xref: /freebsd/sys/vm/vm_page.c (revision b111430e54537a66080df115c3a2950a2922da5a)
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 __FBSDID("$FreeBSD$");
71 
72 #include "opt_vm.h"
73 
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/counter.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
81 #include <sys/lock.h>
82 #include <sys/malloc.h>
83 #include <sys/mman.h>
84 #include <sys/msgbuf.h>
85 #include <sys/mutex.h>
86 #include <sys/proc.h>
87 #include <sys/rwlock.h>
88 #include <sys/sleepqueue.h>
89 #include <sys/sbuf.h>
90 #include <sys/sched.h>
91 #include <sys/smp.h>
92 #include <sys/sysctl.h>
93 #include <sys/vmmeter.h>
94 #include <sys/vnode.h>
95 
96 #include <vm/vm.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_domainset.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_pagequeue.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/vm_extern.h>
111 #include <vm/vm_dumpset.h>
112 #include <vm/uma.h>
113 #include <vm/uma_int.h>
114 
115 #include <machine/md_var.h>
116 
117 struct vm_domain vm_dom[MAXMEMDOM];
118 
119 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
120 
121 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
122 
123 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
124 /* The following fields are protected by the domainset lock. */
125 domainset_t __exclusive_cache_line vm_min_domains;
126 domainset_t __exclusive_cache_line vm_severe_domains;
127 static int vm_min_waiters;
128 static int vm_severe_waiters;
129 static int vm_pageproc_waiters;
130 
131 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
132     "VM page statistics");
133 
134 static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
135 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
136     CTLFLAG_RD, &pqstate_commit_retries,
137     "Number of failed per-page atomic queue state updates");
138 
139 static COUNTER_U64_DEFINE_EARLY(queue_ops);
140 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
141     CTLFLAG_RD, &queue_ops,
142     "Number of batched queue operations");
143 
144 static COUNTER_U64_DEFINE_EARLY(queue_nops);
145 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
146     CTLFLAG_RD, &queue_nops,
147     "Number of batched queue operations with no effects");
148 
149 /*
150  * bogus page -- for I/O to/from partially complete buffers,
151  * or for paging into sparsely invalid regions.
152  */
153 vm_page_t bogus_page;
154 
155 vm_page_t vm_page_array;
156 long vm_page_array_size;
157 long first_page;
158 
159 struct bitset *vm_page_dump;
160 long vm_page_dump_pages;
161 
162 static TAILQ_HEAD(, vm_page) blacklist_head;
163 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
164 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
165     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
166 
167 static uma_zone_t fakepg_zone;
168 
169 static void vm_page_alloc_check(vm_page_t m);
170 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
171     vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
172 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
173 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
174 static bool vm_page_free_prep(vm_page_t m);
175 static void vm_page_free_toq(vm_page_t m);
176 static void vm_page_init(void *dummy);
177 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
178     vm_pindex_t pindex, vm_page_t mpred);
179 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
180     vm_page_t mpred);
181 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
182     const uint16_t nflag);
183 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
184     vm_page_t m_run, vm_paddr_t high);
185 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
186 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
187     int req);
188 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
189     int flags);
190 static void vm_page_zone_release(void *arg, void **store, int cnt);
191 
192 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
193 
194 static void
195 vm_page_init(void *dummy)
196 {
197 
198 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
199 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
200 	bogus_page = vm_page_alloc_noobj(VM_ALLOC_WIRED);
201 }
202 
203 /*
204  * The cache page zone is initialized later since we need to be able to allocate
205  * pages before UMA is fully initialized.
206  */
207 static void
208 vm_page_init_cache_zones(void *dummy __unused)
209 {
210 	struct vm_domain *vmd;
211 	struct vm_pgcache *pgcache;
212 	int cache, domain, maxcache, pool;
213 
214 	maxcache = 0;
215 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
216 	maxcache *= mp_ncpus;
217 	for (domain = 0; domain < vm_ndomains; domain++) {
218 		vmd = VM_DOMAIN(domain);
219 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
220 			pgcache = &vmd->vmd_pgcache[pool];
221 			pgcache->domain = domain;
222 			pgcache->pool = pool;
223 			pgcache->zone = uma_zcache_create("vm pgcache",
224 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
225 			    vm_page_zone_import, vm_page_zone_release, pgcache,
226 			    UMA_ZONE_VM);
227 
228 			/*
229 			 * Limit each pool's zone to 0.1% of the pages in the
230 			 * domain.
231 			 */
232 			cache = maxcache != 0 ? maxcache :
233 			    vmd->vmd_page_count / 1000;
234 			uma_zone_set_maxcache(pgcache->zone, cache);
235 		}
236 	}
237 }
238 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
239 
240 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
241 #if PAGE_SIZE == 32768
242 #ifdef CTASSERT
243 CTASSERT(sizeof(u_long) >= 8);
244 #endif
245 #endif
246 
247 /*
248  *	vm_set_page_size:
249  *
250  *	Sets the page size, perhaps based upon the memory
251  *	size.  Must be called before any use of page-size
252  *	dependent functions.
253  */
254 void
255 vm_set_page_size(void)
256 {
257 	if (vm_cnt.v_page_size == 0)
258 		vm_cnt.v_page_size = PAGE_SIZE;
259 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
260 		panic("vm_set_page_size: page size not a power of two");
261 }
262 
263 /*
264  *	vm_page_blacklist_next:
265  *
266  *	Find the next entry in the provided string of blacklist
267  *	addresses.  Entries are separated by space, comma, or newline.
268  *	If an invalid integer is encountered then the rest of the
269  *	string is skipped.  Updates the list pointer to the next
270  *	character, or NULL if the string is exhausted or invalid.
271  */
272 static vm_paddr_t
273 vm_page_blacklist_next(char **list, char *end)
274 {
275 	vm_paddr_t bad;
276 	char *cp, *pos;
277 
278 	if (list == NULL || *list == NULL)
279 		return (0);
280 	if (**list =='\0') {
281 		*list = NULL;
282 		return (0);
283 	}
284 
285 	/*
286 	 * If there's no end pointer then the buffer is coming from
287 	 * the kenv and we know it's null-terminated.
288 	 */
289 	if (end == NULL)
290 		end = *list + strlen(*list);
291 
292 	/* Ensure that strtoq() won't walk off the end */
293 	if (*end != '\0') {
294 		if (*end == '\n' || *end == ' ' || *end  == ',')
295 			*end = '\0';
296 		else {
297 			printf("Blacklist not terminated, skipping\n");
298 			*list = NULL;
299 			return (0);
300 		}
301 	}
302 
303 	for (pos = *list; *pos != '\0'; pos = cp) {
304 		bad = strtoq(pos, &cp, 0);
305 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
306 			if (bad == 0) {
307 				if (++cp < end)
308 					continue;
309 				else
310 					break;
311 			}
312 		} else
313 			break;
314 		if (*cp == '\0' || ++cp >= end)
315 			*list = NULL;
316 		else
317 			*list = cp;
318 		return (trunc_page(bad));
319 	}
320 	printf("Garbage in RAM blacklist, skipping\n");
321 	*list = NULL;
322 	return (0);
323 }
324 
325 bool
326 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
327 {
328 	struct vm_domain *vmd;
329 	vm_page_t m;
330 	int ret;
331 
332 	m = vm_phys_paddr_to_vm_page(pa);
333 	if (m == NULL)
334 		return (true); /* page does not exist, no failure */
335 
336 	vmd = vm_pagequeue_domain(m);
337 	vm_domain_free_lock(vmd);
338 	ret = vm_phys_unfree_page(m);
339 	vm_domain_free_unlock(vmd);
340 	if (ret != 0) {
341 		vm_domain_freecnt_inc(vmd, -1);
342 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
343 		if (verbose)
344 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
345 	}
346 	return (ret);
347 }
348 
349 /*
350  *	vm_page_blacklist_check:
351  *
352  *	Iterate through the provided string of blacklist addresses, pulling
353  *	each entry out of the physical allocator free list and putting it
354  *	onto a list for reporting via the vm.page_blacklist sysctl.
355  */
356 static void
357 vm_page_blacklist_check(char *list, char *end)
358 {
359 	vm_paddr_t pa;
360 	char *next;
361 
362 	next = list;
363 	while (next != NULL) {
364 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
365 			continue;
366 		vm_page_blacklist_add(pa, bootverbose);
367 	}
368 }
369 
370 /*
371  *	vm_page_blacklist_load:
372  *
373  *	Search for a special module named "ram_blacklist".  It'll be a
374  *	plain text file provided by the user via the loader directive
375  *	of the same name.
376  */
377 static void
378 vm_page_blacklist_load(char **list, char **end)
379 {
380 	void *mod;
381 	u_char *ptr;
382 	u_int len;
383 
384 	mod = NULL;
385 	ptr = NULL;
386 
387 	mod = preload_search_by_type("ram_blacklist");
388 	if (mod != NULL) {
389 		ptr = preload_fetch_addr(mod);
390 		len = preload_fetch_size(mod);
391         }
392 	*list = ptr;
393 	if (ptr != NULL)
394 		*end = ptr + len;
395 	else
396 		*end = NULL;
397 	return;
398 }
399 
400 static int
401 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
402 {
403 	vm_page_t m;
404 	struct sbuf sbuf;
405 	int error, first;
406 
407 	first = 1;
408 	error = sysctl_wire_old_buffer(req, 0);
409 	if (error != 0)
410 		return (error);
411 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
412 	TAILQ_FOREACH(m, &blacklist_head, listq) {
413 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
414 		    (uintmax_t)m->phys_addr);
415 		first = 0;
416 	}
417 	error = sbuf_finish(&sbuf);
418 	sbuf_delete(&sbuf);
419 	return (error);
420 }
421 
422 /*
423  * Initialize a dummy page for use in scans of the specified paging queue.
424  * In principle, this function only needs to set the flag PG_MARKER.
425  * Nonetheless, it write busies the page as a safety precaution.
426  */
427 void
428 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
429 {
430 
431 	bzero(marker, sizeof(*marker));
432 	marker->flags = PG_MARKER;
433 	marker->a.flags = aflags;
434 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
435 	marker->a.queue = queue;
436 }
437 
438 static void
439 vm_page_domain_init(int domain)
440 {
441 	struct vm_domain *vmd;
442 	struct vm_pagequeue *pq;
443 	int i;
444 
445 	vmd = VM_DOMAIN(domain);
446 	bzero(vmd, sizeof(*vmd));
447 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
448 	    "vm inactive pagequeue";
449 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
450 	    "vm active pagequeue";
451 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
452 	    "vm laundry pagequeue";
453 	*__DECONST(const char **,
454 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
455 	    "vm unswappable pagequeue";
456 	vmd->vmd_domain = domain;
457 	vmd->vmd_page_count = 0;
458 	vmd->vmd_free_count = 0;
459 	vmd->vmd_segs = 0;
460 	vmd->vmd_oom = FALSE;
461 	for (i = 0; i < PQ_COUNT; i++) {
462 		pq = &vmd->vmd_pagequeues[i];
463 		TAILQ_INIT(&pq->pq_pl);
464 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
465 		    MTX_DEF | MTX_DUPOK);
466 		pq->pq_pdpages = 0;
467 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
468 	}
469 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
470 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
471 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
472 
473 	/*
474 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
475 	 * insertions.
476 	 */
477 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
478 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
479 	    &vmd->vmd_inacthead, plinks.q);
480 
481 	/*
482 	 * The clock pages are used to implement active queue scanning without
483 	 * requeues.  Scans start at clock[0], which is advanced after the scan
484 	 * ends.  When the two clock hands meet, they are reset and scanning
485 	 * resumes from the head of the queue.
486 	 */
487 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
488 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
489 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
490 	    &vmd->vmd_clock[0], plinks.q);
491 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
492 	    &vmd->vmd_clock[1], plinks.q);
493 }
494 
495 /*
496  * Initialize a physical page in preparation for adding it to the free
497  * lists.
498  */
499 void
500 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
501 {
502 
503 	m->object = NULL;
504 	m->ref_count = 0;
505 	m->busy_lock = VPB_FREED;
506 	m->flags = m->a.flags = 0;
507 	m->phys_addr = pa;
508 	m->a.queue = PQ_NONE;
509 	m->psind = 0;
510 	m->segind = segind;
511 	m->order = VM_NFREEORDER;
512 	m->pool = VM_FREEPOOL_DEFAULT;
513 	m->valid = m->dirty = 0;
514 	pmap_page_init(m);
515 }
516 
517 #ifndef PMAP_HAS_PAGE_ARRAY
518 static vm_paddr_t
519 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
520 {
521 	vm_paddr_t new_end;
522 
523 	/*
524 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
525 	 * However, because this page is allocated from KVM, out-of-bounds
526 	 * accesses using the direct map will not be trapped.
527 	 */
528 	*vaddr += PAGE_SIZE;
529 
530 	/*
531 	 * Allocate physical memory for the page structures, and map it.
532 	 */
533 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
534 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
535 	    VM_PROT_READ | VM_PROT_WRITE);
536 	vm_page_array_size = page_range;
537 
538 	return (new_end);
539 }
540 #endif
541 
542 /*
543  *	vm_page_startup:
544  *
545  *	Initializes the resident memory module.  Allocates physical memory for
546  *	bootstrapping UMA and some data structures that are used to manage
547  *	physical pages.  Initializes these structures, and populates the free
548  *	page queues.
549  */
550 vm_offset_t
551 vm_page_startup(vm_offset_t vaddr)
552 {
553 	struct vm_phys_seg *seg;
554 	struct vm_domain *vmd;
555 	vm_page_t m;
556 	char *list, *listend;
557 	vm_paddr_t end, high_avail, low_avail, new_end, size;
558 	vm_paddr_t page_range __unused;
559 	vm_paddr_t last_pa, pa, startp, endp;
560 	u_long pagecount;
561 #if MINIDUMP_PAGE_TRACKING
562 	u_long vm_page_dump_size;
563 #endif
564 	int biggestone, i, segind;
565 #ifdef WITNESS
566 	vm_offset_t mapped;
567 	int witness_size;
568 #endif
569 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
570 	long ii;
571 #endif
572 
573 	vaddr = round_page(vaddr);
574 
575 	vm_phys_early_startup();
576 	biggestone = vm_phys_avail_largest();
577 	end = phys_avail[biggestone+1];
578 
579 	/*
580 	 * Initialize the page and queue locks.
581 	 */
582 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
583 	for (i = 0; i < PA_LOCK_COUNT; i++)
584 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
585 	for (i = 0; i < vm_ndomains; i++)
586 		vm_page_domain_init(i);
587 
588 	new_end = end;
589 #ifdef WITNESS
590 	witness_size = round_page(witness_startup_count());
591 	new_end -= witness_size;
592 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
593 	    VM_PROT_READ | VM_PROT_WRITE);
594 	bzero((void *)mapped, witness_size);
595 	witness_startup((void *)mapped);
596 #endif
597 
598 #if MINIDUMP_PAGE_TRACKING
599 	/*
600 	 * Allocate a bitmap to indicate that a random physical page
601 	 * needs to be included in a minidump.
602 	 *
603 	 * The amd64 port needs this to indicate which direct map pages
604 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
605 	 *
606 	 * However, i386 still needs this workspace internally within the
607 	 * minidump code.  In theory, they are not needed on i386, but are
608 	 * included should the sf_buf code decide to use them.
609 	 */
610 	last_pa = 0;
611 	vm_page_dump_pages = 0;
612 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
613 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
614 		    dump_avail[i] / PAGE_SIZE;
615 		if (dump_avail[i + 1] > last_pa)
616 			last_pa = dump_avail[i + 1];
617 	}
618 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
619 	new_end -= vm_page_dump_size;
620 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
621 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
622 	bzero((void *)vm_page_dump, vm_page_dump_size);
623 #else
624 	(void)last_pa;
625 #endif
626 #if defined(__aarch64__) || defined(__amd64__) || \
627     defined(__riscv) || defined(__powerpc64__)
628 	/*
629 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
630 	 * in a crash dump.  When pmap_map() uses the direct map, they are
631 	 * not automatically included.
632 	 */
633 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
634 		dump_add_page(pa);
635 #endif
636 	phys_avail[biggestone + 1] = new_end;
637 #ifdef __amd64__
638 	/*
639 	 * Request that the physical pages underlying the message buffer be
640 	 * included in a crash dump.  Since the message buffer is accessed
641 	 * through the direct map, they are not automatically included.
642 	 */
643 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
644 	last_pa = pa + round_page(msgbufsize);
645 	while (pa < last_pa) {
646 		dump_add_page(pa);
647 		pa += PAGE_SIZE;
648 	}
649 #endif
650 	/*
651 	 * Compute the number of pages of memory that will be available for
652 	 * use, taking into account the overhead of a page structure per page.
653 	 * In other words, solve
654 	 *	"available physical memory" - round_page(page_range *
655 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
656 	 * for page_range.
657 	 */
658 	low_avail = phys_avail[0];
659 	high_avail = phys_avail[1];
660 	for (i = 0; i < vm_phys_nsegs; i++) {
661 		if (vm_phys_segs[i].start < low_avail)
662 			low_avail = vm_phys_segs[i].start;
663 		if (vm_phys_segs[i].end > high_avail)
664 			high_avail = vm_phys_segs[i].end;
665 	}
666 	/* Skip the first chunk.  It is already accounted for. */
667 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
668 		if (phys_avail[i] < low_avail)
669 			low_avail = phys_avail[i];
670 		if (phys_avail[i + 1] > high_avail)
671 			high_avail = phys_avail[i + 1];
672 	}
673 	first_page = low_avail / PAGE_SIZE;
674 #ifdef VM_PHYSSEG_SPARSE
675 	size = 0;
676 	for (i = 0; i < vm_phys_nsegs; i++)
677 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
678 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
679 		size += phys_avail[i + 1] - phys_avail[i];
680 #elif defined(VM_PHYSSEG_DENSE)
681 	size = high_avail - low_avail;
682 #else
683 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
684 #endif
685 
686 #ifdef PMAP_HAS_PAGE_ARRAY
687 	pmap_page_array_startup(size / PAGE_SIZE);
688 	biggestone = vm_phys_avail_largest();
689 	end = new_end = phys_avail[biggestone + 1];
690 #else
691 #ifdef VM_PHYSSEG_DENSE
692 	/*
693 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
694 	 * the overhead of a page structure per page only if vm_page_array is
695 	 * allocated from the last physical memory chunk.  Otherwise, we must
696 	 * allocate page structures representing the physical memory
697 	 * underlying vm_page_array, even though they will not be used.
698 	 */
699 	if (new_end != high_avail)
700 		page_range = size / PAGE_SIZE;
701 	else
702 #endif
703 	{
704 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
705 
706 		/*
707 		 * If the partial bytes remaining are large enough for
708 		 * a page (PAGE_SIZE) without a corresponding
709 		 * 'struct vm_page', then new_end will contain an
710 		 * extra page after subtracting the length of the VM
711 		 * page array.  Compensate by subtracting an extra
712 		 * page from new_end.
713 		 */
714 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
715 			if (new_end == high_avail)
716 				high_avail -= PAGE_SIZE;
717 			new_end -= PAGE_SIZE;
718 		}
719 	}
720 	end = new_end;
721 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
722 #endif
723 
724 #if VM_NRESERVLEVEL > 0
725 	/*
726 	 * Allocate physical memory for the reservation management system's
727 	 * data structures, and map it.
728 	 */
729 	new_end = vm_reserv_startup(&vaddr, new_end);
730 #endif
731 #if defined(__aarch64__) || defined(__amd64__) || \
732     defined(__riscv) || defined(__powerpc64__)
733 	/*
734 	 * Include vm_page_array and vm_reserv_array in a crash dump.
735 	 */
736 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
737 		dump_add_page(pa);
738 #endif
739 	phys_avail[biggestone + 1] = new_end;
740 
741 	/*
742 	 * Add physical memory segments corresponding to the available
743 	 * physical pages.
744 	 */
745 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
746 		if (vm_phys_avail_size(i) != 0)
747 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
748 
749 	/*
750 	 * Initialize the physical memory allocator.
751 	 */
752 	vm_phys_init();
753 
754 	/*
755 	 * Initialize the page structures and add every available page to the
756 	 * physical memory allocator's free lists.
757 	 */
758 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
759 	for (ii = 0; ii < vm_page_array_size; ii++) {
760 		m = &vm_page_array[ii];
761 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
762 		m->flags = PG_FICTITIOUS;
763 	}
764 #endif
765 	vm_cnt.v_page_count = 0;
766 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
767 		seg = &vm_phys_segs[segind];
768 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
769 		    m++, pa += PAGE_SIZE)
770 			vm_page_init_page(m, pa, segind);
771 
772 		/*
773 		 * Add the segment's pages that are covered by one of
774 		 * phys_avail's ranges to the free lists.
775 		 */
776 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
777 			if (seg->end <= phys_avail[i] ||
778 			    seg->start >= phys_avail[i + 1])
779 				continue;
780 
781 			startp = MAX(seg->start, phys_avail[i]);
782 			endp = MIN(seg->end, phys_avail[i + 1]);
783 			pagecount = (u_long)atop(endp - startp);
784 			if (pagecount == 0)
785 				continue;
786 
787 			m = seg->first_page + atop(startp - seg->start);
788 			vmd = VM_DOMAIN(seg->domain);
789 			vm_domain_free_lock(vmd);
790 			vm_phys_enqueue_contig(m, pagecount);
791 			vm_domain_free_unlock(vmd);
792 			vm_domain_freecnt_inc(vmd, pagecount);
793 			vm_cnt.v_page_count += (u_int)pagecount;
794 			vmd->vmd_page_count += (u_int)pagecount;
795 			vmd->vmd_segs |= 1UL << segind;
796 		}
797 	}
798 
799 	/*
800 	 * Remove blacklisted pages from the physical memory allocator.
801 	 */
802 	TAILQ_INIT(&blacklist_head);
803 	vm_page_blacklist_load(&list, &listend);
804 	vm_page_blacklist_check(list, listend);
805 
806 	list = kern_getenv("vm.blacklist");
807 	vm_page_blacklist_check(list, NULL);
808 
809 	freeenv(list);
810 #if VM_NRESERVLEVEL > 0
811 	/*
812 	 * Initialize the reservation management system.
813 	 */
814 	vm_reserv_init();
815 #endif
816 
817 	return (vaddr);
818 }
819 
820 void
821 vm_page_reference(vm_page_t m)
822 {
823 
824 	vm_page_aflag_set(m, PGA_REFERENCED);
825 }
826 
827 /*
828  *	vm_page_trybusy
829  *
830  *	Helper routine for grab functions to trylock busy.
831  *
832  *	Returns true on success and false on failure.
833  */
834 static bool
835 vm_page_trybusy(vm_page_t m, int allocflags)
836 {
837 
838 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
839 		return (vm_page_trysbusy(m));
840 	else
841 		return (vm_page_tryxbusy(m));
842 }
843 
844 /*
845  *	vm_page_tryacquire
846  *
847  *	Helper routine for grab functions to trylock busy and wire.
848  *
849  *	Returns true on success and false on failure.
850  */
851 static inline bool
852 vm_page_tryacquire(vm_page_t m, int allocflags)
853 {
854 	bool locked;
855 
856 	locked = vm_page_trybusy(m, allocflags);
857 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
858 		vm_page_wire(m);
859 	return (locked);
860 }
861 
862 /*
863  *	vm_page_busy_acquire:
864  *
865  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
866  *	and drop the object lock if necessary.
867  */
868 bool
869 vm_page_busy_acquire(vm_page_t m, int allocflags)
870 {
871 	vm_object_t obj;
872 	bool locked;
873 
874 	/*
875 	 * The page-specific object must be cached because page
876 	 * identity can change during the sleep, causing the
877 	 * re-lock of a different object.
878 	 * It is assumed that a reference to the object is already
879 	 * held by the callers.
880 	 */
881 	obj = atomic_load_ptr(&m->object);
882 	for (;;) {
883 		if (vm_page_tryacquire(m, allocflags))
884 			return (true);
885 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
886 			return (false);
887 		if (obj != NULL)
888 			locked = VM_OBJECT_WOWNED(obj);
889 		else
890 			locked = false;
891 		MPASS(locked || vm_page_wired(m));
892 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
893 		    locked) && locked)
894 			VM_OBJECT_WLOCK(obj);
895 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
896 			return (false);
897 		KASSERT(m->object == obj || m->object == NULL,
898 		    ("vm_page_busy_acquire: page %p does not belong to %p",
899 		    m, obj));
900 	}
901 }
902 
903 /*
904  *	vm_page_busy_downgrade:
905  *
906  *	Downgrade an exclusive busy page into a single shared busy page.
907  */
908 void
909 vm_page_busy_downgrade(vm_page_t m)
910 {
911 	u_int x;
912 
913 	vm_page_assert_xbusied(m);
914 
915 	x = vm_page_busy_fetch(m);
916 	for (;;) {
917 		if (atomic_fcmpset_rel_int(&m->busy_lock,
918 		    &x, VPB_SHARERS_WORD(1)))
919 			break;
920 	}
921 	if ((x & VPB_BIT_WAITERS) != 0)
922 		wakeup(m);
923 }
924 
925 /*
926  *
927  *	vm_page_busy_tryupgrade:
928  *
929  *	Attempt to upgrade a single shared busy into an exclusive busy.
930  */
931 int
932 vm_page_busy_tryupgrade(vm_page_t m)
933 {
934 	u_int ce, x;
935 
936 	vm_page_assert_sbusied(m);
937 
938 	x = vm_page_busy_fetch(m);
939 	ce = VPB_CURTHREAD_EXCLUSIVE;
940 	for (;;) {
941 		if (VPB_SHARERS(x) > 1)
942 			return (0);
943 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
944 		    ("vm_page_busy_tryupgrade: invalid lock state"));
945 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
946 		    ce | (x & VPB_BIT_WAITERS)))
947 			continue;
948 		return (1);
949 	}
950 }
951 
952 /*
953  *	vm_page_sbusied:
954  *
955  *	Return a positive value if the page is shared busied, 0 otherwise.
956  */
957 int
958 vm_page_sbusied(vm_page_t m)
959 {
960 	u_int x;
961 
962 	x = vm_page_busy_fetch(m);
963 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
964 }
965 
966 /*
967  *	vm_page_sunbusy:
968  *
969  *	Shared unbusy a page.
970  */
971 void
972 vm_page_sunbusy(vm_page_t m)
973 {
974 	u_int x;
975 
976 	vm_page_assert_sbusied(m);
977 
978 	x = vm_page_busy_fetch(m);
979 	for (;;) {
980 		KASSERT(x != VPB_FREED,
981 		    ("vm_page_sunbusy: Unlocking freed page."));
982 		if (VPB_SHARERS(x) > 1) {
983 			if (atomic_fcmpset_int(&m->busy_lock, &x,
984 			    x - VPB_ONE_SHARER))
985 				break;
986 			continue;
987 		}
988 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
989 		    ("vm_page_sunbusy: invalid lock state"));
990 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
991 			continue;
992 		if ((x & VPB_BIT_WAITERS) == 0)
993 			break;
994 		wakeup(m);
995 		break;
996 	}
997 }
998 
999 /*
1000  *	vm_page_busy_sleep:
1001  *
1002  *	Sleep if the page is busy, using the page pointer as wchan.
1003  *	This is used to implement the hard-path of the busying mechanism.
1004  *
1005  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1006  *	will not sleep if the page is shared-busy.
1007  *
1008  *	The object lock must be held on entry.
1009  *
1010  *	Returns true if it slept and dropped the object lock, or false
1011  *	if there was no sleep and the lock is still held.
1012  */
1013 bool
1014 vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
1015 {
1016 	vm_object_t obj;
1017 
1018 	obj = m->object;
1019 	VM_OBJECT_ASSERT_LOCKED(obj);
1020 
1021 	return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
1022 	    true));
1023 }
1024 
1025 /*
1026  *	vm_page_busy_sleep_unlocked:
1027  *
1028  *	Sleep if the page is busy, using the page pointer as wchan.
1029  *	This is used to implement the hard-path of busying mechanism.
1030  *
1031  *	If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
1032  *	will not sleep if the page is shared-busy.
1033  *
1034  *	The object lock must not be held on entry.  The operation will
1035  *	return if the page changes identity.
1036  */
1037 void
1038 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1039     const char *wmesg, int allocflags)
1040 {
1041 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1042 
1043 	(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
1044 }
1045 
1046 /*
1047  *	_vm_page_busy_sleep:
1048  *
1049  *	Internal busy sleep function.  Verifies the page identity and
1050  *	lockstate against parameters.  Returns true if it sleeps and
1051  *	false otherwise.
1052  *
1053  *	allocflags uses VM_ALLOC_* flags to specify the lock required.
1054  *
1055  *	If locked is true the lock will be dropped for any true returns
1056  *	and held for any false returns.
1057  */
1058 static bool
1059 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1060     const char *wmesg, int allocflags, bool locked)
1061 {
1062 	bool xsleep;
1063 	u_int x;
1064 
1065 	/*
1066 	 * If the object is busy we must wait for that to drain to zero
1067 	 * before trying the page again.
1068 	 */
1069 	if (obj != NULL && vm_object_busied(obj)) {
1070 		if (locked)
1071 			VM_OBJECT_DROP(obj);
1072 		vm_object_busy_wait(obj, wmesg);
1073 		return (true);
1074 	}
1075 
1076 	if (!vm_page_busied(m))
1077 		return (false);
1078 
1079 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1080 	sleepq_lock(m);
1081 	x = vm_page_busy_fetch(m);
1082 	do {
1083 		/*
1084 		 * If the page changes objects or becomes unlocked we can
1085 		 * simply return.
1086 		 */
1087 		if (x == VPB_UNBUSIED ||
1088 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1089 		    m->object != obj || m->pindex != pindex) {
1090 			sleepq_release(m);
1091 			return (false);
1092 		}
1093 		if ((x & VPB_BIT_WAITERS) != 0)
1094 			break;
1095 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1096 	if (locked)
1097 		VM_OBJECT_DROP(obj);
1098 	DROP_GIANT();
1099 	sleepq_add(m, NULL, wmesg, 0, 0);
1100 	sleepq_wait(m, PVM);
1101 	PICKUP_GIANT();
1102 	return (true);
1103 }
1104 
1105 /*
1106  *	vm_page_trysbusy:
1107  *
1108  *	Try to shared busy a page.
1109  *	If the operation succeeds 1 is returned otherwise 0.
1110  *	The operation never sleeps.
1111  */
1112 int
1113 vm_page_trysbusy(vm_page_t m)
1114 {
1115 	vm_object_t obj;
1116 	u_int x;
1117 
1118 	obj = m->object;
1119 	x = vm_page_busy_fetch(m);
1120 	for (;;) {
1121 		if ((x & VPB_BIT_SHARED) == 0)
1122 			return (0);
1123 		/*
1124 		 * Reduce the window for transient busies that will trigger
1125 		 * false negatives in vm_page_ps_test().
1126 		 */
1127 		if (obj != NULL && vm_object_busied(obj))
1128 			return (0);
1129 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1130 		    x + VPB_ONE_SHARER))
1131 			break;
1132 	}
1133 
1134 	/* Refetch the object now that we're guaranteed that it is stable. */
1135 	obj = m->object;
1136 	if (obj != NULL && vm_object_busied(obj)) {
1137 		vm_page_sunbusy(m);
1138 		return (0);
1139 	}
1140 	return (1);
1141 }
1142 
1143 /*
1144  *	vm_page_tryxbusy:
1145  *
1146  *	Try to exclusive busy a page.
1147  *	If the operation succeeds 1 is returned otherwise 0.
1148  *	The operation never sleeps.
1149  */
1150 int
1151 vm_page_tryxbusy(vm_page_t m)
1152 {
1153 	vm_object_t obj;
1154 
1155         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1156             VPB_CURTHREAD_EXCLUSIVE) == 0)
1157 		return (0);
1158 
1159 	obj = m->object;
1160 	if (obj != NULL && vm_object_busied(obj)) {
1161 		vm_page_xunbusy(m);
1162 		return (0);
1163 	}
1164 	return (1);
1165 }
1166 
1167 static void
1168 vm_page_xunbusy_hard_tail(vm_page_t m)
1169 {
1170 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1171 	/* Wake the waiter. */
1172 	wakeup(m);
1173 }
1174 
1175 /*
1176  *	vm_page_xunbusy_hard:
1177  *
1178  *	Called when unbusy has failed because there is a waiter.
1179  */
1180 void
1181 vm_page_xunbusy_hard(vm_page_t m)
1182 {
1183 	vm_page_assert_xbusied(m);
1184 	vm_page_xunbusy_hard_tail(m);
1185 }
1186 
1187 void
1188 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1189 {
1190 	vm_page_assert_xbusied_unchecked(m);
1191 	vm_page_xunbusy_hard_tail(m);
1192 }
1193 
1194 static void
1195 vm_page_busy_free(vm_page_t m)
1196 {
1197 	u_int x;
1198 
1199 	atomic_thread_fence_rel();
1200 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1201 	if ((x & VPB_BIT_WAITERS) != 0)
1202 		wakeup(m);
1203 }
1204 
1205 /*
1206  *	vm_page_unhold_pages:
1207  *
1208  *	Unhold each of the pages that is referenced by the given array.
1209  */
1210 void
1211 vm_page_unhold_pages(vm_page_t *ma, int count)
1212 {
1213 
1214 	for (; count != 0; count--) {
1215 		vm_page_unwire(*ma, PQ_ACTIVE);
1216 		ma++;
1217 	}
1218 }
1219 
1220 vm_page_t
1221 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1222 {
1223 	vm_page_t m;
1224 
1225 #ifdef VM_PHYSSEG_SPARSE
1226 	m = vm_phys_paddr_to_vm_page(pa);
1227 	if (m == NULL)
1228 		m = vm_phys_fictitious_to_vm_page(pa);
1229 	return (m);
1230 #elif defined(VM_PHYSSEG_DENSE)
1231 	long pi;
1232 
1233 	pi = atop(pa);
1234 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1235 		m = &vm_page_array[pi - first_page];
1236 		return (m);
1237 	}
1238 	return (vm_phys_fictitious_to_vm_page(pa));
1239 #else
1240 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1241 #endif
1242 }
1243 
1244 /*
1245  *	vm_page_getfake:
1246  *
1247  *	Create a fictitious page with the specified physical address and
1248  *	memory attribute.  The memory attribute is the only the machine-
1249  *	dependent aspect of a fictitious page that must be initialized.
1250  */
1251 vm_page_t
1252 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1253 {
1254 	vm_page_t m;
1255 
1256 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1257 	vm_page_initfake(m, paddr, memattr);
1258 	return (m);
1259 }
1260 
1261 void
1262 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1263 {
1264 
1265 	if ((m->flags & PG_FICTITIOUS) != 0) {
1266 		/*
1267 		 * The page's memattr might have changed since the
1268 		 * previous initialization.  Update the pmap to the
1269 		 * new memattr.
1270 		 */
1271 		goto memattr;
1272 	}
1273 	m->phys_addr = paddr;
1274 	m->a.queue = PQ_NONE;
1275 	/* Fictitious pages don't use "segind". */
1276 	m->flags = PG_FICTITIOUS;
1277 	/* Fictitious pages don't use "order" or "pool". */
1278 	m->oflags = VPO_UNMANAGED;
1279 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1280 	/* Fictitious pages are unevictable. */
1281 	m->ref_count = 1;
1282 	pmap_page_init(m);
1283 memattr:
1284 	pmap_page_set_memattr(m, memattr);
1285 }
1286 
1287 /*
1288  *	vm_page_putfake:
1289  *
1290  *	Release a fictitious page.
1291  */
1292 void
1293 vm_page_putfake(vm_page_t m)
1294 {
1295 
1296 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1297 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1298 	    ("vm_page_putfake: bad page %p", m));
1299 	vm_page_assert_xbusied(m);
1300 	vm_page_busy_free(m);
1301 	uma_zfree(fakepg_zone, m);
1302 }
1303 
1304 /*
1305  *	vm_page_updatefake:
1306  *
1307  *	Update the given fictitious page to the specified physical address and
1308  *	memory attribute.
1309  */
1310 void
1311 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1312 {
1313 
1314 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1315 	    ("vm_page_updatefake: bad page %p", m));
1316 	m->phys_addr = paddr;
1317 	pmap_page_set_memattr(m, memattr);
1318 }
1319 
1320 /*
1321  *	vm_page_free:
1322  *
1323  *	Free a page.
1324  */
1325 void
1326 vm_page_free(vm_page_t m)
1327 {
1328 
1329 	m->flags &= ~PG_ZERO;
1330 	vm_page_free_toq(m);
1331 }
1332 
1333 /*
1334  *	vm_page_free_zero:
1335  *
1336  *	Free a page to the zerod-pages queue
1337  */
1338 void
1339 vm_page_free_zero(vm_page_t m)
1340 {
1341 
1342 	m->flags |= PG_ZERO;
1343 	vm_page_free_toq(m);
1344 }
1345 
1346 /*
1347  * Unbusy and handle the page queueing for a page from a getpages request that
1348  * was optionally read ahead or behind.
1349  */
1350 void
1351 vm_page_readahead_finish(vm_page_t m)
1352 {
1353 
1354 	/* We shouldn't put invalid pages on queues. */
1355 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1356 
1357 	/*
1358 	 * Since the page is not the actually needed one, whether it should
1359 	 * be activated or deactivated is not obvious.  Empirical results
1360 	 * have shown that deactivating the page is usually the best choice,
1361 	 * unless the page is wanted by another thread.
1362 	 */
1363 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1364 		vm_page_activate(m);
1365 	else
1366 		vm_page_deactivate(m);
1367 	vm_page_xunbusy_unchecked(m);
1368 }
1369 
1370 /*
1371  * Destroy the identity of an invalid page and free it if possible.
1372  * This is intended to be used when reading a page from backing store fails.
1373  */
1374 void
1375 vm_page_free_invalid(vm_page_t m)
1376 {
1377 
1378 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1379 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1380 	KASSERT(m->object != NULL, ("page %p has no object", m));
1381 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1382 
1383 	/*
1384 	 * We may be attempting to free the page as part of the handling for an
1385 	 * I/O error, in which case the page was xbusied by a different thread.
1386 	 */
1387 	vm_page_xbusy_claim(m);
1388 
1389 	/*
1390 	 * If someone has wired this page while the object lock
1391 	 * was not held, then the thread that unwires is responsible
1392 	 * for freeing the page.  Otherwise just free the page now.
1393 	 * The wire count of this unmapped page cannot change while
1394 	 * we have the page xbusy and the page's object wlocked.
1395 	 */
1396 	if (vm_page_remove(m))
1397 		vm_page_free(m);
1398 }
1399 
1400 /*
1401  *	vm_page_dirty_KBI:		[ internal use only ]
1402  *
1403  *	Set all bits in the page's dirty field.
1404  *
1405  *	The object containing the specified page must be locked if the
1406  *	call is made from the machine-independent layer.
1407  *
1408  *	See vm_page_clear_dirty_mask().
1409  *
1410  *	This function should only be called by vm_page_dirty().
1411  */
1412 void
1413 vm_page_dirty_KBI(vm_page_t m)
1414 {
1415 
1416 	/* Refer to this operation by its public name. */
1417 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1418 	m->dirty = VM_PAGE_BITS_ALL;
1419 }
1420 
1421 /*
1422  *	vm_page_insert:		[ internal use only ]
1423  *
1424  *	Inserts the given mem entry into the object and object list.
1425  *
1426  *	The object must be locked.
1427  */
1428 int
1429 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1430 {
1431 	vm_page_t mpred;
1432 
1433 	VM_OBJECT_ASSERT_WLOCKED(object);
1434 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1435 	return (vm_page_insert_after(m, object, pindex, mpred));
1436 }
1437 
1438 /*
1439  *	vm_page_insert_after:
1440  *
1441  *	Inserts the page "m" into the specified object at offset "pindex".
1442  *
1443  *	The page "mpred" must immediately precede the offset "pindex" within
1444  *	the specified object.
1445  *
1446  *	The object must be locked.
1447  */
1448 static int
1449 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1450     vm_page_t mpred)
1451 {
1452 	vm_page_t msucc;
1453 
1454 	VM_OBJECT_ASSERT_WLOCKED(object);
1455 	KASSERT(m->object == NULL,
1456 	    ("vm_page_insert_after: page already inserted"));
1457 	if (mpred != NULL) {
1458 		KASSERT(mpred->object == object,
1459 		    ("vm_page_insert_after: object doesn't contain mpred"));
1460 		KASSERT(mpred->pindex < pindex,
1461 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1462 		msucc = TAILQ_NEXT(mpred, listq);
1463 	} else
1464 		msucc = TAILQ_FIRST(&object->memq);
1465 	if (msucc != NULL)
1466 		KASSERT(msucc->pindex > pindex,
1467 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1468 
1469 	/*
1470 	 * Record the object/offset pair in this page.
1471 	 */
1472 	m->object = object;
1473 	m->pindex = pindex;
1474 	m->ref_count |= VPRC_OBJREF;
1475 
1476 	/*
1477 	 * Now link into the object's ordered list of backed pages.
1478 	 */
1479 	if (vm_radix_insert(&object->rtree, m)) {
1480 		m->object = NULL;
1481 		m->pindex = 0;
1482 		m->ref_count &= ~VPRC_OBJREF;
1483 		return (1);
1484 	}
1485 	vm_page_insert_radixdone(m, object, mpred);
1486 	return (0);
1487 }
1488 
1489 /*
1490  *	vm_page_insert_radixdone:
1491  *
1492  *	Complete page "m" insertion into the specified object after the
1493  *	radix trie hooking.
1494  *
1495  *	The page "mpred" must precede the offset "m->pindex" within the
1496  *	specified object.
1497  *
1498  *	The object must be locked.
1499  */
1500 static void
1501 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1502 {
1503 
1504 	VM_OBJECT_ASSERT_WLOCKED(object);
1505 	KASSERT(object != NULL && m->object == object,
1506 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1507 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1508 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1509 	if (mpred != NULL) {
1510 		KASSERT(mpred->object == object,
1511 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1512 		KASSERT(mpred->pindex < m->pindex,
1513 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1514 	}
1515 
1516 	if (mpred != NULL)
1517 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1518 	else
1519 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1520 
1521 	/*
1522 	 * Show that the object has one more resident page.
1523 	 */
1524 	object->resident_page_count++;
1525 
1526 	/*
1527 	 * Hold the vnode until the last page is released.
1528 	 */
1529 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1530 		vhold(object->handle);
1531 
1532 	/*
1533 	 * Since we are inserting a new and possibly dirty page,
1534 	 * update the object's generation count.
1535 	 */
1536 	if (pmap_page_is_write_mapped(m))
1537 		vm_object_set_writeable_dirty(object);
1538 }
1539 
1540 /*
1541  * Do the work to remove a page from its object.  The caller is responsible for
1542  * updating the page's fields to reflect this removal.
1543  */
1544 static void
1545 vm_page_object_remove(vm_page_t m)
1546 {
1547 	vm_object_t object;
1548 	vm_page_t mrem;
1549 
1550 	vm_page_assert_xbusied(m);
1551 	object = m->object;
1552 	VM_OBJECT_ASSERT_WLOCKED(object);
1553 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1554 	    ("page %p is missing its object ref", m));
1555 
1556 	/* Deferred free of swap space. */
1557 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1558 		vm_pager_page_unswapped(m);
1559 
1560 	m->object = NULL;
1561 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1562 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1563 
1564 	/*
1565 	 * Now remove from the object's list of backed pages.
1566 	 */
1567 	TAILQ_REMOVE(&object->memq, m, listq);
1568 
1569 	/*
1570 	 * And show that the object has one fewer resident page.
1571 	 */
1572 	object->resident_page_count--;
1573 
1574 	/*
1575 	 * The vnode may now be recycled.
1576 	 */
1577 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1578 		vdrop(object->handle);
1579 }
1580 
1581 /*
1582  *	vm_page_remove:
1583  *
1584  *	Removes the specified page from its containing object, but does not
1585  *	invalidate any backing storage.  Returns true if the object's reference
1586  *	was the last reference to the page, and false otherwise.
1587  *
1588  *	The object must be locked and the page must be exclusively busied.
1589  *	The exclusive busy will be released on return.  If this is not the
1590  *	final ref and the caller does not hold a wire reference it may not
1591  *	continue to access the page.
1592  */
1593 bool
1594 vm_page_remove(vm_page_t m)
1595 {
1596 	bool dropped;
1597 
1598 	dropped = vm_page_remove_xbusy(m);
1599 	vm_page_xunbusy(m);
1600 
1601 	return (dropped);
1602 }
1603 
1604 /*
1605  *	vm_page_remove_xbusy
1606  *
1607  *	Removes the page but leaves the xbusy held.  Returns true if this
1608  *	removed the final ref and false otherwise.
1609  */
1610 bool
1611 vm_page_remove_xbusy(vm_page_t m)
1612 {
1613 
1614 	vm_page_object_remove(m);
1615 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1616 }
1617 
1618 /*
1619  *	vm_page_lookup:
1620  *
1621  *	Returns the page associated with the object/offset
1622  *	pair specified; if none is found, NULL is returned.
1623  *
1624  *	The object must be locked.
1625  */
1626 vm_page_t
1627 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1628 {
1629 
1630 	VM_OBJECT_ASSERT_LOCKED(object);
1631 	return (vm_radix_lookup(&object->rtree, pindex));
1632 }
1633 
1634 /*
1635  *	vm_page_lookup_unlocked:
1636  *
1637  *	Returns the page associated with the object/offset pair specified;
1638  *	if none is found, NULL is returned.  The page may be no longer be
1639  *	present in the object at the time that this function returns.  Only
1640  *	useful for opportunistic checks such as inmem().
1641  */
1642 vm_page_t
1643 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1644 {
1645 
1646 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1647 }
1648 
1649 /*
1650  *	vm_page_relookup:
1651  *
1652  *	Returns a page that must already have been busied by
1653  *	the caller.  Used for bogus page replacement.
1654  */
1655 vm_page_t
1656 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1657 {
1658 	vm_page_t m;
1659 
1660 	m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1661 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1662 	    m->object == object && m->pindex == pindex,
1663 	    ("vm_page_relookup: Invalid page %p", m));
1664 	return (m);
1665 }
1666 
1667 /*
1668  * This should only be used by lockless functions for releasing transient
1669  * incorrect acquires.  The page may have been freed after we acquired a
1670  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1671  * further to do.
1672  */
1673 static void
1674 vm_page_busy_release(vm_page_t m)
1675 {
1676 	u_int x;
1677 
1678 	x = vm_page_busy_fetch(m);
1679 	for (;;) {
1680 		if (x == VPB_FREED)
1681 			break;
1682 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1683 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1684 			    x - VPB_ONE_SHARER))
1685 				break;
1686 			continue;
1687 		}
1688 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1689 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1690 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1691 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1692 			continue;
1693 		if ((x & VPB_BIT_WAITERS) != 0)
1694 			wakeup(m);
1695 		break;
1696 	}
1697 }
1698 
1699 /*
1700  *	vm_page_find_least:
1701  *
1702  *	Returns the page associated with the object with least pindex
1703  *	greater than or equal to the parameter pindex, or NULL.
1704  *
1705  *	The object must be locked.
1706  */
1707 vm_page_t
1708 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1709 {
1710 	vm_page_t m;
1711 
1712 	VM_OBJECT_ASSERT_LOCKED(object);
1713 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1714 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1715 	return (m);
1716 }
1717 
1718 /*
1719  * Returns the given page's successor (by pindex) within the object if it is
1720  * resident; if none is found, NULL is returned.
1721  *
1722  * The object must be locked.
1723  */
1724 vm_page_t
1725 vm_page_next(vm_page_t m)
1726 {
1727 	vm_page_t next;
1728 
1729 	VM_OBJECT_ASSERT_LOCKED(m->object);
1730 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1731 		MPASS(next->object == m->object);
1732 		if (next->pindex != m->pindex + 1)
1733 			next = NULL;
1734 	}
1735 	return (next);
1736 }
1737 
1738 /*
1739  * Returns the given page's predecessor (by pindex) within the object if it is
1740  * resident; if none is found, NULL is returned.
1741  *
1742  * The object must be locked.
1743  */
1744 vm_page_t
1745 vm_page_prev(vm_page_t m)
1746 {
1747 	vm_page_t prev;
1748 
1749 	VM_OBJECT_ASSERT_LOCKED(m->object);
1750 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1751 		MPASS(prev->object == m->object);
1752 		if (prev->pindex != m->pindex - 1)
1753 			prev = NULL;
1754 	}
1755 	return (prev);
1756 }
1757 
1758 /*
1759  * Uses the page mnew as a replacement for an existing page at index
1760  * pindex which must be already present in the object.
1761  *
1762  * Both pages must be exclusively busied on enter.  The old page is
1763  * unbusied on exit.
1764  *
1765  * A return value of true means mold is now free.  If this is not the
1766  * final ref and the caller does not hold a wire reference it may not
1767  * continue to access the page.
1768  */
1769 static bool
1770 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1771     vm_page_t mold)
1772 {
1773 	vm_page_t mret;
1774 	bool dropped;
1775 
1776 	VM_OBJECT_ASSERT_WLOCKED(object);
1777 	vm_page_assert_xbusied(mold);
1778 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1779 	    ("vm_page_replace: page %p already in object", mnew));
1780 
1781 	/*
1782 	 * This function mostly follows vm_page_insert() and
1783 	 * vm_page_remove() without the radix, object count and vnode
1784 	 * dance.  Double check such functions for more comments.
1785 	 */
1786 
1787 	mnew->object = object;
1788 	mnew->pindex = pindex;
1789 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1790 	mret = vm_radix_replace(&object->rtree, mnew);
1791 	KASSERT(mret == mold,
1792 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1793 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1794 	    (mnew->oflags & VPO_UNMANAGED),
1795 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1796 
1797 	/* Keep the resident page list in sorted order. */
1798 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1799 	TAILQ_REMOVE(&object->memq, mold, listq);
1800 	mold->object = NULL;
1801 
1802 	/*
1803 	 * The object's resident_page_count does not change because we have
1804 	 * swapped one page for another, but the generation count should
1805 	 * change if the page is dirty.
1806 	 */
1807 	if (pmap_page_is_write_mapped(mnew))
1808 		vm_object_set_writeable_dirty(object);
1809 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1810 	vm_page_xunbusy(mold);
1811 
1812 	return (dropped);
1813 }
1814 
1815 void
1816 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1817     vm_page_t mold)
1818 {
1819 
1820 	vm_page_assert_xbusied(mnew);
1821 
1822 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1823 		vm_page_free(mold);
1824 }
1825 
1826 /*
1827  *	vm_page_rename:
1828  *
1829  *	Move the given memory entry from its
1830  *	current object to the specified target object/offset.
1831  *
1832  *	Note: swap associated with the page must be invalidated by the move.  We
1833  *	      have to do this for several reasons:  (1) we aren't freeing the
1834  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1835  *	      moving the page from object A to B, and will then later move
1836  *	      the backing store from A to B and we can't have a conflict.
1837  *
1838  *	Note: we *always* dirty the page.  It is necessary both for the
1839  *	      fact that we moved it, and because we may be invalidating
1840  *	      swap.
1841  *
1842  *	The objects must be locked.
1843  */
1844 int
1845 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1846 {
1847 	vm_page_t mpred;
1848 	vm_pindex_t opidx;
1849 
1850 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1851 
1852 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1853 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1854 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1855 	    ("vm_page_rename: pindex already renamed"));
1856 
1857 	/*
1858 	 * Create a custom version of vm_page_insert() which does not depend
1859 	 * by m_prev and can cheat on the implementation aspects of the
1860 	 * function.
1861 	 */
1862 	opidx = m->pindex;
1863 	m->pindex = new_pindex;
1864 	if (vm_radix_insert(&new_object->rtree, m)) {
1865 		m->pindex = opidx;
1866 		return (1);
1867 	}
1868 
1869 	/*
1870 	 * The operation cannot fail anymore.  The removal must happen before
1871 	 * the listq iterator is tainted.
1872 	 */
1873 	m->pindex = opidx;
1874 	vm_page_object_remove(m);
1875 
1876 	/* Return back to the new pindex to complete vm_page_insert(). */
1877 	m->pindex = new_pindex;
1878 	m->object = new_object;
1879 
1880 	vm_page_insert_radixdone(m, new_object, mpred);
1881 	vm_page_dirty(m);
1882 	return (0);
1883 }
1884 
1885 /*
1886  *	vm_page_alloc:
1887  *
1888  *	Allocate and return a page that is associated with the specified
1889  *	object and offset pair.  By default, this page is exclusive busied.
1890  *
1891  *	The caller must always specify an allocation class.
1892  *
1893  *	allocation classes:
1894  *	VM_ALLOC_NORMAL		normal process request
1895  *	VM_ALLOC_SYSTEM		system *really* needs a page
1896  *	VM_ALLOC_INTERRUPT	interrupt time request
1897  *
1898  *	optional allocation flags:
1899  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1900  *				intends to allocate
1901  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1902  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1903  *	VM_ALLOC_SBUSY		shared busy the allocated page
1904  *	VM_ALLOC_WIRED		wire the allocated page
1905  *	VM_ALLOC_ZERO		prefer a zeroed page
1906  */
1907 vm_page_t
1908 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1909 {
1910 
1911 	return (vm_page_alloc_after(object, pindex, req,
1912 	    vm_radix_lookup_le(&object->rtree, pindex)));
1913 }
1914 
1915 vm_page_t
1916 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1917     int req)
1918 {
1919 
1920 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1921 	    vm_radix_lookup_le(&object->rtree, pindex)));
1922 }
1923 
1924 /*
1925  * Allocate a page in the specified object with the given page index.  To
1926  * optimize insertion of the page into the object, the caller must also specifiy
1927  * the resident page in the object with largest index smaller than the given
1928  * page index, or NULL if no such page exists.
1929  */
1930 vm_page_t
1931 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1932     int req, vm_page_t mpred)
1933 {
1934 	struct vm_domainset_iter di;
1935 	vm_page_t m;
1936 	int domain;
1937 
1938 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1939 	do {
1940 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1941 		    mpred);
1942 		if (m != NULL)
1943 			break;
1944 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1945 
1946 	return (m);
1947 }
1948 
1949 /*
1950  * Returns true if the number of free pages exceeds the minimum
1951  * for the request class and false otherwise.
1952  */
1953 static int
1954 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1955 {
1956 	u_int limit, old, new;
1957 
1958 	if (req_class == VM_ALLOC_INTERRUPT)
1959 		limit = 0;
1960 	else if (req_class == VM_ALLOC_SYSTEM)
1961 		limit = vmd->vmd_interrupt_free_min;
1962 	else
1963 		limit = vmd->vmd_free_reserved;
1964 
1965 	/*
1966 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1967 	 */
1968 	limit += npages;
1969 	old = vmd->vmd_free_count;
1970 	do {
1971 		if (old < limit)
1972 			return (0);
1973 		new = old - npages;
1974 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1975 
1976 	/* Wake the page daemon if we've crossed the threshold. */
1977 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1978 		pagedaemon_wakeup(vmd->vmd_domain);
1979 
1980 	/* Only update bitsets on transitions. */
1981 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1982 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1983 		vm_domain_set(vmd);
1984 
1985 	return (1);
1986 }
1987 
1988 int
1989 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1990 {
1991 	int req_class;
1992 
1993 	/*
1994 	 * The page daemon is allowed to dig deeper into the free page list.
1995 	 */
1996 	req_class = req & VM_ALLOC_CLASS_MASK;
1997 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1998 		req_class = VM_ALLOC_SYSTEM;
1999 	return (_vm_domain_allocate(vmd, req_class, npages));
2000 }
2001 
2002 vm_page_t
2003 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2004     int req, vm_page_t mpred)
2005 {
2006 	struct vm_domain *vmd;
2007 	vm_page_t m;
2008 	int flags;
2009 
2010 #define	VPA_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |	\
2011 			 VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY |		\
2012 			 VM_ALLOC_SBUSY | VM_ALLOC_WIRED |		\
2013 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
2014 	KASSERT((req & ~VPA_FLAGS) == 0,
2015 	    ("invalid request %#x", req));
2016 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2017 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2018 	    ("invalid request %#x", req));
2019 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2020 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2021 	    (uintmax_t)pindex));
2022 	VM_OBJECT_ASSERT_WLOCKED(object);
2023 
2024 	flags = 0;
2025 	m = NULL;
2026 again:
2027 #if VM_NRESERVLEVEL > 0
2028 	/*
2029 	 * Can we allocate the page from a reservation?
2030 	 */
2031 	if (vm_object_reserv(object) &&
2032 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2033 	    NULL) {
2034 		goto found;
2035 	}
2036 #endif
2037 	vmd = VM_DOMAIN(domain);
2038 	if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
2039 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
2040 		    M_NOWAIT | M_NOVM);
2041 		if (m != NULL) {
2042 			flags |= PG_PCPU_CACHE;
2043 			goto found;
2044 		}
2045 	}
2046 	if (vm_domain_allocate(vmd, req, 1)) {
2047 		/*
2048 		 * If not, allocate it from the free page queues.
2049 		 */
2050 		vm_domain_free_lock(vmd);
2051 		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT, 0);
2052 		vm_domain_free_unlock(vmd);
2053 		if (m == NULL) {
2054 			vm_domain_freecnt_inc(vmd, 1);
2055 #if VM_NRESERVLEVEL > 0
2056 			if (vm_reserv_reclaim_inactive(domain))
2057 				goto again;
2058 #endif
2059 		}
2060 	}
2061 	if (m == NULL) {
2062 		/*
2063 		 * Not allocatable, give up.
2064 		 */
2065 		if (vm_domain_alloc_fail(vmd, object, req))
2066 			goto again;
2067 		return (NULL);
2068 	}
2069 
2070 	/*
2071 	 * At this point we had better have found a good page.
2072 	 */
2073 found:
2074 	vm_page_dequeue(m);
2075 	vm_page_alloc_check(m);
2076 
2077 	/*
2078 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2079 	 */
2080 	flags |= m->flags & PG_ZERO;
2081 	if ((req & VM_ALLOC_NODUMP) != 0)
2082 		flags |= PG_NODUMP;
2083 	m->flags = flags;
2084 	m->a.flags = 0;
2085 	m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2086 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2087 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2088 	else if ((req & VM_ALLOC_SBUSY) != 0)
2089 		m->busy_lock = VPB_SHARERS_WORD(1);
2090 	else
2091 		m->busy_lock = VPB_UNBUSIED;
2092 	if (req & VM_ALLOC_WIRED) {
2093 		vm_wire_add(1);
2094 		m->ref_count = 1;
2095 	}
2096 	m->a.act_count = 0;
2097 
2098 	if (vm_page_insert_after(m, object, pindex, mpred)) {
2099 		if (req & VM_ALLOC_WIRED) {
2100 			vm_wire_sub(1);
2101 			m->ref_count = 0;
2102 		}
2103 		KASSERT(m->object == NULL, ("page %p has object", m));
2104 		m->oflags = VPO_UNMANAGED;
2105 		m->busy_lock = VPB_UNBUSIED;
2106 		/* Don't change PG_ZERO. */
2107 		vm_page_free_toq(m);
2108 		if (req & VM_ALLOC_WAITFAIL) {
2109 			VM_OBJECT_WUNLOCK(object);
2110 			vm_radix_wait();
2111 			VM_OBJECT_WLOCK(object);
2112 		}
2113 		return (NULL);
2114 	}
2115 
2116 	/* Ignore device objects; the pager sets "memattr" for them. */
2117 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2118 	    (object->flags & OBJ_FICTITIOUS) == 0)
2119 		pmap_page_set_memattr(m, object->memattr);
2120 
2121 	return (m);
2122 }
2123 
2124 /*
2125  *	vm_page_alloc_contig:
2126  *
2127  *	Allocate a contiguous set of physical pages of the given size "npages"
2128  *	from the free lists.  All of the physical pages must be at or above
2129  *	the given physical address "low" and below the given physical address
2130  *	"high".  The given value "alignment" determines the alignment of the
2131  *	first physical page in the set.  If the given value "boundary" is
2132  *	non-zero, then the set of physical pages cannot cross any physical
2133  *	address boundary that is a multiple of that value.  Both "alignment"
2134  *	and "boundary" must be a power of two.
2135  *
2136  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2137  *	then the memory attribute setting for the physical pages is configured
2138  *	to the object's memory attribute setting.  Otherwise, the memory
2139  *	attribute setting for the physical pages is configured to "memattr",
2140  *	overriding the object's memory attribute setting.  However, if the
2141  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2142  *	memory attribute setting for the physical pages cannot be configured
2143  *	to VM_MEMATTR_DEFAULT.
2144  *
2145  *	The specified object may not contain fictitious pages.
2146  *
2147  *	The caller must always specify an allocation class.
2148  *
2149  *	allocation classes:
2150  *	VM_ALLOC_NORMAL		normal process request
2151  *	VM_ALLOC_SYSTEM		system *really* needs a page
2152  *	VM_ALLOC_INTERRUPT	interrupt time request
2153  *
2154  *	optional allocation flags:
2155  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2156  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2157  *	VM_ALLOC_SBUSY		shared busy the allocated page
2158  *	VM_ALLOC_WIRED		wire the allocated page
2159  *	VM_ALLOC_ZERO		prefer a zeroed page
2160  */
2161 vm_page_t
2162 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2163     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2164     vm_paddr_t boundary, vm_memattr_t memattr)
2165 {
2166 	struct vm_domainset_iter di;
2167 	vm_page_t m;
2168 	int domain;
2169 
2170 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2171 	do {
2172 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2173 		    npages, low, high, alignment, boundary, memattr);
2174 		if (m != NULL)
2175 			break;
2176 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2177 
2178 	return (m);
2179 }
2180 
2181 static vm_page_t
2182 vm_page_find_contig_domain(int domain, int req, u_long npages, vm_paddr_t low,
2183     vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2184 {
2185 	struct vm_domain *vmd;
2186 	vm_page_t m_ret;
2187 
2188 	/*
2189 	 * Can we allocate the pages without the number of free pages falling
2190 	 * below the lower bound for the allocation class?
2191 	 */
2192 	vmd = VM_DOMAIN(domain);
2193 	if (!vm_domain_allocate(vmd, req, npages))
2194 		return (NULL);
2195 	/*
2196 	 * Try to allocate the pages from the free page queues.
2197 	 */
2198 	vm_domain_free_lock(vmd);
2199 	m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2200 	    alignment, boundary);
2201 	vm_domain_free_unlock(vmd);
2202 	if (m_ret != NULL)
2203 		return (m_ret);
2204 #if VM_NRESERVLEVEL > 0
2205 	/*
2206 	 * Try to break a reservation to allocate the pages.
2207 	 */
2208 	if ((req & VM_ALLOC_NORECLAIM) == 0) {
2209 		m_ret = vm_reserv_reclaim_contig(domain, npages, low,
2210 	            high, alignment, boundary);
2211 		if (m_ret != NULL)
2212 			return (m_ret);
2213 	}
2214 #endif
2215 	vm_domain_freecnt_inc(vmd, npages);
2216 	return (NULL);
2217 }
2218 
2219 vm_page_t
2220 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2221     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2222     vm_paddr_t boundary, vm_memattr_t memattr)
2223 {
2224 	vm_page_t m, m_ret, mpred;
2225 	u_int busy_lock, flags, oflags;
2226 
2227 #define	VPAC_FLAGS	(VPA_FLAGS | VM_ALLOC_NORECLAIM)
2228 	KASSERT((req & ~VPAC_FLAGS) == 0,
2229 	    ("invalid request %#x", req));
2230 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2231 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2232 	    ("invalid request %#x", req));
2233 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2234 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2235 	    ("invalid request %#x", req));
2236 	VM_OBJECT_ASSERT_WLOCKED(object);
2237 	KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2238 	    ("vm_page_alloc_contig: object %p has fictitious pages",
2239 	    object));
2240 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2241 
2242 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
2243 	KASSERT(mpred == NULL || mpred->pindex != pindex,
2244 	    ("vm_page_alloc_contig: pindex already allocated"));
2245 	for (;;) {
2246 #if VM_NRESERVLEVEL > 0
2247 		/*
2248 		 * Can we allocate the pages from a reservation?
2249 		 */
2250 		if (vm_object_reserv(object) &&
2251 		    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2252 		    mpred, npages, low, high, alignment, boundary)) != NULL) {
2253 			break;
2254 		}
2255 #endif
2256 		if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2257 		    low, high, alignment, boundary)) != NULL)
2258 			break;
2259 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
2260 			return (NULL);
2261 	}
2262 	for (m = m_ret; m < &m_ret[npages]; m++) {
2263 		vm_page_dequeue(m);
2264 		vm_page_alloc_check(m);
2265 	}
2266 
2267 	/*
2268 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2269 	 */
2270 	flags = PG_ZERO;
2271 	if ((req & VM_ALLOC_NODUMP) != 0)
2272 		flags |= PG_NODUMP;
2273 	oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
2274 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
2275 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2276 	else if ((req & VM_ALLOC_SBUSY) != 0)
2277 		busy_lock = VPB_SHARERS_WORD(1);
2278 	else
2279 		busy_lock = VPB_UNBUSIED;
2280 	if ((req & VM_ALLOC_WIRED) != 0)
2281 		vm_wire_add(npages);
2282 	if (object->memattr != VM_MEMATTR_DEFAULT &&
2283 	    memattr == VM_MEMATTR_DEFAULT)
2284 		memattr = object->memattr;
2285 	for (m = m_ret; m < &m_ret[npages]; m++) {
2286 		m->a.flags = 0;
2287 		m->flags = (m->flags | PG_NODUMP) & flags;
2288 		m->busy_lock = busy_lock;
2289 		if ((req & VM_ALLOC_WIRED) != 0)
2290 			m->ref_count = 1;
2291 		m->a.act_count = 0;
2292 		m->oflags = oflags;
2293 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2294 			if ((req & VM_ALLOC_WIRED) != 0)
2295 				vm_wire_sub(npages);
2296 			KASSERT(m->object == NULL,
2297 			    ("page %p has object", m));
2298 			mpred = m;
2299 			for (m = m_ret; m < &m_ret[npages]; m++) {
2300 				if (m <= mpred &&
2301 				    (req & VM_ALLOC_WIRED) != 0)
2302 					m->ref_count = 0;
2303 				m->oflags = VPO_UNMANAGED;
2304 				m->busy_lock = VPB_UNBUSIED;
2305 				/* Don't change PG_ZERO. */
2306 				vm_page_free_toq(m);
2307 			}
2308 			if (req & VM_ALLOC_WAITFAIL) {
2309 				VM_OBJECT_WUNLOCK(object);
2310 				vm_radix_wait();
2311 				VM_OBJECT_WLOCK(object);
2312 			}
2313 			return (NULL);
2314 		}
2315 		mpred = m;
2316 		if (memattr != VM_MEMATTR_DEFAULT)
2317 			pmap_page_set_memattr(m, memattr);
2318 		pindex++;
2319 	}
2320 	return (m_ret);
2321 }
2322 
2323 /*
2324  * Allocate a physical page that is not intended to be inserted into a VM
2325  * object.  If the "freelist" parameter is not equal to VM_NFREELIST, then only
2326  * pages from the specified vm_phys freelist will be returned.
2327  */
2328 static __always_inline vm_page_t
2329 _vm_page_alloc_noobj_domain(int domain, const int freelist, int req)
2330 {
2331 	struct vm_domain *vmd;
2332 	vm_page_t m;
2333 	int flags;
2334 
2335 #define	VPAN_FLAGS	(VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL |      \
2336 			 VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK |		\
2337 			 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED |		\
2338 			 VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
2339 	KASSERT((req & ~VPAN_FLAGS) == 0,
2340 	    ("invalid request %#x", req));
2341 
2342 	flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0;
2343 	vmd = VM_DOMAIN(domain);
2344 again:
2345 	if (freelist == VM_NFREELIST &&
2346 	    vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
2347 		m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
2348 		    M_NOWAIT | M_NOVM);
2349 		if (m != NULL) {
2350 			flags |= PG_PCPU_CACHE;
2351 			goto found;
2352 		}
2353 	}
2354 
2355 	if (vm_domain_allocate(vmd, req, 1)) {
2356 		vm_domain_free_lock(vmd);
2357 		if (freelist == VM_NFREELIST)
2358 			m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
2359 		else
2360 			m = vm_phys_alloc_freelist_pages(domain, freelist,
2361 			    VM_FREEPOOL_DIRECT, 0);
2362 		vm_domain_free_unlock(vmd);
2363 		if (m == NULL) {
2364 			vm_domain_freecnt_inc(vmd, 1);
2365 #if VM_NRESERVLEVEL > 0
2366 			if (freelist == VM_NFREELIST &&
2367 			    vm_reserv_reclaim_inactive(domain))
2368 				goto again;
2369 #endif
2370 		}
2371 	}
2372 	if (m == NULL) {
2373 		if (vm_domain_alloc_fail(vmd, NULL, req))
2374 			goto again;
2375 		return (NULL);
2376 	}
2377 
2378 found:
2379 	vm_page_dequeue(m);
2380 	vm_page_alloc_check(m);
2381 
2382 	/*
2383 	 * Consumers should not rely on a useful default pindex value.
2384 	 */
2385 	m->pindex = 0xdeadc0dedeadc0de;
2386 	m->flags = (m->flags & PG_ZERO) | flags;
2387 	m->a.flags = 0;
2388 	m->oflags = VPO_UNMANAGED;
2389 	m->busy_lock = VPB_UNBUSIED;
2390 	if ((req & VM_ALLOC_WIRED) != 0) {
2391 		vm_wire_add(1);
2392 		m->ref_count = 1;
2393 	}
2394 
2395 	if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2396 		pmap_zero_page(m);
2397 
2398 	return (m);
2399 }
2400 
2401 vm_page_t
2402 vm_page_alloc_freelist(int freelist, int req)
2403 {
2404 	struct vm_domainset_iter di;
2405 	vm_page_t m;
2406 	int domain;
2407 
2408 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2409 	do {
2410 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2411 		if (m != NULL)
2412 			break;
2413 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2414 
2415 	return (m);
2416 }
2417 
2418 vm_page_t
2419 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2420 {
2421 	KASSERT(freelist >= 0 && freelist < VM_NFREELIST,
2422 	    ("%s: invalid freelist %d", __func__, freelist));
2423 
2424 	return (_vm_page_alloc_noobj_domain(domain, freelist, req));
2425 }
2426 
2427 vm_page_t
2428 vm_page_alloc_noobj(int req)
2429 {
2430 	struct vm_domainset_iter di;
2431 	vm_page_t m;
2432 	int domain;
2433 
2434 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2435 	do {
2436 		m = vm_page_alloc_noobj_domain(domain, req);
2437 		if (m != NULL)
2438 			break;
2439 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2440 
2441 	return (m);
2442 }
2443 
2444 vm_page_t
2445 vm_page_alloc_noobj_domain(int domain, int req)
2446 {
2447 	return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req));
2448 }
2449 
2450 vm_page_t
2451 vm_page_alloc_noobj_contig(int req, u_long npages, vm_paddr_t low,
2452     vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2453     vm_memattr_t memattr)
2454 {
2455 	struct vm_domainset_iter di;
2456 	vm_page_t m;
2457 	int domain;
2458 
2459 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2460 	do {
2461 		m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
2462 		    high, alignment, boundary, memattr);
2463 		if (m != NULL)
2464 			break;
2465 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2466 
2467 	return (m);
2468 }
2469 
2470 vm_page_t
2471 vm_page_alloc_noobj_contig_domain(int domain, int req, u_long npages,
2472     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
2473     vm_memattr_t memattr)
2474 {
2475 	vm_page_t m, m_ret;
2476 	u_int flags;
2477 
2478 #define	VPANC_FLAGS	(VPAN_FLAGS | VM_ALLOC_NORECLAIM)
2479 	KASSERT((req & ~VPANC_FLAGS) == 0,
2480 	    ("invalid request %#x", req));
2481 	KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
2482 	    (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
2483 	    ("invalid request %#x", req));
2484 	KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2485 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2486 	    ("invalid request %#x", req));
2487 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2488 
2489 	while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
2490 	    low, high, alignment, boundary)) == NULL) {
2491 		if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
2492 			return (NULL);
2493 	}
2494 
2495 	/*
2496 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2497 	 */
2498 	flags = PG_ZERO;
2499 	if ((req & VM_ALLOC_NODUMP) != 0)
2500 		flags |= PG_NODUMP;
2501 	if ((req & VM_ALLOC_WIRED) != 0)
2502 		vm_wire_add(npages);
2503 	for (m = m_ret; m < &m_ret[npages]; m++) {
2504 		vm_page_dequeue(m);
2505 		vm_page_alloc_check(m);
2506 
2507 		/*
2508 		 * Consumers should not rely on a useful default pindex value.
2509 		 */
2510 		m->pindex = 0xdeadc0dedeadc0de;
2511 		m->a.flags = 0;
2512 		m->flags = (m->flags | PG_NODUMP) & flags;
2513 		m->busy_lock = VPB_UNBUSIED;
2514 		if ((req & VM_ALLOC_WIRED) != 0)
2515 			m->ref_count = 1;
2516 		m->a.act_count = 0;
2517 		m->oflags = VPO_UNMANAGED;
2518 
2519 		/*
2520 		 * Zero the page before updating any mappings since the page is
2521 		 * not yet shared with any devices which might require the
2522 		 * non-default memory attribute.  pmap_page_set_memattr()
2523 		 * flushes data caches before returning.
2524 		 */
2525 		if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
2526 			pmap_zero_page(m);
2527 		if (memattr != VM_MEMATTR_DEFAULT)
2528 			pmap_page_set_memattr(m, memattr);
2529 	}
2530 	return (m_ret);
2531 }
2532 
2533 /*
2534  * Check a page that has been freshly dequeued from a freelist.
2535  */
2536 static void
2537 vm_page_alloc_check(vm_page_t m)
2538 {
2539 
2540 	KASSERT(m->object == NULL, ("page %p has object", m));
2541 	KASSERT(m->a.queue == PQ_NONE &&
2542 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2543 	    ("page %p has unexpected queue %d, flags %#x",
2544 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2545 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2546 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2547 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2548 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2549 	    ("page %p has unexpected memattr %d",
2550 	    m, pmap_page_get_memattr(m)));
2551 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2552 	pmap_vm_page_alloc_check(m);
2553 }
2554 
2555 static int
2556 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2557 {
2558 	struct vm_domain *vmd;
2559 	struct vm_pgcache *pgcache;
2560 	int i;
2561 
2562 	pgcache = arg;
2563 	vmd = VM_DOMAIN(pgcache->domain);
2564 
2565 	/*
2566 	 * The page daemon should avoid creating extra memory pressure since its
2567 	 * main purpose is to replenish the store of free pages.
2568 	 */
2569 	if (vmd->vmd_severeset || curproc == pageproc ||
2570 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2571 		return (0);
2572 	domain = vmd->vmd_domain;
2573 	vm_domain_free_lock(vmd);
2574 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2575 	    (vm_page_t *)store);
2576 	vm_domain_free_unlock(vmd);
2577 	if (cnt != i)
2578 		vm_domain_freecnt_inc(vmd, cnt - i);
2579 
2580 	return (i);
2581 }
2582 
2583 static void
2584 vm_page_zone_release(void *arg, void **store, int cnt)
2585 {
2586 	struct vm_domain *vmd;
2587 	struct vm_pgcache *pgcache;
2588 	vm_page_t m;
2589 	int i;
2590 
2591 	pgcache = arg;
2592 	vmd = VM_DOMAIN(pgcache->domain);
2593 	vm_domain_free_lock(vmd);
2594 	for (i = 0; i < cnt; i++) {
2595 		m = (vm_page_t)store[i];
2596 		vm_phys_free_pages(m, 0);
2597 	}
2598 	vm_domain_free_unlock(vmd);
2599 	vm_domain_freecnt_inc(vmd, cnt);
2600 }
2601 
2602 #define	VPSC_ANY	0	/* No restrictions. */
2603 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2604 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2605 
2606 /*
2607  *	vm_page_scan_contig:
2608  *
2609  *	Scan vm_page_array[] between the specified entries "m_start" and
2610  *	"m_end" for a run of contiguous physical pages that satisfy the
2611  *	specified conditions, and return the lowest page in the run.  The
2612  *	specified "alignment" determines the alignment of the lowest physical
2613  *	page in the run.  If the specified "boundary" is non-zero, then the
2614  *	run of physical pages cannot span a physical address that is a
2615  *	multiple of "boundary".
2616  *
2617  *	"m_end" is never dereferenced, so it need not point to a vm_page
2618  *	structure within vm_page_array[].
2619  *
2620  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2621  *	span a hole (or discontiguity) in the physical address space.  Both
2622  *	"alignment" and "boundary" must be a power of two.
2623  */
2624 vm_page_t
2625 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2626     u_long alignment, vm_paddr_t boundary, int options)
2627 {
2628 	vm_object_t object;
2629 	vm_paddr_t pa;
2630 	vm_page_t m, m_run;
2631 #if VM_NRESERVLEVEL > 0
2632 	int level;
2633 #endif
2634 	int m_inc, order, run_ext, run_len;
2635 
2636 	KASSERT(npages > 0, ("npages is 0"));
2637 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2638 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2639 	m_run = NULL;
2640 	run_len = 0;
2641 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2642 		KASSERT((m->flags & PG_MARKER) == 0,
2643 		    ("page %p is PG_MARKER", m));
2644 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2645 		    ("fictitious page %p has invalid ref count", m));
2646 
2647 		/*
2648 		 * If the current page would be the start of a run, check its
2649 		 * physical address against the end, alignment, and boundary
2650 		 * conditions.  If it doesn't satisfy these conditions, either
2651 		 * terminate the scan or advance to the next page that
2652 		 * satisfies the failed condition.
2653 		 */
2654 		if (run_len == 0) {
2655 			KASSERT(m_run == NULL, ("m_run != NULL"));
2656 			if (m + npages > m_end)
2657 				break;
2658 			pa = VM_PAGE_TO_PHYS(m);
2659 			if (!vm_addr_align_ok(pa, alignment)) {
2660 				m_inc = atop(roundup2(pa, alignment) - pa);
2661 				continue;
2662 			}
2663 			if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
2664 				m_inc = atop(roundup2(pa, boundary) - pa);
2665 				continue;
2666 			}
2667 		} else
2668 			KASSERT(m_run != NULL, ("m_run == NULL"));
2669 
2670 retry:
2671 		m_inc = 1;
2672 		if (vm_page_wired(m))
2673 			run_ext = 0;
2674 #if VM_NRESERVLEVEL > 0
2675 		else if ((level = vm_reserv_level(m)) >= 0 &&
2676 		    (options & VPSC_NORESERV) != 0) {
2677 			run_ext = 0;
2678 			/* Advance to the end of the reservation. */
2679 			pa = VM_PAGE_TO_PHYS(m);
2680 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2681 			    pa);
2682 		}
2683 #endif
2684 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2685 			/*
2686 			 * The page is considered eligible for relocation if
2687 			 * and only if it could be laundered or reclaimed by
2688 			 * the page daemon.
2689 			 */
2690 			VM_OBJECT_RLOCK(object);
2691 			if (object != m->object) {
2692 				VM_OBJECT_RUNLOCK(object);
2693 				goto retry;
2694 			}
2695 			/* Don't care: PG_NODUMP, PG_ZERO. */
2696 			if (object->type != OBJT_DEFAULT &&
2697 			    (object->flags & OBJ_SWAP) == 0 &&
2698 			    object->type != OBJT_VNODE) {
2699 				run_ext = 0;
2700 #if VM_NRESERVLEVEL > 0
2701 			} else if ((options & VPSC_NOSUPER) != 0 &&
2702 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2703 				run_ext = 0;
2704 				/* Advance to the end of the superpage. */
2705 				pa = VM_PAGE_TO_PHYS(m);
2706 				m_inc = atop(roundup2(pa + 1,
2707 				    vm_reserv_size(level)) - pa);
2708 #endif
2709 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2710 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2711 				/*
2712 				 * The page is allocated but eligible for
2713 				 * relocation.  Extend the current run by one
2714 				 * page.
2715 				 */
2716 				KASSERT(pmap_page_get_memattr(m) ==
2717 				    VM_MEMATTR_DEFAULT,
2718 				    ("page %p has an unexpected memattr", m));
2719 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2720 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2721 				    ("page %p has unexpected oflags", m));
2722 				/* Don't care: PGA_NOSYNC. */
2723 				run_ext = 1;
2724 			} else
2725 				run_ext = 0;
2726 			VM_OBJECT_RUNLOCK(object);
2727 #if VM_NRESERVLEVEL > 0
2728 		} else if (level >= 0) {
2729 			/*
2730 			 * The page is reserved but not yet allocated.  In
2731 			 * other words, it is still free.  Extend the current
2732 			 * run by one page.
2733 			 */
2734 			run_ext = 1;
2735 #endif
2736 		} else if ((order = m->order) < VM_NFREEORDER) {
2737 			/*
2738 			 * The page is enqueued in the physical memory
2739 			 * allocator's free page queues.  Moreover, it is the
2740 			 * first page in a power-of-two-sized run of
2741 			 * contiguous free pages.  Add these pages to the end
2742 			 * of the current run, and jump ahead.
2743 			 */
2744 			run_ext = 1 << order;
2745 			m_inc = 1 << order;
2746 		} else {
2747 			/*
2748 			 * Skip the page for one of the following reasons: (1)
2749 			 * It is enqueued in the physical memory allocator's
2750 			 * free page queues.  However, it is not the first
2751 			 * page in a run of contiguous free pages.  (This case
2752 			 * rarely occurs because the scan is performed in
2753 			 * ascending order.) (2) It is not reserved, and it is
2754 			 * transitioning from free to allocated.  (Conversely,
2755 			 * the transition from allocated to free for managed
2756 			 * pages is blocked by the page busy lock.) (3) It is
2757 			 * allocated but not contained by an object and not
2758 			 * wired, e.g., allocated by Xen's balloon driver.
2759 			 */
2760 			run_ext = 0;
2761 		}
2762 
2763 		/*
2764 		 * Extend or reset the current run of pages.
2765 		 */
2766 		if (run_ext > 0) {
2767 			if (run_len == 0)
2768 				m_run = m;
2769 			run_len += run_ext;
2770 		} else {
2771 			if (run_len > 0) {
2772 				m_run = NULL;
2773 				run_len = 0;
2774 			}
2775 		}
2776 	}
2777 	if (run_len >= npages)
2778 		return (m_run);
2779 	return (NULL);
2780 }
2781 
2782 /*
2783  *	vm_page_reclaim_run:
2784  *
2785  *	Try to relocate each of the allocated virtual pages within the
2786  *	specified run of physical pages to a new physical address.  Free the
2787  *	physical pages underlying the relocated virtual pages.  A virtual page
2788  *	is relocatable if and only if it could be laundered or reclaimed by
2789  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2790  *	physical address above "high".
2791  *
2792  *	Returns 0 if every physical page within the run was already free or
2793  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2794  *	value indicating why the last attempt to relocate a virtual page was
2795  *	unsuccessful.
2796  *
2797  *	"req_class" must be an allocation class.
2798  */
2799 static int
2800 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2801     vm_paddr_t high)
2802 {
2803 	struct vm_domain *vmd;
2804 	struct spglist free;
2805 	vm_object_t object;
2806 	vm_paddr_t pa;
2807 	vm_page_t m, m_end, m_new;
2808 	int error, order, req;
2809 
2810 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2811 	    ("req_class is not an allocation class"));
2812 	SLIST_INIT(&free);
2813 	error = 0;
2814 	m = m_run;
2815 	m_end = m_run + npages;
2816 	for (; error == 0 && m < m_end; m++) {
2817 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2818 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2819 
2820 		/*
2821 		 * Racily check for wirings.  Races are handled once the object
2822 		 * lock is held and the page is unmapped.
2823 		 */
2824 		if (vm_page_wired(m))
2825 			error = EBUSY;
2826 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2827 			/*
2828 			 * The page is relocated if and only if it could be
2829 			 * laundered or reclaimed by the page daemon.
2830 			 */
2831 			VM_OBJECT_WLOCK(object);
2832 			/* Don't care: PG_NODUMP, PG_ZERO. */
2833 			if (m->object != object ||
2834 			    (object->type != OBJT_DEFAULT &&
2835 			    (object->flags & OBJ_SWAP) == 0 &&
2836 			    object->type != OBJT_VNODE))
2837 				error = EINVAL;
2838 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2839 				error = EINVAL;
2840 			else if (vm_page_queue(m) != PQ_NONE &&
2841 			    vm_page_tryxbusy(m) != 0) {
2842 				if (vm_page_wired(m)) {
2843 					vm_page_xunbusy(m);
2844 					error = EBUSY;
2845 					goto unlock;
2846 				}
2847 				KASSERT(pmap_page_get_memattr(m) ==
2848 				    VM_MEMATTR_DEFAULT,
2849 				    ("page %p has an unexpected memattr", m));
2850 				KASSERT(m->oflags == 0,
2851 				    ("page %p has unexpected oflags", m));
2852 				/* Don't care: PGA_NOSYNC. */
2853 				if (!vm_page_none_valid(m)) {
2854 					/*
2855 					 * First, try to allocate a new page
2856 					 * that is above "high".  Failing
2857 					 * that, try to allocate a new page
2858 					 * that is below "m_run".  Allocate
2859 					 * the new page between the end of
2860 					 * "m_run" and "high" only as a last
2861 					 * resort.
2862 					 */
2863 					req = req_class;
2864 					if ((m->flags & PG_NODUMP) != 0)
2865 						req |= VM_ALLOC_NODUMP;
2866 					if (trunc_page(high) !=
2867 					    ~(vm_paddr_t)PAGE_MASK) {
2868 						m_new =
2869 						    vm_page_alloc_noobj_contig(
2870 						    req, 1, round_page(high),
2871 						    ~(vm_paddr_t)0, PAGE_SIZE,
2872 						    0, VM_MEMATTR_DEFAULT);
2873 					} else
2874 						m_new = NULL;
2875 					if (m_new == NULL) {
2876 						pa = VM_PAGE_TO_PHYS(m_run);
2877 						m_new =
2878 						    vm_page_alloc_noobj_contig(
2879 						    req, 1, 0, pa - 1,
2880 						    PAGE_SIZE, 0,
2881 						    VM_MEMATTR_DEFAULT);
2882 					}
2883 					if (m_new == NULL) {
2884 						pa += ptoa(npages);
2885 						m_new =
2886 						    vm_page_alloc_noobj_contig(
2887 						    req, 1, pa, high, PAGE_SIZE,
2888 						    0, VM_MEMATTR_DEFAULT);
2889 					}
2890 					if (m_new == NULL) {
2891 						vm_page_xunbusy(m);
2892 						error = ENOMEM;
2893 						goto unlock;
2894 					}
2895 
2896 					/*
2897 					 * Unmap the page and check for new
2898 					 * wirings that may have been acquired
2899 					 * through a pmap lookup.
2900 					 */
2901 					if (object->ref_count != 0 &&
2902 					    !vm_page_try_remove_all(m)) {
2903 						vm_page_xunbusy(m);
2904 						vm_page_free(m_new);
2905 						error = EBUSY;
2906 						goto unlock;
2907 					}
2908 
2909 					/*
2910 					 * Replace "m" with the new page.  For
2911 					 * vm_page_replace(), "m" must be busy
2912 					 * and dequeued.  Finally, change "m"
2913 					 * as if vm_page_free() was called.
2914 					 */
2915 					m_new->a.flags = m->a.flags &
2916 					    ~PGA_QUEUE_STATE_MASK;
2917 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2918 					    ("page %p is managed", m_new));
2919 					m_new->oflags = 0;
2920 					pmap_copy_page(m, m_new);
2921 					m_new->valid = m->valid;
2922 					m_new->dirty = m->dirty;
2923 					m->flags &= ~PG_ZERO;
2924 					vm_page_dequeue(m);
2925 					if (vm_page_replace_hold(m_new, object,
2926 					    m->pindex, m) &&
2927 					    vm_page_free_prep(m))
2928 						SLIST_INSERT_HEAD(&free, m,
2929 						    plinks.s.ss);
2930 
2931 					/*
2932 					 * The new page must be deactivated
2933 					 * before the object is unlocked.
2934 					 */
2935 					vm_page_deactivate(m_new);
2936 				} else {
2937 					m->flags &= ~PG_ZERO;
2938 					vm_page_dequeue(m);
2939 					if (vm_page_free_prep(m))
2940 						SLIST_INSERT_HEAD(&free, m,
2941 						    plinks.s.ss);
2942 					KASSERT(m->dirty == 0,
2943 					    ("page %p is dirty", m));
2944 				}
2945 			} else
2946 				error = EBUSY;
2947 unlock:
2948 			VM_OBJECT_WUNLOCK(object);
2949 		} else {
2950 			MPASS(vm_page_domain(m) == domain);
2951 			vmd = VM_DOMAIN(domain);
2952 			vm_domain_free_lock(vmd);
2953 			order = m->order;
2954 			if (order < VM_NFREEORDER) {
2955 				/*
2956 				 * The page is enqueued in the physical memory
2957 				 * allocator's free page queues.  Moreover, it
2958 				 * is the first page in a power-of-two-sized
2959 				 * run of contiguous free pages.  Jump ahead
2960 				 * to the last page within that run, and
2961 				 * continue from there.
2962 				 */
2963 				m += (1 << order) - 1;
2964 			}
2965 #if VM_NRESERVLEVEL > 0
2966 			else if (vm_reserv_is_page_free(m))
2967 				order = 0;
2968 #endif
2969 			vm_domain_free_unlock(vmd);
2970 			if (order == VM_NFREEORDER)
2971 				error = EINVAL;
2972 		}
2973 	}
2974 	if ((m = SLIST_FIRST(&free)) != NULL) {
2975 		int cnt;
2976 
2977 		vmd = VM_DOMAIN(domain);
2978 		cnt = 0;
2979 		vm_domain_free_lock(vmd);
2980 		do {
2981 			MPASS(vm_page_domain(m) == domain);
2982 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2983 			vm_phys_free_pages(m, 0);
2984 			cnt++;
2985 		} while ((m = SLIST_FIRST(&free)) != NULL);
2986 		vm_domain_free_unlock(vmd);
2987 		vm_domain_freecnt_inc(vmd, cnt);
2988 	}
2989 	return (error);
2990 }
2991 
2992 #define	NRUNS	16
2993 
2994 CTASSERT(powerof2(NRUNS));
2995 
2996 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2997 
2998 #define	MIN_RECLAIM	8
2999 
3000 /*
3001  *	vm_page_reclaim_contig:
3002  *
3003  *	Reclaim allocated, contiguous physical memory satisfying the specified
3004  *	conditions by relocating the virtual pages using that physical memory.
3005  *	Returns true if reclamation is successful and false otherwise.  Since
3006  *	relocation requires the allocation of physical pages, reclamation may
3007  *	fail due to a shortage of free pages.  When reclamation fails, callers
3008  *	are expected to perform vm_wait() before retrying a failed allocation
3009  *	operation, e.g., vm_page_alloc_contig().
3010  *
3011  *	The caller must always specify an allocation class through "req".
3012  *
3013  *	allocation classes:
3014  *	VM_ALLOC_NORMAL		normal process request
3015  *	VM_ALLOC_SYSTEM		system *really* needs a page
3016  *	VM_ALLOC_INTERRUPT	interrupt time request
3017  *
3018  *	The optional allocation flags are ignored.
3019  *
3020  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
3021  *	must be a power of two.
3022  */
3023 bool
3024 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
3025     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
3026 {
3027 	struct vm_domain *vmd;
3028 	vm_paddr_t curr_low;
3029 	vm_page_t m_run, m_runs[NRUNS];
3030 	u_long count, minalign, reclaimed;
3031 	int error, i, options, req_class;
3032 
3033 	KASSERT(npages > 0, ("npages is 0"));
3034 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
3035 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
3036 
3037 	/*
3038 	 * The caller will attempt an allocation after some runs have been
3039 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
3040 	 * of vm_phys_alloc_contig(), round up the requested length to the next
3041 	 * power of two or maximum chunk size, and ensure that each run is
3042 	 * suitably aligned.
3043 	 */
3044 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
3045 	npages = roundup2(npages, minalign);
3046 	if (alignment < ptoa(minalign))
3047 		alignment = ptoa(minalign);
3048 
3049 	/*
3050 	 * The page daemon is allowed to dig deeper into the free page list.
3051 	 */
3052 	req_class = req & VM_ALLOC_CLASS_MASK;
3053 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
3054 		req_class = VM_ALLOC_SYSTEM;
3055 
3056 	/*
3057 	 * Return if the number of free pages cannot satisfy the requested
3058 	 * allocation.
3059 	 */
3060 	vmd = VM_DOMAIN(domain);
3061 	count = vmd->vmd_free_count;
3062 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3063 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3064 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3065 		return (false);
3066 
3067 	/*
3068 	 * Scan up to three times, relaxing the restrictions ("options") on
3069 	 * the reclamation of reservations and superpages each time.
3070 	 */
3071 	for (options = VPSC_NORESERV;;) {
3072 		/*
3073 		 * Find the highest runs that satisfy the given constraints
3074 		 * and restrictions, and record them in "m_runs".
3075 		 */
3076 		curr_low = low;
3077 		count = 0;
3078 		for (;;) {
3079 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
3080 			    high, alignment, boundary, options);
3081 			if (m_run == NULL)
3082 				break;
3083 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
3084 			m_runs[RUN_INDEX(count)] = m_run;
3085 			count++;
3086 		}
3087 
3088 		/*
3089 		 * Reclaim the highest runs in LIFO (descending) order until
3090 		 * the number of reclaimed pages, "reclaimed", is at least
3091 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
3092 		 * reclamation is idempotent, and runs will (likely) recur
3093 		 * from one scan to the next as restrictions are relaxed.
3094 		 */
3095 		reclaimed = 0;
3096 		for (i = 0; count > 0 && i < NRUNS; i++) {
3097 			count--;
3098 			m_run = m_runs[RUN_INDEX(count)];
3099 			error = vm_page_reclaim_run(req_class, domain, npages,
3100 			    m_run, high);
3101 			if (error == 0) {
3102 				reclaimed += npages;
3103 				if (reclaimed >= MIN_RECLAIM)
3104 					return (true);
3105 			}
3106 		}
3107 
3108 		/*
3109 		 * Either relax the restrictions on the next scan or return if
3110 		 * the last scan had no restrictions.
3111 		 */
3112 		if (options == VPSC_NORESERV)
3113 			options = VPSC_NOSUPER;
3114 		else if (options == VPSC_NOSUPER)
3115 			options = VPSC_ANY;
3116 		else if (options == VPSC_ANY)
3117 			return (reclaimed != 0);
3118 	}
3119 }
3120 
3121 bool
3122 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3123     u_long alignment, vm_paddr_t boundary)
3124 {
3125 	struct vm_domainset_iter di;
3126 	int domain;
3127 	bool ret;
3128 
3129 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3130 	do {
3131 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3132 		    high, alignment, boundary);
3133 		if (ret)
3134 			break;
3135 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3136 
3137 	return (ret);
3138 }
3139 
3140 /*
3141  * Set the domain in the appropriate page level domainset.
3142  */
3143 void
3144 vm_domain_set(struct vm_domain *vmd)
3145 {
3146 
3147 	mtx_lock(&vm_domainset_lock);
3148 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3149 		vmd->vmd_minset = 1;
3150 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3151 	}
3152 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3153 		vmd->vmd_severeset = 1;
3154 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3155 	}
3156 	mtx_unlock(&vm_domainset_lock);
3157 }
3158 
3159 /*
3160  * Clear the domain from the appropriate page level domainset.
3161  */
3162 void
3163 vm_domain_clear(struct vm_domain *vmd)
3164 {
3165 
3166 	mtx_lock(&vm_domainset_lock);
3167 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3168 		vmd->vmd_minset = 0;
3169 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3170 		if (vm_min_waiters != 0) {
3171 			vm_min_waiters = 0;
3172 			wakeup(&vm_min_domains);
3173 		}
3174 	}
3175 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3176 		vmd->vmd_severeset = 0;
3177 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3178 		if (vm_severe_waiters != 0) {
3179 			vm_severe_waiters = 0;
3180 			wakeup(&vm_severe_domains);
3181 		}
3182 	}
3183 
3184 	/*
3185 	 * If pageout daemon needs pages, then tell it that there are
3186 	 * some free.
3187 	 */
3188 	if (vmd->vmd_pageout_pages_needed &&
3189 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3190 		wakeup(&vmd->vmd_pageout_pages_needed);
3191 		vmd->vmd_pageout_pages_needed = 0;
3192 	}
3193 
3194 	/* See comments in vm_wait_doms(). */
3195 	if (vm_pageproc_waiters) {
3196 		vm_pageproc_waiters = 0;
3197 		wakeup(&vm_pageproc_waiters);
3198 	}
3199 	mtx_unlock(&vm_domainset_lock);
3200 }
3201 
3202 /*
3203  * Wait for free pages to exceed the min threshold globally.
3204  */
3205 void
3206 vm_wait_min(void)
3207 {
3208 
3209 	mtx_lock(&vm_domainset_lock);
3210 	while (vm_page_count_min()) {
3211 		vm_min_waiters++;
3212 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3213 	}
3214 	mtx_unlock(&vm_domainset_lock);
3215 }
3216 
3217 /*
3218  * Wait for free pages to exceed the severe threshold globally.
3219  */
3220 void
3221 vm_wait_severe(void)
3222 {
3223 
3224 	mtx_lock(&vm_domainset_lock);
3225 	while (vm_page_count_severe()) {
3226 		vm_severe_waiters++;
3227 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3228 		    "vmwait", 0);
3229 	}
3230 	mtx_unlock(&vm_domainset_lock);
3231 }
3232 
3233 u_int
3234 vm_wait_count(void)
3235 {
3236 
3237 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3238 }
3239 
3240 int
3241 vm_wait_doms(const domainset_t *wdoms, int mflags)
3242 {
3243 	int error;
3244 
3245 	error = 0;
3246 
3247 	/*
3248 	 * We use racey wakeup synchronization to avoid expensive global
3249 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3250 	 * To handle this, we only sleep for one tick in this instance.  It
3251 	 * is expected that most allocations for the pageproc will come from
3252 	 * kmem or vm_page_grab* which will use the more specific and
3253 	 * race-free vm_wait_domain().
3254 	 */
3255 	if (curproc == pageproc) {
3256 		mtx_lock(&vm_domainset_lock);
3257 		vm_pageproc_waiters++;
3258 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3259 		    PVM | PDROP | mflags, "pageprocwait", 1);
3260 	} else {
3261 		/*
3262 		 * XXX Ideally we would wait only until the allocation could
3263 		 * be satisfied.  This condition can cause new allocators to
3264 		 * consume all freed pages while old allocators wait.
3265 		 */
3266 		mtx_lock(&vm_domainset_lock);
3267 		if (vm_page_count_min_set(wdoms)) {
3268 			vm_min_waiters++;
3269 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3270 			    PVM | PDROP | mflags, "vmwait", 0);
3271 		} else
3272 			mtx_unlock(&vm_domainset_lock);
3273 	}
3274 	return (error);
3275 }
3276 
3277 /*
3278  *	vm_wait_domain:
3279  *
3280  *	Sleep until free pages are available for allocation.
3281  *	- Called in various places after failed memory allocations.
3282  */
3283 void
3284 vm_wait_domain(int domain)
3285 {
3286 	struct vm_domain *vmd;
3287 	domainset_t wdom;
3288 
3289 	vmd = VM_DOMAIN(domain);
3290 	vm_domain_free_assert_unlocked(vmd);
3291 
3292 	if (curproc == pageproc) {
3293 		mtx_lock(&vm_domainset_lock);
3294 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3295 			vmd->vmd_pageout_pages_needed = 1;
3296 			msleep(&vmd->vmd_pageout_pages_needed,
3297 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3298 		} else
3299 			mtx_unlock(&vm_domainset_lock);
3300 	} else {
3301 		if (pageproc == NULL)
3302 			panic("vm_wait in early boot");
3303 		DOMAINSET_ZERO(&wdom);
3304 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3305 		vm_wait_doms(&wdom, 0);
3306 	}
3307 }
3308 
3309 static int
3310 vm_wait_flags(vm_object_t obj, int mflags)
3311 {
3312 	struct domainset *d;
3313 
3314 	d = NULL;
3315 
3316 	/*
3317 	 * Carefully fetch pointers only once: the struct domainset
3318 	 * itself is ummutable but the pointer might change.
3319 	 */
3320 	if (obj != NULL)
3321 		d = obj->domain.dr_policy;
3322 	if (d == NULL)
3323 		d = curthread->td_domain.dr_policy;
3324 
3325 	return (vm_wait_doms(&d->ds_mask, mflags));
3326 }
3327 
3328 /*
3329  *	vm_wait:
3330  *
3331  *	Sleep until free pages are available for allocation in the
3332  *	affinity domains of the obj.  If obj is NULL, the domain set
3333  *	for the calling thread is used.
3334  *	Called in various places after failed memory allocations.
3335  */
3336 void
3337 vm_wait(vm_object_t obj)
3338 {
3339 	(void)vm_wait_flags(obj, 0);
3340 }
3341 
3342 int
3343 vm_wait_intr(vm_object_t obj)
3344 {
3345 	return (vm_wait_flags(obj, PCATCH));
3346 }
3347 
3348 /*
3349  *	vm_domain_alloc_fail:
3350  *
3351  *	Called when a page allocation function fails.  Informs the
3352  *	pagedaemon and performs the requested wait.  Requires the
3353  *	domain_free and object lock on entry.  Returns with the
3354  *	object lock held and free lock released.  Returns an error when
3355  *	retry is necessary.
3356  *
3357  */
3358 static int
3359 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3360 {
3361 
3362 	vm_domain_free_assert_unlocked(vmd);
3363 
3364 	atomic_add_int(&vmd->vmd_pageout_deficit,
3365 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3366 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3367 		if (object != NULL)
3368 			VM_OBJECT_WUNLOCK(object);
3369 		vm_wait_domain(vmd->vmd_domain);
3370 		if (object != NULL)
3371 			VM_OBJECT_WLOCK(object);
3372 		if (req & VM_ALLOC_WAITOK)
3373 			return (EAGAIN);
3374 	}
3375 
3376 	return (0);
3377 }
3378 
3379 /*
3380  *	vm_waitpfault:
3381  *
3382  *	Sleep until free pages are available for allocation.
3383  *	- Called only in vm_fault so that processes page faulting
3384  *	  can be easily tracked.
3385  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3386  *	  processes will be able to grab memory first.  Do not change
3387  *	  this balance without careful testing first.
3388  */
3389 void
3390 vm_waitpfault(struct domainset *dset, int timo)
3391 {
3392 
3393 	/*
3394 	 * XXX Ideally we would wait only until the allocation could
3395 	 * be satisfied.  This condition can cause new allocators to
3396 	 * consume all freed pages while old allocators wait.
3397 	 */
3398 	mtx_lock(&vm_domainset_lock);
3399 	if (vm_page_count_min_set(&dset->ds_mask)) {
3400 		vm_min_waiters++;
3401 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3402 		    "pfault", timo);
3403 	} else
3404 		mtx_unlock(&vm_domainset_lock);
3405 }
3406 
3407 static struct vm_pagequeue *
3408 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3409 {
3410 
3411 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3412 }
3413 
3414 #ifdef INVARIANTS
3415 static struct vm_pagequeue *
3416 vm_page_pagequeue(vm_page_t m)
3417 {
3418 
3419 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3420 }
3421 #endif
3422 
3423 static __always_inline bool
3424 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3425 {
3426 	vm_page_astate_t tmp;
3427 
3428 	tmp = *old;
3429 	do {
3430 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3431 			return (true);
3432 		counter_u64_add(pqstate_commit_retries, 1);
3433 	} while (old->_bits == tmp._bits);
3434 
3435 	return (false);
3436 }
3437 
3438 /*
3439  * Do the work of committing a queue state update that moves the page out of
3440  * its current queue.
3441  */
3442 static bool
3443 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3444     vm_page_astate_t *old, vm_page_astate_t new)
3445 {
3446 	vm_page_t next;
3447 
3448 	vm_pagequeue_assert_locked(pq);
3449 	KASSERT(vm_page_pagequeue(m) == pq,
3450 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3451 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3452 	    ("%s: invalid queue indices %d %d",
3453 	    __func__, old->queue, new.queue));
3454 
3455 	/*
3456 	 * Once the queue index of the page changes there is nothing
3457 	 * synchronizing with further updates to the page's physical
3458 	 * queue state.  Therefore we must speculatively remove the page
3459 	 * from the queue now and be prepared to roll back if the queue
3460 	 * state update fails.  If the page is not physically enqueued then
3461 	 * we just update its queue index.
3462 	 */
3463 	if ((old->flags & PGA_ENQUEUED) != 0) {
3464 		new.flags &= ~PGA_ENQUEUED;
3465 		next = TAILQ_NEXT(m, plinks.q);
3466 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3467 		vm_pagequeue_cnt_dec(pq);
3468 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3469 			if (next == NULL)
3470 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3471 			else
3472 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3473 			vm_pagequeue_cnt_inc(pq);
3474 			return (false);
3475 		} else {
3476 			return (true);
3477 		}
3478 	} else {
3479 		return (vm_page_pqstate_fcmpset(m, old, new));
3480 	}
3481 }
3482 
3483 static bool
3484 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3485     vm_page_astate_t new)
3486 {
3487 	struct vm_pagequeue *pq;
3488 	vm_page_astate_t as;
3489 	bool ret;
3490 
3491 	pq = _vm_page_pagequeue(m, old->queue);
3492 
3493 	/*
3494 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3495 	 * corresponding page queue lock is held.
3496 	 */
3497 	vm_pagequeue_lock(pq);
3498 	as = vm_page_astate_load(m);
3499 	if (__predict_false(as._bits != old->_bits)) {
3500 		*old = as;
3501 		ret = false;
3502 	} else {
3503 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3504 	}
3505 	vm_pagequeue_unlock(pq);
3506 	return (ret);
3507 }
3508 
3509 /*
3510  * Commit a queue state update that enqueues or requeues a page.
3511  */
3512 static bool
3513 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3514     vm_page_astate_t *old, vm_page_astate_t new)
3515 {
3516 	struct vm_domain *vmd;
3517 
3518 	vm_pagequeue_assert_locked(pq);
3519 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3520 	    ("%s: invalid queue indices %d %d",
3521 	    __func__, old->queue, new.queue));
3522 
3523 	new.flags |= PGA_ENQUEUED;
3524 	if (!vm_page_pqstate_fcmpset(m, old, new))
3525 		return (false);
3526 
3527 	if ((old->flags & PGA_ENQUEUED) != 0)
3528 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3529 	else
3530 		vm_pagequeue_cnt_inc(pq);
3531 
3532 	/*
3533 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3534 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3535 	 * applied, even if it was set first.
3536 	 */
3537 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3538 		vmd = vm_pagequeue_domain(m);
3539 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3540 		    ("%s: invalid page queue for page %p", __func__, m));
3541 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3542 	} else {
3543 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3544 	}
3545 	return (true);
3546 }
3547 
3548 /*
3549  * Commit a queue state update that encodes a request for a deferred queue
3550  * operation.
3551  */
3552 static bool
3553 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3554     vm_page_astate_t new)
3555 {
3556 
3557 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3558 	    ("%s: invalid state, queue %d flags %x",
3559 	    __func__, new.queue, new.flags));
3560 
3561 	if (old->_bits != new._bits &&
3562 	    !vm_page_pqstate_fcmpset(m, old, new))
3563 		return (false);
3564 	vm_page_pqbatch_submit(m, new.queue);
3565 	return (true);
3566 }
3567 
3568 /*
3569  * A generic queue state update function.  This handles more cases than the
3570  * specialized functions above.
3571  */
3572 bool
3573 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3574 {
3575 
3576 	if (old->_bits == new._bits)
3577 		return (true);
3578 
3579 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3580 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3581 			return (false);
3582 		if (new.queue != PQ_NONE)
3583 			vm_page_pqbatch_submit(m, new.queue);
3584 	} else {
3585 		if (!vm_page_pqstate_fcmpset(m, old, new))
3586 			return (false);
3587 		if (new.queue != PQ_NONE &&
3588 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3589 			vm_page_pqbatch_submit(m, new.queue);
3590 	}
3591 	return (true);
3592 }
3593 
3594 /*
3595  * Apply deferred queue state updates to a page.
3596  */
3597 static inline void
3598 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3599 {
3600 	vm_page_astate_t new, old;
3601 
3602 	CRITICAL_ASSERT(curthread);
3603 	vm_pagequeue_assert_locked(pq);
3604 	KASSERT(queue < PQ_COUNT,
3605 	    ("%s: invalid queue index %d", __func__, queue));
3606 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3607 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3608 
3609 	for (old = vm_page_astate_load(m);;) {
3610 		if (__predict_false(old.queue != queue ||
3611 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3612 			counter_u64_add(queue_nops, 1);
3613 			break;
3614 		}
3615 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3616 		    ("%s: page %p is unmanaged", __func__, m));
3617 
3618 		new = old;
3619 		if ((old.flags & PGA_DEQUEUE) != 0) {
3620 			new.flags &= ~PGA_QUEUE_OP_MASK;
3621 			new.queue = PQ_NONE;
3622 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3623 			    m, &old, new))) {
3624 				counter_u64_add(queue_ops, 1);
3625 				break;
3626 			}
3627 		} else {
3628 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3629 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3630 			    m, &old, new))) {
3631 				counter_u64_add(queue_ops, 1);
3632 				break;
3633 			}
3634 		}
3635 	}
3636 }
3637 
3638 static void
3639 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3640     uint8_t queue)
3641 {
3642 	int i;
3643 
3644 	for (i = 0; i < bq->bq_cnt; i++)
3645 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3646 	vm_batchqueue_init(bq);
3647 }
3648 
3649 /*
3650  *	vm_page_pqbatch_submit:		[ internal use only ]
3651  *
3652  *	Enqueue a page in the specified page queue's batched work queue.
3653  *	The caller must have encoded the requested operation in the page
3654  *	structure's a.flags field.
3655  */
3656 void
3657 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3658 {
3659 	struct vm_batchqueue *bq;
3660 	struct vm_pagequeue *pq;
3661 	int domain;
3662 
3663 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3664 
3665 	domain = vm_page_domain(m);
3666 	critical_enter();
3667 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3668 	if (vm_batchqueue_insert(bq, m)) {
3669 		critical_exit();
3670 		return;
3671 	}
3672 	critical_exit();
3673 
3674 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3675 	vm_pagequeue_lock(pq);
3676 	critical_enter();
3677 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3678 	vm_pqbatch_process(pq, bq, queue);
3679 	vm_pqbatch_process_page(pq, m, queue);
3680 	vm_pagequeue_unlock(pq);
3681 	critical_exit();
3682 }
3683 
3684 /*
3685  *	vm_page_pqbatch_drain:		[ internal use only ]
3686  *
3687  *	Force all per-CPU page queue batch queues to be drained.  This is
3688  *	intended for use in severe memory shortages, to ensure that pages
3689  *	do not remain stuck in the batch queues.
3690  */
3691 void
3692 vm_page_pqbatch_drain(void)
3693 {
3694 	struct thread *td;
3695 	struct vm_domain *vmd;
3696 	struct vm_pagequeue *pq;
3697 	int cpu, domain, queue;
3698 
3699 	td = curthread;
3700 	CPU_FOREACH(cpu) {
3701 		thread_lock(td);
3702 		sched_bind(td, cpu);
3703 		thread_unlock(td);
3704 
3705 		for (domain = 0; domain < vm_ndomains; domain++) {
3706 			vmd = VM_DOMAIN(domain);
3707 			for (queue = 0; queue < PQ_COUNT; queue++) {
3708 				pq = &vmd->vmd_pagequeues[queue];
3709 				vm_pagequeue_lock(pq);
3710 				critical_enter();
3711 				vm_pqbatch_process(pq,
3712 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3713 				critical_exit();
3714 				vm_pagequeue_unlock(pq);
3715 			}
3716 		}
3717 	}
3718 	thread_lock(td);
3719 	sched_unbind(td);
3720 	thread_unlock(td);
3721 }
3722 
3723 /*
3724  *	vm_page_dequeue_deferred:	[ internal use only ]
3725  *
3726  *	Request removal of the given page from its current page
3727  *	queue.  Physical removal from the queue may be deferred
3728  *	indefinitely.
3729  */
3730 void
3731 vm_page_dequeue_deferred(vm_page_t m)
3732 {
3733 	vm_page_astate_t new, old;
3734 
3735 	old = vm_page_astate_load(m);
3736 	do {
3737 		if (old.queue == PQ_NONE) {
3738 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3739 			    ("%s: page %p has unexpected queue state",
3740 			    __func__, m));
3741 			break;
3742 		}
3743 		new = old;
3744 		new.flags |= PGA_DEQUEUE;
3745 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3746 }
3747 
3748 /*
3749  *	vm_page_dequeue:
3750  *
3751  *	Remove the page from whichever page queue it's in, if any, before
3752  *	returning.
3753  */
3754 void
3755 vm_page_dequeue(vm_page_t m)
3756 {
3757 	vm_page_astate_t new, old;
3758 
3759 	old = vm_page_astate_load(m);
3760 	do {
3761 		if (old.queue == PQ_NONE) {
3762 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3763 			    ("%s: page %p has unexpected queue state",
3764 			    __func__, m));
3765 			break;
3766 		}
3767 		new = old;
3768 		new.flags &= ~PGA_QUEUE_OP_MASK;
3769 		new.queue = PQ_NONE;
3770 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3771 
3772 }
3773 
3774 /*
3775  * Schedule the given page for insertion into the specified page queue.
3776  * Physical insertion of the page may be deferred indefinitely.
3777  */
3778 static void
3779 vm_page_enqueue(vm_page_t m, uint8_t queue)
3780 {
3781 
3782 	KASSERT(m->a.queue == PQ_NONE &&
3783 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3784 	    ("%s: page %p is already enqueued", __func__, m));
3785 	KASSERT(m->ref_count > 0,
3786 	    ("%s: page %p does not carry any references", __func__, m));
3787 
3788 	m->a.queue = queue;
3789 	if ((m->a.flags & PGA_REQUEUE) == 0)
3790 		vm_page_aflag_set(m, PGA_REQUEUE);
3791 	vm_page_pqbatch_submit(m, queue);
3792 }
3793 
3794 /*
3795  *	vm_page_free_prep:
3796  *
3797  *	Prepares the given page to be put on the free list,
3798  *	disassociating it from any VM object. The caller may return
3799  *	the page to the free list only if this function returns true.
3800  *
3801  *	The object, if it exists, must be locked, and then the page must
3802  *	be xbusy.  Otherwise the page must be not busied.  A managed
3803  *	page must be unmapped.
3804  */
3805 static bool
3806 vm_page_free_prep(vm_page_t m)
3807 {
3808 
3809 	/*
3810 	 * Synchronize with threads that have dropped a reference to this
3811 	 * page.
3812 	 */
3813 	atomic_thread_fence_acq();
3814 
3815 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3816 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3817 		uint64_t *p;
3818 		int i;
3819 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3820 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3821 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3822 			    m, i, (uintmax_t)*p));
3823 	}
3824 #endif
3825 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3826 		KASSERT(!pmap_page_is_mapped(m),
3827 		    ("vm_page_free_prep: freeing mapped page %p", m));
3828 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3829 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3830 	} else {
3831 		KASSERT(m->a.queue == PQ_NONE,
3832 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3833 	}
3834 	VM_CNT_INC(v_tfree);
3835 
3836 	if (m->object != NULL) {
3837 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3838 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
3839 		    ("vm_page_free_prep: managed flag mismatch for page %p",
3840 		    m));
3841 		vm_page_assert_xbusied(m);
3842 
3843 		/*
3844 		 * The object reference can be released without an atomic
3845 		 * operation.
3846 		 */
3847 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3848 		    m->ref_count == VPRC_OBJREF,
3849 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3850 		    m, m->ref_count));
3851 		vm_page_object_remove(m);
3852 		m->ref_count -= VPRC_OBJREF;
3853 	} else
3854 		vm_page_assert_unbusied(m);
3855 
3856 	vm_page_busy_free(m);
3857 
3858 	/*
3859 	 * If fictitious remove object association and
3860 	 * return.
3861 	 */
3862 	if ((m->flags & PG_FICTITIOUS) != 0) {
3863 		KASSERT(m->ref_count == 1,
3864 		    ("fictitious page %p is referenced", m));
3865 		KASSERT(m->a.queue == PQ_NONE,
3866 		    ("fictitious page %p is queued", m));
3867 		return (false);
3868 	}
3869 
3870 	/*
3871 	 * Pages need not be dequeued before they are returned to the physical
3872 	 * memory allocator, but they must at least be marked for a deferred
3873 	 * dequeue.
3874 	 */
3875 	if ((m->oflags & VPO_UNMANAGED) == 0)
3876 		vm_page_dequeue_deferred(m);
3877 
3878 	m->valid = 0;
3879 	vm_page_undirty(m);
3880 
3881 	if (m->ref_count != 0)
3882 		panic("vm_page_free_prep: page %p has references", m);
3883 
3884 	/*
3885 	 * Restore the default memory attribute to the page.
3886 	 */
3887 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3888 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3889 
3890 #if VM_NRESERVLEVEL > 0
3891 	/*
3892 	 * Determine whether the page belongs to a reservation.  If the page was
3893 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3894 	 * as an optimization, we avoid the check in that case.
3895 	 */
3896 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3897 		return (false);
3898 #endif
3899 
3900 	return (true);
3901 }
3902 
3903 /*
3904  *	vm_page_free_toq:
3905  *
3906  *	Returns the given page to the free list, disassociating it
3907  *	from any VM object.
3908  *
3909  *	The object must be locked.  The page must be exclusively busied if it
3910  *	belongs to an object.
3911  */
3912 static void
3913 vm_page_free_toq(vm_page_t m)
3914 {
3915 	struct vm_domain *vmd;
3916 	uma_zone_t zone;
3917 
3918 	if (!vm_page_free_prep(m))
3919 		return;
3920 
3921 	vmd = vm_pagequeue_domain(m);
3922 	zone = vmd->vmd_pgcache[m->pool].zone;
3923 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3924 		uma_zfree(zone, m);
3925 		return;
3926 	}
3927 	vm_domain_free_lock(vmd);
3928 	vm_phys_free_pages(m, 0);
3929 	vm_domain_free_unlock(vmd);
3930 	vm_domain_freecnt_inc(vmd, 1);
3931 }
3932 
3933 /*
3934  *	vm_page_free_pages_toq:
3935  *
3936  *	Returns a list of pages to the free list, disassociating it
3937  *	from any VM object.  In other words, this is equivalent to
3938  *	calling vm_page_free_toq() for each page of a list of VM objects.
3939  */
3940 void
3941 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3942 {
3943 	vm_page_t m;
3944 	int count;
3945 
3946 	if (SLIST_EMPTY(free))
3947 		return;
3948 
3949 	count = 0;
3950 	while ((m = SLIST_FIRST(free)) != NULL) {
3951 		count++;
3952 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3953 		vm_page_free_toq(m);
3954 	}
3955 
3956 	if (update_wire_count)
3957 		vm_wire_sub(count);
3958 }
3959 
3960 /*
3961  * Mark this page as wired down.  For managed pages, this prevents reclamation
3962  * by the page daemon, or when the containing object, if any, is destroyed.
3963  */
3964 void
3965 vm_page_wire(vm_page_t m)
3966 {
3967 	u_int old;
3968 
3969 #ifdef INVARIANTS
3970 	if (m->object != NULL && !vm_page_busied(m) &&
3971 	    !vm_object_busied(m->object))
3972 		VM_OBJECT_ASSERT_LOCKED(m->object);
3973 #endif
3974 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3975 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
3976 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
3977 
3978 	old = atomic_fetchadd_int(&m->ref_count, 1);
3979 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3980 	    ("vm_page_wire: counter overflow for page %p", m));
3981 	if (VPRC_WIRE_COUNT(old) == 0) {
3982 		if ((m->oflags & VPO_UNMANAGED) == 0)
3983 			vm_page_aflag_set(m, PGA_DEQUEUE);
3984 		vm_wire_add(1);
3985 	}
3986 }
3987 
3988 /*
3989  * Attempt to wire a mapped page following a pmap lookup of that page.
3990  * This may fail if a thread is concurrently tearing down mappings of the page.
3991  * The transient failure is acceptable because it translates to the
3992  * failure of the caller pmap_extract_and_hold(), which should be then
3993  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3994  */
3995 bool
3996 vm_page_wire_mapped(vm_page_t m)
3997 {
3998 	u_int old;
3999 
4000 	old = m->ref_count;
4001 	do {
4002 		KASSERT(old > 0,
4003 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
4004 		if ((old & VPRC_BLOCKED) != 0)
4005 			return (false);
4006 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
4007 
4008 	if (VPRC_WIRE_COUNT(old) == 0) {
4009 		if ((m->oflags & VPO_UNMANAGED) == 0)
4010 			vm_page_aflag_set(m, PGA_DEQUEUE);
4011 		vm_wire_add(1);
4012 	}
4013 	return (true);
4014 }
4015 
4016 /*
4017  * Release a wiring reference to a managed page.  If the page still belongs to
4018  * an object, update its position in the page queues to reflect the reference.
4019  * If the wiring was the last reference to the page, free the page.
4020  */
4021 static void
4022 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
4023 {
4024 	u_int old;
4025 
4026 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4027 	    ("%s: page %p is unmanaged", __func__, m));
4028 
4029 	/*
4030 	 * Update LRU state before releasing the wiring reference.
4031 	 * Use a release store when updating the reference count to
4032 	 * synchronize with vm_page_free_prep().
4033 	 */
4034 	old = m->ref_count;
4035 	do {
4036 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4037 		    ("vm_page_unwire: wire count underflow for page %p", m));
4038 
4039 		if (old > VPRC_OBJREF + 1) {
4040 			/*
4041 			 * The page has at least one other wiring reference.  An
4042 			 * earlier iteration of this loop may have called
4043 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
4044 			 * re-set it if necessary.
4045 			 */
4046 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
4047 				vm_page_aflag_set(m, PGA_DEQUEUE);
4048 		} else if (old == VPRC_OBJREF + 1) {
4049 			/*
4050 			 * This is the last wiring.  Clear PGA_DEQUEUE and
4051 			 * update the page's queue state to reflect the
4052 			 * reference.  If the page does not belong to an object
4053 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
4054 			 * clear leftover queue state.
4055 			 */
4056 			vm_page_release_toq(m, nqueue, noreuse);
4057 		} else if (old == 1) {
4058 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4059 		}
4060 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4061 
4062 	if (VPRC_WIRE_COUNT(old) == 1) {
4063 		vm_wire_sub(1);
4064 		if (old == 1)
4065 			vm_page_free(m);
4066 	}
4067 }
4068 
4069 /*
4070  * Release one wiring of the specified page, potentially allowing it to be
4071  * paged out.
4072  *
4073  * Only managed pages belonging to an object can be paged out.  If the number
4074  * of wirings transitions to zero and the page is eligible for page out, then
4075  * the page is added to the specified paging queue.  If the released wiring
4076  * represented the last reference to the page, the page is freed.
4077  */
4078 void
4079 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4080 {
4081 
4082 	KASSERT(nqueue < PQ_COUNT,
4083 	    ("vm_page_unwire: invalid queue %u request for page %p",
4084 	    nqueue, m));
4085 
4086 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4087 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4088 			vm_page_free(m);
4089 		return;
4090 	}
4091 	vm_page_unwire_managed(m, nqueue, false);
4092 }
4093 
4094 /*
4095  * Unwire a page without (re-)inserting it into a page queue.  It is up
4096  * to the caller to enqueue, requeue, or free the page as appropriate.
4097  * In most cases involving managed pages, vm_page_unwire() should be used
4098  * instead.
4099  */
4100 bool
4101 vm_page_unwire_noq(vm_page_t m)
4102 {
4103 	u_int old;
4104 
4105 	old = vm_page_drop(m, 1);
4106 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4107 	    ("%s: counter underflow for page %p", __func__,  m));
4108 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4109 	    ("%s: missing ref on fictitious page %p", __func__, m));
4110 
4111 	if (VPRC_WIRE_COUNT(old) > 1)
4112 		return (false);
4113 	if ((m->oflags & VPO_UNMANAGED) == 0)
4114 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4115 	vm_wire_sub(1);
4116 	return (true);
4117 }
4118 
4119 /*
4120  * Ensure that the page ends up in the specified page queue.  If the page is
4121  * active or being moved to the active queue, ensure that its act_count is
4122  * at least ACT_INIT but do not otherwise mess with it.
4123  */
4124 static __always_inline void
4125 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4126 {
4127 	vm_page_astate_t old, new;
4128 
4129 	KASSERT(m->ref_count > 0,
4130 	    ("%s: page %p does not carry any references", __func__, m));
4131 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4132 	    ("%s: invalid flags %x", __func__, nflag));
4133 
4134 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4135 		return;
4136 
4137 	old = vm_page_astate_load(m);
4138 	do {
4139 		if ((old.flags & PGA_DEQUEUE) != 0)
4140 			break;
4141 		new = old;
4142 		new.flags &= ~PGA_QUEUE_OP_MASK;
4143 		if (nqueue == PQ_ACTIVE)
4144 			new.act_count = max(old.act_count, ACT_INIT);
4145 		if (old.queue == nqueue) {
4146 			if (nqueue != PQ_ACTIVE)
4147 				new.flags |= nflag;
4148 		} else {
4149 			new.flags |= nflag;
4150 			new.queue = nqueue;
4151 		}
4152 	} while (!vm_page_pqstate_commit(m, &old, new));
4153 }
4154 
4155 /*
4156  * Put the specified page on the active list (if appropriate).
4157  */
4158 void
4159 vm_page_activate(vm_page_t m)
4160 {
4161 
4162 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4163 }
4164 
4165 /*
4166  * Move the specified page to the tail of the inactive queue, or requeue
4167  * the page if it is already in the inactive queue.
4168  */
4169 void
4170 vm_page_deactivate(vm_page_t m)
4171 {
4172 
4173 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4174 }
4175 
4176 void
4177 vm_page_deactivate_noreuse(vm_page_t m)
4178 {
4179 
4180 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4181 }
4182 
4183 /*
4184  * Put a page in the laundry, or requeue it if it is already there.
4185  */
4186 void
4187 vm_page_launder(vm_page_t m)
4188 {
4189 
4190 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4191 }
4192 
4193 /*
4194  * Put a page in the PQ_UNSWAPPABLE holding queue.
4195  */
4196 void
4197 vm_page_unswappable(vm_page_t m)
4198 {
4199 
4200 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4201 	    ("page %p already unswappable", m));
4202 
4203 	vm_page_dequeue(m);
4204 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4205 }
4206 
4207 /*
4208  * Release a page back to the page queues in preparation for unwiring.
4209  */
4210 static void
4211 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4212 {
4213 	vm_page_astate_t old, new;
4214 	uint16_t nflag;
4215 
4216 	/*
4217 	 * Use a check of the valid bits to determine whether we should
4218 	 * accelerate reclamation of the page.  The object lock might not be
4219 	 * held here, in which case the check is racy.  At worst we will either
4220 	 * accelerate reclamation of a valid page and violate LRU, or
4221 	 * unnecessarily defer reclamation of an invalid page.
4222 	 *
4223 	 * If we were asked to not cache the page, place it near the head of the
4224 	 * inactive queue so that is reclaimed sooner.
4225 	 */
4226 	if (noreuse || m->valid == 0) {
4227 		nqueue = PQ_INACTIVE;
4228 		nflag = PGA_REQUEUE_HEAD;
4229 	} else {
4230 		nflag = PGA_REQUEUE;
4231 	}
4232 
4233 	old = vm_page_astate_load(m);
4234 	do {
4235 		new = old;
4236 
4237 		/*
4238 		 * If the page is already in the active queue and we are not
4239 		 * trying to accelerate reclamation, simply mark it as
4240 		 * referenced and avoid any queue operations.
4241 		 */
4242 		new.flags &= ~PGA_QUEUE_OP_MASK;
4243 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4244 			new.flags |= PGA_REFERENCED;
4245 		else {
4246 			new.flags |= nflag;
4247 			new.queue = nqueue;
4248 		}
4249 	} while (!vm_page_pqstate_commit(m, &old, new));
4250 }
4251 
4252 /*
4253  * Unwire a page and either attempt to free it or re-add it to the page queues.
4254  */
4255 void
4256 vm_page_release(vm_page_t m, int flags)
4257 {
4258 	vm_object_t object;
4259 
4260 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4261 	    ("vm_page_release: page %p is unmanaged", m));
4262 
4263 	if ((flags & VPR_TRYFREE) != 0) {
4264 		for (;;) {
4265 			object = atomic_load_ptr(&m->object);
4266 			if (object == NULL)
4267 				break;
4268 			/* Depends on type-stability. */
4269 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4270 				break;
4271 			if (object == m->object) {
4272 				vm_page_release_locked(m, flags);
4273 				VM_OBJECT_WUNLOCK(object);
4274 				return;
4275 			}
4276 			VM_OBJECT_WUNLOCK(object);
4277 		}
4278 	}
4279 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4280 }
4281 
4282 /* See vm_page_release(). */
4283 void
4284 vm_page_release_locked(vm_page_t m, int flags)
4285 {
4286 
4287 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4288 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4289 	    ("vm_page_release_locked: page %p is unmanaged", m));
4290 
4291 	if (vm_page_unwire_noq(m)) {
4292 		if ((flags & VPR_TRYFREE) != 0 &&
4293 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4294 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4295 			/*
4296 			 * An unlocked lookup may have wired the page before the
4297 			 * busy lock was acquired, in which case the page must
4298 			 * not be freed.
4299 			 */
4300 			if (__predict_true(!vm_page_wired(m))) {
4301 				vm_page_free(m);
4302 				return;
4303 			}
4304 			vm_page_xunbusy(m);
4305 		} else {
4306 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4307 		}
4308 	}
4309 }
4310 
4311 static bool
4312 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4313 {
4314 	u_int old;
4315 
4316 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4317 	    ("vm_page_try_blocked_op: page %p has no object", m));
4318 	KASSERT(vm_page_busied(m),
4319 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4320 	VM_OBJECT_ASSERT_LOCKED(m->object);
4321 
4322 	old = m->ref_count;
4323 	do {
4324 		KASSERT(old != 0,
4325 		    ("vm_page_try_blocked_op: page %p has no references", m));
4326 		if (VPRC_WIRE_COUNT(old) != 0)
4327 			return (false);
4328 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4329 
4330 	(op)(m);
4331 
4332 	/*
4333 	 * If the object is read-locked, new wirings may be created via an
4334 	 * object lookup.
4335 	 */
4336 	old = vm_page_drop(m, VPRC_BLOCKED);
4337 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4338 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4339 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4340 	    old, m));
4341 	return (true);
4342 }
4343 
4344 /*
4345  * Atomically check for wirings and remove all mappings of the page.
4346  */
4347 bool
4348 vm_page_try_remove_all(vm_page_t m)
4349 {
4350 
4351 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4352 }
4353 
4354 /*
4355  * Atomically check for wirings and remove all writeable mappings of the page.
4356  */
4357 bool
4358 vm_page_try_remove_write(vm_page_t m)
4359 {
4360 
4361 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4362 }
4363 
4364 /*
4365  * vm_page_advise
4366  *
4367  * 	Apply the specified advice to the given page.
4368  */
4369 void
4370 vm_page_advise(vm_page_t m, int advice)
4371 {
4372 
4373 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4374 	vm_page_assert_xbusied(m);
4375 
4376 	if (advice == MADV_FREE)
4377 		/*
4378 		 * Mark the page clean.  This will allow the page to be freed
4379 		 * without first paging it out.  MADV_FREE pages are often
4380 		 * quickly reused by malloc(3), so we do not do anything that
4381 		 * would result in a page fault on a later access.
4382 		 */
4383 		vm_page_undirty(m);
4384 	else if (advice != MADV_DONTNEED) {
4385 		if (advice == MADV_WILLNEED)
4386 			vm_page_activate(m);
4387 		return;
4388 	}
4389 
4390 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4391 		vm_page_dirty(m);
4392 
4393 	/*
4394 	 * Clear any references to the page.  Otherwise, the page daemon will
4395 	 * immediately reactivate the page.
4396 	 */
4397 	vm_page_aflag_clear(m, PGA_REFERENCED);
4398 
4399 	/*
4400 	 * Place clean pages near the head of the inactive queue rather than
4401 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4402 	 * the page will be reused quickly.  Dirty pages not already in the
4403 	 * laundry are moved there.
4404 	 */
4405 	if (m->dirty == 0)
4406 		vm_page_deactivate_noreuse(m);
4407 	else if (!vm_page_in_laundry(m))
4408 		vm_page_launder(m);
4409 }
4410 
4411 /*
4412  *	vm_page_grab_release
4413  *
4414  *	Helper routine for grab functions to release busy on return.
4415  */
4416 static inline void
4417 vm_page_grab_release(vm_page_t m, int allocflags)
4418 {
4419 
4420 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4421 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4422 			vm_page_sunbusy(m);
4423 		else
4424 			vm_page_xunbusy(m);
4425 	}
4426 }
4427 
4428 /*
4429  *	vm_page_grab_sleep
4430  *
4431  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4432  *	if the caller should retry and false otherwise.
4433  *
4434  *	If the object is locked on entry the object will be unlocked with
4435  *	false returns and still locked but possibly having been dropped
4436  *	with true returns.
4437  */
4438 static bool
4439 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4440     const char *wmesg, int allocflags, bool locked)
4441 {
4442 
4443 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4444 		return (false);
4445 
4446 	/*
4447 	 * Reference the page before unlocking and sleeping so that
4448 	 * the page daemon is less likely to reclaim it.
4449 	 */
4450 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4451 		vm_page_reference(m);
4452 
4453 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4454 	    locked)
4455 		VM_OBJECT_WLOCK(object);
4456 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4457 		return (false);
4458 
4459 	return (true);
4460 }
4461 
4462 /*
4463  * Assert that the grab flags are valid.
4464  */
4465 static inline void
4466 vm_page_grab_check(int allocflags)
4467 {
4468 
4469 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4470 	    (allocflags & VM_ALLOC_WIRED) != 0,
4471 	    ("vm_page_grab*: the pages must be busied or wired"));
4472 
4473 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4474 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4475 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4476 }
4477 
4478 /*
4479  * Calculate the page allocation flags for grab.
4480  */
4481 static inline int
4482 vm_page_grab_pflags(int allocflags)
4483 {
4484 	int pflags;
4485 
4486 	pflags = allocflags &
4487 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4488 	    VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
4489 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4490 		pflags |= VM_ALLOC_WAITFAIL;
4491 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4492 		pflags |= VM_ALLOC_SBUSY;
4493 
4494 	return (pflags);
4495 }
4496 
4497 /*
4498  * Grab a page, waiting until we are waken up due to the page
4499  * changing state.  We keep on waiting, if the page continues
4500  * to be in the object.  If the page doesn't exist, first allocate it
4501  * and then conditionally zero it.
4502  *
4503  * This routine may sleep.
4504  *
4505  * The object must be locked on entry.  The lock will, however, be released
4506  * and reacquired if the routine sleeps.
4507  */
4508 vm_page_t
4509 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4510 {
4511 	vm_page_t m;
4512 
4513 	VM_OBJECT_ASSERT_WLOCKED(object);
4514 	vm_page_grab_check(allocflags);
4515 
4516 retrylookup:
4517 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4518 		if (!vm_page_tryacquire(m, allocflags)) {
4519 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4520 			    allocflags, true))
4521 				goto retrylookup;
4522 			return (NULL);
4523 		}
4524 		goto out;
4525 	}
4526 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4527 		return (NULL);
4528 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4529 	if (m == NULL) {
4530 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4531 			return (NULL);
4532 		goto retrylookup;
4533 	}
4534 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4535 		pmap_zero_page(m);
4536 
4537 out:
4538 	vm_page_grab_release(m, allocflags);
4539 
4540 	return (m);
4541 }
4542 
4543 /*
4544  * Locklessly attempt to acquire a page given a (object, pindex) tuple
4545  * and an optional previous page to avoid the radix lookup.  The resulting
4546  * page will be validated against the identity tuple and busied or wired
4547  * as requested.  A NULL *mp return guarantees that the page was not in
4548  * radix at the time of the call but callers must perform higher level
4549  * synchronization or retry the operation under a lock if they require
4550  * an atomic answer.  This is the only lock free validation routine,
4551  * other routines can depend on the resulting page state.
4552  *
4553  * The return value indicates whether the operation failed due to caller
4554  * flags.  The return is tri-state with mp:
4555  *
4556  * (true, *mp != NULL) - The operation was successful.
4557  * (true, *mp == NULL) - The page was not found in tree.
4558  * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4559  */
4560 static bool
4561 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4562     vm_page_t prev, vm_page_t *mp, int allocflags)
4563 {
4564 	vm_page_t m;
4565 
4566 	vm_page_grab_check(allocflags);
4567 	MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4568 
4569 	*mp = NULL;
4570 	for (;;) {
4571 		/*
4572 		 * We may see a false NULL here because the previous page
4573 		 * has been removed or just inserted and the list is loaded
4574 		 * without barriers.  Switch to radix to verify.
4575 		 */
4576 		if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4577 		    QMD_IS_TRASHED(m) || m->pindex != pindex ||
4578 		    atomic_load_ptr(&m->object) != object) {
4579 			prev = NULL;
4580 			/*
4581 			 * This guarantees the result is instantaneously
4582 			 * correct.
4583 			 */
4584 			m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4585 		}
4586 		if (m == NULL)
4587 			return (true);
4588 		if (vm_page_trybusy(m, allocflags)) {
4589 			if (m->object == object && m->pindex == pindex)
4590 				break;
4591 			/* relookup. */
4592 			vm_page_busy_release(m);
4593 			cpu_spinwait();
4594 			continue;
4595 		}
4596 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4597 		    allocflags, false))
4598 			return (false);
4599 	}
4600 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4601 		vm_page_wire(m);
4602 	vm_page_grab_release(m, allocflags);
4603 	*mp = m;
4604 	return (true);
4605 }
4606 
4607 /*
4608  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4609  * is not set.
4610  */
4611 vm_page_t
4612 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4613 {
4614 	vm_page_t m;
4615 
4616 	vm_page_grab_check(allocflags);
4617 
4618 	if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4619 		return (NULL);
4620 	if (m != NULL)
4621 		return (m);
4622 
4623 	/*
4624 	 * The radix lockless lookup should never return a false negative
4625 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4626 	 * was no page present at the instant of the call.  A NOCREAT caller
4627 	 * must handle create races gracefully.
4628 	 */
4629 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4630 		return (NULL);
4631 
4632 	VM_OBJECT_WLOCK(object);
4633 	m = vm_page_grab(object, pindex, allocflags);
4634 	VM_OBJECT_WUNLOCK(object);
4635 
4636 	return (m);
4637 }
4638 
4639 /*
4640  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4641  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4642  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4643  * in simultaneously.  Additional pages will be left on a paging queue but
4644  * will neither be wired nor busy regardless of allocflags.
4645  */
4646 int
4647 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4648 {
4649 	vm_page_t m;
4650 	vm_page_t ma[VM_INITIAL_PAGEIN];
4651 	int after, i, pflags, rv;
4652 
4653 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4654 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4655 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4656 	KASSERT((allocflags &
4657 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4658 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4659 	VM_OBJECT_ASSERT_WLOCKED(object);
4660 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4661 	    VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
4662 	pflags |= VM_ALLOC_WAITFAIL;
4663 
4664 retrylookup:
4665 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4666 		/*
4667 		 * If the page is fully valid it can only become invalid
4668 		 * with the object lock held.  If it is not valid it can
4669 		 * become valid with the busy lock held.  Therefore, we
4670 		 * may unnecessarily lock the exclusive busy here if we
4671 		 * race with I/O completion not using the object lock.
4672 		 * However, we will not end up with an invalid page and a
4673 		 * shared lock.
4674 		 */
4675 		if (!vm_page_trybusy(m,
4676 		    vm_page_all_valid(m) ? allocflags : 0)) {
4677 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4678 			    allocflags, true);
4679 			goto retrylookup;
4680 		}
4681 		if (vm_page_all_valid(m))
4682 			goto out;
4683 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4684 			vm_page_busy_release(m);
4685 			*mp = NULL;
4686 			return (VM_PAGER_FAIL);
4687 		}
4688 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4689 		*mp = NULL;
4690 		return (VM_PAGER_FAIL);
4691 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4692 		goto retrylookup;
4693 	}
4694 
4695 	vm_page_assert_xbusied(m);
4696 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4697 		after = MIN(after, VM_INITIAL_PAGEIN);
4698 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4699 		after = MAX(after, 1);
4700 		ma[0] = m;
4701 		for (i = 1; i < after; i++) {
4702 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4703 				if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4704 					break;
4705 			} else {
4706 				ma[i] = vm_page_alloc(object, m->pindex + i,
4707 				    VM_ALLOC_NORMAL);
4708 				if (ma[i] == NULL)
4709 					break;
4710 			}
4711 		}
4712 		after = i;
4713 		vm_object_pip_add(object, after);
4714 		VM_OBJECT_WUNLOCK(object);
4715 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4716 		VM_OBJECT_WLOCK(object);
4717 		vm_object_pip_wakeupn(object, after);
4718 		/* Pager may have replaced a page. */
4719 		m = ma[0];
4720 		if (rv != VM_PAGER_OK) {
4721 			for (i = 0; i < after; i++) {
4722 				if (!vm_page_wired(ma[i]))
4723 					vm_page_free(ma[i]);
4724 				else
4725 					vm_page_xunbusy(ma[i]);
4726 			}
4727 			*mp = NULL;
4728 			return (rv);
4729 		}
4730 		for (i = 1; i < after; i++)
4731 			vm_page_readahead_finish(ma[i]);
4732 		MPASS(vm_page_all_valid(m));
4733 	} else {
4734 		vm_page_zero_invalid(m, TRUE);
4735 	}
4736 out:
4737 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4738 		vm_page_wire(m);
4739 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4740 		vm_page_busy_downgrade(m);
4741 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4742 		vm_page_busy_release(m);
4743 	*mp = m;
4744 	return (VM_PAGER_OK);
4745 }
4746 
4747 /*
4748  * Locklessly grab a valid page.  If the page is not valid or not yet
4749  * allocated this will fall back to the object lock method.
4750  */
4751 int
4752 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4753     vm_pindex_t pindex, int allocflags)
4754 {
4755 	vm_page_t m;
4756 	int flags;
4757 	int error;
4758 
4759 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4760 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4761 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4762 	    "mismatch"));
4763 	KASSERT((allocflags &
4764 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4765 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4766 
4767 	/*
4768 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
4769 	 * before we can inspect the valid field and return a wired page.
4770 	 */
4771 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4772 	if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4773 		return (VM_PAGER_FAIL);
4774 	if ((m = *mp) != NULL) {
4775 		if (vm_page_all_valid(m)) {
4776 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4777 				vm_page_wire(m);
4778 			vm_page_grab_release(m, allocflags);
4779 			return (VM_PAGER_OK);
4780 		}
4781 		vm_page_busy_release(m);
4782 	}
4783 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4784 		*mp = NULL;
4785 		return (VM_PAGER_FAIL);
4786 	}
4787 	VM_OBJECT_WLOCK(object);
4788 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
4789 	VM_OBJECT_WUNLOCK(object);
4790 
4791 	return (error);
4792 }
4793 
4794 /*
4795  * Return the specified range of pages from the given object.  For each
4796  * page offset within the range, if a page already exists within the object
4797  * at that offset and it is busy, then wait for it to change state.  If,
4798  * instead, the page doesn't exist, then allocate it.
4799  *
4800  * The caller must always specify an allocation class.
4801  *
4802  * allocation classes:
4803  *	VM_ALLOC_NORMAL		normal process request
4804  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4805  *
4806  * The caller must always specify that the pages are to be busied and/or
4807  * wired.
4808  *
4809  * optional allocation flags:
4810  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4811  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4812  *	VM_ALLOC_NOWAIT		do not sleep
4813  *	VM_ALLOC_SBUSY		set page to sbusy state
4814  *	VM_ALLOC_WIRED		wire the pages
4815  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4816  *
4817  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4818  * may return a partial prefix of the requested range.
4819  */
4820 int
4821 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4822     vm_page_t *ma, int count)
4823 {
4824 	vm_page_t m, mpred;
4825 	int pflags;
4826 	int i;
4827 
4828 	VM_OBJECT_ASSERT_WLOCKED(object);
4829 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4830 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4831 	KASSERT(count > 0,
4832 	    ("vm_page_grab_pages: invalid page count %d", count));
4833 	vm_page_grab_check(allocflags);
4834 
4835 	pflags = vm_page_grab_pflags(allocflags);
4836 	i = 0;
4837 retrylookup:
4838 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4839 	if (m == NULL || m->pindex != pindex + i) {
4840 		mpred = m;
4841 		m = NULL;
4842 	} else
4843 		mpred = TAILQ_PREV(m, pglist, listq);
4844 	for (; i < count; i++) {
4845 		if (m != NULL) {
4846 			if (!vm_page_tryacquire(m, allocflags)) {
4847 				if (vm_page_grab_sleep(object, m, pindex + i,
4848 				    "grbmaw", allocflags, true))
4849 					goto retrylookup;
4850 				break;
4851 			}
4852 		} else {
4853 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4854 				break;
4855 			m = vm_page_alloc_after(object, pindex + i,
4856 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4857 			if (m == NULL) {
4858 				if ((allocflags & (VM_ALLOC_NOWAIT |
4859 				    VM_ALLOC_WAITFAIL)) != 0)
4860 					break;
4861 				goto retrylookup;
4862 			}
4863 		}
4864 		if (vm_page_none_valid(m) &&
4865 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4866 			if ((m->flags & PG_ZERO) == 0)
4867 				pmap_zero_page(m);
4868 			vm_page_valid(m);
4869 		}
4870 		vm_page_grab_release(m, allocflags);
4871 		ma[i] = mpred = m;
4872 		m = vm_page_next(m);
4873 	}
4874 	return (i);
4875 }
4876 
4877 /*
4878  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
4879  * and will fall back to the locked variant to handle allocation.
4880  */
4881 int
4882 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
4883     int allocflags, vm_page_t *ma, int count)
4884 {
4885 	vm_page_t m, pred;
4886 	int flags;
4887 	int i;
4888 
4889 	KASSERT(count > 0,
4890 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
4891 	vm_page_grab_check(allocflags);
4892 
4893 	/*
4894 	 * Modify flags for lockless acquire to hold the page until we
4895 	 * set it valid if necessary.
4896 	 */
4897 	flags = allocflags & ~VM_ALLOC_NOBUSY;
4898 	pred = NULL;
4899 	for (i = 0; i < count; i++, pindex++) {
4900 		if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
4901 			return (i);
4902 		if (m == NULL)
4903 			break;
4904 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
4905 			if ((m->flags & PG_ZERO) == 0)
4906 				pmap_zero_page(m);
4907 			vm_page_valid(m);
4908 		}
4909 		/* m will still be wired or busy according to flags. */
4910 		vm_page_grab_release(m, allocflags);
4911 		pred = ma[i] = m;
4912 	}
4913 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
4914 		return (i);
4915 	count -= i;
4916 	VM_OBJECT_WLOCK(object);
4917 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
4918 	VM_OBJECT_WUNLOCK(object);
4919 
4920 	return (i);
4921 }
4922 
4923 /*
4924  * Mapping function for valid or dirty bits in a page.
4925  *
4926  * Inputs are required to range within a page.
4927  */
4928 vm_page_bits_t
4929 vm_page_bits(int base, int size)
4930 {
4931 	int first_bit;
4932 	int last_bit;
4933 
4934 	KASSERT(
4935 	    base + size <= PAGE_SIZE,
4936 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4937 	);
4938 
4939 	if (size == 0)		/* handle degenerate case */
4940 		return (0);
4941 
4942 	first_bit = base >> DEV_BSHIFT;
4943 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4944 
4945 	return (((vm_page_bits_t)2 << last_bit) -
4946 	    ((vm_page_bits_t)1 << first_bit));
4947 }
4948 
4949 void
4950 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4951 {
4952 
4953 #if PAGE_SIZE == 32768
4954 	atomic_set_64((uint64_t *)bits, set);
4955 #elif PAGE_SIZE == 16384
4956 	atomic_set_32((uint32_t *)bits, set);
4957 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4958 	atomic_set_16((uint16_t *)bits, set);
4959 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4960 	atomic_set_8((uint8_t *)bits, set);
4961 #else		/* PAGE_SIZE <= 8192 */
4962 	uintptr_t addr;
4963 	int shift;
4964 
4965 	addr = (uintptr_t)bits;
4966 	/*
4967 	 * Use a trick to perform a 32-bit atomic on the
4968 	 * containing aligned word, to not depend on the existence
4969 	 * of atomic_{set, clear}_{8, 16}.
4970 	 */
4971 	shift = addr & (sizeof(uint32_t) - 1);
4972 #if BYTE_ORDER == BIG_ENDIAN
4973 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4974 #else
4975 	shift *= NBBY;
4976 #endif
4977 	addr &= ~(sizeof(uint32_t) - 1);
4978 	atomic_set_32((uint32_t *)addr, set << shift);
4979 #endif		/* PAGE_SIZE */
4980 }
4981 
4982 static inline void
4983 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4984 {
4985 
4986 #if PAGE_SIZE == 32768
4987 	atomic_clear_64((uint64_t *)bits, clear);
4988 #elif PAGE_SIZE == 16384
4989 	atomic_clear_32((uint32_t *)bits, clear);
4990 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4991 	atomic_clear_16((uint16_t *)bits, clear);
4992 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4993 	atomic_clear_8((uint8_t *)bits, clear);
4994 #else		/* PAGE_SIZE <= 8192 */
4995 	uintptr_t addr;
4996 	int shift;
4997 
4998 	addr = (uintptr_t)bits;
4999 	/*
5000 	 * Use a trick to perform a 32-bit atomic on the
5001 	 * containing aligned word, to not depend on the existence
5002 	 * of atomic_{set, clear}_{8, 16}.
5003 	 */
5004 	shift = addr & (sizeof(uint32_t) - 1);
5005 #if BYTE_ORDER == BIG_ENDIAN
5006 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5007 #else
5008 	shift *= NBBY;
5009 #endif
5010 	addr &= ~(sizeof(uint32_t) - 1);
5011 	atomic_clear_32((uint32_t *)addr, clear << shift);
5012 #endif		/* PAGE_SIZE */
5013 }
5014 
5015 static inline vm_page_bits_t
5016 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
5017 {
5018 #if PAGE_SIZE == 32768
5019 	uint64_t old;
5020 
5021 	old = *bits;
5022 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
5023 	return (old);
5024 #elif PAGE_SIZE == 16384
5025 	uint32_t old;
5026 
5027 	old = *bits;
5028 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
5029 	return (old);
5030 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
5031 	uint16_t old;
5032 
5033 	old = *bits;
5034 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
5035 	return (old);
5036 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
5037 	uint8_t old;
5038 
5039 	old = *bits;
5040 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
5041 	return (old);
5042 #else		/* PAGE_SIZE <= 4096*/
5043 	uintptr_t addr;
5044 	uint32_t old, new, mask;
5045 	int shift;
5046 
5047 	addr = (uintptr_t)bits;
5048 	/*
5049 	 * Use a trick to perform a 32-bit atomic on the
5050 	 * containing aligned word, to not depend on the existence
5051 	 * of atomic_{set, swap, clear}_{8, 16}.
5052 	 */
5053 	shift = addr & (sizeof(uint32_t) - 1);
5054 #if BYTE_ORDER == BIG_ENDIAN
5055 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
5056 #else
5057 	shift *= NBBY;
5058 #endif
5059 	addr &= ~(sizeof(uint32_t) - 1);
5060 	mask = VM_PAGE_BITS_ALL << shift;
5061 
5062 	old = *bits;
5063 	do {
5064 		new = old & ~mask;
5065 		new |= newbits << shift;
5066 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5067 	return (old >> shift);
5068 #endif		/* PAGE_SIZE */
5069 }
5070 
5071 /*
5072  *	vm_page_set_valid_range:
5073  *
5074  *	Sets portions of a page valid.  The arguments are expected
5075  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5076  *	of any partial chunks touched by the range.  The invalid portion of
5077  *	such chunks will be zeroed.
5078  *
5079  *	(base + size) must be less then or equal to PAGE_SIZE.
5080  */
5081 void
5082 vm_page_set_valid_range(vm_page_t m, int base, int size)
5083 {
5084 	int endoff, frag;
5085 	vm_page_bits_t pagebits;
5086 
5087 	vm_page_assert_busied(m);
5088 	if (size == 0)	/* handle degenerate case */
5089 		return;
5090 
5091 	/*
5092 	 * If the base is not DEV_BSIZE aligned and the valid
5093 	 * bit is clear, we have to zero out a portion of the
5094 	 * first block.
5095 	 */
5096 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5097 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5098 		pmap_zero_page_area(m, frag, base - frag);
5099 
5100 	/*
5101 	 * If the ending offset is not DEV_BSIZE aligned and the
5102 	 * valid bit is clear, we have to zero out a portion of
5103 	 * the last block.
5104 	 */
5105 	endoff = base + size;
5106 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5107 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5108 		pmap_zero_page_area(m, endoff,
5109 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5110 
5111 	/*
5112 	 * Assert that no previously invalid block that is now being validated
5113 	 * is already dirty.
5114 	 */
5115 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5116 	    ("vm_page_set_valid_range: page %p is dirty", m));
5117 
5118 	/*
5119 	 * Set valid bits inclusive of any overlap.
5120 	 */
5121 	pagebits = vm_page_bits(base, size);
5122 	if (vm_page_xbusied(m))
5123 		m->valid |= pagebits;
5124 	else
5125 		vm_page_bits_set(m, &m->valid, pagebits);
5126 }
5127 
5128 /*
5129  * Set the page dirty bits and free the invalid swap space if
5130  * present.  Returns the previous dirty bits.
5131  */
5132 vm_page_bits_t
5133 vm_page_set_dirty(vm_page_t m)
5134 {
5135 	vm_page_bits_t old;
5136 
5137 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5138 
5139 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5140 		old = m->dirty;
5141 		m->dirty = VM_PAGE_BITS_ALL;
5142 	} else
5143 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5144 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5145 		vm_pager_page_unswapped(m);
5146 
5147 	return (old);
5148 }
5149 
5150 /*
5151  * Clear the given bits from the specified page's dirty field.
5152  */
5153 static __inline void
5154 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5155 {
5156 
5157 	vm_page_assert_busied(m);
5158 
5159 	/*
5160 	 * If the page is xbusied and not write mapped we are the
5161 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5162 	 * layer can call vm_page_dirty() without holding a distinguished
5163 	 * lock.  The combination of page busy and atomic operations
5164 	 * suffice to guarantee consistency of the page dirty field.
5165 	 */
5166 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5167 		m->dirty &= ~pagebits;
5168 	else
5169 		vm_page_bits_clear(m, &m->dirty, pagebits);
5170 }
5171 
5172 /*
5173  *	vm_page_set_validclean:
5174  *
5175  *	Sets portions of a page valid and clean.  The arguments are expected
5176  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5177  *	of any partial chunks touched by the range.  The invalid portion of
5178  *	such chunks will be zero'd.
5179  *
5180  *	(base + size) must be less then or equal to PAGE_SIZE.
5181  */
5182 void
5183 vm_page_set_validclean(vm_page_t m, int base, int size)
5184 {
5185 	vm_page_bits_t oldvalid, pagebits;
5186 	int endoff, frag;
5187 
5188 	vm_page_assert_busied(m);
5189 	if (size == 0)	/* handle degenerate case */
5190 		return;
5191 
5192 	/*
5193 	 * If the base is not DEV_BSIZE aligned and the valid
5194 	 * bit is clear, we have to zero out a portion of the
5195 	 * first block.
5196 	 */
5197 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5198 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5199 		pmap_zero_page_area(m, frag, base - frag);
5200 
5201 	/*
5202 	 * If the ending offset is not DEV_BSIZE aligned and the
5203 	 * valid bit is clear, we have to zero out a portion of
5204 	 * the last block.
5205 	 */
5206 	endoff = base + size;
5207 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5208 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5209 		pmap_zero_page_area(m, endoff,
5210 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5211 
5212 	/*
5213 	 * Set valid, clear dirty bits.  If validating the entire
5214 	 * page we can safely clear the pmap modify bit.  We also
5215 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5216 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5217 	 * be set again.
5218 	 *
5219 	 * We set valid bits inclusive of any overlap, but we can only
5220 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5221 	 * the range.
5222 	 */
5223 	oldvalid = m->valid;
5224 	pagebits = vm_page_bits(base, size);
5225 	if (vm_page_xbusied(m))
5226 		m->valid |= pagebits;
5227 	else
5228 		vm_page_bits_set(m, &m->valid, pagebits);
5229 #if 0	/* NOT YET */
5230 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5231 		frag = DEV_BSIZE - frag;
5232 		base += frag;
5233 		size -= frag;
5234 		if (size < 0)
5235 			size = 0;
5236 	}
5237 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5238 #endif
5239 	if (base == 0 && size == PAGE_SIZE) {
5240 		/*
5241 		 * The page can only be modified within the pmap if it is
5242 		 * mapped, and it can only be mapped if it was previously
5243 		 * fully valid.
5244 		 */
5245 		if (oldvalid == VM_PAGE_BITS_ALL)
5246 			/*
5247 			 * Perform the pmap_clear_modify() first.  Otherwise,
5248 			 * a concurrent pmap operation, such as
5249 			 * pmap_protect(), could clear a modification in the
5250 			 * pmap and set the dirty field on the page before
5251 			 * pmap_clear_modify() had begun and after the dirty
5252 			 * field was cleared here.
5253 			 */
5254 			pmap_clear_modify(m);
5255 		m->dirty = 0;
5256 		vm_page_aflag_clear(m, PGA_NOSYNC);
5257 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5258 		m->dirty &= ~pagebits;
5259 	else
5260 		vm_page_clear_dirty_mask(m, pagebits);
5261 }
5262 
5263 void
5264 vm_page_clear_dirty(vm_page_t m, int base, int size)
5265 {
5266 
5267 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5268 }
5269 
5270 /*
5271  *	vm_page_set_invalid:
5272  *
5273  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5274  *	valid and dirty bits for the effected areas are cleared.
5275  */
5276 void
5277 vm_page_set_invalid(vm_page_t m, int base, int size)
5278 {
5279 	vm_page_bits_t bits;
5280 	vm_object_t object;
5281 
5282 	/*
5283 	 * The object lock is required so that pages can't be mapped
5284 	 * read-only while we're in the process of invalidating them.
5285 	 */
5286 	object = m->object;
5287 	VM_OBJECT_ASSERT_WLOCKED(object);
5288 	vm_page_assert_busied(m);
5289 
5290 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5291 	    size >= object->un_pager.vnp.vnp_size)
5292 		bits = VM_PAGE_BITS_ALL;
5293 	else
5294 		bits = vm_page_bits(base, size);
5295 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5296 		pmap_remove_all(m);
5297 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5298 	    !pmap_page_is_mapped(m),
5299 	    ("vm_page_set_invalid: page %p is mapped", m));
5300 	if (vm_page_xbusied(m)) {
5301 		m->valid &= ~bits;
5302 		m->dirty &= ~bits;
5303 	} else {
5304 		vm_page_bits_clear(m, &m->valid, bits);
5305 		vm_page_bits_clear(m, &m->dirty, bits);
5306 	}
5307 }
5308 
5309 /*
5310  *	vm_page_invalid:
5311  *
5312  *	Invalidates the entire page.  The page must be busy, unmapped, and
5313  *	the enclosing object must be locked.  The object locks protects
5314  *	against concurrent read-only pmap enter which is done without
5315  *	busy.
5316  */
5317 void
5318 vm_page_invalid(vm_page_t m)
5319 {
5320 
5321 	vm_page_assert_busied(m);
5322 	VM_OBJECT_ASSERT_WLOCKED(m->object);
5323 	MPASS(!pmap_page_is_mapped(m));
5324 
5325 	if (vm_page_xbusied(m))
5326 		m->valid = 0;
5327 	else
5328 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5329 }
5330 
5331 /*
5332  * vm_page_zero_invalid()
5333  *
5334  *	The kernel assumes that the invalid portions of a page contain
5335  *	garbage, but such pages can be mapped into memory by user code.
5336  *	When this occurs, we must zero out the non-valid portions of the
5337  *	page so user code sees what it expects.
5338  *
5339  *	Pages are most often semi-valid when the end of a file is mapped
5340  *	into memory and the file's size is not page aligned.
5341  */
5342 void
5343 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5344 {
5345 	int b;
5346 	int i;
5347 
5348 	/*
5349 	 * Scan the valid bits looking for invalid sections that
5350 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5351 	 * valid bit may be set ) have already been zeroed by
5352 	 * vm_page_set_validclean().
5353 	 */
5354 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5355 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5356 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5357 			if (i > b) {
5358 				pmap_zero_page_area(m,
5359 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5360 			}
5361 			b = i + 1;
5362 		}
5363 	}
5364 
5365 	/*
5366 	 * setvalid is TRUE when we can safely set the zero'd areas
5367 	 * as being valid.  We can do this if there are no cache consistancy
5368 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5369 	 */
5370 	if (setvalid)
5371 		vm_page_valid(m);
5372 }
5373 
5374 /*
5375  *	vm_page_is_valid:
5376  *
5377  *	Is (partial) page valid?  Note that the case where size == 0
5378  *	will return FALSE in the degenerate case where the page is
5379  *	entirely invalid, and TRUE otherwise.
5380  *
5381  *	Some callers envoke this routine without the busy lock held and
5382  *	handle races via higher level locks.  Typical callers should
5383  *	hold a busy lock to prevent invalidation.
5384  */
5385 int
5386 vm_page_is_valid(vm_page_t m, int base, int size)
5387 {
5388 	vm_page_bits_t bits;
5389 
5390 	bits = vm_page_bits(base, size);
5391 	return (m->valid != 0 && (m->valid & bits) == bits);
5392 }
5393 
5394 /*
5395  * Returns true if all of the specified predicates are true for the entire
5396  * (super)page and false otherwise.
5397  */
5398 bool
5399 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5400 {
5401 	vm_object_t object;
5402 	int i, npages;
5403 
5404 	object = m->object;
5405 	if (skip_m != NULL && skip_m->object != object)
5406 		return (false);
5407 	VM_OBJECT_ASSERT_LOCKED(object);
5408 	npages = atop(pagesizes[m->psind]);
5409 
5410 	/*
5411 	 * The physically contiguous pages that make up a superpage, i.e., a
5412 	 * page with a page size index ("psind") greater than zero, will
5413 	 * occupy adjacent entries in vm_page_array[].
5414 	 */
5415 	for (i = 0; i < npages; i++) {
5416 		/* Always test object consistency, including "skip_m". */
5417 		if (m[i].object != object)
5418 			return (false);
5419 		if (&m[i] == skip_m)
5420 			continue;
5421 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5422 			return (false);
5423 		if ((flags & PS_ALL_DIRTY) != 0) {
5424 			/*
5425 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5426 			 * might stop this case from spuriously returning
5427 			 * "false".  However, that would require a write lock
5428 			 * on the object containing "m[i]".
5429 			 */
5430 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5431 				return (false);
5432 		}
5433 		if ((flags & PS_ALL_VALID) != 0 &&
5434 		    m[i].valid != VM_PAGE_BITS_ALL)
5435 			return (false);
5436 	}
5437 	return (true);
5438 }
5439 
5440 /*
5441  * Set the page's dirty bits if the page is modified.
5442  */
5443 void
5444 vm_page_test_dirty(vm_page_t m)
5445 {
5446 
5447 	vm_page_assert_busied(m);
5448 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5449 		vm_page_dirty(m);
5450 }
5451 
5452 void
5453 vm_page_valid(vm_page_t m)
5454 {
5455 
5456 	vm_page_assert_busied(m);
5457 	if (vm_page_xbusied(m))
5458 		m->valid = VM_PAGE_BITS_ALL;
5459 	else
5460 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5461 }
5462 
5463 void
5464 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5465 {
5466 
5467 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5468 }
5469 
5470 void
5471 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5472 {
5473 
5474 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5475 }
5476 
5477 int
5478 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5479 {
5480 
5481 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5482 }
5483 
5484 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5485 void
5486 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5487 {
5488 
5489 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5490 }
5491 
5492 void
5493 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5494 {
5495 
5496 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5497 }
5498 #endif
5499 
5500 #ifdef INVARIANTS
5501 void
5502 vm_page_object_busy_assert(vm_page_t m)
5503 {
5504 
5505 	/*
5506 	 * Certain of the page's fields may only be modified by the
5507 	 * holder of a page or object busy.
5508 	 */
5509 	if (m->object != NULL && !vm_page_busied(m))
5510 		VM_OBJECT_ASSERT_BUSY(m->object);
5511 }
5512 
5513 void
5514 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5515 {
5516 
5517 	if ((bits & PGA_WRITEABLE) == 0)
5518 		return;
5519 
5520 	/*
5521 	 * The PGA_WRITEABLE flag can only be set if the page is
5522 	 * managed, is exclusively busied or the object is locked.
5523 	 * Currently, this flag is only set by pmap_enter().
5524 	 */
5525 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5526 	    ("PGA_WRITEABLE on unmanaged page"));
5527 	if (!vm_page_xbusied(m))
5528 		VM_OBJECT_ASSERT_BUSY(m->object);
5529 }
5530 #endif
5531 
5532 #include "opt_ddb.h"
5533 #ifdef DDB
5534 #include <sys/kernel.h>
5535 
5536 #include <ddb/ddb.h>
5537 
5538 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5539 {
5540 
5541 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5542 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5543 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5544 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5545 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5546 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5547 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5548 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5549 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5550 }
5551 
5552 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5553 {
5554 	int dom;
5555 
5556 	db_printf("pq_free %d\n", vm_free_count());
5557 	for (dom = 0; dom < vm_ndomains; dom++) {
5558 		db_printf(
5559     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5560 		    dom,
5561 		    vm_dom[dom].vmd_page_count,
5562 		    vm_dom[dom].vmd_free_count,
5563 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5564 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5565 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5566 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5567 	}
5568 }
5569 
5570 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5571 {
5572 	vm_page_t m;
5573 	boolean_t phys, virt;
5574 
5575 	if (!have_addr) {
5576 		db_printf("show pginfo addr\n");
5577 		return;
5578 	}
5579 
5580 	phys = strchr(modif, 'p') != NULL;
5581 	virt = strchr(modif, 'v') != NULL;
5582 	if (virt)
5583 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5584 	else if (phys)
5585 		m = PHYS_TO_VM_PAGE(addr);
5586 	else
5587 		m = (vm_page_t)addr;
5588 	db_printf(
5589     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5590     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5591 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5592 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5593 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5594 }
5595 #endif /* DDB */
5596