xref: /freebsd/sys/vm/vm_page.c (revision b249ce48ea5560afdcff57e72a9880b7d3132434)
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/uma.h>
112 #include <vm/uma_int.h>
113 
114 #include <machine/md_var.h>
115 
116 extern int	uma_startup_count(int);
117 extern void	uma_startup(void *, int);
118 extern int	vmem_startup_count(void);
119 
120 struct vm_domain vm_dom[MAXMEMDOM];
121 
122 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
123 
124 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
125 
126 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
127 /* The following fields are protected by the domainset lock. */
128 domainset_t __exclusive_cache_line vm_min_domains;
129 domainset_t __exclusive_cache_line vm_severe_domains;
130 static int vm_min_waiters;
131 static int vm_severe_waiters;
132 static int vm_pageproc_waiters;
133 
134 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0,
135     "VM page statistics");
136 
137 static counter_u64_t pqstate_commit_retries = EARLY_COUNTER;
138 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
139     CTLFLAG_RD, &pqstate_commit_retries,
140     "Number of failed per-page atomic queue state updates");
141 
142 static counter_u64_t queue_ops = EARLY_COUNTER;
143 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
144     CTLFLAG_RD, &queue_ops,
145     "Number of batched queue operations");
146 
147 static counter_u64_t queue_nops = EARLY_COUNTER;
148 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
149     CTLFLAG_RD, &queue_nops,
150     "Number of batched queue operations with no effects");
151 
152 static void
153 counter_startup(void)
154 {
155 
156 	pqstate_commit_retries = counter_u64_alloc(M_WAITOK);
157 	queue_ops = counter_u64_alloc(M_WAITOK);
158 	queue_nops = counter_u64_alloc(M_WAITOK);
159 }
160 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
161 
162 /*
163  * bogus page -- for I/O to/from partially complete buffers,
164  * or for paging into sparsely invalid regions.
165  */
166 vm_page_t bogus_page;
167 
168 vm_page_t vm_page_array;
169 long vm_page_array_size;
170 long first_page;
171 
172 static int boot_pages;
173 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
174     &boot_pages, 0,
175     "number of pages allocated for bootstrapping the VM system");
176 
177 static TAILQ_HEAD(, vm_page) blacklist_head;
178 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
179 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
180     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
181 
182 static uma_zone_t fakepg_zone;
183 
184 static void vm_page_alloc_check(vm_page_t m);
185 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
186     const char *wmesg, bool nonshared, bool locked);
187 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
188 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
189 static bool vm_page_free_prep(vm_page_t m);
190 static void vm_page_free_toq(vm_page_t m);
191 static void vm_page_init(void *dummy);
192 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
193     vm_pindex_t pindex, vm_page_t mpred);
194 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
195     vm_page_t mpred);
196 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
197     const uint16_t nflag);
198 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
199     vm_page_t m_run, vm_paddr_t high);
200 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
201 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
202     int req);
203 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
204     int flags);
205 static void vm_page_zone_release(void *arg, void **store, int cnt);
206 
207 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
208 
209 static void
210 vm_page_init(void *dummy)
211 {
212 
213 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
214 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
215 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
216 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
217 }
218 
219 /*
220  * The cache page zone is initialized later since we need to be able to allocate
221  * pages before UMA is fully initialized.
222  */
223 static void
224 vm_page_init_cache_zones(void *dummy __unused)
225 {
226 	struct vm_domain *vmd;
227 	struct vm_pgcache *pgcache;
228 	int cache, domain, maxcache, pool;
229 
230 	maxcache = 0;
231 	TUNABLE_INT_FETCH("vm.pgcache_zone_max", &maxcache);
232 	for (domain = 0; domain < vm_ndomains; domain++) {
233 		vmd = VM_DOMAIN(domain);
234 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
235 			pgcache = &vmd->vmd_pgcache[pool];
236 			pgcache->domain = domain;
237 			pgcache->pool = pool;
238 			pgcache->zone = uma_zcache_create("vm pgcache",
239 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
240 			    vm_page_zone_import, vm_page_zone_release, pgcache,
241 			    UMA_ZONE_VM);
242 
243 			/*
244 			 * Limit each pool's zone to 0.1% of the pages in the
245 			 * domain.
246 			 */
247 			cache = maxcache != 0 ? maxcache :
248 			    vmd->vmd_page_count / 1000;
249 			uma_zone_set_maxcache(pgcache->zone, cache);
250 		}
251 	}
252 }
253 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
254 
255 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
256 #if PAGE_SIZE == 32768
257 #ifdef CTASSERT
258 CTASSERT(sizeof(u_long) >= 8);
259 #endif
260 #endif
261 
262 /*
263  *	vm_set_page_size:
264  *
265  *	Sets the page size, perhaps based upon the memory
266  *	size.  Must be called before any use of page-size
267  *	dependent functions.
268  */
269 void
270 vm_set_page_size(void)
271 {
272 	if (vm_cnt.v_page_size == 0)
273 		vm_cnt.v_page_size = PAGE_SIZE;
274 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
275 		panic("vm_set_page_size: page size not a power of two");
276 }
277 
278 /*
279  *	vm_page_blacklist_next:
280  *
281  *	Find the next entry in the provided string of blacklist
282  *	addresses.  Entries are separated by space, comma, or newline.
283  *	If an invalid integer is encountered then the rest of the
284  *	string is skipped.  Updates the list pointer to the next
285  *	character, or NULL if the string is exhausted or invalid.
286  */
287 static vm_paddr_t
288 vm_page_blacklist_next(char **list, char *end)
289 {
290 	vm_paddr_t bad;
291 	char *cp, *pos;
292 
293 	if (list == NULL || *list == NULL)
294 		return (0);
295 	if (**list =='\0') {
296 		*list = NULL;
297 		return (0);
298 	}
299 
300 	/*
301 	 * If there's no end pointer then the buffer is coming from
302 	 * the kenv and we know it's null-terminated.
303 	 */
304 	if (end == NULL)
305 		end = *list + strlen(*list);
306 
307 	/* Ensure that strtoq() won't walk off the end */
308 	if (*end != '\0') {
309 		if (*end == '\n' || *end == ' ' || *end  == ',')
310 			*end = '\0';
311 		else {
312 			printf("Blacklist not terminated, skipping\n");
313 			*list = NULL;
314 			return (0);
315 		}
316 	}
317 
318 	for (pos = *list; *pos != '\0'; pos = cp) {
319 		bad = strtoq(pos, &cp, 0);
320 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
321 			if (bad == 0) {
322 				if (++cp < end)
323 					continue;
324 				else
325 					break;
326 			}
327 		} else
328 			break;
329 		if (*cp == '\0' || ++cp >= end)
330 			*list = NULL;
331 		else
332 			*list = cp;
333 		return (trunc_page(bad));
334 	}
335 	printf("Garbage in RAM blacklist, skipping\n");
336 	*list = NULL;
337 	return (0);
338 }
339 
340 bool
341 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
342 {
343 	struct vm_domain *vmd;
344 	vm_page_t m;
345 	int ret;
346 
347 	m = vm_phys_paddr_to_vm_page(pa);
348 	if (m == NULL)
349 		return (true); /* page does not exist, no failure */
350 
351 	vmd = vm_pagequeue_domain(m);
352 	vm_domain_free_lock(vmd);
353 	ret = vm_phys_unfree_page(m);
354 	vm_domain_free_unlock(vmd);
355 	if (ret != 0) {
356 		vm_domain_freecnt_inc(vmd, -1);
357 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
358 		if (verbose)
359 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
360 	}
361 	return (ret);
362 }
363 
364 /*
365  *	vm_page_blacklist_check:
366  *
367  *	Iterate through the provided string of blacklist addresses, pulling
368  *	each entry out of the physical allocator free list and putting it
369  *	onto a list for reporting via the vm.page_blacklist sysctl.
370  */
371 static void
372 vm_page_blacklist_check(char *list, char *end)
373 {
374 	vm_paddr_t pa;
375 	char *next;
376 
377 	next = list;
378 	while (next != NULL) {
379 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
380 			continue;
381 		vm_page_blacklist_add(pa, bootverbose);
382 	}
383 }
384 
385 /*
386  *	vm_page_blacklist_load:
387  *
388  *	Search for a special module named "ram_blacklist".  It'll be a
389  *	plain text file provided by the user via the loader directive
390  *	of the same name.
391  */
392 static void
393 vm_page_blacklist_load(char **list, char **end)
394 {
395 	void *mod;
396 	u_char *ptr;
397 	u_int len;
398 
399 	mod = NULL;
400 	ptr = NULL;
401 
402 	mod = preload_search_by_type("ram_blacklist");
403 	if (mod != NULL) {
404 		ptr = preload_fetch_addr(mod);
405 		len = preload_fetch_size(mod);
406         }
407 	*list = ptr;
408 	if (ptr != NULL)
409 		*end = ptr + len;
410 	else
411 		*end = NULL;
412 	return;
413 }
414 
415 static int
416 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
417 {
418 	vm_page_t m;
419 	struct sbuf sbuf;
420 	int error, first;
421 
422 	first = 1;
423 	error = sysctl_wire_old_buffer(req, 0);
424 	if (error != 0)
425 		return (error);
426 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
427 	TAILQ_FOREACH(m, &blacklist_head, listq) {
428 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
429 		    (uintmax_t)m->phys_addr);
430 		first = 0;
431 	}
432 	error = sbuf_finish(&sbuf);
433 	sbuf_delete(&sbuf);
434 	return (error);
435 }
436 
437 /*
438  * Initialize a dummy page for use in scans of the specified paging queue.
439  * In principle, this function only needs to set the flag PG_MARKER.
440  * Nonetheless, it write busies the page as a safety precaution.
441  */
442 static void
443 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
444 {
445 
446 	bzero(marker, sizeof(*marker));
447 	marker->flags = PG_MARKER;
448 	marker->a.flags = aflags;
449 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
450 	marker->a.queue = queue;
451 }
452 
453 static void
454 vm_page_domain_init(int domain)
455 {
456 	struct vm_domain *vmd;
457 	struct vm_pagequeue *pq;
458 	int i;
459 
460 	vmd = VM_DOMAIN(domain);
461 	bzero(vmd, sizeof(*vmd));
462 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
463 	    "vm inactive pagequeue";
464 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
465 	    "vm active pagequeue";
466 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
467 	    "vm laundry pagequeue";
468 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
469 	    "vm unswappable pagequeue";
470 	vmd->vmd_domain = domain;
471 	vmd->vmd_page_count = 0;
472 	vmd->vmd_free_count = 0;
473 	vmd->vmd_segs = 0;
474 	vmd->vmd_oom = FALSE;
475 	for (i = 0; i < PQ_COUNT; i++) {
476 		pq = &vmd->vmd_pagequeues[i];
477 		TAILQ_INIT(&pq->pq_pl);
478 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
479 		    MTX_DEF | MTX_DUPOK);
480 		pq->pq_pdpages = 0;
481 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
482 	}
483 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
484 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
485 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
486 
487 	/*
488 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
489 	 * insertions.
490 	 */
491 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
492 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
493 	    &vmd->vmd_inacthead, plinks.q);
494 
495 	/*
496 	 * The clock pages are used to implement active queue scanning without
497 	 * requeues.  Scans start at clock[0], which is advanced after the scan
498 	 * ends.  When the two clock hands meet, they are reset and scanning
499 	 * resumes from the head of the queue.
500 	 */
501 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
502 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
503 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
504 	    &vmd->vmd_clock[0], plinks.q);
505 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
506 	    &vmd->vmd_clock[1], plinks.q);
507 }
508 
509 /*
510  * Initialize a physical page in preparation for adding it to the free
511  * lists.
512  */
513 static void
514 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
515 {
516 
517 	m->object = NULL;
518 	m->ref_count = 0;
519 	m->busy_lock = VPB_UNBUSIED;
520 	m->flags = m->a.flags = 0;
521 	m->phys_addr = pa;
522 	m->a.queue = PQ_NONE;
523 	m->psind = 0;
524 	m->segind = segind;
525 	m->order = VM_NFREEORDER;
526 	m->pool = VM_FREEPOOL_DEFAULT;
527 	m->valid = m->dirty = 0;
528 	pmap_page_init(m);
529 }
530 
531 #ifndef PMAP_HAS_PAGE_ARRAY
532 static vm_paddr_t
533 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
534 {
535 	vm_paddr_t new_end;
536 
537 	/*
538 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
539 	 * However, because this page is allocated from KVM, out-of-bounds
540 	 * accesses using the direct map will not be trapped.
541 	 */
542 	*vaddr += PAGE_SIZE;
543 
544 	/*
545 	 * Allocate physical memory for the page structures, and map it.
546 	 */
547 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
548 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
549 	    VM_PROT_READ | VM_PROT_WRITE);
550 	vm_page_array_size = page_range;
551 
552 	return (new_end);
553 }
554 #endif
555 
556 /*
557  *	vm_page_startup:
558  *
559  *	Initializes the resident memory module.  Allocates physical memory for
560  *	bootstrapping UMA and some data structures that are used to manage
561  *	physical pages.  Initializes these structures, and populates the free
562  *	page queues.
563  */
564 vm_offset_t
565 vm_page_startup(vm_offset_t vaddr)
566 {
567 	struct vm_phys_seg *seg;
568 	vm_page_t m;
569 	char *list, *listend;
570 	vm_offset_t mapped;
571 	vm_paddr_t end, high_avail, low_avail, new_end, size;
572 	vm_paddr_t page_range __unused;
573 	vm_paddr_t last_pa, pa;
574 	u_long pagecount;
575 	int biggestone, i, segind;
576 #ifdef WITNESS
577 	int witness_size;
578 #endif
579 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
580 	long ii;
581 #endif
582 
583 	vaddr = round_page(vaddr);
584 
585 	vm_phys_early_startup();
586 	biggestone = vm_phys_avail_largest();
587 	end = phys_avail[biggestone+1];
588 
589 	/*
590 	 * Initialize the page and queue locks.
591 	 */
592 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
593 	for (i = 0; i < PA_LOCK_COUNT; i++)
594 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
595 	for (i = 0; i < vm_ndomains; i++)
596 		vm_page_domain_init(i);
597 
598 	/*
599 	 * Allocate memory for use when boot strapping the kernel memory
600 	 * allocator.  Tell UMA how many zones we are going to create
601 	 * before going fully functional.  UMA will add its zones.
602 	 *
603 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
604 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
605 	 */
606 	boot_pages = uma_startup_count(8);
607 
608 #ifndef UMA_MD_SMALL_ALLOC
609 	/* vmem_startup() calls uma_prealloc(). */
610 	boot_pages += vmem_startup_count();
611 	/* vm_map_startup() calls uma_prealloc(). */
612 	boot_pages += howmany(MAX_KMAP,
613 	    slab_ipers(sizeof(struct vm_map), UMA_ALIGN_PTR));
614 
615 	/*
616 	 * Before going fully functional kmem_init() does allocation
617 	 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
618 	 */
619 	boot_pages += 2;
620 #endif
621 	/*
622 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
623 	 * manually fetch the value.
624 	 */
625 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
626 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
627 	new_end = trunc_page(new_end);
628 	mapped = pmap_map(&vaddr, new_end, end,
629 	    VM_PROT_READ | VM_PROT_WRITE);
630 	bzero((void *)mapped, end - new_end);
631 	uma_startup((void *)mapped, boot_pages);
632 
633 #ifdef WITNESS
634 	witness_size = round_page(witness_startup_count());
635 	new_end -= witness_size;
636 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
637 	    VM_PROT_READ | VM_PROT_WRITE);
638 	bzero((void *)mapped, witness_size);
639 	witness_startup((void *)mapped);
640 #endif
641 
642 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
643     defined(__i386__) || defined(__mips__) || defined(__riscv) || \
644     defined(__powerpc64__)
645 	/*
646 	 * Allocate a bitmap to indicate that a random physical page
647 	 * needs to be included in a minidump.
648 	 *
649 	 * The amd64 port needs this to indicate which direct map pages
650 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
651 	 *
652 	 * However, i386 still needs this workspace internally within the
653 	 * minidump code.  In theory, they are not needed on i386, but are
654 	 * included should the sf_buf code decide to use them.
655 	 */
656 	last_pa = 0;
657 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
658 		if (dump_avail[i + 1] > last_pa)
659 			last_pa = dump_avail[i + 1];
660 	page_range = last_pa / PAGE_SIZE;
661 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
662 	new_end -= vm_page_dump_size;
663 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
664 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
665 	bzero((void *)vm_page_dump, vm_page_dump_size);
666 #else
667 	(void)last_pa;
668 #endif
669 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
670     defined(__riscv) || defined(__powerpc64__)
671 	/*
672 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
673 	 * in a crash dump.  When pmap_map() uses the direct map, they are
674 	 * not automatically included.
675 	 */
676 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
677 		dump_add_page(pa);
678 #endif
679 	phys_avail[biggestone + 1] = new_end;
680 #ifdef __amd64__
681 	/*
682 	 * Request that the physical pages underlying the message buffer be
683 	 * included in a crash dump.  Since the message buffer is accessed
684 	 * through the direct map, they are not automatically included.
685 	 */
686 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
687 	last_pa = pa + round_page(msgbufsize);
688 	while (pa < last_pa) {
689 		dump_add_page(pa);
690 		pa += PAGE_SIZE;
691 	}
692 #endif
693 	/*
694 	 * Compute the number of pages of memory that will be available for
695 	 * use, taking into account the overhead of a page structure per page.
696 	 * In other words, solve
697 	 *	"available physical memory" - round_page(page_range *
698 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
699 	 * for page_range.
700 	 */
701 	low_avail = phys_avail[0];
702 	high_avail = phys_avail[1];
703 	for (i = 0; i < vm_phys_nsegs; i++) {
704 		if (vm_phys_segs[i].start < low_avail)
705 			low_avail = vm_phys_segs[i].start;
706 		if (vm_phys_segs[i].end > high_avail)
707 			high_avail = vm_phys_segs[i].end;
708 	}
709 	/* Skip the first chunk.  It is already accounted for. */
710 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
711 		if (phys_avail[i] < low_avail)
712 			low_avail = phys_avail[i];
713 		if (phys_avail[i + 1] > high_avail)
714 			high_avail = phys_avail[i + 1];
715 	}
716 	first_page = low_avail / PAGE_SIZE;
717 #ifdef VM_PHYSSEG_SPARSE
718 	size = 0;
719 	for (i = 0; i < vm_phys_nsegs; i++)
720 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
721 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
722 		size += phys_avail[i + 1] - phys_avail[i];
723 #elif defined(VM_PHYSSEG_DENSE)
724 	size = high_avail - low_avail;
725 #else
726 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
727 #endif
728 
729 #ifdef PMAP_HAS_PAGE_ARRAY
730 	pmap_page_array_startup(size / PAGE_SIZE);
731 	biggestone = vm_phys_avail_largest();
732 	end = new_end = phys_avail[biggestone + 1];
733 #else
734 #ifdef VM_PHYSSEG_DENSE
735 	/*
736 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
737 	 * the overhead of a page structure per page only if vm_page_array is
738 	 * allocated from the last physical memory chunk.  Otherwise, we must
739 	 * allocate page structures representing the physical memory
740 	 * underlying vm_page_array, even though they will not be used.
741 	 */
742 	if (new_end != high_avail)
743 		page_range = size / PAGE_SIZE;
744 	else
745 #endif
746 	{
747 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
748 
749 		/*
750 		 * If the partial bytes remaining are large enough for
751 		 * a page (PAGE_SIZE) without a corresponding
752 		 * 'struct vm_page', then new_end will contain an
753 		 * extra page after subtracting the length of the VM
754 		 * page array.  Compensate by subtracting an extra
755 		 * page from new_end.
756 		 */
757 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
758 			if (new_end == high_avail)
759 				high_avail -= PAGE_SIZE;
760 			new_end -= PAGE_SIZE;
761 		}
762 	}
763 	end = new_end;
764 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
765 #endif
766 
767 #if VM_NRESERVLEVEL > 0
768 	/*
769 	 * Allocate physical memory for the reservation management system's
770 	 * data structures, and map it.
771 	 */
772 	new_end = vm_reserv_startup(&vaddr, new_end);
773 #endif
774 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
775     defined(__riscv) || defined(__powerpc64__)
776 	/*
777 	 * Include vm_page_array and vm_reserv_array in a crash dump.
778 	 */
779 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
780 		dump_add_page(pa);
781 #endif
782 	phys_avail[biggestone + 1] = new_end;
783 
784 	/*
785 	 * Add physical memory segments corresponding to the available
786 	 * physical pages.
787 	 */
788 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
789 		if (vm_phys_avail_size(i) != 0)
790 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
791 
792 	/*
793 	 * Initialize the physical memory allocator.
794 	 */
795 	vm_phys_init();
796 
797 	/*
798 	 * Initialize the page structures and add every available page to the
799 	 * physical memory allocator's free lists.
800 	 */
801 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
802 	for (ii = 0; ii < vm_page_array_size; ii++) {
803 		m = &vm_page_array[ii];
804 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
805 		m->flags = PG_FICTITIOUS;
806 	}
807 #endif
808 	vm_cnt.v_page_count = 0;
809 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
810 		seg = &vm_phys_segs[segind];
811 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
812 		    m++, pa += PAGE_SIZE)
813 			vm_page_init_page(m, pa, segind);
814 
815 		/*
816 		 * Add the segment to the free lists only if it is covered by
817 		 * one of the ranges in phys_avail.  Because we've added the
818 		 * ranges to the vm_phys_segs array, we can assume that each
819 		 * segment is either entirely contained in one of the ranges,
820 		 * or doesn't overlap any of them.
821 		 */
822 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
823 			struct vm_domain *vmd;
824 
825 			if (seg->start < phys_avail[i] ||
826 			    seg->end > phys_avail[i + 1])
827 				continue;
828 
829 			m = seg->first_page;
830 			pagecount = (u_long)atop(seg->end - seg->start);
831 
832 			vmd = VM_DOMAIN(seg->domain);
833 			vm_domain_free_lock(vmd);
834 			vm_phys_enqueue_contig(m, pagecount);
835 			vm_domain_free_unlock(vmd);
836 			vm_domain_freecnt_inc(vmd, pagecount);
837 			vm_cnt.v_page_count += (u_int)pagecount;
838 
839 			vmd = VM_DOMAIN(seg->domain);
840 			vmd->vmd_page_count += (u_int)pagecount;
841 			vmd->vmd_segs |= 1UL << m->segind;
842 			break;
843 		}
844 	}
845 
846 	/*
847 	 * Remove blacklisted pages from the physical memory allocator.
848 	 */
849 	TAILQ_INIT(&blacklist_head);
850 	vm_page_blacklist_load(&list, &listend);
851 	vm_page_blacklist_check(list, listend);
852 
853 	list = kern_getenv("vm.blacklist");
854 	vm_page_blacklist_check(list, NULL);
855 
856 	freeenv(list);
857 #if VM_NRESERVLEVEL > 0
858 	/*
859 	 * Initialize the reservation management system.
860 	 */
861 	vm_reserv_init();
862 #endif
863 
864 	return (vaddr);
865 }
866 
867 void
868 vm_page_reference(vm_page_t m)
869 {
870 
871 	vm_page_aflag_set(m, PGA_REFERENCED);
872 }
873 
874 static bool
875 vm_page_acquire_flags(vm_page_t m, int allocflags)
876 {
877 	bool locked;
878 
879 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
880 		locked = vm_page_trysbusy(m);
881 	else
882 		locked = vm_page_tryxbusy(m);
883 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
884 		vm_page_wire(m);
885 	return (locked);
886 }
887 
888 /*
889  *	vm_page_busy_sleep_flags
890  *
891  *	Sleep for busy according to VM_ALLOC_ parameters.
892  */
893 static bool
894 vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wmesg,
895     int allocflags)
896 {
897 
898 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
899 		return (false);
900 	/*
901 	 * Reference the page before unlocking and
902 	 * sleeping so that the page daemon is less
903 	 * likely to reclaim it.
904 	 */
905 	if ((allocflags & VM_ALLOC_NOCREAT) == 0)
906 		vm_page_aflag_set(m, PGA_REFERENCED);
907 	if (_vm_page_busy_sleep(object, m, wmesg, (allocflags &
908 	    VM_ALLOC_IGN_SBUSY) != 0, true))
909 		VM_OBJECT_WLOCK(object);
910 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
911 		return (false);
912 	return (true);
913 }
914 
915 /*
916  *	vm_page_busy_acquire:
917  *
918  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
919  *	and drop the object lock if necessary.
920  */
921 bool
922 vm_page_busy_acquire(vm_page_t m, int allocflags)
923 {
924 	vm_object_t obj;
925 	bool locked;
926 
927 	/*
928 	 * The page-specific object must be cached because page
929 	 * identity can change during the sleep, causing the
930 	 * re-lock of a different object.
931 	 * It is assumed that a reference to the object is already
932 	 * held by the callers.
933 	 */
934 	obj = m->object;
935 	for (;;) {
936 		if (vm_page_acquire_flags(m, allocflags))
937 			return (true);
938 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
939 			return (false);
940 		if (obj != NULL)
941 			locked = VM_OBJECT_WOWNED(obj);
942 		else
943 			locked = false;
944 		MPASS(locked || vm_page_wired(m));
945 		if (_vm_page_busy_sleep(obj, m, "vmpba",
946 		    (allocflags & VM_ALLOC_SBUSY) != 0, locked))
947 			VM_OBJECT_WLOCK(obj);
948 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
949 			return (false);
950 		KASSERT(m->object == obj || m->object == NULL,
951 		    ("vm_page_busy_acquire: page %p does not belong to %p",
952 		    m, obj));
953 	}
954 }
955 
956 /*
957  *	vm_page_busy_downgrade:
958  *
959  *	Downgrade an exclusive busy page into a single shared busy page.
960  */
961 void
962 vm_page_busy_downgrade(vm_page_t m)
963 {
964 	u_int x;
965 
966 	vm_page_assert_xbusied(m);
967 
968 	x = m->busy_lock;
969 	for (;;) {
970 		if (atomic_fcmpset_rel_int(&m->busy_lock,
971 		    &x, VPB_SHARERS_WORD(1)))
972 			break;
973 	}
974 	if ((x & VPB_BIT_WAITERS) != 0)
975 		wakeup(m);
976 }
977 
978 /*
979  *
980  *	vm_page_busy_tryupgrade:
981  *
982  *	Attempt to upgrade a single shared busy into an exclusive busy.
983  */
984 int
985 vm_page_busy_tryupgrade(vm_page_t m)
986 {
987 	u_int ce, x;
988 
989 	vm_page_assert_sbusied(m);
990 
991 	x = m->busy_lock;
992 	ce = VPB_CURTHREAD_EXCLUSIVE;
993 	for (;;) {
994 		if (VPB_SHARERS(x) > 1)
995 			return (0);
996 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
997 		    ("vm_page_busy_tryupgrade: invalid lock state"));
998 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
999 		    ce | (x & VPB_BIT_WAITERS)))
1000 			continue;
1001 		return (1);
1002 	}
1003 }
1004 
1005 /*
1006  *	vm_page_sbusied:
1007  *
1008  *	Return a positive value if the page is shared busied, 0 otherwise.
1009  */
1010 int
1011 vm_page_sbusied(vm_page_t m)
1012 {
1013 	u_int x;
1014 
1015 	x = m->busy_lock;
1016 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1017 }
1018 
1019 /*
1020  *	vm_page_sunbusy:
1021  *
1022  *	Shared unbusy a page.
1023  */
1024 void
1025 vm_page_sunbusy(vm_page_t m)
1026 {
1027 	u_int x;
1028 
1029 	vm_page_assert_sbusied(m);
1030 
1031 	x = m->busy_lock;
1032 	for (;;) {
1033 		if (VPB_SHARERS(x) > 1) {
1034 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1035 			    x - VPB_ONE_SHARER))
1036 				break;
1037 			continue;
1038 		}
1039 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1040 		    ("vm_page_sunbusy: invalid lock state"));
1041 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1042 			continue;
1043 		if ((x & VPB_BIT_WAITERS) == 0)
1044 			break;
1045 		wakeup(m);
1046 		break;
1047 	}
1048 }
1049 
1050 /*
1051  *	vm_page_busy_sleep:
1052  *
1053  *	Sleep if the page is busy, using the page pointer as wchan.
1054  *	This is used to implement the hard-path of busying mechanism.
1055  *
1056  *	If nonshared is true, sleep only if the page is xbusy.
1057  *
1058  *	The object lock must be held on entry and will be released on exit.
1059  */
1060 void
1061 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1062 {
1063 	vm_object_t obj;
1064 
1065 	obj = m->object;
1066 	VM_OBJECT_ASSERT_LOCKED(obj);
1067 	vm_page_lock_assert(m, MA_NOTOWNED);
1068 
1069 	if (!_vm_page_busy_sleep(obj, m, wmesg, nonshared, true))
1070 		VM_OBJECT_DROP(obj);
1071 }
1072 
1073 /*
1074  *	_vm_page_busy_sleep:
1075  *
1076  *	Internal busy sleep function.
1077  */
1078 static bool
1079 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg,
1080     bool nonshared, bool locked)
1081 {
1082 	u_int x;
1083 
1084 	/*
1085 	 * If the object is busy we must wait for that to drain to zero
1086 	 * before trying the page again.
1087 	 */
1088 	if (obj != NULL && vm_object_busied(obj)) {
1089 		if (locked)
1090 			VM_OBJECT_DROP(obj);
1091 		vm_object_busy_wait(obj, wmesg);
1092 		return (locked);
1093 	}
1094 	sleepq_lock(m);
1095 	x = m->busy_lock;
1096 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
1097 	    ((x & VPB_BIT_WAITERS) == 0 &&
1098 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
1099 		sleepq_release(m);
1100 		return (false);
1101 	}
1102 	if (locked)
1103 		VM_OBJECT_DROP(obj);
1104 	DROP_GIANT();
1105 	sleepq_add(m, NULL, wmesg, 0, 0);
1106 	sleepq_wait(m, PVM);
1107 	PICKUP_GIANT();
1108 	return (locked);
1109 }
1110 
1111 /*
1112  *	vm_page_trysbusy:
1113  *
1114  *	Try to shared busy a page.
1115  *	If the operation succeeds 1 is returned otherwise 0.
1116  *	The operation never sleeps.
1117  */
1118 int
1119 vm_page_trysbusy(vm_page_t m)
1120 {
1121 	vm_object_t obj;
1122 	u_int x;
1123 
1124 	obj = m->object;
1125 	x = m->busy_lock;
1126 	for (;;) {
1127 		if ((x & VPB_BIT_SHARED) == 0)
1128 			return (0);
1129 		/*
1130 		 * Reduce the window for transient busies that will trigger
1131 		 * false negatives in vm_page_ps_test().
1132 		 */
1133 		if (obj != NULL && vm_object_busied(obj))
1134 			return (0);
1135 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1136 		    x + VPB_ONE_SHARER))
1137 			break;
1138 	}
1139 
1140 	/* Refetch the object now that we're guaranteed that it is stable. */
1141 	obj = m->object;
1142 	if (obj != NULL && vm_object_busied(obj)) {
1143 		vm_page_sunbusy(m);
1144 		return (0);
1145 	}
1146 	return (1);
1147 }
1148 
1149 /*
1150  *	vm_page_tryxbusy:
1151  *
1152  *	Try to exclusive busy a page.
1153  *	If the operation succeeds 1 is returned otherwise 0.
1154  *	The operation never sleeps.
1155  */
1156 int
1157 vm_page_tryxbusy(vm_page_t m)
1158 {
1159 	vm_object_t obj;
1160 
1161         if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED,
1162             VPB_CURTHREAD_EXCLUSIVE) == 0)
1163 		return (0);
1164 
1165 	obj = m->object;
1166 	if (obj != NULL && vm_object_busied(obj)) {
1167 		vm_page_xunbusy(m);
1168 		return (0);
1169 	}
1170 	return (1);
1171 }
1172 
1173 static void
1174 vm_page_xunbusy_hard_tail(vm_page_t m)
1175 {
1176 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1177 	/* Wake the waiter. */
1178 	wakeup(m);
1179 }
1180 
1181 /*
1182  *	vm_page_xunbusy_hard:
1183  *
1184  *	Called when unbusy has failed because there is a waiter.
1185  */
1186 void
1187 vm_page_xunbusy_hard(vm_page_t m)
1188 {
1189 	vm_page_assert_xbusied(m);
1190 	vm_page_xunbusy_hard_tail(m);
1191 }
1192 
1193 void
1194 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1195 {
1196 	vm_page_assert_xbusied_unchecked(m);
1197 	vm_page_xunbusy_hard_tail(m);
1198 }
1199 
1200 /*
1201  * Avoid releasing and reacquiring the same page lock.
1202  */
1203 void
1204 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1205 {
1206 	struct mtx *mtx1;
1207 
1208 	mtx1 = vm_page_lockptr(m);
1209 	if (*mtx == mtx1)
1210 		return;
1211 	if (*mtx != NULL)
1212 		mtx_unlock(*mtx);
1213 	*mtx = mtx1;
1214 	mtx_lock(mtx1);
1215 }
1216 
1217 /*
1218  *	vm_page_unhold_pages:
1219  *
1220  *	Unhold each of the pages that is referenced by the given array.
1221  */
1222 void
1223 vm_page_unhold_pages(vm_page_t *ma, int count)
1224 {
1225 
1226 	for (; count != 0; count--) {
1227 		vm_page_unwire(*ma, PQ_ACTIVE);
1228 		ma++;
1229 	}
1230 }
1231 
1232 vm_page_t
1233 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1234 {
1235 	vm_page_t m;
1236 
1237 #ifdef VM_PHYSSEG_SPARSE
1238 	m = vm_phys_paddr_to_vm_page(pa);
1239 	if (m == NULL)
1240 		m = vm_phys_fictitious_to_vm_page(pa);
1241 	return (m);
1242 #elif defined(VM_PHYSSEG_DENSE)
1243 	long pi;
1244 
1245 	pi = atop(pa);
1246 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1247 		m = &vm_page_array[pi - first_page];
1248 		return (m);
1249 	}
1250 	return (vm_phys_fictitious_to_vm_page(pa));
1251 #else
1252 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1253 #endif
1254 }
1255 
1256 /*
1257  *	vm_page_getfake:
1258  *
1259  *	Create a fictitious page with the specified physical address and
1260  *	memory attribute.  The memory attribute is the only the machine-
1261  *	dependent aspect of a fictitious page that must be initialized.
1262  */
1263 vm_page_t
1264 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1265 {
1266 	vm_page_t m;
1267 
1268 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1269 	vm_page_initfake(m, paddr, memattr);
1270 	return (m);
1271 }
1272 
1273 void
1274 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1275 {
1276 
1277 	if ((m->flags & PG_FICTITIOUS) != 0) {
1278 		/*
1279 		 * The page's memattr might have changed since the
1280 		 * previous initialization.  Update the pmap to the
1281 		 * new memattr.
1282 		 */
1283 		goto memattr;
1284 	}
1285 	m->phys_addr = paddr;
1286 	m->a.queue = PQ_NONE;
1287 	/* Fictitious pages don't use "segind". */
1288 	m->flags = PG_FICTITIOUS;
1289 	/* Fictitious pages don't use "order" or "pool". */
1290 	m->oflags = VPO_UNMANAGED;
1291 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1292 	/* Fictitious pages are unevictable. */
1293 	m->ref_count = 1;
1294 	pmap_page_init(m);
1295 memattr:
1296 	pmap_page_set_memattr(m, memattr);
1297 }
1298 
1299 /*
1300  *	vm_page_putfake:
1301  *
1302  *	Release a fictitious page.
1303  */
1304 void
1305 vm_page_putfake(vm_page_t m)
1306 {
1307 
1308 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1309 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1310 	    ("vm_page_putfake: bad page %p", m));
1311 	vm_page_xunbusy(m);
1312 	uma_zfree(fakepg_zone, m);
1313 }
1314 
1315 /*
1316  *	vm_page_updatefake:
1317  *
1318  *	Update the given fictitious page to the specified physical address and
1319  *	memory attribute.
1320  */
1321 void
1322 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1323 {
1324 
1325 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1326 	    ("vm_page_updatefake: bad page %p", m));
1327 	m->phys_addr = paddr;
1328 	pmap_page_set_memattr(m, memattr);
1329 }
1330 
1331 /*
1332  *	vm_page_free:
1333  *
1334  *	Free a page.
1335  */
1336 void
1337 vm_page_free(vm_page_t m)
1338 {
1339 
1340 	m->flags &= ~PG_ZERO;
1341 	vm_page_free_toq(m);
1342 }
1343 
1344 /*
1345  *	vm_page_free_zero:
1346  *
1347  *	Free a page to the zerod-pages queue
1348  */
1349 void
1350 vm_page_free_zero(vm_page_t m)
1351 {
1352 
1353 	m->flags |= PG_ZERO;
1354 	vm_page_free_toq(m);
1355 }
1356 
1357 /*
1358  * Unbusy and handle the page queueing for a page from a getpages request that
1359  * was optionally read ahead or behind.
1360  */
1361 void
1362 vm_page_readahead_finish(vm_page_t m)
1363 {
1364 
1365 	/* We shouldn't put invalid pages on queues. */
1366 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1367 
1368 	/*
1369 	 * Since the page is not the actually needed one, whether it should
1370 	 * be activated or deactivated is not obvious.  Empirical results
1371 	 * have shown that deactivating the page is usually the best choice,
1372 	 * unless the page is wanted by another thread.
1373 	 */
1374 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1375 		vm_page_activate(m);
1376 	else
1377 		vm_page_deactivate(m);
1378 	vm_page_xunbusy_unchecked(m);
1379 }
1380 
1381 /*
1382  *	vm_page_sleep_if_busy:
1383  *
1384  *	Sleep and release the object lock if the page is busied.
1385  *	Returns TRUE if the thread slept.
1386  *
1387  *	The given page must be unlocked and object containing it must
1388  *	be locked.
1389  */
1390 int
1391 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1392 {
1393 	vm_object_t obj;
1394 
1395 	vm_page_lock_assert(m, MA_NOTOWNED);
1396 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1397 
1398 	/*
1399 	 * The page-specific object must be cached because page
1400 	 * identity can change during the sleep, causing the
1401 	 * re-lock of a different object.
1402 	 * It is assumed that a reference to the object is already
1403 	 * held by the callers.
1404 	 */
1405 	obj = m->object;
1406 	if (vm_page_busied(m) || (obj != NULL && obj->busy)) {
1407 		vm_page_busy_sleep(m, msg, false);
1408 		VM_OBJECT_WLOCK(obj);
1409 		return (TRUE);
1410 	}
1411 	return (FALSE);
1412 }
1413 
1414 /*
1415  *	vm_page_sleep_if_xbusy:
1416  *
1417  *	Sleep and release the object lock if the page is xbusied.
1418  *	Returns TRUE if the thread slept.
1419  *
1420  *	The given page must be unlocked and object containing it must
1421  *	be locked.
1422  */
1423 int
1424 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg)
1425 {
1426 	vm_object_t obj;
1427 
1428 	vm_page_lock_assert(m, MA_NOTOWNED);
1429 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1430 
1431 	/*
1432 	 * The page-specific object must be cached because page
1433 	 * identity can change during the sleep, causing the
1434 	 * re-lock of a different object.
1435 	 * It is assumed that a reference to the object is already
1436 	 * held by the callers.
1437 	 */
1438 	obj = m->object;
1439 	if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) {
1440 		vm_page_busy_sleep(m, msg, true);
1441 		VM_OBJECT_WLOCK(obj);
1442 		return (TRUE);
1443 	}
1444 	return (FALSE);
1445 }
1446 
1447 /*
1448  *	vm_page_dirty_KBI:		[ internal use only ]
1449  *
1450  *	Set all bits in the page's dirty field.
1451  *
1452  *	The object containing the specified page must be locked if the
1453  *	call is made from the machine-independent layer.
1454  *
1455  *	See vm_page_clear_dirty_mask().
1456  *
1457  *	This function should only be called by vm_page_dirty().
1458  */
1459 void
1460 vm_page_dirty_KBI(vm_page_t m)
1461 {
1462 
1463 	/* Refer to this operation by its public name. */
1464 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1465 	m->dirty = VM_PAGE_BITS_ALL;
1466 }
1467 
1468 /*
1469  *	vm_page_insert:		[ internal use only ]
1470  *
1471  *	Inserts the given mem entry into the object and object list.
1472  *
1473  *	The object must be locked.
1474  */
1475 int
1476 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1477 {
1478 	vm_page_t mpred;
1479 
1480 	VM_OBJECT_ASSERT_WLOCKED(object);
1481 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1482 	return (vm_page_insert_after(m, object, pindex, mpred));
1483 }
1484 
1485 /*
1486  *	vm_page_insert_after:
1487  *
1488  *	Inserts the page "m" into the specified object at offset "pindex".
1489  *
1490  *	The page "mpred" must immediately precede the offset "pindex" within
1491  *	the specified object.
1492  *
1493  *	The object must be locked.
1494  */
1495 static int
1496 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1497     vm_page_t mpred)
1498 {
1499 	vm_page_t msucc;
1500 
1501 	VM_OBJECT_ASSERT_WLOCKED(object);
1502 	KASSERT(m->object == NULL,
1503 	    ("vm_page_insert_after: page already inserted"));
1504 	if (mpred != NULL) {
1505 		KASSERT(mpred->object == object,
1506 		    ("vm_page_insert_after: object doesn't contain mpred"));
1507 		KASSERT(mpred->pindex < pindex,
1508 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1509 		msucc = TAILQ_NEXT(mpred, listq);
1510 	} else
1511 		msucc = TAILQ_FIRST(&object->memq);
1512 	if (msucc != NULL)
1513 		KASSERT(msucc->pindex > pindex,
1514 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1515 
1516 	/*
1517 	 * Record the object/offset pair in this page.
1518 	 */
1519 	m->object = object;
1520 	m->pindex = pindex;
1521 	m->ref_count |= VPRC_OBJREF;
1522 
1523 	/*
1524 	 * Now link into the object's ordered list of backed pages.
1525 	 */
1526 	if (vm_radix_insert(&object->rtree, m)) {
1527 		m->object = NULL;
1528 		m->pindex = 0;
1529 		m->ref_count &= ~VPRC_OBJREF;
1530 		return (1);
1531 	}
1532 	vm_page_insert_radixdone(m, object, mpred);
1533 	return (0);
1534 }
1535 
1536 /*
1537  *	vm_page_insert_radixdone:
1538  *
1539  *	Complete page "m" insertion into the specified object after the
1540  *	radix trie hooking.
1541  *
1542  *	The page "mpred" must precede the offset "m->pindex" within the
1543  *	specified object.
1544  *
1545  *	The object must be locked.
1546  */
1547 static void
1548 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1549 {
1550 
1551 	VM_OBJECT_ASSERT_WLOCKED(object);
1552 	KASSERT(object != NULL && m->object == object,
1553 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1554 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1555 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1556 	if (mpred != NULL) {
1557 		KASSERT(mpred->object == object,
1558 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1559 		KASSERT(mpred->pindex < m->pindex,
1560 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1561 	}
1562 
1563 	if (mpred != NULL)
1564 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1565 	else
1566 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1567 
1568 	/*
1569 	 * Show that the object has one more resident page.
1570 	 */
1571 	object->resident_page_count++;
1572 
1573 	/*
1574 	 * Hold the vnode until the last page is released.
1575 	 */
1576 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1577 		vhold(object->handle);
1578 
1579 	/*
1580 	 * Since we are inserting a new and possibly dirty page,
1581 	 * update the object's generation count.
1582 	 */
1583 	if (pmap_page_is_write_mapped(m))
1584 		vm_object_set_writeable_dirty(object);
1585 }
1586 
1587 /*
1588  * Do the work to remove a page from its object.  The caller is responsible for
1589  * updating the page's fields to reflect this removal.
1590  */
1591 static void
1592 vm_page_object_remove(vm_page_t m)
1593 {
1594 	vm_object_t object;
1595 	vm_page_t mrem;
1596 
1597 	vm_page_assert_xbusied(m);
1598 	object = m->object;
1599 	VM_OBJECT_ASSERT_WLOCKED(object);
1600 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1601 	    ("page %p is missing its object ref", m));
1602 
1603 	/* Deferred free of swap space. */
1604 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1605 		vm_pager_page_unswapped(m);
1606 
1607 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1608 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1609 
1610 	/*
1611 	 * Now remove from the object's list of backed pages.
1612 	 */
1613 	TAILQ_REMOVE(&object->memq, m, listq);
1614 
1615 	/*
1616 	 * And show that the object has one fewer resident page.
1617 	 */
1618 	object->resident_page_count--;
1619 
1620 	/*
1621 	 * The vnode may now be recycled.
1622 	 */
1623 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1624 		vdrop(object->handle);
1625 }
1626 
1627 /*
1628  *	vm_page_remove:
1629  *
1630  *	Removes the specified page from its containing object, but does not
1631  *	invalidate any backing storage.  Returns true if the object's reference
1632  *	was the last reference to the page, and false otherwise.
1633  *
1634  *	The object must be locked and the page must be exclusively busied.
1635  *	The exclusive busy will be released on return.  If this is not the
1636  *	final ref and the caller does not hold a wire reference it may not
1637  *	continue to access the page.
1638  */
1639 bool
1640 vm_page_remove(vm_page_t m)
1641 {
1642 	bool dropped;
1643 
1644 	dropped = vm_page_remove_xbusy(m);
1645 	vm_page_xunbusy(m);
1646 
1647 	return (dropped);
1648 }
1649 
1650 /*
1651  *	vm_page_remove_xbusy
1652  *
1653  *	Removes the page but leaves the xbusy held.  Returns true if this
1654  *	removed the final ref and false otherwise.
1655  */
1656 bool
1657 vm_page_remove_xbusy(vm_page_t m)
1658 {
1659 
1660 	vm_page_object_remove(m);
1661 	m->object = NULL;
1662 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1663 }
1664 
1665 /*
1666  *	vm_page_lookup:
1667  *
1668  *	Returns the page associated with the object/offset
1669  *	pair specified; if none is found, NULL is returned.
1670  *
1671  *	The object must be locked.
1672  */
1673 vm_page_t
1674 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1675 {
1676 
1677 	VM_OBJECT_ASSERT_LOCKED(object);
1678 	return (vm_radix_lookup(&object->rtree, pindex));
1679 }
1680 
1681 /*
1682  *	vm_page_find_least:
1683  *
1684  *	Returns the page associated with the object with least pindex
1685  *	greater than or equal to the parameter pindex, or NULL.
1686  *
1687  *	The object must be locked.
1688  */
1689 vm_page_t
1690 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1691 {
1692 	vm_page_t m;
1693 
1694 	VM_OBJECT_ASSERT_LOCKED(object);
1695 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1696 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1697 	return (m);
1698 }
1699 
1700 /*
1701  * Returns the given page's successor (by pindex) within the object if it is
1702  * resident; if none is found, NULL is returned.
1703  *
1704  * The object must be locked.
1705  */
1706 vm_page_t
1707 vm_page_next(vm_page_t m)
1708 {
1709 	vm_page_t next;
1710 
1711 	VM_OBJECT_ASSERT_LOCKED(m->object);
1712 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1713 		MPASS(next->object == m->object);
1714 		if (next->pindex != m->pindex + 1)
1715 			next = NULL;
1716 	}
1717 	return (next);
1718 }
1719 
1720 /*
1721  * Returns the given page's predecessor (by pindex) within the object if it is
1722  * resident; if none is found, NULL is returned.
1723  *
1724  * The object must be locked.
1725  */
1726 vm_page_t
1727 vm_page_prev(vm_page_t m)
1728 {
1729 	vm_page_t prev;
1730 
1731 	VM_OBJECT_ASSERT_LOCKED(m->object);
1732 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1733 		MPASS(prev->object == m->object);
1734 		if (prev->pindex != m->pindex - 1)
1735 			prev = NULL;
1736 	}
1737 	return (prev);
1738 }
1739 
1740 /*
1741  * Uses the page mnew as a replacement for an existing page at index
1742  * pindex which must be already present in the object.
1743  *
1744  * Both pages must be exclusively busied on enter.  The old page is
1745  * unbusied on exit.
1746  *
1747  * A return value of true means mold is now free.  If this is not the
1748  * final ref and the caller does not hold a wire reference it may not
1749  * continue to access the page.
1750  */
1751 static bool
1752 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1753     vm_page_t mold)
1754 {
1755 	vm_page_t mret;
1756 	bool dropped;
1757 
1758 	VM_OBJECT_ASSERT_WLOCKED(object);
1759 	vm_page_assert_xbusied(mold);
1760 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1761 	    ("vm_page_replace: page %p already in object", mnew));
1762 
1763 	/*
1764 	 * This function mostly follows vm_page_insert() and
1765 	 * vm_page_remove() without the radix, object count and vnode
1766 	 * dance.  Double check such functions for more comments.
1767 	 */
1768 
1769 	mnew->object = object;
1770 	mnew->pindex = pindex;
1771 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1772 	mret = vm_radix_replace(&object->rtree, mnew);
1773 	KASSERT(mret == mold,
1774 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1775 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1776 	    (mnew->oflags & VPO_UNMANAGED),
1777 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1778 
1779 	/* Keep the resident page list in sorted order. */
1780 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1781 	TAILQ_REMOVE(&object->memq, mold, listq);
1782 	mold->object = NULL;
1783 
1784 	/*
1785 	 * The object's resident_page_count does not change because we have
1786 	 * swapped one page for another, but the generation count should
1787 	 * change if the page is dirty.
1788 	 */
1789 	if (pmap_page_is_write_mapped(mnew))
1790 		vm_object_set_writeable_dirty(object);
1791 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1792 	vm_page_xunbusy(mold);
1793 
1794 	return (dropped);
1795 }
1796 
1797 void
1798 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1799     vm_page_t mold)
1800 {
1801 
1802 	vm_page_assert_xbusied(mnew);
1803 
1804 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1805 		vm_page_free(mold);
1806 }
1807 
1808 /*
1809  *	vm_page_rename:
1810  *
1811  *	Move the given memory entry from its
1812  *	current object to the specified target object/offset.
1813  *
1814  *	Note: swap associated with the page must be invalidated by the move.  We
1815  *	      have to do this for several reasons:  (1) we aren't freeing the
1816  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1817  *	      moving the page from object A to B, and will then later move
1818  *	      the backing store from A to B and we can't have a conflict.
1819  *
1820  *	Note: we *always* dirty the page.  It is necessary both for the
1821  *	      fact that we moved it, and because we may be invalidating
1822  *	      swap.
1823  *
1824  *	The objects must be locked.
1825  */
1826 int
1827 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1828 {
1829 	vm_page_t mpred;
1830 	vm_pindex_t opidx;
1831 
1832 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1833 
1834 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1835 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1836 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1837 	    ("vm_page_rename: pindex already renamed"));
1838 
1839 	/*
1840 	 * Create a custom version of vm_page_insert() which does not depend
1841 	 * by m_prev and can cheat on the implementation aspects of the
1842 	 * function.
1843 	 */
1844 	opidx = m->pindex;
1845 	m->pindex = new_pindex;
1846 	if (vm_radix_insert(&new_object->rtree, m)) {
1847 		m->pindex = opidx;
1848 		return (1);
1849 	}
1850 
1851 	/*
1852 	 * The operation cannot fail anymore.  The removal must happen before
1853 	 * the listq iterator is tainted.
1854 	 */
1855 	m->pindex = opidx;
1856 	vm_page_object_remove(m);
1857 
1858 	/* Return back to the new pindex to complete vm_page_insert(). */
1859 	m->pindex = new_pindex;
1860 	m->object = new_object;
1861 
1862 	vm_page_insert_radixdone(m, new_object, mpred);
1863 	vm_page_dirty(m);
1864 	return (0);
1865 }
1866 
1867 /*
1868  *	vm_page_alloc:
1869  *
1870  *	Allocate and return a page that is associated with the specified
1871  *	object and offset pair.  By default, this page is exclusive busied.
1872  *
1873  *	The caller must always specify an allocation class.
1874  *
1875  *	allocation classes:
1876  *	VM_ALLOC_NORMAL		normal process request
1877  *	VM_ALLOC_SYSTEM		system *really* needs a page
1878  *	VM_ALLOC_INTERRUPT	interrupt time request
1879  *
1880  *	optional allocation flags:
1881  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1882  *				intends to allocate
1883  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1884  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1885  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1886  *				should not be exclusive busy
1887  *	VM_ALLOC_SBUSY		shared busy the allocated page
1888  *	VM_ALLOC_WIRED		wire the allocated page
1889  *	VM_ALLOC_ZERO		prefer a zeroed page
1890  */
1891 vm_page_t
1892 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1893 {
1894 
1895 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1896 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1897 }
1898 
1899 vm_page_t
1900 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1901     int req)
1902 {
1903 
1904 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1905 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1906 	    NULL));
1907 }
1908 
1909 /*
1910  * Allocate a page in the specified object with the given page index.  To
1911  * optimize insertion of the page into the object, the caller must also specifiy
1912  * the resident page in the object with largest index smaller than the given
1913  * page index, or NULL if no such page exists.
1914  */
1915 vm_page_t
1916 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1917     int req, vm_page_t mpred)
1918 {
1919 	struct vm_domainset_iter di;
1920 	vm_page_t m;
1921 	int domain;
1922 
1923 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1924 	do {
1925 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1926 		    mpred);
1927 		if (m != NULL)
1928 			break;
1929 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1930 
1931 	return (m);
1932 }
1933 
1934 /*
1935  * Returns true if the number of free pages exceeds the minimum
1936  * for the request class and false otherwise.
1937  */
1938 static int
1939 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1940 {
1941 	u_int limit, old, new;
1942 
1943 	if (req_class == VM_ALLOC_INTERRUPT)
1944 		limit = 0;
1945 	else if (req_class == VM_ALLOC_SYSTEM)
1946 		limit = vmd->vmd_interrupt_free_min;
1947 	else
1948 		limit = vmd->vmd_free_reserved;
1949 
1950 	/*
1951 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1952 	 */
1953 	limit += npages;
1954 	old = vmd->vmd_free_count;
1955 	do {
1956 		if (old < limit)
1957 			return (0);
1958 		new = old - npages;
1959 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1960 
1961 	/* Wake the page daemon if we've crossed the threshold. */
1962 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1963 		pagedaemon_wakeup(vmd->vmd_domain);
1964 
1965 	/* Only update bitsets on transitions. */
1966 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1967 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1968 		vm_domain_set(vmd);
1969 
1970 	return (1);
1971 }
1972 
1973 int
1974 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1975 {
1976 	int req_class;
1977 
1978 	/*
1979 	 * The page daemon is allowed to dig deeper into the free page list.
1980 	 */
1981 	req_class = req & VM_ALLOC_CLASS_MASK;
1982 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1983 		req_class = VM_ALLOC_SYSTEM;
1984 	return (_vm_domain_allocate(vmd, req_class, npages));
1985 }
1986 
1987 vm_page_t
1988 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1989     int req, vm_page_t mpred)
1990 {
1991 	struct vm_domain *vmd;
1992 	vm_page_t m;
1993 	int flags, pool;
1994 
1995 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1996 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1997 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1998 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1999 	    ("inconsistent object(%p)/req(%x)", object, req));
2000 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2001 	    ("Can't sleep and retry object insertion."));
2002 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2003 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2004 	    (uintmax_t)pindex));
2005 	if (object != NULL)
2006 		VM_OBJECT_ASSERT_WLOCKED(object);
2007 
2008 	flags = 0;
2009 	m = NULL;
2010 	pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2011 again:
2012 #if VM_NRESERVLEVEL > 0
2013 	/*
2014 	 * Can we allocate the page from a reservation?
2015 	 */
2016 	if (vm_object_reserv(object) &&
2017 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2018 	    NULL) {
2019 		domain = vm_phys_domain(m);
2020 		vmd = VM_DOMAIN(domain);
2021 		goto found;
2022 	}
2023 #endif
2024 	vmd = VM_DOMAIN(domain);
2025 	if (vmd->vmd_pgcache[pool].zone != NULL) {
2026 		m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
2027 		if (m != NULL) {
2028 			flags |= PG_PCPU_CACHE;
2029 			goto found;
2030 		}
2031 	}
2032 	if (vm_domain_allocate(vmd, req, 1)) {
2033 		/*
2034 		 * If not, allocate it from the free page queues.
2035 		 */
2036 		vm_domain_free_lock(vmd);
2037 		m = vm_phys_alloc_pages(domain, pool, 0);
2038 		vm_domain_free_unlock(vmd);
2039 		if (m == NULL) {
2040 			vm_domain_freecnt_inc(vmd, 1);
2041 #if VM_NRESERVLEVEL > 0
2042 			if (vm_reserv_reclaim_inactive(domain))
2043 				goto again;
2044 #endif
2045 		}
2046 	}
2047 	if (m == NULL) {
2048 		/*
2049 		 * Not allocatable, give up.
2050 		 */
2051 		if (vm_domain_alloc_fail(vmd, object, req))
2052 			goto again;
2053 		return (NULL);
2054 	}
2055 
2056 	/*
2057 	 * At this point we had better have found a good page.
2058 	 */
2059 found:
2060 	vm_page_dequeue(m);
2061 	vm_page_alloc_check(m);
2062 
2063 	/*
2064 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2065 	 */
2066 	if ((req & VM_ALLOC_ZERO) != 0)
2067 		flags |= (m->flags & PG_ZERO);
2068 	if ((req & VM_ALLOC_NODUMP) != 0)
2069 		flags |= PG_NODUMP;
2070 	m->flags = flags;
2071 	m->a.flags = 0;
2072 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2073 	    VPO_UNMANAGED : 0;
2074 	m->busy_lock = VPB_UNBUSIED;
2075 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2076 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2077 	if ((req & VM_ALLOC_SBUSY) != 0)
2078 		m->busy_lock = VPB_SHARERS_WORD(1);
2079 	if (req & VM_ALLOC_WIRED) {
2080 		vm_wire_add(1);
2081 		m->ref_count = 1;
2082 	}
2083 	m->a.act_count = 0;
2084 
2085 	if (object != NULL) {
2086 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2087 			if (req & VM_ALLOC_WIRED) {
2088 				vm_wire_sub(1);
2089 				m->ref_count = 0;
2090 			}
2091 			KASSERT(m->object == NULL, ("page %p has object", m));
2092 			m->oflags = VPO_UNMANAGED;
2093 			m->busy_lock = VPB_UNBUSIED;
2094 			/* Don't change PG_ZERO. */
2095 			vm_page_free_toq(m);
2096 			if (req & VM_ALLOC_WAITFAIL) {
2097 				VM_OBJECT_WUNLOCK(object);
2098 				vm_radix_wait();
2099 				VM_OBJECT_WLOCK(object);
2100 			}
2101 			return (NULL);
2102 		}
2103 
2104 		/* Ignore device objects; the pager sets "memattr" for them. */
2105 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2106 		    (object->flags & OBJ_FICTITIOUS) == 0)
2107 			pmap_page_set_memattr(m, object->memattr);
2108 	} else
2109 		m->pindex = pindex;
2110 
2111 	return (m);
2112 }
2113 
2114 /*
2115  *	vm_page_alloc_contig:
2116  *
2117  *	Allocate a contiguous set of physical pages of the given size "npages"
2118  *	from the free lists.  All of the physical pages must be at or above
2119  *	the given physical address "low" and below the given physical address
2120  *	"high".  The given value "alignment" determines the alignment of the
2121  *	first physical page in the set.  If the given value "boundary" is
2122  *	non-zero, then the set of physical pages cannot cross any physical
2123  *	address boundary that is a multiple of that value.  Both "alignment"
2124  *	and "boundary" must be a power of two.
2125  *
2126  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2127  *	then the memory attribute setting for the physical pages is configured
2128  *	to the object's memory attribute setting.  Otherwise, the memory
2129  *	attribute setting for the physical pages is configured to "memattr",
2130  *	overriding the object's memory attribute setting.  However, if the
2131  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2132  *	memory attribute setting for the physical pages cannot be configured
2133  *	to VM_MEMATTR_DEFAULT.
2134  *
2135  *	The specified object may not contain fictitious pages.
2136  *
2137  *	The caller must always specify an allocation class.
2138  *
2139  *	allocation classes:
2140  *	VM_ALLOC_NORMAL		normal process request
2141  *	VM_ALLOC_SYSTEM		system *really* needs a page
2142  *	VM_ALLOC_INTERRUPT	interrupt time request
2143  *
2144  *	optional allocation flags:
2145  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2146  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2147  *	VM_ALLOC_NOOBJ		page is not associated with an object and
2148  *				should not be exclusive busy
2149  *	VM_ALLOC_SBUSY		shared busy the allocated page
2150  *	VM_ALLOC_WIRED		wire the allocated page
2151  *	VM_ALLOC_ZERO		prefer a zeroed page
2152  */
2153 vm_page_t
2154 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2155     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2156     vm_paddr_t boundary, vm_memattr_t memattr)
2157 {
2158 	struct vm_domainset_iter di;
2159 	vm_page_t m;
2160 	int domain;
2161 
2162 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2163 	do {
2164 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2165 		    npages, low, high, alignment, boundary, memattr);
2166 		if (m != NULL)
2167 			break;
2168 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2169 
2170 	return (m);
2171 }
2172 
2173 vm_page_t
2174 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2175     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2176     vm_paddr_t boundary, vm_memattr_t memattr)
2177 {
2178 	struct vm_domain *vmd;
2179 	vm_page_t m, m_ret, mpred;
2180 	u_int busy_lock, flags, oflags;
2181 
2182 	mpred = NULL;	/* XXX: pacify gcc */
2183 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2184 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2185 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2186 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2187 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2188 	    req));
2189 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2190 	    ("Can't sleep and retry object insertion."));
2191 	if (object != NULL) {
2192 		VM_OBJECT_ASSERT_WLOCKED(object);
2193 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2194 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2195 		    object));
2196 	}
2197 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2198 
2199 	if (object != NULL) {
2200 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2201 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2202 		    ("vm_page_alloc_contig: pindex already allocated"));
2203 	}
2204 
2205 	/*
2206 	 * Can we allocate the pages without the number of free pages falling
2207 	 * below the lower bound for the allocation class?
2208 	 */
2209 	m_ret = NULL;
2210 again:
2211 #if VM_NRESERVLEVEL > 0
2212 	/*
2213 	 * Can we allocate the pages from a reservation?
2214 	 */
2215 	if (vm_object_reserv(object) &&
2216 	    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2217 	    mpred, npages, low, high, alignment, boundary)) != NULL) {
2218 		domain = vm_phys_domain(m_ret);
2219 		vmd = VM_DOMAIN(domain);
2220 		goto found;
2221 	}
2222 #endif
2223 	vmd = VM_DOMAIN(domain);
2224 	if (vm_domain_allocate(vmd, req, npages)) {
2225 		/*
2226 		 * allocate them from the free page queues.
2227 		 */
2228 		vm_domain_free_lock(vmd);
2229 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2230 		    alignment, boundary);
2231 		vm_domain_free_unlock(vmd);
2232 		if (m_ret == NULL) {
2233 			vm_domain_freecnt_inc(vmd, npages);
2234 #if VM_NRESERVLEVEL > 0
2235 			if (vm_reserv_reclaim_contig(domain, npages, low,
2236 			    high, alignment, boundary))
2237 				goto again;
2238 #endif
2239 		}
2240 	}
2241 	if (m_ret == NULL) {
2242 		if (vm_domain_alloc_fail(vmd, object, req))
2243 			goto again;
2244 		return (NULL);
2245 	}
2246 #if VM_NRESERVLEVEL > 0
2247 found:
2248 #endif
2249 	for (m = m_ret; m < &m_ret[npages]; m++) {
2250 		vm_page_dequeue(m);
2251 		vm_page_alloc_check(m);
2252 	}
2253 
2254 	/*
2255 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2256 	 */
2257 	flags = 0;
2258 	if ((req & VM_ALLOC_ZERO) != 0)
2259 		flags = PG_ZERO;
2260 	if ((req & VM_ALLOC_NODUMP) != 0)
2261 		flags |= PG_NODUMP;
2262 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2263 	    VPO_UNMANAGED : 0;
2264 	busy_lock = VPB_UNBUSIED;
2265 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2266 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2267 	if ((req & VM_ALLOC_SBUSY) != 0)
2268 		busy_lock = VPB_SHARERS_WORD(1);
2269 	if ((req & VM_ALLOC_WIRED) != 0)
2270 		vm_wire_add(npages);
2271 	if (object != NULL) {
2272 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2273 		    memattr == VM_MEMATTR_DEFAULT)
2274 			memattr = object->memattr;
2275 	}
2276 	for (m = m_ret; m < &m_ret[npages]; m++) {
2277 		m->a.flags = 0;
2278 		m->flags = (m->flags | PG_NODUMP) & flags;
2279 		m->busy_lock = busy_lock;
2280 		if ((req & VM_ALLOC_WIRED) != 0)
2281 			m->ref_count = 1;
2282 		m->a.act_count = 0;
2283 		m->oflags = oflags;
2284 		if (object != NULL) {
2285 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2286 				if ((req & VM_ALLOC_WIRED) != 0)
2287 					vm_wire_sub(npages);
2288 				KASSERT(m->object == NULL,
2289 				    ("page %p has object", m));
2290 				mpred = m;
2291 				for (m = m_ret; m < &m_ret[npages]; m++) {
2292 					if (m <= mpred &&
2293 					    (req & VM_ALLOC_WIRED) != 0)
2294 						m->ref_count = 0;
2295 					m->oflags = VPO_UNMANAGED;
2296 					m->busy_lock = VPB_UNBUSIED;
2297 					/* Don't change PG_ZERO. */
2298 					vm_page_free_toq(m);
2299 				}
2300 				if (req & VM_ALLOC_WAITFAIL) {
2301 					VM_OBJECT_WUNLOCK(object);
2302 					vm_radix_wait();
2303 					VM_OBJECT_WLOCK(object);
2304 				}
2305 				return (NULL);
2306 			}
2307 			mpred = m;
2308 		} else
2309 			m->pindex = pindex;
2310 		if (memattr != VM_MEMATTR_DEFAULT)
2311 			pmap_page_set_memattr(m, memattr);
2312 		pindex++;
2313 	}
2314 	return (m_ret);
2315 }
2316 
2317 /*
2318  * Check a page that has been freshly dequeued from a freelist.
2319  */
2320 static void
2321 vm_page_alloc_check(vm_page_t m)
2322 {
2323 
2324 	KASSERT(m->object == NULL, ("page %p has object", m));
2325 	KASSERT(m->a.queue == PQ_NONE &&
2326 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2327 	    ("page %p has unexpected queue %d, flags %#x",
2328 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2329 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2330 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2331 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2332 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2333 	    ("page %p has unexpected memattr %d",
2334 	    m, pmap_page_get_memattr(m)));
2335 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2336 }
2337 
2338 /*
2339  * 	vm_page_alloc_freelist:
2340  *
2341  *	Allocate a physical page from the specified free page list.
2342  *
2343  *	The caller must always specify an allocation class.
2344  *
2345  *	allocation classes:
2346  *	VM_ALLOC_NORMAL		normal process request
2347  *	VM_ALLOC_SYSTEM		system *really* needs a page
2348  *	VM_ALLOC_INTERRUPT	interrupt time request
2349  *
2350  *	optional allocation flags:
2351  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2352  *				intends to allocate
2353  *	VM_ALLOC_WIRED		wire the allocated page
2354  *	VM_ALLOC_ZERO		prefer a zeroed page
2355  */
2356 vm_page_t
2357 vm_page_alloc_freelist(int freelist, int req)
2358 {
2359 	struct vm_domainset_iter di;
2360 	vm_page_t m;
2361 	int domain;
2362 
2363 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2364 	do {
2365 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2366 		if (m != NULL)
2367 			break;
2368 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2369 
2370 	return (m);
2371 }
2372 
2373 vm_page_t
2374 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2375 {
2376 	struct vm_domain *vmd;
2377 	vm_page_t m;
2378 	u_int flags;
2379 
2380 	m = NULL;
2381 	vmd = VM_DOMAIN(domain);
2382 again:
2383 	if (vm_domain_allocate(vmd, req, 1)) {
2384 		vm_domain_free_lock(vmd);
2385 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2386 		    VM_FREEPOOL_DIRECT, 0);
2387 		vm_domain_free_unlock(vmd);
2388 		if (m == NULL)
2389 			vm_domain_freecnt_inc(vmd, 1);
2390 	}
2391 	if (m == NULL) {
2392 		if (vm_domain_alloc_fail(vmd, NULL, req))
2393 			goto again;
2394 		return (NULL);
2395 	}
2396 	vm_page_dequeue(m);
2397 	vm_page_alloc_check(m);
2398 
2399 	/*
2400 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2401 	 */
2402 	m->a.flags = 0;
2403 	flags = 0;
2404 	if ((req & VM_ALLOC_ZERO) != 0)
2405 		flags = PG_ZERO;
2406 	m->flags &= flags;
2407 	if ((req & VM_ALLOC_WIRED) != 0) {
2408 		vm_wire_add(1);
2409 		m->ref_count = 1;
2410 	}
2411 	/* Unmanaged pages don't use "act_count". */
2412 	m->oflags = VPO_UNMANAGED;
2413 	return (m);
2414 }
2415 
2416 static int
2417 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2418 {
2419 	struct vm_domain *vmd;
2420 	struct vm_pgcache *pgcache;
2421 	int i;
2422 
2423 	pgcache = arg;
2424 	vmd = VM_DOMAIN(pgcache->domain);
2425 
2426 	/*
2427 	 * The page daemon should avoid creating extra memory pressure since its
2428 	 * main purpose is to replenish the store of free pages.
2429 	 */
2430 	if (vmd->vmd_severeset || curproc == pageproc ||
2431 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2432 		return (0);
2433 	domain = vmd->vmd_domain;
2434 	vm_domain_free_lock(vmd);
2435 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2436 	    (vm_page_t *)store);
2437 	vm_domain_free_unlock(vmd);
2438 	if (cnt != i)
2439 		vm_domain_freecnt_inc(vmd, cnt - i);
2440 
2441 	return (i);
2442 }
2443 
2444 static void
2445 vm_page_zone_release(void *arg, void **store, int cnt)
2446 {
2447 	struct vm_domain *vmd;
2448 	struct vm_pgcache *pgcache;
2449 	vm_page_t m;
2450 	int i;
2451 
2452 	pgcache = arg;
2453 	vmd = VM_DOMAIN(pgcache->domain);
2454 	vm_domain_free_lock(vmd);
2455 	for (i = 0; i < cnt; i++) {
2456 		m = (vm_page_t)store[i];
2457 		vm_phys_free_pages(m, 0);
2458 	}
2459 	vm_domain_free_unlock(vmd);
2460 	vm_domain_freecnt_inc(vmd, cnt);
2461 }
2462 
2463 #define	VPSC_ANY	0	/* No restrictions. */
2464 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2465 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2466 
2467 /*
2468  *	vm_page_scan_contig:
2469  *
2470  *	Scan vm_page_array[] between the specified entries "m_start" and
2471  *	"m_end" for a run of contiguous physical pages that satisfy the
2472  *	specified conditions, and return the lowest page in the run.  The
2473  *	specified "alignment" determines the alignment of the lowest physical
2474  *	page in the run.  If the specified "boundary" is non-zero, then the
2475  *	run of physical pages cannot span a physical address that is a
2476  *	multiple of "boundary".
2477  *
2478  *	"m_end" is never dereferenced, so it need not point to a vm_page
2479  *	structure within vm_page_array[].
2480  *
2481  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2482  *	span a hole (or discontiguity) in the physical address space.  Both
2483  *	"alignment" and "boundary" must be a power of two.
2484  */
2485 vm_page_t
2486 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2487     u_long alignment, vm_paddr_t boundary, int options)
2488 {
2489 	struct mtx *m_mtx;
2490 	vm_object_t object;
2491 	vm_paddr_t pa;
2492 	vm_page_t m, m_run;
2493 #if VM_NRESERVLEVEL > 0
2494 	int level;
2495 #endif
2496 	int m_inc, order, run_ext, run_len;
2497 
2498 	KASSERT(npages > 0, ("npages is 0"));
2499 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2500 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2501 	m_run = NULL;
2502 	run_len = 0;
2503 	m_mtx = NULL;
2504 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2505 		KASSERT((m->flags & PG_MARKER) == 0,
2506 		    ("page %p is PG_MARKER", m));
2507 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2508 		    ("fictitious page %p has invalid ref count", m));
2509 
2510 		/*
2511 		 * If the current page would be the start of a run, check its
2512 		 * physical address against the end, alignment, and boundary
2513 		 * conditions.  If it doesn't satisfy these conditions, either
2514 		 * terminate the scan or advance to the next page that
2515 		 * satisfies the failed condition.
2516 		 */
2517 		if (run_len == 0) {
2518 			KASSERT(m_run == NULL, ("m_run != NULL"));
2519 			if (m + npages > m_end)
2520 				break;
2521 			pa = VM_PAGE_TO_PHYS(m);
2522 			if ((pa & (alignment - 1)) != 0) {
2523 				m_inc = atop(roundup2(pa, alignment) - pa);
2524 				continue;
2525 			}
2526 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2527 			    boundary) != 0) {
2528 				m_inc = atop(roundup2(pa, boundary) - pa);
2529 				continue;
2530 			}
2531 		} else
2532 			KASSERT(m_run != NULL, ("m_run == NULL"));
2533 
2534 		vm_page_change_lock(m, &m_mtx);
2535 		m_inc = 1;
2536 retry:
2537 		if (vm_page_wired(m))
2538 			run_ext = 0;
2539 #if VM_NRESERVLEVEL > 0
2540 		else if ((level = vm_reserv_level(m)) >= 0 &&
2541 		    (options & VPSC_NORESERV) != 0) {
2542 			run_ext = 0;
2543 			/* Advance to the end of the reservation. */
2544 			pa = VM_PAGE_TO_PHYS(m);
2545 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2546 			    pa);
2547 		}
2548 #endif
2549 		else if ((object = m->object) != NULL) {
2550 			/*
2551 			 * The page is considered eligible for relocation if
2552 			 * and only if it could be laundered or reclaimed by
2553 			 * the page daemon.
2554 			 */
2555 			if (!VM_OBJECT_TRYRLOCK(object)) {
2556 				mtx_unlock(m_mtx);
2557 				VM_OBJECT_RLOCK(object);
2558 				mtx_lock(m_mtx);
2559 				if (m->object != object) {
2560 					/*
2561 					 * The page may have been freed.
2562 					 */
2563 					VM_OBJECT_RUNLOCK(object);
2564 					goto retry;
2565 				}
2566 			}
2567 			/* Don't care: PG_NODUMP, PG_ZERO. */
2568 			if (object->type != OBJT_DEFAULT &&
2569 			    object->type != OBJT_SWAP &&
2570 			    object->type != OBJT_VNODE) {
2571 				run_ext = 0;
2572 #if VM_NRESERVLEVEL > 0
2573 			} else if ((options & VPSC_NOSUPER) != 0 &&
2574 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2575 				run_ext = 0;
2576 				/* Advance to the end of the superpage. */
2577 				pa = VM_PAGE_TO_PHYS(m);
2578 				m_inc = atop(roundup2(pa + 1,
2579 				    vm_reserv_size(level)) - pa);
2580 #endif
2581 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2582 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) &&
2583 			    !vm_page_wired(m)) {
2584 				/*
2585 				 * The page is allocated but eligible for
2586 				 * relocation.  Extend the current run by one
2587 				 * page.
2588 				 */
2589 				KASSERT(pmap_page_get_memattr(m) ==
2590 				    VM_MEMATTR_DEFAULT,
2591 				    ("page %p has an unexpected memattr", m));
2592 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2593 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2594 				    ("page %p has unexpected oflags", m));
2595 				/* Don't care: PGA_NOSYNC. */
2596 				run_ext = 1;
2597 			} else
2598 				run_ext = 0;
2599 			VM_OBJECT_RUNLOCK(object);
2600 #if VM_NRESERVLEVEL > 0
2601 		} else if (level >= 0) {
2602 			/*
2603 			 * The page is reserved but not yet allocated.  In
2604 			 * other words, it is still free.  Extend the current
2605 			 * run by one page.
2606 			 */
2607 			run_ext = 1;
2608 #endif
2609 		} else if ((order = m->order) < VM_NFREEORDER) {
2610 			/*
2611 			 * The page is enqueued in the physical memory
2612 			 * allocator's free page queues.  Moreover, it is the
2613 			 * first page in a power-of-two-sized run of
2614 			 * contiguous free pages.  Add these pages to the end
2615 			 * of the current run, and jump ahead.
2616 			 */
2617 			run_ext = 1 << order;
2618 			m_inc = 1 << order;
2619 		} else {
2620 			/*
2621 			 * Skip the page for one of the following reasons: (1)
2622 			 * It is enqueued in the physical memory allocator's
2623 			 * free page queues.  However, it is not the first
2624 			 * page in a run of contiguous free pages.  (This case
2625 			 * rarely occurs because the scan is performed in
2626 			 * ascending order.) (2) It is not reserved, and it is
2627 			 * transitioning from free to allocated.  (Conversely,
2628 			 * the transition from allocated to free for managed
2629 			 * pages is blocked by the page lock.) (3) It is
2630 			 * allocated but not contained by an object and not
2631 			 * wired, e.g., allocated by Xen's balloon driver.
2632 			 */
2633 			run_ext = 0;
2634 		}
2635 
2636 		/*
2637 		 * Extend or reset the current run of pages.
2638 		 */
2639 		if (run_ext > 0) {
2640 			if (run_len == 0)
2641 				m_run = m;
2642 			run_len += run_ext;
2643 		} else {
2644 			if (run_len > 0) {
2645 				m_run = NULL;
2646 				run_len = 0;
2647 			}
2648 		}
2649 	}
2650 	if (m_mtx != NULL)
2651 		mtx_unlock(m_mtx);
2652 	if (run_len >= npages)
2653 		return (m_run);
2654 	return (NULL);
2655 }
2656 
2657 /*
2658  *	vm_page_reclaim_run:
2659  *
2660  *	Try to relocate each of the allocated virtual pages within the
2661  *	specified run of physical pages to a new physical address.  Free the
2662  *	physical pages underlying the relocated virtual pages.  A virtual page
2663  *	is relocatable if and only if it could be laundered or reclaimed by
2664  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2665  *	physical address above "high".
2666  *
2667  *	Returns 0 if every physical page within the run was already free or
2668  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2669  *	value indicating why the last attempt to relocate a virtual page was
2670  *	unsuccessful.
2671  *
2672  *	"req_class" must be an allocation class.
2673  */
2674 static int
2675 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2676     vm_paddr_t high)
2677 {
2678 	struct vm_domain *vmd;
2679 	struct mtx *m_mtx;
2680 	struct spglist free;
2681 	vm_object_t object;
2682 	vm_paddr_t pa;
2683 	vm_page_t m, m_end, m_new;
2684 	int error, order, req;
2685 
2686 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2687 	    ("req_class is not an allocation class"));
2688 	SLIST_INIT(&free);
2689 	error = 0;
2690 	m = m_run;
2691 	m_end = m_run + npages;
2692 	m_mtx = NULL;
2693 	for (; error == 0 && m < m_end; m++) {
2694 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2695 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2696 
2697 		/*
2698 		 * Avoid releasing and reacquiring the same page lock.
2699 		 */
2700 		vm_page_change_lock(m, &m_mtx);
2701 retry:
2702 		/*
2703 		 * Racily check for wirings.  Races are handled below.
2704 		 */
2705 		if (vm_page_wired(m))
2706 			error = EBUSY;
2707 		else if ((object = m->object) != NULL) {
2708 			/*
2709 			 * The page is relocated if and only if it could be
2710 			 * laundered or reclaimed by the page daemon.
2711 			 */
2712 			if (!VM_OBJECT_TRYWLOCK(object)) {
2713 				mtx_unlock(m_mtx);
2714 				VM_OBJECT_WLOCK(object);
2715 				mtx_lock(m_mtx);
2716 				if (m->object != object) {
2717 					/*
2718 					 * The page may have been freed.
2719 					 */
2720 					VM_OBJECT_WUNLOCK(object);
2721 					goto retry;
2722 				}
2723 			}
2724 			/* Don't care: PG_NODUMP, PG_ZERO. */
2725 			if (object->type != OBJT_DEFAULT &&
2726 			    object->type != OBJT_SWAP &&
2727 			    object->type != OBJT_VNODE)
2728 				error = EINVAL;
2729 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2730 				error = EINVAL;
2731 			else if (vm_page_queue(m) != PQ_NONE &&
2732 			    vm_page_tryxbusy(m) != 0) {
2733 				if (vm_page_wired(m)) {
2734 					vm_page_xunbusy(m);
2735 					error = EBUSY;
2736 					goto unlock;
2737 				}
2738 				KASSERT(pmap_page_get_memattr(m) ==
2739 				    VM_MEMATTR_DEFAULT,
2740 				    ("page %p has an unexpected memattr", m));
2741 				KASSERT(m->oflags == 0,
2742 				    ("page %p has unexpected oflags", m));
2743 				/* Don't care: PGA_NOSYNC. */
2744 				if (!vm_page_none_valid(m)) {
2745 					/*
2746 					 * First, try to allocate a new page
2747 					 * that is above "high".  Failing
2748 					 * that, try to allocate a new page
2749 					 * that is below "m_run".  Allocate
2750 					 * the new page between the end of
2751 					 * "m_run" and "high" only as a last
2752 					 * resort.
2753 					 */
2754 					req = req_class | VM_ALLOC_NOOBJ;
2755 					if ((m->flags & PG_NODUMP) != 0)
2756 						req |= VM_ALLOC_NODUMP;
2757 					if (trunc_page(high) !=
2758 					    ~(vm_paddr_t)PAGE_MASK) {
2759 						m_new = vm_page_alloc_contig(
2760 						    NULL, 0, req, 1,
2761 						    round_page(high),
2762 						    ~(vm_paddr_t)0,
2763 						    PAGE_SIZE, 0,
2764 						    VM_MEMATTR_DEFAULT);
2765 					} else
2766 						m_new = NULL;
2767 					if (m_new == NULL) {
2768 						pa = VM_PAGE_TO_PHYS(m_run);
2769 						m_new = vm_page_alloc_contig(
2770 						    NULL, 0, req, 1,
2771 						    0, pa - 1, PAGE_SIZE, 0,
2772 						    VM_MEMATTR_DEFAULT);
2773 					}
2774 					if (m_new == NULL) {
2775 						pa += ptoa(npages);
2776 						m_new = vm_page_alloc_contig(
2777 						    NULL, 0, req, 1,
2778 						    pa, high, PAGE_SIZE, 0,
2779 						    VM_MEMATTR_DEFAULT);
2780 					}
2781 					if (m_new == NULL) {
2782 						vm_page_xunbusy(m);
2783 						error = ENOMEM;
2784 						goto unlock;
2785 					}
2786 
2787 					/*
2788 					 * Unmap the page and check for new
2789 					 * wirings that may have been acquired
2790 					 * through a pmap lookup.
2791 					 */
2792 					if (object->ref_count != 0 &&
2793 					    !vm_page_try_remove_all(m)) {
2794 						vm_page_xunbusy(m);
2795 						vm_page_free(m_new);
2796 						error = EBUSY;
2797 						goto unlock;
2798 					}
2799 
2800 					/*
2801 					 * Replace "m" with the new page.  For
2802 					 * vm_page_replace(), "m" must be busy
2803 					 * and dequeued.  Finally, change "m"
2804 					 * as if vm_page_free() was called.
2805 					 */
2806 					m_new->a.flags = m->a.flags &
2807 					    ~PGA_QUEUE_STATE_MASK;
2808 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2809 					    ("page %p is managed", m_new));
2810 					m_new->oflags = 0;
2811 					pmap_copy_page(m, m_new);
2812 					m_new->valid = m->valid;
2813 					m_new->dirty = m->dirty;
2814 					m->flags &= ~PG_ZERO;
2815 					vm_page_dequeue(m);
2816 					if (vm_page_replace_hold(m_new, object,
2817 					    m->pindex, m) &&
2818 					    vm_page_free_prep(m))
2819 						SLIST_INSERT_HEAD(&free, m,
2820 						    plinks.s.ss);
2821 
2822 					/*
2823 					 * The new page must be deactivated
2824 					 * before the object is unlocked.
2825 					 */
2826 					vm_page_change_lock(m_new, &m_mtx);
2827 					vm_page_deactivate(m_new);
2828 				} else {
2829 					m->flags &= ~PG_ZERO;
2830 					vm_page_dequeue(m);
2831 					if (vm_page_free_prep(m))
2832 						SLIST_INSERT_HEAD(&free, m,
2833 						    plinks.s.ss);
2834 					KASSERT(m->dirty == 0,
2835 					    ("page %p is dirty", m));
2836 				}
2837 			} else
2838 				error = EBUSY;
2839 unlock:
2840 			VM_OBJECT_WUNLOCK(object);
2841 		} else {
2842 			MPASS(vm_phys_domain(m) == domain);
2843 			vmd = VM_DOMAIN(domain);
2844 			vm_domain_free_lock(vmd);
2845 			order = m->order;
2846 			if (order < VM_NFREEORDER) {
2847 				/*
2848 				 * The page is enqueued in the physical memory
2849 				 * allocator's free page queues.  Moreover, it
2850 				 * is the first page in a power-of-two-sized
2851 				 * run of contiguous free pages.  Jump ahead
2852 				 * to the last page within that run, and
2853 				 * continue from there.
2854 				 */
2855 				m += (1 << order) - 1;
2856 			}
2857 #if VM_NRESERVLEVEL > 0
2858 			else if (vm_reserv_is_page_free(m))
2859 				order = 0;
2860 #endif
2861 			vm_domain_free_unlock(vmd);
2862 			if (order == VM_NFREEORDER)
2863 				error = EINVAL;
2864 		}
2865 	}
2866 	if (m_mtx != NULL)
2867 		mtx_unlock(m_mtx);
2868 	if ((m = SLIST_FIRST(&free)) != NULL) {
2869 		int cnt;
2870 
2871 		vmd = VM_DOMAIN(domain);
2872 		cnt = 0;
2873 		vm_domain_free_lock(vmd);
2874 		do {
2875 			MPASS(vm_phys_domain(m) == domain);
2876 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2877 			vm_phys_free_pages(m, 0);
2878 			cnt++;
2879 		} while ((m = SLIST_FIRST(&free)) != NULL);
2880 		vm_domain_free_unlock(vmd);
2881 		vm_domain_freecnt_inc(vmd, cnt);
2882 	}
2883 	return (error);
2884 }
2885 
2886 #define	NRUNS	16
2887 
2888 CTASSERT(powerof2(NRUNS));
2889 
2890 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2891 
2892 #define	MIN_RECLAIM	8
2893 
2894 /*
2895  *	vm_page_reclaim_contig:
2896  *
2897  *	Reclaim allocated, contiguous physical memory satisfying the specified
2898  *	conditions by relocating the virtual pages using that physical memory.
2899  *	Returns true if reclamation is successful and false otherwise.  Since
2900  *	relocation requires the allocation of physical pages, reclamation may
2901  *	fail due to a shortage of free pages.  When reclamation fails, callers
2902  *	are expected to perform vm_wait() before retrying a failed allocation
2903  *	operation, e.g., vm_page_alloc_contig().
2904  *
2905  *	The caller must always specify an allocation class through "req".
2906  *
2907  *	allocation classes:
2908  *	VM_ALLOC_NORMAL		normal process request
2909  *	VM_ALLOC_SYSTEM		system *really* needs a page
2910  *	VM_ALLOC_INTERRUPT	interrupt time request
2911  *
2912  *	The optional allocation flags are ignored.
2913  *
2914  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2915  *	must be a power of two.
2916  */
2917 bool
2918 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2919     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2920 {
2921 	struct vm_domain *vmd;
2922 	vm_paddr_t curr_low;
2923 	vm_page_t m_run, m_runs[NRUNS];
2924 	u_long count, reclaimed;
2925 	int error, i, options, req_class;
2926 
2927 	KASSERT(npages > 0, ("npages is 0"));
2928 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2929 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2930 	req_class = req & VM_ALLOC_CLASS_MASK;
2931 
2932 	/*
2933 	 * The page daemon is allowed to dig deeper into the free page list.
2934 	 */
2935 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2936 		req_class = VM_ALLOC_SYSTEM;
2937 
2938 	/*
2939 	 * Return if the number of free pages cannot satisfy the requested
2940 	 * allocation.
2941 	 */
2942 	vmd = VM_DOMAIN(domain);
2943 	count = vmd->vmd_free_count;
2944 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2945 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2946 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2947 		return (false);
2948 
2949 	/*
2950 	 * Scan up to three times, relaxing the restrictions ("options") on
2951 	 * the reclamation of reservations and superpages each time.
2952 	 */
2953 	for (options = VPSC_NORESERV;;) {
2954 		/*
2955 		 * Find the highest runs that satisfy the given constraints
2956 		 * and restrictions, and record them in "m_runs".
2957 		 */
2958 		curr_low = low;
2959 		count = 0;
2960 		for (;;) {
2961 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2962 			    high, alignment, boundary, options);
2963 			if (m_run == NULL)
2964 				break;
2965 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2966 			m_runs[RUN_INDEX(count)] = m_run;
2967 			count++;
2968 		}
2969 
2970 		/*
2971 		 * Reclaim the highest runs in LIFO (descending) order until
2972 		 * the number of reclaimed pages, "reclaimed", is at least
2973 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2974 		 * reclamation is idempotent, and runs will (likely) recur
2975 		 * from one scan to the next as restrictions are relaxed.
2976 		 */
2977 		reclaimed = 0;
2978 		for (i = 0; count > 0 && i < NRUNS; i++) {
2979 			count--;
2980 			m_run = m_runs[RUN_INDEX(count)];
2981 			error = vm_page_reclaim_run(req_class, domain, npages,
2982 			    m_run, high);
2983 			if (error == 0) {
2984 				reclaimed += npages;
2985 				if (reclaimed >= MIN_RECLAIM)
2986 					return (true);
2987 			}
2988 		}
2989 
2990 		/*
2991 		 * Either relax the restrictions on the next scan or return if
2992 		 * the last scan had no restrictions.
2993 		 */
2994 		if (options == VPSC_NORESERV)
2995 			options = VPSC_NOSUPER;
2996 		else if (options == VPSC_NOSUPER)
2997 			options = VPSC_ANY;
2998 		else if (options == VPSC_ANY)
2999 			return (reclaimed != 0);
3000 	}
3001 }
3002 
3003 bool
3004 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3005     u_long alignment, vm_paddr_t boundary)
3006 {
3007 	struct vm_domainset_iter di;
3008 	int domain;
3009 	bool ret;
3010 
3011 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3012 	do {
3013 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3014 		    high, alignment, boundary);
3015 		if (ret)
3016 			break;
3017 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3018 
3019 	return (ret);
3020 }
3021 
3022 /*
3023  * Set the domain in the appropriate page level domainset.
3024  */
3025 void
3026 vm_domain_set(struct vm_domain *vmd)
3027 {
3028 
3029 	mtx_lock(&vm_domainset_lock);
3030 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3031 		vmd->vmd_minset = 1;
3032 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3033 	}
3034 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3035 		vmd->vmd_severeset = 1;
3036 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3037 	}
3038 	mtx_unlock(&vm_domainset_lock);
3039 }
3040 
3041 /*
3042  * Clear the domain from the appropriate page level domainset.
3043  */
3044 void
3045 vm_domain_clear(struct vm_domain *vmd)
3046 {
3047 
3048 	mtx_lock(&vm_domainset_lock);
3049 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3050 		vmd->vmd_minset = 0;
3051 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3052 		if (vm_min_waiters != 0) {
3053 			vm_min_waiters = 0;
3054 			wakeup(&vm_min_domains);
3055 		}
3056 	}
3057 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3058 		vmd->vmd_severeset = 0;
3059 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3060 		if (vm_severe_waiters != 0) {
3061 			vm_severe_waiters = 0;
3062 			wakeup(&vm_severe_domains);
3063 		}
3064 	}
3065 
3066 	/*
3067 	 * If pageout daemon needs pages, then tell it that there are
3068 	 * some free.
3069 	 */
3070 	if (vmd->vmd_pageout_pages_needed &&
3071 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3072 		wakeup(&vmd->vmd_pageout_pages_needed);
3073 		vmd->vmd_pageout_pages_needed = 0;
3074 	}
3075 
3076 	/* See comments in vm_wait_doms(). */
3077 	if (vm_pageproc_waiters) {
3078 		vm_pageproc_waiters = 0;
3079 		wakeup(&vm_pageproc_waiters);
3080 	}
3081 	mtx_unlock(&vm_domainset_lock);
3082 }
3083 
3084 /*
3085  * Wait for free pages to exceed the min threshold globally.
3086  */
3087 void
3088 vm_wait_min(void)
3089 {
3090 
3091 	mtx_lock(&vm_domainset_lock);
3092 	while (vm_page_count_min()) {
3093 		vm_min_waiters++;
3094 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3095 	}
3096 	mtx_unlock(&vm_domainset_lock);
3097 }
3098 
3099 /*
3100  * Wait for free pages to exceed the severe threshold globally.
3101  */
3102 void
3103 vm_wait_severe(void)
3104 {
3105 
3106 	mtx_lock(&vm_domainset_lock);
3107 	while (vm_page_count_severe()) {
3108 		vm_severe_waiters++;
3109 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3110 		    "vmwait", 0);
3111 	}
3112 	mtx_unlock(&vm_domainset_lock);
3113 }
3114 
3115 u_int
3116 vm_wait_count(void)
3117 {
3118 
3119 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3120 }
3121 
3122 void
3123 vm_wait_doms(const domainset_t *wdoms)
3124 {
3125 
3126 	/*
3127 	 * We use racey wakeup synchronization to avoid expensive global
3128 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3129 	 * To handle this, we only sleep for one tick in this instance.  It
3130 	 * is expected that most allocations for the pageproc will come from
3131 	 * kmem or vm_page_grab* which will use the more specific and
3132 	 * race-free vm_wait_domain().
3133 	 */
3134 	if (curproc == pageproc) {
3135 		mtx_lock(&vm_domainset_lock);
3136 		vm_pageproc_waiters++;
3137 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3138 		    "pageprocwait", 1);
3139 	} else {
3140 		/*
3141 		 * XXX Ideally we would wait only until the allocation could
3142 		 * be satisfied.  This condition can cause new allocators to
3143 		 * consume all freed pages while old allocators wait.
3144 		 */
3145 		mtx_lock(&vm_domainset_lock);
3146 		if (vm_page_count_min_set(wdoms)) {
3147 			vm_min_waiters++;
3148 			msleep(&vm_min_domains, &vm_domainset_lock,
3149 			    PVM | PDROP, "vmwait", 0);
3150 		} else
3151 			mtx_unlock(&vm_domainset_lock);
3152 	}
3153 }
3154 
3155 /*
3156  *	vm_wait_domain:
3157  *
3158  *	Sleep until free pages are available for allocation.
3159  *	- Called in various places after failed memory allocations.
3160  */
3161 void
3162 vm_wait_domain(int domain)
3163 {
3164 	struct vm_domain *vmd;
3165 	domainset_t wdom;
3166 
3167 	vmd = VM_DOMAIN(domain);
3168 	vm_domain_free_assert_unlocked(vmd);
3169 
3170 	if (curproc == pageproc) {
3171 		mtx_lock(&vm_domainset_lock);
3172 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3173 			vmd->vmd_pageout_pages_needed = 1;
3174 			msleep(&vmd->vmd_pageout_pages_needed,
3175 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3176 		} else
3177 			mtx_unlock(&vm_domainset_lock);
3178 	} else {
3179 		if (pageproc == NULL)
3180 			panic("vm_wait in early boot");
3181 		DOMAINSET_ZERO(&wdom);
3182 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3183 		vm_wait_doms(&wdom);
3184 	}
3185 }
3186 
3187 /*
3188  *	vm_wait:
3189  *
3190  *	Sleep until free pages are available for allocation in the
3191  *	affinity domains of the obj.  If obj is NULL, the domain set
3192  *	for the calling thread is used.
3193  *	Called in various places after failed memory allocations.
3194  */
3195 void
3196 vm_wait(vm_object_t obj)
3197 {
3198 	struct domainset *d;
3199 
3200 	d = NULL;
3201 
3202 	/*
3203 	 * Carefully fetch pointers only once: the struct domainset
3204 	 * itself is ummutable but the pointer might change.
3205 	 */
3206 	if (obj != NULL)
3207 		d = obj->domain.dr_policy;
3208 	if (d == NULL)
3209 		d = curthread->td_domain.dr_policy;
3210 
3211 	vm_wait_doms(&d->ds_mask);
3212 }
3213 
3214 /*
3215  *	vm_domain_alloc_fail:
3216  *
3217  *	Called when a page allocation function fails.  Informs the
3218  *	pagedaemon and performs the requested wait.  Requires the
3219  *	domain_free and object lock on entry.  Returns with the
3220  *	object lock held and free lock released.  Returns an error when
3221  *	retry is necessary.
3222  *
3223  */
3224 static int
3225 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3226 {
3227 
3228 	vm_domain_free_assert_unlocked(vmd);
3229 
3230 	atomic_add_int(&vmd->vmd_pageout_deficit,
3231 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3232 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3233 		if (object != NULL)
3234 			VM_OBJECT_WUNLOCK(object);
3235 		vm_wait_domain(vmd->vmd_domain);
3236 		if (object != NULL)
3237 			VM_OBJECT_WLOCK(object);
3238 		if (req & VM_ALLOC_WAITOK)
3239 			return (EAGAIN);
3240 	}
3241 
3242 	return (0);
3243 }
3244 
3245 /*
3246  *	vm_waitpfault:
3247  *
3248  *	Sleep until free pages are available for allocation.
3249  *	- Called only in vm_fault so that processes page faulting
3250  *	  can be easily tracked.
3251  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3252  *	  processes will be able to grab memory first.  Do not change
3253  *	  this balance without careful testing first.
3254  */
3255 void
3256 vm_waitpfault(struct domainset *dset, int timo)
3257 {
3258 
3259 	/*
3260 	 * XXX Ideally we would wait only until the allocation could
3261 	 * be satisfied.  This condition can cause new allocators to
3262 	 * consume all freed pages while old allocators wait.
3263 	 */
3264 	mtx_lock(&vm_domainset_lock);
3265 	if (vm_page_count_min_set(&dset->ds_mask)) {
3266 		vm_min_waiters++;
3267 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3268 		    "pfault", timo);
3269 	} else
3270 		mtx_unlock(&vm_domainset_lock);
3271 }
3272 
3273 static struct vm_pagequeue *
3274 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3275 {
3276 
3277 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3278 }
3279 
3280 #ifdef INVARIANTS
3281 static struct vm_pagequeue *
3282 vm_page_pagequeue(vm_page_t m)
3283 {
3284 
3285 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3286 }
3287 #endif
3288 
3289 static __always_inline bool
3290 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3291 {
3292 	vm_page_astate_t tmp;
3293 
3294 	tmp = *old;
3295 	do {
3296 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3297 			return (true);
3298 		counter_u64_add(pqstate_commit_retries, 1);
3299 	} while (old->_bits == tmp._bits);
3300 
3301 	return (false);
3302 }
3303 
3304 /*
3305  * Do the work of committing a queue state update that moves the page out of
3306  * its current queue.
3307  */
3308 static bool
3309 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3310     vm_page_astate_t *old, vm_page_astate_t new)
3311 {
3312 	vm_page_t next;
3313 
3314 	vm_pagequeue_assert_locked(pq);
3315 	KASSERT(vm_page_pagequeue(m) == pq,
3316 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3317 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3318 	    ("%s: invalid queue indices %d %d",
3319 	    __func__, old->queue, new.queue));
3320 
3321 	/*
3322 	 * Once the queue index of the page changes there is nothing
3323 	 * synchronizing with further updates to the page's physical
3324 	 * queue state.  Therefore we must speculatively remove the page
3325 	 * from the queue now and be prepared to roll back if the queue
3326 	 * state update fails.  If the page is not physically enqueued then
3327 	 * we just update its queue index.
3328 	 */
3329 	if ((old->flags & PGA_ENQUEUED) != 0) {
3330 		new.flags &= ~PGA_ENQUEUED;
3331 		next = TAILQ_NEXT(m, plinks.q);
3332 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3333 		vm_pagequeue_cnt_dec(pq);
3334 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3335 			if (next == NULL)
3336 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3337 			else
3338 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3339 			vm_pagequeue_cnt_inc(pq);
3340 			return (false);
3341 		} else {
3342 			return (true);
3343 		}
3344 	} else {
3345 		return (vm_page_pqstate_fcmpset(m, old, new));
3346 	}
3347 }
3348 
3349 static bool
3350 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3351     vm_page_astate_t new)
3352 {
3353 	struct vm_pagequeue *pq;
3354 	vm_page_astate_t as;
3355 	bool ret;
3356 
3357 	pq = _vm_page_pagequeue(m, old->queue);
3358 
3359 	/*
3360 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3361 	 * corresponding page queue lock is held.
3362 	 */
3363 	vm_pagequeue_lock(pq);
3364 	as = vm_page_astate_load(m);
3365 	if (__predict_false(as._bits != old->_bits)) {
3366 		*old = as;
3367 		ret = false;
3368 	} else {
3369 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3370 	}
3371 	vm_pagequeue_unlock(pq);
3372 	return (ret);
3373 }
3374 
3375 /*
3376  * Commit a queue state update that enqueues or requeues a page.
3377  */
3378 static bool
3379 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3380     vm_page_astate_t *old, vm_page_astate_t new)
3381 {
3382 	struct vm_domain *vmd;
3383 
3384 	vm_pagequeue_assert_locked(pq);
3385 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3386 	    ("%s: invalid queue indices %d %d",
3387 	    __func__, old->queue, new.queue));
3388 
3389 	new.flags |= PGA_ENQUEUED;
3390 	if (!vm_page_pqstate_fcmpset(m, old, new))
3391 		return (false);
3392 
3393 	if ((old->flags & PGA_ENQUEUED) != 0)
3394 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3395 	else
3396 		vm_pagequeue_cnt_inc(pq);
3397 
3398 	/*
3399 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3400 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3401 	 * applied, even if it was set first.
3402 	 */
3403 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3404 		vmd = vm_pagequeue_domain(m);
3405 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3406 		    ("%s: invalid page queue for page %p", __func__, m));
3407 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3408 	} else {
3409 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3410 	}
3411 	return (true);
3412 }
3413 
3414 /*
3415  * Commit a queue state update that encodes a request for a deferred queue
3416  * operation.
3417  */
3418 static bool
3419 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3420     vm_page_astate_t new)
3421 {
3422 
3423 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3424 	    ("%s: invalid state, queue %d flags %x",
3425 	    __func__, new.queue, new.flags));
3426 
3427 	if (old->_bits != new._bits &&
3428 	    !vm_page_pqstate_fcmpset(m, old, new))
3429 		return (false);
3430 	vm_page_pqbatch_submit(m, new.queue);
3431 	return (true);
3432 }
3433 
3434 /*
3435  * A generic queue state update function.  This handles more cases than the
3436  * specialized functions above.
3437  */
3438 bool
3439 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3440 {
3441 
3442 	if (old->_bits == new._bits)
3443 		return (true);
3444 
3445 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3446 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3447 			return (false);
3448 		if (new.queue != PQ_NONE)
3449 			vm_page_pqbatch_submit(m, new.queue);
3450 	} else {
3451 		if (!vm_page_pqstate_fcmpset(m, old, new))
3452 			return (false);
3453 		if (new.queue != PQ_NONE &&
3454 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3455 			vm_page_pqbatch_submit(m, new.queue);
3456 	}
3457 	return (true);
3458 }
3459 
3460 /*
3461  * Apply deferred queue state updates to a page.
3462  */
3463 static inline void
3464 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3465 {
3466 	vm_page_astate_t new, old;
3467 
3468 	CRITICAL_ASSERT(curthread);
3469 	vm_pagequeue_assert_locked(pq);
3470 	KASSERT(queue < PQ_COUNT,
3471 	    ("%s: invalid queue index %d", __func__, queue));
3472 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3473 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3474 
3475 	for (old = vm_page_astate_load(m);;) {
3476 		if (__predict_false(old.queue != queue ||
3477 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3478 			counter_u64_add(queue_nops, 1);
3479 			break;
3480 		}
3481 		KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0,
3482 		    ("%s: page %p has unexpected queue state", __func__, m));
3483 
3484 		new = old;
3485 		if ((old.flags & PGA_DEQUEUE) != 0) {
3486 			new.flags &= ~PGA_QUEUE_OP_MASK;
3487 			new.queue = PQ_NONE;
3488 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3489 			    m, &old, new))) {
3490 				counter_u64_add(queue_ops, 1);
3491 				break;
3492 			}
3493 		} else {
3494 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3495 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3496 			    m, &old, new))) {
3497 				counter_u64_add(queue_ops, 1);
3498 				break;
3499 			}
3500 		}
3501 	}
3502 }
3503 
3504 static void
3505 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3506     uint8_t queue)
3507 {
3508 	int i;
3509 
3510 	for (i = 0; i < bq->bq_cnt; i++)
3511 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3512 	vm_batchqueue_init(bq);
3513 }
3514 
3515 /*
3516  *	vm_page_pqbatch_submit:		[ internal use only ]
3517  *
3518  *	Enqueue a page in the specified page queue's batched work queue.
3519  *	The caller must have encoded the requested operation in the page
3520  *	structure's a.flags field.
3521  */
3522 void
3523 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3524 {
3525 	struct vm_batchqueue *bq;
3526 	struct vm_pagequeue *pq;
3527 	int domain;
3528 
3529 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3530 	    ("page %p is unmanaged", m));
3531 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3532 
3533 	domain = vm_phys_domain(m);
3534 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3535 
3536 	critical_enter();
3537 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3538 	if (vm_batchqueue_insert(bq, m)) {
3539 		critical_exit();
3540 		return;
3541 	}
3542 	critical_exit();
3543 	vm_pagequeue_lock(pq);
3544 	critical_enter();
3545 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3546 	vm_pqbatch_process(pq, bq, queue);
3547 	vm_pqbatch_process_page(pq, m, queue);
3548 	vm_pagequeue_unlock(pq);
3549 	critical_exit();
3550 }
3551 
3552 /*
3553  *	vm_page_pqbatch_drain:		[ internal use only ]
3554  *
3555  *	Force all per-CPU page queue batch queues to be drained.  This is
3556  *	intended for use in severe memory shortages, to ensure that pages
3557  *	do not remain stuck in the batch queues.
3558  */
3559 void
3560 vm_page_pqbatch_drain(void)
3561 {
3562 	struct thread *td;
3563 	struct vm_domain *vmd;
3564 	struct vm_pagequeue *pq;
3565 	int cpu, domain, queue;
3566 
3567 	td = curthread;
3568 	CPU_FOREACH(cpu) {
3569 		thread_lock(td);
3570 		sched_bind(td, cpu);
3571 		thread_unlock(td);
3572 
3573 		for (domain = 0; domain < vm_ndomains; domain++) {
3574 			vmd = VM_DOMAIN(domain);
3575 			for (queue = 0; queue < PQ_COUNT; queue++) {
3576 				pq = &vmd->vmd_pagequeues[queue];
3577 				vm_pagequeue_lock(pq);
3578 				critical_enter();
3579 				vm_pqbatch_process(pq,
3580 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3581 				critical_exit();
3582 				vm_pagequeue_unlock(pq);
3583 			}
3584 		}
3585 	}
3586 	thread_lock(td);
3587 	sched_unbind(td);
3588 	thread_unlock(td);
3589 }
3590 
3591 /*
3592  *	vm_page_dequeue_deferred:	[ internal use only ]
3593  *
3594  *	Request removal of the given page from its current page
3595  *	queue.  Physical removal from the queue may be deferred
3596  *	indefinitely.
3597  *
3598  *	The page must be locked.
3599  */
3600 void
3601 vm_page_dequeue_deferred(vm_page_t m)
3602 {
3603 	vm_page_astate_t new, old;
3604 
3605 	old = vm_page_astate_load(m);
3606 	do {
3607 		if (old.queue == PQ_NONE) {
3608 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3609 			    ("%s: page %p has unexpected queue state",
3610 			    __func__, m));
3611 			break;
3612 		}
3613 		new = old;
3614 		new.flags |= PGA_DEQUEUE;
3615 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3616 }
3617 
3618 /*
3619  *	vm_page_dequeue:
3620  *
3621  *	Remove the page from whichever page queue it's in, if any, before
3622  *	returning.
3623  */
3624 void
3625 vm_page_dequeue(vm_page_t m)
3626 {
3627 	vm_page_astate_t new, old;
3628 
3629 	old = vm_page_astate_load(m);
3630 	do {
3631 		if (old.queue == PQ_NONE) {
3632 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3633 			    ("%s: page %p has unexpected queue state",
3634 			    __func__, m));
3635 			break;
3636 		}
3637 		new = old;
3638 		new.flags &= ~PGA_QUEUE_OP_MASK;
3639 		new.queue = PQ_NONE;
3640 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3641 
3642 }
3643 
3644 /*
3645  * Schedule the given page for insertion into the specified page queue.
3646  * Physical insertion of the page may be deferred indefinitely.
3647  */
3648 static void
3649 vm_page_enqueue(vm_page_t m, uint8_t queue)
3650 {
3651 
3652 	KASSERT(m->a.queue == PQ_NONE &&
3653 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3654 	    ("%s: page %p is already enqueued", __func__, m));
3655 	KASSERT(m->ref_count > 0,
3656 	    ("%s: page %p does not carry any references", __func__, m));
3657 
3658 	m->a.queue = queue;
3659 	if ((m->a.flags & PGA_REQUEUE) == 0)
3660 		vm_page_aflag_set(m, PGA_REQUEUE);
3661 	vm_page_pqbatch_submit(m, queue);
3662 }
3663 
3664 /*
3665  *	vm_page_free_prep:
3666  *
3667  *	Prepares the given page to be put on the free list,
3668  *	disassociating it from any VM object. The caller may return
3669  *	the page to the free list only if this function returns true.
3670  *
3671  *	The object must be locked.  The page must be locked if it is
3672  *	managed.
3673  */
3674 static bool
3675 vm_page_free_prep(vm_page_t m)
3676 {
3677 
3678 	/*
3679 	 * Synchronize with threads that have dropped a reference to this
3680 	 * page.
3681 	 */
3682 	atomic_thread_fence_acq();
3683 
3684 	if (vm_page_sbusied(m))
3685 		panic("vm_page_free_prep: freeing shared busy page %p", m);
3686 
3687 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3688 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3689 		uint64_t *p;
3690 		int i;
3691 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3692 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3693 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3694 			    m, i, (uintmax_t)*p));
3695 	}
3696 #endif
3697 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3698 		KASSERT(!pmap_page_is_mapped(m),
3699 		    ("vm_page_free_prep: freeing mapped page %p", m));
3700 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3701 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3702 	} else {
3703 		KASSERT(m->a.queue == PQ_NONE,
3704 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3705 	}
3706 	VM_CNT_INC(v_tfree);
3707 
3708 	if (m->object != NULL) {
3709 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3710 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
3711 		    ("vm_page_free_prep: managed flag mismatch for page %p",
3712 		    m));
3713 		vm_page_object_remove(m);
3714 
3715 		/*
3716 		 * The object reference can be released without an atomic
3717 		 * operation.
3718 		 */
3719 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3720 		    m->ref_count == VPRC_OBJREF,
3721 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3722 		    m, m->ref_count));
3723 		m->object = NULL;
3724 		m->ref_count -= VPRC_OBJREF;
3725 		vm_page_xunbusy(m);
3726 	}
3727 
3728 	if (vm_page_xbusied(m))
3729 		panic("vm_page_free_prep: freeing exclusive busy page %p", m);
3730 
3731 	/*
3732 	 * If fictitious remove object association and
3733 	 * return.
3734 	 */
3735 	if ((m->flags & PG_FICTITIOUS) != 0) {
3736 		KASSERT(m->ref_count == 1,
3737 		    ("fictitious page %p is referenced", m));
3738 		KASSERT(m->a.queue == PQ_NONE,
3739 		    ("fictitious page %p is queued", m));
3740 		return (false);
3741 	}
3742 
3743 	/*
3744 	 * Pages need not be dequeued before they are returned to the physical
3745 	 * memory allocator, but they must at least be marked for a deferred
3746 	 * dequeue.
3747 	 */
3748 	if ((m->oflags & VPO_UNMANAGED) == 0)
3749 		vm_page_dequeue_deferred(m);
3750 
3751 	m->valid = 0;
3752 	vm_page_undirty(m);
3753 
3754 	if (m->ref_count != 0)
3755 		panic("vm_page_free_prep: page %p has references", m);
3756 
3757 	/*
3758 	 * Restore the default memory attribute to the page.
3759 	 */
3760 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3761 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3762 
3763 #if VM_NRESERVLEVEL > 0
3764 	/*
3765 	 * Determine whether the page belongs to a reservation.  If the page was
3766 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3767 	 * as an optimization, we avoid the check in that case.
3768 	 */
3769 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3770 		return (false);
3771 #endif
3772 
3773 	return (true);
3774 }
3775 
3776 /*
3777  *	vm_page_free_toq:
3778  *
3779  *	Returns the given page to the free list, disassociating it
3780  *	from any VM object.
3781  *
3782  *	The object must be locked.  The page must be locked if it is
3783  *	managed.
3784  */
3785 static void
3786 vm_page_free_toq(vm_page_t m)
3787 {
3788 	struct vm_domain *vmd;
3789 	uma_zone_t zone;
3790 
3791 	if (!vm_page_free_prep(m))
3792 		return;
3793 
3794 	vmd = vm_pagequeue_domain(m);
3795 	zone = vmd->vmd_pgcache[m->pool].zone;
3796 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3797 		uma_zfree(zone, m);
3798 		return;
3799 	}
3800 	vm_domain_free_lock(vmd);
3801 	vm_phys_free_pages(m, 0);
3802 	vm_domain_free_unlock(vmd);
3803 	vm_domain_freecnt_inc(vmd, 1);
3804 }
3805 
3806 /*
3807  *	vm_page_free_pages_toq:
3808  *
3809  *	Returns a list of pages to the free list, disassociating it
3810  *	from any VM object.  In other words, this is equivalent to
3811  *	calling vm_page_free_toq() for each page of a list of VM objects.
3812  *
3813  *	The objects must be locked.  The pages must be locked if it is
3814  *	managed.
3815  */
3816 void
3817 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3818 {
3819 	vm_page_t m;
3820 	int count;
3821 
3822 	if (SLIST_EMPTY(free))
3823 		return;
3824 
3825 	count = 0;
3826 	while ((m = SLIST_FIRST(free)) != NULL) {
3827 		count++;
3828 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3829 		vm_page_free_toq(m);
3830 	}
3831 
3832 	if (update_wire_count)
3833 		vm_wire_sub(count);
3834 }
3835 
3836 /*
3837  * Mark this page as wired down, preventing reclamation by the page daemon
3838  * or when the containing object is destroyed.
3839  */
3840 void
3841 vm_page_wire(vm_page_t m)
3842 {
3843 	u_int old;
3844 
3845 	KASSERT(m->object != NULL,
3846 	    ("vm_page_wire: page %p does not belong to an object", m));
3847 	if (!vm_page_busied(m) && !vm_object_busied(m->object))
3848 		VM_OBJECT_ASSERT_LOCKED(m->object);
3849 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3850 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
3851 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
3852 
3853 	old = atomic_fetchadd_int(&m->ref_count, 1);
3854 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3855 	    ("vm_page_wire: counter overflow for page %p", m));
3856 	if (VPRC_WIRE_COUNT(old) == 0) {
3857 		if ((m->oflags & VPO_UNMANAGED) == 0)
3858 			vm_page_aflag_set(m, PGA_DEQUEUE);
3859 		vm_wire_add(1);
3860 	}
3861 }
3862 
3863 /*
3864  * Attempt to wire a mapped page following a pmap lookup of that page.
3865  * This may fail if a thread is concurrently tearing down mappings of the page.
3866  * The transient failure is acceptable because it translates to the
3867  * failure of the caller pmap_extract_and_hold(), which should be then
3868  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3869  */
3870 bool
3871 vm_page_wire_mapped(vm_page_t m)
3872 {
3873 	u_int old;
3874 
3875 	old = m->ref_count;
3876 	do {
3877 		KASSERT(old > 0,
3878 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3879 		if ((old & VPRC_BLOCKED) != 0)
3880 			return (false);
3881 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3882 
3883 	if (VPRC_WIRE_COUNT(old) == 0) {
3884 		if ((m->oflags & VPO_UNMANAGED) == 0)
3885 			vm_page_aflag_set(m, PGA_DEQUEUE);
3886 		vm_wire_add(1);
3887 	}
3888 	return (true);
3889 }
3890 
3891 /*
3892  * Release a wiring reference to a managed page.  If the page still belongs to
3893  * an object, update its position in the page queues to reflect the reference.
3894  * If the wiring was the last reference to the page, free the page.
3895  */
3896 static void
3897 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3898 {
3899 	u_int old;
3900 
3901 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3902 	    ("%s: page %p is unmanaged", __func__, m));
3903 
3904 	/*
3905 	 * Update LRU state before releasing the wiring reference.
3906 	 * Use a release store when updating the reference count to
3907 	 * synchronize with vm_page_free_prep().
3908 	 */
3909 	old = m->ref_count;
3910 	do {
3911 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
3912 		    ("vm_page_unwire: wire count underflow for page %p", m));
3913 
3914 		if (old > VPRC_OBJREF + 1) {
3915 			/*
3916 			 * The page has at least one other wiring reference.  An
3917 			 * earlier iteration of this loop may have called
3918 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3919 			 * re-set it if necessary.
3920 			 */
3921 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3922 				vm_page_aflag_set(m, PGA_DEQUEUE);
3923 		} else if (old == VPRC_OBJREF + 1) {
3924 			/*
3925 			 * This is the last wiring.  Clear PGA_DEQUEUE and
3926 			 * update the page's queue state to reflect the
3927 			 * reference.  If the page does not belong to an object
3928 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3929 			 * clear leftover queue state.
3930 			 */
3931 			vm_page_release_toq(m, nqueue, false);
3932 		} else if (old == 1) {
3933 			vm_page_aflag_clear(m, PGA_DEQUEUE);
3934 		}
3935 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3936 
3937 	if (VPRC_WIRE_COUNT(old) == 1) {
3938 		vm_wire_sub(1);
3939 		if (old == 1)
3940 			vm_page_free(m);
3941 	}
3942 }
3943 
3944 /*
3945  * Release one wiring of the specified page, potentially allowing it to be
3946  * paged out.
3947  *
3948  * Only managed pages belonging to an object can be paged out.  If the number
3949  * of wirings transitions to zero and the page is eligible for page out, then
3950  * the page is added to the specified paging queue.  If the released wiring
3951  * represented the last reference to the page, the page is freed.
3952  *
3953  * A managed page must be locked.
3954  */
3955 void
3956 vm_page_unwire(vm_page_t m, uint8_t nqueue)
3957 {
3958 
3959 	KASSERT(nqueue < PQ_COUNT,
3960 	    ("vm_page_unwire: invalid queue %u request for page %p",
3961 	    nqueue, m));
3962 
3963 	if ((m->oflags & VPO_UNMANAGED) != 0) {
3964 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
3965 			vm_page_free(m);
3966 		return;
3967 	}
3968 	vm_page_unwire_managed(m, nqueue, false);
3969 }
3970 
3971 /*
3972  * Unwire a page without (re-)inserting it into a page queue.  It is up
3973  * to the caller to enqueue, requeue, or free the page as appropriate.
3974  * In most cases involving managed pages, vm_page_unwire() should be used
3975  * instead.
3976  */
3977 bool
3978 vm_page_unwire_noq(vm_page_t m)
3979 {
3980 	u_int old;
3981 
3982 	old = vm_page_drop(m, 1);
3983 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
3984 	    ("vm_page_unref: counter underflow for page %p", m));
3985 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3986 	    ("vm_page_unref: missing ref on fictitious page %p", m));
3987 
3988 	if (VPRC_WIRE_COUNT(old) > 1)
3989 		return (false);
3990 	if ((m->oflags & VPO_UNMANAGED) == 0)
3991 		vm_page_aflag_clear(m, PGA_DEQUEUE);
3992 	vm_wire_sub(1);
3993 	return (true);
3994 }
3995 
3996 /*
3997  * Ensure that the page ends up in the specified page queue.  If the page is
3998  * active or being moved to the active queue, ensure that its act_count is
3999  * at least ACT_INIT but do not otherwise mess with it.
4000  *
4001  * A managed page must be locked.
4002  */
4003 static __always_inline void
4004 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4005 {
4006 	vm_page_astate_t old, new;
4007 
4008 	KASSERT(m->ref_count > 0,
4009 	    ("%s: page %p does not carry any references", __func__, m));
4010 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4011 	    ("%s: invalid flags %x", __func__, nflag));
4012 
4013 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4014 		return;
4015 
4016 	old = vm_page_astate_load(m);
4017 	do {
4018 		if ((old.flags & PGA_DEQUEUE) != 0)
4019 			break;
4020 		new = old;
4021 		new.flags &= ~PGA_QUEUE_OP_MASK;
4022 		if (nqueue == PQ_ACTIVE)
4023 			new.act_count = max(old.act_count, ACT_INIT);
4024 		if (old.queue == nqueue) {
4025 			if (nqueue != PQ_ACTIVE)
4026 				new.flags |= nflag;
4027 		} else {
4028 			new.flags |= nflag;
4029 			new.queue = nqueue;
4030 		}
4031 	} while (!vm_page_pqstate_commit(m, &old, new));
4032 }
4033 
4034 /*
4035  * Put the specified page on the active list (if appropriate).
4036  */
4037 void
4038 vm_page_activate(vm_page_t m)
4039 {
4040 
4041 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4042 }
4043 
4044 /*
4045  * Move the specified page to the tail of the inactive queue, or requeue
4046  * the page if it is already in the inactive queue.
4047  */
4048 void
4049 vm_page_deactivate(vm_page_t m)
4050 {
4051 
4052 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4053 }
4054 
4055 void
4056 vm_page_deactivate_noreuse(vm_page_t m)
4057 {
4058 
4059 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4060 }
4061 
4062 /*
4063  * Put a page in the laundry, or requeue it if it is already there.
4064  */
4065 void
4066 vm_page_launder(vm_page_t m)
4067 {
4068 
4069 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4070 }
4071 
4072 /*
4073  * Put a page in the PQ_UNSWAPPABLE holding queue.
4074  */
4075 void
4076 vm_page_unswappable(vm_page_t m)
4077 {
4078 
4079 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4080 	    ("page %p already unswappable", m));
4081 
4082 	vm_page_dequeue(m);
4083 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4084 }
4085 
4086 /*
4087  * Release a page back to the page queues in preparation for unwiring.
4088  */
4089 static void
4090 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4091 {
4092 	vm_page_astate_t old, new;
4093 	uint16_t nflag;
4094 
4095 	/*
4096 	 * Use a check of the valid bits to determine whether we should
4097 	 * accelerate reclamation of the page.  The object lock might not be
4098 	 * held here, in which case the check is racy.  At worst we will either
4099 	 * accelerate reclamation of a valid page and violate LRU, or
4100 	 * unnecessarily defer reclamation of an invalid page.
4101 	 *
4102 	 * If we were asked to not cache the page, place it near the head of the
4103 	 * inactive queue so that is reclaimed sooner.
4104 	 */
4105 	if (noreuse || m->valid == 0) {
4106 		nqueue = PQ_INACTIVE;
4107 		nflag = PGA_REQUEUE_HEAD;
4108 	} else {
4109 		nflag = PGA_REQUEUE;
4110 	}
4111 
4112 	old = vm_page_astate_load(m);
4113 	do {
4114 		new = old;
4115 
4116 		/*
4117 		 * If the page is already in the active queue and we are not
4118 		 * trying to accelerate reclamation, simply mark it as
4119 		 * referenced and avoid any queue operations.
4120 		 */
4121 		new.flags &= ~PGA_QUEUE_OP_MASK;
4122 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4123 			new.flags |= PGA_REFERENCED;
4124 		else {
4125 			new.flags |= nflag;
4126 			new.queue = nqueue;
4127 		}
4128 	} while (!vm_page_pqstate_commit(m, &old, new));
4129 }
4130 
4131 /*
4132  * Unwire a page and either attempt to free it or re-add it to the page queues.
4133  */
4134 void
4135 vm_page_release(vm_page_t m, int flags)
4136 {
4137 	vm_object_t object;
4138 
4139 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4140 	    ("vm_page_release: page %p is unmanaged", m));
4141 
4142 	if ((flags & VPR_TRYFREE) != 0) {
4143 		for (;;) {
4144 			object = (vm_object_t)atomic_load_ptr(&m->object);
4145 			if (object == NULL)
4146 				break;
4147 			/* Depends on type-stability. */
4148 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4149 				break;
4150 			if (object == m->object) {
4151 				vm_page_release_locked(m, flags);
4152 				VM_OBJECT_WUNLOCK(object);
4153 				return;
4154 			}
4155 			VM_OBJECT_WUNLOCK(object);
4156 		}
4157 	}
4158 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4159 }
4160 
4161 /* See vm_page_release(). */
4162 void
4163 vm_page_release_locked(vm_page_t m, int flags)
4164 {
4165 
4166 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4167 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4168 	    ("vm_page_release_locked: page %p is unmanaged", m));
4169 
4170 	if (vm_page_unwire_noq(m)) {
4171 		if ((flags & VPR_TRYFREE) != 0 &&
4172 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4173 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4174 			vm_page_free(m);
4175 		} else {
4176 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4177 		}
4178 	}
4179 }
4180 
4181 static bool
4182 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4183 {
4184 	u_int old;
4185 
4186 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4187 	    ("vm_page_try_blocked_op: page %p has no object", m));
4188 	KASSERT(vm_page_busied(m),
4189 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4190 	VM_OBJECT_ASSERT_LOCKED(m->object);
4191 
4192 	old = m->ref_count;
4193 	do {
4194 		KASSERT(old != 0,
4195 		    ("vm_page_try_blocked_op: page %p has no references", m));
4196 		if (VPRC_WIRE_COUNT(old) != 0)
4197 			return (false);
4198 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4199 
4200 	(op)(m);
4201 
4202 	/*
4203 	 * If the object is read-locked, new wirings may be created via an
4204 	 * object lookup.
4205 	 */
4206 	old = vm_page_drop(m, VPRC_BLOCKED);
4207 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4208 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4209 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4210 	    old, m));
4211 	return (true);
4212 }
4213 
4214 /*
4215  * Atomically check for wirings and remove all mappings of the page.
4216  */
4217 bool
4218 vm_page_try_remove_all(vm_page_t m)
4219 {
4220 
4221 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4222 }
4223 
4224 /*
4225  * Atomically check for wirings and remove all writeable mappings of the page.
4226  */
4227 bool
4228 vm_page_try_remove_write(vm_page_t m)
4229 {
4230 
4231 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4232 }
4233 
4234 /*
4235  * vm_page_advise
4236  *
4237  * 	Apply the specified advice to the given page.
4238  *
4239  *	The object and page must be locked.
4240  */
4241 void
4242 vm_page_advise(vm_page_t m, int advice)
4243 {
4244 
4245 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4246 	if (advice == MADV_FREE)
4247 		/*
4248 		 * Mark the page clean.  This will allow the page to be freed
4249 		 * without first paging it out.  MADV_FREE pages are often
4250 		 * quickly reused by malloc(3), so we do not do anything that
4251 		 * would result in a page fault on a later access.
4252 		 */
4253 		vm_page_undirty(m);
4254 	else if (advice != MADV_DONTNEED) {
4255 		if (advice == MADV_WILLNEED)
4256 			vm_page_activate(m);
4257 		return;
4258 	}
4259 
4260 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4261 		vm_page_dirty(m);
4262 
4263 	/*
4264 	 * Clear any references to the page.  Otherwise, the page daemon will
4265 	 * immediately reactivate the page.
4266 	 */
4267 	vm_page_aflag_clear(m, PGA_REFERENCED);
4268 
4269 	/*
4270 	 * Place clean pages near the head of the inactive queue rather than
4271 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4272 	 * the page will be reused quickly.  Dirty pages not already in the
4273 	 * laundry are moved there.
4274 	 */
4275 	if (m->dirty == 0)
4276 		vm_page_deactivate_noreuse(m);
4277 	else if (!vm_page_in_laundry(m))
4278 		vm_page_launder(m);
4279 }
4280 
4281 static inline int
4282 vm_page_grab_pflags(int allocflags)
4283 {
4284 	int pflags;
4285 
4286 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4287 	    (allocflags & VM_ALLOC_WIRED) != 0,
4288 	    ("vm_page_grab_pflags: the pages must be busied or wired"));
4289 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4290 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4291 	    ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4292 	    "mismatch"));
4293 	pflags = allocflags &
4294 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4295 	    VM_ALLOC_NOBUSY);
4296 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4297 		pflags |= VM_ALLOC_WAITFAIL;
4298 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4299 		pflags |= VM_ALLOC_SBUSY;
4300 
4301 	return (pflags);
4302 }
4303 
4304 /*
4305  * Grab a page, waiting until we are waken up due to the page
4306  * changing state.  We keep on waiting, if the page continues
4307  * to be in the object.  If the page doesn't exist, first allocate it
4308  * and then conditionally zero it.
4309  *
4310  * This routine may sleep.
4311  *
4312  * The object must be locked on entry.  The lock will, however, be released
4313  * and reacquired if the routine sleeps.
4314  */
4315 vm_page_t
4316 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4317 {
4318 	vm_page_t m;
4319 	int pflags;
4320 
4321 	VM_OBJECT_ASSERT_WLOCKED(object);
4322 	pflags = vm_page_grab_pflags(allocflags);
4323 retrylookup:
4324 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4325 		if (!vm_page_acquire_flags(m, allocflags)) {
4326 			if (vm_page_busy_sleep_flags(object, m, "pgrbwt",
4327 			    allocflags))
4328 				goto retrylookup;
4329 			return (NULL);
4330 		}
4331 		goto out;
4332 	}
4333 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4334 		return (NULL);
4335 	m = vm_page_alloc(object, pindex, pflags);
4336 	if (m == NULL) {
4337 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4338 			return (NULL);
4339 		goto retrylookup;
4340 	}
4341 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4342 		pmap_zero_page(m);
4343 
4344 out:
4345 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4346 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4347 			vm_page_sunbusy(m);
4348 		else
4349 			vm_page_xunbusy(m);
4350 	}
4351 	return (m);
4352 }
4353 
4354 /*
4355  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4356  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4357  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4358  * in simultaneously.  Additional pages will be left on a paging queue but
4359  * will neither be wired nor busy regardless of allocflags.
4360  */
4361 int
4362 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4363 {
4364 	vm_page_t m;
4365 	vm_page_t ma[VM_INITIAL_PAGEIN];
4366 	bool sleep, xbusy;
4367 	int after, i, pflags, rv;
4368 
4369 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4370 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4371 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4372 	KASSERT((allocflags &
4373 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4374 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4375 	VM_OBJECT_ASSERT_WLOCKED(object);
4376 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4377 	pflags |= VM_ALLOC_WAITFAIL;
4378 
4379 retrylookup:
4380 	xbusy = false;
4381 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4382 		/*
4383 		 * If the page is fully valid it can only become invalid
4384 		 * with the object lock held.  If it is not valid it can
4385 		 * become valid with the busy lock held.  Therefore, we
4386 		 * may unnecessarily lock the exclusive busy here if we
4387 		 * race with I/O completion not using the object lock.
4388 		 * However, we will not end up with an invalid page and a
4389 		 * shared lock.
4390 		 */
4391 		if (!vm_page_all_valid(m) ||
4392 		    (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4393 			sleep = !vm_page_tryxbusy(m);
4394 			xbusy = true;
4395 		} else
4396 			sleep = !vm_page_trysbusy(m);
4397 		if (sleep) {
4398 			(void)vm_page_busy_sleep_flags(object, m, "pgrbwt",
4399 			    allocflags);
4400 			goto retrylookup;
4401 		}
4402 		if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4403 		   !vm_page_all_valid(m)) {
4404 			if (xbusy)
4405 				vm_page_xunbusy(m);
4406 			else
4407 				vm_page_sunbusy(m);
4408 			*mp = NULL;
4409 			return (VM_PAGER_FAIL);
4410 		}
4411 		if ((allocflags & VM_ALLOC_WIRED) != 0)
4412 			vm_page_wire(m);
4413 		if (vm_page_all_valid(m))
4414 			goto out;
4415 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4416 		*mp = NULL;
4417 		return (VM_PAGER_FAIL);
4418 	} else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4419 		xbusy = true;
4420 	} else {
4421 		goto retrylookup;
4422 	}
4423 
4424 	vm_page_assert_xbusied(m);
4425 	MPASS(xbusy);
4426 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4427 		after = MIN(after, VM_INITIAL_PAGEIN);
4428 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4429 		after = MAX(after, 1);
4430 		ma[0] = m;
4431 		for (i = 1; i < after; i++) {
4432 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4433 				if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4434 					break;
4435 			} else {
4436 				ma[i] = vm_page_alloc(object, m->pindex + i,
4437 				    VM_ALLOC_NORMAL);
4438 				if (ma[i] == NULL)
4439 					break;
4440 			}
4441 		}
4442 		after = i;
4443 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4444 		/* Pager may have replaced a page. */
4445 		m = ma[0];
4446 		if (rv != VM_PAGER_OK) {
4447 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4448 				vm_page_unwire_noq(m);
4449 			for (i = 0; i < after; i++) {
4450 				if (!vm_page_wired(ma[i]))
4451 					vm_page_free(ma[i]);
4452 				else
4453 					vm_page_xunbusy(ma[i]);
4454 			}
4455 			*mp = NULL;
4456 			return (rv);
4457 		}
4458 		for (i = 1; i < after; i++)
4459 			vm_page_readahead_finish(ma[i]);
4460 		MPASS(vm_page_all_valid(m));
4461 	} else {
4462 		vm_page_zero_invalid(m, TRUE);
4463 	}
4464 out:
4465 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4466 		if (xbusy)
4467 			vm_page_xunbusy(m);
4468 		else
4469 			vm_page_sunbusy(m);
4470 	}
4471 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4472 		vm_page_busy_downgrade(m);
4473 	*mp = m;
4474 	return (VM_PAGER_OK);
4475 }
4476 
4477 /*
4478  * Return the specified range of pages from the given object.  For each
4479  * page offset within the range, if a page already exists within the object
4480  * at that offset and it is busy, then wait for it to change state.  If,
4481  * instead, the page doesn't exist, then allocate it.
4482  *
4483  * The caller must always specify an allocation class.
4484  *
4485  * allocation classes:
4486  *	VM_ALLOC_NORMAL		normal process request
4487  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4488  *
4489  * The caller must always specify that the pages are to be busied and/or
4490  * wired.
4491  *
4492  * optional allocation flags:
4493  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4494  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4495  *	VM_ALLOC_NOWAIT		do not sleep
4496  *	VM_ALLOC_SBUSY		set page to sbusy state
4497  *	VM_ALLOC_WIRED		wire the pages
4498  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4499  *
4500  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4501  * may return a partial prefix of the requested range.
4502  */
4503 int
4504 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4505     vm_page_t *ma, int count)
4506 {
4507 	vm_page_t m, mpred;
4508 	int pflags;
4509 	int i;
4510 
4511 	VM_OBJECT_ASSERT_WLOCKED(object);
4512 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4513 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4514 
4515 	pflags = vm_page_grab_pflags(allocflags);
4516 	if (count == 0)
4517 		return (0);
4518 
4519 	i = 0;
4520 retrylookup:
4521 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4522 	if (m == NULL || m->pindex != pindex + i) {
4523 		mpred = m;
4524 		m = NULL;
4525 	} else
4526 		mpred = TAILQ_PREV(m, pglist, listq);
4527 	for (; i < count; i++) {
4528 		if (m != NULL) {
4529 			if (!vm_page_acquire_flags(m, allocflags)) {
4530 				if (vm_page_busy_sleep_flags(object, m,
4531 				    "grbmaw", allocflags))
4532 					goto retrylookup;
4533 				break;
4534 			}
4535 		} else {
4536 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4537 				break;
4538 			m = vm_page_alloc_after(object, pindex + i,
4539 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4540 			if (m == NULL) {
4541 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4542 					break;
4543 				goto retrylookup;
4544 			}
4545 		}
4546 		if (vm_page_none_valid(m) &&
4547 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4548 			if ((m->flags & PG_ZERO) == 0)
4549 				pmap_zero_page(m);
4550 			vm_page_valid(m);
4551 		}
4552 		if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4553 			if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4554 				vm_page_sunbusy(m);
4555 			else
4556 				vm_page_xunbusy(m);
4557 		}
4558 		ma[i] = mpred = m;
4559 		m = vm_page_next(m);
4560 	}
4561 	return (i);
4562 }
4563 
4564 /*
4565  * Mapping function for valid or dirty bits in a page.
4566  *
4567  * Inputs are required to range within a page.
4568  */
4569 vm_page_bits_t
4570 vm_page_bits(int base, int size)
4571 {
4572 	int first_bit;
4573 	int last_bit;
4574 
4575 	KASSERT(
4576 	    base + size <= PAGE_SIZE,
4577 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4578 	);
4579 
4580 	if (size == 0)		/* handle degenerate case */
4581 		return (0);
4582 
4583 	first_bit = base >> DEV_BSHIFT;
4584 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4585 
4586 	return (((vm_page_bits_t)2 << last_bit) -
4587 	    ((vm_page_bits_t)1 << first_bit));
4588 }
4589 
4590 void
4591 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4592 {
4593 
4594 #if PAGE_SIZE == 32768
4595 	atomic_set_64((uint64_t *)bits, set);
4596 #elif PAGE_SIZE == 16384
4597 	atomic_set_32((uint32_t *)bits, set);
4598 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4599 	atomic_set_16((uint16_t *)bits, set);
4600 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4601 	atomic_set_8((uint8_t *)bits, set);
4602 #else		/* PAGE_SIZE <= 8192 */
4603 	uintptr_t addr;
4604 	int shift;
4605 
4606 	addr = (uintptr_t)bits;
4607 	/*
4608 	 * Use a trick to perform a 32-bit atomic on the
4609 	 * containing aligned word, to not depend on the existence
4610 	 * of atomic_{set, clear}_{8, 16}.
4611 	 */
4612 	shift = addr & (sizeof(uint32_t) - 1);
4613 #if BYTE_ORDER == BIG_ENDIAN
4614 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4615 #else
4616 	shift *= NBBY;
4617 #endif
4618 	addr &= ~(sizeof(uint32_t) - 1);
4619 	atomic_set_32((uint32_t *)addr, set << shift);
4620 #endif		/* PAGE_SIZE */
4621 }
4622 
4623 static inline void
4624 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4625 {
4626 
4627 #if PAGE_SIZE == 32768
4628 	atomic_clear_64((uint64_t *)bits, clear);
4629 #elif PAGE_SIZE == 16384
4630 	atomic_clear_32((uint32_t *)bits, clear);
4631 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4632 	atomic_clear_16((uint16_t *)bits, clear);
4633 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4634 	atomic_clear_8((uint8_t *)bits, clear);
4635 #else		/* PAGE_SIZE <= 8192 */
4636 	uintptr_t addr;
4637 	int shift;
4638 
4639 	addr = (uintptr_t)bits;
4640 	/*
4641 	 * Use a trick to perform a 32-bit atomic on the
4642 	 * containing aligned word, to not depend on the existence
4643 	 * of atomic_{set, clear}_{8, 16}.
4644 	 */
4645 	shift = addr & (sizeof(uint32_t) - 1);
4646 #if BYTE_ORDER == BIG_ENDIAN
4647 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4648 #else
4649 	shift *= NBBY;
4650 #endif
4651 	addr &= ~(sizeof(uint32_t) - 1);
4652 	atomic_clear_32((uint32_t *)addr, clear << shift);
4653 #endif		/* PAGE_SIZE */
4654 }
4655 
4656 static inline vm_page_bits_t
4657 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4658 {
4659 #if PAGE_SIZE == 32768
4660 	uint64_t old;
4661 
4662 	old = *bits;
4663 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4664 	return (old);
4665 #elif PAGE_SIZE == 16384
4666 	uint32_t old;
4667 
4668 	old = *bits;
4669 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4670 	return (old);
4671 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4672 	uint16_t old;
4673 
4674 	old = *bits;
4675 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4676 	return (old);
4677 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4678 	uint8_t old;
4679 
4680 	old = *bits;
4681 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4682 	return (old);
4683 #else		/* PAGE_SIZE <= 4096*/
4684 	uintptr_t addr;
4685 	uint32_t old, new, mask;
4686 	int shift;
4687 
4688 	addr = (uintptr_t)bits;
4689 	/*
4690 	 * Use a trick to perform a 32-bit atomic on the
4691 	 * containing aligned word, to not depend on the existence
4692 	 * of atomic_{set, swap, clear}_{8, 16}.
4693 	 */
4694 	shift = addr & (sizeof(uint32_t) - 1);
4695 #if BYTE_ORDER == BIG_ENDIAN
4696 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4697 #else
4698 	shift *= NBBY;
4699 #endif
4700 	addr &= ~(sizeof(uint32_t) - 1);
4701 	mask = VM_PAGE_BITS_ALL << shift;
4702 
4703 	old = *bits;
4704 	do {
4705 		new = old & ~mask;
4706 		new |= newbits << shift;
4707 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4708 	return (old >> shift);
4709 #endif		/* PAGE_SIZE */
4710 }
4711 
4712 /*
4713  *	vm_page_set_valid_range:
4714  *
4715  *	Sets portions of a page valid.  The arguments are expected
4716  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4717  *	of any partial chunks touched by the range.  The invalid portion of
4718  *	such chunks will be zeroed.
4719  *
4720  *	(base + size) must be less then or equal to PAGE_SIZE.
4721  */
4722 void
4723 vm_page_set_valid_range(vm_page_t m, int base, int size)
4724 {
4725 	int endoff, frag;
4726 	vm_page_bits_t pagebits;
4727 
4728 	vm_page_assert_busied(m);
4729 	if (size == 0)	/* handle degenerate case */
4730 		return;
4731 
4732 	/*
4733 	 * If the base is not DEV_BSIZE aligned and the valid
4734 	 * bit is clear, we have to zero out a portion of the
4735 	 * first block.
4736 	 */
4737 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4738 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4739 		pmap_zero_page_area(m, frag, base - frag);
4740 
4741 	/*
4742 	 * If the ending offset is not DEV_BSIZE aligned and the
4743 	 * valid bit is clear, we have to zero out a portion of
4744 	 * the last block.
4745 	 */
4746 	endoff = base + size;
4747 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4748 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4749 		pmap_zero_page_area(m, endoff,
4750 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4751 
4752 	/*
4753 	 * Assert that no previously invalid block that is now being validated
4754 	 * is already dirty.
4755 	 */
4756 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4757 	    ("vm_page_set_valid_range: page %p is dirty", m));
4758 
4759 	/*
4760 	 * Set valid bits inclusive of any overlap.
4761 	 */
4762 	pagebits = vm_page_bits(base, size);
4763 	if (vm_page_xbusied(m))
4764 		m->valid |= pagebits;
4765 	else
4766 		vm_page_bits_set(m, &m->valid, pagebits);
4767 }
4768 
4769 /*
4770  * Set the page dirty bits and free the invalid swap space if
4771  * present.  Returns the previous dirty bits.
4772  */
4773 vm_page_bits_t
4774 vm_page_set_dirty(vm_page_t m)
4775 {
4776 	vm_page_bits_t old;
4777 
4778 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
4779 
4780 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
4781 		old = m->dirty;
4782 		m->dirty = VM_PAGE_BITS_ALL;
4783 	} else
4784 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
4785 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
4786 		vm_pager_page_unswapped(m);
4787 
4788 	return (old);
4789 }
4790 
4791 /*
4792  * Clear the given bits from the specified page's dirty field.
4793  */
4794 static __inline void
4795 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4796 {
4797 
4798 	vm_page_assert_busied(m);
4799 
4800 	/*
4801 	 * If the page is xbusied and not write mapped we are the
4802 	 * only thread that can modify dirty bits.  Otherwise, The pmap
4803 	 * layer can call vm_page_dirty() without holding a distinguished
4804 	 * lock.  The combination of page busy and atomic operations
4805 	 * suffice to guarantee consistency of the page dirty field.
4806 	 */
4807 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4808 		m->dirty &= ~pagebits;
4809 	else
4810 		vm_page_bits_clear(m, &m->dirty, pagebits);
4811 }
4812 
4813 /*
4814  *	vm_page_set_validclean:
4815  *
4816  *	Sets portions of a page valid and clean.  The arguments are expected
4817  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4818  *	of any partial chunks touched by the range.  The invalid portion of
4819  *	such chunks will be zero'd.
4820  *
4821  *	(base + size) must be less then or equal to PAGE_SIZE.
4822  */
4823 void
4824 vm_page_set_validclean(vm_page_t m, int base, int size)
4825 {
4826 	vm_page_bits_t oldvalid, pagebits;
4827 	int endoff, frag;
4828 
4829 	vm_page_assert_busied(m);
4830 	if (size == 0)	/* handle degenerate case */
4831 		return;
4832 
4833 	/*
4834 	 * If the base is not DEV_BSIZE aligned and the valid
4835 	 * bit is clear, we have to zero out a portion of the
4836 	 * first block.
4837 	 */
4838 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4839 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4840 		pmap_zero_page_area(m, frag, base - frag);
4841 
4842 	/*
4843 	 * If the ending offset is not DEV_BSIZE aligned and the
4844 	 * valid bit is clear, we have to zero out a portion of
4845 	 * the last block.
4846 	 */
4847 	endoff = base + size;
4848 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4849 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4850 		pmap_zero_page_area(m, endoff,
4851 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4852 
4853 	/*
4854 	 * Set valid, clear dirty bits.  If validating the entire
4855 	 * page we can safely clear the pmap modify bit.  We also
4856 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
4857 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4858 	 * be set again.
4859 	 *
4860 	 * We set valid bits inclusive of any overlap, but we can only
4861 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4862 	 * the range.
4863 	 */
4864 	oldvalid = m->valid;
4865 	pagebits = vm_page_bits(base, size);
4866 	if (vm_page_xbusied(m))
4867 		m->valid |= pagebits;
4868 	else
4869 		vm_page_bits_set(m, &m->valid, pagebits);
4870 #if 0	/* NOT YET */
4871 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4872 		frag = DEV_BSIZE - frag;
4873 		base += frag;
4874 		size -= frag;
4875 		if (size < 0)
4876 			size = 0;
4877 	}
4878 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4879 #endif
4880 	if (base == 0 && size == PAGE_SIZE) {
4881 		/*
4882 		 * The page can only be modified within the pmap if it is
4883 		 * mapped, and it can only be mapped if it was previously
4884 		 * fully valid.
4885 		 */
4886 		if (oldvalid == VM_PAGE_BITS_ALL)
4887 			/*
4888 			 * Perform the pmap_clear_modify() first.  Otherwise,
4889 			 * a concurrent pmap operation, such as
4890 			 * pmap_protect(), could clear a modification in the
4891 			 * pmap and set the dirty field on the page before
4892 			 * pmap_clear_modify() had begun and after the dirty
4893 			 * field was cleared here.
4894 			 */
4895 			pmap_clear_modify(m);
4896 		m->dirty = 0;
4897 		vm_page_aflag_clear(m, PGA_NOSYNC);
4898 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
4899 		m->dirty &= ~pagebits;
4900 	else
4901 		vm_page_clear_dirty_mask(m, pagebits);
4902 }
4903 
4904 void
4905 vm_page_clear_dirty(vm_page_t m, int base, int size)
4906 {
4907 
4908 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4909 }
4910 
4911 /*
4912  *	vm_page_set_invalid:
4913  *
4914  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4915  *	valid and dirty bits for the effected areas are cleared.
4916  */
4917 void
4918 vm_page_set_invalid(vm_page_t m, int base, int size)
4919 {
4920 	vm_page_bits_t bits;
4921 	vm_object_t object;
4922 
4923 	/*
4924 	 * The object lock is required so that pages can't be mapped
4925 	 * read-only while we're in the process of invalidating them.
4926 	 */
4927 	object = m->object;
4928 	VM_OBJECT_ASSERT_WLOCKED(object);
4929 	vm_page_assert_busied(m);
4930 
4931 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4932 	    size >= object->un_pager.vnp.vnp_size)
4933 		bits = VM_PAGE_BITS_ALL;
4934 	else
4935 		bits = vm_page_bits(base, size);
4936 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
4937 		pmap_remove_all(m);
4938 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
4939 	    !pmap_page_is_mapped(m),
4940 	    ("vm_page_set_invalid: page %p is mapped", m));
4941 	if (vm_page_xbusied(m)) {
4942 		m->valid &= ~bits;
4943 		m->dirty &= ~bits;
4944 	} else {
4945 		vm_page_bits_clear(m, &m->valid, bits);
4946 		vm_page_bits_clear(m, &m->dirty, bits);
4947 	}
4948 }
4949 
4950 /*
4951  *	vm_page_invalid:
4952  *
4953  *	Invalidates the entire page.  The page must be busy, unmapped, and
4954  *	the enclosing object must be locked.  The object locks protects
4955  *	against concurrent read-only pmap enter which is done without
4956  *	busy.
4957  */
4958 void
4959 vm_page_invalid(vm_page_t m)
4960 {
4961 
4962 	vm_page_assert_busied(m);
4963 	VM_OBJECT_ASSERT_LOCKED(m->object);
4964 	MPASS(!pmap_page_is_mapped(m));
4965 
4966 	if (vm_page_xbusied(m))
4967 		m->valid = 0;
4968 	else
4969 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
4970 }
4971 
4972 /*
4973  * vm_page_zero_invalid()
4974  *
4975  *	The kernel assumes that the invalid portions of a page contain
4976  *	garbage, but such pages can be mapped into memory by user code.
4977  *	When this occurs, we must zero out the non-valid portions of the
4978  *	page so user code sees what it expects.
4979  *
4980  *	Pages are most often semi-valid when the end of a file is mapped
4981  *	into memory and the file's size is not page aligned.
4982  */
4983 void
4984 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4985 {
4986 	int b;
4987 	int i;
4988 
4989 	/*
4990 	 * Scan the valid bits looking for invalid sections that
4991 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4992 	 * valid bit may be set ) have already been zeroed by
4993 	 * vm_page_set_validclean().
4994 	 */
4995 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4996 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4997 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4998 			if (i > b) {
4999 				pmap_zero_page_area(m,
5000 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5001 			}
5002 			b = i + 1;
5003 		}
5004 	}
5005 
5006 	/*
5007 	 * setvalid is TRUE when we can safely set the zero'd areas
5008 	 * as being valid.  We can do this if there are no cache consistancy
5009 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5010 	 */
5011 	if (setvalid)
5012 		vm_page_valid(m);
5013 }
5014 
5015 /*
5016  *	vm_page_is_valid:
5017  *
5018  *	Is (partial) page valid?  Note that the case where size == 0
5019  *	will return FALSE in the degenerate case where the page is
5020  *	entirely invalid, and TRUE otherwise.
5021  *
5022  *	Some callers envoke this routine without the busy lock held and
5023  *	handle races via higher level locks.  Typical callers should
5024  *	hold a busy lock to prevent invalidation.
5025  */
5026 int
5027 vm_page_is_valid(vm_page_t m, int base, int size)
5028 {
5029 	vm_page_bits_t bits;
5030 
5031 	bits = vm_page_bits(base, size);
5032 	return (m->valid != 0 && (m->valid & bits) == bits);
5033 }
5034 
5035 /*
5036  * Returns true if all of the specified predicates are true for the entire
5037  * (super)page and false otherwise.
5038  */
5039 bool
5040 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5041 {
5042 	vm_object_t object;
5043 	int i, npages;
5044 
5045 	object = m->object;
5046 	if (skip_m != NULL && skip_m->object != object)
5047 		return (false);
5048 	VM_OBJECT_ASSERT_LOCKED(object);
5049 	npages = atop(pagesizes[m->psind]);
5050 
5051 	/*
5052 	 * The physically contiguous pages that make up a superpage, i.e., a
5053 	 * page with a page size index ("psind") greater than zero, will
5054 	 * occupy adjacent entries in vm_page_array[].
5055 	 */
5056 	for (i = 0; i < npages; i++) {
5057 		/* Always test object consistency, including "skip_m". */
5058 		if (m[i].object != object)
5059 			return (false);
5060 		if (&m[i] == skip_m)
5061 			continue;
5062 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5063 			return (false);
5064 		if ((flags & PS_ALL_DIRTY) != 0) {
5065 			/*
5066 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5067 			 * might stop this case from spuriously returning
5068 			 * "false".  However, that would require a write lock
5069 			 * on the object containing "m[i]".
5070 			 */
5071 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5072 				return (false);
5073 		}
5074 		if ((flags & PS_ALL_VALID) != 0 &&
5075 		    m[i].valid != VM_PAGE_BITS_ALL)
5076 			return (false);
5077 	}
5078 	return (true);
5079 }
5080 
5081 /*
5082  * Set the page's dirty bits if the page is modified.
5083  */
5084 void
5085 vm_page_test_dirty(vm_page_t m)
5086 {
5087 
5088 	vm_page_assert_busied(m);
5089 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5090 		vm_page_dirty(m);
5091 }
5092 
5093 void
5094 vm_page_valid(vm_page_t m)
5095 {
5096 
5097 	vm_page_assert_busied(m);
5098 	if (vm_page_xbusied(m))
5099 		m->valid = VM_PAGE_BITS_ALL;
5100 	else
5101 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5102 }
5103 
5104 void
5105 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5106 {
5107 
5108 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5109 }
5110 
5111 void
5112 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5113 {
5114 
5115 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5116 }
5117 
5118 int
5119 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5120 {
5121 
5122 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5123 }
5124 
5125 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5126 void
5127 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5128 {
5129 
5130 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5131 }
5132 
5133 void
5134 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5135 {
5136 
5137 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5138 }
5139 #endif
5140 
5141 #ifdef INVARIANTS
5142 void
5143 vm_page_object_busy_assert(vm_page_t m)
5144 {
5145 
5146 	/*
5147 	 * Certain of the page's fields may only be modified by the
5148 	 * holder of a page or object busy.
5149 	 */
5150 	if (m->object != NULL && !vm_page_busied(m))
5151 		VM_OBJECT_ASSERT_BUSY(m->object);
5152 }
5153 
5154 void
5155 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5156 {
5157 
5158 	if ((bits & PGA_WRITEABLE) == 0)
5159 		return;
5160 
5161 	/*
5162 	 * The PGA_WRITEABLE flag can only be set if the page is
5163 	 * managed, is exclusively busied or the object is locked.
5164 	 * Currently, this flag is only set by pmap_enter().
5165 	 */
5166 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5167 	    ("PGA_WRITEABLE on unmanaged page"));
5168 	if (!vm_page_xbusied(m))
5169 		VM_OBJECT_ASSERT_BUSY(m->object);
5170 }
5171 #endif
5172 
5173 #include "opt_ddb.h"
5174 #ifdef DDB
5175 #include <sys/kernel.h>
5176 
5177 #include <ddb/ddb.h>
5178 
5179 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5180 {
5181 
5182 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5183 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5184 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5185 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5186 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5187 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5188 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5189 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5190 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5191 }
5192 
5193 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5194 {
5195 	int dom;
5196 
5197 	db_printf("pq_free %d\n", vm_free_count());
5198 	for (dom = 0; dom < vm_ndomains; dom++) {
5199 		db_printf(
5200     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5201 		    dom,
5202 		    vm_dom[dom].vmd_page_count,
5203 		    vm_dom[dom].vmd_free_count,
5204 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5205 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5206 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5207 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5208 	}
5209 }
5210 
5211 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5212 {
5213 	vm_page_t m;
5214 	boolean_t phys, virt;
5215 
5216 	if (!have_addr) {
5217 		db_printf("show pginfo addr\n");
5218 		return;
5219 	}
5220 
5221 	phys = strchr(modif, 'p') != NULL;
5222 	virt = strchr(modif, 'v') != NULL;
5223 	if (virt)
5224 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5225 	else if (phys)
5226 		m = PHYS_TO_VM_PAGE(addr);
5227 	else
5228 		m = (vm_page_t)addr;
5229 	db_printf(
5230     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
5231     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5232 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5233 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5234 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5235 }
5236 #endif /* DDB */
5237