xref: /freebsd/sys/vm/vm_page.c (revision 0610f417a4a664b7e3a8c57de800ae1f326a40e3)
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 queue_ops = EARLY_COUNTER;
138 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
139     CTLFLAG_RD, &queue_ops,
140     "Number of batched queue operations");
141 
142 static counter_u64_t queue_nops = EARLY_COUNTER;
143 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
144     CTLFLAG_RD, &queue_nops,
145     "Number of batched queue operations with no effects");
146 
147 static void
148 counter_startup(void)
149 {
150 
151 	queue_ops = counter_u64_alloc(M_WAITOK);
152 	queue_nops = counter_u64_alloc(M_WAITOK);
153 }
154 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
155 
156 /*
157  * bogus page -- for I/O to/from partially complete buffers,
158  * or for paging into sparsely invalid regions.
159  */
160 vm_page_t bogus_page;
161 
162 vm_page_t vm_page_array;
163 long vm_page_array_size;
164 long first_page;
165 
166 static int boot_pages;
167 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
168     &boot_pages, 0,
169     "number of pages allocated for bootstrapping the VM system");
170 
171 static TAILQ_HEAD(, vm_page) blacklist_head;
172 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
173 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
174     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
175 
176 static uma_zone_t fakepg_zone;
177 
178 static void vm_page_alloc_check(vm_page_t m);
179 static void _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
180     const char *wmesg, bool nonshared, bool locked);
181 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
182 static void vm_page_dequeue_complete(vm_page_t m);
183 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
184 static void vm_page_init(void *dummy);
185 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
186     vm_pindex_t pindex, vm_page_t mpred);
187 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
188     vm_page_t mpred);
189 static void vm_page_mvqueue(vm_page_t m, uint8_t queue);
190 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
191     vm_page_t m_run, vm_paddr_t high);
192 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
193     int req);
194 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
195     int flags);
196 static void vm_page_zone_release(void *arg, void **store, int cnt);
197 
198 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
199 
200 static void
201 vm_page_init(void *dummy)
202 {
203 
204 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
205 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
206 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
207 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
208 }
209 
210 /*
211  * The cache page zone is initialized later since we need to be able to allocate
212  * pages before UMA is fully initialized.
213  */
214 static void
215 vm_page_init_cache_zones(void *dummy __unused)
216 {
217 	struct vm_domain *vmd;
218 	struct vm_pgcache *pgcache;
219 	int cache, domain, maxcache, pool;
220 
221 	maxcache = 0;
222 	TUNABLE_INT_FETCH("vm.pgcache_zone_max", &maxcache);
223 	for (domain = 0; domain < vm_ndomains; domain++) {
224 		vmd = VM_DOMAIN(domain);
225 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
226 			pgcache = &vmd->vmd_pgcache[pool];
227 			pgcache->domain = domain;
228 			pgcache->pool = pool;
229 			pgcache->zone = uma_zcache_create("vm pgcache",
230 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
231 			    vm_page_zone_import, vm_page_zone_release, pgcache,
232 			    UMA_ZONE_VM);
233 
234 			/*
235 			 * Limit each pool's zone to 0.1% of the pages in the
236 			 * domain.
237 			 */
238 			cache = maxcache != 0 ? maxcache :
239 			    vmd->vmd_page_count / 1000;
240 			uma_zone_set_maxcache(pgcache->zone, cache);
241 		}
242 	}
243 }
244 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
245 
246 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
247 #if PAGE_SIZE == 32768
248 #ifdef CTASSERT
249 CTASSERT(sizeof(u_long) >= 8);
250 #endif
251 #endif
252 
253 /*
254  *	vm_set_page_size:
255  *
256  *	Sets the page size, perhaps based upon the memory
257  *	size.  Must be called before any use of page-size
258  *	dependent functions.
259  */
260 void
261 vm_set_page_size(void)
262 {
263 	if (vm_cnt.v_page_size == 0)
264 		vm_cnt.v_page_size = PAGE_SIZE;
265 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
266 		panic("vm_set_page_size: page size not a power of two");
267 }
268 
269 /*
270  *	vm_page_blacklist_next:
271  *
272  *	Find the next entry in the provided string of blacklist
273  *	addresses.  Entries are separated by space, comma, or newline.
274  *	If an invalid integer is encountered then the rest of the
275  *	string is skipped.  Updates the list pointer to the next
276  *	character, or NULL if the string is exhausted or invalid.
277  */
278 static vm_paddr_t
279 vm_page_blacklist_next(char **list, char *end)
280 {
281 	vm_paddr_t bad;
282 	char *cp, *pos;
283 
284 	if (list == NULL || *list == NULL)
285 		return (0);
286 	if (**list =='\0') {
287 		*list = NULL;
288 		return (0);
289 	}
290 
291 	/*
292 	 * If there's no end pointer then the buffer is coming from
293 	 * the kenv and we know it's null-terminated.
294 	 */
295 	if (end == NULL)
296 		end = *list + strlen(*list);
297 
298 	/* Ensure that strtoq() won't walk off the end */
299 	if (*end != '\0') {
300 		if (*end == '\n' || *end == ' ' || *end  == ',')
301 			*end = '\0';
302 		else {
303 			printf("Blacklist not terminated, skipping\n");
304 			*list = NULL;
305 			return (0);
306 		}
307 	}
308 
309 	for (pos = *list; *pos != '\0'; pos = cp) {
310 		bad = strtoq(pos, &cp, 0);
311 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
312 			if (bad == 0) {
313 				if (++cp < end)
314 					continue;
315 				else
316 					break;
317 			}
318 		} else
319 			break;
320 		if (*cp == '\0' || ++cp >= end)
321 			*list = NULL;
322 		else
323 			*list = cp;
324 		return (trunc_page(bad));
325 	}
326 	printf("Garbage in RAM blacklist, skipping\n");
327 	*list = NULL;
328 	return (0);
329 }
330 
331 bool
332 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
333 {
334 	struct vm_domain *vmd;
335 	vm_page_t m;
336 	int ret;
337 
338 	m = vm_phys_paddr_to_vm_page(pa);
339 	if (m == NULL)
340 		return (true); /* page does not exist, no failure */
341 
342 	vmd = vm_pagequeue_domain(m);
343 	vm_domain_free_lock(vmd);
344 	ret = vm_phys_unfree_page(m);
345 	vm_domain_free_unlock(vmd);
346 	if (ret != 0) {
347 		vm_domain_freecnt_inc(vmd, -1);
348 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
349 		if (verbose)
350 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
351 	}
352 	return (ret);
353 }
354 
355 /*
356  *	vm_page_blacklist_check:
357  *
358  *	Iterate through the provided string of blacklist addresses, pulling
359  *	each entry out of the physical allocator free list and putting it
360  *	onto a list for reporting via the vm.page_blacklist sysctl.
361  */
362 static void
363 vm_page_blacklist_check(char *list, char *end)
364 {
365 	vm_paddr_t pa;
366 	char *next;
367 
368 	next = list;
369 	while (next != NULL) {
370 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
371 			continue;
372 		vm_page_blacklist_add(pa, bootverbose);
373 	}
374 }
375 
376 /*
377  *	vm_page_blacklist_load:
378  *
379  *	Search for a special module named "ram_blacklist".  It'll be a
380  *	plain text file provided by the user via the loader directive
381  *	of the same name.
382  */
383 static void
384 vm_page_blacklist_load(char **list, char **end)
385 {
386 	void *mod;
387 	u_char *ptr;
388 	u_int len;
389 
390 	mod = NULL;
391 	ptr = NULL;
392 
393 	mod = preload_search_by_type("ram_blacklist");
394 	if (mod != NULL) {
395 		ptr = preload_fetch_addr(mod);
396 		len = preload_fetch_size(mod);
397         }
398 	*list = ptr;
399 	if (ptr != NULL)
400 		*end = ptr + len;
401 	else
402 		*end = NULL;
403 	return;
404 }
405 
406 static int
407 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
408 {
409 	vm_page_t m;
410 	struct sbuf sbuf;
411 	int error, first;
412 
413 	first = 1;
414 	error = sysctl_wire_old_buffer(req, 0);
415 	if (error != 0)
416 		return (error);
417 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
418 	TAILQ_FOREACH(m, &blacklist_head, listq) {
419 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
420 		    (uintmax_t)m->phys_addr);
421 		first = 0;
422 	}
423 	error = sbuf_finish(&sbuf);
424 	sbuf_delete(&sbuf);
425 	return (error);
426 }
427 
428 /*
429  * Initialize a dummy page for use in scans of the specified paging queue.
430  * In principle, this function only needs to set the flag PG_MARKER.
431  * Nonetheless, it write busies the page as a safety precaution.
432  */
433 static void
434 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
435 {
436 
437 	bzero(marker, sizeof(*marker));
438 	marker->flags = PG_MARKER;
439 	marker->a.flags = aflags;
440 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
441 	marker->a.queue = queue;
442 }
443 
444 static void
445 vm_page_domain_init(int domain)
446 {
447 	struct vm_domain *vmd;
448 	struct vm_pagequeue *pq;
449 	int i;
450 
451 	vmd = VM_DOMAIN(domain);
452 	bzero(vmd, sizeof(*vmd));
453 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
454 	    "vm inactive pagequeue";
455 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
456 	    "vm active pagequeue";
457 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
458 	    "vm laundry pagequeue";
459 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
460 	    "vm unswappable pagequeue";
461 	vmd->vmd_domain = domain;
462 	vmd->vmd_page_count = 0;
463 	vmd->vmd_free_count = 0;
464 	vmd->vmd_segs = 0;
465 	vmd->vmd_oom = FALSE;
466 	for (i = 0; i < PQ_COUNT; i++) {
467 		pq = &vmd->vmd_pagequeues[i];
468 		TAILQ_INIT(&pq->pq_pl);
469 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
470 		    MTX_DEF | MTX_DUPOK);
471 		pq->pq_pdpages = 0;
472 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
473 	}
474 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
475 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
476 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
477 
478 	/*
479 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
480 	 * insertions.
481 	 */
482 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
483 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
484 	    &vmd->vmd_inacthead, plinks.q);
485 
486 	/*
487 	 * The clock pages are used to implement active queue scanning without
488 	 * requeues.  Scans start at clock[0], which is advanced after the scan
489 	 * ends.  When the two clock hands meet, they are reset and scanning
490 	 * resumes from the head of the queue.
491 	 */
492 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
493 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
494 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
495 	    &vmd->vmd_clock[0], plinks.q);
496 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
497 	    &vmd->vmd_clock[1], plinks.q);
498 }
499 
500 /*
501  * Initialize a physical page in preparation for adding it to the free
502  * lists.
503  */
504 static void
505 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
506 {
507 
508 	m->object = NULL;
509 	m->ref_count = 0;
510 	m->busy_lock = VPB_UNBUSIED;
511 	m->flags = m->a.flags = 0;
512 	m->phys_addr = pa;
513 	m->a.queue = PQ_NONE;
514 	m->psind = 0;
515 	m->segind = segind;
516 	m->order = VM_NFREEORDER;
517 	m->pool = VM_FREEPOOL_DEFAULT;
518 	m->valid = m->dirty = 0;
519 	pmap_page_init(m);
520 }
521 
522 #ifndef PMAP_HAS_PAGE_ARRAY
523 static vm_paddr_t
524 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
525 {
526 	vm_paddr_t new_end;
527 
528 	/*
529 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
530 	 * However, because this page is allocated from KVM, out-of-bounds
531 	 * accesses using the direct map will not be trapped.
532 	 */
533 	*vaddr += PAGE_SIZE;
534 
535 	/*
536 	 * Allocate physical memory for the page structures, and map it.
537 	 */
538 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
539 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
540 	    VM_PROT_READ | VM_PROT_WRITE);
541 	vm_page_array_size = page_range;
542 
543 	return (new_end);
544 }
545 #endif
546 
547 /*
548  *	vm_page_startup:
549  *
550  *	Initializes the resident memory module.  Allocates physical memory for
551  *	bootstrapping UMA and some data structures that are used to manage
552  *	physical pages.  Initializes these structures, and populates the free
553  *	page queues.
554  */
555 vm_offset_t
556 vm_page_startup(vm_offset_t vaddr)
557 {
558 	struct vm_phys_seg *seg;
559 	vm_page_t m;
560 	char *list, *listend;
561 	vm_offset_t mapped;
562 	vm_paddr_t end, high_avail, low_avail, new_end, size;
563 	vm_paddr_t page_range __unused;
564 	vm_paddr_t last_pa, pa;
565 	u_long pagecount;
566 	int biggestone, i, segind;
567 #ifdef WITNESS
568 	int witness_size;
569 #endif
570 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
571 	long ii;
572 #endif
573 
574 	vaddr = round_page(vaddr);
575 
576 	vm_phys_early_startup();
577 	biggestone = vm_phys_avail_largest();
578 	end = phys_avail[biggestone+1];
579 
580 	/*
581 	 * Initialize the page and queue locks.
582 	 */
583 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
584 	for (i = 0; i < PA_LOCK_COUNT; i++)
585 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
586 	for (i = 0; i < vm_ndomains; i++)
587 		vm_page_domain_init(i);
588 
589 	/*
590 	 * Allocate memory for use when boot strapping the kernel memory
591 	 * allocator.  Tell UMA how many zones we are going to create
592 	 * before going fully functional.  UMA will add its zones.
593 	 *
594 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
595 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
596 	 */
597 	boot_pages = uma_startup_count(8);
598 
599 #ifndef UMA_MD_SMALL_ALLOC
600 	/* vmem_startup() calls uma_prealloc(). */
601 	boot_pages += vmem_startup_count();
602 	/* vm_map_startup() calls uma_prealloc(). */
603 	boot_pages += howmany(MAX_KMAP,
604 	    slab_ipers(sizeof(struct vm_map), UMA_ALIGN_PTR));
605 
606 	/*
607 	 * Before going fully functional kmem_init() does allocation
608 	 * from "KMAP ENTRY" and vmem_create() does allocation from "vmem".
609 	 */
610 	boot_pages += 2;
611 #endif
612 	/*
613 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
614 	 * manually fetch the value.
615 	 */
616 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
617 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
618 	new_end = trunc_page(new_end);
619 	mapped = pmap_map(&vaddr, new_end, end,
620 	    VM_PROT_READ | VM_PROT_WRITE);
621 	bzero((void *)mapped, end - new_end);
622 	uma_startup((void *)mapped, boot_pages);
623 
624 #ifdef WITNESS
625 	witness_size = round_page(witness_startup_count());
626 	new_end -= witness_size;
627 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
628 	    VM_PROT_READ | VM_PROT_WRITE);
629 	bzero((void *)mapped, witness_size);
630 	witness_startup((void *)mapped);
631 #endif
632 
633 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
634     defined(__i386__) || defined(__mips__) || defined(__riscv) || \
635     defined(__powerpc64__)
636 	/*
637 	 * Allocate a bitmap to indicate that a random physical page
638 	 * needs to be included in a minidump.
639 	 *
640 	 * The amd64 port needs this to indicate which direct map pages
641 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
642 	 *
643 	 * However, i386 still needs this workspace internally within the
644 	 * minidump code.  In theory, they are not needed on i386, but are
645 	 * included should the sf_buf code decide to use them.
646 	 */
647 	last_pa = 0;
648 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
649 		if (dump_avail[i + 1] > last_pa)
650 			last_pa = dump_avail[i + 1];
651 	page_range = last_pa / PAGE_SIZE;
652 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
653 	new_end -= vm_page_dump_size;
654 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
655 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
656 	bzero((void *)vm_page_dump, vm_page_dump_size);
657 #else
658 	(void)last_pa;
659 #endif
660 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
661     defined(__riscv) || defined(__powerpc64__)
662 	/*
663 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
664 	 * in a crash dump.  When pmap_map() uses the direct map, they are
665 	 * not automatically included.
666 	 */
667 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
668 		dump_add_page(pa);
669 #endif
670 	phys_avail[biggestone + 1] = new_end;
671 #ifdef __amd64__
672 	/*
673 	 * Request that the physical pages underlying the message buffer be
674 	 * included in a crash dump.  Since the message buffer is accessed
675 	 * through the direct map, they are not automatically included.
676 	 */
677 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
678 	last_pa = pa + round_page(msgbufsize);
679 	while (pa < last_pa) {
680 		dump_add_page(pa);
681 		pa += PAGE_SIZE;
682 	}
683 #endif
684 	/*
685 	 * Compute the number of pages of memory that will be available for
686 	 * use, taking into account the overhead of a page structure per page.
687 	 * In other words, solve
688 	 *	"available physical memory" - round_page(page_range *
689 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
690 	 * for page_range.
691 	 */
692 	low_avail = phys_avail[0];
693 	high_avail = phys_avail[1];
694 	for (i = 0; i < vm_phys_nsegs; i++) {
695 		if (vm_phys_segs[i].start < low_avail)
696 			low_avail = vm_phys_segs[i].start;
697 		if (vm_phys_segs[i].end > high_avail)
698 			high_avail = vm_phys_segs[i].end;
699 	}
700 	/* Skip the first chunk.  It is already accounted for. */
701 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
702 		if (phys_avail[i] < low_avail)
703 			low_avail = phys_avail[i];
704 		if (phys_avail[i + 1] > high_avail)
705 			high_avail = phys_avail[i + 1];
706 	}
707 	first_page = low_avail / PAGE_SIZE;
708 #ifdef VM_PHYSSEG_SPARSE
709 	size = 0;
710 	for (i = 0; i < vm_phys_nsegs; i++)
711 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
712 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
713 		size += phys_avail[i + 1] - phys_avail[i];
714 #elif defined(VM_PHYSSEG_DENSE)
715 	size = high_avail - low_avail;
716 #else
717 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
718 #endif
719 
720 #ifdef PMAP_HAS_PAGE_ARRAY
721 	pmap_page_array_startup(size / PAGE_SIZE);
722 	biggestone = vm_phys_avail_largest();
723 	end = new_end = phys_avail[biggestone + 1];
724 #else
725 #ifdef VM_PHYSSEG_DENSE
726 	/*
727 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
728 	 * the overhead of a page structure per page only if vm_page_array is
729 	 * allocated from the last physical memory chunk.  Otherwise, we must
730 	 * allocate page structures representing the physical memory
731 	 * underlying vm_page_array, even though they will not be used.
732 	 */
733 	if (new_end != high_avail)
734 		page_range = size / PAGE_SIZE;
735 	else
736 #endif
737 	{
738 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
739 
740 		/*
741 		 * If the partial bytes remaining are large enough for
742 		 * a page (PAGE_SIZE) without a corresponding
743 		 * 'struct vm_page', then new_end will contain an
744 		 * extra page after subtracting the length of the VM
745 		 * page array.  Compensate by subtracting an extra
746 		 * page from new_end.
747 		 */
748 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
749 			if (new_end == high_avail)
750 				high_avail -= PAGE_SIZE;
751 			new_end -= PAGE_SIZE;
752 		}
753 	}
754 	end = new_end;
755 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
756 #endif
757 
758 #if VM_NRESERVLEVEL > 0
759 	/*
760 	 * Allocate physical memory for the reservation management system's
761 	 * data structures, and map it.
762 	 */
763 	new_end = vm_reserv_startup(&vaddr, new_end);
764 #endif
765 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
766     defined(__riscv) || defined(__powerpc64__)
767 	/*
768 	 * Include vm_page_array and vm_reserv_array in a crash dump.
769 	 */
770 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
771 		dump_add_page(pa);
772 #endif
773 	phys_avail[biggestone + 1] = new_end;
774 
775 	/*
776 	 * Add physical memory segments corresponding to the available
777 	 * physical pages.
778 	 */
779 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
780 		if (vm_phys_avail_size(i) != 0)
781 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
782 
783 	/*
784 	 * Initialize the physical memory allocator.
785 	 */
786 	vm_phys_init();
787 
788 	/*
789 	 * Initialize the page structures and add every available page to the
790 	 * physical memory allocator's free lists.
791 	 */
792 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
793 	for (ii = 0; ii < vm_page_array_size; ii++) {
794 		m = &vm_page_array[ii];
795 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
796 		m->flags = PG_FICTITIOUS;
797 	}
798 #endif
799 	vm_cnt.v_page_count = 0;
800 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
801 		seg = &vm_phys_segs[segind];
802 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
803 		    m++, pa += PAGE_SIZE)
804 			vm_page_init_page(m, pa, segind);
805 
806 		/*
807 		 * Add the segment to the free lists only if it is covered by
808 		 * one of the ranges in phys_avail.  Because we've added the
809 		 * ranges to the vm_phys_segs array, we can assume that each
810 		 * segment is either entirely contained in one of the ranges,
811 		 * or doesn't overlap any of them.
812 		 */
813 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
814 			struct vm_domain *vmd;
815 
816 			if (seg->start < phys_avail[i] ||
817 			    seg->end > phys_avail[i + 1])
818 				continue;
819 
820 			m = seg->first_page;
821 			pagecount = (u_long)atop(seg->end - seg->start);
822 
823 			vmd = VM_DOMAIN(seg->domain);
824 			vm_domain_free_lock(vmd);
825 			vm_phys_enqueue_contig(m, pagecount);
826 			vm_domain_free_unlock(vmd);
827 			vm_domain_freecnt_inc(vmd, pagecount);
828 			vm_cnt.v_page_count += (u_int)pagecount;
829 
830 			vmd = VM_DOMAIN(seg->domain);
831 			vmd->vmd_page_count += (u_int)pagecount;
832 			vmd->vmd_segs |= 1UL << m->segind;
833 			break;
834 		}
835 	}
836 
837 	/*
838 	 * Remove blacklisted pages from the physical memory allocator.
839 	 */
840 	TAILQ_INIT(&blacklist_head);
841 	vm_page_blacklist_load(&list, &listend);
842 	vm_page_blacklist_check(list, listend);
843 
844 	list = kern_getenv("vm.blacklist");
845 	vm_page_blacklist_check(list, NULL);
846 
847 	freeenv(list);
848 #if VM_NRESERVLEVEL > 0
849 	/*
850 	 * Initialize the reservation management system.
851 	 */
852 	vm_reserv_init();
853 #endif
854 
855 	return (vaddr);
856 }
857 
858 void
859 vm_page_reference(vm_page_t m)
860 {
861 
862 	vm_page_aflag_set(m, PGA_REFERENCED);
863 }
864 
865 static bool
866 vm_page_acquire_flags(vm_page_t m, int allocflags)
867 {
868 	bool locked;
869 
870 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
871 		locked = vm_page_trysbusy(m);
872 	else
873 		locked = vm_page_tryxbusy(m);
874 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
875 		vm_page_wire(m);
876 	return (locked);
877 }
878 
879 static bool
880 vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wchan,
881     int allocflags)
882 {
883 
884 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
885 		return (false);
886 	/*
887 	 * Reference the page before unlocking and
888 	 * sleeping so that the page daemon is less
889 	 * likely to reclaim it.
890 	 */
891 	if ((allocflags & VM_ALLOC_NOCREAT) == 0)
892 		vm_page_aflag_set(m, PGA_REFERENCED);
893 	vm_page_busy_sleep(m, wchan, (allocflags &
894 	    VM_ALLOC_IGN_SBUSY) != 0);
895 	VM_OBJECT_WLOCK(object);
896 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
897 		return (false);
898 	return (true);
899 }
900 
901 /*
902  *	vm_page_busy_acquire:
903  *
904  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
905  *	and drop the object lock if necessary.
906  */
907 bool
908 vm_page_busy_acquire(vm_page_t m, int allocflags)
909 {
910 	vm_object_t obj;
911 	bool locked;
912 
913 	/*
914 	 * The page-specific object must be cached because page
915 	 * identity can change during the sleep, causing the
916 	 * re-lock of a different object.
917 	 * It is assumed that a reference to the object is already
918 	 * held by the callers.
919 	 */
920 	obj = m->object;
921 	for (;;) {
922 		if (vm_page_acquire_flags(m, allocflags))
923 			return (true);
924 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
925 			return (false);
926 		if (obj != NULL)
927 			locked = VM_OBJECT_WOWNED(obj);
928 		else
929 			locked = false;
930 		MPASS(locked || vm_page_wired(m));
931 		_vm_page_busy_sleep(obj, m, "vmpba",
932 		    (allocflags & VM_ALLOC_SBUSY) != 0, locked);
933 		if (locked)
934 			VM_OBJECT_WLOCK(obj);
935 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
936 			return (false);
937 		KASSERT(m->object == obj || m->object == NULL,
938 		    ("vm_page_busy_acquire: page %p does not belong to %p",
939 		    m, obj));
940 	}
941 }
942 
943 /*
944  *	vm_page_busy_downgrade:
945  *
946  *	Downgrade an exclusive busy page into a single shared busy page.
947  */
948 void
949 vm_page_busy_downgrade(vm_page_t m)
950 {
951 	u_int x;
952 
953 	vm_page_assert_xbusied(m);
954 
955 	x = m->busy_lock;
956 	for (;;) {
957 		if (atomic_fcmpset_rel_int(&m->busy_lock,
958 		    &x, VPB_SHARERS_WORD(1)))
959 			break;
960 	}
961 	if ((x & VPB_BIT_WAITERS) != 0)
962 		wakeup(m);
963 }
964 
965 /*
966  *
967  *	vm_page_busy_tryupgrade:
968  *
969  *	Attempt to upgrade a single shared busy into an exclusive busy.
970  */
971 int
972 vm_page_busy_tryupgrade(vm_page_t m)
973 {
974 	u_int ce, x;
975 
976 	vm_page_assert_sbusied(m);
977 
978 	x = m->busy_lock;
979 	ce = VPB_CURTHREAD_EXCLUSIVE;
980 	for (;;) {
981 		if (VPB_SHARERS(x) > 1)
982 			return (0);
983 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
984 		    ("vm_page_busy_tryupgrade: invalid lock state"));
985 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
986 		    ce | (x & VPB_BIT_WAITERS)))
987 			continue;
988 		return (1);
989 	}
990 }
991 
992 /*
993  *	vm_page_sbusied:
994  *
995  *	Return a positive value if the page is shared busied, 0 otherwise.
996  */
997 int
998 vm_page_sbusied(vm_page_t m)
999 {
1000 	u_int x;
1001 
1002 	x = m->busy_lock;
1003 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1004 }
1005 
1006 /*
1007  *	vm_page_sunbusy:
1008  *
1009  *	Shared unbusy a page.
1010  */
1011 void
1012 vm_page_sunbusy(vm_page_t m)
1013 {
1014 	u_int x;
1015 
1016 	vm_page_assert_sbusied(m);
1017 
1018 	x = m->busy_lock;
1019 	for (;;) {
1020 		if (VPB_SHARERS(x) > 1) {
1021 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1022 			    x - VPB_ONE_SHARER))
1023 				break;
1024 			continue;
1025 		}
1026 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1027 		    ("vm_page_sunbusy: invalid lock state"));
1028 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1029 			continue;
1030 		if ((x & VPB_BIT_WAITERS) == 0)
1031 			break;
1032 		wakeup(m);
1033 		break;
1034 	}
1035 }
1036 
1037 /*
1038  *	vm_page_busy_sleep:
1039  *
1040  *	Sleep if the page is busy, using the page pointer as wchan.
1041  *	This is used to implement the hard-path of busying mechanism.
1042  *
1043  *	If nonshared is true, sleep only if the page is xbusy.
1044  *
1045  *	The object lock must be held on entry and will be released on exit.
1046  */
1047 void
1048 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1049 {
1050 	vm_object_t obj;
1051 
1052 	obj = m->object;
1053 	VM_OBJECT_ASSERT_LOCKED(obj);
1054 	vm_page_lock_assert(m, MA_NOTOWNED);
1055 
1056 	_vm_page_busy_sleep(obj, m, wmesg, nonshared, true);
1057 }
1058 
1059 static void
1060 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg,
1061     bool nonshared, bool locked)
1062 {
1063 	u_int x;
1064 
1065 	/*
1066 	 * If the object is busy we must wait for that to drain to zero
1067 	 * before trying the page again.
1068 	 */
1069 	if (obj != NULL && vm_object_busied(obj)) {
1070 		if (locked)
1071 			VM_OBJECT_DROP(obj);
1072 		vm_object_busy_wait(obj, wmesg);
1073 		return;
1074 	}
1075 	sleepq_lock(m);
1076 	x = m->busy_lock;
1077 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
1078 	    ((x & VPB_BIT_WAITERS) == 0 &&
1079 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
1080 		if (locked)
1081 			VM_OBJECT_DROP(obj);
1082 		sleepq_release(m);
1083 		return;
1084 	}
1085 	if (locked)
1086 		VM_OBJECT_DROP(obj);
1087 	DROP_GIANT();
1088 	sleepq_add(m, NULL, wmesg, 0, 0);
1089 	sleepq_wait(m, PVM);
1090 	PICKUP_GIANT();
1091 }
1092 
1093 /*
1094  *	vm_page_trysbusy:
1095  *
1096  *	Try to shared busy a page.
1097  *	If the operation succeeds 1 is returned otherwise 0.
1098  *	The operation never sleeps.
1099  */
1100 int
1101 vm_page_trysbusy(vm_page_t m)
1102 {
1103 	vm_object_t obj;
1104 	u_int x;
1105 
1106 	obj = m->object;
1107 	x = m->busy_lock;
1108 	for (;;) {
1109 		if ((x & VPB_BIT_SHARED) == 0)
1110 			return (0);
1111 		/*
1112 		 * Reduce the window for transient busies that will trigger
1113 		 * false negatives in vm_page_ps_test().
1114 		 */
1115 		if (obj != NULL && vm_object_busied(obj))
1116 			return (0);
1117 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1118 		    x + VPB_ONE_SHARER))
1119 			break;
1120 	}
1121 
1122 	/* Refetch the object now that we're guaranteed that it is stable. */
1123 	obj = m->object;
1124 	if (obj != NULL && vm_object_busied(obj)) {
1125 		vm_page_sunbusy(m);
1126 		return (0);
1127 	}
1128 	return (1);
1129 }
1130 
1131 /*
1132  *	vm_page_tryxbusy:
1133  *
1134  *	Try to exclusive busy a page.
1135  *	If the operation succeeds 1 is returned otherwise 0.
1136  *	The operation never sleeps.
1137  */
1138 int
1139 vm_page_tryxbusy(vm_page_t m)
1140 {
1141 	vm_object_t obj;
1142 
1143         if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED,
1144             VPB_CURTHREAD_EXCLUSIVE) == 0)
1145 		return (0);
1146 
1147 	obj = m->object;
1148 	if (obj != NULL && vm_object_busied(obj)) {
1149 		vm_page_xunbusy(m);
1150 		return (0);
1151 	}
1152 	return (1);
1153 }
1154 
1155 static void
1156 vm_page_xunbusy_hard_tail(vm_page_t m)
1157 {
1158 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1159 	/* Wake the waiter. */
1160 	wakeup(m);
1161 }
1162 
1163 /*
1164  *	vm_page_xunbusy_hard:
1165  *
1166  *	Called when unbusy has failed because there is a waiter.
1167  */
1168 void
1169 vm_page_xunbusy_hard(vm_page_t m)
1170 {
1171 	vm_page_assert_xbusied(m);
1172 	vm_page_xunbusy_hard_tail(m);
1173 }
1174 
1175 void
1176 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1177 {
1178 	vm_page_assert_xbusied_unchecked(m);
1179 	vm_page_xunbusy_hard_tail(m);
1180 }
1181 
1182 /*
1183  * Avoid releasing and reacquiring the same page lock.
1184  */
1185 void
1186 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1187 {
1188 	struct mtx *mtx1;
1189 
1190 	mtx1 = vm_page_lockptr(m);
1191 	if (*mtx == mtx1)
1192 		return;
1193 	if (*mtx != NULL)
1194 		mtx_unlock(*mtx);
1195 	*mtx = mtx1;
1196 	mtx_lock(mtx1);
1197 }
1198 
1199 /*
1200  *	vm_page_unhold_pages:
1201  *
1202  *	Unhold each of the pages that is referenced by the given array.
1203  */
1204 void
1205 vm_page_unhold_pages(vm_page_t *ma, int count)
1206 {
1207 
1208 	for (; count != 0; count--) {
1209 		vm_page_unwire(*ma, PQ_ACTIVE);
1210 		ma++;
1211 	}
1212 }
1213 
1214 vm_page_t
1215 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1216 {
1217 	vm_page_t m;
1218 
1219 #ifdef VM_PHYSSEG_SPARSE
1220 	m = vm_phys_paddr_to_vm_page(pa);
1221 	if (m == NULL)
1222 		m = vm_phys_fictitious_to_vm_page(pa);
1223 	return (m);
1224 #elif defined(VM_PHYSSEG_DENSE)
1225 	long pi;
1226 
1227 	pi = atop(pa);
1228 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1229 		m = &vm_page_array[pi - first_page];
1230 		return (m);
1231 	}
1232 	return (vm_phys_fictitious_to_vm_page(pa));
1233 #else
1234 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1235 #endif
1236 }
1237 
1238 /*
1239  *	vm_page_getfake:
1240  *
1241  *	Create a fictitious page with the specified physical address and
1242  *	memory attribute.  The memory attribute is the only the machine-
1243  *	dependent aspect of a fictitious page that must be initialized.
1244  */
1245 vm_page_t
1246 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1247 {
1248 	vm_page_t m;
1249 
1250 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1251 	vm_page_initfake(m, paddr, memattr);
1252 	return (m);
1253 }
1254 
1255 void
1256 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1257 {
1258 
1259 	if ((m->flags & PG_FICTITIOUS) != 0) {
1260 		/*
1261 		 * The page's memattr might have changed since the
1262 		 * previous initialization.  Update the pmap to the
1263 		 * new memattr.
1264 		 */
1265 		goto memattr;
1266 	}
1267 	m->phys_addr = paddr;
1268 	m->a.queue = PQ_NONE;
1269 	/* Fictitious pages don't use "segind". */
1270 	m->flags = PG_FICTITIOUS;
1271 	/* Fictitious pages don't use "order" or "pool". */
1272 	m->oflags = VPO_UNMANAGED;
1273 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1274 	/* Fictitious pages are unevictable. */
1275 	m->ref_count = 1;
1276 	pmap_page_init(m);
1277 memattr:
1278 	pmap_page_set_memattr(m, memattr);
1279 }
1280 
1281 /*
1282  *	vm_page_putfake:
1283  *
1284  *	Release a fictitious page.
1285  */
1286 void
1287 vm_page_putfake(vm_page_t m)
1288 {
1289 
1290 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1291 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1292 	    ("vm_page_putfake: bad page %p", m));
1293 	if (vm_page_xbusied(m))
1294 		vm_page_xunbusy(m);
1295 	uma_zfree(fakepg_zone, m);
1296 }
1297 
1298 /*
1299  *	vm_page_updatefake:
1300  *
1301  *	Update the given fictitious page to the specified physical address and
1302  *	memory attribute.
1303  */
1304 void
1305 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1306 {
1307 
1308 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1309 	    ("vm_page_updatefake: bad page %p", m));
1310 	m->phys_addr = paddr;
1311 	pmap_page_set_memattr(m, memattr);
1312 }
1313 
1314 /*
1315  *	vm_page_free:
1316  *
1317  *	Free a page.
1318  */
1319 void
1320 vm_page_free(vm_page_t m)
1321 {
1322 
1323 	m->flags &= ~PG_ZERO;
1324 	vm_page_free_toq(m);
1325 }
1326 
1327 /*
1328  *	vm_page_free_zero:
1329  *
1330  *	Free a page to the zerod-pages queue
1331  */
1332 void
1333 vm_page_free_zero(vm_page_t m)
1334 {
1335 
1336 	m->flags |= PG_ZERO;
1337 	vm_page_free_toq(m);
1338 }
1339 
1340 /*
1341  * Unbusy and handle the page queueing for a page from a getpages request that
1342  * was optionally read ahead or behind.
1343  */
1344 void
1345 vm_page_readahead_finish(vm_page_t m)
1346 {
1347 
1348 	/* We shouldn't put invalid pages on queues. */
1349 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1350 
1351 	/*
1352 	 * Since the page is not the actually needed one, whether it should
1353 	 * be activated or deactivated is not obvious.  Empirical results
1354 	 * have shown that deactivating the page is usually the best choice,
1355 	 * unless the page is wanted by another thread.
1356 	 */
1357 	vm_page_lock(m);
1358 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1359 		vm_page_activate(m);
1360 	else
1361 		vm_page_deactivate(m);
1362 	vm_page_unlock(m);
1363 	vm_page_xunbusy_unchecked(m);
1364 }
1365 
1366 /*
1367  *	vm_page_sleep_if_busy:
1368  *
1369  *	Sleep and release the object lock if the page is busied.
1370  *	Returns TRUE if the thread slept.
1371  *
1372  *	The given page must be unlocked and object containing it must
1373  *	be locked.
1374  */
1375 int
1376 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1377 {
1378 	vm_object_t obj;
1379 
1380 	vm_page_lock_assert(m, MA_NOTOWNED);
1381 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1382 
1383 	/*
1384 	 * The page-specific object must be cached because page
1385 	 * identity can change during the sleep, causing the
1386 	 * re-lock of a different object.
1387 	 * It is assumed that a reference to the object is already
1388 	 * held by the callers.
1389 	 */
1390 	obj = m->object;
1391 	if (vm_page_busied(m) || (obj != NULL && obj->busy)) {
1392 		vm_page_busy_sleep(m, msg, false);
1393 		VM_OBJECT_WLOCK(obj);
1394 		return (TRUE);
1395 	}
1396 	return (FALSE);
1397 }
1398 
1399 /*
1400  *	vm_page_sleep_if_xbusy:
1401  *
1402  *	Sleep and release the object lock if the page is xbusied.
1403  *	Returns TRUE if the thread slept.
1404  *
1405  *	The given page must be unlocked and object containing it must
1406  *	be locked.
1407  */
1408 int
1409 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg)
1410 {
1411 	vm_object_t obj;
1412 
1413 	vm_page_lock_assert(m, MA_NOTOWNED);
1414 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1415 
1416 	/*
1417 	 * The page-specific object must be cached because page
1418 	 * identity can change during the sleep, causing the
1419 	 * re-lock of a different object.
1420 	 * It is assumed that a reference to the object is already
1421 	 * held by the callers.
1422 	 */
1423 	obj = m->object;
1424 	if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) {
1425 		vm_page_busy_sleep(m, msg, true);
1426 		VM_OBJECT_WLOCK(obj);
1427 		return (TRUE);
1428 	}
1429 	return (FALSE);
1430 }
1431 
1432 /*
1433  *	vm_page_dirty_KBI:		[ internal use only ]
1434  *
1435  *	Set all bits in the page's dirty field.
1436  *
1437  *	The object containing the specified page must be locked if the
1438  *	call is made from the machine-independent layer.
1439  *
1440  *	See vm_page_clear_dirty_mask().
1441  *
1442  *	This function should only be called by vm_page_dirty().
1443  */
1444 void
1445 vm_page_dirty_KBI(vm_page_t m)
1446 {
1447 
1448 	/* Refer to this operation by its public name. */
1449 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1450 	m->dirty = VM_PAGE_BITS_ALL;
1451 }
1452 
1453 /*
1454  *	vm_page_insert:		[ internal use only ]
1455  *
1456  *	Inserts the given mem entry into the object and object list.
1457  *
1458  *	The object must be locked.
1459  */
1460 int
1461 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1462 {
1463 	vm_page_t mpred;
1464 
1465 	VM_OBJECT_ASSERT_WLOCKED(object);
1466 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1467 	return (vm_page_insert_after(m, object, pindex, mpred));
1468 }
1469 
1470 /*
1471  *	vm_page_insert_after:
1472  *
1473  *	Inserts the page "m" into the specified object at offset "pindex".
1474  *
1475  *	The page "mpred" must immediately precede the offset "pindex" within
1476  *	the specified object.
1477  *
1478  *	The object must be locked.
1479  */
1480 static int
1481 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1482     vm_page_t mpred)
1483 {
1484 	vm_page_t msucc;
1485 
1486 	VM_OBJECT_ASSERT_WLOCKED(object);
1487 	KASSERT(m->object == NULL,
1488 	    ("vm_page_insert_after: page already inserted"));
1489 	if (mpred != NULL) {
1490 		KASSERT(mpred->object == object,
1491 		    ("vm_page_insert_after: object doesn't contain mpred"));
1492 		KASSERT(mpred->pindex < pindex,
1493 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1494 		msucc = TAILQ_NEXT(mpred, listq);
1495 	} else
1496 		msucc = TAILQ_FIRST(&object->memq);
1497 	if (msucc != NULL)
1498 		KASSERT(msucc->pindex > pindex,
1499 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1500 
1501 	/*
1502 	 * Record the object/offset pair in this page.
1503 	 */
1504 	m->object = object;
1505 	m->pindex = pindex;
1506 	m->ref_count |= VPRC_OBJREF;
1507 
1508 	/*
1509 	 * Now link into the object's ordered list of backed pages.
1510 	 */
1511 	if (vm_radix_insert(&object->rtree, m)) {
1512 		m->object = NULL;
1513 		m->pindex = 0;
1514 		m->ref_count &= ~VPRC_OBJREF;
1515 		return (1);
1516 	}
1517 	vm_page_insert_radixdone(m, object, mpred);
1518 	return (0);
1519 }
1520 
1521 /*
1522  *	vm_page_insert_radixdone:
1523  *
1524  *	Complete page "m" insertion into the specified object after the
1525  *	radix trie hooking.
1526  *
1527  *	The page "mpred" must precede the offset "m->pindex" within the
1528  *	specified object.
1529  *
1530  *	The object must be locked.
1531  */
1532 static void
1533 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1534 {
1535 
1536 	VM_OBJECT_ASSERT_WLOCKED(object);
1537 	KASSERT(object != NULL && m->object == object,
1538 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1539 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1540 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1541 	if (mpred != NULL) {
1542 		KASSERT(mpred->object == object,
1543 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1544 		KASSERT(mpred->pindex < m->pindex,
1545 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1546 	}
1547 
1548 	if (mpred != NULL)
1549 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1550 	else
1551 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1552 
1553 	/*
1554 	 * Show that the object has one more resident page.
1555 	 */
1556 	object->resident_page_count++;
1557 
1558 	/*
1559 	 * Hold the vnode until the last page is released.
1560 	 */
1561 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1562 		vhold(object->handle);
1563 
1564 	/*
1565 	 * Since we are inserting a new and possibly dirty page,
1566 	 * update the object's generation count.
1567 	 */
1568 	if (pmap_page_is_write_mapped(m))
1569 		vm_object_set_writeable_dirty(object);
1570 }
1571 
1572 /*
1573  * Do the work to remove a page from its object.  The caller is responsible for
1574  * updating the page's fields to reflect this removal.
1575  */
1576 static void
1577 vm_page_object_remove(vm_page_t m)
1578 {
1579 	vm_object_t object;
1580 	vm_page_t mrem;
1581 
1582 	object = m->object;
1583 	VM_OBJECT_ASSERT_WLOCKED(object);
1584 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1585 	    ("page %p is missing its object ref", m));
1586 
1587 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1588 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1589 
1590 	/*
1591 	 * Now remove from the object's list of backed pages.
1592 	 */
1593 	TAILQ_REMOVE(&object->memq, m, listq);
1594 
1595 	/*
1596 	 * And show that the object has one fewer resident page.
1597 	 */
1598 	object->resident_page_count--;
1599 
1600 	/*
1601 	 * The vnode may now be recycled.
1602 	 */
1603 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1604 		vdrop(object->handle);
1605 }
1606 
1607 /*
1608  *	vm_page_remove:
1609  *
1610  *	Removes the specified page from its containing object, but does not
1611  *	invalidate any backing storage.  Returns true if the object's reference
1612  *	was the last reference to the page, and false otherwise.
1613  *
1614  *	The object must be locked.
1615  */
1616 bool
1617 vm_page_remove(vm_page_t m)
1618 {
1619 
1620 	vm_page_object_remove(m);
1621 	m->object = NULL;
1622 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1623 }
1624 
1625 /*
1626  *	vm_page_lookup:
1627  *
1628  *	Returns the page associated with the object/offset
1629  *	pair specified; if none is found, NULL is returned.
1630  *
1631  *	The object must be locked.
1632  */
1633 vm_page_t
1634 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1635 {
1636 
1637 	VM_OBJECT_ASSERT_LOCKED(object);
1638 	return (vm_radix_lookup(&object->rtree, pindex));
1639 }
1640 
1641 /*
1642  *	vm_page_find_least:
1643  *
1644  *	Returns the page associated with the object with least pindex
1645  *	greater than or equal to the parameter pindex, or NULL.
1646  *
1647  *	The object must be locked.
1648  */
1649 vm_page_t
1650 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1651 {
1652 	vm_page_t m;
1653 
1654 	VM_OBJECT_ASSERT_LOCKED(object);
1655 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1656 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1657 	return (m);
1658 }
1659 
1660 /*
1661  * Returns the given page's successor (by pindex) within the object if it is
1662  * resident; if none is found, NULL is returned.
1663  *
1664  * The object must be locked.
1665  */
1666 vm_page_t
1667 vm_page_next(vm_page_t m)
1668 {
1669 	vm_page_t next;
1670 
1671 	VM_OBJECT_ASSERT_LOCKED(m->object);
1672 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1673 		MPASS(next->object == m->object);
1674 		if (next->pindex != m->pindex + 1)
1675 			next = NULL;
1676 	}
1677 	return (next);
1678 }
1679 
1680 /*
1681  * Returns the given page's predecessor (by pindex) within the object if it is
1682  * resident; if none is found, NULL is returned.
1683  *
1684  * The object must be locked.
1685  */
1686 vm_page_t
1687 vm_page_prev(vm_page_t m)
1688 {
1689 	vm_page_t prev;
1690 
1691 	VM_OBJECT_ASSERT_LOCKED(m->object);
1692 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1693 		MPASS(prev->object == m->object);
1694 		if (prev->pindex != m->pindex - 1)
1695 			prev = NULL;
1696 	}
1697 	return (prev);
1698 }
1699 
1700 /*
1701  * Uses the page mnew as a replacement for an existing page at index
1702  * pindex which must be already present in the object.
1703  */
1704 vm_page_t
1705 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1706 {
1707 	vm_page_t mold;
1708 
1709 	VM_OBJECT_ASSERT_WLOCKED(object);
1710 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1711 	    ("vm_page_replace: page %p already in object", mnew));
1712 
1713 	/*
1714 	 * This function mostly follows vm_page_insert() and
1715 	 * vm_page_remove() without the radix, object count and vnode
1716 	 * dance.  Double check such functions for more comments.
1717 	 */
1718 
1719 	mnew->object = object;
1720 	mnew->pindex = pindex;
1721 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1722 	mold = vm_radix_replace(&object->rtree, mnew);
1723 
1724 	/* Keep the resident page list in sorted order. */
1725 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1726 	TAILQ_REMOVE(&object->memq, mold, listq);
1727 
1728 	mold->object = NULL;
1729 	atomic_clear_int(&mold->ref_count, VPRC_OBJREF);
1730 	vm_page_xunbusy(mold);
1731 
1732 	/*
1733 	 * The object's resident_page_count does not change because we have
1734 	 * swapped one page for another, but the generation count should
1735 	 * change if the page is dirty.
1736 	 */
1737 	if (pmap_page_is_write_mapped(mnew))
1738 		vm_object_set_writeable_dirty(object);
1739 	return (mold);
1740 }
1741 
1742 /*
1743  *	vm_page_rename:
1744  *
1745  *	Move the given memory entry from its
1746  *	current object to the specified target object/offset.
1747  *
1748  *	Note: swap associated with the page must be invalidated by the move.  We
1749  *	      have to do this for several reasons:  (1) we aren't freeing the
1750  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1751  *	      moving the page from object A to B, and will then later move
1752  *	      the backing store from A to B and we can't have a conflict.
1753  *
1754  *	Note: we *always* dirty the page.  It is necessary both for the
1755  *	      fact that we moved it, and because we may be invalidating
1756  *	      swap.
1757  *
1758  *	The objects must be locked.
1759  */
1760 int
1761 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1762 {
1763 	vm_page_t mpred;
1764 	vm_pindex_t opidx;
1765 
1766 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1767 
1768 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1769 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1770 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1771 	    ("vm_page_rename: pindex already renamed"));
1772 
1773 	/*
1774 	 * Create a custom version of vm_page_insert() which does not depend
1775 	 * by m_prev and can cheat on the implementation aspects of the
1776 	 * function.
1777 	 */
1778 	opidx = m->pindex;
1779 	m->pindex = new_pindex;
1780 	if (vm_radix_insert(&new_object->rtree, m)) {
1781 		m->pindex = opidx;
1782 		return (1);
1783 	}
1784 
1785 	/*
1786 	 * The operation cannot fail anymore.  The removal must happen before
1787 	 * the listq iterator is tainted.
1788 	 */
1789 	m->pindex = opidx;
1790 	vm_page_object_remove(m);
1791 
1792 	/* Return back to the new pindex to complete vm_page_insert(). */
1793 	m->pindex = new_pindex;
1794 	m->object = new_object;
1795 
1796 	vm_page_insert_radixdone(m, new_object, mpred);
1797 	vm_page_dirty(m);
1798 	return (0);
1799 }
1800 
1801 /*
1802  *	vm_page_alloc:
1803  *
1804  *	Allocate and return a page that is associated with the specified
1805  *	object and offset pair.  By default, this page is exclusive busied.
1806  *
1807  *	The caller must always specify an allocation class.
1808  *
1809  *	allocation classes:
1810  *	VM_ALLOC_NORMAL		normal process request
1811  *	VM_ALLOC_SYSTEM		system *really* needs a page
1812  *	VM_ALLOC_INTERRUPT	interrupt time request
1813  *
1814  *	optional allocation flags:
1815  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1816  *				intends to allocate
1817  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1818  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1819  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1820  *				should not be exclusive busy
1821  *	VM_ALLOC_SBUSY		shared busy the allocated page
1822  *	VM_ALLOC_WIRED		wire the allocated page
1823  *	VM_ALLOC_ZERO		prefer a zeroed page
1824  */
1825 vm_page_t
1826 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1827 {
1828 
1829 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1830 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1831 }
1832 
1833 vm_page_t
1834 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1835     int req)
1836 {
1837 
1838 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1839 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1840 	    NULL));
1841 }
1842 
1843 /*
1844  * Allocate a page in the specified object with the given page index.  To
1845  * optimize insertion of the page into the object, the caller must also specifiy
1846  * the resident page in the object with largest index smaller than the given
1847  * page index, or NULL if no such page exists.
1848  */
1849 vm_page_t
1850 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1851     int req, vm_page_t mpred)
1852 {
1853 	struct vm_domainset_iter di;
1854 	vm_page_t m;
1855 	int domain;
1856 
1857 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1858 	do {
1859 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1860 		    mpred);
1861 		if (m != NULL)
1862 			break;
1863 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1864 
1865 	return (m);
1866 }
1867 
1868 /*
1869  * Returns true if the number of free pages exceeds the minimum
1870  * for the request class and false otherwise.
1871  */
1872 static int
1873 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1874 {
1875 	u_int limit, old, new;
1876 
1877 	if (req_class == VM_ALLOC_INTERRUPT)
1878 		limit = 0;
1879 	else if (req_class == VM_ALLOC_SYSTEM)
1880 		limit = vmd->vmd_interrupt_free_min;
1881 	else
1882 		limit = vmd->vmd_free_reserved;
1883 
1884 	/*
1885 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1886 	 */
1887 	limit += npages;
1888 	old = vmd->vmd_free_count;
1889 	do {
1890 		if (old < limit)
1891 			return (0);
1892 		new = old - npages;
1893 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1894 
1895 	/* Wake the page daemon if we've crossed the threshold. */
1896 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1897 		pagedaemon_wakeup(vmd->vmd_domain);
1898 
1899 	/* Only update bitsets on transitions. */
1900 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1901 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1902 		vm_domain_set(vmd);
1903 
1904 	return (1);
1905 }
1906 
1907 int
1908 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1909 {
1910 	int req_class;
1911 
1912 	/*
1913 	 * The page daemon is allowed to dig deeper into the free page list.
1914 	 */
1915 	req_class = req & VM_ALLOC_CLASS_MASK;
1916 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1917 		req_class = VM_ALLOC_SYSTEM;
1918 	return (_vm_domain_allocate(vmd, req_class, npages));
1919 }
1920 
1921 vm_page_t
1922 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1923     int req, vm_page_t mpred)
1924 {
1925 	struct vm_domain *vmd;
1926 	vm_page_t m;
1927 	int flags, pool;
1928 
1929 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1930 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1931 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1932 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1933 	    ("inconsistent object(%p)/req(%x)", object, req));
1934 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1935 	    ("Can't sleep and retry object insertion."));
1936 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1937 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1938 	    (uintmax_t)pindex));
1939 	if (object != NULL)
1940 		VM_OBJECT_ASSERT_WLOCKED(object);
1941 
1942 	flags = 0;
1943 	m = NULL;
1944 	pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
1945 again:
1946 #if VM_NRESERVLEVEL > 0
1947 	/*
1948 	 * Can we allocate the page from a reservation?
1949 	 */
1950 	if (vm_object_reserv(object) &&
1951 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
1952 	    NULL) {
1953 		domain = vm_phys_domain(m);
1954 		vmd = VM_DOMAIN(domain);
1955 		goto found;
1956 	}
1957 #endif
1958 	vmd = VM_DOMAIN(domain);
1959 	if (vmd->vmd_pgcache[pool].zone != NULL) {
1960 		m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
1961 		if (m != NULL) {
1962 			flags |= PG_PCPU_CACHE;
1963 			goto found;
1964 		}
1965 	}
1966 	if (vm_domain_allocate(vmd, req, 1)) {
1967 		/*
1968 		 * If not, allocate it from the free page queues.
1969 		 */
1970 		vm_domain_free_lock(vmd);
1971 		m = vm_phys_alloc_pages(domain, pool, 0);
1972 		vm_domain_free_unlock(vmd);
1973 		if (m == NULL) {
1974 			vm_domain_freecnt_inc(vmd, 1);
1975 #if VM_NRESERVLEVEL > 0
1976 			if (vm_reserv_reclaim_inactive(domain))
1977 				goto again;
1978 #endif
1979 		}
1980 	}
1981 	if (m == NULL) {
1982 		/*
1983 		 * Not allocatable, give up.
1984 		 */
1985 		if (vm_domain_alloc_fail(vmd, object, req))
1986 			goto again;
1987 		return (NULL);
1988 	}
1989 
1990 	/*
1991 	 * At this point we had better have found a good page.
1992 	 */
1993 found:
1994 	vm_page_dequeue(m);
1995 	vm_page_alloc_check(m);
1996 
1997 	/*
1998 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
1999 	 */
2000 	if ((req & VM_ALLOC_ZERO) != 0)
2001 		flags |= (m->flags & PG_ZERO);
2002 	if ((req & VM_ALLOC_NODUMP) != 0)
2003 		flags |= PG_NODUMP;
2004 	m->flags = flags;
2005 	m->a.flags = 0;
2006 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2007 	    VPO_UNMANAGED : 0;
2008 	m->busy_lock = VPB_UNBUSIED;
2009 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2010 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2011 	if ((req & VM_ALLOC_SBUSY) != 0)
2012 		m->busy_lock = VPB_SHARERS_WORD(1);
2013 	if (req & VM_ALLOC_WIRED) {
2014 		/*
2015 		 * The page lock is not required for wiring a page until that
2016 		 * page is inserted into the object.
2017 		 */
2018 		vm_wire_add(1);
2019 		m->ref_count = 1;
2020 	}
2021 	m->a.act_count = 0;
2022 
2023 	if (object != NULL) {
2024 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2025 			if (req & VM_ALLOC_WIRED) {
2026 				vm_wire_sub(1);
2027 				m->ref_count = 0;
2028 			}
2029 			KASSERT(m->object == NULL, ("page %p has object", m));
2030 			m->oflags = VPO_UNMANAGED;
2031 			m->busy_lock = VPB_UNBUSIED;
2032 			/* Don't change PG_ZERO. */
2033 			vm_page_free_toq(m);
2034 			if (req & VM_ALLOC_WAITFAIL) {
2035 				VM_OBJECT_WUNLOCK(object);
2036 				vm_radix_wait();
2037 				VM_OBJECT_WLOCK(object);
2038 			}
2039 			return (NULL);
2040 		}
2041 
2042 		/* Ignore device objects; the pager sets "memattr" for them. */
2043 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2044 		    (object->flags & OBJ_FICTITIOUS) == 0)
2045 			pmap_page_set_memattr(m, object->memattr);
2046 	} else
2047 		m->pindex = pindex;
2048 
2049 	return (m);
2050 }
2051 
2052 /*
2053  *	vm_page_alloc_contig:
2054  *
2055  *	Allocate a contiguous set of physical pages of the given size "npages"
2056  *	from the free lists.  All of the physical pages must be at or above
2057  *	the given physical address "low" and below the given physical address
2058  *	"high".  The given value "alignment" determines the alignment of the
2059  *	first physical page in the set.  If the given value "boundary" is
2060  *	non-zero, then the set of physical pages cannot cross any physical
2061  *	address boundary that is a multiple of that value.  Both "alignment"
2062  *	and "boundary" must be a power of two.
2063  *
2064  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2065  *	then the memory attribute setting for the physical pages is configured
2066  *	to the object's memory attribute setting.  Otherwise, the memory
2067  *	attribute setting for the physical pages is configured to "memattr",
2068  *	overriding the object's memory attribute setting.  However, if the
2069  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2070  *	memory attribute setting for the physical pages cannot be configured
2071  *	to VM_MEMATTR_DEFAULT.
2072  *
2073  *	The specified object may not contain fictitious pages.
2074  *
2075  *	The caller must always specify an allocation class.
2076  *
2077  *	allocation classes:
2078  *	VM_ALLOC_NORMAL		normal process request
2079  *	VM_ALLOC_SYSTEM		system *really* needs a page
2080  *	VM_ALLOC_INTERRUPT	interrupt time request
2081  *
2082  *	optional allocation flags:
2083  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2084  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2085  *	VM_ALLOC_NOOBJ		page is not associated with an object and
2086  *				should not be exclusive busy
2087  *	VM_ALLOC_SBUSY		shared busy the allocated page
2088  *	VM_ALLOC_WIRED		wire the allocated page
2089  *	VM_ALLOC_ZERO		prefer a zeroed page
2090  */
2091 vm_page_t
2092 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2093     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2094     vm_paddr_t boundary, vm_memattr_t memattr)
2095 {
2096 	struct vm_domainset_iter di;
2097 	vm_page_t m;
2098 	int domain;
2099 
2100 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2101 	do {
2102 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2103 		    npages, low, high, alignment, boundary, memattr);
2104 		if (m != NULL)
2105 			break;
2106 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2107 
2108 	return (m);
2109 }
2110 
2111 vm_page_t
2112 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2113     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2114     vm_paddr_t boundary, vm_memattr_t memattr)
2115 {
2116 	struct vm_domain *vmd;
2117 	vm_page_t m, m_ret, mpred;
2118 	u_int busy_lock, flags, oflags;
2119 
2120 	mpred = NULL;	/* XXX: pacify gcc */
2121 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2122 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2123 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2124 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2125 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2126 	    req));
2127 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2128 	    ("Can't sleep and retry object insertion."));
2129 	if (object != NULL) {
2130 		VM_OBJECT_ASSERT_WLOCKED(object);
2131 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2132 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2133 		    object));
2134 	}
2135 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2136 
2137 	if (object != NULL) {
2138 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2139 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2140 		    ("vm_page_alloc_contig: pindex already allocated"));
2141 	}
2142 
2143 	/*
2144 	 * Can we allocate the pages without the number of free pages falling
2145 	 * below the lower bound for the allocation class?
2146 	 */
2147 	m_ret = NULL;
2148 again:
2149 #if VM_NRESERVLEVEL > 0
2150 	/*
2151 	 * Can we allocate the pages from a reservation?
2152 	 */
2153 	if (vm_object_reserv(object) &&
2154 	    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2155 	    mpred, npages, low, high, alignment, boundary)) != NULL) {
2156 		domain = vm_phys_domain(m_ret);
2157 		vmd = VM_DOMAIN(domain);
2158 		goto found;
2159 	}
2160 #endif
2161 	vmd = VM_DOMAIN(domain);
2162 	if (vm_domain_allocate(vmd, req, npages)) {
2163 		/*
2164 		 * allocate them from the free page queues.
2165 		 */
2166 		vm_domain_free_lock(vmd);
2167 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2168 		    alignment, boundary);
2169 		vm_domain_free_unlock(vmd);
2170 		if (m_ret == NULL) {
2171 			vm_domain_freecnt_inc(vmd, npages);
2172 #if VM_NRESERVLEVEL > 0
2173 			if (vm_reserv_reclaim_contig(domain, npages, low,
2174 			    high, alignment, boundary))
2175 				goto again;
2176 #endif
2177 		}
2178 	}
2179 	if (m_ret == NULL) {
2180 		if (vm_domain_alloc_fail(vmd, object, req))
2181 			goto again;
2182 		return (NULL);
2183 	}
2184 #if VM_NRESERVLEVEL > 0
2185 found:
2186 #endif
2187 	for (m = m_ret; m < &m_ret[npages]; m++) {
2188 		vm_page_dequeue(m);
2189 		vm_page_alloc_check(m);
2190 	}
2191 
2192 	/*
2193 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2194 	 */
2195 	flags = 0;
2196 	if ((req & VM_ALLOC_ZERO) != 0)
2197 		flags = PG_ZERO;
2198 	if ((req & VM_ALLOC_NODUMP) != 0)
2199 		flags |= PG_NODUMP;
2200 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2201 	    VPO_UNMANAGED : 0;
2202 	busy_lock = VPB_UNBUSIED;
2203 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2204 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2205 	if ((req & VM_ALLOC_SBUSY) != 0)
2206 		busy_lock = VPB_SHARERS_WORD(1);
2207 	if ((req & VM_ALLOC_WIRED) != 0)
2208 		vm_wire_add(npages);
2209 	if (object != NULL) {
2210 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2211 		    memattr == VM_MEMATTR_DEFAULT)
2212 			memattr = object->memattr;
2213 	}
2214 	for (m = m_ret; m < &m_ret[npages]; m++) {
2215 		m->a.flags = 0;
2216 		m->flags = (m->flags | PG_NODUMP) & flags;
2217 		m->busy_lock = busy_lock;
2218 		if ((req & VM_ALLOC_WIRED) != 0)
2219 			m->ref_count = 1;
2220 		m->a.act_count = 0;
2221 		m->oflags = oflags;
2222 		if (object != NULL) {
2223 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2224 				if ((req & VM_ALLOC_WIRED) != 0)
2225 					vm_wire_sub(npages);
2226 				KASSERT(m->object == NULL,
2227 				    ("page %p has object", m));
2228 				mpred = m;
2229 				for (m = m_ret; m < &m_ret[npages]; m++) {
2230 					if (m <= mpred &&
2231 					    (req & VM_ALLOC_WIRED) != 0)
2232 						m->ref_count = 0;
2233 					m->oflags = VPO_UNMANAGED;
2234 					m->busy_lock = VPB_UNBUSIED;
2235 					/* Don't change PG_ZERO. */
2236 					vm_page_free_toq(m);
2237 				}
2238 				if (req & VM_ALLOC_WAITFAIL) {
2239 					VM_OBJECT_WUNLOCK(object);
2240 					vm_radix_wait();
2241 					VM_OBJECT_WLOCK(object);
2242 				}
2243 				return (NULL);
2244 			}
2245 			mpred = m;
2246 		} else
2247 			m->pindex = pindex;
2248 		if (memattr != VM_MEMATTR_DEFAULT)
2249 			pmap_page_set_memattr(m, memattr);
2250 		pindex++;
2251 	}
2252 	return (m_ret);
2253 }
2254 
2255 /*
2256  * Check a page that has been freshly dequeued from a freelist.
2257  */
2258 static void
2259 vm_page_alloc_check(vm_page_t m)
2260 {
2261 
2262 	KASSERT(m->object == NULL, ("page %p has object", m));
2263 	KASSERT(m->a.queue == PQ_NONE &&
2264 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2265 	    ("page %p has unexpected queue %d, flags %#x",
2266 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2267 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2268 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2269 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2270 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2271 	    ("page %p has unexpected memattr %d",
2272 	    m, pmap_page_get_memattr(m)));
2273 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2274 }
2275 
2276 /*
2277  * 	vm_page_alloc_freelist:
2278  *
2279  *	Allocate a physical page from the specified free page list.
2280  *
2281  *	The caller must always specify an allocation class.
2282  *
2283  *	allocation classes:
2284  *	VM_ALLOC_NORMAL		normal process request
2285  *	VM_ALLOC_SYSTEM		system *really* needs a page
2286  *	VM_ALLOC_INTERRUPT	interrupt time request
2287  *
2288  *	optional allocation flags:
2289  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2290  *				intends to allocate
2291  *	VM_ALLOC_WIRED		wire the allocated page
2292  *	VM_ALLOC_ZERO		prefer a zeroed page
2293  */
2294 vm_page_t
2295 vm_page_alloc_freelist(int freelist, int req)
2296 {
2297 	struct vm_domainset_iter di;
2298 	vm_page_t m;
2299 	int domain;
2300 
2301 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2302 	do {
2303 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2304 		if (m != NULL)
2305 			break;
2306 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2307 
2308 	return (m);
2309 }
2310 
2311 vm_page_t
2312 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2313 {
2314 	struct vm_domain *vmd;
2315 	vm_page_t m;
2316 	u_int flags;
2317 
2318 	m = NULL;
2319 	vmd = VM_DOMAIN(domain);
2320 again:
2321 	if (vm_domain_allocate(vmd, req, 1)) {
2322 		vm_domain_free_lock(vmd);
2323 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2324 		    VM_FREEPOOL_DIRECT, 0);
2325 		vm_domain_free_unlock(vmd);
2326 		if (m == NULL)
2327 			vm_domain_freecnt_inc(vmd, 1);
2328 	}
2329 	if (m == NULL) {
2330 		if (vm_domain_alloc_fail(vmd, NULL, req))
2331 			goto again;
2332 		return (NULL);
2333 	}
2334 	vm_page_dequeue(m);
2335 	vm_page_alloc_check(m);
2336 
2337 	/*
2338 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2339 	 */
2340 	m->a.flags = 0;
2341 	flags = 0;
2342 	if ((req & VM_ALLOC_ZERO) != 0)
2343 		flags = PG_ZERO;
2344 	m->flags &= flags;
2345 	if ((req & VM_ALLOC_WIRED) != 0) {
2346 		/*
2347 		 * The page lock is not required for wiring a page that does
2348 		 * not belong to an object.
2349 		 */
2350 		vm_wire_add(1);
2351 		m->ref_count = 1;
2352 	}
2353 	/* Unmanaged pages don't use "act_count". */
2354 	m->oflags = VPO_UNMANAGED;
2355 	return (m);
2356 }
2357 
2358 static int
2359 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2360 {
2361 	struct vm_domain *vmd;
2362 	struct vm_pgcache *pgcache;
2363 	int i;
2364 
2365 	pgcache = arg;
2366 	vmd = VM_DOMAIN(pgcache->domain);
2367 
2368 	/*
2369 	 * The page daemon should avoid creating extra memory pressure since its
2370 	 * main purpose is to replenish the store of free pages.
2371 	 */
2372 	if (vmd->vmd_severeset || curproc == pageproc ||
2373 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2374 		return (0);
2375 	domain = vmd->vmd_domain;
2376 	vm_domain_free_lock(vmd);
2377 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2378 	    (vm_page_t *)store);
2379 	vm_domain_free_unlock(vmd);
2380 	if (cnt != i)
2381 		vm_domain_freecnt_inc(vmd, cnt - i);
2382 
2383 	return (i);
2384 }
2385 
2386 static void
2387 vm_page_zone_release(void *arg, void **store, int cnt)
2388 {
2389 	struct vm_domain *vmd;
2390 	struct vm_pgcache *pgcache;
2391 	vm_page_t m;
2392 	int i;
2393 
2394 	pgcache = arg;
2395 	vmd = VM_DOMAIN(pgcache->domain);
2396 	vm_domain_free_lock(vmd);
2397 	for (i = 0; i < cnt; i++) {
2398 		m = (vm_page_t)store[i];
2399 		vm_phys_free_pages(m, 0);
2400 	}
2401 	vm_domain_free_unlock(vmd);
2402 	vm_domain_freecnt_inc(vmd, cnt);
2403 }
2404 
2405 #define	VPSC_ANY	0	/* No restrictions. */
2406 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2407 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2408 
2409 /*
2410  *	vm_page_scan_contig:
2411  *
2412  *	Scan vm_page_array[] between the specified entries "m_start" and
2413  *	"m_end" for a run of contiguous physical pages that satisfy the
2414  *	specified conditions, and return the lowest page in the run.  The
2415  *	specified "alignment" determines the alignment of the lowest physical
2416  *	page in the run.  If the specified "boundary" is non-zero, then the
2417  *	run of physical pages cannot span a physical address that is a
2418  *	multiple of "boundary".
2419  *
2420  *	"m_end" is never dereferenced, so it need not point to a vm_page
2421  *	structure within vm_page_array[].
2422  *
2423  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2424  *	span a hole (or discontiguity) in the physical address space.  Both
2425  *	"alignment" and "boundary" must be a power of two.
2426  */
2427 vm_page_t
2428 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2429     u_long alignment, vm_paddr_t boundary, int options)
2430 {
2431 	struct mtx *m_mtx;
2432 	vm_object_t object;
2433 	vm_paddr_t pa;
2434 	vm_page_t m, m_run;
2435 #if VM_NRESERVLEVEL > 0
2436 	int level;
2437 #endif
2438 	int m_inc, order, run_ext, run_len;
2439 
2440 	KASSERT(npages > 0, ("npages is 0"));
2441 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2442 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2443 	m_run = NULL;
2444 	run_len = 0;
2445 	m_mtx = NULL;
2446 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2447 		KASSERT((m->flags & PG_MARKER) == 0,
2448 		    ("page %p is PG_MARKER", m));
2449 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2450 		    ("fictitious page %p has invalid ref count", m));
2451 
2452 		/*
2453 		 * If the current page would be the start of a run, check its
2454 		 * physical address against the end, alignment, and boundary
2455 		 * conditions.  If it doesn't satisfy these conditions, either
2456 		 * terminate the scan or advance to the next page that
2457 		 * satisfies the failed condition.
2458 		 */
2459 		if (run_len == 0) {
2460 			KASSERT(m_run == NULL, ("m_run != NULL"));
2461 			if (m + npages > m_end)
2462 				break;
2463 			pa = VM_PAGE_TO_PHYS(m);
2464 			if ((pa & (alignment - 1)) != 0) {
2465 				m_inc = atop(roundup2(pa, alignment) - pa);
2466 				continue;
2467 			}
2468 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2469 			    boundary) != 0) {
2470 				m_inc = atop(roundup2(pa, boundary) - pa);
2471 				continue;
2472 			}
2473 		} else
2474 			KASSERT(m_run != NULL, ("m_run == NULL"));
2475 
2476 		vm_page_change_lock(m, &m_mtx);
2477 		m_inc = 1;
2478 retry:
2479 		if (vm_page_wired(m))
2480 			run_ext = 0;
2481 #if VM_NRESERVLEVEL > 0
2482 		else if ((level = vm_reserv_level(m)) >= 0 &&
2483 		    (options & VPSC_NORESERV) != 0) {
2484 			run_ext = 0;
2485 			/* Advance to the end of the reservation. */
2486 			pa = VM_PAGE_TO_PHYS(m);
2487 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2488 			    pa);
2489 		}
2490 #endif
2491 		else if ((object = m->object) != NULL) {
2492 			/*
2493 			 * The page is considered eligible for relocation if
2494 			 * and only if it could be laundered or reclaimed by
2495 			 * the page daemon.
2496 			 */
2497 			if (!VM_OBJECT_TRYRLOCK(object)) {
2498 				mtx_unlock(m_mtx);
2499 				VM_OBJECT_RLOCK(object);
2500 				mtx_lock(m_mtx);
2501 				if (m->object != object) {
2502 					/*
2503 					 * The page may have been freed.
2504 					 */
2505 					VM_OBJECT_RUNLOCK(object);
2506 					goto retry;
2507 				}
2508 			}
2509 			/* Don't care: PG_NODUMP, PG_ZERO. */
2510 			if (object->type != OBJT_DEFAULT &&
2511 			    object->type != OBJT_SWAP &&
2512 			    object->type != OBJT_VNODE) {
2513 				run_ext = 0;
2514 #if VM_NRESERVLEVEL > 0
2515 			} else if ((options & VPSC_NOSUPER) != 0 &&
2516 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2517 				run_ext = 0;
2518 				/* Advance to the end of the superpage. */
2519 				pa = VM_PAGE_TO_PHYS(m);
2520 				m_inc = atop(roundup2(pa + 1,
2521 				    vm_reserv_size(level)) - pa);
2522 #endif
2523 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2524 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) &&
2525 			    !vm_page_wired(m)) {
2526 				/*
2527 				 * The page is allocated but eligible for
2528 				 * relocation.  Extend the current run by one
2529 				 * page.
2530 				 */
2531 				KASSERT(pmap_page_get_memattr(m) ==
2532 				    VM_MEMATTR_DEFAULT,
2533 				    ("page %p has an unexpected memattr", m));
2534 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2535 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2536 				    ("page %p has unexpected oflags", m));
2537 				/* Don't care: PGA_NOSYNC. */
2538 				run_ext = 1;
2539 			} else
2540 				run_ext = 0;
2541 			VM_OBJECT_RUNLOCK(object);
2542 #if VM_NRESERVLEVEL > 0
2543 		} else if (level >= 0) {
2544 			/*
2545 			 * The page is reserved but not yet allocated.  In
2546 			 * other words, it is still free.  Extend the current
2547 			 * run by one page.
2548 			 */
2549 			run_ext = 1;
2550 #endif
2551 		} else if ((order = m->order) < VM_NFREEORDER) {
2552 			/*
2553 			 * The page is enqueued in the physical memory
2554 			 * allocator's free page queues.  Moreover, it is the
2555 			 * first page in a power-of-two-sized run of
2556 			 * contiguous free pages.  Add these pages to the end
2557 			 * of the current run, and jump ahead.
2558 			 */
2559 			run_ext = 1 << order;
2560 			m_inc = 1 << order;
2561 		} else {
2562 			/*
2563 			 * Skip the page for one of the following reasons: (1)
2564 			 * It is enqueued in the physical memory allocator's
2565 			 * free page queues.  However, it is not the first
2566 			 * page in a run of contiguous free pages.  (This case
2567 			 * rarely occurs because the scan is performed in
2568 			 * ascending order.) (2) It is not reserved, and it is
2569 			 * transitioning from free to allocated.  (Conversely,
2570 			 * the transition from allocated to free for managed
2571 			 * pages is blocked by the page lock.) (3) It is
2572 			 * allocated but not contained by an object and not
2573 			 * wired, e.g., allocated by Xen's balloon driver.
2574 			 */
2575 			run_ext = 0;
2576 		}
2577 
2578 		/*
2579 		 * Extend or reset the current run of pages.
2580 		 */
2581 		if (run_ext > 0) {
2582 			if (run_len == 0)
2583 				m_run = m;
2584 			run_len += run_ext;
2585 		} else {
2586 			if (run_len > 0) {
2587 				m_run = NULL;
2588 				run_len = 0;
2589 			}
2590 		}
2591 	}
2592 	if (m_mtx != NULL)
2593 		mtx_unlock(m_mtx);
2594 	if (run_len >= npages)
2595 		return (m_run);
2596 	return (NULL);
2597 }
2598 
2599 /*
2600  *	vm_page_reclaim_run:
2601  *
2602  *	Try to relocate each of the allocated virtual pages within the
2603  *	specified run of physical pages to a new physical address.  Free the
2604  *	physical pages underlying the relocated virtual pages.  A virtual page
2605  *	is relocatable if and only if it could be laundered or reclaimed by
2606  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2607  *	physical address above "high".
2608  *
2609  *	Returns 0 if every physical page within the run was already free or
2610  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2611  *	value indicating why the last attempt to relocate a virtual page was
2612  *	unsuccessful.
2613  *
2614  *	"req_class" must be an allocation class.
2615  */
2616 static int
2617 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2618     vm_paddr_t high)
2619 {
2620 	struct vm_domain *vmd;
2621 	struct mtx *m_mtx;
2622 	struct spglist free;
2623 	vm_object_t object;
2624 	vm_paddr_t pa;
2625 	vm_page_t m, m_end, m_new;
2626 	int error, order, req;
2627 
2628 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2629 	    ("req_class is not an allocation class"));
2630 	SLIST_INIT(&free);
2631 	error = 0;
2632 	m = m_run;
2633 	m_end = m_run + npages;
2634 	m_mtx = NULL;
2635 	for (; error == 0 && m < m_end; m++) {
2636 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2637 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2638 
2639 		/*
2640 		 * Avoid releasing and reacquiring the same page lock.
2641 		 */
2642 		vm_page_change_lock(m, &m_mtx);
2643 retry:
2644 		/*
2645 		 * Racily check for wirings.  Races are handled below.
2646 		 */
2647 		if (vm_page_wired(m))
2648 			error = EBUSY;
2649 		else if ((object = m->object) != NULL) {
2650 			/*
2651 			 * The page is relocated if and only if it could be
2652 			 * laundered or reclaimed by the page daemon.
2653 			 */
2654 			if (!VM_OBJECT_TRYWLOCK(object)) {
2655 				mtx_unlock(m_mtx);
2656 				VM_OBJECT_WLOCK(object);
2657 				mtx_lock(m_mtx);
2658 				if (m->object != object) {
2659 					/*
2660 					 * The page may have been freed.
2661 					 */
2662 					VM_OBJECT_WUNLOCK(object);
2663 					goto retry;
2664 				}
2665 			}
2666 			/* Don't care: PG_NODUMP, PG_ZERO. */
2667 			if (object->type != OBJT_DEFAULT &&
2668 			    object->type != OBJT_SWAP &&
2669 			    object->type != OBJT_VNODE)
2670 				error = EINVAL;
2671 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2672 				error = EINVAL;
2673 			else if (vm_page_queue(m) != PQ_NONE &&
2674 			    vm_page_tryxbusy(m) != 0) {
2675 				if (vm_page_wired(m)) {
2676 					vm_page_xunbusy(m);
2677 					error = EBUSY;
2678 					goto unlock;
2679 				}
2680 				KASSERT(pmap_page_get_memattr(m) ==
2681 				    VM_MEMATTR_DEFAULT,
2682 				    ("page %p has an unexpected memattr", m));
2683 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2684 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2685 				    ("page %p has unexpected oflags", m));
2686 				/* Don't care: PGA_NOSYNC. */
2687 				if (!vm_page_none_valid(m)) {
2688 					/*
2689 					 * First, try to allocate a new page
2690 					 * that is above "high".  Failing
2691 					 * that, try to allocate a new page
2692 					 * that is below "m_run".  Allocate
2693 					 * the new page between the end of
2694 					 * "m_run" and "high" only as a last
2695 					 * resort.
2696 					 */
2697 					req = req_class | VM_ALLOC_NOOBJ;
2698 					if ((m->flags & PG_NODUMP) != 0)
2699 						req |= VM_ALLOC_NODUMP;
2700 					if (trunc_page(high) !=
2701 					    ~(vm_paddr_t)PAGE_MASK) {
2702 						m_new = vm_page_alloc_contig(
2703 						    NULL, 0, req, 1,
2704 						    round_page(high),
2705 						    ~(vm_paddr_t)0,
2706 						    PAGE_SIZE, 0,
2707 						    VM_MEMATTR_DEFAULT);
2708 					} else
2709 						m_new = NULL;
2710 					if (m_new == NULL) {
2711 						pa = VM_PAGE_TO_PHYS(m_run);
2712 						m_new = vm_page_alloc_contig(
2713 						    NULL, 0, req, 1,
2714 						    0, pa - 1, PAGE_SIZE, 0,
2715 						    VM_MEMATTR_DEFAULT);
2716 					}
2717 					if (m_new == NULL) {
2718 						pa += ptoa(npages);
2719 						m_new = vm_page_alloc_contig(
2720 						    NULL, 0, req, 1,
2721 						    pa, high, PAGE_SIZE, 0,
2722 						    VM_MEMATTR_DEFAULT);
2723 					}
2724 					if (m_new == NULL) {
2725 						vm_page_xunbusy(m);
2726 						error = ENOMEM;
2727 						goto unlock;
2728 					}
2729 
2730 					/*
2731 					 * Unmap the page and check for new
2732 					 * wirings that may have been acquired
2733 					 * through a pmap lookup.
2734 					 */
2735 					if (object->ref_count != 0 &&
2736 					    !vm_page_try_remove_all(m)) {
2737 						vm_page_free(m_new);
2738 						error = EBUSY;
2739 						goto unlock;
2740 					}
2741 
2742 					/*
2743 					 * Replace "m" with the new page.  For
2744 					 * vm_page_replace(), "m" must be busy
2745 					 * and dequeued.  Finally, change "m"
2746 					 * as if vm_page_free() was called.
2747 					 */
2748 					m_new->a.flags = m->a.flags &
2749 					    ~PGA_QUEUE_STATE_MASK;
2750 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2751 					    ("page %p is managed", m_new));
2752 					pmap_copy_page(m, m_new);
2753 					m_new->valid = m->valid;
2754 					m_new->dirty = m->dirty;
2755 					m->flags &= ~PG_ZERO;
2756 					vm_page_dequeue(m);
2757 					vm_page_replace_checked(m_new, object,
2758 					    m->pindex, m);
2759 					if (vm_page_free_prep(m))
2760 						SLIST_INSERT_HEAD(&free, m,
2761 						    plinks.s.ss);
2762 
2763 					/*
2764 					 * The new page must be deactivated
2765 					 * before the object is unlocked.
2766 					 */
2767 					vm_page_change_lock(m_new, &m_mtx);
2768 					vm_page_deactivate(m_new);
2769 				} else {
2770 					m->flags &= ~PG_ZERO;
2771 					vm_page_dequeue(m);
2772 					if (vm_page_free_prep(m))
2773 						SLIST_INSERT_HEAD(&free, m,
2774 						    plinks.s.ss);
2775 					KASSERT(m->dirty == 0,
2776 					    ("page %p is dirty", m));
2777 				}
2778 			} else
2779 				error = EBUSY;
2780 unlock:
2781 			VM_OBJECT_WUNLOCK(object);
2782 		} else {
2783 			MPASS(vm_phys_domain(m) == domain);
2784 			vmd = VM_DOMAIN(domain);
2785 			vm_domain_free_lock(vmd);
2786 			order = m->order;
2787 			if (order < VM_NFREEORDER) {
2788 				/*
2789 				 * The page is enqueued in the physical memory
2790 				 * allocator's free page queues.  Moreover, it
2791 				 * is the first page in a power-of-two-sized
2792 				 * run of contiguous free pages.  Jump ahead
2793 				 * to the last page within that run, and
2794 				 * continue from there.
2795 				 */
2796 				m += (1 << order) - 1;
2797 			}
2798 #if VM_NRESERVLEVEL > 0
2799 			else if (vm_reserv_is_page_free(m))
2800 				order = 0;
2801 #endif
2802 			vm_domain_free_unlock(vmd);
2803 			if (order == VM_NFREEORDER)
2804 				error = EINVAL;
2805 		}
2806 	}
2807 	if (m_mtx != NULL)
2808 		mtx_unlock(m_mtx);
2809 	if ((m = SLIST_FIRST(&free)) != NULL) {
2810 		int cnt;
2811 
2812 		vmd = VM_DOMAIN(domain);
2813 		cnt = 0;
2814 		vm_domain_free_lock(vmd);
2815 		do {
2816 			MPASS(vm_phys_domain(m) == domain);
2817 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2818 			vm_phys_free_pages(m, 0);
2819 			cnt++;
2820 		} while ((m = SLIST_FIRST(&free)) != NULL);
2821 		vm_domain_free_unlock(vmd);
2822 		vm_domain_freecnt_inc(vmd, cnt);
2823 	}
2824 	return (error);
2825 }
2826 
2827 #define	NRUNS	16
2828 
2829 CTASSERT(powerof2(NRUNS));
2830 
2831 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2832 
2833 #define	MIN_RECLAIM	8
2834 
2835 /*
2836  *	vm_page_reclaim_contig:
2837  *
2838  *	Reclaim allocated, contiguous physical memory satisfying the specified
2839  *	conditions by relocating the virtual pages using that physical memory.
2840  *	Returns true if reclamation is successful and false otherwise.  Since
2841  *	relocation requires the allocation of physical pages, reclamation may
2842  *	fail due to a shortage of free pages.  When reclamation fails, callers
2843  *	are expected to perform vm_wait() before retrying a failed allocation
2844  *	operation, e.g., vm_page_alloc_contig().
2845  *
2846  *	The caller must always specify an allocation class through "req".
2847  *
2848  *	allocation classes:
2849  *	VM_ALLOC_NORMAL		normal process request
2850  *	VM_ALLOC_SYSTEM		system *really* needs a page
2851  *	VM_ALLOC_INTERRUPT	interrupt time request
2852  *
2853  *	The optional allocation flags are ignored.
2854  *
2855  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2856  *	must be a power of two.
2857  */
2858 bool
2859 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2860     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2861 {
2862 	struct vm_domain *vmd;
2863 	vm_paddr_t curr_low;
2864 	vm_page_t m_run, m_runs[NRUNS];
2865 	u_long count, reclaimed;
2866 	int error, i, options, req_class;
2867 
2868 	KASSERT(npages > 0, ("npages is 0"));
2869 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2870 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2871 	req_class = req & VM_ALLOC_CLASS_MASK;
2872 
2873 	/*
2874 	 * The page daemon is allowed to dig deeper into the free page list.
2875 	 */
2876 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2877 		req_class = VM_ALLOC_SYSTEM;
2878 
2879 	/*
2880 	 * Return if the number of free pages cannot satisfy the requested
2881 	 * allocation.
2882 	 */
2883 	vmd = VM_DOMAIN(domain);
2884 	count = vmd->vmd_free_count;
2885 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2886 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2887 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2888 		return (false);
2889 
2890 	/*
2891 	 * Scan up to three times, relaxing the restrictions ("options") on
2892 	 * the reclamation of reservations and superpages each time.
2893 	 */
2894 	for (options = VPSC_NORESERV;;) {
2895 		/*
2896 		 * Find the highest runs that satisfy the given constraints
2897 		 * and restrictions, and record them in "m_runs".
2898 		 */
2899 		curr_low = low;
2900 		count = 0;
2901 		for (;;) {
2902 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2903 			    high, alignment, boundary, options);
2904 			if (m_run == NULL)
2905 				break;
2906 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2907 			m_runs[RUN_INDEX(count)] = m_run;
2908 			count++;
2909 		}
2910 
2911 		/*
2912 		 * Reclaim the highest runs in LIFO (descending) order until
2913 		 * the number of reclaimed pages, "reclaimed", is at least
2914 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2915 		 * reclamation is idempotent, and runs will (likely) recur
2916 		 * from one scan to the next as restrictions are relaxed.
2917 		 */
2918 		reclaimed = 0;
2919 		for (i = 0; count > 0 && i < NRUNS; i++) {
2920 			count--;
2921 			m_run = m_runs[RUN_INDEX(count)];
2922 			error = vm_page_reclaim_run(req_class, domain, npages,
2923 			    m_run, high);
2924 			if (error == 0) {
2925 				reclaimed += npages;
2926 				if (reclaimed >= MIN_RECLAIM)
2927 					return (true);
2928 			}
2929 		}
2930 
2931 		/*
2932 		 * Either relax the restrictions on the next scan or return if
2933 		 * the last scan had no restrictions.
2934 		 */
2935 		if (options == VPSC_NORESERV)
2936 			options = VPSC_NOSUPER;
2937 		else if (options == VPSC_NOSUPER)
2938 			options = VPSC_ANY;
2939 		else if (options == VPSC_ANY)
2940 			return (reclaimed != 0);
2941 	}
2942 }
2943 
2944 bool
2945 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2946     u_long alignment, vm_paddr_t boundary)
2947 {
2948 	struct vm_domainset_iter di;
2949 	int domain;
2950 	bool ret;
2951 
2952 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2953 	do {
2954 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2955 		    high, alignment, boundary);
2956 		if (ret)
2957 			break;
2958 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2959 
2960 	return (ret);
2961 }
2962 
2963 /*
2964  * Set the domain in the appropriate page level domainset.
2965  */
2966 void
2967 vm_domain_set(struct vm_domain *vmd)
2968 {
2969 
2970 	mtx_lock(&vm_domainset_lock);
2971 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2972 		vmd->vmd_minset = 1;
2973 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2974 	}
2975 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2976 		vmd->vmd_severeset = 1;
2977 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2978 	}
2979 	mtx_unlock(&vm_domainset_lock);
2980 }
2981 
2982 /*
2983  * Clear the domain from the appropriate page level domainset.
2984  */
2985 void
2986 vm_domain_clear(struct vm_domain *vmd)
2987 {
2988 
2989 	mtx_lock(&vm_domainset_lock);
2990 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2991 		vmd->vmd_minset = 0;
2992 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2993 		if (vm_min_waiters != 0) {
2994 			vm_min_waiters = 0;
2995 			wakeup(&vm_min_domains);
2996 		}
2997 	}
2998 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
2999 		vmd->vmd_severeset = 0;
3000 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3001 		if (vm_severe_waiters != 0) {
3002 			vm_severe_waiters = 0;
3003 			wakeup(&vm_severe_domains);
3004 		}
3005 	}
3006 
3007 	/*
3008 	 * If pageout daemon needs pages, then tell it that there are
3009 	 * some free.
3010 	 */
3011 	if (vmd->vmd_pageout_pages_needed &&
3012 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3013 		wakeup(&vmd->vmd_pageout_pages_needed);
3014 		vmd->vmd_pageout_pages_needed = 0;
3015 	}
3016 
3017 	/* See comments in vm_wait_doms(). */
3018 	if (vm_pageproc_waiters) {
3019 		vm_pageproc_waiters = 0;
3020 		wakeup(&vm_pageproc_waiters);
3021 	}
3022 	mtx_unlock(&vm_domainset_lock);
3023 }
3024 
3025 /*
3026  * Wait for free pages to exceed the min threshold globally.
3027  */
3028 void
3029 vm_wait_min(void)
3030 {
3031 
3032 	mtx_lock(&vm_domainset_lock);
3033 	while (vm_page_count_min()) {
3034 		vm_min_waiters++;
3035 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3036 	}
3037 	mtx_unlock(&vm_domainset_lock);
3038 }
3039 
3040 /*
3041  * Wait for free pages to exceed the severe threshold globally.
3042  */
3043 void
3044 vm_wait_severe(void)
3045 {
3046 
3047 	mtx_lock(&vm_domainset_lock);
3048 	while (vm_page_count_severe()) {
3049 		vm_severe_waiters++;
3050 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3051 		    "vmwait", 0);
3052 	}
3053 	mtx_unlock(&vm_domainset_lock);
3054 }
3055 
3056 u_int
3057 vm_wait_count(void)
3058 {
3059 
3060 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3061 }
3062 
3063 void
3064 vm_wait_doms(const domainset_t *wdoms)
3065 {
3066 
3067 	/*
3068 	 * We use racey wakeup synchronization to avoid expensive global
3069 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3070 	 * To handle this, we only sleep for one tick in this instance.  It
3071 	 * is expected that most allocations for the pageproc will come from
3072 	 * kmem or vm_page_grab* which will use the more specific and
3073 	 * race-free vm_wait_domain().
3074 	 */
3075 	if (curproc == pageproc) {
3076 		mtx_lock(&vm_domainset_lock);
3077 		vm_pageproc_waiters++;
3078 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3079 		    "pageprocwait", 1);
3080 	} else {
3081 		/*
3082 		 * XXX Ideally we would wait only until the allocation could
3083 		 * be satisfied.  This condition can cause new allocators to
3084 		 * consume all freed pages while old allocators wait.
3085 		 */
3086 		mtx_lock(&vm_domainset_lock);
3087 		if (vm_page_count_min_set(wdoms)) {
3088 			vm_min_waiters++;
3089 			msleep(&vm_min_domains, &vm_domainset_lock,
3090 			    PVM | PDROP, "vmwait", 0);
3091 		} else
3092 			mtx_unlock(&vm_domainset_lock);
3093 	}
3094 }
3095 
3096 /*
3097  *	vm_wait_domain:
3098  *
3099  *	Sleep until free pages are available for allocation.
3100  *	- Called in various places after failed memory allocations.
3101  */
3102 void
3103 vm_wait_domain(int domain)
3104 {
3105 	struct vm_domain *vmd;
3106 	domainset_t wdom;
3107 
3108 	vmd = VM_DOMAIN(domain);
3109 	vm_domain_free_assert_unlocked(vmd);
3110 
3111 	if (curproc == pageproc) {
3112 		mtx_lock(&vm_domainset_lock);
3113 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3114 			vmd->vmd_pageout_pages_needed = 1;
3115 			msleep(&vmd->vmd_pageout_pages_needed,
3116 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3117 		} else
3118 			mtx_unlock(&vm_domainset_lock);
3119 	} else {
3120 		if (pageproc == NULL)
3121 			panic("vm_wait in early boot");
3122 		DOMAINSET_ZERO(&wdom);
3123 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3124 		vm_wait_doms(&wdom);
3125 	}
3126 }
3127 
3128 /*
3129  *	vm_wait:
3130  *
3131  *	Sleep until free pages are available for allocation in the
3132  *	affinity domains of the obj.  If obj is NULL, the domain set
3133  *	for the calling thread is used.
3134  *	Called in various places after failed memory allocations.
3135  */
3136 void
3137 vm_wait(vm_object_t obj)
3138 {
3139 	struct domainset *d;
3140 
3141 	d = NULL;
3142 
3143 	/*
3144 	 * Carefully fetch pointers only once: the struct domainset
3145 	 * itself is ummutable but the pointer might change.
3146 	 */
3147 	if (obj != NULL)
3148 		d = obj->domain.dr_policy;
3149 	if (d == NULL)
3150 		d = curthread->td_domain.dr_policy;
3151 
3152 	vm_wait_doms(&d->ds_mask);
3153 }
3154 
3155 /*
3156  *	vm_domain_alloc_fail:
3157  *
3158  *	Called when a page allocation function fails.  Informs the
3159  *	pagedaemon and performs the requested wait.  Requires the
3160  *	domain_free and object lock on entry.  Returns with the
3161  *	object lock held and free lock released.  Returns an error when
3162  *	retry is necessary.
3163  *
3164  */
3165 static int
3166 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3167 {
3168 
3169 	vm_domain_free_assert_unlocked(vmd);
3170 
3171 	atomic_add_int(&vmd->vmd_pageout_deficit,
3172 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3173 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3174 		if (object != NULL)
3175 			VM_OBJECT_WUNLOCK(object);
3176 		vm_wait_domain(vmd->vmd_domain);
3177 		if (object != NULL)
3178 			VM_OBJECT_WLOCK(object);
3179 		if (req & VM_ALLOC_WAITOK)
3180 			return (EAGAIN);
3181 	}
3182 
3183 	return (0);
3184 }
3185 
3186 /*
3187  *	vm_waitpfault:
3188  *
3189  *	Sleep until free pages are available for allocation.
3190  *	- Called only in vm_fault so that processes page faulting
3191  *	  can be easily tracked.
3192  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3193  *	  processes will be able to grab memory first.  Do not change
3194  *	  this balance without careful testing first.
3195  */
3196 void
3197 vm_waitpfault(struct domainset *dset, int timo)
3198 {
3199 
3200 	/*
3201 	 * XXX Ideally we would wait only until the allocation could
3202 	 * be satisfied.  This condition can cause new allocators to
3203 	 * consume all freed pages while old allocators wait.
3204 	 */
3205 	mtx_lock(&vm_domainset_lock);
3206 	if (vm_page_count_min_set(&dset->ds_mask)) {
3207 		vm_min_waiters++;
3208 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3209 		    "pfault", timo);
3210 	} else
3211 		mtx_unlock(&vm_domainset_lock);
3212 }
3213 
3214 static struct vm_pagequeue *
3215 vm_page_pagequeue(vm_page_t m)
3216 {
3217 
3218 	uint8_t queue;
3219 
3220 	if ((queue = atomic_load_8(&m->a.queue)) == PQ_NONE)
3221 		return (NULL);
3222 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3223 }
3224 
3225 static inline void
3226 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3227 {
3228 	struct vm_domain *vmd;
3229 	uint16_t qflags;
3230 
3231 	CRITICAL_ASSERT(curthread);
3232 	vm_pagequeue_assert_locked(pq);
3233 
3234 	/*
3235 	 * The page daemon is allowed to set m->a.queue = PQ_NONE without
3236 	 * the page queue lock held.  In this case it is about to free the page,
3237 	 * which must not have any queue state.
3238 	 */
3239 	qflags = atomic_load_16(&m->a.flags);
3240 	KASSERT(pq == vm_page_pagequeue(m) ||
3241 	    (qflags & PGA_QUEUE_STATE_MASK) == 0,
3242 	    ("page %p doesn't belong to queue %p but has aflags %#x",
3243 	    m, pq, qflags));
3244 
3245 	if ((qflags & PGA_DEQUEUE) != 0) {
3246 		if (__predict_true((qflags & PGA_ENQUEUED) != 0))
3247 			vm_pagequeue_remove(pq, m);
3248 		vm_page_dequeue_complete(m);
3249 		counter_u64_add(queue_ops, 1);
3250 	} else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3251 		if ((qflags & PGA_ENQUEUED) != 0)
3252 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3253 		else {
3254 			vm_pagequeue_cnt_inc(pq);
3255 			vm_page_aflag_set(m, PGA_ENQUEUED);
3256 		}
3257 
3258 		/*
3259 		 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.
3260 		 * In particular, if both flags are set in close succession,
3261 		 * only PGA_REQUEUE_HEAD will be applied, even if it was set
3262 		 * first.
3263 		 */
3264 		if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3265 			KASSERT(m->a.queue == PQ_INACTIVE,
3266 			    ("head enqueue not supported for page %p", m));
3267 			vmd = vm_pagequeue_domain(m);
3268 			TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3269 		} else
3270 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3271 
3272 		vm_page_aflag_clear(m, qflags & (PGA_REQUEUE |
3273 		    PGA_REQUEUE_HEAD));
3274 		counter_u64_add(queue_ops, 1);
3275 	} else {
3276 		counter_u64_add(queue_nops, 1);
3277 	}
3278 }
3279 
3280 static void
3281 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3282     uint8_t queue)
3283 {
3284 	vm_page_t m;
3285 	int i;
3286 
3287 	for (i = 0; i < bq->bq_cnt; i++) {
3288 		m = bq->bq_pa[i];
3289 		if (__predict_false(m->a.queue != queue))
3290 			continue;
3291 		vm_pqbatch_process_page(pq, m);
3292 	}
3293 	vm_batchqueue_init(bq);
3294 }
3295 
3296 /*
3297  *	vm_page_pqbatch_submit:		[ internal use only ]
3298  *
3299  *	Enqueue a page in the specified page queue's batched work queue.
3300  *	The caller must have encoded the requested operation in the page
3301  *	structure's a.flags field.
3302  */
3303 void
3304 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3305 {
3306 	struct vm_batchqueue *bq;
3307 	struct vm_pagequeue *pq;
3308 	int domain;
3309 
3310 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3311 	    ("page %p is unmanaged", m));
3312 	KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL,
3313 	    ("missing synchronization for page %p", m));
3314 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3315 
3316 	domain = vm_phys_domain(m);
3317 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3318 
3319 	critical_enter();
3320 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3321 	if (vm_batchqueue_insert(bq, m)) {
3322 		critical_exit();
3323 		return;
3324 	}
3325 	critical_exit();
3326 	vm_pagequeue_lock(pq);
3327 	critical_enter();
3328 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3329 	vm_pqbatch_process(pq, bq, queue);
3330 
3331 	/*
3332 	 * The page may have been logically dequeued before we acquired the
3333 	 * page queue lock.  In this case, since we either hold the page lock
3334 	 * or the page is being freed, a different thread cannot be concurrently
3335 	 * enqueuing the page.
3336 	 */
3337 	if (__predict_true(m->a.queue == queue))
3338 		vm_pqbatch_process_page(pq, m);
3339 	else {
3340 		KASSERT(m->a.queue == PQ_NONE,
3341 		    ("invalid queue transition for page %p", m));
3342 		KASSERT((m->a.flags & PGA_ENQUEUED) == 0,
3343 		    ("page %p is enqueued with invalid queue index", m));
3344 	}
3345 	vm_pagequeue_unlock(pq);
3346 	critical_exit();
3347 }
3348 
3349 /*
3350  *	vm_page_pqbatch_drain:		[ internal use only ]
3351  *
3352  *	Force all per-CPU page queue batch queues to be drained.  This is
3353  *	intended for use in severe memory shortages, to ensure that pages
3354  *	do not remain stuck in the batch queues.
3355  */
3356 void
3357 vm_page_pqbatch_drain(void)
3358 {
3359 	struct thread *td;
3360 	struct vm_domain *vmd;
3361 	struct vm_pagequeue *pq;
3362 	int cpu, domain, queue;
3363 
3364 	td = curthread;
3365 	CPU_FOREACH(cpu) {
3366 		thread_lock(td);
3367 		sched_bind(td, cpu);
3368 		thread_unlock(td);
3369 
3370 		for (domain = 0; domain < vm_ndomains; domain++) {
3371 			vmd = VM_DOMAIN(domain);
3372 			for (queue = 0; queue < PQ_COUNT; queue++) {
3373 				pq = &vmd->vmd_pagequeues[queue];
3374 				vm_pagequeue_lock(pq);
3375 				critical_enter();
3376 				vm_pqbatch_process(pq,
3377 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3378 				critical_exit();
3379 				vm_pagequeue_unlock(pq);
3380 			}
3381 		}
3382 	}
3383 	thread_lock(td);
3384 	sched_unbind(td);
3385 	thread_unlock(td);
3386 }
3387 
3388 /*
3389  * Complete the logical removal of a page from a page queue.  We must be
3390  * careful to synchronize with the page daemon, which may be concurrently
3391  * examining the page with only the page lock held.  The page must not be
3392  * in a state where it appears to be logically enqueued.
3393  */
3394 static void
3395 vm_page_dequeue_complete(vm_page_t m)
3396 {
3397 
3398 	m->a.queue = PQ_NONE;
3399 	atomic_thread_fence_rel();
3400 	vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3401 }
3402 
3403 /*
3404  *	vm_page_dequeue_deferred:	[ internal use only ]
3405  *
3406  *	Request removal of the given page from its current page
3407  *	queue.  Physical removal from the queue may be deferred
3408  *	indefinitely.
3409  *
3410  *	The page must be locked.
3411  */
3412 void
3413 vm_page_dequeue_deferred(vm_page_t m)
3414 {
3415 	uint8_t queue;
3416 
3417 	vm_page_assert_locked(m);
3418 
3419 	if ((queue = vm_page_queue(m)) == PQ_NONE)
3420 		return;
3421 
3422 	/*
3423 	 * Set PGA_DEQUEUE if it is not already set to handle a concurrent call
3424 	 * to vm_page_dequeue_deferred_free().  In particular, avoid modifying
3425 	 * the page's queue state once vm_page_dequeue_deferred_free() has been
3426 	 * called.  In the event of a race, two batch queue entries for the page
3427 	 * will be created, but the second will have no effect.
3428 	 */
3429 	if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE))
3430 		vm_page_pqbatch_submit(m, queue);
3431 }
3432 
3433 /*
3434  * A variant of vm_page_dequeue_deferred() that does not assert the page
3435  * lock and is only to be called from vm_page_free_prep().  Because the
3436  * page is being freed, we can assume that nothing other than the page
3437  * daemon is scheduling queue operations on this page, so we get for
3438  * free the mutual exclusion that is otherwise provided by the page lock.
3439  * To handle races, the page daemon must take care to atomically check
3440  * for PGA_DEQUEUE when updating queue state.
3441  */
3442 static void
3443 vm_page_dequeue_deferred_free(vm_page_t m)
3444 {
3445 	uint8_t queue;
3446 
3447 	KASSERT(m->ref_count == 0, ("page %p has references", m));
3448 
3449 	for (;;) {
3450 		if ((m->a.flags & PGA_DEQUEUE) != 0)
3451 			return;
3452 		atomic_thread_fence_acq();
3453 		if ((queue = atomic_load_8(&m->a.queue)) == PQ_NONE)
3454 			return;
3455 		if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE,
3456 		    PGA_DEQUEUE)) {
3457 			vm_page_pqbatch_submit(m, queue);
3458 			break;
3459 		}
3460 	}
3461 }
3462 
3463 /*
3464  *	vm_page_dequeue:
3465  *
3466  *	Remove the page from whichever page queue it's in, if any.
3467  *	The page must either be locked or unallocated.  This constraint
3468  *	ensures that the queue state of the page will remain consistent
3469  *	after this function returns.
3470  */
3471 void
3472 vm_page_dequeue(vm_page_t m)
3473 {
3474 	struct vm_pagequeue *pq, *pq1;
3475 	uint16_t aflags;
3476 
3477 	KASSERT(mtx_owned(vm_page_lockptr(m)) || m->ref_count == 0,
3478 	    ("page %p is allocated and unlocked", m));
3479 
3480 	for (pq = vm_page_pagequeue(m);; pq = pq1) {
3481 		if (pq == NULL) {
3482 			/*
3483 			 * A thread may be concurrently executing
3484 			 * vm_page_dequeue_complete().  Ensure that all queue
3485 			 * state is cleared before we return.
3486 			 */
3487 			aflags = atomic_load_16(&m->a.flags);
3488 			if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3489 				return;
3490 			KASSERT((aflags & PGA_DEQUEUE) != 0,
3491 			    ("page %p has unexpected queue state flags %#x",
3492 			    m, aflags));
3493 
3494 			/*
3495 			 * Busy wait until the thread updating queue state is
3496 			 * finished.  Such a thread must be executing in a
3497 			 * critical section.
3498 			 */
3499 			cpu_spinwait();
3500 			pq1 = vm_page_pagequeue(m);
3501 			continue;
3502 		}
3503 		vm_pagequeue_lock(pq);
3504 		if ((pq1 = vm_page_pagequeue(m)) == pq)
3505 			break;
3506 		vm_pagequeue_unlock(pq);
3507 	}
3508 	KASSERT(pq == vm_page_pagequeue(m),
3509 	    ("%s: page %p migrated directly between queues", __func__, m));
3510 	KASSERT((m->a.flags & PGA_DEQUEUE) != 0 ||
3511 	    mtx_owned(vm_page_lockptr(m)),
3512 	    ("%s: queued unlocked page %p", __func__, m));
3513 
3514 	if ((m->a.flags & PGA_ENQUEUED) != 0)
3515 		vm_pagequeue_remove(pq, m);
3516 	vm_page_dequeue_complete(m);
3517 	vm_pagequeue_unlock(pq);
3518 }
3519 
3520 /*
3521  * Schedule the given page for insertion into the specified page queue.
3522  * Physical insertion of the page may be deferred indefinitely.
3523  */
3524 static void
3525 vm_page_enqueue(vm_page_t m, uint8_t queue)
3526 {
3527 
3528 	vm_page_assert_locked(m);
3529 	KASSERT(m->a.queue == PQ_NONE &&
3530 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3531 	    ("%s: page %p is already enqueued", __func__, m));
3532 	KASSERT(m->ref_count > 0,
3533 	    ("%s: page %p does not carry any references", __func__, m));
3534 
3535 	m->a.queue = queue;
3536 	if ((m->a.flags & PGA_REQUEUE) == 0)
3537 		vm_page_aflag_set(m, PGA_REQUEUE);
3538 	vm_page_pqbatch_submit(m, queue);
3539 }
3540 
3541 /*
3542  *	vm_page_requeue:		[ internal use only ]
3543  *
3544  *	Schedule a requeue of the given page.
3545  *
3546  *	The page must be locked.
3547  */
3548 void
3549 vm_page_requeue(vm_page_t m)
3550 {
3551 
3552 	vm_page_assert_locked(m);
3553 	KASSERT(vm_page_queue(m) != PQ_NONE,
3554 	    ("%s: page %p is not logically enqueued", __func__, m));
3555 	KASSERT(m->ref_count > 0,
3556 	    ("%s: page %p does not carry any references", __func__, m));
3557 
3558 	if ((m->a.flags & PGA_REQUEUE) == 0)
3559 		vm_page_aflag_set(m, PGA_REQUEUE);
3560 	vm_page_pqbatch_submit(m, atomic_load_8(&m->a.queue));
3561 }
3562 
3563 /*
3564  *	vm_page_swapqueue:		[ internal use only ]
3565  *
3566  *	Move the page from one queue to another, or to the tail of its
3567  *	current queue, in the face of a possible concurrent call to
3568  *	vm_page_dequeue_deferred_free().
3569  */
3570 void
3571 vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq)
3572 {
3573 	struct vm_pagequeue *pq;
3574 	vm_page_t next;
3575 	bool queued;
3576 
3577 	KASSERT(oldq < PQ_COUNT && newq < PQ_COUNT && oldq != newq,
3578 	    ("vm_page_swapqueue: invalid queues (%d, %d)", oldq, newq));
3579 	vm_page_assert_locked(m);
3580 
3581 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[oldq];
3582 	vm_pagequeue_lock(pq);
3583 
3584 	/*
3585 	 * The physical queue state might change at any point before the page
3586 	 * queue lock is acquired, so we must verify that we hold the correct
3587 	 * lock before proceeding.
3588 	 */
3589 	if (__predict_false(m->a.queue != oldq)) {
3590 		vm_pagequeue_unlock(pq);
3591 		return;
3592 	}
3593 
3594 	/*
3595 	 * Once the queue index of the page changes, there is nothing
3596 	 * synchronizing with further updates to the physical queue state.
3597 	 * Therefore we must remove the page from the queue now in anticipation
3598 	 * of a successful commit, and be prepared to roll back.
3599 	 */
3600 	if (__predict_true((m->a.flags & PGA_ENQUEUED) != 0)) {
3601 		next = TAILQ_NEXT(m, plinks.q);
3602 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3603 		vm_page_aflag_clear(m, PGA_ENQUEUED);
3604 		queued = true;
3605 	} else {
3606 		queued = false;
3607 	}
3608 
3609 	/*
3610 	 * Atomically update the queue field and set PGA_REQUEUE while
3611 	 * ensuring that PGA_DEQUEUE has not been set.
3612 	 */
3613 	if (__predict_false(!vm_page_pqstate_cmpset(m, oldq, newq, PGA_DEQUEUE,
3614 	    PGA_REQUEUE))) {
3615 		if (queued) {
3616 			vm_page_aflag_set(m, PGA_ENQUEUED);
3617 			if (next != NULL)
3618 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3619 			else
3620 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3621 		}
3622 		vm_pagequeue_unlock(pq);
3623 		return;
3624 	}
3625 	vm_pagequeue_cnt_dec(pq);
3626 	vm_pagequeue_unlock(pq);
3627 	vm_page_pqbatch_submit(m, newq);
3628 }
3629 
3630 /*
3631  *	vm_page_free_prep:
3632  *
3633  *	Prepares the given page to be put on the free list,
3634  *	disassociating it from any VM object. The caller may return
3635  *	the page to the free list only if this function returns true.
3636  *
3637  *	The object must be locked.  The page must be locked if it is
3638  *	managed.
3639  */
3640 bool
3641 vm_page_free_prep(vm_page_t m)
3642 {
3643 
3644 	/*
3645 	 * Synchronize with threads that have dropped a reference to this
3646 	 * page.
3647 	 */
3648 	atomic_thread_fence_acq();
3649 
3650 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3651 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3652 		uint64_t *p;
3653 		int i;
3654 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3655 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3656 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3657 			    m, i, (uintmax_t)*p));
3658 	}
3659 #endif
3660 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3661 		KASSERT(!pmap_page_is_mapped(m),
3662 		    ("vm_page_free_prep: freeing mapped page %p", m));
3663 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3664 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3665 	} else {
3666 		KASSERT(m->a.queue == PQ_NONE,
3667 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3668 	}
3669 	VM_CNT_INC(v_tfree);
3670 
3671 	if (vm_page_sbusied(m))
3672 		panic("vm_page_free_prep: freeing shared busy page %p", m);
3673 
3674 	if (m->object != NULL) {
3675 		vm_page_object_remove(m);
3676 
3677 		/*
3678 		 * The object reference can be released without an atomic
3679 		 * operation.
3680 		 */
3681 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3682 		    m->ref_count == VPRC_OBJREF,
3683 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3684 		    m, m->ref_count));
3685 		m->object = NULL;
3686 		m->ref_count -= VPRC_OBJREF;
3687 	}
3688 
3689 	if (vm_page_xbusied(m))
3690 		vm_page_xunbusy(m);
3691 
3692 	/*
3693 	 * If fictitious remove object association and
3694 	 * return.
3695 	 */
3696 	if ((m->flags & PG_FICTITIOUS) != 0) {
3697 		KASSERT(m->ref_count == 1,
3698 		    ("fictitious page %p is referenced", m));
3699 		KASSERT(m->a.queue == PQ_NONE,
3700 		    ("fictitious page %p is queued", m));
3701 		return (false);
3702 	}
3703 
3704 	/*
3705 	 * Pages need not be dequeued before they are returned to the physical
3706 	 * memory allocator, but they must at least be marked for a deferred
3707 	 * dequeue.
3708 	 */
3709 	if ((m->oflags & VPO_UNMANAGED) == 0)
3710 		vm_page_dequeue_deferred_free(m);
3711 
3712 	m->valid = 0;
3713 	vm_page_undirty(m);
3714 
3715 	if (m->ref_count != 0)
3716 		panic("vm_page_free_prep: page %p has references", m);
3717 
3718 	/*
3719 	 * Restore the default memory attribute to the page.
3720 	 */
3721 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3722 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3723 
3724 #if VM_NRESERVLEVEL > 0
3725 	/*
3726 	 * Determine whether the page belongs to a reservation.  If the page was
3727 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3728 	 * as an optimization, we avoid the check in that case.
3729 	 */
3730 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3731 		return (false);
3732 #endif
3733 
3734 	return (true);
3735 }
3736 
3737 /*
3738  *	vm_page_free_toq:
3739  *
3740  *	Returns the given page to the free list, disassociating it
3741  *	from any VM object.
3742  *
3743  *	The object must be locked.  The page must be locked if it is
3744  *	managed.
3745  */
3746 void
3747 vm_page_free_toq(vm_page_t m)
3748 {
3749 	struct vm_domain *vmd;
3750 	uma_zone_t zone;
3751 
3752 	if (!vm_page_free_prep(m))
3753 		return;
3754 
3755 	vmd = vm_pagequeue_domain(m);
3756 	zone = vmd->vmd_pgcache[m->pool].zone;
3757 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3758 		uma_zfree(zone, m);
3759 		return;
3760 	}
3761 	vm_domain_free_lock(vmd);
3762 	vm_phys_free_pages(m, 0);
3763 	vm_domain_free_unlock(vmd);
3764 	vm_domain_freecnt_inc(vmd, 1);
3765 }
3766 
3767 /*
3768  *	vm_page_free_pages_toq:
3769  *
3770  *	Returns a list of pages to the free list, disassociating it
3771  *	from any VM object.  In other words, this is equivalent to
3772  *	calling vm_page_free_toq() for each page of a list of VM objects.
3773  *
3774  *	The objects must be locked.  The pages must be locked if it is
3775  *	managed.
3776  */
3777 void
3778 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3779 {
3780 	vm_page_t m;
3781 	int count;
3782 
3783 	if (SLIST_EMPTY(free))
3784 		return;
3785 
3786 	count = 0;
3787 	while ((m = SLIST_FIRST(free)) != NULL) {
3788 		count++;
3789 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3790 		vm_page_free_toq(m);
3791 	}
3792 
3793 	if (update_wire_count)
3794 		vm_wire_sub(count);
3795 }
3796 
3797 /*
3798  * Mark this page as wired down, preventing reclamation by the page daemon
3799  * or when the containing object is destroyed.
3800  */
3801 void
3802 vm_page_wire(vm_page_t m)
3803 {
3804 	u_int old;
3805 
3806 	KASSERT(m->object != NULL,
3807 	    ("vm_page_wire: page %p does not belong to an object", m));
3808 	if (!vm_page_busied(m) && !vm_object_busied(m->object))
3809 		VM_OBJECT_ASSERT_LOCKED(m->object);
3810 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3811 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
3812 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
3813 
3814 	old = atomic_fetchadd_int(&m->ref_count, 1);
3815 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3816 	    ("vm_page_wire: counter overflow for page %p", m));
3817 	if (VPRC_WIRE_COUNT(old) == 0)
3818 		vm_wire_add(1);
3819 }
3820 
3821 /*
3822  * Attempt to wire a mapped page following a pmap lookup of that page.
3823  * This may fail if a thread is concurrently tearing down mappings of the page.
3824  * The transient failure is acceptable because it translates to the
3825  * failure of the caller pmap_extract_and_hold(), which should be then
3826  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3827  */
3828 bool
3829 vm_page_wire_mapped(vm_page_t m)
3830 {
3831 	u_int old;
3832 
3833 	old = m->ref_count;
3834 	do {
3835 		KASSERT(old > 0,
3836 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3837 		if ((old & VPRC_BLOCKED) != 0)
3838 			return (false);
3839 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3840 
3841 	if (VPRC_WIRE_COUNT(old) == 0)
3842 		vm_wire_add(1);
3843 	return (true);
3844 }
3845 
3846 /*
3847  * Release one wiring of the specified page, potentially allowing it to be
3848  * paged out.
3849  *
3850  * Only managed pages belonging to an object can be paged out.  If the number
3851  * of wirings transitions to zero and the page is eligible for page out, then
3852  * the page is added to the specified paging queue.  If the released wiring
3853  * represented the last reference to the page, the page is freed.
3854  *
3855  * A managed page must be locked.
3856  */
3857 void
3858 vm_page_unwire(vm_page_t m, uint8_t queue)
3859 {
3860 	u_int old;
3861 	bool locked;
3862 
3863 	KASSERT(queue < PQ_COUNT,
3864 	    ("vm_page_unwire: invalid queue %u request for page %p", queue, m));
3865 
3866 	if ((m->oflags & VPO_UNMANAGED) != 0) {
3867 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
3868 			vm_page_free(m);
3869 		return;
3870 	}
3871 
3872 	/*
3873 	 * Update LRU state before releasing the wiring reference.
3874 	 * We only need to do this once since we hold the page lock.
3875 	 * Use a release store when updating the reference count to
3876 	 * synchronize with vm_page_free_prep().
3877 	 */
3878 	old = m->ref_count;
3879 	locked = false;
3880 	do {
3881 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
3882 		    ("vm_page_unwire: wire count underflow for page %p", m));
3883 		if (!locked && VPRC_WIRE_COUNT(old) == 1) {
3884 			vm_page_lock(m);
3885 			locked = true;
3886 			if (queue == PQ_ACTIVE && vm_page_queue(m) == PQ_ACTIVE)
3887 				vm_page_reference(m);
3888 			else
3889 				vm_page_mvqueue(m, queue);
3890 		}
3891 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3892 
3893 	/*
3894 	 * Release the lock only after the wiring is released, to ensure that
3895 	 * the page daemon does not encounter and dequeue the page while it is
3896 	 * still wired.
3897 	 */
3898 	if (locked)
3899 		vm_page_unlock(m);
3900 
3901 	if (VPRC_WIRE_COUNT(old) == 1) {
3902 		vm_wire_sub(1);
3903 		if (old == 1)
3904 			vm_page_free(m);
3905 	}
3906 }
3907 
3908 /*
3909  * Unwire a page without (re-)inserting it into a page queue.  It is up
3910  * to the caller to enqueue, requeue, or free the page as appropriate.
3911  * In most cases involving managed pages, vm_page_unwire() should be used
3912  * instead.
3913  */
3914 bool
3915 vm_page_unwire_noq(vm_page_t m)
3916 {
3917 	u_int old;
3918 
3919 	old = vm_page_drop(m, 1);
3920 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
3921 	    ("vm_page_unref: counter underflow for page %p", m));
3922 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3923 	    ("vm_page_unref: missing ref on fictitious page %p", m));
3924 
3925 	if (VPRC_WIRE_COUNT(old) > 1)
3926 		return (false);
3927 	vm_wire_sub(1);
3928 	return (true);
3929 }
3930 
3931 /*
3932  * Ensure that the page is in the specified page queue.  If the page is
3933  * active or being moved to the active queue, ensure that its act_count is
3934  * at least ACT_INIT but do not otherwise mess with it.  Otherwise, ensure that
3935  * the page is at the tail of its page queue.
3936  *
3937  * The page may be wired.  The caller should release its wiring reference
3938  * before releasing the page lock, otherwise the page daemon may immediately
3939  * dequeue the page.
3940  *
3941  * A managed page must be locked.
3942  */
3943 static __always_inline void
3944 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue)
3945 {
3946 
3947 	vm_page_assert_locked(m);
3948 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3949 	    ("vm_page_mvqueue: page %p is unmanaged", m));
3950 	KASSERT(m->ref_count > 0,
3951 	    ("%s: page %p does not carry any references", __func__, m));
3952 
3953 	if (vm_page_queue(m) != nqueue) {
3954 		vm_page_dequeue(m);
3955 		vm_page_enqueue(m, nqueue);
3956 	} else if (nqueue != PQ_ACTIVE) {
3957 		vm_page_requeue(m);
3958 	}
3959 
3960 	if (nqueue == PQ_ACTIVE && m->a.act_count < ACT_INIT)
3961 		m->a.act_count = ACT_INIT;
3962 }
3963 
3964 /*
3965  * Put the specified page on the active list (if appropriate).
3966  */
3967 void
3968 vm_page_activate(vm_page_t m)
3969 {
3970 
3971 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3972 		return;
3973 	vm_page_mvqueue(m, PQ_ACTIVE);
3974 }
3975 
3976 /*
3977  * Move the specified page to the tail of the inactive queue, or requeue
3978  * the page if it is already in the inactive queue.
3979  */
3980 void
3981 vm_page_deactivate(vm_page_t m)
3982 {
3983 
3984 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3985 		return;
3986 	vm_page_mvqueue(m, PQ_INACTIVE);
3987 }
3988 
3989 /*
3990  * Move the specified page close to the head of the inactive queue,
3991  * bypassing LRU.  A marker page is used to maintain FIFO ordering.
3992  * As with regular enqueues, we use a per-CPU batch queue to reduce
3993  * contention on the page queue lock.
3994  */
3995 static void
3996 _vm_page_deactivate_noreuse(vm_page_t m)
3997 {
3998 
3999 	vm_page_assert_locked(m);
4000 
4001 	if (!vm_page_inactive(m)) {
4002 		vm_page_dequeue(m);
4003 		m->a.queue = PQ_INACTIVE;
4004 	}
4005 	if ((m->a.flags & PGA_REQUEUE_HEAD) == 0)
4006 		vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
4007 	vm_page_pqbatch_submit(m, PQ_INACTIVE);
4008 }
4009 
4010 void
4011 vm_page_deactivate_noreuse(vm_page_t m)
4012 {
4013 
4014 	KASSERT(m->object != NULL,
4015 	    ("vm_page_deactivate_noreuse: page %p has no object", m));
4016 
4017 	if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_wired(m))
4018 		_vm_page_deactivate_noreuse(m);
4019 }
4020 
4021 /*
4022  * Put a page in the laundry, or requeue it if it is already there.
4023  */
4024 void
4025 vm_page_launder(vm_page_t m)
4026 {
4027 
4028 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4029 		return;
4030 	vm_page_mvqueue(m, PQ_LAUNDRY);
4031 }
4032 
4033 /*
4034  * Put a page in the PQ_UNSWAPPABLE holding queue.
4035  */
4036 void
4037 vm_page_unswappable(vm_page_t m)
4038 {
4039 
4040 	vm_page_assert_locked(m);
4041 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4042 	    ("page %p already unswappable", m));
4043 
4044 	vm_page_dequeue(m);
4045 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4046 }
4047 
4048 static void
4049 vm_page_release_toq(vm_page_t m, int flags)
4050 {
4051 
4052 	vm_page_assert_locked(m);
4053 
4054 	/*
4055 	 * Use a check of the valid bits to determine whether we should
4056 	 * accelerate reclamation of the page.  The object lock might not be
4057 	 * held here, in which case the check is racy.  At worst we will either
4058 	 * accelerate reclamation of a valid page and violate LRU, or
4059 	 * unnecessarily defer reclamation of an invalid page.
4060 	 *
4061 	 * If we were asked to not cache the page, place it near the head of the
4062 	 * inactive queue so that is reclaimed sooner.
4063 	 */
4064 	if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0)
4065 		_vm_page_deactivate_noreuse(m);
4066 	else if (vm_page_active(m))
4067 		vm_page_reference(m);
4068 	else
4069 		vm_page_mvqueue(m, PQ_INACTIVE);
4070 }
4071 
4072 /*
4073  * Unwire a page and either attempt to free it or re-add it to the page queues.
4074  */
4075 void
4076 vm_page_release(vm_page_t m, int flags)
4077 {
4078 	vm_object_t object;
4079 	u_int old;
4080 	bool locked;
4081 
4082 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4083 	    ("vm_page_release: page %p is unmanaged", m));
4084 
4085 	if ((flags & VPR_TRYFREE) != 0) {
4086 		for (;;) {
4087 			object = (vm_object_t)atomic_load_ptr(&m->object);
4088 			if (object == NULL)
4089 				break;
4090 			/* Depends on type-stability. */
4091 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) {
4092 				object = NULL;
4093 				break;
4094 			}
4095 			if (object == m->object)
4096 				break;
4097 			VM_OBJECT_WUNLOCK(object);
4098 		}
4099 		if (__predict_true(object != NULL)) {
4100 			vm_page_release_locked(m, flags);
4101 			VM_OBJECT_WUNLOCK(object);
4102 			return;
4103 		}
4104 	}
4105 
4106 	/*
4107 	 * Update LRU state before releasing the wiring reference.
4108 	 * Use a release store when updating the reference count to
4109 	 * synchronize with vm_page_free_prep().
4110 	 */
4111 	old = m->ref_count;
4112 	locked = false;
4113 	do {
4114 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4115 		    ("vm_page_unwire: wire count underflow for page %p", m));
4116 		if (!locked && VPRC_WIRE_COUNT(old) == 1) {
4117 			vm_page_lock(m);
4118 			locked = true;
4119 			vm_page_release_toq(m, flags);
4120 		}
4121 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4122 
4123 	/*
4124 	 * Release the lock only after the wiring is released, to ensure that
4125 	 * the page daemon does not encounter and dequeue the page while it is
4126 	 * still wired.
4127 	 */
4128 	if (locked)
4129 		vm_page_unlock(m);
4130 
4131 	if (VPRC_WIRE_COUNT(old) == 1) {
4132 		vm_wire_sub(1);
4133 		if (old == 1)
4134 			vm_page_free(m);
4135 	}
4136 }
4137 
4138 /* See vm_page_release(). */
4139 void
4140 vm_page_release_locked(vm_page_t m, int flags)
4141 {
4142 
4143 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4144 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4145 	    ("vm_page_release_locked: page %p is unmanaged", m));
4146 
4147 	if (vm_page_unwire_noq(m)) {
4148 		if ((flags & VPR_TRYFREE) != 0 &&
4149 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4150 		    m->dirty == 0 && !vm_page_busied(m)) {
4151 			vm_page_free(m);
4152 		} else {
4153 			vm_page_lock(m);
4154 			vm_page_release_toq(m, flags);
4155 			vm_page_unlock(m);
4156 		}
4157 	}
4158 }
4159 
4160 static bool
4161 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4162 {
4163 	u_int old;
4164 
4165 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4166 	    ("vm_page_try_blocked_op: page %p has no object", m));
4167 	KASSERT(vm_page_busied(m),
4168 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4169 	VM_OBJECT_ASSERT_LOCKED(m->object);
4170 
4171 	old = m->ref_count;
4172 	do {
4173 		KASSERT(old != 0,
4174 		    ("vm_page_try_blocked_op: page %p has no references", m));
4175 		if (VPRC_WIRE_COUNT(old) != 0)
4176 			return (false);
4177 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4178 
4179 	(op)(m);
4180 
4181 	/*
4182 	 * If the object is read-locked, new wirings may be created via an
4183 	 * object lookup.
4184 	 */
4185 	old = vm_page_drop(m, VPRC_BLOCKED);
4186 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4187 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4188 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4189 	    old, m));
4190 	return (true);
4191 }
4192 
4193 /*
4194  * Atomically check for wirings and remove all mappings of the page.
4195  */
4196 bool
4197 vm_page_try_remove_all(vm_page_t m)
4198 {
4199 
4200 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4201 }
4202 
4203 /*
4204  * Atomically check for wirings and remove all writeable mappings of the page.
4205  */
4206 bool
4207 vm_page_try_remove_write(vm_page_t m)
4208 {
4209 
4210 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4211 }
4212 
4213 /*
4214  * vm_page_advise
4215  *
4216  * 	Apply the specified advice to the given page.
4217  *
4218  *	The object and page must be locked.
4219  */
4220 void
4221 vm_page_advise(vm_page_t m, int advice)
4222 {
4223 
4224 	vm_page_assert_locked(m);
4225 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4226 	if (advice == MADV_FREE)
4227 		/*
4228 		 * Mark the page clean.  This will allow the page to be freed
4229 		 * without first paging it out.  MADV_FREE pages are often
4230 		 * quickly reused by malloc(3), so we do not do anything that
4231 		 * would result in a page fault on a later access.
4232 		 */
4233 		vm_page_undirty(m);
4234 	else if (advice != MADV_DONTNEED) {
4235 		if (advice == MADV_WILLNEED)
4236 			vm_page_activate(m);
4237 		return;
4238 	}
4239 
4240 	/*
4241 	 * Clear any references to the page.  Otherwise, the page daemon will
4242 	 * immediately reactivate the page.
4243 	 */
4244 	vm_page_aflag_clear(m, PGA_REFERENCED);
4245 
4246 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4247 		vm_page_dirty(m);
4248 
4249 	/*
4250 	 * Place clean pages near the head of the inactive queue rather than
4251 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4252 	 * the page will be reused quickly.  Dirty pages not already in the
4253 	 * laundry are moved there.
4254 	 */
4255 	if (m->dirty == 0)
4256 		vm_page_deactivate_noreuse(m);
4257 	else if (!vm_page_in_laundry(m))
4258 		vm_page_launder(m);
4259 }
4260 
4261 static inline int
4262 vm_page_grab_pflags(int allocflags)
4263 {
4264 	int pflags;
4265 
4266 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4267 	    (allocflags & VM_ALLOC_WIRED) != 0,
4268 	    ("vm_page_grab_pflags: the pages must be busied or wired"));
4269 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4270 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4271 	    ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4272 	    "mismatch"));
4273 	pflags = allocflags &
4274 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4275 	    VM_ALLOC_NOBUSY);
4276 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4277 		pflags |= VM_ALLOC_WAITFAIL;
4278 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4279 		pflags |= VM_ALLOC_SBUSY;
4280 
4281 	return (pflags);
4282 }
4283 
4284 /*
4285  * Grab a page, waiting until we are waken up due to the page
4286  * changing state.  We keep on waiting, if the page continues
4287  * to be in the object.  If the page doesn't exist, first allocate it
4288  * and then conditionally zero it.
4289  *
4290  * This routine may sleep.
4291  *
4292  * The object must be locked on entry.  The lock will, however, be released
4293  * and reacquired if the routine sleeps.
4294  */
4295 vm_page_t
4296 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4297 {
4298 	vm_page_t m;
4299 	int pflags;
4300 
4301 	VM_OBJECT_ASSERT_WLOCKED(object);
4302 	pflags = vm_page_grab_pflags(allocflags);
4303 retrylookup:
4304 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4305 		if (!vm_page_acquire_flags(m, allocflags)) {
4306 			if (vm_page_busy_sleep_flags(object, m, "pgrbwt",
4307 			    allocflags))
4308 				goto retrylookup;
4309 			return (NULL);
4310 		}
4311 		goto out;
4312 	}
4313 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4314 		return (NULL);
4315 	m = vm_page_alloc(object, pindex, pflags);
4316 	if (m == NULL) {
4317 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4318 			return (NULL);
4319 		goto retrylookup;
4320 	}
4321 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4322 		pmap_zero_page(m);
4323 
4324 out:
4325 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4326 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4327 			vm_page_sunbusy(m);
4328 		else
4329 			vm_page_xunbusy(m);
4330 	}
4331 	return (m);
4332 }
4333 
4334 /*
4335  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4336  * their pager are zero filled and validated.
4337  */
4338 int
4339 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4340 {
4341 	vm_page_t m;
4342 	bool sleep, xbusy;
4343 	int pflags;
4344 	int rv;
4345 
4346 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4347 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4348 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4349 	KASSERT((allocflags &
4350 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4351 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4352 	VM_OBJECT_ASSERT_WLOCKED(object);
4353 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4354 	pflags |= VM_ALLOC_WAITFAIL;
4355 
4356 retrylookup:
4357 	xbusy = false;
4358 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4359 		/*
4360 		 * If the page is fully valid it can only become invalid
4361 		 * with the object lock held.  If it is not valid it can
4362 		 * become valid with the busy lock held.  Therefore, we
4363 		 * may unnecessarily lock the exclusive busy here if we
4364 		 * race with I/O completion not using the object lock.
4365 		 * However, we will not end up with an invalid page and a
4366 		 * shared lock.
4367 		 */
4368 		if (!vm_page_all_valid(m) ||
4369 		    (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4370 			sleep = !vm_page_tryxbusy(m);
4371 			xbusy = true;
4372 		} else
4373 			sleep = !vm_page_trysbusy(m);
4374 		if (sleep) {
4375 			(void)vm_page_busy_sleep_flags(object, m, "pgrbwt",
4376 			    allocflags);
4377 			goto retrylookup;
4378 		}
4379 		if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4380 		   !vm_page_all_valid(m)) {
4381 			if (xbusy)
4382 				vm_page_xunbusy(m);
4383 			else
4384 				vm_page_sunbusy(m);
4385 			*mp = NULL;
4386 			return (VM_PAGER_FAIL);
4387 		}
4388 		if ((allocflags & VM_ALLOC_WIRED) != 0)
4389 			vm_page_wire(m);
4390 		if (vm_page_all_valid(m))
4391 			goto out;
4392 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4393 		*mp = NULL;
4394 		return (VM_PAGER_FAIL);
4395 	} else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4396 		xbusy = true;
4397 	} else {
4398 		goto retrylookup;
4399 	}
4400 
4401 	vm_page_assert_xbusied(m);
4402 	MPASS(xbusy);
4403 	if (vm_pager_has_page(object, pindex, NULL, NULL)) {
4404 		rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
4405 		if (rv != VM_PAGER_OK) {
4406 			if (allocflags & VM_ALLOC_WIRED)
4407 				vm_page_unwire_noq(m);
4408 			vm_page_free(m);
4409 			*mp = NULL;
4410 			return (rv);
4411 		}
4412 		MPASS(vm_page_all_valid(m));
4413 	} else {
4414 		vm_page_zero_invalid(m, TRUE);
4415 	}
4416 out:
4417 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4418 		if (xbusy)
4419 			vm_page_xunbusy(m);
4420 		else
4421 			vm_page_sunbusy(m);
4422 	}
4423 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4424 		vm_page_busy_downgrade(m);
4425 	*mp = m;
4426 	return (VM_PAGER_OK);
4427 }
4428 
4429 /*
4430  * Return the specified range of pages from the given object.  For each
4431  * page offset within the range, if a page already exists within the object
4432  * at that offset and it is busy, then wait for it to change state.  If,
4433  * instead, the page doesn't exist, then allocate it.
4434  *
4435  * The caller must always specify an allocation class.
4436  *
4437  * allocation classes:
4438  *	VM_ALLOC_NORMAL		normal process request
4439  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4440  *
4441  * The caller must always specify that the pages are to be busied and/or
4442  * wired.
4443  *
4444  * optional allocation flags:
4445  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4446  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4447  *	VM_ALLOC_NOWAIT		do not sleep
4448  *	VM_ALLOC_SBUSY		set page to sbusy state
4449  *	VM_ALLOC_WIRED		wire the pages
4450  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4451  *
4452  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4453  * may return a partial prefix of the requested range.
4454  */
4455 int
4456 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4457     vm_page_t *ma, int count)
4458 {
4459 	vm_page_t m, mpred;
4460 	int pflags;
4461 	int i;
4462 
4463 	VM_OBJECT_ASSERT_WLOCKED(object);
4464 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4465 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4466 
4467 	pflags = vm_page_grab_pflags(allocflags);
4468 	if (count == 0)
4469 		return (0);
4470 
4471 	i = 0;
4472 retrylookup:
4473 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4474 	if (m == NULL || m->pindex != pindex + i) {
4475 		mpred = m;
4476 		m = NULL;
4477 	} else
4478 		mpred = TAILQ_PREV(m, pglist, listq);
4479 	for (; i < count; i++) {
4480 		if (m != NULL) {
4481 			if (!vm_page_acquire_flags(m, allocflags)) {
4482 				if (vm_page_busy_sleep_flags(object, m,
4483 				    "grbmaw", allocflags))
4484 					goto retrylookup;
4485 				break;
4486 			}
4487 		} else {
4488 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4489 				break;
4490 			m = vm_page_alloc_after(object, pindex + i,
4491 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4492 			if (m == NULL) {
4493 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4494 					break;
4495 				goto retrylookup;
4496 			}
4497 		}
4498 		if (vm_page_none_valid(m) &&
4499 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4500 			if ((m->flags & PG_ZERO) == 0)
4501 				pmap_zero_page(m);
4502 			vm_page_valid(m);
4503 		}
4504 		if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4505 			if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4506 				vm_page_sunbusy(m);
4507 			else
4508 				vm_page_xunbusy(m);
4509 		}
4510 		ma[i] = mpred = m;
4511 		m = vm_page_next(m);
4512 	}
4513 	return (i);
4514 }
4515 
4516 /*
4517  * Mapping function for valid or dirty bits in a page.
4518  *
4519  * Inputs are required to range within a page.
4520  */
4521 vm_page_bits_t
4522 vm_page_bits(int base, int size)
4523 {
4524 	int first_bit;
4525 	int last_bit;
4526 
4527 	KASSERT(
4528 	    base + size <= PAGE_SIZE,
4529 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4530 	);
4531 
4532 	if (size == 0)		/* handle degenerate case */
4533 		return (0);
4534 
4535 	first_bit = base >> DEV_BSHIFT;
4536 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4537 
4538 	return (((vm_page_bits_t)2 << last_bit) -
4539 	    ((vm_page_bits_t)1 << first_bit));
4540 }
4541 
4542 void
4543 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4544 {
4545 
4546 #if PAGE_SIZE == 32768
4547 	atomic_set_64((uint64_t *)bits, set);
4548 #elif PAGE_SIZE == 16384
4549 	atomic_set_32((uint32_t *)bits, set);
4550 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4551 	atomic_set_16((uint16_t *)bits, set);
4552 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4553 	atomic_set_8((uint8_t *)bits, set);
4554 #else		/* PAGE_SIZE <= 8192 */
4555 	uintptr_t addr;
4556 	int shift;
4557 
4558 	addr = (uintptr_t)bits;
4559 	/*
4560 	 * Use a trick to perform a 32-bit atomic on the
4561 	 * containing aligned word, to not depend on the existence
4562 	 * of atomic_{set, clear}_{8, 16}.
4563 	 */
4564 	shift = addr & (sizeof(uint32_t) - 1);
4565 #if BYTE_ORDER == BIG_ENDIAN
4566 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4567 #else
4568 	shift *= NBBY;
4569 #endif
4570 	addr &= ~(sizeof(uint32_t) - 1);
4571 	atomic_set_32((uint32_t *)addr, set << shift);
4572 #endif		/* PAGE_SIZE */
4573 }
4574 
4575 static inline void
4576 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4577 {
4578 
4579 #if PAGE_SIZE == 32768
4580 	atomic_clear_64((uint64_t *)bits, clear);
4581 #elif PAGE_SIZE == 16384
4582 	atomic_clear_32((uint32_t *)bits, clear);
4583 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4584 	atomic_clear_16((uint16_t *)bits, clear);
4585 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4586 	atomic_clear_8((uint8_t *)bits, clear);
4587 #else		/* PAGE_SIZE <= 8192 */
4588 	uintptr_t addr;
4589 	int shift;
4590 
4591 	addr = (uintptr_t)bits;
4592 	/*
4593 	 * Use a trick to perform a 32-bit atomic on the
4594 	 * containing aligned word, to not depend on the existence
4595 	 * of atomic_{set, clear}_{8, 16}.
4596 	 */
4597 	shift = addr & (sizeof(uint32_t) - 1);
4598 #if BYTE_ORDER == BIG_ENDIAN
4599 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4600 #else
4601 	shift *= NBBY;
4602 #endif
4603 	addr &= ~(sizeof(uint32_t) - 1);
4604 	atomic_clear_32((uint32_t *)addr, clear << shift);
4605 #endif		/* PAGE_SIZE */
4606 }
4607 
4608 /*
4609  *	vm_page_set_valid_range:
4610  *
4611  *	Sets portions of a page valid.  The arguments are expected
4612  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4613  *	of any partial chunks touched by the range.  The invalid portion of
4614  *	such chunks will be zeroed.
4615  *
4616  *	(base + size) must be less then or equal to PAGE_SIZE.
4617  */
4618 void
4619 vm_page_set_valid_range(vm_page_t m, int base, int size)
4620 {
4621 	int endoff, frag;
4622 	vm_page_bits_t pagebits;
4623 
4624 	vm_page_assert_busied(m);
4625 	if (size == 0)	/* handle degenerate case */
4626 		return;
4627 
4628 	/*
4629 	 * If the base is not DEV_BSIZE aligned and the valid
4630 	 * bit is clear, we have to zero out a portion of the
4631 	 * first block.
4632 	 */
4633 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4634 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4635 		pmap_zero_page_area(m, frag, base - frag);
4636 
4637 	/*
4638 	 * If the ending offset is not DEV_BSIZE aligned and the
4639 	 * valid bit is clear, we have to zero out a portion of
4640 	 * the last block.
4641 	 */
4642 	endoff = base + size;
4643 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4644 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4645 		pmap_zero_page_area(m, endoff,
4646 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4647 
4648 	/*
4649 	 * Assert that no previously invalid block that is now being validated
4650 	 * is already dirty.
4651 	 */
4652 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4653 	    ("vm_page_set_valid_range: page %p is dirty", m));
4654 
4655 	/*
4656 	 * Set valid bits inclusive of any overlap.
4657 	 */
4658 	pagebits = vm_page_bits(base, size);
4659 	if (vm_page_xbusied(m))
4660 		m->valid |= pagebits;
4661 	else
4662 		vm_page_bits_set(m, &m->valid, pagebits);
4663 }
4664 
4665 /*
4666  * Clear the given bits from the specified page's dirty field.
4667  */
4668 static __inline void
4669 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4670 {
4671 
4672 	vm_page_assert_busied(m);
4673 
4674 	/*
4675 	 * If the page is xbusied and not write mapped we are the
4676 	 * only thread that can modify dirty bits.  Otherwise, The pmap
4677 	 * layer can call vm_page_dirty() without holding a distinguished
4678 	 * lock.  The combination of page busy and atomic operations
4679 	 * suffice to guarantee consistency of the page dirty field.
4680 	 */
4681 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4682 		m->dirty &= ~pagebits;
4683 	else
4684 		vm_page_bits_clear(m, &m->dirty, pagebits);
4685 }
4686 
4687 /*
4688  *	vm_page_set_validclean:
4689  *
4690  *	Sets portions of a page valid and clean.  The arguments are expected
4691  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4692  *	of any partial chunks touched by the range.  The invalid portion of
4693  *	such chunks will be zero'd.
4694  *
4695  *	(base + size) must be less then or equal to PAGE_SIZE.
4696  */
4697 void
4698 vm_page_set_validclean(vm_page_t m, int base, int size)
4699 {
4700 	vm_page_bits_t oldvalid, pagebits;
4701 	int endoff, frag;
4702 
4703 	vm_page_assert_busied(m);
4704 	if (size == 0)	/* handle degenerate case */
4705 		return;
4706 
4707 	/*
4708 	 * If the base is not DEV_BSIZE aligned and the valid
4709 	 * bit is clear, we have to zero out a portion of the
4710 	 * first block.
4711 	 */
4712 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4713 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4714 		pmap_zero_page_area(m, frag, base - frag);
4715 
4716 	/*
4717 	 * If the ending offset is not DEV_BSIZE aligned and the
4718 	 * valid bit is clear, we have to zero out a portion of
4719 	 * the last block.
4720 	 */
4721 	endoff = base + size;
4722 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4723 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4724 		pmap_zero_page_area(m, endoff,
4725 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4726 
4727 	/*
4728 	 * Set valid, clear dirty bits.  If validating the entire
4729 	 * page we can safely clear the pmap modify bit.  We also
4730 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
4731 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4732 	 * be set again.
4733 	 *
4734 	 * We set valid bits inclusive of any overlap, but we can only
4735 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4736 	 * the range.
4737 	 */
4738 	oldvalid = m->valid;
4739 	pagebits = vm_page_bits(base, size);
4740 	if (vm_page_xbusied(m))
4741 		m->valid |= pagebits;
4742 	else
4743 		vm_page_bits_set(m, &m->valid, pagebits);
4744 #if 0	/* NOT YET */
4745 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4746 		frag = DEV_BSIZE - frag;
4747 		base += frag;
4748 		size -= frag;
4749 		if (size < 0)
4750 			size = 0;
4751 	}
4752 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4753 #endif
4754 	if (base == 0 && size == PAGE_SIZE) {
4755 		/*
4756 		 * The page can only be modified within the pmap if it is
4757 		 * mapped, and it can only be mapped if it was previously
4758 		 * fully valid.
4759 		 */
4760 		if (oldvalid == VM_PAGE_BITS_ALL)
4761 			/*
4762 			 * Perform the pmap_clear_modify() first.  Otherwise,
4763 			 * a concurrent pmap operation, such as
4764 			 * pmap_protect(), could clear a modification in the
4765 			 * pmap and set the dirty field on the page before
4766 			 * pmap_clear_modify() had begun and after the dirty
4767 			 * field was cleared here.
4768 			 */
4769 			pmap_clear_modify(m);
4770 		m->dirty = 0;
4771 		vm_page_aflag_clear(m, PGA_NOSYNC);
4772 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
4773 		m->dirty &= ~pagebits;
4774 	else
4775 		vm_page_clear_dirty_mask(m, pagebits);
4776 }
4777 
4778 void
4779 vm_page_clear_dirty(vm_page_t m, int base, int size)
4780 {
4781 
4782 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4783 }
4784 
4785 /*
4786  *	vm_page_set_invalid:
4787  *
4788  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4789  *	valid and dirty bits for the effected areas are cleared.
4790  */
4791 void
4792 vm_page_set_invalid(vm_page_t m, int base, int size)
4793 {
4794 	vm_page_bits_t bits;
4795 	vm_object_t object;
4796 
4797 	/*
4798 	 * The object lock is required so that pages can't be mapped
4799 	 * read-only while we're in the process of invalidating them.
4800 	 */
4801 	object = m->object;
4802 	VM_OBJECT_ASSERT_WLOCKED(object);
4803 	vm_page_assert_busied(m);
4804 
4805 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4806 	    size >= object->un_pager.vnp.vnp_size)
4807 		bits = VM_PAGE_BITS_ALL;
4808 	else
4809 		bits = vm_page_bits(base, size);
4810 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
4811 		pmap_remove_all(m);
4812 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
4813 	    !pmap_page_is_mapped(m),
4814 	    ("vm_page_set_invalid: page %p is mapped", m));
4815 	if (vm_page_xbusied(m)) {
4816 		m->valid &= ~bits;
4817 		m->dirty &= ~bits;
4818 	} else {
4819 		vm_page_bits_clear(m, &m->valid, bits);
4820 		vm_page_bits_clear(m, &m->dirty, bits);
4821 	}
4822 }
4823 
4824 /*
4825  *	vm_page_invalid:
4826  *
4827  *	Invalidates the entire page.  The page must be busy, unmapped, and
4828  *	the enclosing object must be locked.  The object locks protects
4829  *	against concurrent read-only pmap enter which is done without
4830  *	busy.
4831  */
4832 void
4833 vm_page_invalid(vm_page_t m)
4834 {
4835 
4836 	vm_page_assert_busied(m);
4837 	VM_OBJECT_ASSERT_LOCKED(m->object);
4838 	MPASS(!pmap_page_is_mapped(m));
4839 
4840 	if (vm_page_xbusied(m))
4841 		m->valid = 0;
4842 	else
4843 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
4844 }
4845 
4846 /*
4847  * vm_page_zero_invalid()
4848  *
4849  *	The kernel assumes that the invalid portions of a page contain
4850  *	garbage, but such pages can be mapped into memory by user code.
4851  *	When this occurs, we must zero out the non-valid portions of the
4852  *	page so user code sees what it expects.
4853  *
4854  *	Pages are most often semi-valid when the end of a file is mapped
4855  *	into memory and the file's size is not page aligned.
4856  */
4857 void
4858 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4859 {
4860 	int b;
4861 	int i;
4862 
4863 	/*
4864 	 * Scan the valid bits looking for invalid sections that
4865 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4866 	 * valid bit may be set ) have already been zeroed by
4867 	 * vm_page_set_validclean().
4868 	 */
4869 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4870 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4871 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4872 			if (i > b) {
4873 				pmap_zero_page_area(m,
4874 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4875 			}
4876 			b = i + 1;
4877 		}
4878 	}
4879 
4880 	/*
4881 	 * setvalid is TRUE when we can safely set the zero'd areas
4882 	 * as being valid.  We can do this if there are no cache consistancy
4883 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
4884 	 */
4885 	if (setvalid)
4886 		vm_page_valid(m);
4887 }
4888 
4889 /*
4890  *	vm_page_is_valid:
4891  *
4892  *	Is (partial) page valid?  Note that the case where size == 0
4893  *	will return FALSE in the degenerate case where the page is
4894  *	entirely invalid, and TRUE otherwise.
4895  *
4896  *	Some callers envoke this routine without the busy lock held and
4897  *	handle races via higher level locks.  Typical callers should
4898  *	hold a busy lock to prevent invalidation.
4899  */
4900 int
4901 vm_page_is_valid(vm_page_t m, int base, int size)
4902 {
4903 	vm_page_bits_t bits;
4904 
4905 	bits = vm_page_bits(base, size);
4906 	return (m->valid != 0 && (m->valid & bits) == bits);
4907 }
4908 
4909 /*
4910  * Returns true if all of the specified predicates are true for the entire
4911  * (super)page and false otherwise.
4912  */
4913 bool
4914 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
4915 {
4916 	vm_object_t object;
4917 	int i, npages;
4918 
4919 	object = m->object;
4920 	if (skip_m != NULL && skip_m->object != object)
4921 		return (false);
4922 	VM_OBJECT_ASSERT_LOCKED(object);
4923 	npages = atop(pagesizes[m->psind]);
4924 
4925 	/*
4926 	 * The physically contiguous pages that make up a superpage, i.e., a
4927 	 * page with a page size index ("psind") greater than zero, will
4928 	 * occupy adjacent entries in vm_page_array[].
4929 	 */
4930 	for (i = 0; i < npages; i++) {
4931 		/* Always test object consistency, including "skip_m". */
4932 		if (m[i].object != object)
4933 			return (false);
4934 		if (&m[i] == skip_m)
4935 			continue;
4936 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
4937 			return (false);
4938 		if ((flags & PS_ALL_DIRTY) != 0) {
4939 			/*
4940 			 * Calling vm_page_test_dirty() or pmap_is_modified()
4941 			 * might stop this case from spuriously returning
4942 			 * "false".  However, that would require a write lock
4943 			 * on the object containing "m[i]".
4944 			 */
4945 			if (m[i].dirty != VM_PAGE_BITS_ALL)
4946 				return (false);
4947 		}
4948 		if ((flags & PS_ALL_VALID) != 0 &&
4949 		    m[i].valid != VM_PAGE_BITS_ALL)
4950 			return (false);
4951 	}
4952 	return (true);
4953 }
4954 
4955 /*
4956  * Set the page's dirty bits if the page is modified.
4957  */
4958 void
4959 vm_page_test_dirty(vm_page_t m)
4960 {
4961 
4962 	vm_page_assert_busied(m);
4963 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
4964 		vm_page_dirty(m);
4965 }
4966 
4967 void
4968 vm_page_valid(vm_page_t m)
4969 {
4970 
4971 	vm_page_assert_busied(m);
4972 	if (vm_page_xbusied(m))
4973 		m->valid = VM_PAGE_BITS_ALL;
4974 	else
4975 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
4976 }
4977 
4978 void
4979 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
4980 {
4981 
4982 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
4983 }
4984 
4985 void
4986 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
4987 {
4988 
4989 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
4990 }
4991 
4992 int
4993 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
4994 {
4995 
4996 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
4997 }
4998 
4999 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5000 void
5001 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5002 {
5003 
5004 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5005 }
5006 
5007 void
5008 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5009 {
5010 
5011 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5012 }
5013 #endif
5014 
5015 #ifdef INVARIANTS
5016 void
5017 vm_page_object_busy_assert(vm_page_t m)
5018 {
5019 
5020 	/*
5021 	 * Certain of the page's fields may only be modified by the
5022 	 * holder of a page or object busy.
5023 	 */
5024 	if (m->object != NULL && !vm_page_busied(m))
5025 		VM_OBJECT_ASSERT_BUSY(m->object);
5026 }
5027 
5028 void
5029 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5030 {
5031 
5032 	if ((bits & PGA_WRITEABLE) == 0)
5033 		return;
5034 
5035 	/*
5036 	 * The PGA_WRITEABLE flag can only be set if the page is
5037 	 * managed, is exclusively busied or the object is locked.
5038 	 * Currently, this flag is only set by pmap_enter().
5039 	 */
5040 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5041 	    ("PGA_WRITEABLE on unmanaged page"));
5042 	if (!vm_page_xbusied(m))
5043 		VM_OBJECT_ASSERT_BUSY(m->object);
5044 }
5045 #endif
5046 
5047 #include "opt_ddb.h"
5048 #ifdef DDB
5049 #include <sys/kernel.h>
5050 
5051 #include <ddb/ddb.h>
5052 
5053 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5054 {
5055 
5056 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5057 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5058 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5059 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5060 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5061 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5062 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5063 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5064 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5065 }
5066 
5067 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5068 {
5069 	int dom;
5070 
5071 	db_printf("pq_free %d\n", vm_free_count());
5072 	for (dom = 0; dom < vm_ndomains; dom++) {
5073 		db_printf(
5074     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5075 		    dom,
5076 		    vm_dom[dom].vmd_page_count,
5077 		    vm_dom[dom].vmd_free_count,
5078 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5079 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5080 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5081 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5082 	}
5083 }
5084 
5085 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5086 {
5087 	vm_page_t m;
5088 	boolean_t phys, virt;
5089 
5090 	if (!have_addr) {
5091 		db_printf("show pginfo addr\n");
5092 		return;
5093 	}
5094 
5095 	phys = strchr(modif, 'p') != NULL;
5096 	virt = strchr(modif, 'v') != NULL;
5097 	if (virt)
5098 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5099 	else if (phys)
5100 		m = PHYS_TO_VM_PAGE(addr);
5101 	else
5102 		m = (vm_page_t)addr;
5103 	db_printf(
5104     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
5105     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5106 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5107 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5108 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5109 }
5110 #endif /* DDB */
5111