ_NONE &&
2550 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2551 	    ("page %p has unexpected queue %d, flags %#x",
2552 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2553 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2554 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2555 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2556 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2557 	    ("page %p has unexpected memattr %d",
2558 	    m, pmap_page_get_memattr(m)));
2559 	KASSERT(vm_page_none_valid(m), ("free page %p is valid", m));
2560 	pmap_vm_page_alloc_check(m);
2561 }
2562 
2563 static int
2564 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2565 {
2566 	struct vm_domain *vmd;
2567 	struct vm_pgcache *pgcache;
2568 	int i;
2569 
2570 	pgcache = arg;
2571 	vmd = VM_DOMAIN(pgcache->domain);
2572 
2573 	/*
2574 	 * The page daemon should avoid creating extra memory pressure since its
2575 	 * main purpose is to replenish the store of free pages.
2576 	 */
2577 	if (vmd->vmd_severeset || curproc == pageproc ||
2578 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2579 		return (0);
2580 	domain = vmd->vmd_domain;
2581 	vm_domain_free_lock(vmd);
2582 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2583 	    (vm_page_t *)store);
2584 	vm_domain_free_unlock(vmd);
2585 	if (cnt != i)
2586 		vm_domain_freecnt_inc(vmd, cnt - i);
2587 
2588 	return (i);
2589 }
2590 
2591 static void
2592 vm_page_zone_release(void *arg, void **store, int cnt)
2593 {
2594 	struct vm_domain *vmd;
2595 	struct vm_pgcache *pgcache;
2596 	vm_page_t m;
2597 	int i;
2598 
2599 	pgcache = arg;
2600 	vmd = VM_DOMAIN(pgcache->domain);
2601 	vm_domain_free_lock(vmd);
2602 	for (i = 0; i < cnt; i++) {
2603 		m = (vm_page_t)store[i];
2604 		vm_phys_free_pages(m, 0);
2605 	}
2606 	vm_domain_free_unlock(vmd);
2607 	vm_domain_freecnt_inc(vmd, cnt);
2608 }
2609 
2610 #define	VPSC_ANY	0	/* No restrictions. */
2611 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2612 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2613 
2614 /*
2615  *	vm_page_scan_contig:
2616  *
2617  *	Scan vm_page_array[] between the specified entries "m_start" and
2618  *	"m_end" for a run of contiguous physical pages that satisfy the
2619  *	specified conditions, and return the lowest page in the run.  The
2620  *	specified "alignment" determines the alignment of the lowest physical
2621  *	page in the run.  If the specified "boundary" is non-zero, then the
2622  *	run of physical pages cannot span a physical address that is a
2623  *	multiple of "boundary".
2624  *
2625  *	"m_end" is never dereferenced, so it need not point to a vm_page
2626  *	structure within vm_page_array[].
2627  *
2628  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2629  *	span a hole (or discontiguity) in the physical address space.  Both
2630  *	"alignment" and "boundary" must be a power of two.
2631  */
2632 static vm_page_t
2633 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2634     u_long alignment, vm_paddr_t boundary, int options)
2635 {
2636 	vm_object_t object;
2637 	vm_paddr_t pa;
2638 	vm_page_t m, m_run;
2639 #if VM_NRESERVLEVEL > 0
2640 	int level;
2641 #endif
2642 	int m_inc, order, run_ext, run_len;
2643 
2644 	KASSERT(npages > 0, ("npages is 0"));
2645 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2646 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2647 	m_run = NULL;
2648 	run_len = 0;
2649 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2650 		KASSERT((m->flags & PG_MARKER) == 0,
2651 		    ("page %p is PG_MARKER", m));
2652 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2653 		    ("fictitious page %p has invalid ref count", m));
2654 
2655 		/*
2656 		 * If the current page would be the start of a run, check its
2657 		 * physical address against the end, alignment, and boundary
2658 		 * conditions.  If it doesn't satisfy these conditions, either
2659 		 * terminate the scan or advance to the next page that
2660 		 * satisfies the failed condition.
2661 		 */
2662 		if (run_len == 0) {
2663 			KASSERT(m_run == NULL, ("m_run != NULL"));
2664 			if (m + npages > m_end)
2665 				break;
2666 			pa = VM_PAGE_TO_PHYS(m);
2667 			if (!vm_addr_align_ok(pa, alignment)) {
2668 				m_inc = atop(roundup2(pa, alignment) - pa);
2669 				continue;
2670 			}
2671 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2672 				m_inc = atop(roundup2(pa, boundary) - pa);
2673 				continue;
2674 			}
2675 		} else
2676 			KASSERT(m_run != NULL, ("m_run == NULL"));
2677 
2678 retry:
2679 		m_inc = 1;
2680 		if (vm_page_wired(m))
2681 			run_ext = 0;
2682 #if VM_NRESERVLEVEL > 0
2683 		else if ((level = vm_reserv_level(m)) >= 0 &&
2684 		    (options & VPSC_NORESERV) != 0) {
2685 			run_ext = 0;
2686 			/* Advance to the end of the reservation. */
2687 			pa = VM_PAGE_TO_PHYS(m);
2688 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2689 			    pa);
2690 		}
2691 #endif
2692 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2693 			/*
2694 			 * The page is considered eligible for relocation if
2695 			 * and only if it could be laundered or reclaimed by
2696 			 * the page daemon.
2697 			 */
2698 			VM_OBJECT_RLOCK(object);
2699 			if (object != m->object) {
2700 				VM_OBJECT_RUNLOCK(object);
2701 				goto retry;
2702 			}
2703 			/* Don't care: PG_NODUMP, PG_ZERO. */
2704 			if ((object->flags & OBJ_SWAP) == 0 &&
2705 			    object->type != OBJT_VNODE) {
2706 				run_ext = 0;
2707 #if VM_NRESERVLEVEL > 0
2708 			} else if ((options & VPSC_NOSUPER) != 0 &&
2709 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2710 				run_ext = 0;
2711 				/* Advance to the end of the superpage. */
2712 				pa = VM_PAGE_TO_PHYS(m);
2713 				m_inc = atop(roundup2(pa + 1,
2714 				    vm_reserv_size(level)) - pa);
2715 #endif
2716 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2717 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2718 				/*
2719 				 * The page is allocated but eligible for
2720 				 * relocation.  Extend the current run by one
2721 				 * page.
2722 				 */
2723 				KASSERT(pmap_page_get_memattr(m) ==
2724 				    VM_MEMATTR_DEFAULT,
2725 				    ("page %p has an unexpected memattr", m));
2726 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2727 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2728 				    ("page %p has unexpected oflags", m));
2729 				/* Don't care: PGA_NOSYNC. */
2730 				run_ext = 1;
2731 			} else
2732 				run_ext = 0;
2733 			VM_OBJECT_RUNLOCK(object);
2734 #if VM_NRESERVLEVEL > 0
2735 		} else if (level >= 0) {
2736 			/*
2737 			 * The page is reserved but not yet allocated.  In
2738 			 * other words, it is still free.  Extend the current
2739 			 * run by one page.
2740 			 */
2741 			run_ext = 1;
2742 #endif
2743 		} else if ((order = m->order) < VM_NFREEORDER) {
2744 			/*
2745 			 * The page is enqueued in the physical memory
2746 			 * allocator's free page queues.  Moreover, it is the
2747 			 * first page in a power-of-two-sized run of
2748 			 * contiguous free pages.  Add these pages to the end
2749 			 * of the current run, and jump ahead.
2750 			 */
2751 			run_ext = 1 << order;
2752 			m_inc = 1 << order;
2753 		} else {
2754 			/*
2755 			 * Skip the page for one of the following reasons: (1)
2756 			 * It is enqueued in the physical memory allocator's
2757 			 * free page queues.  However, it is not the first
2758 			 * page in a run of contiguous free pages.  (This case
2759 			 * rarely occurs because the scan is performed in
2760 			 * ascending order.) (2) It is not reserved, and it is
2761 			 * transitioning from free to allocated.  (Conversely,
2762 			 * the transition from allocated to free for managed
2763 			 * pages is blocked by the page busy lock.) (3) It is
2764 			 * allocated but not contained by an object and not
2765 			 * wired, e.g., allocated by Xen's balloon driver.
2766 			 */
2767 			run_ext = 0;
2768 		}
2769 
2770 		/*
2771 		 * Extend or reset the current run of pages.
2772 		 */
2773 		if (run_ext > 0) {
2774 			if (run_len == 0)
2775 				m_run = m;
2776 			run_len += run_ext;
2777 		} else {
2778 			if (run_len > 0) {
2779 				m_run = NULL;
2780 				run_len = 0;
2781 			}
2782 		}
2783 	}
2784 	if (run_len >= npages)
2785 		return (m_run);
2786 	return (NULL);
2787 }
2788 
2789 /*
2790  *	vm_page_reclaim_run:
2791  *
2792  *	Try to relocate each of the allocated virtual pages within the
2793  *	specified run of physical pages to a new physical address.  Free the
2794  *	physical pages underlying the relocated virtual pages.  A virtual page
2795  *	is relocatable if and only if it could be laundered or reclaimed by
2796  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2797  *	physical address above "high".
2798  *
2799  *	Returns 0 if every physical page within the run was already free or
2800  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2801  *	value indicating why the last attempt to relocate a virtual page was
2802  *	unsuccessful.
2803  *
2804  *	"req_class" must be an allocation class.
2805  */
2806 static int
2807 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2808     vm_paddr_t high)
2809 {
2810 	struct vm_domain *vmd;
2811 	struct spglist free;
2812 	vm_object_t object;
2813 	vm_paddr_t pa;
2814 	vm_page_t m, m_end, m_new;
2815 	int error, order, req;
2816 
2817 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2818 	    ("req_class is not an allocation class"));
2819 	SLIST_INIT(&free);
2820 	error = 0;
2821 	m = m_run;
2822 	m_end = m_run + npages;
2823 	for (; error == 0 && m < m_end; m++) {
2824 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2825 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2826 
2827 		/*
2828 		 * Racily check for wirings.  Races are handled once the object
2829 		 * lock is held and the page is unmapped.
2830 		 */
2831 		if (vm_page_wired(m))
2832 			error = EBUSY;
2833 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2834 			/*
2835 			 * The page is relocated if and only if it could be
2836 			 * laundered or reclaimed by the page daemon.
2837 			 */
2838 			VM_OBJECT_WLOCK(object);
2839 			/* Don't care: PG_NODUMP, PG_ZERO. */
2840 			if (m->object != object ||
2841 			    ((object->flags & OBJ_SWAP) == 0 &&
2842 			    object->type != OBJT_VNODE))
2843 				error = EINVAL;
2844 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2845 				error = EINVAL;
2846 			else if (vm_page_queue(m) != PQ_NONE &&
2847 			    vm_page_tryxbusy(m) != 0) {
2848 				if (vm_page_wired(m)) {
2849 					vm_page_xunbusy(m);
2850 					error = EBUSY;
2851 					goto unlock;
2852 				}
2853 				KASSERT(pmap_page_get_memattr(m) ==
2854 				    VM_MEMATTR_DEFAULT,
2855 				    ("page %p has an unexpected memattr", m));
2856 				KASSERT(m->oflags == 0,
2857 				    ("page %p has unexpected oflags", m));
2858 				/* Don't care: PGA_NOSYNC. */
2859 				if (!vm_page_none_valid(m)) {
2860 					/*
2861 					 * First, try to allocate a new page
2862 					 * that is above "high".  Failing
2863 					 * that, try to allocate a new page
2864 					 * that is below "m_run".  Allocate
2865 					 * the new page between the end of
2866 					 * "m_run" and "high" only as a last
2867 					 * resort.
2868 					 */
2869 					req = req_class;
2870 					if ((m->flags & PG_NODUMP) != 0)
2871 						req |= VM_ALLOC_NODUMP;
2872 					if (trunc_page(high) !=
2873 					    ~(vm_paddr_t)PAGE_MASK) {
2874 						m_new =
2875 						    vm_page_alloc_noobj_contig(
2876 						    req, 1, round_page(high),
2877 						    ~(vm_paddr_t)0, PAGE_SIZE,
2878 						    0, VM_MEMATTR_DEFAULT);
2879 					} else
2880 						m_new = NULL;
2881 					if (m_new == NULL) {
2882 						pa = VM_PAGE_TO_PHYS(m_run);
2883 						m_new =
2884 						    vm_page_alloc_noobj_contig(
2885 						    req, 1, 0, pa - 1,
2886 						    PAGE_SIZE, 0,
2887 						    VM_MEMATTR_DEFAULT);
2888 					}
2889 					if (m_new == NULL) {
2890 						pa += ptoa(npages);
2891 						m_new =
2892 						    vm_page_alloc_noobj_contig(
2893 						    req, 1, pa, high, PAGE_SIZE,
2894 						    0, VM_MEMATTR_DEFAULT);
2895 					}
2896 					if (m_new == NULL) {
2897 						vm_page_xunbusy(m);
2898 						error = ENOMEM;
2899 						goto unlock;
2900 					}
2901 
2902 					/*
2903 					 * Unmap the page and check for new
2904 					 * wirings that may have been acquired
2905 					 * through a pmap lookup.
2906 					 */
2907 					if (object->ref_count != 0 &&
2908 					    !vm_page_try_remove_all(m)) {
2909 						vm_page_xunbusy(m);
2910 						vm_page_free(m_new);
2911 						error = EBUSY;
2912 						goto unlock;
2913 					}
2914 
2915 					/*
2916 					 * Replace "m" with the new page.  For
2917 					 * vm_page_replace(), "m" must be busy
2918 					 * and dequeued.  Finally, change "m"
2919 					 * as if vm_page_free() was called.
2920 					 */
2921 					m_new->a.flags = m->a.flags &
2922 					    ~PGA_QUEUE_STATE_MASK;
2923 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2924 					    ("page %p is managed", m_new));
2925 					m_new->oflags = 0;
2926 					pmap_copy_page(m, m_new);
2927 					m_new->valid = m->valid;
2928 					m_new->dirty = m->dirty;
2929 					m->flags &= ~PG_ZERO;
2930 					vm_page_dequeue(m);
2931 					if (vm_page_replace_hold(m_new, object,
2932 					    m->pindex, m) &&
2933 					    vm_page_free_prep(m))
2934 						SLIST_INSERT_HEAD(&free, m,
2935 						    plinks.s.ss);
2936 
2937 					/*
2938 					 * The new page must be deactivated
2939 					 * before the object is unlocked.
2940 					 */
2941 					vm_page_deactivate(m_new);
2942 				} else {
2943 					m->flags &= ~PG_ZERO;
2944 					vm_page_dequeue(m);
2945 					if (vm_page_free_prep(m))
2946 						SLIST_INSERT_HEAD(&free, m,
2947 						    plinks.s.ss);
2948 					KASSERT(m->dirty == 0,
2949 					    ("page %p is dirty", m));
2950 				}
2951 			} else
2952 				error = EBUSY;
2953 unlock:
2954 			VM_OBJECT_WUNLOCK(object);
2955 		} else {
2956 			MPASS(vm_page_domain(m) == domain);
2957 			vmd = VM_DOMAIN(domain);
2958 			vm_domain_free_lock(vmd);
2959 			order = m->order;
2960 			if (order < VM_NFREEORDER) {
2961 				/*
2962 				 * The page is enqueued in the physical memory
2963 				 * allocator's free page queues.  Moreover, it
2964 				 * is the first page in a power-of-two-sized
2965 				 * run of contiguous free pages.  Jump ahead
2966 				 * to the last page within that run, and
2967 				 * continue from there.
2968 				 */
2969 				m += (1 << order) - 1;
2970 			}
2971 #if VM_NRESERVLEVEL > 0
2972 			else if (vm_reserv_is_page_free(m))
2973 				order = 0;
2974 #endif
2975 			vm_domain_free_unlock(vmd);
2976 			if (order == VM_NFREEORDER)
2977 				error = EINVAL;
2978 		}
2979 	}
2980 	if ((m = SLIST_FIRST(&free)) != NULL) {
2981 		int cnt;
2982 
2983 		vmd = VM_DOMAIN(domain);
2984 		cnt = 0;
2985 		vm_domain_free_lock(vmd);
2986 		do {
2987 			MPASS(vm_page_domain(m) == domain);
2988 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2989 			vm_phys_free_pages(m, 0);
2990 			cnt++;
2991 		} while ((m = SLIST_FIRST(&free)) != NULL);
2992 		vm_domain_free_unlock(vmd);
2993 		vm_domain_freecnt_inc(vmd, cnt);
2994 	}
2995 	return (error);
2996 }
2997 
2998 #define	NRUNS	16
2999 
3000 #define	RUN_INDEX(count, nruns)	((count) % (nruns))
3001 
3002 #define	MIN_RECLAIM	8
3003 
3004 /*
3005  *	vm_page_reclaim_contig:
3006  *
3007  *	Reclaim allocated, contiguous physical memory satisfying the specified
3008  *	conditions by relocating the virtual pages using that physical memory.
3009  *	Returns true if reclamation is successful and false otherwise.  Since
3010  *	relocation requires the allocation of physical pages, reclamation may
3011  *	fail due to a shortage of free pages.  When reclamation fails, callers
3012  *	are expected to perform vm_wait() before retrying a failed allocation
3013  *	operation, e.g., vm_page_alloc_contig().
3014  *
3015  *	The caller must always specify an allocation class through "req".
3016  *
3017  *	allocation classes:
3018  *	VM_ALLOC_NORMAL		normal process request
3019  *	VM_ALLOC_SYSTEM		system *really* needs a page
3020  *	VM_ALLOC_INTERRUPT	interrupt time request
3021  *
3022  *	The optional allocation flags are ignored.
3023  *
3024  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3025  *	must be a power of two.
3026  */
3027 bool
3028 vm_page_reclaim_contig_domain_ext(int domain, int req, u_long npages,
3029     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
3030     int desired_runs)
3031 {
3032 	struct vm_domain *vmd;
3033 	vm_page_t bounds[2], m_run, _m_runs[NRUNS], *m_runs;
3034 	u_long count, minalign, reclaimed;
3035 	int error, i, min_reclaim, nruns, options, req_class, segind;
3036 	bool ret;
3037 
3038 	KASSERT(npages > 0, ("npages is 0"));
3039 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3040 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3041 
3042 	ret = false;
3043 
3044 	/*
3045 	 * If the caller wants to reclaim multiple runs, try to allocate
3046 	 * space to store the runs.  If that fails, fall back to the old
3047 	 * behavior of just reclaiming MIN_RECLAIM pages.
3048 	 */
3049 	if (desired_runs > 1)
3050 		m_runs = malloc((NRUNS + desired_runs) * sizeof(*m_runs),
3051 		    M_TEMP, M_NOWAIT);
3052 	else
3053 		m_runs = NULL;
3054 
3055 	if (m_runs == NULL) {
3056 		m_runs = _m_runs;
3057 		nruns = NRUNS;
3058 	} else {
3059 		nruns = NRUNS + desired_runs - 1;
3060 	}
3061 	min_reclaim = MAX(desired_runs * npages, MIN_RECLAIM);
3062 
3063 	/*
3064 	 * The caller will attempt an allocation after some runs have been
3065 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3066 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3067 	 * power of two or maximum chunk size, and ensure that each run is
3068 	 * suitably aligned.
3069 	 */
3070 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3071 	npages = roundup2(npages, minalign);
3072 	if (alignment < ptoa(minalign))
3073 		alignment = ptoa(minalign);
3074 
3075 	/*
3076 	 * The page daemon is allowed to dig deeper into the free page list.
3077 	 */
3078 	req_class = req & VM_ALLOC_CLASS_MASK;
3079 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3080 		req_class = VM_ALLOC_SYSTEM;
3081 
3082 	/*
3083 	 * Return if the number of free pages cannot satisfy the requested
3084 	 * allocation.
3085 	 */
3086 	vmd = VM_DOMAIN(domain);
3087 	count = vmd->vmd_free_count;
3088 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3089 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3090 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3091 		goto done;
3092 
3093 	/*
3094 	 * Scan up to three times, relaxing the restrictions ("options") on
3095 	 * the reclamation of reservations and superpages each time.
3096 	 */
3097 	for (options = VPSC_NORESERV;;) {
3098 		/*
3099 		 * Find the highest runs that satisfy the given constraints
3100 		 * and restrictions, and record them in "m_runs".
3101 		 */
3102 		count = 0;
3103 		segind = vm_phys_lookup_segind(low);
3104 		while ((segind = vm_phys_find_range(bounds, segind, domain,
3105 		    npages, low, high)) != -1) {
3106 			while ((m_run = vm_page_scan_contig(npages, bounds[0],
3107 			    bounds[1], alignment, boundary, options))) {
3108 				bounds[0] = m_run + npages;
3109 				m_runs[RUN_INDEX(count, nruns)] = m_run;
3110 				count++;
3111 			}
3112 			segind++;
3113 		}
3114 
3115 		/*
3116 		 * Reclaim the highest runs in LIFO (descending) order until
3117 		 * the number of reclaimed pages, "reclaimed", is at least
3118 		 * "min_reclaim".  Reset "reclaimed" each time because each
3119 		 * reclamation is idempotent, and runs will (likely) recur
3120 		 * from one scan to the next as restrictions are relaxed.
3121 		 */
3122 		reclaimed = 0;
3123 		for (i = 0; count > 0 && i < nruns; i++) {
3124 			count--;
3125 			m_run = m_runs[RUN_INDEX(count, nruns)];
3126 			error = vm_page_reclaim_run(req_class, domain, npages,
3127 			    m_run, high);
3128 			if (error == 0) {
3129 				reclaimed += npages;
3130 				if (reclaimed >= min_reclaim) {
3131 					ret = true;
3132 					goto done;
3133 				}
3134 			}
3135 		}
3136 
3137 		/*
3138 		 * Either relax the restrictions on the next scan or return if
3139 		 * the last scan had no restrictions.
3140 		 */
3141 		if (options == VPSC_NORESERV)
3142 			options = VPSC_NOSUPER;
3143 		else if (options == VPSC_NOSUPER)
3144 			options = VPSC_ANY;
3145 		else if (options == VPSC_ANY) {
3146 			ret = reclaimed != 0;
3147 			goto done;
3148 		}
3149 	}
3150 done:
3151 	if (m_runs != _m_runs)
3152 		free(m_runs, M_TEMP);
3153 	return (ret);
3154 }
3155 
3156 bool
3157 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3158     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3159 {
3160 	return (vm_page_reclaim_contig_domain_ext(domain, req, npages, low, high,
3161 	    alignment, boundary, 1));
3162 }
3163 
3164 bool
3165 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3166     u_long alignment, vm_paddr_t boundary)
3167 {
3168 	struct vm_domainset_iter di;
3169 	int domain;
3170 	bool ret;
3171 
3172 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3173 	do {
3174 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3175 		    high, alignment, boundary);
3176 		if (ret)
3177 			break;
3178 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3179 
3180 	return (ret);
3181 }
3182 
3183 /*
3184  * Set the domain in the appropriate page level domainset.
3185  */
3186 void
3187 vm_domain_set(struct vm_domain *vmd)
3188 {
3189 
3190 	mtx_lock(&vm_domainset_lock);
3191 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3192 		vmd->vmd_minset = 1;
3193 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3194 	}
3195 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3196 		vmd->vmd_severeset = 1;
3197 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3198 	}
3199 	mtx_unlock(&vm_domainset_lock);
3200 }
3201 
3202 /*
3203  * Clear the domain from the appropriate page level domainset.
3204  */
3205 void
3206 vm_domain_clear(struct vm_domain *vmd)
3207 {
3208 
3209 	mtx_lock(&vm_domainset_lock);
3210 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3211 		vmd->vmd_minset = 0;
3212 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3213 		if (vm_min_waiters != 0) {
3214 			vm_min_waiters = 0;
3215 			wakeup(&vm_min_domains);
3216 		}
3217 	}
3218 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3219 		vmd->vmd_severeset = 0;
3220 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3221 		if (vm_severe_waiters != 0) {
3222 			vm_severe_waiters = 0;
3223 			wakeup(&vm_severe_domains);
3224 		}
3225 	}
3226 
3227 	/*
3228 	 * If pageout daemon needs pages, then tell it that there are
3229 	 * some free.
3230 	 */
3231 	if (vmd->vmd_pageout_pages_needed &&
3232 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3233 		wakeup(&vmd->vmd_pageout_pages_needed);
3234 		vmd->vmd_pageout_pages_needed = 0;
3235 	}
3236 
3237 	/* See comments in vm_wait_doms(). */
3238 	if (vm_pageproc_waiters) {
3239 		vm_pageproc_waiters = 0;
3240 		wakeup(&vm_pageproc_waiters);
3241 	}
3242 	mtx_unlock(&vm_domainset_lock);
3243 }
3244 
3245 /*
3246  * Wait for free pages to exceed the min threshold globally.
3247  */
3248 void
3249 vm_wait_min(void)
3250 {
3251 
3252 	mtx_lock(&vm_domainset_lock);
3253 	while (vm_page_count_min()) {
3254 		vm_min_waiters++;
3255 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3256 	}
3257 	mtx_unlock(&vm_domainset_lock);
3258 }
3259 
3260 /*
3261  * Wait for free pages to exceed the severe threshold globally.
3262  */
3263 void
3264 vm_wait_severe(void)
3265 {
3266 
3267 	mtx_lock(&vm_domainset_lock);
3268 	while (vm_page_count_severe()) {
3269 		vm_severe_waiters++;
3270 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3271 		    "vmwait", 0);
3272 	}
3273 	mtx_unlock(&vm_domainset_lock);
3274 }
3275 
3276 u_int
3277 vm_wait_count(void)
3278 {
3279 
3280 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3281 }
3282 
3283 int
3284 vm_wait_doms(const domainset_t *wdoms, int mflags)
3285 {
3286 	int error;
3287 
3288 	error = 0;
3289 
3290 	/*
3291 	 * We use racey wakeup synchronization to avoid expensive global
3292 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3293 	 * To handle this, we only sleep for one tick in this instance.  It
3294 	 * is expected that most allocations for the pageproc will come from
3295 	 * kmem or vm_page_grab* which will use the more specific and
3296 	 * race-free vm_wait_domain().
3297 	 */
3298 	if (curproc == pageproc) {
3299 		mtx_lock(&vm_domainset_lock);
3300 		vm_pageproc_waiters++;
3301 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3302 		    PVM | PDROP | mflags, "pageprocwait", 1);
3303 	} else {
3304 		/*
3305 		 * XXX Ideally we would wait only until the allocation could
3306 		 * be satisfied.  This condition can cause new allocators to
3307 		 * consume all freed pages while old allocators wait.
3308 		 */
3309 		mtx_lock(&vm_domainset_lock);
3310 		if (vm_page_count_min_set(wdoms)) {
3311 			if (pageproc == NULL)
3312 				panic("vm_wait in early boot");
3313 			vm_min_waiters++;
3314 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3315 			    PVM | PDROP | mflags, "vmwait", 0);
3316 		} else
3317 			mtx_unlock(&vm_domainset_lock);
3318 	}
3319 	return (error);
3320 }
3321 
3322 /*
3323  *	vm_wait_domain:
3324  *
3325  *	Sleep until free pages are available for allocation.
3326  *	- Called in various places after failed memory allocations.
3327  */
3328 void
3329 vm_wait_domain(int domain)
3330 {
3331 	struct vm_domain *vmd;
3332 	domainset_t wdom;
3333 
3334 	vmd = VM_DOMAIN(domain);
3335 	vm_domain_free_assert_unlocked(vmd);
3336 
3337 	if (curproc == pageproc) {
3338 		mtx_lock(&vm_domainset_lock);
3339 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3340 			vmd->vmd_pageout_pages_needed = 1;
3341 			msleep(&vmd->vmd_pageout_pages_needed,
3342 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3343 		} else
3344 			mtx_unlock(&vm_domainset_lock);
3345 	} else {
3346 		DOMAINSET_ZERO(&wdom);
3347 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3348 		vm_wait_doms(&wdom, 0);
3349 	}
3350 }
3351 
3352 static int
3353 vm_wait_flags(vm_object_t obj, int mflags)
3354 {
3355 	struct domainset *d;
3356 
3357 	d = NULL;
3358 
3359 	/*
3360 	 * Carefully fetch pointers only once: the struct domainset
3361 	 * itself is ummutable but the pointer might change.
3362 	 */
3363 	if (obj != NULL)
3364 		d = obj->domain.dr_policy;
3365 	if (d == NULL)
3366 		d = curthread->td_domain.dr_policy;
3367 
3368 	return (vm_wait_doms(&d->ds_mask, mflags));
3369 }
3370 
3371 /*
3372  *	vm_wait:
3373  *
3374  *	Sleep until free pages are available for allocation in the
3375  *	affinity domains of the obj.  If obj is NULL, the domain set
3376  *	for the calling thread is used.
3377  *	Called in various places after failed memory allocations.
3378  */
3379 void
3380 vm_wait(vm_object_t obj)
3381 {
3382 	(void)vm_wait_flags(obj, 0);
3383 }
3384 
3385 int
3386 vm_wait_intr(vm_object_t obj)
3387 {
3388 	return (vm_wait_flags(obj, PCATCH));
3389 }
3390 
3391 /*
3392  *	vm_domain_alloc_fail:
3393  *
3394  *	Called when a page allocation function fails.  Informs the
3395  *	pagedaemon and performs the requested wait.  Requires the
3396  *	domain_free and object lock on entry.  Returns with the
3397  *	object lock held and free lock released.  Returns an error when
3398  *	retry is necessary.
3399  *
3400  */
3401 static int
3402 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3403 {
3404 
3405 	vm_domain_free_assert_unlocked(vmd);
3406 
3407 	atomic_add_int(&vmd->vmd_pageout_deficit,
3408 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3409 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3410 		if (object != NULL)
3411 			VM_OBJECT_WUNLOCK(object);
3412 		vm_wait_domain(vmd->vmd_domain);
3413 		if (object != NULL)
3414 			VM_OBJECT_WLOCK(object);
3415 		if (req & VM_ALLOC_WAITOK)
3416 			return (EAGAIN);
3417 	}
3418 
3419 	return (0);
3420 }
3421 
3422 /*
3423  *	vm_waitpfault:
3424  *
3425  *	Sleep until free pages are available for allocation.
3426  *	- Called only in vm_fault so that processes page faulting
3427  *	  can be easily tracked.
3428  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3429  *	  processes will be able to grab memory first.  Do not change
3430  *	  this balance without careful testing first.
3431  */
3432 void
3433 vm_waitpfault(struct domainset *dset, int timo)
3434 {
3435 
3436 	/*
3437 	 * XXX Ideally we would wait only until the allocation could
3438 	 * be satisfied.  This condition can cause new allocators to
3439 	 * consume all freed pages while old allocators wait.
3440 	 */
3441 	mtx_lock(&vm_domainset_lock);
3442 	if (vm_page_count_min_set(&dset->ds_mask)) {
3443 		vm_min_waiters++;
3444 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3445 		    "pfault", timo);
3446 	} else
3447 		mtx_unlock(&vm_domainset_lock);
3448 }
3449 
3450 static struct vm_pagequeue *
3451 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3452 {
3453 
3454 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3455 }
3456 
3457 #ifdef INVARIANTS
3458 static struct vm_pagequeue *
3459 vm_page_pagequeue(vm_page_t m)
3460 {
3461 
3462 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3463 }
3464 #endif
3465 
3466 static __always_inline bool
3467 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3468 {
3469 	vm_page_astate_t tmp;
3470 
3471 	tmp = *old;
3472 	do {
3473 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3474 			return (true);
3475 		counter_u64_add(pqstate_commit_retries, 1);
3476 	} while (old->_bits == tmp._bits);
3477 
3478 	return (false);
3479 }
3480 
3481 /*
3482  * Do the work of committing a queue state update that moves the page out of
3483  * its current queue.
3484  */
3485 static bool
3486 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3487     vm_page_astate_t *old, vm_page_astate_t new)
3488 {
3489 	vm_page_t next;
3490 
3491 	vm_pagequeue_assert_locked(pq);
3492 	KASSERT(vm_page_pagequeue(m) == pq,
3493 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3494 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3495 	    ("%s: invalid queue indices %d %d",
3496 	    __func__, old->queue, new.queue));
3497 
3498 	/*
3499 	 * Once the queue index of the page changes there is nothing
3500 	 * synchronizing with further updates to the page's physical
3501 	 * queue state.  Therefore we must speculatively remove the page
3502 	 * from the queue now and be prepared to roll back if the queue
3503 	 * state update fails.  If the page is not physically enqueued then
3504 	 * we just update its queue index.
3505 	 */
3506 	if ((old->flags & PGA_ENQUEUED) != 0) {
3507 		new.flags &= ~PGA_ENQUEUED;
3508 		next = TAILQ_NEXT(m, plinks.q);
3509 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3510 		vm_pagequeue_cnt_dec(pq);
3511 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3512 			if (next == NULL)
3513 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3514 			else
3515 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3516 			vm_pagequeue_cnt_inc(pq);
3517 			return (false);
3518 		} else {
3519 			return (true);
3520 		}
3521 	} else {
3522 		return (vm_page_pqstate_fcmpset(m, old, new));
3523 	}
3524 }
3525 
3526 static bool
3527 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3528     vm_page_astate_t new)
3529 {
3530 	struct vm_pagequeue *pq;
3531 	vm_page_astate_t as;
3532 	bool ret;
3533 
3534 	pq = _vm_page_pagequeue(m, old->queue);
3535 
3536 	/*
3537 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3538 	 * corresponding page queue lock is held.
3539 	 */
3540 	vm_pagequeue_lock(pq);
3541 	as = vm_page_astate_load(m);
3542 	if (__predict_false(as._bits != old->_bits)) {
3543 		*old = as;
3544 		ret = false;
3545 	} else {
3546 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3547 	}
3548 	vm_pagequeue_unlock(pq);
3549 	return (ret);
3550 }
3551 
3552 /*
3553  * Commit a queue state update that enqueues or requeues a page.
3554  */
3555 static bool
3556 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3557     vm_page_astate_t *old, vm_page_astate_t new)
3558 {
3559 	struct vm_domain *vmd;
3560 
3561 	vm_pagequeue_assert_locked(pq);
3562 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3563 	    ("%s: invalid queue indices %d %d",
3564 	    __func__, old->queue, new.queue));
3565 
3566 	new.flags |= PGA_ENQUEUED;
3567 	if (!vm_page_pqstate_fcmpset(m, old, new))
3568 		return (false);
3569 
3570 	if ((old->flags & PGA_ENQUEUED) != 0)
3571 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3572 	else
3573 		vm_pagequeue_cnt_inc(pq);
3574 
3575 	/*
3576 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3577 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3578 	 * applied, even if it was set first.
3579 	 */
3580 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3581 		vmd = vm_pagequeue_domain(m);
3582 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3583 		    ("%s: invalid page queue for page %p", __func__, m));
3584 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3585 	} else {
3586 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3587 	}
3588 	return (true);
3589 }
3590 
3591 /*
3592  * Commit a queue state update that encodes a request for a deferred queue
3593  * operation.
3594  */
3595 static bool
3596 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3597     vm_page_astate_t new)
3598 {
3599 
3600 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3601 	    ("%s: invalid state, queue %d flags %x",
3602 	    __func__, new.queue, new.flags));
3603 
3604 	if (old->_bits != new._bits &&
3605 	    !vm_page_pqstate_fcmpset(m, old, new))
3606 		return (false);
3607 	vm_page_pqbatch_submit(m, new.queue);
3608 	return (true);
3609 }
3610 
3611 /*
3612  * A generic queue state update function.  This handles more cases than the
3613  * specialized functions above.
3614  */
3615 bool
3616 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3617 {
3618 
3619 	if (old->_bits == new._bits)
3620 		return (true);
3621 
3622 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3623 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3624 			return (false);
3625 		if (new.queue != PQ_NONE)
3626 			vm_page_pqbatch_submit(m, new.queue);
3627 	} else {
3628 		if (!vm_page_pqstate_fcmpset(m, old, new))
3629 			return (false);
3630 		if (new.queue != PQ_NONE &&
3631 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3632 			vm_page_pqbatch_submit(m, new.queue);
3633 	}
3634 	return (true);
3635 }
3636 
3637 /*
3638  * Apply deferred queue state updates to a page.
3639  */
3640 static inline void
3641 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3642 {
3643 	vm_page_astate_t new, old;
3644 
3645 	CRITICAL_ASSERT(curthread);
3646 	vm_pagequeue_assert_locked(pq);
3647 	KASSERT(queue < PQ_COUNT,
3648 	    ("%s: invalid queue index %d", __func__, queue));
3649 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3650 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3651 
3652 	for (old = vm_page_astate_load(m);;) {
3653 		if (__predict_false(old.queue != queue ||
3654 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3655 			counter_u64_add(queue_nops, 1);
3656 			break;
3657 		}
3658 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3659 		    ("%s: page %p is unmanaged", __func__, m));
3660 
3661 		new = old;
3662 		if ((old.flags & PGA_DEQUEUE) != 0) {
3663 			new.flags &= ~PGA_QUEUE_OP_MASK;
3664 			new.queue = PQ_NONE;
3665 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3666 			    m, &old, new))) {
3667 				counter_u64_add(queue_ops, 1);
3668 				break;
3669 			}
3670 		} else {
3671 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3672 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3673 			    m, &old, new))) {
3674 				counter_u64_add(queue_ops, 1);
3675 				break;
3676 			}
3677 		}
3678 	}
3679 }
3680 
3681 static void
3682 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3683     uint8_t queue)
3684 {
3685 	int i;
3686 
3687 	for (i = 0; i < bq->bq_cnt; i++)
3688 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3689 	vm_batchqueue_init(bq);
3690 }
3691 
3692 /*
3693  *	vm_page_pqbatch_submit:		[ internal use only ]
3694  *
3695  *	Enqueue a page in the specified page queue's batched work queue.
3696  *	The caller must have encoded the requested operation in the page
3697  *	structure's a.flags field.
3698  */
3699 void
3700 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3701 {
3702 	struct vm_batchqueue *bq;
3703 	struct vm_pagequeue *pq;
3704 	int domain, slots_remaining;
3705 
3706 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3707 
3708 	domain = vm_page_domain(m);
3709 	critical_enter();
3710 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3711 	slots_remaining = vm_batchqueue_insert(bq, m);
3712 	if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
3713 		/* keep building the bq */
3714 		critical_exit();
3715 		return;
3716 	} else if (slots_remaining > 0 ) {
3717 		/* Try to process the bq if we can get the lock */
3718 		pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3719 		if (vm_pagequeue_trylock(pq)) {
3720 			vm_pqbatch_process(pq, bq, queue);
3721 			vm_pagequeue_unlock(pq);
3722 		}
3723 		critical_exit();
3724 		return;
3725 	}
3726 	critical_exit();
3727 
3728 	/* if we make it here, the bq is full so wait for the lock */
3729 
3730 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3731 	vm_pagequeue_lock(pq);
3732 	critical_enter();
3733 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3734 	vm_pqbatch_process(pq, bq, queue);
3735 	vm_pqbatch_process_page(pq, m, queue);
3736 	vm_pagequeue_unlock(pq);
3737 	critical_exit();
3738 }
3739 
3740 /*
3741  *	vm_page_pqbatch_drain:		[ internal use only ]
3742  *
3743  *	Force all per-CPU page queue batch queues to be drained.  This is
3744  *	intended for use in severe memory shortages, to ensure that pages
3745  *	do not remain stuck in the batch queues.
3746  */
3747 void
3748 vm_page_pqbatch_drain(void)
3749 {
3750 	struct thread *td;
3751 	struct vm_domain *vmd;
3752 	struct vm_pagequeue *pq;
3753 	int cpu, domain, queue;
3754 
3755 	td = curthread;
3756 	CPU_FOREACH(cpu) {
3757 		thread_lock(td);
3758 		sched_bind(td, cpu);
3759 		thread_unlock(td);
3760 
3761 		for (domain = 0; domain < vm_ndomains; domain++) {
3762 			vmd = VM_DOMAIN(domain);
3763 			for (queue = 0; queue < PQ_COUNT; queue++) {
3764 				pq = &vmd->vmd_pagequeues[queue];
3765 				vm_pagequeue_lock(pq);
3766 				critical_enter();
3767 				vm_pqbatch_process(pq,
3768 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3769 				critical_exit();
3770 				vm_pagequeue_unlock(pq);
3771 			}
3772 		}
3773 	}
3774 	thread_lock(td);
3775 	sched_unbind(td);
3776 	thread_unlock(td);
3777 }
3778 
3779 /*
3780  *	vm_page_dequeue_deferred:	[ internal use only ]
3781  *
3782  *	Request removal of the given page from its current page
3783  *	queue.  Physical removal from the queue may be deferred
3784  *	indefinitely.
3785  */
3786 void
3787 vm_page_dequeue_deferred(vm_page_t m)
3788 {
3789 	vm_page_astate_t new, old;
3790 
3791 	old = vm_page_astate_load(m);
3792 	do {
3793 		if (old.queue == PQ_NONE) {
3794 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3795 			    ("%s: page %p has unexpected queue state",
3796 			    __func__, m));
3797 			break;
3798 		}
3799 		new = old;
3800 		new.flags |= PGA_DEQUEUE;
3801 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3802 }
3803 
3804 /*
3805  *	vm_page_dequeue:
3806  *
3807  *	Remove the page from whichever page queue it's in, if any, before
3808  *	returning.
3809  */
3810 void
3811 vm_page_dequeue(vm_page_t m)
3812 {
3813 	vm_page_astate_t new, old;
3814 
3815 	old = vm_page_astate_load(m);
3816 	do {
3817 		if (old.queue == PQ_NONE) {
3818 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3819 			    ("%s: page %p has unexpected queue state",
3820 			    __func__, m));
3821 			break;
3822 		}
3823 		new = old;
3824 		new.flags &= ~PGA_QUEUE_OP_MASK;
3825 		new.queue = PQ_NONE;
3826 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3827 
3828 }
3829 
3830 /*
3831  * Schedule the given page for insertion into the specified page queue.
3832  * Physical insertion of the page may be deferred indefinitely.
3833  */
3834 static void
3835 vm_page_enqueue(vm_page_t m, uint8_t queue)
3836 {
3837 
3838 	KASSERT(m->a.queue == PQ_NONE &&
3839 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3840 	    ("%s: page %p is already enqueued", __func__, m));
3841 	KASSERT(m->ref_count > 0,
3842 	    ("%s: page %p does not carry any references", __func__, m));
3843 
3844 	m->a.queue = queue;
3845 	if ((m->a.flags & PGA_REQUEUE) == 0)
3846 		vm_page_aflag_set(m, PGA_REQUEUE);
3847 	vm_page_pqbatch_submit(m, queue);
3848 }
3849 
3850 /*
3851  *	vm_page_free_prep:
3852  *
3853  *	Prepares the given page to be put on the free list,
3854  *	disassociating it from any VM object. The caller may return
3855  *	the page to the free list only if this function returns true.
3856  *
3857  *	The object, if it exists, must be locked, and then the page must
3858  *	be xbusy.  Otherwise the page must be not busied.  A managed
3859  *	page must be unmapped.
3860  */
3861 static bool
3862 vm_page_free_prep(vm_page_t m)
3863 {
3864 
3865 	/*
3866 	 * Synchronize with threads that have dropped a reference to this
3867 	 * page.
3868 	 */
3869 	atomic_thread_fence_acq();
3870 
3871 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3872 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3873 		uint64_t *p;
3874 		int i;
3875 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3876 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3877 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3878 			    m, i, (uintmax_t)*p));
3879 	}
3880 #endif
3881 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3882 		KASSERT(!pmap_page_is_mapped(m),
3883 		    ("vm_page_free_prep: freeing mapped page %p", m));
3884 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3885 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3886 	} else {
3887 		KASSERT(m->a.queue == PQ_NONE,
3888 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3889 	}
3890 	VM_CNT_INC(v_tfree);
3891 
3892 	if (m->object != NULL) {
3893 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3894 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
3895 		    ("vm_page_free_prep: managed flag mismatch for page %p",
3896 		    m));
3897 		vm_page_assert_xbusied(m);
3898 
3899 		/*
3900 		 * The object reference can be released without an atomic
3901 		 * operation.
3902 		 */
3903 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3904 		    m->ref_count == VPRC_OBJREF,
3905 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3906 		    m, m->ref_count));
3907 		vm_page_object_remove(m);
3908 		m->ref_count -= VPRC_OBJREF;
3909 	} else
3910 		vm_page_assert_unbusied(m);
3911 
3912 	vm_page_busy_free(m);
3913 
3914 	/*
3915 	 * If fictitious remove object association and
3916 	 * return.
3917 	 */
3918 	if ((m->flags & PG_FICTITIOUS) != 0) {
3919 		KASSERT(m->ref_count == 1,
3920 		    ("fictitious page %p is referenced", m));
3921 		KASSERT(m->a.queue == PQ_NONE,
3922 		    ("fictitious page %p is queued", m));
3923 		return (false);
3924 	}
3925 
3926 	/*
3927 	 * Pages need not be dequeued before they are returned to the physical
3928 	 * memory allocator, but they must at least be marked for a deferred
3929 	 * dequeue.
3930 	 */
3931 	if ((m->oflags & VPO_UNMANAGED) == 0)
3932 		vm_page_dequeue_deferred(m);
3933 
3934 	m->valid = 0;
3935 	vm_page_undirty(m);
3936 
3937 	if (m->ref_count != 0)
3938 		panic("vm_page_free_prep: page %p has references", m);
3939 
3940 	/*
3941 	 * Restore the default memory attribute to the page.
3942 	 */
3943 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3944 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3945 
3946 #if VM_NRESERVLEVEL > 0
3947 	/*
3948 	 * Determine whether the page belongs to a reservation.  If the page was
3949 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3950 	 * as an optimization, we avoid the check in that case.
3951 	 */
3952 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3953 		return (false);
3954 #endif
3955 
3956 	return (true);
3957 }
3958 
3959 /*
3960  *	vm_page_free_toq:
3961  *
3962  *	Returns the given page to the free list, disassociating it
3963  *	from any VM object.
3964  *
3965  *	The object must be locked.  The page must be exclusively busied if it
3966  *	belongs to an object.
3967  */
3968 static void
3969 vm_page_free_toq(vm_page_t m)
3970 {
3971 	struct vm_domain *vmd;
3972 	uma_zone_t zone;
3973 
3974 	if (!vm_page_free_prep(m))
3975 		return;
3976 
3977 	vmd = vm_pagequeue_domain(m);
3978 	zone = vmd->vmd_pgcache[m->pool].zone;
3979 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3980 		uma_zfree(zone, m);
3981 		return;
3982 	}
3983 	vm_domain_free_lock(vmd);
3984 	vm_phys_free_pages(m, 0);
3985 	vm_domain_free_unlock(vmd);
3986 	vm_domain_freecnt_inc(vmd, 1);
3987 }
3988 
3989 /*
3990  *	vm_page_free_pages_toq:
3991  *
3992  *	Returns a list of pages to the free list, disassociating it
3993  *	from any VM object.  In other words, this is equivalent to
3994  *	calling vm_page_free_toq() for each page of a list of VM objects.
3995  */
3996 void
3997 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3998 {
3999 	vm_page_t m;
4000 	int count;
4001 
4002 	if (SLIST_EMPTY(free))
4003 		return;
4004 
4005 	count = 0;
4006 	while ((m = SLIST_FIRST(free)) != NULL) {
4007 		count++;
4008 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
4009 		vm_page_free_toq(m);
4010 	}
4011 
4012 	if (update_wire_count)
4013 		vm_wire_sub(count);
4014 }
4015 
4016 /*
4017  * Mark this page as wired down.  For managed pages, this prevents reclamation
4018  * by the page daemon, or when the containing object, if any, is destroyed.
4019  */
4020 void
4021 vm_page_wire(vm_page_t m)
4022 {
4023 	u_int old;
4024 
4025 #ifdef INVARIANTS
4026 	if (m->object != NULL && !vm_page_busied(m) &&
4027 	    !vm_object_busied(m->object))
4028 		VM_OBJECT_ASSERT_LOCKED(m->object);
4029 #endif
4030 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
4031 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
4032 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
4033 
4034 	old = atomic_fetchadd_int(&m->ref_count, 1);
4035 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
4036 	    ("vm_page_wire: counter overflow for page %p", m));
4037 	if (VPRC_WIRE_COUNT(old) == 0) {
4038 		if ((m->oflags & VPO_UNMANAGED) == 0)
4039 			vm_page_aflag_set(m, PGA_DEQUEUE);
4040 		vm_wire_add(1);
4041 	}
4042 }
4043 
4044 /*
4045  * Attempt to wire a mapped page following a pmap lookup of that page.
4046  * This may fail if a thread is concurrently tearing down mappings of the page.
4047  * The transient failure is acceptable because it translates to the
4048  * failure of the caller pmap_extract_and_hold(), which should be then
4049  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
4050  */
4051 bool
4052 vm_page_wire_mapped(vm_page_t m)
4053 {
4054 	u_int old;
4055 
4056 	old = m->ref_count;
4057 	do {
4058 		KASSERT(old > 0,
4059 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4060 		if ((old & VPRC_BLOCKED) != 0)
4061 			return (false);
4062 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4063 
4064 	if (VPRC_WIRE_COUNT(old) == 0) {
4065 		if ((m->oflags & VPO_UNMANAGED) == 0)
4066 			vm_page_aflag_set(m, PGA_DEQUEUE);
4067 		vm_wire_add(1);
4068 	}
4069 	return (true);
4070 }
4071 
4072 /*
4073  * Release a wiring reference to a managed page.  If the page still belongs to
4074  * an object, update its position in the page queues to reflect the reference.
4075  * If the wiring was the last reference to the page, free the page.
4076  */
4077 static void
4078 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4079 {
4080 	u_int old;
4081 
4082 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4083 	    ("%s: page %p is unmanaged", __func__, m));
4084 
4085 	/*
4086 	 * Update LRU state before releasing the wiring reference.
4087 	 * Use a release store when updating the reference count to
4088 	 * synchronize with vm_page_free_prep().
4089 	 */
4090 	old = m->ref_count;
4091 	do {
4092 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4093 		    ("vm_page_unwire: wire count underflow for page %p", m));
4094 
4095 		if (old > VPRC_OBJREF + 1) {
4096 			/*
4097 			 * The page has at least one other wiring reference.  An
4098 			 * earlier iteration of this loop may have called
4099 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4100 			 * re-set it if necessary.
4101 			 */
4102 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4103 				vm_page_aflag_set(m, PGA_DEQUEUE);
4104 		} else if (old == VPRC_OBJREF + 1) {
4105 			/*
4106 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4107 			 * update the page's queue state to reflect the
4108 			 * reference.  If the page does not belong to an object
4109 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4110 			 * clear leftover queue state.
4111 			 */
4112 			vm_page_release_toq(m, nqueue, noreuse);
4113 		} else if (old == 1) {
4114 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4115 		}
4116 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4117 
4118 	if (VPRC_WIRE_COUNT(old) == 1) {
4119 		vm_wire_sub(1);
4120 		if (old == 1)
4121 			vm_page_free(m);
4122 	}
4123 }
4124 
4125 /*
4126  * Release one wiring of the specified page, potentially allowing it to be
4127  * paged out.
4128  *
4129  * Only managed pages belonging to an object can be paged out.  If the number
4130  * of wirings transitions to zero and the page is eligible for page out, then
4131  * the page is added to the specified paging queue.  If the released wiring
4132  * represented the last reference to the page, the page is freed.
4133  */
4134 void
4135 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4136 {
4137 
4138 	KASSERT(nqueue < PQ_COUNT,
4139 	    ("vm_page_unwire: invalid queue %u request for page %p",
4140 	    nqueue, m));
4141 
4142 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4143 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4144 			vm_page_free(m);
4145 		return;
4146 	}
4147 	vm_page_unwire_managed(m, nqueue, false);
4148 }
4149 
4150 /*
4151  * Unwire a page without (re-)inserting it into a page queue.  It is up
4152  * to the caller to enqueue, requeue, or free the page as appropriate.
4153  * In most cases involving managed pages, vm_page_unwire() should be used
4154  * instead.
4155  */
4156 bool
4157 vm_page_unwire_noq(vm_page_t m)
4158 {
4159 	u_int old;
4160 
4161 	old = vm_page_drop(m, 1);
4162 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4163 	    ("%s: counter underflow for page %p", __func__,  m));
4164 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4165 	    ("%s: missing ref on fictitious page %p", __func__, m));
4166 
4167 	if (VPRC_WIRE_COUNT(old) > 1)
4168 		return (false);
4169 	if ((m->oflags & VPO_UNMANAGED) == 0)
4170 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4171 	vm_wire_sub(1);
4172 	return (true);
4173 }
4174 
4175 /*
4176  * Ensure that the page ends up in the specified page queue.  If the page is
4177  * active or being moved to the active queue, ensure that its act_count is
4178  * at least ACT_INIT but do not otherwise mess with it.
4179  */
4180 static __always_inline void
4181 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4182 {
4183 	vm_page_astate_t old, new;
4184 
4185 	KASSERT(m->ref_count > 0,
4186 	    ("%s: page %p does not carry any references", __func__, m));
4187 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4188 	    ("%s: invalid flags %x", __func__, nflag));
4189 
4190 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4191 		return;
4192 
4193 	old = vm_page_astate_load(m);
4194 	do {
4195 		if ((old.flags & PGA_DEQUEUE) != 0)
4196 			break;
4197 		new = old;
4198 		new.flags &= ~PGA_QUEUE_OP_MASK;
4199 		if (nqueue == PQ_ACTIVE)
4200 			new.act_count = max(old.act_count, ACT_INIT);
4201 		if (old.queue == nqueue) {
4202 			/*
4203 			 * There is no need to requeue pages already in the
4204 			 * active queue.
4205 			 */
4206 			if (nqueue != PQ_ACTIVE ||
4207 			    (old.flags & PGA_ENQUEUED) == 0)
4208 				new.flags |= nflag;
4209 		} else {
4210 			new.flags |= nflag;
4211 			new.queue = nqueue;
4212 		}
4213 	} while (!vm_page_pqstate_commit(m, &old, new));
4214 }
4215 
4216 /*
4217  * Put the specified page on the active list (if appropriate).
4218  */
4219 void
4220 vm_page_activate(vm_page_t m)
4221 {
4222 
4223 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4224 }
4225 
4226 /*
4227  * Move the specified page to the tail of the inactive queue, or requeue
4228  * the page if it is already in the inactive queue.
4229  */
4230 void
4231 vm_page_deactivate(vm_page_t m)
4232 {
4233 
4234 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4235 }
4236 
4237 void
4238 vm_page_deactivate_noreuse(vm_page_t m)
4239 {
4240 
4241 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4242 }
4243 
4244 /*
4245  * Put a page in the laundry, or requeue it if it is already there.
4246  */
4247 void
4248 vm_page_launder(vm_page_t m)
4249 {
4250 
4251 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4252 }
4253 
4254 /*
4255  * Put a page in the PQ_UNSWAPPABLE holding queue.
4256  */
4257 void
4258 vm_page_unswappable(vm_page_t m)
4259 {
4260 
4261 	VM_OBJECT_ASSERT_LOCKED(m->object);
4262 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4263 	    ("page %p already unswappable", m));
4264 
4265 	vm_page_dequeue(m);
4266 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4267 }
4268 
4269 /*
4270  * Release a page back to the page queues in preparation for unwiring.
4271  */
4272 static void
4273 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4274 {
4275 	vm_page_astate_t old, new;
4276 	uint16_t nflag;
4277 
4278 	/*
4279 	 * Use a check of the valid bits to determine whether we should
4280 	 * accelerate reclamation of the page.  The object lock might not be
4281 	 * held here, in which case the check is racy.  At worst we will either
4282 	 * accelerate reclamation of a valid page and violate LRU, or
4283 	 * unnecessarily defer reclamation of an invalid page.
4284 	 *
4285 	 * If we were asked to not cache the page, place it near the head of the
4286 	 * inactive queue so that is reclaimed sooner.
4287 	 */
4288 	if (noreuse || vm_page_none_valid(m)) {
4289 		nqueue = PQ_INACTIVE;
4290 		nflag = PGA_REQUEUE_HEAD;
4291 	} else {
4292 		nflag = PGA_REQUEUE;
4293 	}
4294 
4295 	old = vm_page_astate_load(m);
4296 	do {
4297 		new = old;
4298 
4299 		/*
4300 		 * If the page is already in the active queue and we are not
4301 		 * trying to accelerate reclamation, simply mark it as
4302 		 * referenced and avoid any queue operations.
4303 		 */
4304 		new.flags &= ~PGA_QUEUE_OP_MASK;
4305 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
4306 		    (old.flags & PGA_ENQUEUED) != 0)
4307 			new.flags |= PGA_REFERENCED;
4308 		else {
4309 			new.flags |= nflag;
4310 			new.queue = nqueue;
4311 		}
4312 	} while (!vm_page_pqstate_commit(m, &old, new));
4313 }
4314 
4315 /*
4316  * Unwire a page and either attempt to free it or re-add it to the page queues.
4317  */
4318 void
4319 vm_page_release(vm_page_t m, int flags)
4320 {
4321 	vm_object_t object;
4322 
4323 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4324 	    ("vm_page_release: page %p is unmanaged", m));
4325 
4326 	if ((flags & VPR_TRYFREE) != 0) {
4327 		for (;;) {
4328 			object = atomic_load_ptr(&m->object);
4329 			if (object == NULL)
4330 				break;
4331 			/* Depends on type-stability. */
4332 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4333 				break;
4334 			if (object == m->object) {
4335 				vm_page_release_locked(m, flags);
4336 				VM_OBJECT_WUNLOCK(object);
4337 				return;
4338 			}
4339 			VM_OBJECT_WUNLOCK(object);
4340 		}
4341 	}
4342 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4343 }
4344 
4345 /* See vm_page_release(). */
4346 void
4347 vm_page_release_locked(vm_page_t m, int flags)
4348 {
4349 
4350 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4351 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4352 	    ("vm_page_release_locked: page %p is unmanaged", m));
4353 
4354 	if (vm_page_unwire_noq(m)) {
4355 		if ((flags & VPR_TRYFREE) != 0 &&
4356 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4357 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4358 			/*
4359 			 * An unlocked lookup may have wired the page before the
4360 			 * busy lock was acquired, in which case the page must
4361 			 * not be freed.
4362 			 */
4363 			if (__predict_true(!vm_page_wired(m))) {
4364 				vm_page_free(m);
4365 				return;
4366 			}
4367 			vm_page_xunbusy(m);
4368 		} else {
4369 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4370 		}
4371 	}
4372 }
4373 
4374 static bool
4375 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4376 {
4377 	u_int old;
4378 
4379 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4380 	    ("vm_page_try_blocked_op: page %p has no object", m));
4381 	KASSERT(vm_page_busied(m),
4382 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4383 	VM_OBJECT_ASSERT_LOCKED(m->object);
4384 
4385 	old = m->ref_count;
4386 	do {
4387 		KASSERT(old != 0,
4388 		    ("vm_page_try_blocked_op: page %p has no references", m));
4389 		if (VPRC_WIRE_COUNT(old) != 0)
4390 			return (false);
4391 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4392 
4393 	(op)(m);
4394 
4395 	/*
4396 	 * If the object is read-locked, new wirings may be created via an
4397 	 * object lookup.
4398 	 */
4399 	old = vm_page_drop(m, VPRC_BLOCKED);
4400 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4401 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4402 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4403 	    old, m));
4404 	return (true);
4405 }
4406 
4407 /*
4408  * Atomically check for wirings and remove all mappings of the page.
4409  */
4410 bool
4411 vm_page_try_remove_all(vm_page_t m)
4412 {
4413 
4414 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4415 }
4416 
4417 /*
4418  * Atomically check for wirings and remove all writeable mappings of the page.
4419  */
4420 bool
4421 vm_page_try_remove_write(vm_page_t m)
4422 {
4423 
4424 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4425 }
4426 
4427 /*
4428  * vm_page_advise
4429  *
4430  * 	Apply the specified advice to the given page.
4431  */
4432 void
4433 vm_page_advise(vm_page_t m, int advice)
4434 {
4435 
4436 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4437 	vm_page_assert_xbusied(m);
4438 
4439 	if (advice == MADV_FREE)
4440 		/*
4441 		 * Mark the page clean.  This will allow the page to be freed
4442 		 * without first paging it out.  MADV_FREE pages are often
4443 		 * quickly reused by malloc(3), so we do not do anything that
4444 		 * would result in a page fault on a later access.
4445 		 */
4446 		vm_page_undirty(m);
4447 	else if (advice != MADV_DONTNEED) {
4448 		if (advice == MADV_WILLNEED)
4449 			vm_page_activate(m);
4450 		return;
4451 	}
4452 
4453 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4454 		vm_page_dirty(m);
4455 
4456 	/*
4457 	 * Clear any references to the page.  Otherwise, the page daemon will
4458 	 * immediately reactivate the page.
4459 	 */
4460 	vm_page_aflag_clear(m, PGA_REFERENCED);
4461 
4462 	/*
4463 	 * Place clean pages near the head of the inactive queue rather than
4464 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4465 	 * the page will be reused quickly.  Dirty pages not already in the
4466 	 * laundry are moved there.
4467 	 */
4468 	if (m->dirty == 0)
4469 		vm_page_deactivate_noreuse(m);
4470 	else if (!vm_page_in_laundry(m))
4471 		vm_page_launder(m);
4472 }
4473 
4474 /*
4475  *	vm_page_grab_release
4476  *
4477  *	Helper routine for grab functions to release busy on return.
4478  */
4479 static inline void
4480 vm_page_grab_release(vm_page_t m, int allocflags)
4481 {
4482 
4483 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4484 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4485 			vm_page_sunbusy(m);
4486 		else
4487 			vm_page_xunbusy(m);
4488 	}
4489 }
4490 
4491 /*
4492  *	vm_page_grab_sleep
4493  *
4494  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4495  *	if the caller should retry and false otherwise.
4496  *
4497  *	If the object is locked on entry the object will be unlocked with
4498  *	false returns and still locked but possibly having been dropped
4499  *	with true returns.
4500  */
4501 static bool
4502 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4503     const char *wmesg, int allocflags, bool locked)
4504 {
4505 
4506 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4507 		return (false);
4508 
4509 	/*
4510 	 * Reference the page before unlocking and sleeping so that
4511 	 * the page daemon is less likely to reclaim it.
4512 	 */
4513 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4514 		vm_page_reference(m);
4515 
4516 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4517 	    locked)
4518 		VM_OBJECT_WLOCK(object);
4519 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4520 		return (false);
4521 
4522 	return (true);
4523 }
4524 
4525 /*
4526  * Assert that the grab flags are valid.
4527  */
4528 static inline void
4529 vm_page_grab_check(int allocflags)
4530 {
4531 
4532 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4533 	    (allocflags & VM_ALLOC_WIRED) != 0,
4534 	    ("vm_page_grab*: the pages must be busied or wired"));
4535 
4536 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4537 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4538 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4539 }
4540 
4541 /*
4542  * Calculate the page allocation flags for grab.
4543  */
4544 static inline int
4545 vm_page_grab_pflags(int allocflags)
4546 {
4547 	int pflags;
4548 
4549 	pflags = allocflags &
4550 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4551 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4552 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4553 		pflags |= VM_ALLOC_WAITFAIL;
4554 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4555 		pflags |= VM_ALLOC_SBUSY;
4556 
4557 	return (pflags);
4558 }
4559 
4560 /*
4561  * Grab a page, waiting until we are waken up due to the page
4562  * changing state.  We keep on waiting, if the page continues
4563  * to be in the object.  If the page doesn't exist, first allocate it
4564  * and then conditionally zero it.
4565  *
4566  * This routine may sleep.
4567  *
4568  * The object must be locked on entry.  The lock will, however, be released
4569  * and reacquired if the routine sleeps.
4570  */
4571 vm_page_t
4572 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4573 {
4574 	vm_page_t m;
4575 
4576 	VM_OBJECT_ASSERT_WLOCKED(object);
4577 	vm_page_grab_check(allocflags);
4578 
4579 retrylookup:
4580 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4581 		if (!vm_page_tryacquire(m, allocflags)) {
4582 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4583 			    allocflags, true))
4584 				goto retrylookup;
4585 			return (NULL);
4586 		}
4587 		goto out;
4588 	}
4589 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4590 		return (NULL);
4591 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4592 	if (m == NULL) {
4593 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4594 			return (NULL);
4595 		goto retrylookup;
4596 	}
4597 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4598 		pmap_zero_page(m);
4599 
4600 out:
4601 	vm_page_grab_release(m, allocflags);
4602 
4603 	return (m);
4604 }
4605 
4606 /*
4607  * Locklessly attempt to acquire a page given a (object, pindex) tuple
4608  * and an optional previous page to avoid the radix lookup.  The resulting
4609  * page will be validated against the identity tuple and busied or wired
4610  * as requested.  A NULL *mp return guarantees that the page was not in
4611  * radix at the time of the call but callers must perform higher level
4612  * synchronization or retry the operation under a lock if they require
4613  * an atomic answer.  This is the only lock free validation routine,
4614  * other routines can depend on the resulting page state.
4615  *
4616  * The return value indicates whether the operation failed due to caller
4617  * flags.  The return is tri-state with mp:
4618  *
4619  * (true, *mp != NULL) - The operation was successful.
4620  * (true, *mp == NULL) - The page was not found in tree.
4621  * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4622  */
4623 static bool
4624 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4625     vm_page_t prev, vm_page_t *mp, int allocflags)
4626 {
4627 	vm_page_t m;
4628 
4629 	vm_page_grab_check(allocflags);
4630 	MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4631 
4632 	*mp = NULL;
4633 	for (;;) {
4634 		/*
4635 		 * We may see a false NULL here because the previous page
4636 		 * has been removed or just inserted and the list is loaded
4637 		 * without barriers.  Switch to radix to verify.
4638 		 */
4639 		if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4640 		    QMD_IS_TRASHED(m) || m->pindex != pindex ||
4641 		    atomic_load_ptr(&m->object) != object) {
4642 			prev = NULL;
4643 			/*
4644 			 * This guarantees the result is instantaneously
4645 			 * correct.
4646 			 */
4647 			m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4648 		}
4649 		if (m == NULL)
4650 			return (true);
4651 		if (vm_page_trybusy(m, allocflags)) {
4652 			if (m->object == object && m->pindex == pindex)
4653 				break;
4654 			/* relookup. */
4655 			vm_page_busy_release(m);
4656 			cpu_spinwait();
4657 			continue;
4658 		}
4659 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4660 		    allocflags, false))
4661 			return (false);
4662 	}
4663 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4664 		vm_page_wire(m);
4665 	vm_page_grab_release(m, allocflags);
4666 	*mp = m;
4667 	return (true);
4668 }
4669 
4670 /*
4671  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4672  * is not set.
4673  */
4674 vm_page_t
4675 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4676 {
4677 	vm_page_t m;
4678 
4679 	vm_page_grab_check(allocflags);
4680 
4681 	if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4682 		return (NULL);
4683 	if (m != NULL)
4684 		return (m);
4685 
4686 	/*
4687 	 * The radix lockless lookup should never return a false negative
4688 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4689 	 * was no page present at the instant of the call.  A NOCREAT caller
4690 	 * must handle create races gracefully.
4691 	 */
4692 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4693 		return (NULL);
4694 
4695 	VM_OBJECT_WLOCK(object);
4696 	m = vm_page_grab(object, pindex, allocflags);
4697 	VM_OBJECT_WUNLOCK(object);
4698 
4699 	return (m);
4700 }
4701 
4702 /*
4703  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4704  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4705  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4706  * in simultaneously.  Additional pages will be left on a paging queue but
4707  * will neither be wired nor busy regardless of allocflags.
4708  */
4709 int
4710 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4711 {
4712 	vm_page_t m;
4713 	vm_page_t ma[VM_INITIAL_PAGEIN];
4714 	int after, i, pflags, rv;
4715 
4716 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4717 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4718 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4719 	KASSERT((allocflags &
4720 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4721 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4722 	VM_OBJECT_ASSERT_WLOCKED(object);
4723 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4724 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4725 	pflags |= VM_ALLOC_WAITFAIL;
4726 
4727 retrylookup:
4728 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4729 		/*
4730 		 * If the page is fully valid it can only become invalid
4731 		 * with the object lock held.  If it is not valid it can
4732 		 * become valid with the busy lock held.  Therefore, we
4733 		 * may unnecessarily lock the exclusive busy here if we
4734 		 * race with I/O completion not using the object lock.
4735 		 * However, we will not end up with an invalid page and a
4736 		 * shared lock.
4737 		 */
4738 		if (!vm_page_trybusy(m,
4739 		    vm_page_all_valid(m) ? allocflags : 0)) {
4740 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4741 			    allocflags, true);
4742 			goto retrylookup;
4743 		}
4744 		if (vm_page_all_valid(m))
4745 			goto out;
4746 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4747 			vm_page_busy_release(m);
4748 			*mp = NULL;
4749 			return (VM_PAGER_FAIL);
4750 		}
4751 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4752 		*mp = NULL;
4753 		return (VM_PAGER_FAIL);
4754 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4755 		if (!vm_pager_can_alloc_page(object, pindex)) {
4756 			*mp = NULL;
4757 			return (VM_PAGER_AGAIN);
4758 		}
4759 		goto retrylookup;
4760 	}
4761 
4762 	vm_page_assert_xbusied(m);
4763 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4764 		after = MIN(after, VM_INITIAL_PAGEIN);
4765 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4766 		after = MAX(after, 1);
4767 		ma[0] = m;
4768 		for (i = 1; i < after; i++) {
4769 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4770 				if (vm_page_any_valid(ma[i]) ||
4771 				    !vm_page_tryxbusy(ma[i]))
4772 					break;
4773 			} else {
4774 				ma[i] = vm_page_alloc(object, m->pindex + i,
4775 				    VM_ALLOC_NORMAL);
4776 				if (ma[i] == NULL)
4777 					break;
4778 			}
4779 		}
4780 		after = i;
4781 		vm_object_pip_add(object, after);
4782 		VM_OBJECT_WUNLOCK(object);
4783 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4784 		VM_OBJECT_WLOCK(object);
4785 		vm_object_pip_wakeupn(object, after);
4786 		/* Pager may have replaced a page. */
4787 		m = ma[0];
4788 		if (rv != VM_PAGER_OK) {
4789 			for (i = 0; i < after; i++) {
4790 				if (!vm_page_wired(ma[i]))
4791 					vm_page_free(ma[i]);
4792 				else
4793 					vm_page_xunbusy(ma[i]);
4794 			}
4795 			*mp = NULL;
4796 			return (rv);
4797 		}
4798 		for (i = 1; i < after; i++)
4799 			vm_page_readahead_finish(ma[i]);
4800 		MPASS(vm_page_all_valid(m));
4801 	} else {
4802 		vm_page_zero_invalid(m, TRUE);
4803 	}
4804 out:
4805 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4806 		vm_page_wire(m);
4807 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4808 		vm_page_busy_downgrade(m);
4809 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4810 		vm_page_busy_release(m);
4811 	*mp = m;
4812 	return (VM_PAGER_OK);
4813 }
4814 
4815 /*
4816  * Locklessly grab a valid page.  If the page is not valid or not yet
4817  * allocated this will fall back to the object lock method.
4818  */
4819 int
4820 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4821     vm_pindex_t pindex, int allocflags)
4822 {
4823 	vm_page_t m;
4824 	int flags;
4825 	int error;
4826 
4827 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4828 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4829 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4830 	    "mismatch"));
4831 	KASSERT((allocflags &
4832 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4833 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4834 
4835 	/*
4836 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
4837 	 * before we can inspect the valid field and return a wired page.
4838 	 */
4839 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4840 	if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4841 		return (VM_PAGER_FAIL);
4842 	if ((m = *mp) != NULL) {
4843 		if (vm_page_all_valid(m)) {
4844 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4845 				vm_page_wire(m);
4846 			vm_page_grab_release(m, allocflags);
4847 			return (VM_PAGER_OK);
4848 		}
4849 		vm_page_busy_release(m);
4850 	}
4851 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4852 		*mp = NULL;
4853 		return (VM_PAGER_FAIL);
4854 	}
4855 	VM_OBJECT_WLOCK(object);
4856 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
4857 	VM_OBJECT_WUNLOCK(object);
4858 
4859 	return (error);
4860 }
4861 
4862 /*
4863  * Return the specified range of pages from the given object.  For each
4864  * page offset within the range, if a page already exists within the object
4865  * at that offset and it is busy, then wait for it to change state.  If,
4866  * instead, the page doesn't exist, then allocate it.
4867  *
4868  * The caller must always specify an allocation class.
4869  *
4870  * allocation classes:
4871  *	VM_ALLOC_NORMAL		normal process request
4872  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4873  *
4874  * The caller must always specify that the pages are to be busied and/or
4875  * wired.
4876  *
4877  * optional allocation flags:
4878  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4879  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4880  *	VM_ALLOC_NOWAIT		do not sleep
4881  *	VM_ALLOC_SBUSY		set page to sbusy state
4882  *	VM_ALLOC_WIRED		wire the pages
4883  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4884  *
4885  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4886  * may return a partial prefix of the requested range.
4887  */
4888 int
4889 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4890     vm_page_t *ma, int count)
4891 {
4892 	vm_page_t m, mpred;
4893 	int pflags;
4894 	int i;
4895 
4896 	VM_OBJECT_ASSERT_WLOCKED(object);
4897 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4898 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4899 	KASSERT(count > 0,
4900 	    ("vm_page_grab_pages: invalid page count %d", count));
4901 	vm_page_grab_check(allocflags);
4902 
4903 	pflags = vm_page_grab_pflags(allocflags);
4904 	i = 0;
4905 retrylookup:
4906 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4907 	if (m == NULL || m->pindex != pindex + i) {
4908 		mpred = m;
4909 		m = NULL;
4910 	} else
4911 		mpred = TAILQ_PREV(m, pglist, listq);
4912 	for (; i < count; i++) {
4913 		if (m != NULL) {
4914 			if (!vm_page_tryacquire(m, allocflags)) {
4915 				if (vm_page_grab_sleep(object, m, pindex + i,
4916 				    "grbmaw", allocflags, true))
4917 					goto retrylookup;
4918 				break;
4919 			}
4920 		} else {
4921 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4922 				break;
4923 			m = vm_page_alloc_after(object, pindex + i,
4924 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4925 			if (m == NULL) {
4926 				if ((allocflags & (VM_ALLOC_NOWAIT |
4927 				    VM_ALLOC_WAITFAIL)) != 0)
4928 					break;
4929 				goto retrylookup;
4930 			}
4931 		}
4932 		if (vm_page_none_valid(m) &&
4933 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4934 			if ((m->flags & PG_ZERO) == 0)
4935 				pmap_zero_page(m);
4936 			vm_page_valid(m);
4937 		}
4938 		vm_page_grab_release(m, allocflags);
4939 		ma[i] = mpred = m;
4940 		m = vm_page_next(m);
4941 	}
4942 	return (i);
4943 }
4944 
4945 /*
4946  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
4947  * and will fall back to the locked variant to handle allocation.
4948  */
4949 int
4950 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
4951     int allocflags, vm_page_t *ma, int count)
4952 {
4953 	vm_page_t m, pred;
4954 	int flags;
4955 	int i;
4956 
4957 	KASSERT(count > 0,
4958 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
4959 	vm_page_grab_check(allocflags);
4960 
4961 	/*
4962 	 * Modify flags for lockless acquire to hold the page until we
4963 	 * set it valid if necessary.
4964 	 */
4965 	flags = allocflags & ~VM_ALLOC_NOBUSY;
4966 	pred = NULL;
4967 	for (i = 0; i < count; i++, pindex++) {
4968 		if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
4969 			return (i);
4970 		if (m == NULL)
4971 			break;
4972 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
4973 			if ((m->flags & PG_ZERO) == 0)
4974 				pmap_zero_page(m);
4975 			vm_page_valid(m);
4976 		}
4977 		/* m will still be wired or busy according to flags. */
4978 		vm_page_grab_release(m, allocflags);
4979 		pred = ma[i] = m;
4980 	}
4981 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
4982 		return (i);
4983 	count -= i;
4984 	VM_OBJECT_WLOCK(object);
4985 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
4986 	VM_OBJECT_WUNLOCK(object);
4987 
4988 	return (i);
4989 }
4990 
4991 /*
4992  * Mapping function for valid or dirty bits in a page.
4993  *
4994  * Inputs are required to range within a page.
4995  */
4996 vm_page_bits_t
4997 vm_page_bits(int base, int size)
4998 {
4999 	int first_bit;
5000 	int last_bit;
5001 
5002 	KASSERT(
5003 	    base + size <= PAGE_SIZE,
5004 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
5005 	);
5006 
5007 	if (size == 0)		/* handle degenerate case */
5008 		return (0);
5009 
5010 	first_bit = base >> DEV_BSHIFT;
5011 	last_bit = (base + size - 1) >> DEV_BSHIFT;
5012 
5013 	return (((vm_page_bits_t)2 << last_bit) -
5014 	    ((vm_page_bits_t)1 << first_bit));
5015 }
5016 
5017 void
5018 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
5019 {
5020 
5021 #if PAGE_SIZE == 32768
5022 	atomic_set_64((uint64_t *)bits, set);
5023 #elif PAGE_SIZE == 16384
5024 	atomic_set_32((uint32_t *)bits, set);
5025 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
5026 	atomic_set_16((uint16_t *)bits, set);
5027 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
5028 	atomic_set_8((uint8_t *)bits, set);
5029 #else		/* PAGE_SIZE <= 8192 */
5030 	uintptr_t addr;
5031 	int shift;
5032 
5033 	addr = (uintptr_t)bits;
5034 	/*
5035 	 * Use a trick to perform a 32-bit atomic on the
5036 	 * containing aligned word, to not depend on the existence
5037 	 * of atomic_{set, clear}_{8, 16}.
5038 	 */
5039 	shift = addr & (sizeof(uint32_t) - 1);
5040 #if BYTE_ORDER == BIG_ENDIAN
5041 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5042 #else
5043 	shift *= NBBY;
5044 #endif
5045 	addr &= ~(sizeof(uint32_t) - 1);
5046 	atomic_set_32((uint32_t *)addr, set << shift);
5047 #endif		/* PAGE_SIZE */
5048 }
5049 
5050 static inline void
5051 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
5052 {
5053 
5054 #if PAGE_SIZE == 32768
5055 	atomic_clear_64((uint64_t *)bits, clear);
5056 #elif PAGE_SIZE == 16384
5057 	atomic_clear_32((uint32_t *)bits, clear);
5058 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
5059 	atomic_clear_16((uint16_t *)bits, clear);
5060 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
5061 	atomic_clear_8((uint8_t *)bits, clear);
5062 #else		/* PAGE_SIZE <= 8192 */
5063 	uintptr_t addr;
5064 	int shift;
5065 
5066 	addr = (uintptr_t)bits;
5067 	/*
5068 	 * Use a trick to perform a 32-bit atomic on the
5069 	 * containing aligned word, to not depend on the existence
5070 	 * of atomic_{set, clear}_{8, 16}.
5071 	 */
5072 	shift = addr & (sizeof(uint32_t) - 1);
5073 #if BYTE_ORDER == BIG_ENDIAN
5074 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5075 #else
5076 	shift *= NBBY;
5077 #endif
5078 	addr &= ~(sizeof(uint32_t) - 1);
5079 	atomic_clear_32((uint32_t *)addr, clear << shift);
5080 #endif		/* PAGE_SIZE */
5081 }
5082 
5083 static inline vm_page_bits_t
5084 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5085 {
5086 #if PAGE_SIZE == 32768
5087 	uint64_t old;
5088 
5089 	old = *bits;
5090 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5091 	return (old);
5092 #elif PAGE_SIZE == 16384
5093 	uint32_t old;
5094 
5095 	old = *bits;
5096 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5097 	return (old);
5098 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5099 	uint16_t old;
5100 
5101 	old = *bits;
5102 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5103 	return (old);
5104 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5105 	uint8_t old;
5106 
5107 	old = *bits;
5108 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5109 	return (old);
5110 #else		/* PAGE_SIZE <= 4096*/
5111 	uintptr_t addr;
5112 	uint32_t old, new, mask;
5113 	int shift;
5114 
5115 	addr = (uintptr_t)bits;
5116 	/*
5117 	 * Use a trick to perform a 32-bit atomic on the
5118 	 * containing aligned word, to not depend on the existence
5119 	 * of atomic_{set, swap, clear}_{8, 16}.
5120 	 */
5121 	shift = addr & (sizeof(uint32_t) - 1);
5122 #if BYTE_ORDER == BIG_ENDIAN
5123 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5124 #else
5125 	shift *= NBBY;
5126 #endif
5127 	addr &= ~(sizeof(uint32_t) - 1);
5128 	mask = VM_PAGE_BITS_ALL << shift;
5129 
5130 	old = *bits;
5131 	do {
5132 		new = old & ~mask;
5133 		new |= newbits << shift;
5134 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5135 	return (old >> shift);
5136 #endif		/* PAGE_SIZE */
5137 }
5138 
5139 /*
5140  *	vm_page_set_valid_range:
5141  *
5142  *	Sets portions of a page valid.  The arguments are expected
5143  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5144  *	of any partial chunks touched by the range.  The invalid portion of
5145  *	such chunks will be zeroed.
5146  *
5147  *	(base + size) must be less then or equal to PAGE_SIZE.
5148  */
5149 void
5150 vm_page_set_valid_range(vm_page_t m, int base, int size)
5151 {
5152 	int endoff, frag;
5153 	vm_page_bits_t pagebits;
5154 
5155 	vm_page_assert_busied(m);
5156 	if (size == 0)	/* handle degenerate case */
5157 		return;
5158 
5159 	/*
5160 	 * If the base is not DEV_BSIZE aligned and the valid
5161 	 * bit is clear, we have to zero out a portion of the
5162 	 * first block.
5163 	 */
5164 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5165 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5166 		pmap_zero_page_area(m, frag, base - frag);
5167 
5168 	/*
5169 	 * If the ending offset is not DEV_BSIZE aligned and the
5170 	 * valid bit is clear, we have to zero out a portion of
5171 	 * the last block.
5172 	 */
5173 	endoff = base + size;
5174 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5175 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5176 		pmap_zero_page_area(m, endoff,
5177 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5178 
5179 	/*
5180 	 * Assert that no previously invalid block that is now being validated
5181 	 * is already dirty.
5182 	 */
5183 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5184 	    ("vm_page_set_valid_range: page %p is dirty", m));
5185 
5186 	/*
5187 	 * Set valid bits inclusive of any overlap.
5188 	 */
5189 	pagebits = vm_page_bits(base, size);
5190 	if (vm_page_xbusied(m))
5191 		m->valid |= pagebits;
5192 	else
5193 		vm_page_bits_set(m, &m->valid, pagebits);
5194 }
5195 
5196 /*
5197  * Set the page dirty bits and free the invalid swap space if
5198  * present.  Returns the previous dirty bits.
5199  */
5200 vm_page_bits_t
5201 vm_page_set_dirty(vm_page_t m)
5202 {
5203 	vm_page_bits_t old;
5204 
5205 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5206 
5207 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5208 		old = m->dirty;
5209 		m->dirty = VM_PAGE_BITS_ALL;
5210 	} else
5211 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5212 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5213 		vm_pager_page_unswapped(m);
5214 
5215 	return (old);
5216 }
5217 
5218 /*
5219  * Clear the given bits from the specified page's dirty field.
5220  */
5221 static __inline void
5222 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5223 {
5224 
5225 	vm_page_assert_busied(m);
5226 
5227 	/*
5228 	 * If the page is xbusied and not write mapped we are the
5229 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5230 	 * layer can call vm_page_dirty() without holding a distinguished
5231 	 * lock.  The combination of page busy and atomic operations
5232 	 * suffice to guarantee consistency of the page dirty field.
5233 	 */
5234 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5235 		m->dirty &= ~pagebits;
5236 	else
5237 		vm_page_bits_clear(m, &m->dirty, pagebits);
5238 }
5239 
5240 /*
5241  *	vm_page_set_validclean:
5242  *
5243  *	Sets portions of a page valid and clean.  The arguments are expected
5244  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5245  *	of any partial chunks touched by the range.  The invalid portion of
5246  *	such chunks will be zero'd.
5247  *
5248  *	(base + size) must be less then or equal to PAGE_SIZE.
5249  */
5250 void
5251 vm_page_set_validclean(vm_page_t m, int base, int size)
5252 {
5253 	vm_page_bits_t oldvalid, pagebits;
5254 	int endoff, frag;
5255 
5256 	vm_page_assert_busied(m);
5257 	if (size == 0)	/* handle degenerate case */
5258 		return;
5259 
5260 	/*
5261 	 * If the base is not DEV_BSIZE aligned and the valid
5262 	 * bit is clear, we have to zero out a portion of the
5263 	 * first block.
5264 	 */
5265 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5266 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5267 		pmap_zero_page_area(m, frag, base - frag);
5268 
5269 	/*
5270 	 * If the ending offset is not DEV_BSIZE aligned and the
5271 	 * valid bit is clear, we have to zero out a portion of
5272 	 * the last block.
5273 	 */
5274 	endoff = base + size;
5275 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5276 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5277 		pmap_zero_page_area(m, endoff,
5278 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5279 
5280 	/*
5281 	 * Set valid, clear dirty bits.  If validating the entire
5282 	 * page we can safely clear the pmap modify bit.  We also
5283 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5284 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5285 	 * be set again.
5286 	 *
5287 	 * We set valid bits inclusive of any overlap, but we can only
5288 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5289 	 * the range.
5290 	 */
5291 	oldvalid = m->valid;
5292 	pagebits = vm_page_bits(base, size);
5293 	if (vm_page_xbusied(m))
5294 		m->valid |= pagebits;
5295 	else
5296 		vm_page_bits_set(m, &m->valid, pagebits);
5297 #if 0	/* NOT YET */
5298 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5299 		frag = DEV_BSIZE - frag;
5300 		base += frag;
5301 		size -= frag;
5302 		if (size < 0)
5303 			size = 0;
5304 	}
5305 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5306 #endif
5307 	if (base == 0 && size == PAGE_SIZE) {
5308 		/*
5309 		 * The page can only be modified within the pmap if it is
5310 		 * mapped, and it can only be mapped if it was previously
5311 		 * fully valid.
5312 		 */
5313 		if (oldvalid == VM_PAGE_BITS_ALL)
5314 			/*
5315 			 * Perform the pmap_clear_modify() first.  Otherwise,
5316 			 * a concurrent pmap operation, such as
5317 			 * pmap_protect(), could clear a modification in the
5318 			 * pmap and set the dirty field on the page before
5319 			 * pmap_clear_modify() had begun and after the dirty
5320 			 * field was cleared here.
5321 			 */
5322 			pmap_clear_modify(m);
5323 		m->dirty = 0;
5324 		vm_page_aflag_clear(m, PGA_NOSYNC);
5325 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5326 		m->dirty &= ~pagebits;
5327 	else
5328 		vm_page_clear_dirty_mask(m, pagebits);
5329 }
5330 
5331 void
5332 vm_page_clear_dirty(vm_page_t m, int base, int size)
5333 {
5334 
5335 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5336 }
5337 
5338 /*
5339  *	vm_page_set_invalid:
5340  *
5341  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5342  *	valid and dirty bits for the effected areas are cleared.
5343  */
5344 void
5345 vm_page_set_invalid(vm_page_t m, int base, int size)
5346 {
5347 	vm_page_bits_t bits;
5348 	vm_object_t object;
5349 
5350 	/*
5351 	 * The object lock is required so that pages can't be mapped
5352 	 * read-only while we're in the process of invalidating them.
5353 	 */
5354 	object = m->object;
5355 	VM_OBJECT_ASSERT_WLOCKED(object);
5356 	vm_page_assert_busied(m);
5357 
5358 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5359 	    size >= object->un_pager.vnp.vnp_size)
5360 		bits = VM_PAGE_BITS_ALL;
5361 	else
5362 		bits = vm_page_bits(base, size);
5363 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5364 		pmap_remove_all(m);
5365 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5366 	    !pmap_page_is_mapped(m),
5367 	    ("vm_page_set_invalid: page %p is mapped", m));
5368 	if (vm_page_xbusied(m)) {
5369 		m->valid &= ~bits;
5370 		m->dirty &= ~bits;
5371 	} else {
5372 		vm_page_bits_clear(m, &m->valid, bits);
5373 		vm_page_bits_clear(m, &m->dirty, bits);
5374 	}
5375 }
5376 
5377 /*
5378  *	vm_page_invalid:
5379  *
5380  *	Invalidates the entire page.  The page must be busy, unmapped, and
5381  *	the enclosing object must be locked.  The object locks protects
5382  *	against concurrent read-only pmap enter which is done without
5383  *	busy.
5384  */
5385 void
5386 vm_page_invalid(vm_page_t m)
5387 {
5388 
5389 	vm_page_assert_busied(m);
5390 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5391 	MPASS(!pmap_page_is_mapped(m));
5392 
5393 	if (vm_page_xbusied(m))
5394 		m->valid = 0;
5395 	else
5396 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5397 }
5398 
5399 /*
5400  * vm_page_zero_invalid()
5401  *
5402  *	The kernel assumes that the invalid portions of a page contain
5403  *	garbage, but such pages can be mapped into memory by user code.
5404  *	When this occurs, we must zero out the non-valid portions of the
5405  *	page so user code sees what it expects.
5406  *
5407  *	Pages are most often semi-valid when the end of a file is mapped
5408  *	into memory and the file's size is not page aligned.
5409  */
5410 void
5411 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5412 {
5413 	int b;
5414 	int i;
5415 
5416 	/*
5417 	 * Scan the valid bits looking for invalid sections that
5418 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5419 	 * valid bit may be set ) have already been zeroed by
5420 	 * vm_page_set_validclean().
5421 	 */
5422 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5423 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5424 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5425 			if (i > b) {
5426 				pmap_zero_page_area(m,
5427 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5428 			}
5429 			b = i + 1;
5430 		}
5431 	}
5432 
5433 	/*
5434 	 * setvalid is TRUE when we can safely set the zero'd areas
5435 	 * as being valid.  We can do this if there are no cache consistency
5436 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5437 	 */
5438 	if (setvalid)
5439 		vm_page_valid(m);
5440 }
5441 
5442 /*
5443  *	vm_page_is_valid:
5444  *
5445  *	Is (partial) page valid?  Note that the case where size == 0
5446  *	will return FALSE in the degenerate case where the page is
5447  *	entirely invalid, and TRUE otherwise.
5448  *
5449  *	Some callers envoke this routine without the busy lock held and
5450  *	handle races via higher level locks.  Typical callers should
5451  *	hold a busy lock to prevent invalidation.
5452  */
5453 int
5454 vm_page_is_valid(vm_page_t m, int base, int size)
5455 {
5456 	vm_page_bits_t bits;
5457 
5458 	bits = vm_page_bits(base, size);
5459 	return (vm_page_any_valid(m) && (m->valid & bits) == bits);
5460 }
5461 
5462 /*
5463  * Returns true if all of the specified predicates are true for the entire
5464  * (super)page and false otherwise.
5465  */
5466 bool
5467 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5468 {
5469 	vm_object_t object;
5470 	int i, npages;
5471 
5472 	object = m->object;
5473 	if (skip_m != NULL && skip_m->object != object)
5474 		return (false);
5475 	VM_OBJECT_ASSERT_LOCKED(object);
5476 	npages = atop(pagesizes[m->psind]);
5477 
5478 	/*
5479 	 * The physically contiguous pages that make up a superpage, i.e., a
5480 	 * page with a page size index ("psind") greater than zero, will
5481 	 * occupy adjacent entries in vm_page_array[].
5482 	 */
5483 	for (i = 0; i < npages; i++) {
5484 		/* Always test object consistency, including "skip_m". */
5485 		if (m[i].object != object)
5486 			return (false);
5487 		if (&m[i] == skip_m)
5488 			continue;
5489 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5490 			return (false);
5491 		if ((flags & PS_ALL_DIRTY) != 0) {
5492 			/*
5493 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5494 			 * might stop this case from spuriously returning
5495 			 * "false".  However, that would require a write lock
5496 			 * on the object containing "m[i]".
5497 			 */
5498 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5499 				return (false);
5500 		}
5501 		if ((flags & PS_ALL_VALID) != 0 &&
5502 		    m[i].valid != VM_PAGE_BITS_ALL)
5503 			return (false);
5504 	}
5505 	return (true);
5506 }
5507 
5508 /*
5509  * Set the page's dirty bits if the page is modified.
5510  */
5511 void
5512 vm_page_test_dirty(vm_page_t m)
5513 {
5514 
5515 	vm_page_assert_busied(m);
5516 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5517 		vm_page_dirty(m);
5518 }
5519 
5520 void
5521 vm_page_valid(vm_page_t m)
5522 {
5523 
5524 	vm_page_assert_busied(m);
5525 	if (vm_page_xbusied(m))
5526 		m->valid = VM_PAGE_BITS_ALL;
5527 	else
5528 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5529 }
5530 
5531 void
5532 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5533 {
5534 
5535 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5536 }
5537 
5538 void
5539 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5540 {
5541 
5542 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5543 }
5544 
5545 int
5546 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5547 {
5548 
5549 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5550 }
5551 
5552 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5553 void
5554 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5555 {
5556 
5557 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5558 }
5559 
5560 void
5561 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5562 {
5563 
5564 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5565 }
5566 #endif
5567 
5568 #ifdef INVARIANTS
5569 void
5570 vm_page_object_busy_assert(vm_page_t m)
5571 {
5572 
5573 	/*
5574 	 * Certain of the page's fields may only be modified by the
5575 	 * holder of a page or object busy.
5576 	 */
5577 	if (m->object != NULL && !vm_page_busied(m))
5578 		VM_OBJECT_ASSERT_BUSY(m->object);
5579 }
5580 
5581 void
5582 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5583 {
5584 
5585 	if ((bits & PGA_WRITEABLE) == 0)
5586 		return;
5587 
5588 	/*
5589 	 * The PGA_WRITEABLE flag can only be set if the page is
5590 	 * managed, is exclusively busied or the object is locked.
5591 	 * Currently, this flag is only set by pmap_enter().
5592 	 */
5593 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5594 	    ("PGA_WRITEABLE on unmanaged page"));
5595 	if (!vm_page_xbusied(m))
5596 		VM_OBJECT_ASSERT_BUSY(m->object);
5597 }
5598 #endif
5599 
5600 #include "opt_ddb.h"
5601 #ifdef DDB
5602 #include <sys/kernel.h>
5603 
5604 #include <ddb/ddb.h>
5605 
5606 DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
5607 {
5608 
5609 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5610 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5611 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5612 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5613 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5614 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5615 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5616 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5617 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5618 }
5619 
5620 DB_SHOW_COMMAND_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
5621 {
5622 	int dom;
5623 
5624 	db_printf("pq_free %d\n", vm_free_count());
5625 	for (dom = 0; dom < vm_ndomains; dom++) {
5626 		db_printf(
5627     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5628 		    dom,
5629 		    vm_dom[dom].vmd_page_count,
5630 		    vm_dom[dom].vmd_free_count,
5631 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5632 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5633 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5634 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5635 	}
5636 }
5637 
5638 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5639 {
5640 	vm_page_t m;
5641 	boolean_t phys, virt;
5642 
5643 	if (!have_addr) {
5644 		db_printf("show pginfo addr\n");
5645 		return;
5646 	}
5647 
5648 	phys = strchr(modif, 'p') != NULL;
5649 	virt = strchr(modif, 'v') != NULL;
5650 	if (virt)
5651 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5652 	else if (phys)
5653 		m = PHYS_TO_VM_PAGE(addr);
5654 	else
5655 		m = (vm_page_t)addr;
5656 	db_printf(
5657     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5658     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5659 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5660 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5661 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5662 }
5663 #endif /* DDB */
5664