xref: /freebsd/sys/vm/vm_page.c (revision 0321a7990b277702fa0b4f8366121bf53d03cb64)
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(NULL, 0, VM_ALLOC_NOOBJ |
201 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
202 }
203 
204 /*
205  * The cache page zone is initialized later since we need to be able to allocate
206  * pages before UMA is fully initialized.
207  */
208 static void
209 vm_page_init_cache_zones(void *dummy __unused)
210 {
211 	struct vm_domain *vmd;
212 	struct vm_pgcache *pgcache;
213 	int cache, domain, maxcache, pool;
214 
215 	maxcache = 0;
216 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
217 	maxcache *= mp_ncpus;
218 	for (domain = 0; domain < vm_ndomains; domain++) {
219 		vmd = VM_DOMAIN(domain);
220 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
221 			pgcache = &vmd->vmd_pgcache[pool];
222 			pgcache->domain = domain;
223 			pgcache->pool = pool;
224 			pgcache->zone = uma_zcache_create("vm pgcache",
225 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
226 			    vm_page_zone_import, vm_page_zone_release, pgcache,
227 			    UMA_ZONE_VM);
228 
229 			/*
230 			 * Limit each pool's zone to 0.1% of the pages in the
231 			 * domain.
232 			 */
233 			cache = maxcache != 0 ? maxcache :
234 			    vmd->vmd_page_count / 1000;
235 			uma_zone_set_maxcache(pgcache->zone, cache);
236 		}
237 	}
238 }
239 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
240 
241 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
242 #if PAGE_SIZE == 32768
243 #ifdef CTASSERT
244 CTASSERT(sizeof(u_long) >= 8);
245 #endif
246 #endif
247 
248 /*
249  *	vm_set_page_size:
250  *
251  *	Sets the page size, perhaps based upon the memory
252  *	size.  Must be called before any use of page-size
253  *	dependent functions.
254  */
255 void
256 vm_set_page_size(void)
257 {
258 	if (vm_cnt.v_page_size == 0)
259 		vm_cnt.v_page_size = PAGE_SIZE;
260 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
261 		panic("vm_set_page_size: page size not a power of two");
262 }
263 
264 /*
265  *	vm_page_blacklist_next:
266  *
267  *	Find the next entry in the provided string of blacklist
268  *	addresses.  Entries are separated by space, comma, or newline.
269  *	If an invalid integer is encountered then the rest of the
270  *	string is skipped.  Updates the list pointer to the next
271  *	character, or NULL if the string is exhausted or invalid.
272  */
273 static vm_paddr_t
274 vm_page_blacklist_next(char **list, char *end)
275 {
276 	vm_paddr_t bad;
277 	char *cp, *pos;
278 
279 	if (list == NULL || *list == NULL)
280 		return (0);
281 	if (**list =='\0') {
282 		*list = NULL;
283 		return (0);
284 	}
285 
286 	/*
287 	 * If there's no end pointer then the buffer is coming from
288 	 * the kenv and we know it's null-terminated.
289 	 */
290 	if (end == NULL)
291 		end = *list + strlen(*list);
292 
293 	/* Ensure that strtoq() won't walk off the end */
294 	if (*end != '\0') {
295 		if (*end == '\n' || *end == ' ' || *end  == ',')
296 			*end = '\0';
297 		else {
298 			printf("Blacklist not terminated, skipping\n");
299 			*list = NULL;
300 			return (0);
301 		}
302 	}
303 
304 	for (pos = *list; *pos != '\0'; pos = cp) {
305 		bad = strtoq(pos, &cp, 0);
306 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
307 			if (bad == 0) {
308 				if (++cp < end)
309 					continue;
310 				else
311 					break;
312 			}
313 		} else
314 			break;
315 		if (*cp == '\0' || ++cp >= end)
316 			*list = NULL;
317 		else
318 			*list = cp;
319 		return (trunc_page(bad));
320 	}
321 	printf("Garbage in RAM blacklist, skipping\n");
322 	*list = NULL;
323 	return (0);
324 }
325 
326 bool
327 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
328 {
329 	struct vm_domain *vmd;
330 	vm_page_t m;
331 	int ret;
332 
333 	m = vm_phys_paddr_to_vm_page(pa);
334 	if (m == NULL)
335 		return (true); /* page does not exist, no failure */
336 
337 	vmd = vm_pagequeue_domain(m);
338 	vm_domain_free_lock(vmd);
339 	ret = vm_phys_unfree_page(m);
340 	vm_domain_free_unlock(vmd);
341 	if (ret != 0) {
342 		vm_domain_freecnt_inc(vmd, -1);
343 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
344 		if (verbose)
345 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
346 	}
347 	return (ret);
348 }
349 
350 /*
351  *	vm_page_blacklist_check:
352  *
353  *	Iterate through the provided string of blacklist addresses, pulling
354  *	each entry out of the physical allocator free list and putting it
355  *	onto a list for reporting via the vm.page_blacklist sysctl.
356  */
357 static void
358 vm_page_blacklist_check(char *list, char *end)
359 {
360 	vm_paddr_t pa;
361 	char *next;
362 
363 	next = list;
364 	while (next != NULL) {
365 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
366 			continue;
367 		vm_page_blacklist_add(pa, bootverbose);
368 	}
369 }
370 
371 /*
372  *	vm_page_blacklist_load:
373  *
374  *	Search for a special module named "ram_blacklist".  It'll be a
375  *	plain text file provided by the user via the loader directive
376  *	of the same name.
377  */
378 static void
379 vm_page_blacklist_load(char **list, char **end)
380 {
381 	void *mod;
382 	u_char *ptr;
383 	u_int len;
384 
385 	mod = NULL;
386 	ptr = NULL;
387 
388 	mod = preload_search_by_type("ram_blacklist");
389 	if (mod != NULL) {
390 		ptr = preload_fetch_addr(mod);
391 		len = preload_fetch_size(mod);
392         }
393 	*list = ptr;
394 	if (ptr != NULL)
395 		*end = ptr + len;
396 	else
397 		*end = NULL;
398 	return;
399 }
400 
401 static int
402 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
403 {
404 	vm_page_t m;
405 	struct sbuf sbuf;
406 	int error, first;
407 
408 	first = 1;
409 	error = sysctl_wire_old_buffer(req, 0);
410 	if (error != 0)
411 		return (error);
412 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
413 	TAILQ_FOREACH(m, &blacklist_head, listq) {
414 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
415 		    (uintmax_t)m->phys_addr);
416 		first = 0;
417 	}
418 	error = sbuf_finish(&sbuf);
419 	sbuf_delete(&sbuf);
420 	return (error);
421 }
422 
423 /*
424  * Initialize a dummy page for use in scans of the specified paging queue.
425  * In principle, this function only needs to set the flag PG_MARKER.
426  * Nonetheless, it write busies the page as a safety precaution.
427  */
428 void
429 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
430 {
431 
432 	bzero(marker, sizeof(*marker));
433 	marker->flags = PG_MARKER;
434 	marker->a.flags = aflags;
435 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
436 	marker->a.queue = queue;
437 }
438 
439 static void
440 vm_page_domain_init(int domain)
441 {
442 	struct vm_domain *vmd;
443 	struct vm_pagequeue *pq;
444 	int i;
445 
446 	vmd = VM_DOMAIN(domain);
447 	bzero(vmd, sizeof(*vmd));
448 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
449 	    "vm inactive pagequeue";
450 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
451 	    "vm active pagequeue";
452 	*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
453 	    "vm laundry pagequeue";
454 	*__DECONST(const char **,
455 	    &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
456 	    "vm unswappable pagequeue";
457 	vmd->vmd_domain = domain;
458 	vmd->vmd_page_count = 0;
459 	vmd->vmd_free_count = 0;
460 	vmd->vmd_segs = 0;
461 	vmd->vmd_oom = FALSE;
462 	for (i = 0; i < PQ_COUNT; i++) {
463 		pq = &vmd->vmd_pagequeues[i];
464 		TAILQ_INIT(&pq->pq_pl);
465 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
466 		    MTX_DEF | MTX_DUPOK);
467 		pq->pq_pdpages = 0;
468 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
469 	}
470 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
471 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
472 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
473 
474 	/*
475 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
476 	 * insertions.
477 	 */
478 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
479 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
480 	    &vmd->vmd_inacthead, plinks.q);
481 
482 	/*
483 	 * The clock pages are used to implement active queue scanning without
484 	 * requeues.  Scans start at clock[0], which is advanced after the scan
485 	 * ends.  When the two clock hands meet, they are reset and scanning
486 	 * resumes from the head of the queue.
487 	 */
488 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
489 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
490 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
491 	    &vmd->vmd_clock[0], plinks.q);
492 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
493 	    &vmd->vmd_clock[1], plinks.q);
494 }
495 
496 /*
497  * Initialize a physical page in preparation for adding it to the free
498  * lists.
499  */
500 void
501 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
502 {
503 
504 	m->object = NULL;
505 	m->ref_count = 0;
506 	m->busy_lock = VPB_FREED;
507 	m->flags = m->a.flags = 0;
508 	m->phys_addr = pa;
509 	m->a.queue = PQ_NONE;
510 	m->psind = 0;
511 	m->segind = segind;
512 	m->order = VM_NFREEORDER;
513 	m->pool = VM_FREEPOOL_DEFAULT;
514 	m->valid = m->dirty = 0;
515 	pmap_page_init(m);
516 }
517 
518 #ifndef PMAP_HAS_PAGE_ARRAY
519 static vm_paddr_t
520 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
521 {
522 	vm_paddr_t new_end;
523 
524 	/*
525 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
526 	 * However, because this page is allocated from KVM, out-of-bounds
527 	 * accesses using the direct map will not be trapped.
528 	 */
529 	*vaddr += PAGE_SIZE;
530 
531 	/*
532 	 * Allocate physical memory for the page structures, and map it.
533 	 */
534 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
535 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
536 	    VM_PROT_READ | VM_PROT_WRITE);
537 	vm_page_array_size = page_range;
538 
539 	return (new_end);
540 }
541 #endif
542 
543 /*
544  *	vm_page_startup:
545  *
546  *	Initializes the resident memory module.  Allocates physical memory for
547  *	bootstrapping UMA and some data structures that are used to manage
548  *	physical pages.  Initializes these structures, and populates the free
549  *	page queues.
550  */
551 vm_offset_t
552 vm_page_startup(vm_offset_t vaddr)
553 {
554 	struct vm_phys_seg *seg;
555 	struct vm_domain *vmd;
556 	vm_page_t m;
557 	char *list, *listend;
558 	vm_paddr_t end, high_avail, low_avail, new_end, size;
559 	vm_paddr_t page_range __unused;
560 	vm_paddr_t last_pa, pa, startp, endp;
561 	u_long pagecount;
562 #if MINIDUMP_PAGE_TRACKING
563 	u_long vm_page_dump_size;
564 #endif
565 	int biggestone, i, segind;
566 #ifdef WITNESS
567 	vm_offset_t mapped;
568 	int witness_size;
569 #endif
570 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
571 	long ii;
572 #endif
573 
574 	vaddr = round_page(vaddr);
575 
576 	vm_phys_early_startup();
577 	biggestone = vm_phys_avail_largest();
578 	end = phys_avail[biggestone+1];
579 
580 	/*
581 	 * Initialize the page and queue locks.
582 	 */
583 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
584 	for (i = 0; i < PA_LOCK_COUNT; i++)
585 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
586 	for (i = 0; i < vm_ndomains; i++)
587 		vm_page_domain_init(i);
588 
589 	new_end = end;
590 #ifdef WITNESS
591 	witness_size = round_page(witness_startup_count());
592 	new_end -= witness_size;
593 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
594 	    VM_PROT_READ | VM_PROT_WRITE);
595 	bzero((void *)mapped, witness_size);
596 	witness_startup((void *)mapped);
597 #endif
598 
599 #if MINIDUMP_PAGE_TRACKING
600 	/*
601 	 * Allocate a bitmap to indicate that a random physical page
602 	 * needs to be included in a minidump.
603 	 *
604 	 * The amd64 port needs this to indicate which direct map pages
605 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
606 	 *
607 	 * However, i386 still needs this workspace internally within the
608 	 * minidump code.  In theory, they are not needed on i386, but are
609 	 * included should the sf_buf code decide to use them.
610 	 */
611 	last_pa = 0;
612 	vm_page_dump_pages = 0;
613 	for (i = 0; dump_avail[i + 1] != 0; i += 2) {
614 		vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
615 		    dump_avail[i] / PAGE_SIZE;
616 		if (dump_avail[i + 1] > last_pa)
617 			last_pa = dump_avail[i + 1];
618 	}
619 	vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
620 	new_end -= vm_page_dump_size;
621 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
622 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
623 	bzero((void *)vm_page_dump, vm_page_dump_size);
624 #else
625 	(void)last_pa;
626 #endif
627 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
628     defined(__riscv) || defined(__powerpc64__)
629 	/*
630 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
631 	 * in a crash dump.  When pmap_map() uses the direct map, they are
632 	 * not automatically included.
633 	 */
634 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
635 		dump_add_page(pa);
636 #endif
637 	phys_avail[biggestone + 1] = new_end;
638 #ifdef __amd64__
639 	/*
640 	 * Request that the physical pages underlying the message buffer be
641 	 * included in a crash dump.  Since the message buffer is accessed
642 	 * through the direct map, they are not automatically included.
643 	 */
644 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
645 	last_pa = pa + round_page(msgbufsize);
646 	while (pa < last_pa) {
647 		dump_add_page(pa);
648 		pa += PAGE_SIZE;
649 	}
650 #endif
651 	/*
652 	 * Compute the number of pages of memory that will be available for
653 	 * use, taking into account the overhead of a page structure per page.
654 	 * In other words, solve
655 	 *	"available physical memory" - round_page(page_range *
656 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
657 	 * for page_range.
658 	 */
659 	low_avail = phys_avail[0];
660 	high_avail = phys_avail[1];
661 	for (i = 0; i < vm_phys_nsegs; i++) {
662 		if (vm_phys_segs[i].start < low_avail)
663 			low_avail = vm_phys_segs[i].start;
664 		if (vm_phys_segs[i].end > high_avail)
665 			high_avail = vm_phys_segs[i].end;
666 	}
667 	/* Skip the first chunk.  It is already accounted for. */
668 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
669 		if (phys_avail[i] < low_avail)
670 			low_avail = phys_avail[i];
671 		if (phys_avail[i + 1] > high_avail)
672 			high_avail = phys_avail[i + 1];
673 	}
674 	first_page = low_avail / PAGE_SIZE;
675 #ifdef VM_PHYSSEG_SPARSE
676 	size = 0;
677 	for (i = 0; i < vm_phys_nsegs; i++)
678 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
679 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
680 		size += phys_avail[i + 1] - phys_avail[i];
681 #elif defined(VM_PHYSSEG_DENSE)
682 	size = high_avail - low_avail;
683 #else
684 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
685 #endif
686 
687 #ifdef PMAP_HAS_PAGE_ARRAY
688 	pmap_page_array_startup(size / PAGE_SIZE);
689 	biggestone = vm_phys_avail_largest();
690 	end = new_end = phys_avail[biggestone + 1];
691 #else
692 #ifdef VM_PHYSSEG_DENSE
693 	/*
694 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
695 	 * the overhead of a page structure per page only if vm_page_array is
696 	 * allocated from the last physical memory chunk.  Otherwise, we must
697 	 * allocate page structures representing the physical memory
698 	 * underlying vm_page_array, even though they will not be used.
699 	 */
700 	if (new_end != high_avail)
701 		page_range = size / PAGE_SIZE;
702 	else
703 #endif
704 	{
705 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
706 
707 		/*
708 		 * If the partial bytes remaining are large enough for
709 		 * a page (PAGE_SIZE) without a corresponding
710 		 * 'struct vm_page', then new_end will contain an
711 		 * extra page after subtracting the length of the VM
712 		 * page array.  Compensate by subtracting an extra
713 		 * page from new_end.
714 		 */
715 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
716 			if (new_end == high_avail)
717 				high_avail -= PAGE_SIZE;
718 			new_end -= PAGE_SIZE;
719 		}
720 	}
721 	end = new_end;
722 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
723 #endif
724 
725 #if VM_NRESERVLEVEL > 0
726 	/*
727 	 * Allocate physical memory for the reservation management system's
728 	 * data structures, and map it.
729 	 */
730 	new_end = vm_reserv_startup(&vaddr, new_end);
731 #endif
732 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
733     defined(__riscv) || defined(__powerpc64__)
734 	/*
735 	 * Include vm_page_array and vm_reserv_array in a crash dump.
736 	 */
737 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
738 		dump_add_page(pa);
739 #endif
740 	phys_avail[biggestone + 1] = new_end;
741 
742 	/*
743 	 * Add physical memory segments corresponding to the available
744 	 * physical pages.
745 	 */
746 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
747 		if (vm_phys_avail_size(i) != 0)
748 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
749 
750 	/*
751 	 * Initialize the physical memory allocator.
752 	 */
753 	vm_phys_init();
754 
755 	/*
756 	 * Initialize the page structures and add every available page to the
757 	 * physical memory allocator's free lists.
758 	 */
759 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
760 	for (ii = 0; ii < vm_page_array_size; ii++) {
761 		m = &vm_page_array[ii];
762 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
763 		m->flags = PG_FICTITIOUS;
764 	}
765 #endif
766 	vm_cnt.v_page_count = 0;
767 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
768 		seg = &vm_phys_segs[segind];
769 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
770 		    m++, pa += PAGE_SIZE)
771 			vm_page_init_page(m, pa, segind);
772 
773 		/*
774 		 * Add the segment's pages that are covered by one of
775 		 * phys_avail's ranges to the free lists.
776 		 */
777 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
778 			if (seg->end <= phys_avail[i] ||
779 			    seg->start >= phys_avail[i + 1])
780 				continue;
781 
782 			startp = MAX(seg->start, phys_avail[i]);
783 			endp = MIN(seg->end, phys_avail[i + 1]);
784 			pagecount = (u_long)atop(endp - startp);
785 			if (pagecount == 0)
786 				continue;
787 
788 			m = seg->first_page + atop(startp - seg->start);
789 			vmd = VM_DOMAIN(seg->domain);
790 			vm_domain_free_lock(vmd);
791 			vm_phys_enqueue_contig(m, pagecount);
792 			vm_domain_free_unlock(vmd);
793 			vm_domain_freecnt_inc(vmd, pagecount);
794 			vm_cnt.v_page_count += (u_int)pagecount;
795 			vmd->vmd_page_count += (u_int)pagecount;
796 			vmd->vmd_segs |= 1UL << segind;
797 		}
798 	}
799 
800 	/*
801 	 * Remove blacklisted pages from the physical memory allocator.
802 	 */
803 	TAILQ_INIT(&blacklist_head);
804 	vm_page_blacklist_load(&list, &listend);
805 	vm_page_blacklist_check(list, listend);
806 
807 	list = kern_getenv("vm.blacklist");
808 	vm_page_blacklist_check(list, NULL);
809 
810 	freeenv(list);
811 #if VM_NRESERVLEVEL > 0
812 	/*
813 	 * Initialize the reservation management system.
814 	 */
815 	vm_reserv_init();
816 #endif
817 
818 	return (vaddr);
819 }
820 
821 void
822 vm_page_reference(vm_page_t m)
823 {
824 
825 	vm_page_aflag_set(m, PGA_REFERENCED);
826 }
827 
828 /*
829  *	vm_page_trybusy
830  *
831  *	Helper routine for grab functions to trylock busy.
832  *
833  *	Returns true on success and false on failure.
834  */
835 static bool
836 vm_page_trybusy(vm_page_t m, int allocflags)
837 {
838 
839 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
840 		return (vm_page_trysbusy(m));
841 	else
842 		return (vm_page_tryxbusy(m));
843 }
844 
845 /*
846  *	vm_page_tryacquire
847  *
848  *	Helper routine for grab functions to trylock busy and wire.
849  *
850  *	Returns true on success and false on failure.
851  */
852 static inline bool
853 vm_page_tryacquire(vm_page_t m, int allocflags)
854 {
855 	bool locked;
856 
857 	locked = vm_page_trybusy(m, allocflags);
858 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
859 		vm_page_wire(m);
860 	return (locked);
861 }
862 
863 /*
864  *	vm_page_busy_acquire:
865  *
866  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
867  *	and drop the object lock if necessary.
868  */
869 bool
870 vm_page_busy_acquire(vm_page_t m, int allocflags)
871 {
872 	vm_object_t obj;
873 	bool locked;
874 
875 	/*
876 	 * The page-specific object must be cached because page
877 	 * identity can change during the sleep, causing the
878 	 * re-lock of a different object.
879 	 * It is assumed that a reference to the object is already
880 	 * held by the callers.
881 	 */
882 	obj = atomic_load_ptr(&m->object);
883 	for (;;) {
884 		if (vm_page_tryacquire(m, allocflags))
885 			return (true);
886 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
887 			return (false);
888 		if (obj != NULL)
889 			locked = VM_OBJECT_WOWNED(obj);
890 		else
891 			locked = false;
892 		MPASS(locked || vm_page_wired(m));
893 		if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
894 		    locked) && locked)
895 			VM_OBJECT_WLOCK(obj);
896 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
897 			return (false);
898 		KASSERT(m->object == obj || m->object == NULL,
899 		    ("vm_page_busy_acquire: page %p does not belong to %p",
900 		    m, obj));
901 	}
902 }
903 
904 /*
905  *	vm_page_busy_downgrade:
906  *
907  *	Downgrade an exclusive busy page into a single shared busy page.
908  */
909 void
910 vm_page_busy_downgrade(vm_page_t m)
911 {
912 	u_int x;
913 
914 	vm_page_assert_xbusied(m);
915 
916 	x = vm_page_busy_fetch(m);
917 	for (;;) {
918 		if (atomic_fcmpset_rel_int(&m->busy_lock,
919 		    &x, VPB_SHARERS_WORD(1)))
920 			break;
921 	}
922 	if ((x & VPB_BIT_WAITERS) != 0)
923 		wakeup(m);
924 }
925 
926 /*
927  *
928  *	vm_page_busy_tryupgrade:
929  *
930  *	Attempt to upgrade a single shared busy into an exclusive busy.
931  */
932 int
933 vm_page_busy_tryupgrade(vm_page_t m)
934 {
935 	u_int ce, x;
936 
937 	vm_page_assert_sbusied(m);
938 
939 	x = vm_page_busy_fetch(m);
940 	ce = VPB_CURTHREAD_EXCLUSIVE;
941 	for (;;) {
942 		if (VPB_SHARERS(x) > 1)
943 			return (0);
944 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
945 		    ("vm_page_busy_tryupgrade: invalid lock state"));
946 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
947 		    ce | (x & VPB_BIT_WAITERS)))
948 			continue;
949 		return (1);
950 	}
951 }
952 
953 /*
954  *	vm_page_sbusied:
955  *
956  *	Return a positive value if the page is shared busied, 0 otherwise.
957  */
958 int
959 vm_page_sbusied(vm_page_t m)
960 {
961 	u_int x;
962 
963 	x = vm_page_busy_fetch(m);
964 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
965 }
966 
967 /*
968  *	vm_page_sunbusy:
969  *
970  *	Shared unbusy a page.
971  */
972 void
973 vm_page_sunbusy(vm_page_t m)
974 {
975 	u_int x;
976 
977 	vm_page_assert_sbusied(m);
978 
979 	x = vm_page_busy_fetch(m);
980 	for (;;) {
981 		KASSERT(x != VPB_FREED,
982 		    ("vm_page_sunbusy: Unlocking freed page."));
983 		if (VPB_SHARERS(x) > 1) {
984 			if (atomic_fcmpset_int(&m->busy_lock, &x,
985 			    x - VPB_ONE_SHARER))
986 				break;
987 			continue;
988 		}
989 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
990 		    ("vm_page_sunbusy: invalid lock state"));
991 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
992 			continue;
993 		if ((x & VPB_BIT_WAITERS) == 0)
994 			break;
995 		wakeup(m);
996 		break;
997 	}
998 }
999 
1000 /*
1001  *	vm_page_busy_sleep:
1002  *
1003  *	Sleep if the page is busy, using the page pointer as wchan.
1004  *	This is used to implement the hard-path of busying mechanism.
1005  *
1006  *	If nonshared is true, sleep only if the page is xbusy.
1007  *
1008  *	The object lock must be held on entry and will be released on exit.
1009  */
1010 void
1011 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1012 {
1013 	vm_object_t obj;
1014 
1015 	obj = m->object;
1016 	VM_OBJECT_ASSERT_LOCKED(obj);
1017 	vm_page_lock_assert(m, MA_NOTOWNED);
1018 
1019 	if (!_vm_page_busy_sleep(obj, m, m->pindex, wmesg,
1020 	    nonshared ? VM_ALLOC_SBUSY : 0 , true))
1021 		VM_OBJECT_DROP(obj);
1022 }
1023 
1024 /*
1025  *	vm_page_busy_sleep_unlocked:
1026  *
1027  *	Sleep if the page is busy, using the page pointer as wchan.
1028  *	This is used to implement the hard-path of busying mechanism.
1029  *
1030  *	If nonshared is true, sleep only if the page is xbusy.
1031  *
1032  *	The object lock must not be held on entry.  The operation will
1033  *	return if the page changes identity.
1034  */
1035 void
1036 vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1037     const char *wmesg, bool nonshared)
1038 {
1039 
1040 	VM_OBJECT_ASSERT_UNLOCKED(obj);
1041 	vm_page_lock_assert(m, MA_NOTOWNED);
1042 
1043 	_vm_page_busy_sleep(obj, m, pindex, wmesg,
1044 	    nonshared ? VM_ALLOC_SBUSY : 0, false);
1045 }
1046 
1047 /*
1048  *	_vm_page_busy_sleep:
1049  *
1050  *	Internal busy sleep function.  Verifies the page identity and
1051  *	lockstate against parameters.  Returns true if it sleeps and
1052  *	false otherwise.
1053  *
1054  *	If locked is true the lock will be dropped for any true returns
1055  *	and held for any false returns.
1056  */
1057 static bool
1058 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
1059     const char *wmesg, int allocflags, bool locked)
1060 {
1061 	bool xsleep;
1062 	u_int x;
1063 
1064 	/*
1065 	 * If the object is busy we must wait for that to drain to zero
1066 	 * before trying the page again.
1067 	 */
1068 	if (obj != NULL && vm_object_busied(obj)) {
1069 		if (locked)
1070 			VM_OBJECT_DROP(obj);
1071 		vm_object_busy_wait(obj, wmesg);
1072 		return (true);
1073 	}
1074 
1075 	if (!vm_page_busied(m))
1076 		return (false);
1077 
1078 	xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
1079 	sleepq_lock(m);
1080 	x = vm_page_busy_fetch(m);
1081 	do {
1082 		/*
1083 		 * If the page changes objects or becomes unlocked we can
1084 		 * simply return.
1085 		 */
1086 		if (x == VPB_UNBUSIED ||
1087 		    (xsleep && (x & VPB_BIT_SHARED) != 0) ||
1088 		    m->object != obj || m->pindex != pindex) {
1089 			sleepq_release(m);
1090 			return (false);
1091 		}
1092 		if ((x & VPB_BIT_WAITERS) != 0)
1093 			break;
1094 	} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
1095 	if (locked)
1096 		VM_OBJECT_DROP(obj);
1097 	DROP_GIANT();
1098 	sleepq_add(m, NULL, wmesg, 0, 0);
1099 	sleepq_wait(m, PVM);
1100 	PICKUP_GIANT();
1101 	return (true);
1102 }
1103 
1104 /*
1105  *	vm_page_trysbusy:
1106  *
1107  *	Try to shared busy a page.
1108  *	If the operation succeeds 1 is returned otherwise 0.
1109  *	The operation never sleeps.
1110  */
1111 int
1112 vm_page_trysbusy(vm_page_t m)
1113 {
1114 	vm_object_t obj;
1115 	u_int x;
1116 
1117 	obj = m->object;
1118 	x = vm_page_busy_fetch(m);
1119 	for (;;) {
1120 		if ((x & VPB_BIT_SHARED) == 0)
1121 			return (0);
1122 		/*
1123 		 * Reduce the window for transient busies that will trigger
1124 		 * false negatives in vm_page_ps_test().
1125 		 */
1126 		if (obj != NULL && vm_object_busied(obj))
1127 			return (0);
1128 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1129 		    x + VPB_ONE_SHARER))
1130 			break;
1131 	}
1132 
1133 	/* Refetch the object now that we're guaranteed that it is stable. */
1134 	obj = m->object;
1135 	if (obj != NULL && vm_object_busied(obj)) {
1136 		vm_page_sunbusy(m);
1137 		return (0);
1138 	}
1139 	return (1);
1140 }
1141 
1142 /*
1143  *	vm_page_tryxbusy:
1144  *
1145  *	Try to exclusive busy a page.
1146  *	If the operation succeeds 1 is returned otherwise 0.
1147  *	The operation never sleeps.
1148  */
1149 int
1150 vm_page_tryxbusy(vm_page_t m)
1151 {
1152 	vm_object_t obj;
1153 
1154         if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
1155             VPB_CURTHREAD_EXCLUSIVE) == 0)
1156 		return (0);
1157 
1158 	obj = m->object;
1159 	if (obj != NULL && vm_object_busied(obj)) {
1160 		vm_page_xunbusy(m);
1161 		return (0);
1162 	}
1163 	return (1);
1164 }
1165 
1166 static void
1167 vm_page_xunbusy_hard_tail(vm_page_t m)
1168 {
1169 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1170 	/* Wake the waiter. */
1171 	wakeup(m);
1172 }
1173 
1174 /*
1175  *	vm_page_xunbusy_hard:
1176  *
1177  *	Called when unbusy has failed because there is a waiter.
1178  */
1179 void
1180 vm_page_xunbusy_hard(vm_page_t m)
1181 {
1182 	vm_page_assert_xbusied(m);
1183 	vm_page_xunbusy_hard_tail(m);
1184 }
1185 
1186 void
1187 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1188 {
1189 	vm_page_assert_xbusied_unchecked(m);
1190 	vm_page_xunbusy_hard_tail(m);
1191 }
1192 
1193 static void
1194 vm_page_busy_free(vm_page_t m)
1195 {
1196 	u_int x;
1197 
1198 	atomic_thread_fence_rel();
1199 	x = atomic_swap_int(&m->busy_lock, VPB_FREED);
1200 	if ((x & VPB_BIT_WAITERS) != 0)
1201 		wakeup(m);
1202 }
1203 
1204 /*
1205  *	vm_page_unhold_pages:
1206  *
1207  *	Unhold each of the pages that is referenced by the given array.
1208  */
1209 void
1210 vm_page_unhold_pages(vm_page_t *ma, int count)
1211 {
1212 
1213 	for (; count != 0; count--) {
1214 		vm_page_unwire(*ma, PQ_ACTIVE);
1215 		ma++;
1216 	}
1217 }
1218 
1219 vm_page_t
1220 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1221 {
1222 	vm_page_t m;
1223 
1224 #ifdef VM_PHYSSEG_SPARSE
1225 	m = vm_phys_paddr_to_vm_page(pa);
1226 	if (m == NULL)
1227 		m = vm_phys_fictitious_to_vm_page(pa);
1228 	return (m);
1229 #elif defined(VM_PHYSSEG_DENSE)
1230 	long pi;
1231 
1232 	pi = atop(pa);
1233 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1234 		m = &vm_page_array[pi - first_page];
1235 		return (m);
1236 	}
1237 	return (vm_phys_fictitious_to_vm_page(pa));
1238 #else
1239 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1240 #endif
1241 }
1242 
1243 /*
1244  *	vm_page_getfake:
1245  *
1246  *	Create a fictitious page with the specified physical address and
1247  *	memory attribute.  The memory attribute is the only the machine-
1248  *	dependent aspect of a fictitious page that must be initialized.
1249  */
1250 vm_page_t
1251 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1252 {
1253 	vm_page_t m;
1254 
1255 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1256 	vm_page_initfake(m, paddr, memattr);
1257 	return (m);
1258 }
1259 
1260 void
1261 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1262 {
1263 
1264 	if ((m->flags & PG_FICTITIOUS) != 0) {
1265 		/*
1266 		 * The page's memattr might have changed since the
1267 		 * previous initialization.  Update the pmap to the
1268 		 * new memattr.
1269 		 */
1270 		goto memattr;
1271 	}
1272 	m->phys_addr = paddr;
1273 	m->a.queue = PQ_NONE;
1274 	/* Fictitious pages don't use "segind". */
1275 	m->flags = PG_FICTITIOUS;
1276 	/* Fictitious pages don't use "order" or "pool". */
1277 	m->oflags = VPO_UNMANAGED;
1278 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1279 	/* Fictitious pages are unevictable. */
1280 	m->ref_count = 1;
1281 	pmap_page_init(m);
1282 memattr:
1283 	pmap_page_set_memattr(m, memattr);
1284 }
1285 
1286 /*
1287  *	vm_page_putfake:
1288  *
1289  *	Release a fictitious page.
1290  */
1291 void
1292 vm_page_putfake(vm_page_t m)
1293 {
1294 
1295 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1296 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1297 	    ("vm_page_putfake: bad page %p", m));
1298 	vm_page_assert_xbusied(m);
1299 	vm_page_busy_free(m);
1300 	uma_zfree(fakepg_zone, m);
1301 }
1302 
1303 /*
1304  *	vm_page_updatefake:
1305  *
1306  *	Update the given fictitious page to the specified physical address and
1307  *	memory attribute.
1308  */
1309 void
1310 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1311 {
1312 
1313 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1314 	    ("vm_page_updatefake: bad page %p", m));
1315 	m->phys_addr = paddr;
1316 	pmap_page_set_memattr(m, memattr);
1317 }
1318 
1319 /*
1320  *	vm_page_free:
1321  *
1322  *	Free a page.
1323  */
1324 void
1325 vm_page_free(vm_page_t m)
1326 {
1327 
1328 	m->flags &= ~PG_ZERO;
1329 	vm_page_free_toq(m);
1330 }
1331 
1332 /*
1333  *	vm_page_free_zero:
1334  *
1335  *	Free a page to the zerod-pages queue
1336  */
1337 void
1338 vm_page_free_zero(vm_page_t m)
1339 {
1340 
1341 	m->flags |= PG_ZERO;
1342 	vm_page_free_toq(m);
1343 }
1344 
1345 /*
1346  * Unbusy and handle the page queueing for a page from a getpages request that
1347  * was optionally read ahead or behind.
1348  */
1349 void
1350 vm_page_readahead_finish(vm_page_t m)
1351 {
1352 
1353 	/* We shouldn't put invalid pages on queues. */
1354 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1355 
1356 	/*
1357 	 * Since the page is not the actually needed one, whether it should
1358 	 * be activated or deactivated is not obvious.  Empirical results
1359 	 * have shown that deactivating the page is usually the best choice,
1360 	 * unless the page is wanted by another thread.
1361 	 */
1362 	if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
1363 		vm_page_activate(m);
1364 	else
1365 		vm_page_deactivate(m);
1366 	vm_page_xunbusy_unchecked(m);
1367 }
1368 
1369 /*
1370  * Destroy the identity of an invalid page and free it if possible.
1371  * This is intended to be used when reading a page from backing store fails.
1372  */
1373 void
1374 vm_page_free_invalid(vm_page_t m)
1375 {
1376 
1377 	KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
1378 	KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
1379 	KASSERT(m->object != NULL, ("page %p has no object", m));
1380 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1381 
1382 	/*
1383 	 * We may be attempting to free the page as part of the handling for an
1384 	 * I/O error, in which case the page was xbusied by a different thread.
1385 	 */
1386 	vm_page_xbusy_claim(m);
1387 
1388 	/*
1389 	 * If someone has wired this page while the object lock
1390 	 * was not held, then the thread that unwires is responsible
1391 	 * for freeing the page.  Otherwise just free the page now.
1392 	 * The wire count of this unmapped page cannot change while
1393 	 * we have the page xbusy and the page's object wlocked.
1394 	 */
1395 	if (vm_page_remove(m))
1396 		vm_page_free(m);
1397 }
1398 
1399 /*
1400  *	vm_page_sleep_if_busy:
1401  *
1402  *	Sleep and release the object lock if the page is busied.
1403  *	Returns TRUE if the thread slept.
1404  *
1405  *	The given page must be unlocked and object containing it must
1406  *	be locked.
1407  */
1408 int
1409 vm_page_sleep_if_busy(vm_page_t m, const char *wmesg)
1410 {
1411 	vm_object_t obj;
1412 
1413 	vm_page_lock_assert(m, MA_NOTOWNED);
1414 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1415 
1416 	/*
1417 	 * The page-specific object must be cached because page
1418 	 * identity can change during the sleep, causing the
1419 	 * re-lock of a different object.
1420 	 * It is assumed that a reference to the object is already
1421 	 * held by the callers.
1422 	 */
1423 	obj = m->object;
1424 	if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, 0, true)) {
1425 		VM_OBJECT_WLOCK(obj);
1426 		return (TRUE);
1427 	}
1428 	return (FALSE);
1429 }
1430 
1431 /*
1432  *	vm_page_sleep_if_xbusy:
1433  *
1434  *	Sleep and release the object lock if the page is xbusied.
1435  *	Returns TRUE if the thread slept.
1436  *
1437  *	The given page must be unlocked and object containing it must
1438  *	be locked.
1439  */
1440 int
1441 vm_page_sleep_if_xbusy(vm_page_t m, const char *wmesg)
1442 {
1443 	vm_object_t obj;
1444 
1445 	vm_page_lock_assert(m, MA_NOTOWNED);
1446 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1447 
1448 	/*
1449 	 * The page-specific object must be cached because page
1450 	 * identity can change during the sleep, causing the
1451 	 * re-lock of a different object.
1452 	 * It is assumed that a reference to the object is already
1453 	 * held by the callers.
1454 	 */
1455 	obj = m->object;
1456 	if (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, VM_ALLOC_SBUSY,
1457 	    true)) {
1458 		VM_OBJECT_WLOCK(obj);
1459 		return (TRUE);
1460 	}
1461 	return (FALSE);
1462 }
1463 
1464 /*
1465  *	vm_page_dirty_KBI:		[ internal use only ]
1466  *
1467  *	Set all bits in the page's dirty field.
1468  *
1469  *	The object containing the specified page must be locked if the
1470  *	call is made from the machine-independent layer.
1471  *
1472  *	See vm_page_clear_dirty_mask().
1473  *
1474  *	This function should only be called by vm_page_dirty().
1475  */
1476 void
1477 vm_page_dirty_KBI(vm_page_t m)
1478 {
1479 
1480 	/* Refer to this operation by its public name. */
1481 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1482 	m->dirty = VM_PAGE_BITS_ALL;
1483 }
1484 
1485 /*
1486  *	vm_page_insert:		[ internal use only ]
1487  *
1488  *	Inserts the given mem entry into the object and object list.
1489  *
1490  *	The object must be locked.
1491  */
1492 int
1493 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1494 {
1495 	vm_page_t mpred;
1496 
1497 	VM_OBJECT_ASSERT_WLOCKED(object);
1498 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1499 	return (vm_page_insert_after(m, object, pindex, mpred));
1500 }
1501 
1502 /*
1503  *	vm_page_insert_after:
1504  *
1505  *	Inserts the page "m" into the specified object at offset "pindex".
1506  *
1507  *	The page "mpred" must immediately precede the offset "pindex" within
1508  *	the specified object.
1509  *
1510  *	The object must be locked.
1511  */
1512 static int
1513 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1514     vm_page_t mpred)
1515 {
1516 	vm_page_t msucc;
1517 
1518 	VM_OBJECT_ASSERT_WLOCKED(object);
1519 	KASSERT(m->object == NULL,
1520 	    ("vm_page_insert_after: page already inserted"));
1521 	if (mpred != NULL) {
1522 		KASSERT(mpred->object == object,
1523 		    ("vm_page_insert_after: object doesn't contain mpred"));
1524 		KASSERT(mpred->pindex < pindex,
1525 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1526 		msucc = TAILQ_NEXT(mpred, listq);
1527 	} else
1528 		msucc = TAILQ_FIRST(&object->memq);
1529 	if (msucc != NULL)
1530 		KASSERT(msucc->pindex > pindex,
1531 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1532 
1533 	/*
1534 	 * Record the object/offset pair in this page.
1535 	 */
1536 	m->object = object;
1537 	m->pindex = pindex;
1538 	m->ref_count |= VPRC_OBJREF;
1539 
1540 	/*
1541 	 * Now link into the object's ordered list of backed pages.
1542 	 */
1543 	if (vm_radix_insert(&object->rtree, m)) {
1544 		m->object = NULL;
1545 		m->pindex = 0;
1546 		m->ref_count &= ~VPRC_OBJREF;
1547 		return (1);
1548 	}
1549 	vm_page_insert_radixdone(m, object, mpred);
1550 	return (0);
1551 }
1552 
1553 /*
1554  *	vm_page_insert_radixdone:
1555  *
1556  *	Complete page "m" insertion into the specified object after the
1557  *	radix trie hooking.
1558  *
1559  *	The page "mpred" must precede the offset "m->pindex" within the
1560  *	specified object.
1561  *
1562  *	The object must be locked.
1563  */
1564 static void
1565 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1566 {
1567 
1568 	VM_OBJECT_ASSERT_WLOCKED(object);
1569 	KASSERT(object != NULL && m->object == object,
1570 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1571 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1572 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1573 	if (mpred != NULL) {
1574 		KASSERT(mpred->object == object,
1575 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1576 		KASSERT(mpred->pindex < m->pindex,
1577 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1578 	}
1579 
1580 	if (mpred != NULL)
1581 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1582 	else
1583 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1584 
1585 	/*
1586 	 * Show that the object has one more resident page.
1587 	 */
1588 	object->resident_page_count++;
1589 
1590 	/*
1591 	 * Hold the vnode until the last page is released.
1592 	 */
1593 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1594 		vhold(object->handle);
1595 
1596 	/*
1597 	 * Since we are inserting a new and possibly dirty page,
1598 	 * update the object's generation count.
1599 	 */
1600 	if (pmap_page_is_write_mapped(m))
1601 		vm_object_set_writeable_dirty(object);
1602 }
1603 
1604 /*
1605  * Do the work to remove a page from its object.  The caller is responsible for
1606  * updating the page's fields to reflect this removal.
1607  */
1608 static void
1609 vm_page_object_remove(vm_page_t m)
1610 {
1611 	vm_object_t object;
1612 	vm_page_t mrem;
1613 
1614 	vm_page_assert_xbusied(m);
1615 	object = m->object;
1616 	VM_OBJECT_ASSERT_WLOCKED(object);
1617 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1618 	    ("page %p is missing its object ref", m));
1619 
1620 	/* Deferred free of swap space. */
1621 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1622 		vm_pager_page_unswapped(m);
1623 
1624 	m->object = NULL;
1625 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1626 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1627 
1628 	/*
1629 	 * Now remove from the object's list of backed pages.
1630 	 */
1631 	TAILQ_REMOVE(&object->memq, m, listq);
1632 
1633 	/*
1634 	 * And show that the object has one fewer resident page.
1635 	 */
1636 	object->resident_page_count--;
1637 
1638 	/*
1639 	 * The vnode may now be recycled.
1640 	 */
1641 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1642 		vdrop(object->handle);
1643 }
1644 
1645 /*
1646  *	vm_page_remove:
1647  *
1648  *	Removes the specified page from its containing object, but does not
1649  *	invalidate any backing storage.  Returns true if the object's reference
1650  *	was the last reference to the page, and false otherwise.
1651  *
1652  *	The object must be locked and the page must be exclusively busied.
1653  *	The exclusive busy will be released on return.  If this is not the
1654  *	final ref and the caller does not hold a wire reference it may not
1655  *	continue to access the page.
1656  */
1657 bool
1658 vm_page_remove(vm_page_t m)
1659 {
1660 	bool dropped;
1661 
1662 	dropped = vm_page_remove_xbusy(m);
1663 	vm_page_xunbusy(m);
1664 
1665 	return (dropped);
1666 }
1667 
1668 /*
1669  *	vm_page_remove_xbusy
1670  *
1671  *	Removes the page but leaves the xbusy held.  Returns true if this
1672  *	removed the final ref and false otherwise.
1673  */
1674 bool
1675 vm_page_remove_xbusy(vm_page_t m)
1676 {
1677 
1678 	vm_page_object_remove(m);
1679 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1680 }
1681 
1682 /*
1683  *	vm_page_lookup:
1684  *
1685  *	Returns the page associated with the object/offset
1686  *	pair specified; if none is found, NULL is returned.
1687  *
1688  *	The object must be locked.
1689  */
1690 vm_page_t
1691 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1692 {
1693 
1694 	VM_OBJECT_ASSERT_LOCKED(object);
1695 	return (vm_radix_lookup(&object->rtree, pindex));
1696 }
1697 
1698 /*
1699  *	vm_page_lookup_unlocked:
1700  *
1701  *	Returns the page associated with the object/offset pair specified;
1702  *	if none is found, NULL is returned.  The page may be no longer be
1703  *	present in the object at the time that this function returns.  Only
1704  *	useful for opportunistic checks such as inmem().
1705  */
1706 vm_page_t
1707 vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
1708 {
1709 
1710 	return (vm_radix_lookup_unlocked(&object->rtree, pindex));
1711 }
1712 
1713 /*
1714  *	vm_page_relookup:
1715  *
1716  *	Returns a page that must already have been busied by
1717  *	the caller.  Used for bogus page replacement.
1718  */
1719 vm_page_t
1720 vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
1721 {
1722 	vm_page_t m;
1723 
1724 	m = vm_radix_lookup_unlocked(&object->rtree, pindex);
1725 	KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
1726 	    m->object == object && m->pindex == pindex,
1727 	    ("vm_page_relookup: Invalid page %p", m));
1728 	return (m);
1729 }
1730 
1731 /*
1732  * This should only be used by lockless functions for releasing transient
1733  * incorrect acquires.  The page may have been freed after we acquired a
1734  * busy lock.  In this case busy_lock == VPB_FREED and we have nothing
1735  * further to do.
1736  */
1737 static void
1738 vm_page_busy_release(vm_page_t m)
1739 {
1740 	u_int x;
1741 
1742 	x = vm_page_busy_fetch(m);
1743 	for (;;) {
1744 		if (x == VPB_FREED)
1745 			break;
1746 		if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
1747 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1748 			    x - VPB_ONE_SHARER))
1749 				break;
1750 			continue;
1751 		}
1752 		KASSERT((x & VPB_BIT_SHARED) != 0 ||
1753 		    (x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
1754 		    ("vm_page_busy_release: %p xbusy not owned.", m));
1755 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1756 			continue;
1757 		if ((x & VPB_BIT_WAITERS) != 0)
1758 			wakeup(m);
1759 		break;
1760 	}
1761 }
1762 
1763 /*
1764  *	vm_page_find_least:
1765  *
1766  *	Returns the page associated with the object with least pindex
1767  *	greater than or equal to the parameter pindex, or NULL.
1768  *
1769  *	The object must be locked.
1770  */
1771 vm_page_t
1772 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1773 {
1774 	vm_page_t m;
1775 
1776 	VM_OBJECT_ASSERT_LOCKED(object);
1777 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1778 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1779 	return (m);
1780 }
1781 
1782 /*
1783  * Returns the given page's successor (by pindex) within the object if it is
1784  * resident; if none is found, NULL is returned.
1785  *
1786  * The object must be locked.
1787  */
1788 vm_page_t
1789 vm_page_next(vm_page_t m)
1790 {
1791 	vm_page_t next;
1792 
1793 	VM_OBJECT_ASSERT_LOCKED(m->object);
1794 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1795 		MPASS(next->object == m->object);
1796 		if (next->pindex != m->pindex + 1)
1797 			next = NULL;
1798 	}
1799 	return (next);
1800 }
1801 
1802 /*
1803  * Returns the given page's predecessor (by pindex) within the object if it is
1804  * resident; if none is found, NULL is returned.
1805  *
1806  * The object must be locked.
1807  */
1808 vm_page_t
1809 vm_page_prev(vm_page_t m)
1810 {
1811 	vm_page_t prev;
1812 
1813 	VM_OBJECT_ASSERT_LOCKED(m->object);
1814 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1815 		MPASS(prev->object == m->object);
1816 		if (prev->pindex != m->pindex - 1)
1817 			prev = NULL;
1818 	}
1819 	return (prev);
1820 }
1821 
1822 /*
1823  * Uses the page mnew as a replacement for an existing page at index
1824  * pindex which must be already present in the object.
1825  *
1826  * Both pages must be exclusively busied on enter.  The old page is
1827  * unbusied on exit.
1828  *
1829  * A return value of true means mold is now free.  If this is not the
1830  * final ref and the caller does not hold a wire reference it may not
1831  * continue to access the page.
1832  */
1833 static bool
1834 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1835     vm_page_t mold)
1836 {
1837 	vm_page_t mret;
1838 	bool dropped;
1839 
1840 	VM_OBJECT_ASSERT_WLOCKED(object);
1841 	vm_page_assert_xbusied(mold);
1842 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1843 	    ("vm_page_replace: page %p already in object", mnew));
1844 
1845 	/*
1846 	 * This function mostly follows vm_page_insert() and
1847 	 * vm_page_remove() without the radix, object count and vnode
1848 	 * dance.  Double check such functions for more comments.
1849 	 */
1850 
1851 	mnew->object = object;
1852 	mnew->pindex = pindex;
1853 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1854 	mret = vm_radix_replace(&object->rtree, mnew);
1855 	KASSERT(mret == mold,
1856 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1857 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1858 	    (mnew->oflags & VPO_UNMANAGED),
1859 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1860 
1861 	/* Keep the resident page list in sorted order. */
1862 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1863 	TAILQ_REMOVE(&object->memq, mold, listq);
1864 	mold->object = NULL;
1865 
1866 	/*
1867 	 * The object's resident_page_count does not change because we have
1868 	 * swapped one page for another, but the generation count should
1869 	 * change if the page is dirty.
1870 	 */
1871 	if (pmap_page_is_write_mapped(mnew))
1872 		vm_object_set_writeable_dirty(object);
1873 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1874 	vm_page_xunbusy(mold);
1875 
1876 	return (dropped);
1877 }
1878 
1879 void
1880 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1881     vm_page_t mold)
1882 {
1883 
1884 	vm_page_assert_xbusied(mnew);
1885 
1886 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1887 		vm_page_free(mold);
1888 }
1889 
1890 /*
1891  *	vm_page_rename:
1892  *
1893  *	Move the given memory entry from its
1894  *	current object to the specified target object/offset.
1895  *
1896  *	Note: swap associated with the page must be invalidated by the move.  We
1897  *	      have to do this for several reasons:  (1) we aren't freeing the
1898  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1899  *	      moving the page from object A to B, and will then later move
1900  *	      the backing store from A to B and we can't have a conflict.
1901  *
1902  *	Note: we *always* dirty the page.  It is necessary both for the
1903  *	      fact that we moved it, and because we may be invalidating
1904  *	      swap.
1905  *
1906  *	The objects must be locked.
1907  */
1908 int
1909 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1910 {
1911 	vm_page_t mpred;
1912 	vm_pindex_t opidx;
1913 
1914 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1915 
1916 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1917 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1918 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1919 	    ("vm_page_rename: pindex already renamed"));
1920 
1921 	/*
1922 	 * Create a custom version of vm_page_insert() which does not depend
1923 	 * by m_prev and can cheat on the implementation aspects of the
1924 	 * function.
1925 	 */
1926 	opidx = m->pindex;
1927 	m->pindex = new_pindex;
1928 	if (vm_radix_insert(&new_object->rtree, m)) {
1929 		m->pindex = opidx;
1930 		return (1);
1931 	}
1932 
1933 	/*
1934 	 * The operation cannot fail anymore.  The removal must happen before
1935 	 * the listq iterator is tainted.
1936 	 */
1937 	m->pindex = opidx;
1938 	vm_page_object_remove(m);
1939 
1940 	/* Return back to the new pindex to complete vm_page_insert(). */
1941 	m->pindex = new_pindex;
1942 	m->object = new_object;
1943 
1944 	vm_page_insert_radixdone(m, new_object, mpred);
1945 	vm_page_dirty(m);
1946 	return (0);
1947 }
1948 
1949 /*
1950  *	vm_page_alloc:
1951  *
1952  *	Allocate and return a page that is associated with the specified
1953  *	object and offset pair.  By default, this page is exclusive busied.
1954  *
1955  *	The caller must always specify an allocation class.
1956  *
1957  *	allocation classes:
1958  *	VM_ALLOC_NORMAL		normal process request
1959  *	VM_ALLOC_SYSTEM		system *really* needs a page
1960  *	VM_ALLOC_INTERRUPT	interrupt time request
1961  *
1962  *	optional allocation flags:
1963  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1964  *				intends to allocate
1965  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1966  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1967  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1968  *				should not be exclusive busy
1969  *	VM_ALLOC_SBUSY		shared busy the allocated page
1970  *	VM_ALLOC_WIRED		wire the allocated page
1971  *	VM_ALLOC_ZERO		prefer a zeroed page
1972  */
1973 vm_page_t
1974 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1975 {
1976 
1977 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1978 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1979 }
1980 
1981 vm_page_t
1982 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1983     int req)
1984 {
1985 
1986 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1987 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1988 	    NULL));
1989 }
1990 
1991 /*
1992  * Allocate a page in the specified object with the given page index.  To
1993  * optimize insertion of the page into the object, the caller must also specifiy
1994  * the resident page in the object with largest index smaller than the given
1995  * page index, or NULL if no such page exists.
1996  */
1997 vm_page_t
1998 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1999     int req, vm_page_t mpred)
2000 {
2001 	struct vm_domainset_iter di;
2002 	vm_page_t m;
2003 	int domain;
2004 
2005 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2006 	do {
2007 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
2008 		    mpred);
2009 		if (m != NULL)
2010 			break;
2011 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2012 
2013 	return (m);
2014 }
2015 
2016 /*
2017  * Returns true if the number of free pages exceeds the minimum
2018  * for the request class and false otherwise.
2019  */
2020 static int
2021 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
2022 {
2023 	u_int limit, old, new;
2024 
2025 	if (req_class == VM_ALLOC_INTERRUPT)
2026 		limit = 0;
2027 	else if (req_class == VM_ALLOC_SYSTEM)
2028 		limit = vmd->vmd_interrupt_free_min;
2029 	else
2030 		limit = vmd->vmd_free_reserved;
2031 
2032 	/*
2033 	 * Attempt to reserve the pages.  Fail if we're below the limit.
2034 	 */
2035 	limit += npages;
2036 	old = vmd->vmd_free_count;
2037 	do {
2038 		if (old < limit)
2039 			return (0);
2040 		new = old - npages;
2041 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
2042 
2043 	/* Wake the page daemon if we've crossed the threshold. */
2044 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
2045 		pagedaemon_wakeup(vmd->vmd_domain);
2046 
2047 	/* Only update bitsets on transitions. */
2048 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
2049 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
2050 		vm_domain_set(vmd);
2051 
2052 	return (1);
2053 }
2054 
2055 int
2056 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
2057 {
2058 	int req_class;
2059 
2060 	/*
2061 	 * The page daemon is allowed to dig deeper into the free page list.
2062 	 */
2063 	req_class = req & VM_ALLOC_CLASS_MASK;
2064 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2065 		req_class = VM_ALLOC_SYSTEM;
2066 	return (_vm_domain_allocate(vmd, req_class, npages));
2067 }
2068 
2069 vm_page_t
2070 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
2071     int req, vm_page_t mpred)
2072 {
2073 	struct vm_domain *vmd;
2074 	vm_page_t m;
2075 	int flags, pool;
2076 
2077 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2078 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2079 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2080 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2081 	    ("inconsistent object(%p)/req(%x)", object, req));
2082 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2083 	    ("Can't sleep and retry object insertion."));
2084 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2085 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2086 	    (uintmax_t)pindex));
2087 	if (object != NULL)
2088 		VM_OBJECT_ASSERT_WLOCKED(object);
2089 
2090 	flags = 0;
2091 	m = NULL;
2092 	pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2093 again:
2094 #if VM_NRESERVLEVEL > 0
2095 	/*
2096 	 * Can we allocate the page from a reservation?
2097 	 */
2098 	if (vm_object_reserv(object) &&
2099 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2100 	    NULL) {
2101 		goto found;
2102 	}
2103 #endif
2104 	vmd = VM_DOMAIN(domain);
2105 	if (vmd->vmd_pgcache[pool].zone != NULL) {
2106 		m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT | M_NOVM);
2107 		if (m != NULL) {
2108 			flags |= PG_PCPU_CACHE;
2109 			goto found;
2110 		}
2111 	}
2112 	if (vm_domain_allocate(vmd, req, 1)) {
2113 		/*
2114 		 * If not, allocate it from the free page queues.
2115 		 */
2116 		vm_domain_free_lock(vmd);
2117 		m = vm_phys_alloc_pages(domain, pool, 0);
2118 		vm_domain_free_unlock(vmd);
2119 		if (m == NULL) {
2120 			vm_domain_freecnt_inc(vmd, 1);
2121 #if VM_NRESERVLEVEL > 0
2122 			if (vm_reserv_reclaim_inactive(domain))
2123 				goto again;
2124 #endif
2125 		}
2126 	}
2127 	if (m == NULL) {
2128 		/*
2129 		 * Not allocatable, give up.
2130 		 */
2131 		if (vm_domain_alloc_fail(vmd, object, req))
2132 			goto again;
2133 		return (NULL);
2134 	}
2135 
2136 	/*
2137 	 * At this point we had better have found a good page.
2138 	 */
2139 found:
2140 	vm_page_dequeue(m);
2141 	vm_page_alloc_check(m);
2142 
2143 	/*
2144 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2145 	 */
2146 	if ((req & VM_ALLOC_ZERO) != 0)
2147 		flags |= (m->flags & PG_ZERO);
2148 	if ((req & VM_ALLOC_NODUMP) != 0)
2149 		flags |= PG_NODUMP;
2150 	m->flags = flags;
2151 	m->a.flags = 0;
2152 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2153 	    VPO_UNMANAGED : 0;
2154 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2155 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2156 	else if ((req & VM_ALLOC_SBUSY) != 0)
2157 		m->busy_lock = VPB_SHARERS_WORD(1);
2158 	else
2159 		m->busy_lock = VPB_UNBUSIED;
2160 	if (req & VM_ALLOC_WIRED) {
2161 		vm_wire_add(1);
2162 		m->ref_count = 1;
2163 	}
2164 	m->a.act_count = 0;
2165 
2166 	if (object != NULL) {
2167 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2168 			if (req & VM_ALLOC_WIRED) {
2169 				vm_wire_sub(1);
2170 				m->ref_count = 0;
2171 			}
2172 			KASSERT(m->object == NULL, ("page %p has object", m));
2173 			m->oflags = VPO_UNMANAGED;
2174 			m->busy_lock = VPB_UNBUSIED;
2175 			/* Don't change PG_ZERO. */
2176 			vm_page_free_toq(m);
2177 			if (req & VM_ALLOC_WAITFAIL) {
2178 				VM_OBJECT_WUNLOCK(object);
2179 				vm_radix_wait();
2180 				VM_OBJECT_WLOCK(object);
2181 			}
2182 			return (NULL);
2183 		}
2184 
2185 		/* Ignore device objects; the pager sets "memattr" for them. */
2186 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2187 		    (object->flags & OBJ_FICTITIOUS) == 0)
2188 			pmap_page_set_memattr(m, object->memattr);
2189 	} else
2190 		m->pindex = pindex;
2191 
2192 	return (m);
2193 }
2194 
2195 /*
2196  *	vm_page_alloc_contig:
2197  *
2198  *	Allocate a contiguous set of physical pages of the given size "npages"
2199  *	from the free lists.  All of the physical pages must be at or above
2200  *	the given physical address "low" and below the given physical address
2201  *	"high".  The given value "alignment" determines the alignment of the
2202  *	first physical page in the set.  If the given value "boundary" is
2203  *	non-zero, then the set of physical pages cannot cross any physical
2204  *	address boundary that is a multiple of that value.  Both "alignment"
2205  *	and "boundary" must be a power of two.
2206  *
2207  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2208  *	then the memory attribute setting for the physical pages is configured
2209  *	to the object's memory attribute setting.  Otherwise, the memory
2210  *	attribute setting for the physical pages is configured to "memattr",
2211  *	overriding the object's memory attribute setting.  However, if the
2212  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2213  *	memory attribute setting for the physical pages cannot be configured
2214  *	to VM_MEMATTR_DEFAULT.
2215  *
2216  *	The specified object may not contain fictitious pages.
2217  *
2218  *	The caller must always specify an allocation class.
2219  *
2220  *	allocation classes:
2221  *	VM_ALLOC_NORMAL		normal process request
2222  *	VM_ALLOC_SYSTEM		system *really* needs a page
2223  *	VM_ALLOC_INTERRUPT	interrupt time request
2224  *
2225  *	optional allocation flags:
2226  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2227  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2228  *	VM_ALLOC_NOOBJ		page is not associated with an object and
2229  *				should not be exclusive busy
2230  *	VM_ALLOC_SBUSY		shared busy the allocated page
2231  *	VM_ALLOC_WIRED		wire the allocated page
2232  *	VM_ALLOC_ZERO		prefer a zeroed page
2233  */
2234 vm_page_t
2235 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2236     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2237     vm_paddr_t boundary, vm_memattr_t memattr)
2238 {
2239 	struct vm_domainset_iter di;
2240 	vm_page_t m;
2241 	int domain;
2242 
2243 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2244 	do {
2245 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2246 		    npages, low, high, alignment, boundary, memattr);
2247 		if (m != NULL)
2248 			break;
2249 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2250 
2251 	return (m);
2252 }
2253 
2254 vm_page_t
2255 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2256     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2257     vm_paddr_t boundary, vm_memattr_t memattr)
2258 {
2259 	struct vm_domain *vmd;
2260 	vm_page_t m, m_ret, mpred;
2261 	u_int busy_lock, flags, oflags;
2262 
2263 	mpred = NULL;	/* XXX: pacify gcc */
2264 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2265 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2266 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2267 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2268 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2269 	    req));
2270 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2271 	    ("Can't sleep and retry object insertion."));
2272 	if (object != NULL) {
2273 		VM_OBJECT_ASSERT_WLOCKED(object);
2274 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2275 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2276 		    object));
2277 	}
2278 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2279 
2280 	if (object != NULL) {
2281 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2282 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2283 		    ("vm_page_alloc_contig: pindex already allocated"));
2284 	}
2285 
2286 	/*
2287 	 * Can we allocate the pages without the number of free pages falling
2288 	 * below the lower bound for the allocation class?
2289 	 */
2290 	m_ret = NULL;
2291 again:
2292 #if VM_NRESERVLEVEL > 0
2293 	/*
2294 	 * Can we allocate the pages from a reservation?
2295 	 */
2296 	if (vm_object_reserv(object) &&
2297 	    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2298 	    mpred, npages, low, high, alignment, boundary)) != NULL) {
2299 		goto found;
2300 	}
2301 #endif
2302 	vmd = VM_DOMAIN(domain);
2303 	if (vm_domain_allocate(vmd, req, npages)) {
2304 		/*
2305 		 * allocate them from the free page queues.
2306 		 */
2307 		vm_domain_free_lock(vmd);
2308 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2309 		    alignment, boundary);
2310 		vm_domain_free_unlock(vmd);
2311 		if (m_ret == NULL) {
2312 			vm_domain_freecnt_inc(vmd, npages);
2313 #if VM_NRESERVLEVEL > 0
2314 			if (vm_reserv_reclaim_contig(domain, npages, low,
2315 			    high, alignment, boundary))
2316 				goto again;
2317 #endif
2318 		}
2319 	}
2320 	if (m_ret == NULL) {
2321 		if (vm_domain_alloc_fail(vmd, object, req))
2322 			goto again;
2323 		return (NULL);
2324 	}
2325 #if VM_NRESERVLEVEL > 0
2326 found:
2327 #endif
2328 	for (m = m_ret; m < &m_ret[npages]; m++) {
2329 		vm_page_dequeue(m);
2330 		vm_page_alloc_check(m);
2331 	}
2332 
2333 	/*
2334 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2335 	 */
2336 	flags = 0;
2337 	if ((req & VM_ALLOC_ZERO) != 0)
2338 		flags = PG_ZERO;
2339 	if ((req & VM_ALLOC_NODUMP) != 0)
2340 		flags |= PG_NODUMP;
2341 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2342 	    VPO_UNMANAGED : 0;
2343 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2344 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2345 	else if ((req & VM_ALLOC_SBUSY) != 0)
2346 		busy_lock = VPB_SHARERS_WORD(1);
2347 	else
2348 		busy_lock = VPB_UNBUSIED;
2349 	if ((req & VM_ALLOC_WIRED) != 0)
2350 		vm_wire_add(npages);
2351 	if (object != NULL) {
2352 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2353 		    memattr == VM_MEMATTR_DEFAULT)
2354 			memattr = object->memattr;
2355 	}
2356 	for (m = m_ret; m < &m_ret[npages]; m++) {
2357 		m->a.flags = 0;
2358 		m->flags = (m->flags | PG_NODUMP) & flags;
2359 		m->busy_lock = busy_lock;
2360 		if ((req & VM_ALLOC_WIRED) != 0)
2361 			m->ref_count = 1;
2362 		m->a.act_count = 0;
2363 		m->oflags = oflags;
2364 		if (object != NULL) {
2365 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2366 				if ((req & VM_ALLOC_WIRED) != 0)
2367 					vm_wire_sub(npages);
2368 				KASSERT(m->object == NULL,
2369 				    ("page %p has object", m));
2370 				mpred = m;
2371 				for (m = m_ret; m < &m_ret[npages]; m++) {
2372 					if (m <= mpred &&
2373 					    (req & VM_ALLOC_WIRED) != 0)
2374 						m->ref_count = 0;
2375 					m->oflags = VPO_UNMANAGED;
2376 					m->busy_lock = VPB_UNBUSIED;
2377 					/* Don't change PG_ZERO. */
2378 					vm_page_free_toq(m);
2379 				}
2380 				if (req & VM_ALLOC_WAITFAIL) {
2381 					VM_OBJECT_WUNLOCK(object);
2382 					vm_radix_wait();
2383 					VM_OBJECT_WLOCK(object);
2384 				}
2385 				return (NULL);
2386 			}
2387 			mpred = m;
2388 		} else
2389 			m->pindex = pindex;
2390 		if (memattr != VM_MEMATTR_DEFAULT)
2391 			pmap_page_set_memattr(m, memattr);
2392 		pindex++;
2393 	}
2394 	return (m_ret);
2395 }
2396 
2397 /*
2398  * Check a page that has been freshly dequeued from a freelist.
2399  */
2400 static void
2401 vm_page_alloc_check(vm_page_t m)
2402 {
2403 
2404 	KASSERT(m->object == NULL, ("page %p has object", m));
2405 	KASSERT(m->a.queue == PQ_NONE &&
2406 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2407 	    ("page %p has unexpected queue %d, flags %#x",
2408 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2409 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2410 	KASSERT(vm_page_busy_freed(m), ("page %p is not freed", m));
2411 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2412 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2413 	    ("page %p has unexpected memattr %d",
2414 	    m, pmap_page_get_memattr(m)));
2415 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2416 	pmap_vm_page_alloc_check(m);
2417 }
2418 
2419 /*
2420  * 	vm_page_alloc_freelist:
2421  *
2422  *	Allocate a physical page from the specified free page list.
2423  *
2424  *	The caller must always specify an allocation class.
2425  *
2426  *	allocation classes:
2427  *	VM_ALLOC_NORMAL		normal process request
2428  *	VM_ALLOC_SYSTEM		system *really* needs a page
2429  *	VM_ALLOC_INTERRUPT	interrupt time request
2430  *
2431  *	optional allocation flags:
2432  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2433  *				intends to allocate
2434  *	VM_ALLOC_WIRED		wire the allocated page
2435  *	VM_ALLOC_ZERO		prefer a zeroed page
2436  */
2437 vm_page_t
2438 vm_page_alloc_freelist(int freelist, int req)
2439 {
2440 	struct vm_domainset_iter di;
2441 	vm_page_t m;
2442 	int domain;
2443 
2444 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2445 	do {
2446 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2447 		if (m != NULL)
2448 			break;
2449 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2450 
2451 	return (m);
2452 }
2453 
2454 vm_page_t
2455 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2456 {
2457 	struct vm_domain *vmd;
2458 	vm_page_t m;
2459 	u_int flags;
2460 
2461 	m = NULL;
2462 	vmd = VM_DOMAIN(domain);
2463 again:
2464 	if (vm_domain_allocate(vmd, req, 1)) {
2465 		vm_domain_free_lock(vmd);
2466 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2467 		    VM_FREEPOOL_DIRECT, 0);
2468 		vm_domain_free_unlock(vmd);
2469 		if (m == NULL)
2470 			vm_domain_freecnt_inc(vmd, 1);
2471 	}
2472 	if (m == NULL) {
2473 		if (vm_domain_alloc_fail(vmd, NULL, req))
2474 			goto again;
2475 		return (NULL);
2476 	}
2477 	vm_page_dequeue(m);
2478 	vm_page_alloc_check(m);
2479 
2480 	/*
2481 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2482 	 */
2483 	m->a.flags = 0;
2484 	flags = 0;
2485 	if ((req & VM_ALLOC_ZERO) != 0)
2486 		flags = PG_ZERO;
2487 	m->flags &= flags;
2488 	if ((req & VM_ALLOC_WIRED) != 0) {
2489 		vm_wire_add(1);
2490 		m->ref_count = 1;
2491 	}
2492 	/* Unmanaged pages don't use "act_count". */
2493 	m->oflags = VPO_UNMANAGED;
2494 	return (m);
2495 }
2496 
2497 static int
2498 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2499 {
2500 	struct vm_domain *vmd;
2501 	struct vm_pgcache *pgcache;
2502 	int i;
2503 
2504 	pgcache = arg;
2505 	vmd = VM_DOMAIN(pgcache->domain);
2506 
2507 	/*
2508 	 * The page daemon should avoid creating extra memory pressure since its
2509 	 * main purpose is to replenish the store of free pages.
2510 	 */
2511 	if (vmd->vmd_severeset || curproc == pageproc ||
2512 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2513 		return (0);
2514 	domain = vmd->vmd_domain;
2515 	vm_domain_free_lock(vmd);
2516 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2517 	    (vm_page_t *)store);
2518 	vm_domain_free_unlock(vmd);
2519 	if (cnt != i)
2520 		vm_domain_freecnt_inc(vmd, cnt - i);
2521 
2522 	return (i);
2523 }
2524 
2525 static void
2526 vm_page_zone_release(void *arg, void **store, int cnt)
2527 {
2528 	struct vm_domain *vmd;
2529 	struct vm_pgcache *pgcache;
2530 	vm_page_t m;
2531 	int i;
2532 
2533 	pgcache = arg;
2534 	vmd = VM_DOMAIN(pgcache->domain);
2535 	vm_domain_free_lock(vmd);
2536 	for (i = 0; i < cnt; i++) {
2537 		m = (vm_page_t)store[i];
2538 		vm_phys_free_pages(m, 0);
2539 	}
2540 	vm_domain_free_unlock(vmd);
2541 	vm_domain_freecnt_inc(vmd, cnt);
2542 }
2543 
2544 #define	VPSC_ANY	0	/* No restrictions. */
2545 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2546 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2547 
2548 /*
2549  *	vm_page_scan_contig:
2550  *
2551  *	Scan vm_page_array[] between the specified entries "m_start" and
2552  *	"m_end" for a run of contiguous physical pages that satisfy the
2553  *	specified conditions, and return the lowest page in the run.  The
2554  *	specified "alignment" determines the alignment of the lowest physical
2555  *	page in the run.  If the specified "boundary" is non-zero, then the
2556  *	run of physical pages cannot span a physical address that is a
2557  *	multiple of "boundary".
2558  *
2559  *	"m_end" is never dereferenced, so it need not point to a vm_page
2560  *	structure within vm_page_array[].
2561  *
2562  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2563  *	span a hole (or discontiguity) in the physical address space.  Both
2564  *	"alignment" and "boundary" must be a power of two.
2565  */
2566 vm_page_t
2567 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2568     u_long alignment, vm_paddr_t boundary, int options)
2569 {
2570 	vm_object_t object;
2571 	vm_paddr_t pa;
2572 	vm_page_t m, m_run;
2573 #if VM_NRESERVLEVEL > 0
2574 	int level;
2575 #endif
2576 	int m_inc, order, run_ext, run_len;
2577 
2578 	KASSERT(npages > 0, ("npages is 0"));
2579 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2580 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2581 	m_run = NULL;
2582 	run_len = 0;
2583 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2584 		KASSERT((m->flags & PG_MARKER) == 0,
2585 		    ("page %p is PG_MARKER", m));
2586 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2587 		    ("fictitious page %p has invalid ref count", m));
2588 
2589 		/*
2590 		 * If the current page would be the start of a run, check its
2591 		 * physical address against the end, alignment, and boundary
2592 		 * conditions.  If it doesn't satisfy these conditions, either
2593 		 * terminate the scan or advance to the next page that
2594 		 * satisfies the failed condition.
2595 		 */
2596 		if (run_len == 0) {
2597 			KASSERT(m_run == NULL, ("m_run != NULL"));
2598 			if (m + npages > m_end)
2599 				break;
2600 			pa = VM_PAGE_TO_PHYS(m);
2601 			if ((pa & (alignment - 1)) != 0) {
2602 				m_inc = atop(roundup2(pa, alignment) - pa);
2603 				continue;
2604 			}
2605 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2606 			    boundary) != 0) {
2607 				m_inc = atop(roundup2(pa, boundary) - pa);
2608 				continue;
2609 			}
2610 		} else
2611 			KASSERT(m_run != NULL, ("m_run == NULL"));
2612 
2613 retry:
2614 		m_inc = 1;
2615 		if (vm_page_wired(m))
2616 			run_ext = 0;
2617 #if VM_NRESERVLEVEL > 0
2618 		else if ((level = vm_reserv_level(m)) >= 0 &&
2619 		    (options & VPSC_NORESERV) != 0) {
2620 			run_ext = 0;
2621 			/* Advance to the end of the reservation. */
2622 			pa = VM_PAGE_TO_PHYS(m);
2623 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2624 			    pa);
2625 		}
2626 #endif
2627 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2628 			/*
2629 			 * The page is considered eligible for relocation if
2630 			 * and only if it could be laundered or reclaimed by
2631 			 * the page daemon.
2632 			 */
2633 			VM_OBJECT_RLOCK(object);
2634 			if (object != m->object) {
2635 				VM_OBJECT_RUNLOCK(object);
2636 				goto retry;
2637 			}
2638 			/* Don't care: PG_NODUMP, PG_ZERO. */
2639 			if (object->type != OBJT_DEFAULT &&
2640 			    (object->flags & OBJ_SWAP) == 0 &&
2641 			    object->type != OBJT_VNODE) {
2642 				run_ext = 0;
2643 #if VM_NRESERVLEVEL > 0
2644 			} else if ((options & VPSC_NOSUPER) != 0 &&
2645 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2646 				run_ext = 0;
2647 				/* Advance to the end of the superpage. */
2648 				pa = VM_PAGE_TO_PHYS(m);
2649 				m_inc = atop(roundup2(pa + 1,
2650 				    vm_reserv_size(level)) - pa);
2651 #endif
2652 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2653 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) {
2654 				/*
2655 				 * The page is allocated but eligible for
2656 				 * relocation.  Extend the current run by one
2657 				 * page.
2658 				 */
2659 				KASSERT(pmap_page_get_memattr(m) ==
2660 				    VM_MEMATTR_DEFAULT,
2661 				    ("page %p has an unexpected memattr", m));
2662 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2663 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2664 				    ("page %p has unexpected oflags", m));
2665 				/* Don't care: PGA_NOSYNC. */
2666 				run_ext = 1;
2667 			} else
2668 				run_ext = 0;
2669 			VM_OBJECT_RUNLOCK(object);
2670 #if VM_NRESERVLEVEL > 0
2671 		} else if (level >= 0) {
2672 			/*
2673 			 * The page is reserved but not yet allocated.  In
2674 			 * other words, it is still free.  Extend the current
2675 			 * run by one page.
2676 			 */
2677 			run_ext = 1;
2678 #endif
2679 		} else if ((order = m->order) < VM_NFREEORDER) {
2680 			/*
2681 			 * The page is enqueued in the physical memory
2682 			 * allocator's free page queues.  Moreover, it is the
2683 			 * first page in a power-of-two-sized run of
2684 			 * contiguous free pages.  Add these pages to the end
2685 			 * of the current run, and jump ahead.
2686 			 */
2687 			run_ext = 1 << order;
2688 			m_inc = 1 << order;
2689 		} else {
2690 			/*
2691 			 * Skip the page for one of the following reasons: (1)
2692 			 * It is enqueued in the physical memory allocator's
2693 			 * free page queues.  However, it is not the first
2694 			 * page in a run of contiguous free pages.  (This case
2695 			 * rarely occurs because the scan is performed in
2696 			 * ascending order.) (2) It is not reserved, and it is
2697 			 * transitioning from free to allocated.  (Conversely,
2698 			 * the transition from allocated to free for managed
2699 			 * pages is blocked by the page busy lock.) (3) It is
2700 			 * allocated but not contained by an object and not
2701 			 * wired, e.g., allocated by Xen's balloon driver.
2702 			 */
2703 			run_ext = 0;
2704 		}
2705 
2706 		/*
2707 		 * Extend or reset the current run of pages.
2708 		 */
2709 		if (run_ext > 0) {
2710 			if (run_len == 0)
2711 				m_run = m;
2712 			run_len += run_ext;
2713 		} else {
2714 			if (run_len > 0) {
2715 				m_run = NULL;
2716 				run_len = 0;
2717 			}
2718 		}
2719 	}
2720 	if (run_len >= npages)
2721 		return (m_run);
2722 	return (NULL);
2723 }
2724 
2725 /*
2726  *	vm_page_reclaim_run:
2727  *
2728  *	Try to relocate each of the allocated virtual pages within the
2729  *	specified run of physical pages to a new physical address.  Free the
2730  *	physical pages underlying the relocated virtual pages.  A virtual page
2731  *	is relocatable if and only if it could be laundered or reclaimed by
2732  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2733  *	physical address above "high".
2734  *
2735  *	Returns 0 if every physical page within the run was already free or
2736  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2737  *	value indicating why the last attempt to relocate a virtual page was
2738  *	unsuccessful.
2739  *
2740  *	"req_class" must be an allocation class.
2741  */
2742 static int
2743 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2744     vm_paddr_t high)
2745 {
2746 	struct vm_domain *vmd;
2747 	struct spglist free;
2748 	vm_object_t object;
2749 	vm_paddr_t pa;
2750 	vm_page_t m, m_end, m_new;
2751 	int error, order, req;
2752 
2753 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2754 	    ("req_class is not an allocation class"));
2755 	SLIST_INIT(&free);
2756 	error = 0;
2757 	m = m_run;
2758 	m_end = m_run + npages;
2759 	for (; error == 0 && m < m_end; m++) {
2760 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2761 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2762 
2763 		/*
2764 		 * Racily check for wirings.  Races are handled once the object
2765 		 * lock is held and the page is unmapped.
2766 		 */
2767 		if (vm_page_wired(m))
2768 			error = EBUSY;
2769 		else if ((object = atomic_load_ptr(&m->object)) != NULL) {
2770 			/*
2771 			 * The page is relocated if and only if it could be
2772 			 * laundered or reclaimed by the page daemon.
2773 			 */
2774 			VM_OBJECT_WLOCK(object);
2775 			/* Don't care: PG_NODUMP, PG_ZERO. */
2776 			if (m->object != object ||
2777 			    (object->type != OBJT_DEFAULT &&
2778 			    (object->flags & OBJ_SWAP) == 0 &&
2779 			    object->type != OBJT_VNODE))
2780 				error = EINVAL;
2781 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2782 				error = EINVAL;
2783 			else if (vm_page_queue(m) != PQ_NONE &&
2784 			    vm_page_tryxbusy(m) != 0) {
2785 				if (vm_page_wired(m)) {
2786 					vm_page_xunbusy(m);
2787 					error = EBUSY;
2788 					goto unlock;
2789 				}
2790 				KASSERT(pmap_page_get_memattr(m) ==
2791 				    VM_MEMATTR_DEFAULT,
2792 				    ("page %p has an unexpected memattr", m));
2793 				KASSERT(m->oflags == 0,
2794 				    ("page %p has unexpected oflags", m));
2795 				/* Don't care: PGA_NOSYNC. */
2796 				if (!vm_page_none_valid(m)) {
2797 					/*
2798 					 * First, try to allocate a new page
2799 					 * that is above "high".  Failing
2800 					 * that, try to allocate a new page
2801 					 * that is below "m_run".  Allocate
2802 					 * the new page between the end of
2803 					 * "m_run" and "high" only as a last
2804 					 * resort.
2805 					 */
2806 					req = req_class | VM_ALLOC_NOOBJ;
2807 					if ((m->flags & PG_NODUMP) != 0)
2808 						req |= VM_ALLOC_NODUMP;
2809 					if (trunc_page(high) !=
2810 					    ~(vm_paddr_t)PAGE_MASK) {
2811 						m_new = vm_page_alloc_contig(
2812 						    NULL, 0, req, 1,
2813 						    round_page(high),
2814 						    ~(vm_paddr_t)0,
2815 						    PAGE_SIZE, 0,
2816 						    VM_MEMATTR_DEFAULT);
2817 					} else
2818 						m_new = NULL;
2819 					if (m_new == NULL) {
2820 						pa = VM_PAGE_TO_PHYS(m_run);
2821 						m_new = vm_page_alloc_contig(
2822 						    NULL, 0, req, 1,
2823 						    0, pa - 1, PAGE_SIZE, 0,
2824 						    VM_MEMATTR_DEFAULT);
2825 					}
2826 					if (m_new == NULL) {
2827 						pa += ptoa(npages);
2828 						m_new = vm_page_alloc_contig(
2829 						    NULL, 0, req, 1,
2830 						    pa, high, PAGE_SIZE, 0,
2831 						    VM_MEMATTR_DEFAULT);
2832 					}
2833 					if (m_new == NULL) {
2834 						vm_page_xunbusy(m);
2835 						error = ENOMEM;
2836 						goto unlock;
2837 					}
2838 
2839 					/*
2840 					 * Unmap the page and check for new
2841 					 * wirings that may have been acquired
2842 					 * through a pmap lookup.
2843 					 */
2844 					if (object->ref_count != 0 &&
2845 					    !vm_page_try_remove_all(m)) {
2846 						vm_page_xunbusy(m);
2847 						vm_page_free(m_new);
2848 						error = EBUSY;
2849 						goto unlock;
2850 					}
2851 
2852 					/*
2853 					 * Replace "m" with the new page.  For
2854 					 * vm_page_replace(), "m" must be busy
2855 					 * and dequeued.  Finally, change "m"
2856 					 * as if vm_page_free() was called.
2857 					 */
2858 					m_new->a.flags = m->a.flags &
2859 					    ~PGA_QUEUE_STATE_MASK;
2860 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2861 					    ("page %p is managed", m_new));
2862 					m_new->oflags = 0;
2863 					pmap_copy_page(m, m_new);
2864 					m_new->valid = m->valid;
2865 					m_new->dirty = m->dirty;
2866 					m->flags &= ~PG_ZERO;
2867 					vm_page_dequeue(m);
2868 					if (vm_page_replace_hold(m_new, object,
2869 					    m->pindex, m) &&
2870 					    vm_page_free_prep(m))
2871 						SLIST_INSERT_HEAD(&free, m,
2872 						    plinks.s.ss);
2873 
2874 					/*
2875 					 * The new page must be deactivated
2876 					 * before the object is unlocked.
2877 					 */
2878 					vm_page_deactivate(m_new);
2879 				} else {
2880 					m->flags &= ~PG_ZERO;
2881 					vm_page_dequeue(m);
2882 					if (vm_page_free_prep(m))
2883 						SLIST_INSERT_HEAD(&free, m,
2884 						    plinks.s.ss);
2885 					KASSERT(m->dirty == 0,
2886 					    ("page %p is dirty", m));
2887 				}
2888 			} else
2889 				error = EBUSY;
2890 unlock:
2891 			VM_OBJECT_WUNLOCK(object);
2892 		} else {
2893 			MPASS(vm_page_domain(m) == domain);
2894 			vmd = VM_DOMAIN(domain);
2895 			vm_domain_free_lock(vmd);
2896 			order = m->order;
2897 			if (order < VM_NFREEORDER) {
2898 				/*
2899 				 * The page is enqueued in the physical memory
2900 				 * allocator's free page queues.  Moreover, it
2901 				 * is the first page in a power-of-two-sized
2902 				 * run of contiguous free pages.  Jump ahead
2903 				 * to the last page within that run, and
2904 				 * continue from there.
2905 				 */
2906 				m += (1 << order) - 1;
2907 			}
2908 #if VM_NRESERVLEVEL > 0
2909 			else if (vm_reserv_is_page_free(m))
2910 				order = 0;
2911 #endif
2912 			vm_domain_free_unlock(vmd);
2913 			if (order == VM_NFREEORDER)
2914 				error = EINVAL;
2915 		}
2916 	}
2917 	if ((m = SLIST_FIRST(&free)) != NULL) {
2918 		int cnt;
2919 
2920 		vmd = VM_DOMAIN(domain);
2921 		cnt = 0;
2922 		vm_domain_free_lock(vmd);
2923 		do {
2924 			MPASS(vm_page_domain(m) == domain);
2925 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2926 			vm_phys_free_pages(m, 0);
2927 			cnt++;
2928 		} while ((m = SLIST_FIRST(&free)) != NULL);
2929 		vm_domain_free_unlock(vmd);
2930 		vm_domain_freecnt_inc(vmd, cnt);
2931 	}
2932 	return (error);
2933 }
2934 
2935 #define	NRUNS	16
2936 
2937 CTASSERT(powerof2(NRUNS));
2938 
2939 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2940 
2941 #define	MIN_RECLAIM	8
2942 
2943 /*
2944  *	vm_page_reclaim_contig:
2945  *
2946  *	Reclaim allocated, contiguous physical memory satisfying the specified
2947  *	conditions by relocating the virtual pages using that physical memory.
2948  *	Returns true if reclamation is successful and false otherwise.  Since
2949  *	relocation requires the allocation of physical pages, reclamation may
2950  *	fail due to a shortage of free pages.  When reclamation fails, callers
2951  *	are expected to perform vm_wait() before retrying a failed allocation
2952  *	operation, e.g., vm_page_alloc_contig().
2953  *
2954  *	The caller must always specify an allocation class through "req".
2955  *
2956  *	allocation classes:
2957  *	VM_ALLOC_NORMAL		normal process request
2958  *	VM_ALLOC_SYSTEM		system *really* needs a page
2959  *	VM_ALLOC_INTERRUPT	interrupt time request
2960  *
2961  *	The optional allocation flags are ignored.
2962  *
2963  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2964  *	must be a power of two.
2965  */
2966 bool
2967 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2968     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2969 {
2970 	struct vm_domain *vmd;
2971 	vm_paddr_t curr_low;
2972 	vm_page_t m_run, m_runs[NRUNS];
2973 	u_long count, minalign, reclaimed;
2974 	int error, i, options, req_class;
2975 
2976 	KASSERT(npages > 0, ("npages is 0"));
2977 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2978 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2979 
2980 	/*
2981 	 * The caller will attempt an allocation after some runs have been
2982 	 * reclaimed and added to the vm_phys buddy lists.  Due to limitations
2983 	 * of vm_phys_alloc_contig(), round up the requested length to the next
2984 	 * power of two or maximum chunk size, and ensure that each run is
2985 	 * suitably aligned.
2986 	 */
2987 	minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
2988 	npages = roundup2(npages, minalign);
2989 	if (alignment < ptoa(minalign))
2990 		alignment = ptoa(minalign);
2991 
2992 	/*
2993 	 * The page daemon is allowed to dig deeper into the free page list.
2994 	 */
2995 	req_class = req & VM_ALLOC_CLASS_MASK;
2996 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2997 		req_class = VM_ALLOC_SYSTEM;
2998 
2999 	/*
3000 	 * Return if the number of free pages cannot satisfy the requested
3001 	 * allocation.
3002 	 */
3003 	vmd = VM_DOMAIN(domain);
3004 	count = vmd->vmd_free_count;
3005 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
3006 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
3007 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
3008 		return (false);
3009 
3010 	/*
3011 	 * Scan up to three times, relaxing the restrictions ("options") on
3012 	 * the reclamation of reservations and superpages each time.
3013 	 */
3014 	for (options = VPSC_NORESERV;;) {
3015 		/*
3016 		 * Find the highest runs that satisfy the given constraints
3017 		 * and restrictions, and record them in "m_runs".
3018 		 */
3019 		curr_low = low;
3020 		count = 0;
3021 		for (;;) {
3022 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
3023 			    high, alignment, boundary, options);
3024 			if (m_run == NULL)
3025 				break;
3026 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
3027 			m_runs[RUN_INDEX(count)] = m_run;
3028 			count++;
3029 		}
3030 
3031 		/*
3032 		 * Reclaim the highest runs in LIFO (descending) order until
3033 		 * the number of reclaimed pages, "reclaimed", is at least
3034 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
3035 		 * reclamation is idempotent, and runs will (likely) recur
3036 		 * from one scan to the next as restrictions are relaxed.
3037 		 */
3038 		reclaimed = 0;
3039 		for (i = 0; count > 0 && i < NRUNS; i++) {
3040 			count--;
3041 			m_run = m_runs[RUN_INDEX(count)];
3042 			error = vm_page_reclaim_run(req_class, domain, npages,
3043 			    m_run, high);
3044 			if (error == 0) {
3045 				reclaimed += npages;
3046 				if (reclaimed >= MIN_RECLAIM)
3047 					return (true);
3048 			}
3049 		}
3050 
3051 		/*
3052 		 * Either relax the restrictions on the next scan or return if
3053 		 * the last scan had no restrictions.
3054 		 */
3055 		if (options == VPSC_NORESERV)
3056 			options = VPSC_NOSUPER;
3057 		else if (options == VPSC_NOSUPER)
3058 			options = VPSC_ANY;
3059 		else if (options == VPSC_ANY)
3060 			return (reclaimed != 0);
3061 	}
3062 }
3063 
3064 bool
3065 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3066     u_long alignment, vm_paddr_t boundary)
3067 {
3068 	struct vm_domainset_iter di;
3069 	int domain;
3070 	bool ret;
3071 
3072 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3073 	do {
3074 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3075 		    high, alignment, boundary);
3076 		if (ret)
3077 			break;
3078 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3079 
3080 	return (ret);
3081 }
3082 
3083 /*
3084  * Set the domain in the appropriate page level domainset.
3085  */
3086 void
3087 vm_domain_set(struct vm_domain *vmd)
3088 {
3089 
3090 	mtx_lock(&vm_domainset_lock);
3091 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3092 		vmd->vmd_minset = 1;
3093 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3094 	}
3095 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3096 		vmd->vmd_severeset = 1;
3097 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3098 	}
3099 	mtx_unlock(&vm_domainset_lock);
3100 }
3101 
3102 /*
3103  * Clear the domain from the appropriate page level domainset.
3104  */
3105 void
3106 vm_domain_clear(struct vm_domain *vmd)
3107 {
3108 
3109 	mtx_lock(&vm_domainset_lock);
3110 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3111 		vmd->vmd_minset = 0;
3112 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3113 		if (vm_min_waiters != 0) {
3114 			vm_min_waiters = 0;
3115 			wakeup(&vm_min_domains);
3116 		}
3117 	}
3118 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3119 		vmd->vmd_severeset = 0;
3120 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3121 		if (vm_severe_waiters != 0) {
3122 			vm_severe_waiters = 0;
3123 			wakeup(&vm_severe_domains);
3124 		}
3125 	}
3126 
3127 	/*
3128 	 * If pageout daemon needs pages, then tell it that there are
3129 	 * some free.
3130 	 */
3131 	if (vmd->vmd_pageout_pages_needed &&
3132 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3133 		wakeup(&vmd->vmd_pageout_pages_needed);
3134 		vmd->vmd_pageout_pages_needed = 0;
3135 	}
3136 
3137 	/* See comments in vm_wait_doms(). */
3138 	if (vm_pageproc_waiters) {
3139 		vm_pageproc_waiters = 0;
3140 		wakeup(&vm_pageproc_waiters);
3141 	}
3142 	mtx_unlock(&vm_domainset_lock);
3143 }
3144 
3145 /*
3146  * Wait for free pages to exceed the min threshold globally.
3147  */
3148 void
3149 vm_wait_min(void)
3150 {
3151 
3152 	mtx_lock(&vm_domainset_lock);
3153 	while (vm_page_count_min()) {
3154 		vm_min_waiters++;
3155 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3156 	}
3157 	mtx_unlock(&vm_domainset_lock);
3158 }
3159 
3160 /*
3161  * Wait for free pages to exceed the severe threshold globally.
3162  */
3163 void
3164 vm_wait_severe(void)
3165 {
3166 
3167 	mtx_lock(&vm_domainset_lock);
3168 	while (vm_page_count_severe()) {
3169 		vm_severe_waiters++;
3170 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3171 		    "vmwait", 0);
3172 	}
3173 	mtx_unlock(&vm_domainset_lock);
3174 }
3175 
3176 u_int
3177 vm_wait_count(void)
3178 {
3179 
3180 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3181 }
3182 
3183 int
3184 vm_wait_doms(const domainset_t *wdoms, int mflags)
3185 {
3186 	int error;
3187 
3188 	error = 0;
3189 
3190 	/*
3191 	 * We use racey wakeup synchronization to avoid expensive global
3192 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3193 	 * To handle this, we only sleep for one tick in this instance.  It
3194 	 * is expected that most allocations for the pageproc will come from
3195 	 * kmem or vm_page_grab* which will use the more specific and
3196 	 * race-free vm_wait_domain().
3197 	 */
3198 	if (curproc == pageproc) {
3199 		mtx_lock(&vm_domainset_lock);
3200 		vm_pageproc_waiters++;
3201 		error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
3202 		    PVM | PDROP | mflags, "pageprocwait", 1);
3203 	} else {
3204 		/*
3205 		 * XXX Ideally we would wait only until the allocation could
3206 		 * be satisfied.  This condition can cause new allocators to
3207 		 * consume all freed pages while old allocators wait.
3208 		 */
3209 		mtx_lock(&vm_domainset_lock);
3210 		if (vm_page_count_min_set(wdoms)) {
3211 			vm_min_waiters++;
3212 			error = msleep(&vm_min_domains, &vm_domainset_lock,
3213 			    PVM | PDROP | mflags, "vmwait", 0);
3214 		} else
3215 			mtx_unlock(&vm_domainset_lock);
3216 	}
3217 	return (error);
3218 }
3219 
3220 /*
3221  *	vm_wait_domain:
3222  *
3223  *	Sleep until free pages are available for allocation.
3224  *	- Called in various places after failed memory allocations.
3225  */
3226 void
3227 vm_wait_domain(int domain)
3228 {
3229 	struct vm_domain *vmd;
3230 	domainset_t wdom;
3231 
3232 	vmd = VM_DOMAIN(domain);
3233 	vm_domain_free_assert_unlocked(vmd);
3234 
3235 	if (curproc == pageproc) {
3236 		mtx_lock(&vm_domainset_lock);
3237 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3238 			vmd->vmd_pageout_pages_needed = 1;
3239 			msleep(&vmd->vmd_pageout_pages_needed,
3240 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3241 		} else
3242 			mtx_unlock(&vm_domainset_lock);
3243 	} else {
3244 		if (pageproc == NULL)
3245 			panic("vm_wait in early boot");
3246 		DOMAINSET_ZERO(&wdom);
3247 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3248 		vm_wait_doms(&wdom, 0);
3249 	}
3250 }
3251 
3252 static int
3253 vm_wait_flags(vm_object_t obj, int mflags)
3254 {
3255 	struct domainset *d;
3256 
3257 	d = NULL;
3258 
3259 	/*
3260 	 * Carefully fetch pointers only once: the struct domainset
3261 	 * itself is ummutable but the pointer might change.
3262 	 */
3263 	if (obj != NULL)
3264 		d = obj->domain.dr_policy;
3265 	if (d == NULL)
3266 		d = curthread->td_domain.dr_policy;
3267 
3268 	return (vm_wait_doms(&d->ds_mask, mflags));
3269 }
3270 
3271 /*
3272  *	vm_wait:
3273  *
3274  *	Sleep until free pages are available for allocation in the
3275  *	affinity domains of the obj.  If obj is NULL, the domain set
3276  *	for the calling thread is used.
3277  *	Called in various places after failed memory allocations.
3278  */
3279 void
3280 vm_wait(vm_object_t obj)
3281 {
3282 	(void)vm_wait_flags(obj, 0);
3283 }
3284 
3285 int
3286 vm_wait_intr(vm_object_t obj)
3287 {
3288 	return (vm_wait_flags(obj, PCATCH));
3289 }
3290 
3291 /*
3292  *	vm_domain_alloc_fail:
3293  *
3294  *	Called when a page allocation function fails.  Informs the
3295  *	pagedaemon and performs the requested wait.  Requires the
3296  *	domain_free and object lock on entry.  Returns with the
3297  *	object lock held and free lock released.  Returns an error when
3298  *	retry is necessary.
3299  *
3300  */
3301 static int
3302 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3303 {
3304 
3305 	vm_domain_free_assert_unlocked(vmd);
3306 
3307 	atomic_add_int(&vmd->vmd_pageout_deficit,
3308 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3309 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3310 		if (object != NULL)
3311 			VM_OBJECT_WUNLOCK(object);
3312 		vm_wait_domain(vmd->vmd_domain);
3313 		if (object != NULL)
3314 			VM_OBJECT_WLOCK(object);
3315 		if (req & VM_ALLOC_WAITOK)
3316 			return (EAGAIN);
3317 	}
3318 
3319 	return (0);
3320 }
3321 
3322 /*
3323  *	vm_waitpfault:
3324  *
3325  *	Sleep until free pages are available for allocation.
3326  *	- Called only in vm_fault so that processes page faulting
3327  *	  can be easily tracked.
3328  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3329  *	  processes will be able to grab memory first.  Do not change
3330  *	  this balance without careful testing first.
3331  */
3332 void
3333 vm_waitpfault(struct domainset *dset, int timo)
3334 {
3335 
3336 	/*
3337 	 * XXX Ideally we would wait only until the allocation could
3338 	 * be satisfied.  This condition can cause new allocators to
3339 	 * consume all freed pages while old allocators wait.
3340 	 */
3341 	mtx_lock(&vm_domainset_lock);
3342 	if (vm_page_count_min_set(&dset->ds_mask)) {
3343 		vm_min_waiters++;
3344 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3345 		    "pfault", timo);
3346 	} else
3347 		mtx_unlock(&vm_domainset_lock);
3348 }
3349 
3350 static struct vm_pagequeue *
3351 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3352 {
3353 
3354 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3355 }
3356 
3357 #ifdef INVARIANTS
3358 static struct vm_pagequeue *
3359 vm_page_pagequeue(vm_page_t m)
3360 {
3361 
3362 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3363 }
3364 #endif
3365 
3366 static __always_inline bool
3367 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3368 {
3369 	vm_page_astate_t tmp;
3370 
3371 	tmp = *old;
3372 	do {
3373 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3374 			return (true);
3375 		counter_u64_add(pqstate_commit_retries, 1);
3376 	} while (old->_bits == tmp._bits);
3377 
3378 	return (false);
3379 }
3380 
3381 /*
3382  * Do the work of committing a queue state update that moves the page out of
3383  * its current queue.
3384  */
3385 static bool
3386 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3387     vm_page_astate_t *old, vm_page_astate_t new)
3388 {
3389 	vm_page_t next;
3390 
3391 	vm_pagequeue_assert_locked(pq);
3392 	KASSERT(vm_page_pagequeue(m) == pq,
3393 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3394 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3395 	    ("%s: invalid queue indices %d %d",
3396 	    __func__, old->queue, new.queue));
3397 
3398 	/*
3399 	 * Once the queue index of the page changes there is nothing
3400 	 * synchronizing with further updates to the page's physical
3401 	 * queue state.  Therefore we must speculatively remove the page
3402 	 * from the queue now and be prepared to roll back if the queue
3403 	 * state update fails.  If the page is not physically enqueued then
3404 	 * we just update its queue index.
3405 	 */
3406 	if ((old->flags & PGA_ENQUEUED) != 0) {
3407 		new.flags &= ~PGA_ENQUEUED;
3408 		next = TAILQ_NEXT(m, plinks.q);
3409 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3410 		vm_pagequeue_cnt_dec(pq);
3411 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3412 			if (next == NULL)
3413 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3414 			else
3415 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3416 			vm_pagequeue_cnt_inc(pq);
3417 			return (false);
3418 		} else {
3419 			return (true);
3420 		}
3421 	} else {
3422 		return (vm_page_pqstate_fcmpset(m, old, new));
3423 	}
3424 }
3425 
3426 static bool
3427 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3428     vm_page_astate_t new)
3429 {
3430 	struct vm_pagequeue *pq;
3431 	vm_page_astate_t as;
3432 	bool ret;
3433 
3434 	pq = _vm_page_pagequeue(m, old->queue);
3435 
3436 	/*
3437 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3438 	 * corresponding page queue lock is held.
3439 	 */
3440 	vm_pagequeue_lock(pq);
3441 	as = vm_page_astate_load(m);
3442 	if (__predict_false(as._bits != old->_bits)) {
3443 		*old = as;
3444 		ret = false;
3445 	} else {
3446 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3447 	}
3448 	vm_pagequeue_unlock(pq);
3449 	return (ret);
3450 }
3451 
3452 /*
3453  * Commit a queue state update that enqueues or requeues a page.
3454  */
3455 static bool
3456 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3457     vm_page_astate_t *old, vm_page_astate_t new)
3458 {
3459 	struct vm_domain *vmd;
3460 
3461 	vm_pagequeue_assert_locked(pq);
3462 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3463 	    ("%s: invalid queue indices %d %d",
3464 	    __func__, old->queue, new.queue));
3465 
3466 	new.flags |= PGA_ENQUEUED;
3467 	if (!vm_page_pqstate_fcmpset(m, old, new))
3468 		return (false);
3469 
3470 	if ((old->flags & PGA_ENQUEUED) != 0)
3471 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3472 	else
3473 		vm_pagequeue_cnt_inc(pq);
3474 
3475 	/*
3476 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3477 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3478 	 * applied, even if it was set first.
3479 	 */
3480 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3481 		vmd = vm_pagequeue_domain(m);
3482 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3483 		    ("%s: invalid page queue for page %p", __func__, m));
3484 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3485 	} else {
3486 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3487 	}
3488 	return (true);
3489 }
3490 
3491 /*
3492  * Commit a queue state update that encodes a request for a deferred queue
3493  * operation.
3494  */
3495 static bool
3496 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3497     vm_page_astate_t new)
3498 {
3499 
3500 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3501 	    ("%s: invalid state, queue %d flags %x",
3502 	    __func__, new.queue, new.flags));
3503 
3504 	if (old->_bits != new._bits &&
3505 	    !vm_page_pqstate_fcmpset(m, old, new))
3506 		return (false);
3507 	vm_page_pqbatch_submit(m, new.queue);
3508 	return (true);
3509 }
3510 
3511 /*
3512  * A generic queue state update function.  This handles more cases than the
3513  * specialized functions above.
3514  */
3515 bool
3516 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3517 {
3518 
3519 	if (old->_bits == new._bits)
3520 		return (true);
3521 
3522 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3523 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3524 			return (false);
3525 		if (new.queue != PQ_NONE)
3526 			vm_page_pqbatch_submit(m, new.queue);
3527 	} else {
3528 		if (!vm_page_pqstate_fcmpset(m, old, new))
3529 			return (false);
3530 		if (new.queue != PQ_NONE &&
3531 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3532 			vm_page_pqbatch_submit(m, new.queue);
3533 	}
3534 	return (true);
3535 }
3536 
3537 /*
3538  * Apply deferred queue state updates to a page.
3539  */
3540 static inline void
3541 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3542 {
3543 	vm_page_astate_t new, old;
3544 
3545 	CRITICAL_ASSERT(curthread);
3546 	vm_pagequeue_assert_locked(pq);
3547 	KASSERT(queue < PQ_COUNT,
3548 	    ("%s: invalid queue index %d", __func__, queue));
3549 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3550 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3551 
3552 	for (old = vm_page_astate_load(m);;) {
3553 		if (__predict_false(old.queue != queue ||
3554 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3555 			counter_u64_add(queue_nops, 1);
3556 			break;
3557 		}
3558 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3559 		    ("%s: page %p is unmanaged", __func__, m));
3560 
3561 		new = old;
3562 		if ((old.flags & PGA_DEQUEUE) != 0) {
3563 			new.flags &= ~PGA_QUEUE_OP_MASK;
3564 			new.queue = PQ_NONE;
3565 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3566 			    m, &old, new))) {
3567 				counter_u64_add(queue_ops, 1);
3568 				break;
3569 			}
3570 		} else {
3571 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3572 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3573 			    m, &old, new))) {
3574 				counter_u64_add(queue_ops, 1);
3575 				break;
3576 			}
3577 		}
3578 	}
3579 }
3580 
3581 static void
3582 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3583     uint8_t queue)
3584 {
3585 	int i;
3586 
3587 	for (i = 0; i < bq->bq_cnt; i++)
3588 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3589 	vm_batchqueue_init(bq);
3590 }
3591 
3592 /*
3593  *	vm_page_pqbatch_submit:		[ internal use only ]
3594  *
3595  *	Enqueue a page in the specified page queue's batched work queue.
3596  *	The caller must have encoded the requested operation in the page
3597  *	structure's a.flags field.
3598  */
3599 void
3600 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3601 {
3602 	struct vm_batchqueue *bq;
3603 	struct vm_pagequeue *pq;
3604 	int domain;
3605 
3606 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3607 
3608 	domain = vm_page_domain(m);
3609 	critical_enter();
3610 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3611 	if (vm_batchqueue_insert(bq, m)) {
3612 		critical_exit();
3613 		return;
3614 	}
3615 	critical_exit();
3616 
3617 	pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
3618 	vm_pagequeue_lock(pq);
3619 	critical_enter();
3620 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3621 	vm_pqbatch_process(pq, bq, queue);
3622 	vm_pqbatch_process_page(pq, m, queue);
3623 	vm_pagequeue_unlock(pq);
3624 	critical_exit();
3625 }
3626 
3627 /*
3628  *	vm_page_pqbatch_drain:		[ internal use only ]
3629  *
3630  *	Force all per-CPU page queue batch queues to be drained.  This is
3631  *	intended for use in severe memory shortages, to ensure that pages
3632  *	do not remain stuck in the batch queues.
3633  */
3634 void
3635 vm_page_pqbatch_drain(void)
3636 {
3637 	struct thread *td;
3638 	struct vm_domain *vmd;
3639 	struct vm_pagequeue *pq;
3640 	int cpu, domain, queue;
3641 
3642 	td = curthread;
3643 	CPU_FOREACH(cpu) {
3644 		thread_lock(td);
3645 		sched_bind(td, cpu);
3646 		thread_unlock(td);
3647 
3648 		for (domain = 0; domain < vm_ndomains; domain++) {
3649 			vmd = VM_DOMAIN(domain);
3650 			for (queue = 0; queue < PQ_COUNT; queue++) {
3651 				pq = &vmd->vmd_pagequeues[queue];
3652 				vm_pagequeue_lock(pq);
3653 				critical_enter();
3654 				vm_pqbatch_process(pq,
3655 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3656 				critical_exit();
3657 				vm_pagequeue_unlock(pq);
3658 			}
3659 		}
3660 	}
3661 	thread_lock(td);
3662 	sched_unbind(td);
3663 	thread_unlock(td);
3664 }
3665 
3666 /*
3667  *	vm_page_dequeue_deferred:	[ internal use only ]
3668  *
3669  *	Request removal of the given page from its current page
3670  *	queue.  Physical removal from the queue may be deferred
3671  *	indefinitely.
3672  */
3673 void
3674 vm_page_dequeue_deferred(vm_page_t m)
3675 {
3676 	vm_page_astate_t new, old;
3677 
3678 	old = vm_page_astate_load(m);
3679 	do {
3680 		if (old.queue == PQ_NONE) {
3681 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3682 			    ("%s: page %p has unexpected queue state",
3683 			    __func__, m));
3684 			break;
3685 		}
3686 		new = old;
3687 		new.flags |= PGA_DEQUEUE;
3688 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3689 }
3690 
3691 /*
3692  *	vm_page_dequeue:
3693  *
3694  *	Remove the page from whichever page queue it's in, if any, before
3695  *	returning.
3696  */
3697 void
3698 vm_page_dequeue(vm_page_t m)
3699 {
3700 	vm_page_astate_t new, old;
3701 
3702 	old = vm_page_astate_load(m);
3703 	do {
3704 		if (old.queue == PQ_NONE) {
3705 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3706 			    ("%s: page %p has unexpected queue state",
3707 			    __func__, m));
3708 			break;
3709 		}
3710 		new = old;
3711 		new.flags &= ~PGA_QUEUE_OP_MASK;
3712 		new.queue = PQ_NONE;
3713 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3714 
3715 }
3716 
3717 /*
3718  * Schedule the given page for insertion into the specified page queue.
3719  * Physical insertion of the page may be deferred indefinitely.
3720  */
3721 static void
3722 vm_page_enqueue(vm_page_t m, uint8_t queue)
3723 {
3724 
3725 	KASSERT(m->a.queue == PQ_NONE &&
3726 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3727 	    ("%s: page %p is already enqueued", __func__, m));
3728 	KASSERT(m->ref_count > 0,
3729 	    ("%s: page %p does not carry any references", __func__, m));
3730 
3731 	m->a.queue = queue;
3732 	if ((m->a.flags & PGA_REQUEUE) == 0)
3733 		vm_page_aflag_set(m, PGA_REQUEUE);
3734 	vm_page_pqbatch_submit(m, queue);
3735 }
3736 
3737 /*
3738  *	vm_page_free_prep:
3739  *
3740  *	Prepares the given page to be put on the free list,
3741  *	disassociating it from any VM object. The caller may return
3742  *	the page to the free list only if this function returns true.
3743  *
3744  *	The object, if it exists, must be locked, and then the page must
3745  *	be xbusy.  Otherwise the page must be not busied.  A managed
3746  *	page must be unmapped.
3747  */
3748 static bool
3749 vm_page_free_prep(vm_page_t m)
3750 {
3751 
3752 	/*
3753 	 * Synchronize with threads that have dropped a reference to this
3754 	 * page.
3755 	 */
3756 	atomic_thread_fence_acq();
3757 
3758 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3759 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3760 		uint64_t *p;
3761 		int i;
3762 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3763 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3764 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3765 			    m, i, (uintmax_t)*p));
3766 	}
3767 #endif
3768 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3769 		KASSERT(!pmap_page_is_mapped(m),
3770 		    ("vm_page_free_prep: freeing mapped page %p", m));
3771 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3772 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3773 	} else {
3774 		KASSERT(m->a.queue == PQ_NONE,
3775 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3776 	}
3777 	VM_CNT_INC(v_tfree);
3778 
3779 	if (m->object != NULL) {
3780 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3781 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
3782 		    ("vm_page_free_prep: managed flag mismatch for page %p",
3783 		    m));
3784 		vm_page_assert_xbusied(m);
3785 
3786 		/*
3787 		 * The object reference can be released without an atomic
3788 		 * operation.
3789 		 */
3790 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3791 		    m->ref_count == VPRC_OBJREF,
3792 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3793 		    m, m->ref_count));
3794 		vm_page_object_remove(m);
3795 		m->ref_count -= VPRC_OBJREF;
3796 	} else
3797 		vm_page_assert_unbusied(m);
3798 
3799 	vm_page_busy_free(m);
3800 
3801 	/*
3802 	 * If fictitious remove object association and
3803 	 * return.
3804 	 */
3805 	if ((m->flags & PG_FICTITIOUS) != 0) {
3806 		KASSERT(m->ref_count == 1,
3807 		    ("fictitious page %p is referenced", m));
3808 		KASSERT(m->a.queue == PQ_NONE,
3809 		    ("fictitious page %p is queued", m));
3810 		return (false);
3811 	}
3812 
3813 	/*
3814 	 * Pages need not be dequeued before they are returned to the physical
3815 	 * memory allocator, but they must at least be marked for a deferred
3816 	 * dequeue.
3817 	 */
3818 	if ((m->oflags & VPO_UNMANAGED) == 0)
3819 		vm_page_dequeue_deferred(m);
3820 
3821 	m->valid = 0;
3822 	vm_page_undirty(m);
3823 
3824 	if (m->ref_count != 0)
3825 		panic("vm_page_free_prep: page %p has references", m);
3826 
3827 	/*
3828 	 * Restore the default memory attribute to the page.
3829 	 */
3830 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3831 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3832 
3833 #if VM_NRESERVLEVEL > 0
3834 	/*
3835 	 * Determine whether the page belongs to a reservation.  If the page was
3836 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3837 	 * as an optimization, we avoid the check in that case.
3838 	 */
3839 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3840 		return (false);
3841 #endif
3842 
3843 	return (true);
3844 }
3845 
3846 /*
3847  *	vm_page_free_toq:
3848  *
3849  *	Returns the given page to the free list, disassociating it
3850  *	from any VM object.
3851  *
3852  *	The object must be locked.  The page must be exclusively busied if it
3853  *	belongs to an object.
3854  */
3855 static void
3856 vm_page_free_toq(vm_page_t m)
3857 {
3858 	struct vm_domain *vmd;
3859 	uma_zone_t zone;
3860 
3861 	if (!vm_page_free_prep(m))
3862 		return;
3863 
3864 	vmd = vm_pagequeue_domain(m);
3865 	zone = vmd->vmd_pgcache[m->pool].zone;
3866 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3867 		uma_zfree(zone, m);
3868 		return;
3869 	}
3870 	vm_domain_free_lock(vmd);
3871 	vm_phys_free_pages(m, 0);
3872 	vm_domain_free_unlock(vmd);
3873 	vm_domain_freecnt_inc(vmd, 1);
3874 }
3875 
3876 /*
3877  *	vm_page_free_pages_toq:
3878  *
3879  *	Returns a list of pages to the free list, disassociating it
3880  *	from any VM object.  In other words, this is equivalent to
3881  *	calling vm_page_free_toq() for each page of a list of VM objects.
3882  */
3883 void
3884 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3885 {
3886 	vm_page_t m;
3887 	int count;
3888 
3889 	if (SLIST_EMPTY(free))
3890 		return;
3891 
3892 	count = 0;
3893 	while ((m = SLIST_FIRST(free)) != NULL) {
3894 		count++;
3895 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3896 		vm_page_free_toq(m);
3897 	}
3898 
3899 	if (update_wire_count)
3900 		vm_wire_sub(count);
3901 }
3902 
3903 /*
3904  * Mark this page as wired down.  For managed pages, this prevents reclamation
3905  * by the page daemon, or when the containing object, if any, is destroyed.
3906  */
3907 void
3908 vm_page_wire(vm_page_t m)
3909 {
3910 	u_int old;
3911 
3912 #ifdef INVARIANTS
3913 	if (m->object != NULL && !vm_page_busied(m) &&
3914 	    !vm_object_busied(m->object))
3915 		VM_OBJECT_ASSERT_LOCKED(m->object);
3916 #endif
3917 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3918 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
3919 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
3920 
3921 	old = atomic_fetchadd_int(&m->ref_count, 1);
3922 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3923 	    ("vm_page_wire: counter overflow for page %p", m));
3924 	if (VPRC_WIRE_COUNT(old) == 0) {
3925 		if ((m->oflags & VPO_UNMANAGED) == 0)
3926 			vm_page_aflag_set(m, PGA_DEQUEUE);
3927 		vm_wire_add(1);
3928 	}
3929 }
3930 
3931 /*
3932  * Attempt to wire a mapped page following a pmap lookup of that page.
3933  * This may fail if a thread is concurrently tearing down mappings of the page.
3934  * The transient failure is acceptable because it translates to the
3935  * failure of the caller pmap_extract_and_hold(), which should be then
3936  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3937  */
3938 bool
3939 vm_page_wire_mapped(vm_page_t m)
3940 {
3941 	u_int old;
3942 
3943 	old = m->ref_count;
3944 	do {
3945 		KASSERT(old > 0,
3946 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3947 		if ((old & VPRC_BLOCKED) != 0)
3948 			return (false);
3949 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3950 
3951 	if (VPRC_WIRE_COUNT(old) == 0) {
3952 		if ((m->oflags & VPO_UNMANAGED) == 0)
3953 			vm_page_aflag_set(m, PGA_DEQUEUE);
3954 		vm_wire_add(1);
3955 	}
3956 	return (true);
3957 }
3958 
3959 /*
3960  * Release a wiring reference to a managed page.  If the page still belongs to
3961  * an object, update its position in the page queues to reflect the reference.
3962  * If the wiring was the last reference to the page, free the page.
3963  */
3964 static void
3965 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3966 {
3967 	u_int old;
3968 
3969 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3970 	    ("%s: page %p is unmanaged", __func__, m));
3971 
3972 	/*
3973 	 * Update LRU state before releasing the wiring reference.
3974 	 * Use a release store when updating the reference count to
3975 	 * synchronize with vm_page_free_prep().
3976 	 */
3977 	old = m->ref_count;
3978 	do {
3979 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
3980 		    ("vm_page_unwire: wire count underflow for page %p", m));
3981 
3982 		if (old > VPRC_OBJREF + 1) {
3983 			/*
3984 			 * The page has at least one other wiring reference.  An
3985 			 * earlier iteration of this loop may have called
3986 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3987 			 * re-set it if necessary.
3988 			 */
3989 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3990 				vm_page_aflag_set(m, PGA_DEQUEUE);
3991 		} else if (old == VPRC_OBJREF + 1) {
3992 			/*
3993 			 * This is the last wiring.  Clear PGA_DEQUEUE and
3994 			 * update the page's queue state to reflect the
3995 			 * reference.  If the page does not belong to an object
3996 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3997 			 * clear leftover queue state.
3998 			 */
3999 			vm_page_release_toq(m, nqueue, noreuse);
4000 		} else if (old == 1) {
4001 			vm_page_aflag_clear(m, PGA_DEQUEUE);
4002 		}
4003 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4004 
4005 	if (VPRC_WIRE_COUNT(old) == 1) {
4006 		vm_wire_sub(1);
4007 		if (old == 1)
4008 			vm_page_free(m);
4009 	}
4010 }
4011 
4012 /*
4013  * Release one wiring of the specified page, potentially allowing it to be
4014  * paged out.
4015  *
4016  * Only managed pages belonging to an object can be paged out.  If the number
4017  * of wirings transitions to zero and the page is eligible for page out, then
4018  * the page is added to the specified paging queue.  If the released wiring
4019  * represented the last reference to the page, the page is freed.
4020  */
4021 void
4022 vm_page_unwire(vm_page_t m, uint8_t nqueue)
4023 {
4024 
4025 	KASSERT(nqueue < PQ_COUNT,
4026 	    ("vm_page_unwire: invalid queue %u request for page %p",
4027 	    nqueue, m));
4028 
4029 	if ((m->oflags & VPO_UNMANAGED) != 0) {
4030 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
4031 			vm_page_free(m);
4032 		return;
4033 	}
4034 	vm_page_unwire_managed(m, nqueue, false);
4035 }
4036 
4037 /*
4038  * Unwire a page without (re-)inserting it into a page queue.  It is up
4039  * to the caller to enqueue, requeue, or free the page as appropriate.
4040  * In most cases involving managed pages, vm_page_unwire() should be used
4041  * instead.
4042  */
4043 bool
4044 vm_page_unwire_noq(vm_page_t m)
4045 {
4046 	u_int old;
4047 
4048 	old = vm_page_drop(m, 1);
4049 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
4050 	    ("%s: counter underflow for page %p", __func__,  m));
4051 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
4052 	    ("%s: missing ref on fictitious page %p", __func__, m));
4053 
4054 	if (VPRC_WIRE_COUNT(old) > 1)
4055 		return (false);
4056 	if ((m->oflags & VPO_UNMANAGED) == 0)
4057 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4058 	vm_wire_sub(1);
4059 	return (true);
4060 }
4061 
4062 /*
4063  * Ensure that the page ends up in the specified page queue.  If the page is
4064  * active or being moved to the active queue, ensure that its act_count is
4065  * at least ACT_INIT but do not otherwise mess with it.
4066  */
4067 static __always_inline void
4068 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4069 {
4070 	vm_page_astate_t old, new;
4071 
4072 	KASSERT(m->ref_count > 0,
4073 	    ("%s: page %p does not carry any references", __func__, m));
4074 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4075 	    ("%s: invalid flags %x", __func__, nflag));
4076 
4077 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4078 		return;
4079 
4080 	old = vm_page_astate_load(m);
4081 	do {
4082 		if ((old.flags & PGA_DEQUEUE) != 0)
4083 			break;
4084 		new = old;
4085 		new.flags &= ~PGA_QUEUE_OP_MASK;
4086 		if (nqueue == PQ_ACTIVE)
4087 			new.act_count = max(old.act_count, ACT_INIT);
4088 		if (old.queue == nqueue) {
4089 			if (nqueue != PQ_ACTIVE)
4090 				new.flags |= nflag;
4091 		} else {
4092 			new.flags |= nflag;
4093 			new.queue = nqueue;
4094 		}
4095 	} while (!vm_page_pqstate_commit(m, &old, new));
4096 }
4097 
4098 /*
4099  * Put the specified page on the active list (if appropriate).
4100  */
4101 void
4102 vm_page_activate(vm_page_t m)
4103 {
4104 
4105 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4106 }
4107 
4108 /*
4109  * Move the specified page to the tail of the inactive queue, or requeue
4110  * the page if it is already in the inactive queue.
4111  */
4112 void
4113 vm_page_deactivate(vm_page_t m)
4114 {
4115 
4116 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4117 }
4118 
4119 void
4120 vm_page_deactivate_noreuse(vm_page_t m)
4121 {
4122 
4123 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4124 }
4125 
4126 /*
4127  * Put a page in the laundry, or requeue it if it is already there.
4128  */
4129 void
4130 vm_page_launder(vm_page_t m)
4131 {
4132 
4133 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4134 }
4135 
4136 /*
4137  * Put a page in the PQ_UNSWAPPABLE holding queue.
4138  */
4139 void
4140 vm_page_unswappable(vm_page_t m)
4141 {
4142 
4143 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4144 	    ("page %p already unswappable", m));
4145 
4146 	vm_page_dequeue(m);
4147 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4148 }
4149 
4150 /*
4151  * Release a page back to the page queues in preparation for unwiring.
4152  */
4153 static void
4154 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4155 {
4156 	vm_page_astate_t old, new;
4157 	uint16_t nflag;
4158 
4159 	/*
4160 	 * Use a check of the valid bits to determine whether we should
4161 	 * accelerate reclamation of the page.  The object lock might not be
4162 	 * held here, in which case the check is racy.  At worst we will either
4163 	 * accelerate reclamation of a valid page and violate LRU, or
4164 	 * unnecessarily defer reclamation of an invalid page.
4165 	 *
4166 	 * If we were asked to not cache the page, place it near the head of the
4167 	 * inactive queue so that is reclaimed sooner.
4168 	 */
4169 	if (noreuse || m->valid == 0) {
4170 		nqueue = PQ_INACTIVE;
4171 		nflag = PGA_REQUEUE_HEAD;
4172 	} else {
4173 		nflag = PGA_REQUEUE;
4174 	}
4175 
4176 	old = vm_page_astate_load(m);
4177 	do {
4178 		new = old;
4179 
4180 		/*
4181 		 * If the page is already in the active queue and we are not
4182 		 * trying to accelerate reclamation, simply mark it as
4183 		 * referenced and avoid any queue operations.
4184 		 */
4185 		new.flags &= ~PGA_QUEUE_OP_MASK;
4186 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4187 			new.flags |= PGA_REFERENCED;
4188 		else {
4189 			new.flags |= nflag;
4190 			new.queue = nqueue;
4191 		}
4192 	} while (!vm_page_pqstate_commit(m, &old, new));
4193 }
4194 
4195 /*
4196  * Unwire a page and either attempt to free it or re-add it to the page queues.
4197  */
4198 void
4199 vm_page_release(vm_page_t m, int flags)
4200 {
4201 	vm_object_t object;
4202 
4203 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4204 	    ("vm_page_release: page %p is unmanaged", m));
4205 
4206 	if ((flags & VPR_TRYFREE) != 0) {
4207 		for (;;) {
4208 			object = atomic_load_ptr(&m->object);
4209 			if (object == NULL)
4210 				break;
4211 			/* Depends on type-stability. */
4212 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4213 				break;
4214 			if (object == m->object) {
4215 				vm_page_release_locked(m, flags);
4216 				VM_OBJECT_WUNLOCK(object);
4217 				return;
4218 			}
4219 			VM_OBJECT_WUNLOCK(object);
4220 		}
4221 	}
4222 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4223 }
4224 
4225 /* See vm_page_release(). */
4226 void
4227 vm_page_release_locked(vm_page_t m, int flags)
4228 {
4229 
4230 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4231 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4232 	    ("vm_page_release_locked: page %p is unmanaged", m));
4233 
4234 	if (vm_page_unwire_noq(m)) {
4235 		if ((flags & VPR_TRYFREE) != 0 &&
4236 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4237 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4238 			/*
4239 			 * An unlocked lookup may have wired the page before the
4240 			 * busy lock was acquired, in which case the page must
4241 			 * not be freed.
4242 			 */
4243 			if (__predict_true(!vm_page_wired(m))) {
4244 				vm_page_free(m);
4245 				return;
4246 			}
4247 			vm_page_xunbusy(m);
4248 		} else {
4249 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4250 		}
4251 	}
4252 }
4253 
4254 static bool
4255 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4256 {
4257 	u_int old;
4258 
4259 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4260 	    ("vm_page_try_blocked_op: page %p has no object", m));
4261 	KASSERT(vm_page_busied(m),
4262 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4263 	VM_OBJECT_ASSERT_LOCKED(m->object);
4264 
4265 	old = m->ref_count;
4266 	do {
4267 		KASSERT(old != 0,
4268 		    ("vm_page_try_blocked_op: page %p has no references", m));
4269 		if (VPRC_WIRE_COUNT(old) != 0)
4270 			return (false);
4271 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4272 
4273 	(op)(m);
4274 
4275 	/*
4276 	 * If the object is read-locked, new wirings may be created via an
4277 	 * object lookup.
4278 	 */
4279 	old = vm_page_drop(m, VPRC_BLOCKED);
4280 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4281 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4282 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4283 	    old, m));
4284 	return (true);
4285 }
4286 
4287 /*
4288  * Atomically check for wirings and remove all mappings of the page.
4289  */
4290 bool
4291 vm_page_try_remove_all(vm_page_t m)
4292 {
4293 
4294 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4295 }
4296 
4297 /*
4298  * Atomically check for wirings and remove all writeable mappings of the page.
4299  */
4300 bool
4301 vm_page_try_remove_write(vm_page_t m)
4302 {
4303 
4304 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4305 }
4306 
4307 /*
4308  * vm_page_advise
4309  *
4310  * 	Apply the specified advice to the given page.
4311  */
4312 void
4313 vm_page_advise(vm_page_t m, int advice)
4314 {
4315 
4316 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4317 	vm_page_assert_xbusied(m);
4318 
4319 	if (advice == MADV_FREE)
4320 		/*
4321 		 * Mark the page clean.  This will allow the page to be freed
4322 		 * without first paging it out.  MADV_FREE pages are often
4323 		 * quickly reused by malloc(3), so we do not do anything that
4324 		 * would result in a page fault on a later access.
4325 		 */
4326 		vm_page_undirty(m);
4327 	else if (advice != MADV_DONTNEED) {
4328 		if (advice == MADV_WILLNEED)
4329 			vm_page_activate(m);
4330 		return;
4331 	}
4332 
4333 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4334 		vm_page_dirty(m);
4335 
4336 	/*
4337 	 * Clear any references to the page.  Otherwise, the page daemon will
4338 	 * immediately reactivate the page.
4339 	 */
4340 	vm_page_aflag_clear(m, PGA_REFERENCED);
4341 
4342 	/*
4343 	 * Place clean pages near the head of the inactive queue rather than
4344 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4345 	 * the page will be reused quickly.  Dirty pages not already in the
4346 	 * laundry are moved there.
4347 	 */
4348 	if (m->dirty == 0)
4349 		vm_page_deactivate_noreuse(m);
4350 	else if (!vm_page_in_laundry(m))
4351 		vm_page_launder(m);
4352 }
4353 
4354 /*
4355  *	vm_page_grab_release
4356  *
4357  *	Helper routine for grab functions to release busy on return.
4358  */
4359 static inline void
4360 vm_page_grab_release(vm_page_t m, int allocflags)
4361 {
4362 
4363 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4364 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4365 			vm_page_sunbusy(m);
4366 		else
4367 			vm_page_xunbusy(m);
4368 	}
4369 }
4370 
4371 /*
4372  *	vm_page_grab_sleep
4373  *
4374  *	Sleep for busy according to VM_ALLOC_ parameters.  Returns true
4375  *	if the caller should retry and false otherwise.
4376  *
4377  *	If the object is locked on entry the object will be unlocked with
4378  *	false returns and still locked but possibly having been dropped
4379  *	with true returns.
4380  */
4381 static bool
4382 vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
4383     const char *wmesg, int allocflags, bool locked)
4384 {
4385 
4386 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4387 		return (false);
4388 
4389 	/*
4390 	 * Reference the page before unlocking and sleeping so that
4391 	 * the page daemon is less likely to reclaim it.
4392 	 */
4393 	if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
4394 		vm_page_reference(m);
4395 
4396 	if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
4397 	    locked)
4398 		VM_OBJECT_WLOCK(object);
4399 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
4400 		return (false);
4401 
4402 	return (true);
4403 }
4404 
4405 /*
4406  * Assert that the grab flags are valid.
4407  */
4408 static inline void
4409 vm_page_grab_check(int allocflags)
4410 {
4411 
4412 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4413 	    (allocflags & VM_ALLOC_WIRED) != 0,
4414 	    ("vm_page_grab*: the pages must be busied or wired"));
4415 
4416 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4417 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4418 	    ("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4419 }
4420 
4421 /*
4422  * Calculate the page allocation flags for grab.
4423  */
4424 static inline int
4425 vm_page_grab_pflags(int allocflags)
4426 {
4427 	int pflags;
4428 
4429 	pflags = allocflags &
4430 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4431 	    VM_ALLOC_NOBUSY);
4432 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4433 		pflags |= VM_ALLOC_WAITFAIL;
4434 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4435 		pflags |= VM_ALLOC_SBUSY;
4436 
4437 	return (pflags);
4438 }
4439 
4440 /*
4441  * Grab a page, waiting until we are waken up due to the page
4442  * changing state.  We keep on waiting, if the page continues
4443  * to be in the object.  If the page doesn't exist, first allocate it
4444  * and then conditionally zero it.
4445  *
4446  * This routine may sleep.
4447  *
4448  * The object must be locked on entry.  The lock will, however, be released
4449  * and reacquired if the routine sleeps.
4450  */
4451 vm_page_t
4452 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4453 {
4454 	vm_page_t m;
4455 
4456 	VM_OBJECT_ASSERT_WLOCKED(object);
4457 	vm_page_grab_check(allocflags);
4458 
4459 retrylookup:
4460 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4461 		if (!vm_page_tryacquire(m, allocflags)) {
4462 			if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4463 			    allocflags, true))
4464 				goto retrylookup;
4465 			return (NULL);
4466 		}
4467 		goto out;
4468 	}
4469 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4470 		return (NULL);
4471 	m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
4472 	if (m == NULL) {
4473 		if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
4474 			return (NULL);
4475 		goto retrylookup;
4476 	}
4477 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4478 		pmap_zero_page(m);
4479 
4480 out:
4481 	vm_page_grab_release(m, allocflags);
4482 
4483 	return (m);
4484 }
4485 
4486 /*
4487  * Locklessly attempt to acquire a page given a (object, pindex) tuple
4488  * and an optional previous page to avoid the radix lookup.  The resulting
4489  * page will be validated against the identity tuple and busied or wired
4490  * as requested.  A NULL *mp return guarantees that the page was not in
4491  * radix at the time of the call but callers must perform higher level
4492  * synchronization or retry the operation under a lock if they require
4493  * an atomic answer.  This is the only lock free validation routine,
4494  * other routines can depend on the resulting page state.
4495  *
4496  * The return value indicates whether the operation failed due to caller
4497  * flags.  The return is tri-state with mp:
4498  *
4499  * (true, *mp != NULL) - The operation was successful.
4500  * (true, *mp == NULL) - The page was not found in tree.
4501  * (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
4502  */
4503 static bool
4504 vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
4505     vm_page_t prev, vm_page_t *mp, int allocflags)
4506 {
4507 	vm_page_t m;
4508 
4509 	vm_page_grab_check(allocflags);
4510 	MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
4511 
4512 	*mp = NULL;
4513 	for (;;) {
4514 		/*
4515 		 * We may see a false NULL here because the previous page
4516 		 * has been removed or just inserted and the list is loaded
4517 		 * without barriers.  Switch to radix to verify.
4518 		 */
4519 		if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
4520 		    QMD_IS_TRASHED(m) || m->pindex != pindex ||
4521 		    atomic_load_ptr(&m->object) != object) {
4522 			prev = NULL;
4523 			/*
4524 			 * This guarantees the result is instantaneously
4525 			 * correct.
4526 			 */
4527 			m = vm_radix_lookup_unlocked(&object->rtree, pindex);
4528 		}
4529 		if (m == NULL)
4530 			return (true);
4531 		if (vm_page_trybusy(m, allocflags)) {
4532 			if (m->object == object && m->pindex == pindex)
4533 				break;
4534 			/* relookup. */
4535 			vm_page_busy_release(m);
4536 			cpu_spinwait();
4537 			continue;
4538 		}
4539 		if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
4540 		    allocflags, false))
4541 			return (false);
4542 	}
4543 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4544 		vm_page_wire(m);
4545 	vm_page_grab_release(m, allocflags);
4546 	*mp = m;
4547 	return (true);
4548 }
4549 
4550 /*
4551  * Try to locklessly grab a page and fall back to the object lock if NOCREAT
4552  * is not set.
4553  */
4554 vm_page_t
4555 vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
4556 {
4557 	vm_page_t m;
4558 
4559 	vm_page_grab_check(allocflags);
4560 
4561 	if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
4562 		return (NULL);
4563 	if (m != NULL)
4564 		return (m);
4565 
4566 	/*
4567 	 * The radix lockless lookup should never return a false negative
4568 	 * errors.  If the user specifies NOCREAT they are guaranteed there
4569 	 * was no page present at the instant of the call.  A NOCREAT caller
4570 	 * must handle create races gracefully.
4571 	 */
4572 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4573 		return (NULL);
4574 
4575 	VM_OBJECT_WLOCK(object);
4576 	m = vm_page_grab(object, pindex, allocflags);
4577 	VM_OBJECT_WUNLOCK(object);
4578 
4579 	return (m);
4580 }
4581 
4582 /*
4583  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4584  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4585  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4586  * in simultaneously.  Additional pages will be left on a paging queue but
4587  * will neither be wired nor busy regardless of allocflags.
4588  */
4589 int
4590 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4591 {
4592 	vm_page_t m;
4593 	vm_page_t ma[VM_INITIAL_PAGEIN];
4594 	int after, i, pflags, rv;
4595 
4596 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4597 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4598 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4599 	KASSERT((allocflags &
4600 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4601 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4602 	VM_OBJECT_ASSERT_WLOCKED(object);
4603 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
4604 	    VM_ALLOC_WIRED);
4605 	pflags |= VM_ALLOC_WAITFAIL;
4606 
4607 retrylookup:
4608 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4609 		/*
4610 		 * If the page is fully valid it can only become invalid
4611 		 * with the object lock held.  If it is not valid it can
4612 		 * become valid with the busy lock held.  Therefore, we
4613 		 * may unnecessarily lock the exclusive busy here if we
4614 		 * race with I/O completion not using the object lock.
4615 		 * However, we will not end up with an invalid page and a
4616 		 * shared lock.
4617 		 */
4618 		if (!vm_page_trybusy(m,
4619 		    vm_page_all_valid(m) ? allocflags : 0)) {
4620 			(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
4621 			    allocflags, true);
4622 			goto retrylookup;
4623 		}
4624 		if (vm_page_all_valid(m))
4625 			goto out;
4626 		if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4627 			vm_page_busy_release(m);
4628 			*mp = NULL;
4629 			return (VM_PAGER_FAIL);
4630 		}
4631 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4632 		*mp = NULL;
4633 		return (VM_PAGER_FAIL);
4634 	} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
4635 		goto retrylookup;
4636 	}
4637 
4638 	vm_page_assert_xbusied(m);
4639 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4640 		after = MIN(after, VM_INITIAL_PAGEIN);
4641 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4642 		after = MAX(after, 1);
4643 		ma[0] = m;
4644 		for (i = 1; i < after; i++) {
4645 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4646 				if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4647 					break;
4648 			} else {
4649 				ma[i] = vm_page_alloc(object, m->pindex + i,
4650 				    VM_ALLOC_NORMAL);
4651 				if (ma[i] == NULL)
4652 					break;
4653 			}
4654 		}
4655 		after = i;
4656 		vm_object_pip_add(object, after);
4657 		VM_OBJECT_WUNLOCK(object);
4658 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4659 		VM_OBJECT_WLOCK(object);
4660 		vm_object_pip_wakeupn(object, after);
4661 		/* Pager may have replaced a page. */
4662 		m = ma[0];
4663 		if (rv != VM_PAGER_OK) {
4664 			for (i = 0; i < after; i++) {
4665 				if (!vm_page_wired(ma[i]))
4666 					vm_page_free(ma[i]);
4667 				else
4668 					vm_page_xunbusy(ma[i]);
4669 			}
4670 			*mp = NULL;
4671 			return (rv);
4672 		}
4673 		for (i = 1; i < after; i++)
4674 			vm_page_readahead_finish(ma[i]);
4675 		MPASS(vm_page_all_valid(m));
4676 	} else {
4677 		vm_page_zero_invalid(m, TRUE);
4678 	}
4679 out:
4680 	if ((allocflags & VM_ALLOC_WIRED) != 0)
4681 		vm_page_wire(m);
4682 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
4683 		vm_page_busy_downgrade(m);
4684 	else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
4685 		vm_page_busy_release(m);
4686 	*mp = m;
4687 	return (VM_PAGER_OK);
4688 }
4689 
4690 /*
4691  * Locklessly grab a valid page.  If the page is not valid or not yet
4692  * allocated this will fall back to the object lock method.
4693  */
4694 int
4695 vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
4696     vm_pindex_t pindex, int allocflags)
4697 {
4698 	vm_page_t m;
4699 	int flags;
4700 	int error;
4701 
4702 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4703 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4704 	    ("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4705 	    "mismatch"));
4706 	KASSERT((allocflags &
4707 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4708 	    ("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
4709 
4710 	/*
4711 	 * Attempt a lockless lookup and busy.  We need at least an sbusy
4712 	 * before we can inspect the valid field and return a wired page.
4713 	 */
4714 	flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
4715 	if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
4716 		return (VM_PAGER_FAIL);
4717 	if ((m = *mp) != NULL) {
4718 		if (vm_page_all_valid(m)) {
4719 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4720 				vm_page_wire(m);
4721 			vm_page_grab_release(m, allocflags);
4722 			return (VM_PAGER_OK);
4723 		}
4724 		vm_page_busy_release(m);
4725 	}
4726 	if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4727 		*mp = NULL;
4728 		return (VM_PAGER_FAIL);
4729 	}
4730 	VM_OBJECT_WLOCK(object);
4731 	error = vm_page_grab_valid(mp, object, pindex, allocflags);
4732 	VM_OBJECT_WUNLOCK(object);
4733 
4734 	return (error);
4735 }
4736 
4737 /*
4738  * Return the specified range of pages from the given object.  For each
4739  * page offset within the range, if a page already exists within the object
4740  * at that offset and it is busy, then wait for it to change state.  If,
4741  * instead, the page doesn't exist, then allocate it.
4742  *
4743  * The caller must always specify an allocation class.
4744  *
4745  * allocation classes:
4746  *	VM_ALLOC_NORMAL		normal process request
4747  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4748  *
4749  * The caller must always specify that the pages are to be busied and/or
4750  * wired.
4751  *
4752  * optional allocation flags:
4753  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4754  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4755  *	VM_ALLOC_NOWAIT		do not sleep
4756  *	VM_ALLOC_SBUSY		set page to sbusy state
4757  *	VM_ALLOC_WIRED		wire the pages
4758  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4759  *
4760  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4761  * may return a partial prefix of the requested range.
4762  */
4763 int
4764 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4765     vm_page_t *ma, int count)
4766 {
4767 	vm_page_t m, mpred;
4768 	int pflags;
4769 	int i;
4770 
4771 	VM_OBJECT_ASSERT_WLOCKED(object);
4772 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4773 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4774 	KASSERT(count > 0,
4775 	    ("vm_page_grab_pages: invalid page count %d", count));
4776 	vm_page_grab_check(allocflags);
4777 
4778 	pflags = vm_page_grab_pflags(allocflags);
4779 	i = 0;
4780 retrylookup:
4781 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4782 	if (m == NULL || m->pindex != pindex + i) {
4783 		mpred = m;
4784 		m = NULL;
4785 	} else
4786 		mpred = TAILQ_PREV(m, pglist, listq);
4787 	for (; i < count; i++) {
4788 		if (m != NULL) {
4789 			if (!vm_page_tryacquire(m, allocflags)) {
4790 				if (vm_page_grab_sleep(object, m, pindex + i,
4791 				    "grbmaw", allocflags, true))
4792 					goto retrylookup;
4793 				break;
4794 			}
4795 		} else {
4796 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4797 				break;
4798 			m = vm_page_alloc_after(object, pindex + i,
4799 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4800 			if (m == NULL) {
4801 				if ((allocflags & (VM_ALLOC_NOWAIT |
4802 				    VM_ALLOC_WAITFAIL)) != 0)
4803 					break;
4804 				goto retrylookup;
4805 			}
4806 		}
4807 		if (vm_page_none_valid(m) &&
4808 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4809 			if ((m->flags & PG_ZERO) == 0)
4810 				pmap_zero_page(m);
4811 			vm_page_valid(m);
4812 		}
4813 		vm_page_grab_release(m, allocflags);
4814 		ma[i] = mpred = m;
4815 		m = vm_page_next(m);
4816 	}
4817 	return (i);
4818 }
4819 
4820 /*
4821  * Unlocked variant of vm_page_grab_pages().  This accepts the same flags
4822  * and will fall back to the locked variant to handle allocation.
4823  */
4824 int
4825 vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
4826     int allocflags, vm_page_t *ma, int count)
4827 {
4828 	vm_page_t m, pred;
4829 	int flags;
4830 	int i;
4831 
4832 	KASSERT(count > 0,
4833 	    ("vm_page_grab_pages_unlocked: invalid page count %d", count));
4834 	vm_page_grab_check(allocflags);
4835 
4836 	/*
4837 	 * Modify flags for lockless acquire to hold the page until we
4838 	 * set it valid if necessary.
4839 	 */
4840 	flags = allocflags & ~VM_ALLOC_NOBUSY;
4841 	pred = NULL;
4842 	for (i = 0; i < count; i++, pindex++) {
4843 		if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
4844 			return (i);
4845 		if (m == NULL)
4846 			break;
4847 		if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
4848 			if ((m->flags & PG_ZERO) == 0)
4849 				pmap_zero_page(m);
4850 			vm_page_valid(m);
4851 		}
4852 		/* m will still be wired or busy according to flags. */
4853 		vm_page_grab_release(m, allocflags);
4854 		pred = ma[i] = m;
4855 	}
4856 	if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
4857 		return (i);
4858 	count -= i;
4859 	VM_OBJECT_WLOCK(object);
4860 	i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
4861 	VM_OBJECT_WUNLOCK(object);
4862 
4863 	return (i);
4864 }
4865 
4866 /*
4867  * Mapping function for valid or dirty bits in a page.
4868  *
4869  * Inputs are required to range within a page.
4870  */
4871 vm_page_bits_t
4872 vm_page_bits(int base, int size)
4873 {
4874 	int first_bit;
4875 	int last_bit;
4876 
4877 	KASSERT(
4878 	    base + size <= PAGE_SIZE,
4879 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4880 	);
4881 
4882 	if (size == 0)		/* handle degenerate case */
4883 		return (0);
4884 
4885 	first_bit = base >> DEV_BSHIFT;
4886 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4887 
4888 	return (((vm_page_bits_t)2 << last_bit) -
4889 	    ((vm_page_bits_t)1 << first_bit));
4890 }
4891 
4892 void
4893 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4894 {
4895 
4896 #if PAGE_SIZE == 32768
4897 	atomic_set_64((uint64_t *)bits, set);
4898 #elif PAGE_SIZE == 16384
4899 	atomic_set_32((uint32_t *)bits, set);
4900 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4901 	atomic_set_16((uint16_t *)bits, set);
4902 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4903 	atomic_set_8((uint8_t *)bits, set);
4904 #else		/* PAGE_SIZE <= 8192 */
4905 	uintptr_t addr;
4906 	int shift;
4907 
4908 	addr = (uintptr_t)bits;
4909 	/*
4910 	 * Use a trick to perform a 32-bit atomic on the
4911 	 * containing aligned word, to not depend on the existence
4912 	 * of atomic_{set, clear}_{8, 16}.
4913 	 */
4914 	shift = addr & (sizeof(uint32_t) - 1);
4915 #if BYTE_ORDER == BIG_ENDIAN
4916 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4917 #else
4918 	shift *= NBBY;
4919 #endif
4920 	addr &= ~(sizeof(uint32_t) - 1);
4921 	atomic_set_32((uint32_t *)addr, set << shift);
4922 #endif		/* PAGE_SIZE */
4923 }
4924 
4925 static inline void
4926 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4927 {
4928 
4929 #if PAGE_SIZE == 32768
4930 	atomic_clear_64((uint64_t *)bits, clear);
4931 #elif PAGE_SIZE == 16384
4932 	atomic_clear_32((uint32_t *)bits, clear);
4933 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4934 	atomic_clear_16((uint16_t *)bits, clear);
4935 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4936 	atomic_clear_8((uint8_t *)bits, clear);
4937 #else		/* PAGE_SIZE <= 8192 */
4938 	uintptr_t addr;
4939 	int shift;
4940 
4941 	addr = (uintptr_t)bits;
4942 	/*
4943 	 * Use a trick to perform a 32-bit atomic on the
4944 	 * containing aligned word, to not depend on the existence
4945 	 * of atomic_{set, clear}_{8, 16}.
4946 	 */
4947 	shift = addr & (sizeof(uint32_t) - 1);
4948 #if BYTE_ORDER == BIG_ENDIAN
4949 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4950 #else
4951 	shift *= NBBY;
4952 #endif
4953 	addr &= ~(sizeof(uint32_t) - 1);
4954 	atomic_clear_32((uint32_t *)addr, clear << shift);
4955 #endif		/* PAGE_SIZE */
4956 }
4957 
4958 static inline vm_page_bits_t
4959 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4960 {
4961 #if PAGE_SIZE == 32768
4962 	uint64_t old;
4963 
4964 	old = *bits;
4965 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4966 	return (old);
4967 #elif PAGE_SIZE == 16384
4968 	uint32_t old;
4969 
4970 	old = *bits;
4971 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4972 	return (old);
4973 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4974 	uint16_t old;
4975 
4976 	old = *bits;
4977 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4978 	return (old);
4979 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4980 	uint8_t old;
4981 
4982 	old = *bits;
4983 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4984 	return (old);
4985 #else		/* PAGE_SIZE <= 4096*/
4986 	uintptr_t addr;
4987 	uint32_t old, new, mask;
4988 	int shift;
4989 
4990 	addr = (uintptr_t)bits;
4991 	/*
4992 	 * Use a trick to perform a 32-bit atomic on the
4993 	 * containing aligned word, to not depend on the existence
4994 	 * of atomic_{set, swap, clear}_{8, 16}.
4995 	 */
4996 	shift = addr & (sizeof(uint32_t) - 1);
4997 #if BYTE_ORDER == BIG_ENDIAN
4998 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4999 #else
5000 	shift *= NBBY;
5001 #endif
5002 	addr &= ~(sizeof(uint32_t) - 1);
5003 	mask = VM_PAGE_BITS_ALL << shift;
5004 
5005 	old = *bits;
5006 	do {
5007 		new = old & ~mask;
5008 		new |= newbits << shift;
5009 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
5010 	return (old >> shift);
5011 #endif		/* PAGE_SIZE */
5012 }
5013 
5014 /*
5015  *	vm_page_set_valid_range:
5016  *
5017  *	Sets portions of a page valid.  The arguments are expected
5018  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5019  *	of any partial chunks touched by the range.  The invalid portion of
5020  *	such chunks will be zeroed.
5021  *
5022  *	(base + size) must be less then or equal to PAGE_SIZE.
5023  */
5024 void
5025 vm_page_set_valid_range(vm_page_t m, int base, int size)
5026 {
5027 	int endoff, frag;
5028 	vm_page_bits_t pagebits;
5029 
5030 	vm_page_assert_busied(m);
5031 	if (size == 0)	/* handle degenerate case */
5032 		return;
5033 
5034 	/*
5035 	 * If the base is not DEV_BSIZE aligned and the valid
5036 	 * bit is clear, we have to zero out a portion of the
5037 	 * first block.
5038 	 */
5039 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5040 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
5041 		pmap_zero_page_area(m, frag, base - frag);
5042 
5043 	/*
5044 	 * If the ending offset is not DEV_BSIZE aligned and the
5045 	 * valid bit is clear, we have to zero out a portion of
5046 	 * the last block.
5047 	 */
5048 	endoff = base + size;
5049 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5050 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
5051 		pmap_zero_page_area(m, endoff,
5052 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5053 
5054 	/*
5055 	 * Assert that no previously invalid block that is now being validated
5056 	 * is already dirty.
5057 	 */
5058 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
5059 	    ("vm_page_set_valid_range: page %p is dirty", m));
5060 
5061 	/*
5062 	 * Set valid bits inclusive of any overlap.
5063 	 */
5064 	pagebits = vm_page_bits(base, size);
5065 	if (vm_page_xbusied(m))
5066 		m->valid |= pagebits;
5067 	else
5068 		vm_page_bits_set(m, &m->valid, pagebits);
5069 }
5070 
5071 /*
5072  * Set the page dirty bits and free the invalid swap space if
5073  * present.  Returns the previous dirty bits.
5074  */
5075 vm_page_bits_t
5076 vm_page_set_dirty(vm_page_t m)
5077 {
5078 	vm_page_bits_t old;
5079 
5080 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
5081 
5082 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
5083 		old = m->dirty;
5084 		m->dirty = VM_PAGE_BITS_ALL;
5085 	} else
5086 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
5087 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
5088 		vm_pager_page_unswapped(m);
5089 
5090 	return (old);
5091 }
5092 
5093 /*
5094  * Clear the given bits from the specified page's dirty field.
5095  */
5096 static __inline void
5097 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
5098 {
5099 
5100 	vm_page_assert_busied(m);
5101 
5102 	/*
5103 	 * If the page is xbusied and not write mapped we are the
5104 	 * only thread that can modify dirty bits.  Otherwise, The pmap
5105 	 * layer can call vm_page_dirty() without holding a distinguished
5106 	 * lock.  The combination of page busy and atomic operations
5107 	 * suffice to guarantee consistency of the page dirty field.
5108 	 */
5109 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
5110 		m->dirty &= ~pagebits;
5111 	else
5112 		vm_page_bits_clear(m, &m->dirty, pagebits);
5113 }
5114 
5115 /*
5116  *	vm_page_set_validclean:
5117  *
5118  *	Sets portions of a page valid and clean.  The arguments are expected
5119  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
5120  *	of any partial chunks touched by the range.  The invalid portion of
5121  *	such chunks will be zero'd.
5122  *
5123  *	(base + size) must be less then or equal to PAGE_SIZE.
5124  */
5125 void
5126 vm_page_set_validclean(vm_page_t m, int base, int size)
5127 {
5128 	vm_page_bits_t oldvalid, pagebits;
5129 	int endoff, frag;
5130 
5131 	vm_page_assert_busied(m);
5132 	if (size == 0)	/* handle degenerate case */
5133 		return;
5134 
5135 	/*
5136 	 * If the base is not DEV_BSIZE aligned and the valid
5137 	 * bit is clear, we have to zero out a portion of the
5138 	 * first block.
5139 	 */
5140 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
5141 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
5142 		pmap_zero_page_area(m, frag, base - frag);
5143 
5144 	/*
5145 	 * If the ending offset is not DEV_BSIZE aligned and the
5146 	 * valid bit is clear, we have to zero out a portion of
5147 	 * the last block.
5148 	 */
5149 	endoff = base + size;
5150 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
5151 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
5152 		pmap_zero_page_area(m, endoff,
5153 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
5154 
5155 	/*
5156 	 * Set valid, clear dirty bits.  If validating the entire
5157 	 * page we can safely clear the pmap modify bit.  We also
5158 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
5159 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
5160 	 * be set again.
5161 	 *
5162 	 * We set valid bits inclusive of any overlap, but we can only
5163 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
5164 	 * the range.
5165 	 */
5166 	oldvalid = m->valid;
5167 	pagebits = vm_page_bits(base, size);
5168 	if (vm_page_xbusied(m))
5169 		m->valid |= pagebits;
5170 	else
5171 		vm_page_bits_set(m, &m->valid, pagebits);
5172 #if 0	/* NOT YET */
5173 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
5174 		frag = DEV_BSIZE - frag;
5175 		base += frag;
5176 		size -= frag;
5177 		if (size < 0)
5178 			size = 0;
5179 	}
5180 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
5181 #endif
5182 	if (base == 0 && size == PAGE_SIZE) {
5183 		/*
5184 		 * The page can only be modified within the pmap if it is
5185 		 * mapped, and it can only be mapped if it was previously
5186 		 * fully valid.
5187 		 */
5188 		if (oldvalid == VM_PAGE_BITS_ALL)
5189 			/*
5190 			 * Perform the pmap_clear_modify() first.  Otherwise,
5191 			 * a concurrent pmap operation, such as
5192 			 * pmap_protect(), could clear a modification in the
5193 			 * pmap and set the dirty field on the page before
5194 			 * pmap_clear_modify() had begun and after the dirty
5195 			 * field was cleared here.
5196 			 */
5197 			pmap_clear_modify(m);
5198 		m->dirty = 0;
5199 		vm_page_aflag_clear(m, PGA_NOSYNC);
5200 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
5201 		m->dirty &= ~pagebits;
5202 	else
5203 		vm_page_clear_dirty_mask(m, pagebits);
5204 }
5205 
5206 void
5207 vm_page_clear_dirty(vm_page_t m, int base, int size)
5208 {
5209 
5210 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
5211 }
5212 
5213 /*
5214  *	vm_page_set_invalid:
5215  *
5216  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
5217  *	valid and dirty bits for the effected areas are cleared.
5218  */
5219 void
5220 vm_page_set_invalid(vm_page_t m, int base, int size)
5221 {
5222 	vm_page_bits_t bits;
5223 	vm_object_t object;
5224 
5225 	/*
5226 	 * The object lock is required so that pages can't be mapped
5227 	 * read-only while we're in the process of invalidating them.
5228 	 */
5229 	object = m->object;
5230 	VM_OBJECT_ASSERT_WLOCKED(object);
5231 	vm_page_assert_busied(m);
5232 
5233 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
5234 	    size >= object->un_pager.vnp.vnp_size)
5235 		bits = VM_PAGE_BITS_ALL;
5236 	else
5237 		bits = vm_page_bits(base, size);
5238 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
5239 		pmap_remove_all(m);
5240 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
5241 	    !pmap_page_is_mapped(m),
5242 	    ("vm_page_set_invalid: page %p is mapped", m));
5243 	if (vm_page_xbusied(m)) {
5244 		m->valid &= ~bits;
5245 		m->dirty &= ~bits;
5246 	} else {
5247 		vm_page_bits_clear(m, &m->valid, bits);
5248 		vm_page_bits_clear(m, &m->dirty, bits);
5249 	}
5250 }
5251 
5252 /*
5253  *	vm_page_invalid:
5254  *
5255  *	Invalidates the entire page.  The page must be busy, unmapped, and
5256  *	the enclosing object must be locked.  The object locks protects
5257  *	against concurrent read-only pmap enter which is done without
5258  *	busy.
5259  */
5260 void
5261 vm_page_invalid(vm_page_t m)
5262 {
5263 
5264 	vm_page_assert_busied(m);
5265 	VM_OBJECT_ASSERT_LOCKED(m->object);
5266 	MPASS(!pmap_page_is_mapped(m));
5267 
5268 	if (vm_page_xbusied(m))
5269 		m->valid = 0;
5270 	else
5271 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
5272 }
5273 
5274 /*
5275  * vm_page_zero_invalid()
5276  *
5277  *	The kernel assumes that the invalid portions of a page contain
5278  *	garbage, but such pages can be mapped into memory by user code.
5279  *	When this occurs, we must zero out the non-valid portions of the
5280  *	page so user code sees what it expects.
5281  *
5282  *	Pages are most often semi-valid when the end of a file is mapped
5283  *	into memory and the file's size is not page aligned.
5284  */
5285 void
5286 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
5287 {
5288 	int b;
5289 	int i;
5290 
5291 	/*
5292 	 * Scan the valid bits looking for invalid sections that
5293 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5294 	 * valid bit may be set ) have already been zeroed by
5295 	 * vm_page_set_validclean().
5296 	 */
5297 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5298 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5299 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5300 			if (i > b) {
5301 				pmap_zero_page_area(m,
5302 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5303 			}
5304 			b = i + 1;
5305 		}
5306 	}
5307 
5308 	/*
5309 	 * setvalid is TRUE when we can safely set the zero'd areas
5310 	 * as being valid.  We can do this if there are no cache consistancy
5311 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5312 	 */
5313 	if (setvalid)
5314 		vm_page_valid(m);
5315 }
5316 
5317 /*
5318  *	vm_page_is_valid:
5319  *
5320  *	Is (partial) page valid?  Note that the case where size == 0
5321  *	will return FALSE in the degenerate case where the page is
5322  *	entirely invalid, and TRUE otherwise.
5323  *
5324  *	Some callers envoke this routine without the busy lock held and
5325  *	handle races via higher level locks.  Typical callers should
5326  *	hold a busy lock to prevent invalidation.
5327  */
5328 int
5329 vm_page_is_valid(vm_page_t m, int base, int size)
5330 {
5331 	vm_page_bits_t bits;
5332 
5333 	bits = vm_page_bits(base, size);
5334 	return (m->valid != 0 && (m->valid & bits) == bits);
5335 }
5336 
5337 /*
5338  * Returns true if all of the specified predicates are true for the entire
5339  * (super)page and false otherwise.
5340  */
5341 bool
5342 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5343 {
5344 	vm_object_t object;
5345 	int i, npages;
5346 
5347 	object = m->object;
5348 	if (skip_m != NULL && skip_m->object != object)
5349 		return (false);
5350 	VM_OBJECT_ASSERT_LOCKED(object);
5351 	npages = atop(pagesizes[m->psind]);
5352 
5353 	/*
5354 	 * The physically contiguous pages that make up a superpage, i.e., a
5355 	 * page with a page size index ("psind") greater than zero, will
5356 	 * occupy adjacent entries in vm_page_array[].
5357 	 */
5358 	for (i = 0; i < npages; i++) {
5359 		/* Always test object consistency, including "skip_m". */
5360 		if (m[i].object != object)
5361 			return (false);
5362 		if (&m[i] == skip_m)
5363 			continue;
5364 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5365 			return (false);
5366 		if ((flags & PS_ALL_DIRTY) != 0) {
5367 			/*
5368 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5369 			 * might stop this case from spuriously returning
5370 			 * "false".  However, that would require a write lock
5371 			 * on the object containing "m[i]".
5372 			 */
5373 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5374 				return (false);
5375 		}
5376 		if ((flags & PS_ALL_VALID) != 0 &&
5377 		    m[i].valid != VM_PAGE_BITS_ALL)
5378 			return (false);
5379 	}
5380 	return (true);
5381 }
5382 
5383 /*
5384  * Set the page's dirty bits if the page is modified.
5385  */
5386 void
5387 vm_page_test_dirty(vm_page_t m)
5388 {
5389 
5390 	vm_page_assert_busied(m);
5391 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5392 		vm_page_dirty(m);
5393 }
5394 
5395 void
5396 vm_page_valid(vm_page_t m)
5397 {
5398 
5399 	vm_page_assert_busied(m);
5400 	if (vm_page_xbusied(m))
5401 		m->valid = VM_PAGE_BITS_ALL;
5402 	else
5403 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5404 }
5405 
5406 void
5407 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5408 {
5409 
5410 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5411 }
5412 
5413 void
5414 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5415 {
5416 
5417 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5418 }
5419 
5420 int
5421 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5422 {
5423 
5424 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5425 }
5426 
5427 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5428 void
5429 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5430 {
5431 
5432 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5433 }
5434 
5435 void
5436 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5437 {
5438 
5439 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5440 }
5441 #endif
5442 
5443 #ifdef INVARIANTS
5444 void
5445 vm_page_object_busy_assert(vm_page_t m)
5446 {
5447 
5448 	/*
5449 	 * Certain of the page's fields may only be modified by the
5450 	 * holder of a page or object busy.
5451 	 */
5452 	if (m->object != NULL && !vm_page_busied(m))
5453 		VM_OBJECT_ASSERT_BUSY(m->object);
5454 }
5455 
5456 void
5457 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5458 {
5459 
5460 	if ((bits & PGA_WRITEABLE) == 0)
5461 		return;
5462 
5463 	/*
5464 	 * The PGA_WRITEABLE flag can only be set if the page is
5465 	 * managed, is exclusively busied or the object is locked.
5466 	 * Currently, this flag is only set by pmap_enter().
5467 	 */
5468 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5469 	    ("PGA_WRITEABLE on unmanaged page"));
5470 	if (!vm_page_xbusied(m))
5471 		VM_OBJECT_ASSERT_BUSY(m->object);
5472 }
5473 #endif
5474 
5475 #include "opt_ddb.h"
5476 #ifdef DDB
5477 #include <sys/kernel.h>
5478 
5479 #include <ddb/ddb.h>
5480 
5481 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5482 {
5483 
5484 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5485 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5486 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5487 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5488 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5489 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5490 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5491 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5492 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5493 }
5494 
5495 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5496 {
5497 	int dom;
5498 
5499 	db_printf("pq_free %d\n", vm_free_count());
5500 	for (dom = 0; dom < vm_ndomains; dom++) {
5501 		db_printf(
5502     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5503 		    dom,
5504 		    vm_dom[dom].vmd_page_count,
5505 		    vm_dom[dom].vmd_free_count,
5506 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5507 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5508 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5509 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5510 	}
5511 }
5512 
5513 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5514 {
5515 	vm_page_t m;
5516 	boolean_t phys, virt;
5517 
5518 	if (!have_addr) {
5519 		db_printf("show pginfo addr\n");
5520 		return;
5521 	}
5522 
5523 	phys = strchr(modif, 'p') != NULL;
5524 	virt = strchr(modif, 'v') != NULL;
5525 	if (virt)
5526 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5527 	else if (phys)
5528 		m = PHYS_TO_VM_PAGE(addr);
5529 	else
5530 		m = (vm_page_t)addr;
5531 	db_printf(
5532     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref 0x%x\n"
5533     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5534 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5535 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5536 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5537 }
5538 #endif /* DDB */
5539