xref: /freebsd/sys/vm/vm_page.c (revision b5f20658ee91b62296384ec68cd1fc82a4fbe4bb)
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 	/* Deferred free of swap space. */
1588 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1589 		vm_pager_page_unswapped(m);
1590 
1591 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1592 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1593 
1594 	/*
1595 	 * Now remove from the object's list of backed pages.
1596 	 */
1597 	TAILQ_REMOVE(&object->memq, m, listq);
1598 
1599 	/*
1600 	 * And show that the object has one fewer resident page.
1601 	 */
1602 	object->resident_page_count--;
1603 
1604 	/*
1605 	 * The vnode may now be recycled.
1606 	 */
1607 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1608 		vdrop(object->handle);
1609 }
1610 
1611 /*
1612  *	vm_page_remove:
1613  *
1614  *	Removes the specified page from its containing object, but does not
1615  *	invalidate any backing storage.  Returns true if the object's reference
1616  *	was the last reference to the page, and false otherwise.
1617  *
1618  *	The object must be locked.
1619  */
1620 bool
1621 vm_page_remove(vm_page_t m)
1622 {
1623 
1624 	vm_page_object_remove(m);
1625 	m->object = NULL;
1626 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1627 }
1628 
1629 /*
1630  *	vm_page_lookup:
1631  *
1632  *	Returns the page associated with the object/offset
1633  *	pair specified; if none is found, NULL is returned.
1634  *
1635  *	The object must be locked.
1636  */
1637 vm_page_t
1638 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1639 {
1640 
1641 	VM_OBJECT_ASSERT_LOCKED(object);
1642 	return (vm_radix_lookup(&object->rtree, pindex));
1643 }
1644 
1645 /*
1646  *	vm_page_find_least:
1647  *
1648  *	Returns the page associated with the object with least pindex
1649  *	greater than or equal to the parameter pindex, or NULL.
1650  *
1651  *	The object must be locked.
1652  */
1653 vm_page_t
1654 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1655 {
1656 	vm_page_t m;
1657 
1658 	VM_OBJECT_ASSERT_LOCKED(object);
1659 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1660 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1661 	return (m);
1662 }
1663 
1664 /*
1665  * Returns the given page's successor (by pindex) within the object if it is
1666  * resident; if none is found, NULL is returned.
1667  *
1668  * The object must be locked.
1669  */
1670 vm_page_t
1671 vm_page_next(vm_page_t m)
1672 {
1673 	vm_page_t next;
1674 
1675 	VM_OBJECT_ASSERT_LOCKED(m->object);
1676 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1677 		MPASS(next->object == m->object);
1678 		if (next->pindex != m->pindex + 1)
1679 			next = NULL;
1680 	}
1681 	return (next);
1682 }
1683 
1684 /*
1685  * Returns the given page's predecessor (by pindex) within the object if it is
1686  * resident; if none is found, NULL is returned.
1687  *
1688  * The object must be locked.
1689  */
1690 vm_page_t
1691 vm_page_prev(vm_page_t m)
1692 {
1693 	vm_page_t prev;
1694 
1695 	VM_OBJECT_ASSERT_LOCKED(m->object);
1696 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1697 		MPASS(prev->object == m->object);
1698 		if (prev->pindex != m->pindex - 1)
1699 			prev = NULL;
1700 	}
1701 	return (prev);
1702 }
1703 
1704 /*
1705  * Uses the page mnew as a replacement for an existing page at index
1706  * pindex which must be already present in the object.
1707  */
1708 vm_page_t
1709 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1710 {
1711 	vm_page_t mold;
1712 
1713 	VM_OBJECT_ASSERT_WLOCKED(object);
1714 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1715 	    ("vm_page_replace: page %p already in object", mnew));
1716 
1717 	/*
1718 	 * This function mostly follows vm_page_insert() and
1719 	 * vm_page_remove() without the radix, object count and vnode
1720 	 * dance.  Double check such functions for more comments.
1721 	 */
1722 
1723 	mnew->object = object;
1724 	mnew->pindex = pindex;
1725 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1726 	mold = vm_radix_replace(&object->rtree, mnew);
1727 
1728 	/* Keep the resident page list in sorted order. */
1729 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1730 	TAILQ_REMOVE(&object->memq, mold, listq);
1731 
1732 	mold->object = NULL;
1733 	atomic_clear_int(&mold->ref_count, VPRC_OBJREF);
1734 	vm_page_xunbusy(mold);
1735 
1736 	/*
1737 	 * The object's resident_page_count does not change because we have
1738 	 * swapped one page for another, but the generation count should
1739 	 * change if the page is dirty.
1740 	 */
1741 	if (pmap_page_is_write_mapped(mnew))
1742 		vm_object_set_writeable_dirty(object);
1743 	return (mold);
1744 }
1745 
1746 /*
1747  *	vm_page_rename:
1748  *
1749  *	Move the given memory entry from its
1750  *	current object to the specified target object/offset.
1751  *
1752  *	Note: swap associated with the page must be invalidated by the move.  We
1753  *	      have to do this for several reasons:  (1) we aren't freeing the
1754  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1755  *	      moving the page from object A to B, and will then later move
1756  *	      the backing store from A to B and we can't have a conflict.
1757  *
1758  *	Note: we *always* dirty the page.  It is necessary both for the
1759  *	      fact that we moved it, and because we may be invalidating
1760  *	      swap.
1761  *
1762  *	The objects must be locked.
1763  */
1764 int
1765 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1766 {
1767 	vm_page_t mpred;
1768 	vm_pindex_t opidx;
1769 
1770 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1771 
1772 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1773 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1774 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1775 	    ("vm_page_rename: pindex already renamed"));
1776 
1777 	/*
1778 	 * Create a custom version of vm_page_insert() which does not depend
1779 	 * by m_prev and can cheat on the implementation aspects of the
1780 	 * function.
1781 	 */
1782 	opidx = m->pindex;
1783 	m->pindex = new_pindex;
1784 	if (vm_radix_insert(&new_object->rtree, m)) {
1785 		m->pindex = opidx;
1786 		return (1);
1787 	}
1788 
1789 	/*
1790 	 * The operation cannot fail anymore.  The removal must happen before
1791 	 * the listq iterator is tainted.
1792 	 */
1793 	m->pindex = opidx;
1794 	vm_page_object_remove(m);
1795 
1796 	/* Return back to the new pindex to complete vm_page_insert(). */
1797 	m->pindex = new_pindex;
1798 	m->object = new_object;
1799 
1800 	vm_page_insert_radixdone(m, new_object, mpred);
1801 	vm_page_dirty(m);
1802 	return (0);
1803 }
1804 
1805 /*
1806  *	vm_page_alloc:
1807  *
1808  *	Allocate and return a page that is associated with the specified
1809  *	object and offset pair.  By default, this page is exclusive busied.
1810  *
1811  *	The caller must always specify an allocation class.
1812  *
1813  *	allocation classes:
1814  *	VM_ALLOC_NORMAL		normal process request
1815  *	VM_ALLOC_SYSTEM		system *really* needs a page
1816  *	VM_ALLOC_INTERRUPT	interrupt time request
1817  *
1818  *	optional allocation flags:
1819  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1820  *				intends to allocate
1821  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1822  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1823  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1824  *				should not be exclusive busy
1825  *	VM_ALLOC_SBUSY		shared busy the allocated page
1826  *	VM_ALLOC_WIRED		wire the allocated page
1827  *	VM_ALLOC_ZERO		prefer a zeroed page
1828  */
1829 vm_page_t
1830 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1831 {
1832 
1833 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1834 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1835 }
1836 
1837 vm_page_t
1838 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1839     int req)
1840 {
1841 
1842 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1843 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1844 	    NULL));
1845 }
1846 
1847 /*
1848  * Allocate a page in the specified object with the given page index.  To
1849  * optimize insertion of the page into the object, the caller must also specifiy
1850  * the resident page in the object with largest index smaller than the given
1851  * page index, or NULL if no such page exists.
1852  */
1853 vm_page_t
1854 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1855     int req, vm_page_t mpred)
1856 {
1857 	struct vm_domainset_iter di;
1858 	vm_page_t m;
1859 	int domain;
1860 
1861 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1862 	do {
1863 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1864 		    mpred);
1865 		if (m != NULL)
1866 			break;
1867 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1868 
1869 	return (m);
1870 }
1871 
1872 /*
1873  * Returns true if the number of free pages exceeds the minimum
1874  * for the request class and false otherwise.
1875  */
1876 static int
1877 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1878 {
1879 	u_int limit, old, new;
1880 
1881 	if (req_class == VM_ALLOC_INTERRUPT)
1882 		limit = 0;
1883 	else if (req_class == VM_ALLOC_SYSTEM)
1884 		limit = vmd->vmd_interrupt_free_min;
1885 	else
1886 		limit = vmd->vmd_free_reserved;
1887 
1888 	/*
1889 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1890 	 */
1891 	limit += npages;
1892 	old = vmd->vmd_free_count;
1893 	do {
1894 		if (old < limit)
1895 			return (0);
1896 		new = old - npages;
1897 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1898 
1899 	/* Wake the page daemon if we've crossed the threshold. */
1900 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1901 		pagedaemon_wakeup(vmd->vmd_domain);
1902 
1903 	/* Only update bitsets on transitions. */
1904 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1905 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1906 		vm_domain_set(vmd);
1907 
1908 	return (1);
1909 }
1910 
1911 int
1912 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1913 {
1914 	int req_class;
1915 
1916 	/*
1917 	 * The page daemon is allowed to dig deeper into the free page list.
1918 	 */
1919 	req_class = req & VM_ALLOC_CLASS_MASK;
1920 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1921 		req_class = VM_ALLOC_SYSTEM;
1922 	return (_vm_domain_allocate(vmd, req_class, npages));
1923 }
1924 
1925 vm_page_t
1926 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1927     int req, vm_page_t mpred)
1928 {
1929 	struct vm_domain *vmd;
1930 	vm_page_t m;
1931 	int flags, pool;
1932 
1933 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1934 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1935 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1936 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1937 	    ("inconsistent object(%p)/req(%x)", object, req));
1938 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
1939 	    ("Can't sleep and retry object insertion."));
1940 	KASSERT(mpred == NULL || mpred->pindex < pindex,
1941 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
1942 	    (uintmax_t)pindex));
1943 	if (object != NULL)
1944 		VM_OBJECT_ASSERT_WLOCKED(object);
1945 
1946 	flags = 0;
1947 	m = NULL;
1948 	pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
1949 again:
1950 #if VM_NRESERVLEVEL > 0
1951 	/*
1952 	 * Can we allocate the page from a reservation?
1953 	 */
1954 	if (vm_object_reserv(object) &&
1955 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
1956 	    NULL) {
1957 		domain = vm_phys_domain(m);
1958 		vmd = VM_DOMAIN(domain);
1959 		goto found;
1960 	}
1961 #endif
1962 	vmd = VM_DOMAIN(domain);
1963 	if (vmd->vmd_pgcache[pool].zone != NULL) {
1964 		m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
1965 		if (m != NULL) {
1966 			flags |= PG_PCPU_CACHE;
1967 			goto found;
1968 		}
1969 	}
1970 	if (vm_domain_allocate(vmd, req, 1)) {
1971 		/*
1972 		 * If not, allocate it from the free page queues.
1973 		 */
1974 		vm_domain_free_lock(vmd);
1975 		m = vm_phys_alloc_pages(domain, pool, 0);
1976 		vm_domain_free_unlock(vmd);
1977 		if (m == NULL) {
1978 			vm_domain_freecnt_inc(vmd, 1);
1979 #if VM_NRESERVLEVEL > 0
1980 			if (vm_reserv_reclaim_inactive(domain))
1981 				goto again;
1982 #endif
1983 		}
1984 	}
1985 	if (m == NULL) {
1986 		/*
1987 		 * Not allocatable, give up.
1988 		 */
1989 		if (vm_domain_alloc_fail(vmd, object, req))
1990 			goto again;
1991 		return (NULL);
1992 	}
1993 
1994 	/*
1995 	 * At this point we had better have found a good page.
1996 	 */
1997 found:
1998 	vm_page_dequeue(m);
1999 	vm_page_alloc_check(m);
2000 
2001 	/*
2002 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2003 	 */
2004 	if ((req & VM_ALLOC_ZERO) != 0)
2005 		flags |= (m->flags & PG_ZERO);
2006 	if ((req & VM_ALLOC_NODUMP) != 0)
2007 		flags |= PG_NODUMP;
2008 	m->flags = flags;
2009 	m->a.flags = 0;
2010 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2011 	    VPO_UNMANAGED : 0;
2012 	m->busy_lock = VPB_UNBUSIED;
2013 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2014 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2015 	if ((req & VM_ALLOC_SBUSY) != 0)
2016 		m->busy_lock = VPB_SHARERS_WORD(1);
2017 	if (req & VM_ALLOC_WIRED) {
2018 		/*
2019 		 * The page lock is not required for wiring a page until that
2020 		 * page is inserted into the object.
2021 		 */
2022 		vm_wire_add(1);
2023 		m->ref_count = 1;
2024 	}
2025 	m->a.act_count = 0;
2026 
2027 	if (object != NULL) {
2028 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2029 			if (req & VM_ALLOC_WIRED) {
2030 				vm_wire_sub(1);
2031 				m->ref_count = 0;
2032 			}
2033 			KASSERT(m->object == NULL, ("page %p has object", m));
2034 			m->oflags = VPO_UNMANAGED;
2035 			m->busy_lock = VPB_UNBUSIED;
2036 			/* Don't change PG_ZERO. */
2037 			vm_page_free_toq(m);
2038 			if (req & VM_ALLOC_WAITFAIL) {
2039 				VM_OBJECT_WUNLOCK(object);
2040 				vm_radix_wait();
2041 				VM_OBJECT_WLOCK(object);
2042 			}
2043 			return (NULL);
2044 		}
2045 
2046 		/* Ignore device objects; the pager sets "memattr" for them. */
2047 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2048 		    (object->flags & OBJ_FICTITIOUS) == 0)
2049 			pmap_page_set_memattr(m, object->memattr);
2050 	} else
2051 		m->pindex = pindex;
2052 
2053 	return (m);
2054 }
2055 
2056 /*
2057  *	vm_page_alloc_contig:
2058  *
2059  *	Allocate a contiguous set of physical pages of the given size "npages"
2060  *	from the free lists.  All of the physical pages must be at or above
2061  *	the given physical address "low" and below the given physical address
2062  *	"high".  The given value "alignment" determines the alignment of the
2063  *	first physical page in the set.  If the given value "boundary" is
2064  *	non-zero, then the set of physical pages cannot cross any physical
2065  *	address boundary that is a multiple of that value.  Both "alignment"
2066  *	and "boundary" must be a power of two.
2067  *
2068  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2069  *	then the memory attribute setting for the physical pages is configured
2070  *	to the object's memory attribute setting.  Otherwise, the memory
2071  *	attribute setting for the physical pages is configured to "memattr",
2072  *	overriding the object's memory attribute setting.  However, if the
2073  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2074  *	memory attribute setting for the physical pages cannot be configured
2075  *	to VM_MEMATTR_DEFAULT.
2076  *
2077  *	The specified object may not contain fictitious pages.
2078  *
2079  *	The caller must always specify an allocation class.
2080  *
2081  *	allocation classes:
2082  *	VM_ALLOC_NORMAL		normal process request
2083  *	VM_ALLOC_SYSTEM		system *really* needs a page
2084  *	VM_ALLOC_INTERRUPT	interrupt time request
2085  *
2086  *	optional allocation flags:
2087  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2088  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2089  *	VM_ALLOC_NOOBJ		page is not associated with an object and
2090  *				should not be exclusive busy
2091  *	VM_ALLOC_SBUSY		shared busy the allocated page
2092  *	VM_ALLOC_WIRED		wire the allocated page
2093  *	VM_ALLOC_ZERO		prefer a zeroed page
2094  */
2095 vm_page_t
2096 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2097     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2098     vm_paddr_t boundary, vm_memattr_t memattr)
2099 {
2100 	struct vm_domainset_iter di;
2101 	vm_page_t m;
2102 	int domain;
2103 
2104 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2105 	do {
2106 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2107 		    npages, low, high, alignment, boundary, memattr);
2108 		if (m != NULL)
2109 			break;
2110 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2111 
2112 	return (m);
2113 }
2114 
2115 vm_page_t
2116 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2117     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2118     vm_paddr_t boundary, vm_memattr_t memattr)
2119 {
2120 	struct vm_domain *vmd;
2121 	vm_page_t m, m_ret, mpred;
2122 	u_int busy_lock, flags, oflags;
2123 
2124 	mpred = NULL;	/* XXX: pacify gcc */
2125 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2126 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2127 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2128 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2129 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2130 	    req));
2131 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2132 	    ("Can't sleep and retry object insertion."));
2133 	if (object != NULL) {
2134 		VM_OBJECT_ASSERT_WLOCKED(object);
2135 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2136 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2137 		    object));
2138 	}
2139 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2140 
2141 	if (object != NULL) {
2142 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2143 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2144 		    ("vm_page_alloc_contig: pindex already allocated"));
2145 	}
2146 
2147 	/*
2148 	 * Can we allocate the pages without the number of free pages falling
2149 	 * below the lower bound for the allocation class?
2150 	 */
2151 	m_ret = NULL;
2152 again:
2153 #if VM_NRESERVLEVEL > 0
2154 	/*
2155 	 * Can we allocate the pages from a reservation?
2156 	 */
2157 	if (vm_object_reserv(object) &&
2158 	    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2159 	    mpred, npages, low, high, alignment, boundary)) != NULL) {
2160 		domain = vm_phys_domain(m_ret);
2161 		vmd = VM_DOMAIN(domain);
2162 		goto found;
2163 	}
2164 #endif
2165 	vmd = VM_DOMAIN(domain);
2166 	if (vm_domain_allocate(vmd, req, npages)) {
2167 		/*
2168 		 * allocate them from the free page queues.
2169 		 */
2170 		vm_domain_free_lock(vmd);
2171 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2172 		    alignment, boundary);
2173 		vm_domain_free_unlock(vmd);
2174 		if (m_ret == NULL) {
2175 			vm_domain_freecnt_inc(vmd, npages);
2176 #if VM_NRESERVLEVEL > 0
2177 			if (vm_reserv_reclaim_contig(domain, npages, low,
2178 			    high, alignment, boundary))
2179 				goto again;
2180 #endif
2181 		}
2182 	}
2183 	if (m_ret == NULL) {
2184 		if (vm_domain_alloc_fail(vmd, object, req))
2185 			goto again;
2186 		return (NULL);
2187 	}
2188 #if VM_NRESERVLEVEL > 0
2189 found:
2190 #endif
2191 	for (m = m_ret; m < &m_ret[npages]; m++) {
2192 		vm_page_dequeue(m);
2193 		vm_page_alloc_check(m);
2194 	}
2195 
2196 	/*
2197 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2198 	 */
2199 	flags = 0;
2200 	if ((req & VM_ALLOC_ZERO) != 0)
2201 		flags = PG_ZERO;
2202 	if ((req & VM_ALLOC_NODUMP) != 0)
2203 		flags |= PG_NODUMP;
2204 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2205 	    VPO_UNMANAGED : 0;
2206 	busy_lock = VPB_UNBUSIED;
2207 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2208 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2209 	if ((req & VM_ALLOC_SBUSY) != 0)
2210 		busy_lock = VPB_SHARERS_WORD(1);
2211 	if ((req & VM_ALLOC_WIRED) != 0)
2212 		vm_wire_add(npages);
2213 	if (object != NULL) {
2214 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2215 		    memattr == VM_MEMATTR_DEFAULT)
2216 			memattr = object->memattr;
2217 	}
2218 	for (m = m_ret; m < &m_ret[npages]; m++) {
2219 		m->a.flags = 0;
2220 		m->flags = (m->flags | PG_NODUMP) & flags;
2221 		m->busy_lock = busy_lock;
2222 		if ((req & VM_ALLOC_WIRED) != 0)
2223 			m->ref_count = 1;
2224 		m->a.act_count = 0;
2225 		m->oflags = oflags;
2226 		if (object != NULL) {
2227 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2228 				if ((req & VM_ALLOC_WIRED) != 0)
2229 					vm_wire_sub(npages);
2230 				KASSERT(m->object == NULL,
2231 				    ("page %p has object", m));
2232 				mpred = m;
2233 				for (m = m_ret; m < &m_ret[npages]; m++) {
2234 					if (m <= mpred &&
2235 					    (req & VM_ALLOC_WIRED) != 0)
2236 						m->ref_count = 0;
2237 					m->oflags = VPO_UNMANAGED;
2238 					m->busy_lock = VPB_UNBUSIED;
2239 					/* Don't change PG_ZERO. */
2240 					vm_page_free_toq(m);
2241 				}
2242 				if (req & VM_ALLOC_WAITFAIL) {
2243 					VM_OBJECT_WUNLOCK(object);
2244 					vm_radix_wait();
2245 					VM_OBJECT_WLOCK(object);
2246 				}
2247 				return (NULL);
2248 			}
2249 			mpred = m;
2250 		} else
2251 			m->pindex = pindex;
2252 		if (memattr != VM_MEMATTR_DEFAULT)
2253 			pmap_page_set_memattr(m, memattr);
2254 		pindex++;
2255 	}
2256 	return (m_ret);
2257 }
2258 
2259 /*
2260  * Check a page that has been freshly dequeued from a freelist.
2261  */
2262 static void
2263 vm_page_alloc_check(vm_page_t m)
2264 {
2265 
2266 	KASSERT(m->object == NULL, ("page %p has object", m));
2267 	KASSERT(m->a.queue == PQ_NONE &&
2268 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2269 	    ("page %p has unexpected queue %d, flags %#x",
2270 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2271 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2272 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2273 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2274 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2275 	    ("page %p has unexpected memattr %d",
2276 	    m, pmap_page_get_memattr(m)));
2277 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2278 }
2279 
2280 /*
2281  * 	vm_page_alloc_freelist:
2282  *
2283  *	Allocate a physical page from the specified free page list.
2284  *
2285  *	The caller must always specify an allocation class.
2286  *
2287  *	allocation classes:
2288  *	VM_ALLOC_NORMAL		normal process request
2289  *	VM_ALLOC_SYSTEM		system *really* needs a page
2290  *	VM_ALLOC_INTERRUPT	interrupt time request
2291  *
2292  *	optional allocation flags:
2293  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2294  *				intends to allocate
2295  *	VM_ALLOC_WIRED		wire the allocated page
2296  *	VM_ALLOC_ZERO		prefer a zeroed page
2297  */
2298 vm_page_t
2299 vm_page_alloc_freelist(int freelist, int req)
2300 {
2301 	struct vm_domainset_iter di;
2302 	vm_page_t m;
2303 	int domain;
2304 
2305 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2306 	do {
2307 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2308 		if (m != NULL)
2309 			break;
2310 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2311 
2312 	return (m);
2313 }
2314 
2315 vm_page_t
2316 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2317 {
2318 	struct vm_domain *vmd;
2319 	vm_page_t m;
2320 	u_int flags;
2321 
2322 	m = NULL;
2323 	vmd = VM_DOMAIN(domain);
2324 again:
2325 	if (vm_domain_allocate(vmd, req, 1)) {
2326 		vm_domain_free_lock(vmd);
2327 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2328 		    VM_FREEPOOL_DIRECT, 0);
2329 		vm_domain_free_unlock(vmd);
2330 		if (m == NULL)
2331 			vm_domain_freecnt_inc(vmd, 1);
2332 	}
2333 	if (m == NULL) {
2334 		if (vm_domain_alloc_fail(vmd, NULL, req))
2335 			goto again;
2336 		return (NULL);
2337 	}
2338 	vm_page_dequeue(m);
2339 	vm_page_alloc_check(m);
2340 
2341 	/*
2342 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2343 	 */
2344 	m->a.flags = 0;
2345 	flags = 0;
2346 	if ((req & VM_ALLOC_ZERO) != 0)
2347 		flags = PG_ZERO;
2348 	m->flags &= flags;
2349 	if ((req & VM_ALLOC_WIRED) != 0) {
2350 		/*
2351 		 * The page lock is not required for wiring a page that does
2352 		 * not belong to an object.
2353 		 */
2354 		vm_wire_add(1);
2355 		m->ref_count = 1;
2356 	}
2357 	/* Unmanaged pages don't use "act_count". */
2358 	m->oflags = VPO_UNMANAGED;
2359 	return (m);
2360 }
2361 
2362 static int
2363 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2364 {
2365 	struct vm_domain *vmd;
2366 	struct vm_pgcache *pgcache;
2367 	int i;
2368 
2369 	pgcache = arg;
2370 	vmd = VM_DOMAIN(pgcache->domain);
2371 
2372 	/*
2373 	 * The page daemon should avoid creating extra memory pressure since its
2374 	 * main purpose is to replenish the store of free pages.
2375 	 */
2376 	if (vmd->vmd_severeset || curproc == pageproc ||
2377 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2378 		return (0);
2379 	domain = vmd->vmd_domain;
2380 	vm_domain_free_lock(vmd);
2381 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2382 	    (vm_page_t *)store);
2383 	vm_domain_free_unlock(vmd);
2384 	if (cnt != i)
2385 		vm_domain_freecnt_inc(vmd, cnt - i);
2386 
2387 	return (i);
2388 }
2389 
2390 static void
2391 vm_page_zone_release(void *arg, void **store, int cnt)
2392 {
2393 	struct vm_domain *vmd;
2394 	struct vm_pgcache *pgcache;
2395 	vm_page_t m;
2396 	int i;
2397 
2398 	pgcache = arg;
2399 	vmd = VM_DOMAIN(pgcache->domain);
2400 	vm_domain_free_lock(vmd);
2401 	for (i = 0; i < cnt; i++) {
2402 		m = (vm_page_t)store[i];
2403 		vm_phys_free_pages(m, 0);
2404 	}
2405 	vm_domain_free_unlock(vmd);
2406 	vm_domain_freecnt_inc(vmd, cnt);
2407 }
2408 
2409 #define	VPSC_ANY	0	/* No restrictions. */
2410 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2411 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2412 
2413 /*
2414  *	vm_page_scan_contig:
2415  *
2416  *	Scan vm_page_array[] between the specified entries "m_start" and
2417  *	"m_end" for a run of contiguous physical pages that satisfy the
2418  *	specified conditions, and return the lowest page in the run.  The
2419  *	specified "alignment" determines the alignment of the lowest physical
2420  *	page in the run.  If the specified "boundary" is non-zero, then the
2421  *	run of physical pages cannot span a physical address that is a
2422  *	multiple of "boundary".
2423  *
2424  *	"m_end" is never dereferenced, so it need not point to a vm_page
2425  *	structure within vm_page_array[].
2426  *
2427  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2428  *	span a hole (or discontiguity) in the physical address space.  Both
2429  *	"alignment" and "boundary" must be a power of two.
2430  */
2431 vm_page_t
2432 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2433     u_long alignment, vm_paddr_t boundary, int options)
2434 {
2435 	struct mtx *m_mtx;
2436 	vm_object_t object;
2437 	vm_paddr_t pa;
2438 	vm_page_t m, m_run;
2439 #if VM_NRESERVLEVEL > 0
2440 	int level;
2441 #endif
2442 	int m_inc, order, run_ext, run_len;
2443 
2444 	KASSERT(npages > 0, ("npages is 0"));
2445 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2446 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2447 	m_run = NULL;
2448 	run_len = 0;
2449 	m_mtx = NULL;
2450 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2451 		KASSERT((m->flags & PG_MARKER) == 0,
2452 		    ("page %p is PG_MARKER", m));
2453 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2454 		    ("fictitious page %p has invalid ref count", m));
2455 
2456 		/*
2457 		 * If the current page would be the start of a run, check its
2458 		 * physical address against the end, alignment, and boundary
2459 		 * conditions.  If it doesn't satisfy these conditions, either
2460 		 * terminate the scan or advance to the next page that
2461 		 * satisfies the failed condition.
2462 		 */
2463 		if (run_len == 0) {
2464 			KASSERT(m_run == NULL, ("m_run != NULL"));
2465 			if (m + npages > m_end)
2466 				break;
2467 			pa = VM_PAGE_TO_PHYS(m);
2468 			if ((pa & (alignment - 1)) != 0) {
2469 				m_inc = atop(roundup2(pa, alignment) - pa);
2470 				continue;
2471 			}
2472 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2473 			    boundary) != 0) {
2474 				m_inc = atop(roundup2(pa, boundary) - pa);
2475 				continue;
2476 			}
2477 		} else
2478 			KASSERT(m_run != NULL, ("m_run == NULL"));
2479 
2480 		vm_page_change_lock(m, &m_mtx);
2481 		m_inc = 1;
2482 retry:
2483 		if (vm_page_wired(m))
2484 			run_ext = 0;
2485 #if VM_NRESERVLEVEL > 0
2486 		else if ((level = vm_reserv_level(m)) >= 0 &&
2487 		    (options & VPSC_NORESERV) != 0) {
2488 			run_ext = 0;
2489 			/* Advance to the end of the reservation. */
2490 			pa = VM_PAGE_TO_PHYS(m);
2491 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2492 			    pa);
2493 		}
2494 #endif
2495 		else if ((object = m->object) != NULL) {
2496 			/*
2497 			 * The page is considered eligible for relocation if
2498 			 * and only if it could be laundered or reclaimed by
2499 			 * the page daemon.
2500 			 */
2501 			if (!VM_OBJECT_TRYRLOCK(object)) {
2502 				mtx_unlock(m_mtx);
2503 				VM_OBJECT_RLOCK(object);
2504 				mtx_lock(m_mtx);
2505 				if (m->object != object) {
2506 					/*
2507 					 * The page may have been freed.
2508 					 */
2509 					VM_OBJECT_RUNLOCK(object);
2510 					goto retry;
2511 				}
2512 			}
2513 			/* Don't care: PG_NODUMP, PG_ZERO. */
2514 			if (object->type != OBJT_DEFAULT &&
2515 			    object->type != OBJT_SWAP &&
2516 			    object->type != OBJT_VNODE) {
2517 				run_ext = 0;
2518 #if VM_NRESERVLEVEL > 0
2519 			} else if ((options & VPSC_NOSUPER) != 0 &&
2520 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2521 				run_ext = 0;
2522 				/* Advance to the end of the superpage. */
2523 				pa = VM_PAGE_TO_PHYS(m);
2524 				m_inc = atop(roundup2(pa + 1,
2525 				    vm_reserv_size(level)) - pa);
2526 #endif
2527 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2528 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) &&
2529 			    !vm_page_wired(m)) {
2530 				/*
2531 				 * The page is allocated but eligible for
2532 				 * relocation.  Extend the current run by one
2533 				 * page.
2534 				 */
2535 				KASSERT(pmap_page_get_memattr(m) ==
2536 				    VM_MEMATTR_DEFAULT,
2537 				    ("page %p has an unexpected memattr", m));
2538 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2539 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2540 				    ("page %p has unexpected oflags", m));
2541 				/* Don't care: PGA_NOSYNC. */
2542 				run_ext = 1;
2543 			} else
2544 				run_ext = 0;
2545 			VM_OBJECT_RUNLOCK(object);
2546 #if VM_NRESERVLEVEL > 0
2547 		} else if (level >= 0) {
2548 			/*
2549 			 * The page is reserved but not yet allocated.  In
2550 			 * other words, it is still free.  Extend the current
2551 			 * run by one page.
2552 			 */
2553 			run_ext = 1;
2554 #endif
2555 		} else if ((order = m->order) < VM_NFREEORDER) {
2556 			/*
2557 			 * The page is enqueued in the physical memory
2558 			 * allocator's free page queues.  Moreover, it is the
2559 			 * first page in a power-of-two-sized run of
2560 			 * contiguous free pages.  Add these pages to the end
2561 			 * of the current run, and jump ahead.
2562 			 */
2563 			run_ext = 1 << order;
2564 			m_inc = 1 << order;
2565 		} else {
2566 			/*
2567 			 * Skip the page for one of the following reasons: (1)
2568 			 * It is enqueued in the physical memory allocator's
2569 			 * free page queues.  However, it is not the first
2570 			 * page in a run of contiguous free pages.  (This case
2571 			 * rarely occurs because the scan is performed in
2572 			 * ascending order.) (2) It is not reserved, and it is
2573 			 * transitioning from free to allocated.  (Conversely,
2574 			 * the transition from allocated to free for managed
2575 			 * pages is blocked by the page lock.) (3) It is
2576 			 * allocated but not contained by an object and not
2577 			 * wired, e.g., allocated by Xen's balloon driver.
2578 			 */
2579 			run_ext = 0;
2580 		}
2581 
2582 		/*
2583 		 * Extend or reset the current run of pages.
2584 		 */
2585 		if (run_ext > 0) {
2586 			if (run_len == 0)
2587 				m_run = m;
2588 			run_len += run_ext;
2589 		} else {
2590 			if (run_len > 0) {
2591 				m_run = NULL;
2592 				run_len = 0;
2593 			}
2594 		}
2595 	}
2596 	if (m_mtx != NULL)
2597 		mtx_unlock(m_mtx);
2598 	if (run_len >= npages)
2599 		return (m_run);
2600 	return (NULL);
2601 }
2602 
2603 /*
2604  *	vm_page_reclaim_run:
2605  *
2606  *	Try to relocate each of the allocated virtual pages within the
2607  *	specified run of physical pages to a new physical address.  Free the
2608  *	physical pages underlying the relocated virtual pages.  A virtual page
2609  *	is relocatable if and only if it could be laundered or reclaimed by
2610  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2611  *	physical address above "high".
2612  *
2613  *	Returns 0 if every physical page within the run was already free or
2614  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2615  *	value indicating why the last attempt to relocate a virtual page was
2616  *	unsuccessful.
2617  *
2618  *	"req_class" must be an allocation class.
2619  */
2620 static int
2621 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2622     vm_paddr_t high)
2623 {
2624 	struct vm_domain *vmd;
2625 	struct mtx *m_mtx;
2626 	struct spglist free;
2627 	vm_object_t object;
2628 	vm_paddr_t pa;
2629 	vm_page_t m, m_end, m_new;
2630 	int error, order, req;
2631 
2632 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2633 	    ("req_class is not an allocation class"));
2634 	SLIST_INIT(&free);
2635 	error = 0;
2636 	m = m_run;
2637 	m_end = m_run + npages;
2638 	m_mtx = NULL;
2639 	for (; error == 0 && m < m_end; m++) {
2640 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2641 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2642 
2643 		/*
2644 		 * Avoid releasing and reacquiring the same page lock.
2645 		 */
2646 		vm_page_change_lock(m, &m_mtx);
2647 retry:
2648 		/*
2649 		 * Racily check for wirings.  Races are handled below.
2650 		 */
2651 		if (vm_page_wired(m))
2652 			error = EBUSY;
2653 		else if ((object = m->object) != NULL) {
2654 			/*
2655 			 * The page is relocated if and only if it could be
2656 			 * laundered or reclaimed by the page daemon.
2657 			 */
2658 			if (!VM_OBJECT_TRYWLOCK(object)) {
2659 				mtx_unlock(m_mtx);
2660 				VM_OBJECT_WLOCK(object);
2661 				mtx_lock(m_mtx);
2662 				if (m->object != object) {
2663 					/*
2664 					 * The page may have been freed.
2665 					 */
2666 					VM_OBJECT_WUNLOCK(object);
2667 					goto retry;
2668 				}
2669 			}
2670 			/* Don't care: PG_NODUMP, PG_ZERO. */
2671 			if (object->type != OBJT_DEFAULT &&
2672 			    object->type != OBJT_SWAP &&
2673 			    object->type != OBJT_VNODE)
2674 				error = EINVAL;
2675 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2676 				error = EINVAL;
2677 			else if (vm_page_queue(m) != PQ_NONE &&
2678 			    vm_page_tryxbusy(m) != 0) {
2679 				if (vm_page_wired(m)) {
2680 					vm_page_xunbusy(m);
2681 					error = EBUSY;
2682 					goto unlock;
2683 				}
2684 				KASSERT(pmap_page_get_memattr(m) ==
2685 				    VM_MEMATTR_DEFAULT,
2686 				    ("page %p has an unexpected memattr", m));
2687 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2688 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2689 				    ("page %p has unexpected oflags", m));
2690 				/* Don't care: PGA_NOSYNC. */
2691 				if (!vm_page_none_valid(m)) {
2692 					/*
2693 					 * First, try to allocate a new page
2694 					 * that is above "high".  Failing
2695 					 * that, try to allocate a new page
2696 					 * that is below "m_run".  Allocate
2697 					 * the new page between the end of
2698 					 * "m_run" and "high" only as a last
2699 					 * resort.
2700 					 */
2701 					req = req_class | VM_ALLOC_NOOBJ;
2702 					if ((m->flags & PG_NODUMP) != 0)
2703 						req |= VM_ALLOC_NODUMP;
2704 					if (trunc_page(high) !=
2705 					    ~(vm_paddr_t)PAGE_MASK) {
2706 						m_new = vm_page_alloc_contig(
2707 						    NULL, 0, req, 1,
2708 						    round_page(high),
2709 						    ~(vm_paddr_t)0,
2710 						    PAGE_SIZE, 0,
2711 						    VM_MEMATTR_DEFAULT);
2712 					} else
2713 						m_new = NULL;
2714 					if (m_new == NULL) {
2715 						pa = VM_PAGE_TO_PHYS(m_run);
2716 						m_new = vm_page_alloc_contig(
2717 						    NULL, 0, req, 1,
2718 						    0, pa - 1, PAGE_SIZE, 0,
2719 						    VM_MEMATTR_DEFAULT);
2720 					}
2721 					if (m_new == NULL) {
2722 						pa += ptoa(npages);
2723 						m_new = vm_page_alloc_contig(
2724 						    NULL, 0, req, 1,
2725 						    pa, high, PAGE_SIZE, 0,
2726 						    VM_MEMATTR_DEFAULT);
2727 					}
2728 					if (m_new == NULL) {
2729 						vm_page_xunbusy(m);
2730 						error = ENOMEM;
2731 						goto unlock;
2732 					}
2733 
2734 					/*
2735 					 * Unmap the page and check for new
2736 					 * wirings that may have been acquired
2737 					 * through a pmap lookup.
2738 					 */
2739 					if (object->ref_count != 0 &&
2740 					    !vm_page_try_remove_all(m)) {
2741 						vm_page_free(m_new);
2742 						error = EBUSY;
2743 						goto unlock;
2744 					}
2745 
2746 					/*
2747 					 * Replace "m" with the new page.  For
2748 					 * vm_page_replace(), "m" must be busy
2749 					 * and dequeued.  Finally, change "m"
2750 					 * as if vm_page_free() was called.
2751 					 */
2752 					m_new->a.flags = m->a.flags &
2753 					    ~PGA_QUEUE_STATE_MASK;
2754 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2755 					    ("page %p is managed", m_new));
2756 					pmap_copy_page(m, m_new);
2757 					m_new->valid = m->valid;
2758 					m_new->dirty = m->dirty;
2759 					m->flags &= ~PG_ZERO;
2760 					vm_page_dequeue(m);
2761 					vm_page_replace_checked(m_new, object,
2762 					    m->pindex, m);
2763 					if (vm_page_free_prep(m))
2764 						SLIST_INSERT_HEAD(&free, m,
2765 						    plinks.s.ss);
2766 
2767 					/*
2768 					 * The new page must be deactivated
2769 					 * before the object is unlocked.
2770 					 */
2771 					vm_page_change_lock(m_new, &m_mtx);
2772 					vm_page_deactivate(m_new);
2773 				} else {
2774 					m->flags &= ~PG_ZERO;
2775 					vm_page_dequeue(m);
2776 					if (vm_page_free_prep(m))
2777 						SLIST_INSERT_HEAD(&free, m,
2778 						    plinks.s.ss);
2779 					KASSERT(m->dirty == 0,
2780 					    ("page %p is dirty", m));
2781 				}
2782 			} else
2783 				error = EBUSY;
2784 unlock:
2785 			VM_OBJECT_WUNLOCK(object);
2786 		} else {
2787 			MPASS(vm_phys_domain(m) == domain);
2788 			vmd = VM_DOMAIN(domain);
2789 			vm_domain_free_lock(vmd);
2790 			order = m->order;
2791 			if (order < VM_NFREEORDER) {
2792 				/*
2793 				 * The page is enqueued in the physical memory
2794 				 * allocator's free page queues.  Moreover, it
2795 				 * is the first page in a power-of-two-sized
2796 				 * run of contiguous free pages.  Jump ahead
2797 				 * to the last page within that run, and
2798 				 * continue from there.
2799 				 */
2800 				m += (1 << order) - 1;
2801 			}
2802 #if VM_NRESERVLEVEL > 0
2803 			else if (vm_reserv_is_page_free(m))
2804 				order = 0;
2805 #endif
2806 			vm_domain_free_unlock(vmd);
2807 			if (order == VM_NFREEORDER)
2808 				error = EINVAL;
2809 		}
2810 	}
2811 	if (m_mtx != NULL)
2812 		mtx_unlock(m_mtx);
2813 	if ((m = SLIST_FIRST(&free)) != NULL) {
2814 		int cnt;
2815 
2816 		vmd = VM_DOMAIN(domain);
2817 		cnt = 0;
2818 		vm_domain_free_lock(vmd);
2819 		do {
2820 			MPASS(vm_phys_domain(m) == domain);
2821 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2822 			vm_phys_free_pages(m, 0);
2823 			cnt++;
2824 		} while ((m = SLIST_FIRST(&free)) != NULL);
2825 		vm_domain_free_unlock(vmd);
2826 		vm_domain_freecnt_inc(vmd, cnt);
2827 	}
2828 	return (error);
2829 }
2830 
2831 #define	NRUNS	16
2832 
2833 CTASSERT(powerof2(NRUNS));
2834 
2835 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2836 
2837 #define	MIN_RECLAIM	8
2838 
2839 /*
2840  *	vm_page_reclaim_contig:
2841  *
2842  *	Reclaim allocated, contiguous physical memory satisfying the specified
2843  *	conditions by relocating the virtual pages using that physical memory.
2844  *	Returns true if reclamation is successful and false otherwise.  Since
2845  *	relocation requires the allocation of physical pages, reclamation may
2846  *	fail due to a shortage of free pages.  When reclamation fails, callers
2847  *	are expected to perform vm_wait() before retrying a failed allocation
2848  *	operation, e.g., vm_page_alloc_contig().
2849  *
2850  *	The caller must always specify an allocation class through "req".
2851  *
2852  *	allocation classes:
2853  *	VM_ALLOC_NORMAL		normal process request
2854  *	VM_ALLOC_SYSTEM		system *really* needs a page
2855  *	VM_ALLOC_INTERRUPT	interrupt time request
2856  *
2857  *	The optional allocation flags are ignored.
2858  *
2859  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2860  *	must be a power of two.
2861  */
2862 bool
2863 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2864     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2865 {
2866 	struct vm_domain *vmd;
2867 	vm_paddr_t curr_low;
2868 	vm_page_t m_run, m_runs[NRUNS];
2869 	u_long count, reclaimed;
2870 	int error, i, options, req_class;
2871 
2872 	KASSERT(npages > 0, ("npages is 0"));
2873 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2874 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2875 	req_class = req & VM_ALLOC_CLASS_MASK;
2876 
2877 	/*
2878 	 * The page daemon is allowed to dig deeper into the free page list.
2879 	 */
2880 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2881 		req_class = VM_ALLOC_SYSTEM;
2882 
2883 	/*
2884 	 * Return if the number of free pages cannot satisfy the requested
2885 	 * allocation.
2886 	 */
2887 	vmd = VM_DOMAIN(domain);
2888 	count = vmd->vmd_free_count;
2889 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2890 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2891 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2892 		return (false);
2893 
2894 	/*
2895 	 * Scan up to three times, relaxing the restrictions ("options") on
2896 	 * the reclamation of reservations and superpages each time.
2897 	 */
2898 	for (options = VPSC_NORESERV;;) {
2899 		/*
2900 		 * Find the highest runs that satisfy the given constraints
2901 		 * and restrictions, and record them in "m_runs".
2902 		 */
2903 		curr_low = low;
2904 		count = 0;
2905 		for (;;) {
2906 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2907 			    high, alignment, boundary, options);
2908 			if (m_run == NULL)
2909 				break;
2910 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2911 			m_runs[RUN_INDEX(count)] = m_run;
2912 			count++;
2913 		}
2914 
2915 		/*
2916 		 * Reclaim the highest runs in LIFO (descending) order until
2917 		 * the number of reclaimed pages, "reclaimed", is at least
2918 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2919 		 * reclamation is idempotent, and runs will (likely) recur
2920 		 * from one scan to the next as restrictions are relaxed.
2921 		 */
2922 		reclaimed = 0;
2923 		for (i = 0; count > 0 && i < NRUNS; i++) {
2924 			count--;
2925 			m_run = m_runs[RUN_INDEX(count)];
2926 			error = vm_page_reclaim_run(req_class, domain, npages,
2927 			    m_run, high);
2928 			if (error == 0) {
2929 				reclaimed += npages;
2930 				if (reclaimed >= MIN_RECLAIM)
2931 					return (true);
2932 			}
2933 		}
2934 
2935 		/*
2936 		 * Either relax the restrictions on the next scan or return if
2937 		 * the last scan had no restrictions.
2938 		 */
2939 		if (options == VPSC_NORESERV)
2940 			options = VPSC_NOSUPER;
2941 		else if (options == VPSC_NOSUPER)
2942 			options = VPSC_ANY;
2943 		else if (options == VPSC_ANY)
2944 			return (reclaimed != 0);
2945 	}
2946 }
2947 
2948 bool
2949 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2950     u_long alignment, vm_paddr_t boundary)
2951 {
2952 	struct vm_domainset_iter di;
2953 	int domain;
2954 	bool ret;
2955 
2956 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2957 	do {
2958 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
2959 		    high, alignment, boundary);
2960 		if (ret)
2961 			break;
2962 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2963 
2964 	return (ret);
2965 }
2966 
2967 /*
2968  * Set the domain in the appropriate page level domainset.
2969  */
2970 void
2971 vm_domain_set(struct vm_domain *vmd)
2972 {
2973 
2974 	mtx_lock(&vm_domainset_lock);
2975 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
2976 		vmd->vmd_minset = 1;
2977 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
2978 	}
2979 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
2980 		vmd->vmd_severeset = 1;
2981 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
2982 	}
2983 	mtx_unlock(&vm_domainset_lock);
2984 }
2985 
2986 /*
2987  * Clear the domain from the appropriate page level domainset.
2988  */
2989 void
2990 vm_domain_clear(struct vm_domain *vmd)
2991 {
2992 
2993 	mtx_lock(&vm_domainset_lock);
2994 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
2995 		vmd->vmd_minset = 0;
2996 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
2997 		if (vm_min_waiters != 0) {
2998 			vm_min_waiters = 0;
2999 			wakeup(&vm_min_domains);
3000 		}
3001 	}
3002 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3003 		vmd->vmd_severeset = 0;
3004 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3005 		if (vm_severe_waiters != 0) {
3006 			vm_severe_waiters = 0;
3007 			wakeup(&vm_severe_domains);
3008 		}
3009 	}
3010 
3011 	/*
3012 	 * If pageout daemon needs pages, then tell it that there are
3013 	 * some free.
3014 	 */
3015 	if (vmd->vmd_pageout_pages_needed &&
3016 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3017 		wakeup(&vmd->vmd_pageout_pages_needed);
3018 		vmd->vmd_pageout_pages_needed = 0;
3019 	}
3020 
3021 	/* See comments in vm_wait_doms(). */
3022 	if (vm_pageproc_waiters) {
3023 		vm_pageproc_waiters = 0;
3024 		wakeup(&vm_pageproc_waiters);
3025 	}
3026 	mtx_unlock(&vm_domainset_lock);
3027 }
3028 
3029 /*
3030  * Wait for free pages to exceed the min threshold globally.
3031  */
3032 void
3033 vm_wait_min(void)
3034 {
3035 
3036 	mtx_lock(&vm_domainset_lock);
3037 	while (vm_page_count_min()) {
3038 		vm_min_waiters++;
3039 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3040 	}
3041 	mtx_unlock(&vm_domainset_lock);
3042 }
3043 
3044 /*
3045  * Wait for free pages to exceed the severe threshold globally.
3046  */
3047 void
3048 vm_wait_severe(void)
3049 {
3050 
3051 	mtx_lock(&vm_domainset_lock);
3052 	while (vm_page_count_severe()) {
3053 		vm_severe_waiters++;
3054 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3055 		    "vmwait", 0);
3056 	}
3057 	mtx_unlock(&vm_domainset_lock);
3058 }
3059 
3060 u_int
3061 vm_wait_count(void)
3062 {
3063 
3064 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3065 }
3066 
3067 void
3068 vm_wait_doms(const domainset_t *wdoms)
3069 {
3070 
3071 	/*
3072 	 * We use racey wakeup synchronization to avoid expensive global
3073 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3074 	 * To handle this, we only sleep for one tick in this instance.  It
3075 	 * is expected that most allocations for the pageproc will come from
3076 	 * kmem or vm_page_grab* which will use the more specific and
3077 	 * race-free vm_wait_domain().
3078 	 */
3079 	if (curproc == pageproc) {
3080 		mtx_lock(&vm_domainset_lock);
3081 		vm_pageproc_waiters++;
3082 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3083 		    "pageprocwait", 1);
3084 	} else {
3085 		/*
3086 		 * XXX Ideally we would wait only until the allocation could
3087 		 * be satisfied.  This condition can cause new allocators to
3088 		 * consume all freed pages while old allocators wait.
3089 		 */
3090 		mtx_lock(&vm_domainset_lock);
3091 		if (vm_page_count_min_set(wdoms)) {
3092 			vm_min_waiters++;
3093 			msleep(&vm_min_domains, &vm_domainset_lock,
3094 			    PVM | PDROP, "vmwait", 0);
3095 		} else
3096 			mtx_unlock(&vm_domainset_lock);
3097 	}
3098 }
3099 
3100 /*
3101  *	vm_wait_domain:
3102  *
3103  *	Sleep until free pages are available for allocation.
3104  *	- Called in various places after failed memory allocations.
3105  */
3106 void
3107 vm_wait_domain(int domain)
3108 {
3109 	struct vm_domain *vmd;
3110 	domainset_t wdom;
3111 
3112 	vmd = VM_DOMAIN(domain);
3113 	vm_domain_free_assert_unlocked(vmd);
3114 
3115 	if (curproc == pageproc) {
3116 		mtx_lock(&vm_domainset_lock);
3117 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3118 			vmd->vmd_pageout_pages_needed = 1;
3119 			msleep(&vmd->vmd_pageout_pages_needed,
3120 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3121 		} else
3122 			mtx_unlock(&vm_domainset_lock);
3123 	} else {
3124 		if (pageproc == NULL)
3125 			panic("vm_wait in early boot");
3126 		DOMAINSET_ZERO(&wdom);
3127 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3128 		vm_wait_doms(&wdom);
3129 	}
3130 }
3131 
3132 /*
3133  *	vm_wait:
3134  *
3135  *	Sleep until free pages are available for allocation in the
3136  *	affinity domains of the obj.  If obj is NULL, the domain set
3137  *	for the calling thread is used.
3138  *	Called in various places after failed memory allocations.
3139  */
3140 void
3141 vm_wait(vm_object_t obj)
3142 {
3143 	struct domainset *d;
3144 
3145 	d = NULL;
3146 
3147 	/*
3148 	 * Carefully fetch pointers only once: the struct domainset
3149 	 * itself is ummutable but the pointer might change.
3150 	 */
3151 	if (obj != NULL)
3152 		d = obj->domain.dr_policy;
3153 	if (d == NULL)
3154 		d = curthread->td_domain.dr_policy;
3155 
3156 	vm_wait_doms(&d->ds_mask);
3157 }
3158 
3159 /*
3160  *	vm_domain_alloc_fail:
3161  *
3162  *	Called when a page allocation function fails.  Informs the
3163  *	pagedaemon and performs the requested wait.  Requires the
3164  *	domain_free and object lock on entry.  Returns with the
3165  *	object lock held and free lock released.  Returns an error when
3166  *	retry is necessary.
3167  *
3168  */
3169 static int
3170 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3171 {
3172 
3173 	vm_domain_free_assert_unlocked(vmd);
3174 
3175 	atomic_add_int(&vmd->vmd_pageout_deficit,
3176 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3177 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3178 		if (object != NULL)
3179 			VM_OBJECT_WUNLOCK(object);
3180 		vm_wait_domain(vmd->vmd_domain);
3181 		if (object != NULL)
3182 			VM_OBJECT_WLOCK(object);
3183 		if (req & VM_ALLOC_WAITOK)
3184 			return (EAGAIN);
3185 	}
3186 
3187 	return (0);
3188 }
3189 
3190 /*
3191  *	vm_waitpfault:
3192  *
3193  *	Sleep until free pages are available for allocation.
3194  *	- Called only in vm_fault so that processes page faulting
3195  *	  can be easily tracked.
3196  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3197  *	  processes will be able to grab memory first.  Do not change
3198  *	  this balance without careful testing first.
3199  */
3200 void
3201 vm_waitpfault(struct domainset *dset, int timo)
3202 {
3203 
3204 	/*
3205 	 * XXX Ideally we would wait only until the allocation could
3206 	 * be satisfied.  This condition can cause new allocators to
3207 	 * consume all freed pages while old allocators wait.
3208 	 */
3209 	mtx_lock(&vm_domainset_lock);
3210 	if (vm_page_count_min_set(&dset->ds_mask)) {
3211 		vm_min_waiters++;
3212 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3213 		    "pfault", timo);
3214 	} else
3215 		mtx_unlock(&vm_domainset_lock);
3216 }
3217 
3218 static struct vm_pagequeue *
3219 vm_page_pagequeue(vm_page_t m)
3220 {
3221 
3222 	uint8_t queue;
3223 
3224 	if ((queue = atomic_load_8(&m->a.queue)) == PQ_NONE)
3225 		return (NULL);
3226 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3227 }
3228 
3229 static inline void
3230 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m)
3231 {
3232 	struct vm_domain *vmd;
3233 	uint16_t qflags;
3234 
3235 	CRITICAL_ASSERT(curthread);
3236 	vm_pagequeue_assert_locked(pq);
3237 
3238 	/*
3239 	 * The page daemon is allowed to set m->a.queue = PQ_NONE without
3240 	 * the page queue lock held.  In this case it is about to free the page,
3241 	 * which must not have any queue state.
3242 	 */
3243 	qflags = atomic_load_16(&m->a.flags);
3244 	KASSERT(pq == vm_page_pagequeue(m) ||
3245 	    (qflags & PGA_QUEUE_STATE_MASK) == 0,
3246 	    ("page %p doesn't belong to queue %p but has aflags %#x",
3247 	    m, pq, qflags));
3248 
3249 	if ((qflags & PGA_DEQUEUE) != 0) {
3250 		if (__predict_true((qflags & PGA_ENQUEUED) != 0))
3251 			vm_pagequeue_remove(pq, m);
3252 		vm_page_dequeue_complete(m);
3253 		counter_u64_add(queue_ops, 1);
3254 	} else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) {
3255 		if ((qflags & PGA_ENQUEUED) != 0)
3256 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3257 		else {
3258 			vm_pagequeue_cnt_inc(pq);
3259 			vm_page_aflag_set(m, PGA_ENQUEUED);
3260 		}
3261 
3262 		/*
3263 		 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.
3264 		 * In particular, if both flags are set in close succession,
3265 		 * only PGA_REQUEUE_HEAD will be applied, even if it was set
3266 		 * first.
3267 		 */
3268 		if ((qflags & PGA_REQUEUE_HEAD) != 0) {
3269 			KASSERT(m->a.queue == PQ_INACTIVE,
3270 			    ("head enqueue not supported for page %p", m));
3271 			vmd = vm_pagequeue_domain(m);
3272 			TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3273 		} else
3274 			TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3275 
3276 		vm_page_aflag_clear(m, qflags & (PGA_REQUEUE |
3277 		    PGA_REQUEUE_HEAD));
3278 		counter_u64_add(queue_ops, 1);
3279 	} else {
3280 		counter_u64_add(queue_nops, 1);
3281 	}
3282 }
3283 
3284 static void
3285 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3286     uint8_t queue)
3287 {
3288 	vm_page_t m;
3289 	int i;
3290 
3291 	for (i = 0; i < bq->bq_cnt; i++) {
3292 		m = bq->bq_pa[i];
3293 		if (__predict_false(m->a.queue != queue))
3294 			continue;
3295 		vm_pqbatch_process_page(pq, m);
3296 	}
3297 	vm_batchqueue_init(bq);
3298 }
3299 
3300 /*
3301  *	vm_page_pqbatch_submit:		[ internal use only ]
3302  *
3303  *	Enqueue a page in the specified page queue's batched work queue.
3304  *	The caller must have encoded the requested operation in the page
3305  *	structure's a.flags field.
3306  */
3307 void
3308 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3309 {
3310 	struct vm_batchqueue *bq;
3311 	struct vm_pagequeue *pq;
3312 	int domain;
3313 
3314 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3315 	    ("page %p is unmanaged", m));
3316 	KASSERT(mtx_owned(vm_page_lockptr(m)) || m->object == NULL,
3317 	    ("missing synchronization for page %p", m));
3318 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3319 
3320 	domain = vm_phys_domain(m);
3321 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3322 
3323 	critical_enter();
3324 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3325 	if (vm_batchqueue_insert(bq, m)) {
3326 		critical_exit();
3327 		return;
3328 	}
3329 	critical_exit();
3330 	vm_pagequeue_lock(pq);
3331 	critical_enter();
3332 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3333 	vm_pqbatch_process(pq, bq, queue);
3334 
3335 	/*
3336 	 * The page may have been logically dequeued before we acquired the
3337 	 * page queue lock.  In this case, since we either hold the page lock
3338 	 * or the page is being freed, a different thread cannot be concurrently
3339 	 * enqueuing the page.
3340 	 */
3341 	if (__predict_true(m->a.queue == queue))
3342 		vm_pqbatch_process_page(pq, m);
3343 	else {
3344 		KASSERT(m->a.queue == PQ_NONE,
3345 		    ("invalid queue transition for page %p", m));
3346 		KASSERT((m->a.flags & PGA_ENQUEUED) == 0,
3347 		    ("page %p is enqueued with invalid queue index", m));
3348 	}
3349 	vm_pagequeue_unlock(pq);
3350 	critical_exit();
3351 }
3352 
3353 /*
3354  *	vm_page_pqbatch_drain:		[ internal use only ]
3355  *
3356  *	Force all per-CPU page queue batch queues to be drained.  This is
3357  *	intended for use in severe memory shortages, to ensure that pages
3358  *	do not remain stuck in the batch queues.
3359  */
3360 void
3361 vm_page_pqbatch_drain(void)
3362 {
3363 	struct thread *td;
3364 	struct vm_domain *vmd;
3365 	struct vm_pagequeue *pq;
3366 	int cpu, domain, queue;
3367 
3368 	td = curthread;
3369 	CPU_FOREACH(cpu) {
3370 		thread_lock(td);
3371 		sched_bind(td, cpu);
3372 		thread_unlock(td);
3373 
3374 		for (domain = 0; domain < vm_ndomains; domain++) {
3375 			vmd = VM_DOMAIN(domain);
3376 			for (queue = 0; queue < PQ_COUNT; queue++) {
3377 				pq = &vmd->vmd_pagequeues[queue];
3378 				vm_pagequeue_lock(pq);
3379 				critical_enter();
3380 				vm_pqbatch_process(pq,
3381 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3382 				critical_exit();
3383 				vm_pagequeue_unlock(pq);
3384 			}
3385 		}
3386 	}
3387 	thread_lock(td);
3388 	sched_unbind(td);
3389 	thread_unlock(td);
3390 }
3391 
3392 /*
3393  * Complete the logical removal of a page from a page queue.  We must be
3394  * careful to synchronize with the page daemon, which may be concurrently
3395  * examining the page with only the page lock held.  The page must not be
3396  * in a state where it appears to be logically enqueued.
3397  */
3398 static void
3399 vm_page_dequeue_complete(vm_page_t m)
3400 {
3401 
3402 	m->a.queue = PQ_NONE;
3403 	atomic_thread_fence_rel();
3404 	vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK);
3405 }
3406 
3407 /*
3408  *	vm_page_dequeue_deferred:	[ internal use only ]
3409  *
3410  *	Request removal of the given page from its current page
3411  *	queue.  Physical removal from the queue may be deferred
3412  *	indefinitely.
3413  *
3414  *	The page must be locked.
3415  */
3416 void
3417 vm_page_dequeue_deferred(vm_page_t m)
3418 {
3419 	uint8_t queue;
3420 
3421 	vm_page_assert_locked(m);
3422 
3423 	if ((queue = vm_page_queue(m)) == PQ_NONE)
3424 		return;
3425 
3426 	/*
3427 	 * Set PGA_DEQUEUE if it is not already set to handle a concurrent call
3428 	 * to vm_page_dequeue_deferred_free().  In particular, avoid modifying
3429 	 * the page's queue state once vm_page_dequeue_deferred_free() has been
3430 	 * called.  In the event of a race, two batch queue entries for the page
3431 	 * will be created, but the second will have no effect.
3432 	 */
3433 	if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE, PGA_DEQUEUE))
3434 		vm_page_pqbatch_submit(m, queue);
3435 }
3436 
3437 /*
3438  * A variant of vm_page_dequeue_deferred() that does not assert the page
3439  * lock and is only to be called from vm_page_free_prep().  Because the
3440  * page is being freed, we can assume that nothing other than the page
3441  * daemon is scheduling queue operations on this page, so we get for
3442  * free the mutual exclusion that is otherwise provided by the page lock.
3443  * To handle races, the page daemon must take care to atomically check
3444  * for PGA_DEQUEUE when updating queue state.
3445  */
3446 static void
3447 vm_page_dequeue_deferred_free(vm_page_t m)
3448 {
3449 	uint8_t queue;
3450 
3451 	KASSERT(m->ref_count == 0, ("page %p has references", m));
3452 
3453 	for (;;) {
3454 		if ((m->a.flags & PGA_DEQUEUE) != 0)
3455 			return;
3456 		atomic_thread_fence_acq();
3457 		if ((queue = atomic_load_8(&m->a.queue)) == PQ_NONE)
3458 			return;
3459 		if (vm_page_pqstate_cmpset(m, queue, queue, PGA_DEQUEUE,
3460 		    PGA_DEQUEUE)) {
3461 			vm_page_pqbatch_submit(m, queue);
3462 			break;
3463 		}
3464 	}
3465 }
3466 
3467 /*
3468  *	vm_page_dequeue:
3469  *
3470  *	Remove the page from whichever page queue it's in, if any.
3471  *	The page must either be locked or unallocated.  This constraint
3472  *	ensures that the queue state of the page will remain consistent
3473  *	after this function returns.
3474  */
3475 void
3476 vm_page_dequeue(vm_page_t m)
3477 {
3478 	struct vm_pagequeue *pq, *pq1;
3479 	uint16_t aflags;
3480 
3481 	KASSERT(mtx_owned(vm_page_lockptr(m)) || m->ref_count == 0,
3482 	    ("page %p is allocated and unlocked", m));
3483 
3484 	for (pq = vm_page_pagequeue(m);; pq = pq1) {
3485 		if (pq == NULL) {
3486 			/*
3487 			 * A thread may be concurrently executing
3488 			 * vm_page_dequeue_complete().  Ensure that all queue
3489 			 * state is cleared before we return.
3490 			 */
3491 			aflags = atomic_load_16(&m->a.flags);
3492 			if ((aflags & PGA_QUEUE_STATE_MASK) == 0)
3493 				return;
3494 			KASSERT((aflags & PGA_DEQUEUE) != 0,
3495 			    ("page %p has unexpected queue state flags %#x",
3496 			    m, aflags));
3497 
3498 			/*
3499 			 * Busy wait until the thread updating queue state is
3500 			 * finished.  Such a thread must be executing in a
3501 			 * critical section.
3502 			 */
3503 			cpu_spinwait();
3504 			pq1 = vm_page_pagequeue(m);
3505 			continue;
3506 		}
3507 		vm_pagequeue_lock(pq);
3508 		if ((pq1 = vm_page_pagequeue(m)) == pq)
3509 			break;
3510 		vm_pagequeue_unlock(pq);
3511 	}
3512 	KASSERT(pq == vm_page_pagequeue(m),
3513 	    ("%s: page %p migrated directly between queues", __func__, m));
3514 	KASSERT((m->a.flags & PGA_DEQUEUE) != 0 ||
3515 	    mtx_owned(vm_page_lockptr(m)),
3516 	    ("%s: queued unlocked page %p", __func__, m));
3517 
3518 	if ((m->a.flags & PGA_ENQUEUED) != 0)
3519 		vm_pagequeue_remove(pq, m);
3520 	vm_page_dequeue_complete(m);
3521 	vm_pagequeue_unlock(pq);
3522 }
3523 
3524 /*
3525  * Schedule the given page for insertion into the specified page queue.
3526  * Physical insertion of the page may be deferred indefinitely.
3527  */
3528 static void
3529 vm_page_enqueue(vm_page_t m, uint8_t queue)
3530 {
3531 
3532 	vm_page_assert_locked(m);
3533 	KASSERT(m->a.queue == PQ_NONE &&
3534 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3535 	    ("%s: page %p is already enqueued", __func__, m));
3536 	KASSERT(m->ref_count > 0,
3537 	    ("%s: page %p does not carry any references", __func__, m));
3538 
3539 	m->a.queue = queue;
3540 	if ((m->a.flags & PGA_REQUEUE) == 0)
3541 		vm_page_aflag_set(m, PGA_REQUEUE);
3542 	vm_page_pqbatch_submit(m, queue);
3543 }
3544 
3545 /*
3546  *	vm_page_requeue:		[ internal use only ]
3547  *
3548  *	Schedule a requeue of the given page.
3549  *
3550  *	The page must be locked.
3551  */
3552 void
3553 vm_page_requeue(vm_page_t m)
3554 {
3555 
3556 	vm_page_assert_locked(m);
3557 	KASSERT(vm_page_queue(m) != PQ_NONE,
3558 	    ("%s: page %p is not logically enqueued", __func__, m));
3559 	KASSERT(m->ref_count > 0,
3560 	    ("%s: page %p does not carry any references", __func__, m));
3561 
3562 	if ((m->a.flags & PGA_REQUEUE) == 0)
3563 		vm_page_aflag_set(m, PGA_REQUEUE);
3564 	vm_page_pqbatch_submit(m, atomic_load_8(&m->a.queue));
3565 }
3566 
3567 /*
3568  *	vm_page_swapqueue:		[ internal use only ]
3569  *
3570  *	Move the page from one queue to another, or to the tail of its
3571  *	current queue, in the face of a possible concurrent call to
3572  *	vm_page_dequeue_deferred_free().
3573  */
3574 void
3575 vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq)
3576 {
3577 	struct vm_pagequeue *pq;
3578 	vm_page_t next;
3579 	bool queued;
3580 
3581 	KASSERT(oldq < PQ_COUNT && newq < PQ_COUNT && oldq != newq,
3582 	    ("vm_page_swapqueue: invalid queues (%d, %d)", oldq, newq));
3583 	vm_page_assert_locked(m);
3584 
3585 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[oldq];
3586 	vm_pagequeue_lock(pq);
3587 
3588 	/*
3589 	 * The physical queue state might change at any point before the page
3590 	 * queue lock is acquired, so we must verify that we hold the correct
3591 	 * lock before proceeding.
3592 	 */
3593 	if (__predict_false(m->a.queue != oldq)) {
3594 		vm_pagequeue_unlock(pq);
3595 		return;
3596 	}
3597 
3598 	/*
3599 	 * Once the queue index of the page changes, there is nothing
3600 	 * synchronizing with further updates to the physical queue state.
3601 	 * Therefore we must remove the page from the queue now in anticipation
3602 	 * of a successful commit, and be prepared to roll back.
3603 	 */
3604 	if (__predict_true((m->a.flags & PGA_ENQUEUED) != 0)) {
3605 		next = TAILQ_NEXT(m, plinks.q);
3606 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3607 		vm_page_aflag_clear(m, PGA_ENQUEUED);
3608 		queued = true;
3609 	} else {
3610 		queued = false;
3611 	}
3612 
3613 	/*
3614 	 * Atomically update the queue field and set PGA_REQUEUE while
3615 	 * ensuring that PGA_DEQUEUE has not been set.
3616 	 */
3617 	if (__predict_false(!vm_page_pqstate_cmpset(m, oldq, newq, PGA_DEQUEUE,
3618 	    PGA_REQUEUE))) {
3619 		if (queued) {
3620 			vm_page_aflag_set(m, PGA_ENQUEUED);
3621 			if (next != NULL)
3622 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3623 			else
3624 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3625 		}
3626 		vm_pagequeue_unlock(pq);
3627 		return;
3628 	}
3629 	vm_pagequeue_cnt_dec(pq);
3630 	vm_pagequeue_unlock(pq);
3631 	vm_page_pqbatch_submit(m, newq);
3632 }
3633 
3634 /*
3635  *	vm_page_free_prep:
3636  *
3637  *	Prepares the given page to be put on the free list,
3638  *	disassociating it from any VM object. The caller may return
3639  *	the page to the free list only if this function returns true.
3640  *
3641  *	The object must be locked.  The page must be locked if it is
3642  *	managed.
3643  */
3644 bool
3645 vm_page_free_prep(vm_page_t m)
3646 {
3647 
3648 	/*
3649 	 * Synchronize with threads that have dropped a reference to this
3650 	 * page.
3651 	 */
3652 	atomic_thread_fence_acq();
3653 
3654 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3655 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3656 		uint64_t *p;
3657 		int i;
3658 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3659 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3660 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3661 			    m, i, (uintmax_t)*p));
3662 	}
3663 #endif
3664 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3665 		KASSERT(!pmap_page_is_mapped(m),
3666 		    ("vm_page_free_prep: freeing mapped page %p", m));
3667 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3668 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3669 	} else {
3670 		KASSERT(m->a.queue == PQ_NONE,
3671 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3672 	}
3673 	VM_CNT_INC(v_tfree);
3674 
3675 	if (vm_page_sbusied(m))
3676 		panic("vm_page_free_prep: freeing shared busy page %p", m);
3677 
3678 	if (m->object != NULL) {
3679 		vm_page_object_remove(m);
3680 
3681 		/*
3682 		 * The object reference can be released without an atomic
3683 		 * operation.
3684 		 */
3685 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3686 		    m->ref_count == VPRC_OBJREF,
3687 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3688 		    m, m->ref_count));
3689 		m->object = NULL;
3690 		m->ref_count -= VPRC_OBJREF;
3691 	}
3692 
3693 	if (vm_page_xbusied(m))
3694 		vm_page_xunbusy(m);
3695 
3696 	/*
3697 	 * If fictitious remove object association and
3698 	 * return.
3699 	 */
3700 	if ((m->flags & PG_FICTITIOUS) != 0) {
3701 		KASSERT(m->ref_count == 1,
3702 		    ("fictitious page %p is referenced", m));
3703 		KASSERT(m->a.queue == PQ_NONE,
3704 		    ("fictitious page %p is queued", m));
3705 		return (false);
3706 	}
3707 
3708 	/*
3709 	 * Pages need not be dequeued before they are returned to the physical
3710 	 * memory allocator, but they must at least be marked for a deferred
3711 	 * dequeue.
3712 	 */
3713 	if ((m->oflags & VPO_UNMANAGED) == 0)
3714 		vm_page_dequeue_deferred_free(m);
3715 
3716 	m->valid = 0;
3717 	vm_page_undirty(m);
3718 
3719 	if (m->ref_count != 0)
3720 		panic("vm_page_free_prep: page %p has references", m);
3721 
3722 	/*
3723 	 * Restore the default memory attribute to the page.
3724 	 */
3725 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3726 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3727 
3728 #if VM_NRESERVLEVEL > 0
3729 	/*
3730 	 * Determine whether the page belongs to a reservation.  If the page was
3731 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3732 	 * as an optimization, we avoid the check in that case.
3733 	 */
3734 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3735 		return (false);
3736 #endif
3737 
3738 	return (true);
3739 }
3740 
3741 /*
3742  *	vm_page_free_toq:
3743  *
3744  *	Returns the given page to the free list, disassociating it
3745  *	from any VM object.
3746  *
3747  *	The object must be locked.  The page must be locked if it is
3748  *	managed.
3749  */
3750 void
3751 vm_page_free_toq(vm_page_t m)
3752 {
3753 	struct vm_domain *vmd;
3754 	uma_zone_t zone;
3755 
3756 	if (!vm_page_free_prep(m))
3757 		return;
3758 
3759 	vmd = vm_pagequeue_domain(m);
3760 	zone = vmd->vmd_pgcache[m->pool].zone;
3761 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3762 		uma_zfree(zone, m);
3763 		return;
3764 	}
3765 	vm_domain_free_lock(vmd);
3766 	vm_phys_free_pages(m, 0);
3767 	vm_domain_free_unlock(vmd);
3768 	vm_domain_freecnt_inc(vmd, 1);
3769 }
3770 
3771 /*
3772  *	vm_page_free_pages_toq:
3773  *
3774  *	Returns a list of pages to the free list, disassociating it
3775  *	from any VM object.  In other words, this is equivalent to
3776  *	calling vm_page_free_toq() for each page of a list of VM objects.
3777  *
3778  *	The objects must be locked.  The pages must be locked if it is
3779  *	managed.
3780  */
3781 void
3782 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3783 {
3784 	vm_page_t m;
3785 	int count;
3786 
3787 	if (SLIST_EMPTY(free))
3788 		return;
3789 
3790 	count = 0;
3791 	while ((m = SLIST_FIRST(free)) != NULL) {
3792 		count++;
3793 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3794 		vm_page_free_toq(m);
3795 	}
3796 
3797 	if (update_wire_count)
3798 		vm_wire_sub(count);
3799 }
3800 
3801 /*
3802  * Mark this page as wired down, preventing reclamation by the page daemon
3803  * or when the containing object is destroyed.
3804  */
3805 void
3806 vm_page_wire(vm_page_t m)
3807 {
3808 	u_int old;
3809 
3810 	KASSERT(m->object != NULL,
3811 	    ("vm_page_wire: page %p does not belong to an object", m));
3812 	if (!vm_page_busied(m) && !vm_object_busied(m->object))
3813 		VM_OBJECT_ASSERT_LOCKED(m->object);
3814 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3815 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
3816 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
3817 
3818 	old = atomic_fetchadd_int(&m->ref_count, 1);
3819 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3820 	    ("vm_page_wire: counter overflow for page %p", m));
3821 	if (VPRC_WIRE_COUNT(old) == 0)
3822 		vm_wire_add(1);
3823 }
3824 
3825 /*
3826  * Attempt to wire a mapped page following a pmap lookup of that page.
3827  * This may fail if a thread is concurrently tearing down mappings of the page.
3828  * The transient failure is acceptable because it translates to the
3829  * failure of the caller pmap_extract_and_hold(), which should be then
3830  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3831  */
3832 bool
3833 vm_page_wire_mapped(vm_page_t m)
3834 {
3835 	u_int old;
3836 
3837 	old = m->ref_count;
3838 	do {
3839 		KASSERT(old > 0,
3840 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3841 		if ((old & VPRC_BLOCKED) != 0)
3842 			return (false);
3843 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3844 
3845 	if (VPRC_WIRE_COUNT(old) == 0)
3846 		vm_wire_add(1);
3847 	return (true);
3848 }
3849 
3850 /*
3851  * Release one wiring of the specified page, potentially allowing it to be
3852  * paged out.
3853  *
3854  * Only managed pages belonging to an object can be paged out.  If the number
3855  * of wirings transitions to zero and the page is eligible for page out, then
3856  * the page is added to the specified paging queue.  If the released wiring
3857  * represented the last reference to the page, the page is freed.
3858  *
3859  * A managed page must be locked.
3860  */
3861 void
3862 vm_page_unwire(vm_page_t m, uint8_t queue)
3863 {
3864 	u_int old;
3865 	bool locked;
3866 
3867 	KASSERT(queue < PQ_COUNT,
3868 	    ("vm_page_unwire: invalid queue %u request for page %p", queue, m));
3869 
3870 	if ((m->oflags & VPO_UNMANAGED) != 0) {
3871 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
3872 			vm_page_free(m);
3873 		return;
3874 	}
3875 
3876 	/*
3877 	 * Update LRU state before releasing the wiring reference.
3878 	 * We only need to do this once since we hold the page lock.
3879 	 * Use a release store when updating the reference count to
3880 	 * synchronize with vm_page_free_prep().
3881 	 */
3882 	old = m->ref_count;
3883 	locked = false;
3884 	do {
3885 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
3886 		    ("vm_page_unwire: wire count underflow for page %p", m));
3887 		if (!locked && VPRC_WIRE_COUNT(old) == 1) {
3888 			vm_page_lock(m);
3889 			locked = true;
3890 			if (queue == PQ_ACTIVE && vm_page_queue(m) == PQ_ACTIVE)
3891 				vm_page_reference(m);
3892 			else
3893 				vm_page_mvqueue(m, queue);
3894 		}
3895 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3896 
3897 	/*
3898 	 * Release the lock only after the wiring is released, to ensure that
3899 	 * the page daemon does not encounter and dequeue the page while it is
3900 	 * still wired.
3901 	 */
3902 	if (locked)
3903 		vm_page_unlock(m);
3904 
3905 	if (VPRC_WIRE_COUNT(old) == 1) {
3906 		vm_wire_sub(1);
3907 		if (old == 1)
3908 			vm_page_free(m);
3909 	}
3910 }
3911 
3912 /*
3913  * Unwire a page without (re-)inserting it into a page queue.  It is up
3914  * to the caller to enqueue, requeue, or free the page as appropriate.
3915  * In most cases involving managed pages, vm_page_unwire() should be used
3916  * instead.
3917  */
3918 bool
3919 vm_page_unwire_noq(vm_page_t m)
3920 {
3921 	u_int old;
3922 
3923 	old = vm_page_drop(m, 1);
3924 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
3925 	    ("vm_page_unref: counter underflow for page %p", m));
3926 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3927 	    ("vm_page_unref: missing ref on fictitious page %p", m));
3928 
3929 	if (VPRC_WIRE_COUNT(old) > 1)
3930 		return (false);
3931 	vm_wire_sub(1);
3932 	return (true);
3933 }
3934 
3935 /*
3936  * Ensure that the page is in the specified page queue.  If the page is
3937  * active or being moved to the active queue, ensure that its act_count is
3938  * at least ACT_INIT but do not otherwise mess with it.  Otherwise, ensure that
3939  * the page is at the tail of its page queue.
3940  *
3941  * The page may be wired.  The caller should release its wiring reference
3942  * before releasing the page lock, otherwise the page daemon may immediately
3943  * dequeue the page.
3944  *
3945  * A managed page must be locked.
3946  */
3947 static __always_inline void
3948 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue)
3949 {
3950 
3951 	vm_page_assert_locked(m);
3952 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3953 	    ("vm_page_mvqueue: page %p is unmanaged", m));
3954 	KASSERT(m->ref_count > 0,
3955 	    ("%s: page %p does not carry any references", __func__, m));
3956 
3957 	if (vm_page_queue(m) != nqueue) {
3958 		vm_page_dequeue(m);
3959 		vm_page_enqueue(m, nqueue);
3960 	} else if (nqueue != PQ_ACTIVE) {
3961 		vm_page_requeue(m);
3962 	}
3963 
3964 	if (nqueue == PQ_ACTIVE && m->a.act_count < ACT_INIT)
3965 		m->a.act_count = ACT_INIT;
3966 }
3967 
3968 /*
3969  * Put the specified page on the active list (if appropriate).
3970  */
3971 void
3972 vm_page_activate(vm_page_t m)
3973 {
3974 
3975 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3976 		return;
3977 	vm_page_mvqueue(m, PQ_ACTIVE);
3978 }
3979 
3980 /*
3981  * Move the specified page to the tail of the inactive queue, or requeue
3982  * the page if it is already in the inactive queue.
3983  */
3984 void
3985 vm_page_deactivate(vm_page_t m)
3986 {
3987 
3988 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
3989 		return;
3990 	vm_page_mvqueue(m, PQ_INACTIVE);
3991 }
3992 
3993 /*
3994  * Move the specified page close to the head of the inactive queue,
3995  * bypassing LRU.  A marker page is used to maintain FIFO ordering.
3996  * As with regular enqueues, we use a per-CPU batch queue to reduce
3997  * contention on the page queue lock.
3998  */
3999 static void
4000 _vm_page_deactivate_noreuse(vm_page_t m)
4001 {
4002 
4003 	vm_page_assert_locked(m);
4004 
4005 	if (!vm_page_inactive(m)) {
4006 		vm_page_dequeue(m);
4007 		m->a.queue = PQ_INACTIVE;
4008 	}
4009 	if ((m->a.flags & PGA_REQUEUE_HEAD) == 0)
4010 		vm_page_aflag_set(m, PGA_REQUEUE_HEAD);
4011 	vm_page_pqbatch_submit(m, PQ_INACTIVE);
4012 }
4013 
4014 void
4015 vm_page_deactivate_noreuse(vm_page_t m)
4016 {
4017 
4018 	KASSERT(m->object != NULL,
4019 	    ("vm_page_deactivate_noreuse: page %p has no object", m));
4020 
4021 	if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_wired(m))
4022 		_vm_page_deactivate_noreuse(m);
4023 }
4024 
4025 /*
4026  * Put a page in the laundry, or requeue it if it is already there.
4027  */
4028 void
4029 vm_page_launder(vm_page_t m)
4030 {
4031 
4032 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4033 		return;
4034 	vm_page_mvqueue(m, PQ_LAUNDRY);
4035 }
4036 
4037 /*
4038  * Put a page in the PQ_UNSWAPPABLE holding queue.
4039  */
4040 void
4041 vm_page_unswappable(vm_page_t m)
4042 {
4043 
4044 	vm_page_assert_locked(m);
4045 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4046 	    ("page %p already unswappable", m));
4047 
4048 	vm_page_dequeue(m);
4049 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4050 }
4051 
4052 static void
4053 vm_page_release_toq(vm_page_t m, int flags)
4054 {
4055 
4056 	vm_page_assert_locked(m);
4057 
4058 	/*
4059 	 * Use a check of the valid bits to determine whether we should
4060 	 * accelerate reclamation of the page.  The object lock might not be
4061 	 * held here, in which case the check is racy.  At worst we will either
4062 	 * accelerate reclamation of a valid page and violate LRU, or
4063 	 * unnecessarily defer reclamation of an invalid page.
4064 	 *
4065 	 * If we were asked to not cache the page, place it near the head of the
4066 	 * inactive queue so that is reclaimed sooner.
4067 	 */
4068 	if ((flags & (VPR_TRYFREE | VPR_NOREUSE)) != 0 || m->valid == 0)
4069 		_vm_page_deactivate_noreuse(m);
4070 	else if (vm_page_active(m))
4071 		vm_page_reference(m);
4072 	else
4073 		vm_page_mvqueue(m, PQ_INACTIVE);
4074 }
4075 
4076 /*
4077  * Unwire a page and either attempt to free it or re-add it to the page queues.
4078  */
4079 void
4080 vm_page_release(vm_page_t m, int flags)
4081 {
4082 	vm_object_t object;
4083 	u_int old;
4084 	bool locked;
4085 
4086 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4087 	    ("vm_page_release: page %p is unmanaged", m));
4088 
4089 	if ((flags & VPR_TRYFREE) != 0) {
4090 		for (;;) {
4091 			object = (vm_object_t)atomic_load_ptr(&m->object);
4092 			if (object == NULL)
4093 				break;
4094 			/* Depends on type-stability. */
4095 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) {
4096 				object = NULL;
4097 				break;
4098 			}
4099 			if (object == m->object)
4100 				break;
4101 			VM_OBJECT_WUNLOCK(object);
4102 		}
4103 		if (__predict_true(object != NULL)) {
4104 			vm_page_release_locked(m, flags);
4105 			VM_OBJECT_WUNLOCK(object);
4106 			return;
4107 		}
4108 	}
4109 
4110 	/*
4111 	 * Update LRU state before releasing the wiring reference.
4112 	 * Use a release store when updating the reference count to
4113 	 * synchronize with vm_page_free_prep().
4114 	 */
4115 	old = m->ref_count;
4116 	locked = false;
4117 	do {
4118 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
4119 		    ("vm_page_unwire: wire count underflow for page %p", m));
4120 		if (!locked && VPRC_WIRE_COUNT(old) == 1) {
4121 			vm_page_lock(m);
4122 			locked = true;
4123 			vm_page_release_toq(m, flags);
4124 		}
4125 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
4126 
4127 	/*
4128 	 * Release the lock only after the wiring is released, to ensure that
4129 	 * the page daemon does not encounter and dequeue the page while it is
4130 	 * still wired.
4131 	 */
4132 	if (locked)
4133 		vm_page_unlock(m);
4134 
4135 	if (VPRC_WIRE_COUNT(old) == 1) {
4136 		vm_wire_sub(1);
4137 		if (old == 1)
4138 			vm_page_free(m);
4139 	}
4140 }
4141 
4142 /* See vm_page_release(). */
4143 void
4144 vm_page_release_locked(vm_page_t m, int flags)
4145 {
4146 
4147 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4148 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4149 	    ("vm_page_release_locked: page %p is unmanaged", m));
4150 
4151 	if (vm_page_unwire_noq(m)) {
4152 		if ((flags & VPR_TRYFREE) != 0 &&
4153 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4154 		    m->dirty == 0 && !vm_page_busied(m)) {
4155 			vm_page_free(m);
4156 		} else {
4157 			vm_page_lock(m);
4158 			vm_page_release_toq(m, flags);
4159 			vm_page_unlock(m);
4160 		}
4161 	}
4162 }
4163 
4164 static bool
4165 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4166 {
4167 	u_int old;
4168 
4169 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4170 	    ("vm_page_try_blocked_op: page %p has no object", m));
4171 	KASSERT(vm_page_busied(m),
4172 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4173 	VM_OBJECT_ASSERT_LOCKED(m->object);
4174 
4175 	old = m->ref_count;
4176 	do {
4177 		KASSERT(old != 0,
4178 		    ("vm_page_try_blocked_op: page %p has no references", m));
4179 		if (VPRC_WIRE_COUNT(old) != 0)
4180 			return (false);
4181 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4182 
4183 	(op)(m);
4184 
4185 	/*
4186 	 * If the object is read-locked, new wirings may be created via an
4187 	 * object lookup.
4188 	 */
4189 	old = vm_page_drop(m, VPRC_BLOCKED);
4190 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4191 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4192 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4193 	    old, m));
4194 	return (true);
4195 }
4196 
4197 /*
4198  * Atomically check for wirings and remove all mappings of the page.
4199  */
4200 bool
4201 vm_page_try_remove_all(vm_page_t m)
4202 {
4203 
4204 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4205 }
4206 
4207 /*
4208  * Atomically check for wirings and remove all writeable mappings of the page.
4209  */
4210 bool
4211 vm_page_try_remove_write(vm_page_t m)
4212 {
4213 
4214 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4215 }
4216 
4217 /*
4218  * vm_page_advise
4219  *
4220  * 	Apply the specified advice to the given page.
4221  *
4222  *	The object and page must be locked.
4223  */
4224 void
4225 vm_page_advise(vm_page_t m, int advice)
4226 {
4227 
4228 	vm_page_assert_locked(m);
4229 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4230 	if (advice == MADV_FREE)
4231 		/*
4232 		 * Mark the page clean.  This will allow the page to be freed
4233 		 * without first paging it out.  MADV_FREE pages are often
4234 		 * quickly reused by malloc(3), so we do not do anything that
4235 		 * would result in a page fault on a later access.
4236 		 */
4237 		vm_page_undirty(m);
4238 	else if (advice != MADV_DONTNEED) {
4239 		if (advice == MADV_WILLNEED)
4240 			vm_page_activate(m);
4241 		return;
4242 	}
4243 
4244 	/*
4245 	 * Clear any references to the page.  Otherwise, the page daemon will
4246 	 * immediately reactivate the page.
4247 	 */
4248 	vm_page_aflag_clear(m, PGA_REFERENCED);
4249 
4250 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4251 		vm_page_dirty(m);
4252 
4253 	/*
4254 	 * Place clean pages near the head of the inactive queue rather than
4255 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4256 	 * the page will be reused quickly.  Dirty pages not already in the
4257 	 * laundry are moved there.
4258 	 */
4259 	if (m->dirty == 0)
4260 		vm_page_deactivate_noreuse(m);
4261 	else if (!vm_page_in_laundry(m))
4262 		vm_page_launder(m);
4263 }
4264 
4265 static inline int
4266 vm_page_grab_pflags(int allocflags)
4267 {
4268 	int pflags;
4269 
4270 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4271 	    (allocflags & VM_ALLOC_WIRED) != 0,
4272 	    ("vm_page_grab_pflags: the pages must be busied or wired"));
4273 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4274 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4275 	    ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4276 	    "mismatch"));
4277 	pflags = allocflags &
4278 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4279 	    VM_ALLOC_NOBUSY);
4280 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4281 		pflags |= VM_ALLOC_WAITFAIL;
4282 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4283 		pflags |= VM_ALLOC_SBUSY;
4284 
4285 	return (pflags);
4286 }
4287 
4288 /*
4289  * Grab a page, waiting until we are waken up due to the page
4290  * changing state.  We keep on waiting, if the page continues
4291  * to be in the object.  If the page doesn't exist, first allocate it
4292  * and then conditionally zero it.
4293  *
4294  * This routine may sleep.
4295  *
4296  * The object must be locked on entry.  The lock will, however, be released
4297  * and reacquired if the routine sleeps.
4298  */
4299 vm_page_t
4300 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4301 {
4302 	vm_page_t m;
4303 	int pflags;
4304 
4305 	VM_OBJECT_ASSERT_WLOCKED(object);
4306 	pflags = vm_page_grab_pflags(allocflags);
4307 retrylookup:
4308 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4309 		if (!vm_page_acquire_flags(m, allocflags)) {
4310 			if (vm_page_busy_sleep_flags(object, m, "pgrbwt",
4311 			    allocflags))
4312 				goto retrylookup;
4313 			return (NULL);
4314 		}
4315 		goto out;
4316 	}
4317 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4318 		return (NULL);
4319 	m = vm_page_alloc(object, pindex, pflags);
4320 	if (m == NULL) {
4321 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4322 			return (NULL);
4323 		goto retrylookup;
4324 	}
4325 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4326 		pmap_zero_page(m);
4327 
4328 out:
4329 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4330 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4331 			vm_page_sunbusy(m);
4332 		else
4333 			vm_page_xunbusy(m);
4334 	}
4335 	return (m);
4336 }
4337 
4338 /*
4339  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4340  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4341  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4342  * in simultaneously.  Additional pages will be left on a paging queue but
4343  * will neither be wired nor busy regardless of allocflags.
4344  */
4345 int
4346 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4347 {
4348 	vm_page_t m;
4349 	vm_page_t ma[VM_INITIAL_PAGEIN];
4350 	bool sleep, xbusy;
4351 	int after, i, pflags, rv;
4352 
4353 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4354 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4355 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4356 	KASSERT((allocflags &
4357 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4358 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4359 	VM_OBJECT_ASSERT_WLOCKED(object);
4360 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4361 	pflags |= VM_ALLOC_WAITFAIL;
4362 
4363 retrylookup:
4364 	xbusy = false;
4365 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4366 		/*
4367 		 * If the page is fully valid it can only become invalid
4368 		 * with the object lock held.  If it is not valid it can
4369 		 * become valid with the busy lock held.  Therefore, we
4370 		 * may unnecessarily lock the exclusive busy here if we
4371 		 * race with I/O completion not using the object lock.
4372 		 * However, we will not end up with an invalid page and a
4373 		 * shared lock.
4374 		 */
4375 		if (!vm_page_all_valid(m) ||
4376 		    (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4377 			sleep = !vm_page_tryxbusy(m);
4378 			xbusy = true;
4379 		} else
4380 			sleep = !vm_page_trysbusy(m);
4381 		if (sleep) {
4382 			(void)vm_page_busy_sleep_flags(object, m, "pgrbwt",
4383 			    allocflags);
4384 			goto retrylookup;
4385 		}
4386 		if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4387 		   !vm_page_all_valid(m)) {
4388 			if (xbusy)
4389 				vm_page_xunbusy(m);
4390 			else
4391 				vm_page_sunbusy(m);
4392 			*mp = NULL;
4393 			return (VM_PAGER_FAIL);
4394 		}
4395 		if ((allocflags & VM_ALLOC_WIRED) != 0)
4396 			vm_page_wire(m);
4397 		if (vm_page_all_valid(m))
4398 			goto out;
4399 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4400 		*mp = NULL;
4401 		return (VM_PAGER_FAIL);
4402 	} else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4403 		xbusy = true;
4404 	} else {
4405 		goto retrylookup;
4406 	}
4407 
4408 	vm_page_assert_xbusied(m);
4409 	MPASS(xbusy);
4410 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4411 		after = MIN(after, VM_INITIAL_PAGEIN);
4412 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4413 		after = MAX(after, 1);
4414 		ma[0] = m;
4415 		for (i = 1; i < after; i++) {
4416 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4417 				if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4418 					break;
4419 			} else {
4420 				ma[i] = vm_page_alloc(object, m->pindex + i,
4421 				    VM_ALLOC_NORMAL);
4422 				if (ma[i] == NULL)
4423 					break;
4424 			}
4425 		}
4426 		after = i;
4427 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4428 		/* Pager may have replaced a page. */
4429 		m = ma[0];
4430 		if (rv != VM_PAGER_OK) {
4431 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4432 				vm_page_unwire_noq(m);
4433 			for (i = 0; i < after; i++) {
4434 				if (!vm_page_wired(ma[i]))
4435 					vm_page_free(ma[i]);
4436 				else
4437 					vm_page_xunbusy(ma[i]);
4438 			}
4439 			*mp = NULL;
4440 			return (rv);
4441 		}
4442 		for (i = 1; i < after; i++)
4443 			vm_page_readahead_finish(ma[i]);
4444 		MPASS(vm_page_all_valid(m));
4445 	} else {
4446 		vm_page_zero_invalid(m, TRUE);
4447 	}
4448 out:
4449 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4450 		if (xbusy)
4451 			vm_page_xunbusy(m);
4452 		else
4453 			vm_page_sunbusy(m);
4454 	}
4455 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4456 		vm_page_busy_downgrade(m);
4457 	*mp = m;
4458 	return (VM_PAGER_OK);
4459 }
4460 
4461 /*
4462  * Return the specified range of pages from the given object.  For each
4463  * page offset within the range, if a page already exists within the object
4464  * at that offset and it is busy, then wait for it to change state.  If,
4465  * instead, the page doesn't exist, then allocate it.
4466  *
4467  * The caller must always specify an allocation class.
4468  *
4469  * allocation classes:
4470  *	VM_ALLOC_NORMAL		normal process request
4471  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4472  *
4473  * The caller must always specify that the pages are to be busied and/or
4474  * wired.
4475  *
4476  * optional allocation flags:
4477  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4478  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4479  *	VM_ALLOC_NOWAIT		do not sleep
4480  *	VM_ALLOC_SBUSY		set page to sbusy state
4481  *	VM_ALLOC_WIRED		wire the pages
4482  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4483  *
4484  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4485  * may return a partial prefix of the requested range.
4486  */
4487 int
4488 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4489     vm_page_t *ma, int count)
4490 {
4491 	vm_page_t m, mpred;
4492 	int pflags;
4493 	int i;
4494 
4495 	VM_OBJECT_ASSERT_WLOCKED(object);
4496 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4497 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4498 
4499 	pflags = vm_page_grab_pflags(allocflags);
4500 	if (count == 0)
4501 		return (0);
4502 
4503 	i = 0;
4504 retrylookup:
4505 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4506 	if (m == NULL || m->pindex != pindex + i) {
4507 		mpred = m;
4508 		m = NULL;
4509 	} else
4510 		mpred = TAILQ_PREV(m, pglist, listq);
4511 	for (; i < count; i++) {
4512 		if (m != NULL) {
4513 			if (!vm_page_acquire_flags(m, allocflags)) {
4514 				if (vm_page_busy_sleep_flags(object, m,
4515 				    "grbmaw", allocflags))
4516 					goto retrylookup;
4517 				break;
4518 			}
4519 		} else {
4520 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4521 				break;
4522 			m = vm_page_alloc_after(object, pindex + i,
4523 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4524 			if (m == NULL) {
4525 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4526 					break;
4527 				goto retrylookup;
4528 			}
4529 		}
4530 		if (vm_page_none_valid(m) &&
4531 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4532 			if ((m->flags & PG_ZERO) == 0)
4533 				pmap_zero_page(m);
4534 			vm_page_valid(m);
4535 		}
4536 		if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4537 			if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4538 				vm_page_sunbusy(m);
4539 			else
4540 				vm_page_xunbusy(m);
4541 		}
4542 		ma[i] = mpred = m;
4543 		m = vm_page_next(m);
4544 	}
4545 	return (i);
4546 }
4547 
4548 /*
4549  * Mapping function for valid or dirty bits in a page.
4550  *
4551  * Inputs are required to range within a page.
4552  */
4553 vm_page_bits_t
4554 vm_page_bits(int base, int size)
4555 {
4556 	int first_bit;
4557 	int last_bit;
4558 
4559 	KASSERT(
4560 	    base + size <= PAGE_SIZE,
4561 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4562 	);
4563 
4564 	if (size == 0)		/* handle degenerate case */
4565 		return (0);
4566 
4567 	first_bit = base >> DEV_BSHIFT;
4568 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4569 
4570 	return (((vm_page_bits_t)2 << last_bit) -
4571 	    ((vm_page_bits_t)1 << first_bit));
4572 }
4573 
4574 void
4575 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4576 {
4577 
4578 #if PAGE_SIZE == 32768
4579 	atomic_set_64((uint64_t *)bits, set);
4580 #elif PAGE_SIZE == 16384
4581 	atomic_set_32((uint32_t *)bits, set);
4582 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4583 	atomic_set_16((uint16_t *)bits, set);
4584 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4585 	atomic_set_8((uint8_t *)bits, set);
4586 #else		/* PAGE_SIZE <= 8192 */
4587 	uintptr_t addr;
4588 	int shift;
4589 
4590 	addr = (uintptr_t)bits;
4591 	/*
4592 	 * Use a trick to perform a 32-bit atomic on the
4593 	 * containing aligned word, to not depend on the existence
4594 	 * of atomic_{set, clear}_{8, 16}.
4595 	 */
4596 	shift = addr & (sizeof(uint32_t) - 1);
4597 #if BYTE_ORDER == BIG_ENDIAN
4598 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4599 #else
4600 	shift *= NBBY;
4601 #endif
4602 	addr &= ~(sizeof(uint32_t) - 1);
4603 	atomic_set_32((uint32_t *)addr, set << shift);
4604 #endif		/* PAGE_SIZE */
4605 }
4606 
4607 static inline void
4608 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4609 {
4610 
4611 #if PAGE_SIZE == 32768
4612 	atomic_clear_64((uint64_t *)bits, clear);
4613 #elif PAGE_SIZE == 16384
4614 	atomic_clear_32((uint32_t *)bits, clear);
4615 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4616 	atomic_clear_16((uint16_t *)bits, clear);
4617 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4618 	atomic_clear_8((uint8_t *)bits, clear);
4619 #else		/* PAGE_SIZE <= 8192 */
4620 	uintptr_t addr;
4621 	int shift;
4622 
4623 	addr = (uintptr_t)bits;
4624 	/*
4625 	 * Use a trick to perform a 32-bit atomic on the
4626 	 * containing aligned word, to not depend on the existence
4627 	 * of atomic_{set, clear}_{8, 16}.
4628 	 */
4629 	shift = addr & (sizeof(uint32_t) - 1);
4630 #if BYTE_ORDER == BIG_ENDIAN
4631 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4632 #else
4633 	shift *= NBBY;
4634 #endif
4635 	addr &= ~(sizeof(uint32_t) - 1);
4636 	atomic_clear_32((uint32_t *)addr, clear << shift);
4637 #endif		/* PAGE_SIZE */
4638 }
4639 
4640 static inline vm_page_bits_t
4641 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4642 {
4643 #if PAGE_SIZE == 32768
4644 	uint64_t old;
4645 
4646 	old = *bits;
4647 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4648 	return (old);
4649 #elif PAGE_SIZE == 16384
4650 	uint32_t old;
4651 
4652 	old = *bits;
4653 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4654 	return (old);
4655 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4656 	uint16_t old;
4657 
4658 	old = *bits;
4659 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4660 	return (old);
4661 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4662 	uint8_t old;
4663 
4664 	old = *bits;
4665 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4666 	return (old);
4667 #else		/* PAGE_SIZE <= 4096*/
4668 	uintptr_t addr;
4669 	uint32_t old, new, mask;
4670 	int shift;
4671 
4672 	addr = (uintptr_t)bits;
4673 	/*
4674 	 * Use a trick to perform a 32-bit atomic on the
4675 	 * containing aligned word, to not depend on the existence
4676 	 * of atomic_{set, swap, clear}_{8, 16}.
4677 	 */
4678 	shift = addr & (sizeof(uint32_t) - 1);
4679 #if BYTE_ORDER == BIG_ENDIAN
4680 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4681 #else
4682 	shift *= NBBY;
4683 #endif
4684 	addr &= ~(sizeof(uint32_t) - 1);
4685 	mask = VM_PAGE_BITS_ALL << shift;
4686 
4687 	old = *bits;
4688 	do {
4689 		new = old & ~mask;
4690 		new |= newbits << shift;
4691 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4692 	return (old >> shift);
4693 #endif		/* PAGE_SIZE */
4694 }
4695 
4696 /*
4697  *	vm_page_set_valid_range:
4698  *
4699  *	Sets portions of a page valid.  The arguments are expected
4700  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4701  *	of any partial chunks touched by the range.  The invalid portion of
4702  *	such chunks will be zeroed.
4703  *
4704  *	(base + size) must be less then or equal to PAGE_SIZE.
4705  */
4706 void
4707 vm_page_set_valid_range(vm_page_t m, int base, int size)
4708 {
4709 	int endoff, frag;
4710 	vm_page_bits_t pagebits;
4711 
4712 	vm_page_assert_busied(m);
4713 	if (size == 0)	/* handle degenerate case */
4714 		return;
4715 
4716 	/*
4717 	 * If the base is not DEV_BSIZE aligned and the valid
4718 	 * bit is clear, we have to zero out a portion of the
4719 	 * first block.
4720 	 */
4721 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4722 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4723 		pmap_zero_page_area(m, frag, base - frag);
4724 
4725 	/*
4726 	 * If the ending offset is not DEV_BSIZE aligned and the
4727 	 * valid bit is clear, we have to zero out a portion of
4728 	 * the last block.
4729 	 */
4730 	endoff = base + size;
4731 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4732 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4733 		pmap_zero_page_area(m, endoff,
4734 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4735 
4736 	/*
4737 	 * Assert that no previously invalid block that is now being validated
4738 	 * is already dirty.
4739 	 */
4740 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4741 	    ("vm_page_set_valid_range: page %p is dirty", m));
4742 
4743 	/*
4744 	 * Set valid bits inclusive of any overlap.
4745 	 */
4746 	pagebits = vm_page_bits(base, size);
4747 	if (vm_page_xbusied(m))
4748 		m->valid |= pagebits;
4749 	else
4750 		vm_page_bits_set(m, &m->valid, pagebits);
4751 }
4752 
4753 /*
4754  * Set the page dirty bits and free the invalid swap space if
4755  * present.  Returns the previous dirty bits.
4756  */
4757 vm_page_bits_t
4758 vm_page_set_dirty(vm_page_t m)
4759 {
4760 	vm_page_bits_t old;
4761 
4762 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
4763 
4764 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
4765 		old = m->dirty;
4766 		m->dirty = VM_PAGE_BITS_ALL;
4767 	} else
4768 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
4769 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
4770 		vm_pager_page_unswapped(m);
4771 
4772 	return (old);
4773 }
4774 
4775 /*
4776  * Clear the given bits from the specified page's dirty field.
4777  */
4778 static __inline void
4779 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4780 {
4781 
4782 	vm_page_assert_busied(m);
4783 
4784 	/*
4785 	 * If the page is xbusied and not write mapped we are the
4786 	 * only thread that can modify dirty bits.  Otherwise, The pmap
4787 	 * layer can call vm_page_dirty() without holding a distinguished
4788 	 * lock.  The combination of page busy and atomic operations
4789 	 * suffice to guarantee consistency of the page dirty field.
4790 	 */
4791 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4792 		m->dirty &= ~pagebits;
4793 	else
4794 		vm_page_bits_clear(m, &m->dirty, pagebits);
4795 }
4796 
4797 /*
4798  *	vm_page_set_validclean:
4799  *
4800  *	Sets portions of a page valid and clean.  The arguments are expected
4801  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4802  *	of any partial chunks touched by the range.  The invalid portion of
4803  *	such chunks will be zero'd.
4804  *
4805  *	(base + size) must be less then or equal to PAGE_SIZE.
4806  */
4807 void
4808 vm_page_set_validclean(vm_page_t m, int base, int size)
4809 {
4810 	vm_page_bits_t oldvalid, pagebits;
4811 	int endoff, frag;
4812 
4813 	vm_page_assert_busied(m);
4814 	if (size == 0)	/* handle degenerate case */
4815 		return;
4816 
4817 	/*
4818 	 * If the base is not DEV_BSIZE aligned and the valid
4819 	 * bit is clear, we have to zero out a portion of the
4820 	 * first block.
4821 	 */
4822 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4823 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4824 		pmap_zero_page_area(m, frag, base - frag);
4825 
4826 	/*
4827 	 * If the ending offset is not DEV_BSIZE aligned and the
4828 	 * valid bit is clear, we have to zero out a portion of
4829 	 * the last block.
4830 	 */
4831 	endoff = base + size;
4832 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4833 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4834 		pmap_zero_page_area(m, endoff,
4835 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4836 
4837 	/*
4838 	 * Set valid, clear dirty bits.  If validating the entire
4839 	 * page we can safely clear the pmap modify bit.  We also
4840 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
4841 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4842 	 * be set again.
4843 	 *
4844 	 * We set valid bits inclusive of any overlap, but we can only
4845 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4846 	 * the range.
4847 	 */
4848 	oldvalid = m->valid;
4849 	pagebits = vm_page_bits(base, size);
4850 	if (vm_page_xbusied(m))
4851 		m->valid |= pagebits;
4852 	else
4853 		vm_page_bits_set(m, &m->valid, pagebits);
4854 #if 0	/* NOT YET */
4855 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4856 		frag = DEV_BSIZE - frag;
4857 		base += frag;
4858 		size -= frag;
4859 		if (size < 0)
4860 			size = 0;
4861 	}
4862 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4863 #endif
4864 	if (base == 0 && size == PAGE_SIZE) {
4865 		/*
4866 		 * The page can only be modified within the pmap if it is
4867 		 * mapped, and it can only be mapped if it was previously
4868 		 * fully valid.
4869 		 */
4870 		if (oldvalid == VM_PAGE_BITS_ALL)
4871 			/*
4872 			 * Perform the pmap_clear_modify() first.  Otherwise,
4873 			 * a concurrent pmap operation, such as
4874 			 * pmap_protect(), could clear a modification in the
4875 			 * pmap and set the dirty field on the page before
4876 			 * pmap_clear_modify() had begun and after the dirty
4877 			 * field was cleared here.
4878 			 */
4879 			pmap_clear_modify(m);
4880 		m->dirty = 0;
4881 		vm_page_aflag_clear(m, PGA_NOSYNC);
4882 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
4883 		m->dirty &= ~pagebits;
4884 	else
4885 		vm_page_clear_dirty_mask(m, pagebits);
4886 }
4887 
4888 void
4889 vm_page_clear_dirty(vm_page_t m, int base, int size)
4890 {
4891 
4892 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4893 }
4894 
4895 /*
4896  *	vm_page_set_invalid:
4897  *
4898  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4899  *	valid and dirty bits for the effected areas are cleared.
4900  */
4901 void
4902 vm_page_set_invalid(vm_page_t m, int base, int size)
4903 {
4904 	vm_page_bits_t bits;
4905 	vm_object_t object;
4906 
4907 	/*
4908 	 * The object lock is required so that pages can't be mapped
4909 	 * read-only while we're in the process of invalidating them.
4910 	 */
4911 	object = m->object;
4912 	VM_OBJECT_ASSERT_WLOCKED(object);
4913 	vm_page_assert_busied(m);
4914 
4915 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4916 	    size >= object->un_pager.vnp.vnp_size)
4917 		bits = VM_PAGE_BITS_ALL;
4918 	else
4919 		bits = vm_page_bits(base, size);
4920 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
4921 		pmap_remove_all(m);
4922 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
4923 	    !pmap_page_is_mapped(m),
4924 	    ("vm_page_set_invalid: page %p is mapped", m));
4925 	if (vm_page_xbusied(m)) {
4926 		m->valid &= ~bits;
4927 		m->dirty &= ~bits;
4928 	} else {
4929 		vm_page_bits_clear(m, &m->valid, bits);
4930 		vm_page_bits_clear(m, &m->dirty, bits);
4931 	}
4932 }
4933 
4934 /*
4935  *	vm_page_invalid:
4936  *
4937  *	Invalidates the entire page.  The page must be busy, unmapped, and
4938  *	the enclosing object must be locked.  The object locks protects
4939  *	against concurrent read-only pmap enter which is done without
4940  *	busy.
4941  */
4942 void
4943 vm_page_invalid(vm_page_t m)
4944 {
4945 
4946 	vm_page_assert_busied(m);
4947 	VM_OBJECT_ASSERT_LOCKED(m->object);
4948 	MPASS(!pmap_page_is_mapped(m));
4949 
4950 	if (vm_page_xbusied(m))
4951 		m->valid = 0;
4952 	else
4953 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
4954 }
4955 
4956 /*
4957  * vm_page_zero_invalid()
4958  *
4959  *	The kernel assumes that the invalid portions of a page contain
4960  *	garbage, but such pages can be mapped into memory by user code.
4961  *	When this occurs, we must zero out the non-valid portions of the
4962  *	page so user code sees what it expects.
4963  *
4964  *	Pages are most often semi-valid when the end of a file is mapped
4965  *	into memory and the file's size is not page aligned.
4966  */
4967 void
4968 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4969 {
4970 	int b;
4971 	int i;
4972 
4973 	/*
4974 	 * Scan the valid bits looking for invalid sections that
4975 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
4976 	 * valid bit may be set ) have already been zeroed by
4977 	 * vm_page_set_validclean().
4978 	 */
4979 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
4980 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
4981 		    (m->valid & ((vm_page_bits_t)1 << i))) {
4982 			if (i > b) {
4983 				pmap_zero_page_area(m,
4984 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
4985 			}
4986 			b = i + 1;
4987 		}
4988 	}
4989 
4990 	/*
4991 	 * setvalid is TRUE when we can safely set the zero'd areas
4992 	 * as being valid.  We can do this if there are no cache consistancy
4993 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
4994 	 */
4995 	if (setvalid)
4996 		vm_page_valid(m);
4997 }
4998 
4999 /*
5000  *	vm_page_is_valid:
5001  *
5002  *	Is (partial) page valid?  Note that the case where size == 0
5003  *	will return FALSE in the degenerate case where the page is
5004  *	entirely invalid, and TRUE otherwise.
5005  *
5006  *	Some callers envoke this routine without the busy lock held and
5007  *	handle races via higher level locks.  Typical callers should
5008  *	hold a busy lock to prevent invalidation.
5009  */
5010 int
5011 vm_page_is_valid(vm_page_t m, int base, int size)
5012 {
5013 	vm_page_bits_t bits;
5014 
5015 	bits = vm_page_bits(base, size);
5016 	return (m->valid != 0 && (m->valid & bits) == bits);
5017 }
5018 
5019 /*
5020  * Returns true if all of the specified predicates are true for the entire
5021  * (super)page and false otherwise.
5022  */
5023 bool
5024 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5025 {
5026 	vm_object_t object;
5027 	int i, npages;
5028 
5029 	object = m->object;
5030 	if (skip_m != NULL && skip_m->object != object)
5031 		return (false);
5032 	VM_OBJECT_ASSERT_LOCKED(object);
5033 	npages = atop(pagesizes[m->psind]);
5034 
5035 	/*
5036 	 * The physically contiguous pages that make up a superpage, i.e., a
5037 	 * page with a page size index ("psind") greater than zero, will
5038 	 * occupy adjacent entries in vm_page_array[].
5039 	 */
5040 	for (i = 0; i < npages; i++) {
5041 		/* Always test object consistency, including "skip_m". */
5042 		if (m[i].object != object)
5043 			return (false);
5044 		if (&m[i] == skip_m)
5045 			continue;
5046 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5047 			return (false);
5048 		if ((flags & PS_ALL_DIRTY) != 0) {
5049 			/*
5050 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5051 			 * might stop this case from spuriously returning
5052 			 * "false".  However, that would require a write lock
5053 			 * on the object containing "m[i]".
5054 			 */
5055 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5056 				return (false);
5057 		}
5058 		if ((flags & PS_ALL_VALID) != 0 &&
5059 		    m[i].valid != VM_PAGE_BITS_ALL)
5060 			return (false);
5061 	}
5062 	return (true);
5063 }
5064 
5065 /*
5066  * Set the page's dirty bits if the page is modified.
5067  */
5068 void
5069 vm_page_test_dirty(vm_page_t m)
5070 {
5071 
5072 	vm_page_assert_busied(m);
5073 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5074 		vm_page_dirty(m);
5075 }
5076 
5077 void
5078 vm_page_valid(vm_page_t m)
5079 {
5080 
5081 	vm_page_assert_busied(m);
5082 	if (vm_page_xbusied(m))
5083 		m->valid = VM_PAGE_BITS_ALL;
5084 	else
5085 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5086 }
5087 
5088 void
5089 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5090 {
5091 
5092 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5093 }
5094 
5095 void
5096 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5097 {
5098 
5099 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5100 }
5101 
5102 int
5103 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5104 {
5105 
5106 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5107 }
5108 
5109 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5110 void
5111 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5112 {
5113 
5114 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5115 }
5116 
5117 void
5118 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5119 {
5120 
5121 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5122 }
5123 #endif
5124 
5125 #ifdef INVARIANTS
5126 void
5127 vm_page_object_busy_assert(vm_page_t m)
5128 {
5129 
5130 	/*
5131 	 * Certain of the page's fields may only be modified by the
5132 	 * holder of a page or object busy.
5133 	 */
5134 	if (m->object != NULL && !vm_page_busied(m))
5135 		VM_OBJECT_ASSERT_BUSY(m->object);
5136 }
5137 
5138 void
5139 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5140 {
5141 
5142 	if ((bits & PGA_WRITEABLE) == 0)
5143 		return;
5144 
5145 	/*
5146 	 * The PGA_WRITEABLE flag can only be set if the page is
5147 	 * managed, is exclusively busied or the object is locked.
5148 	 * Currently, this flag is only set by pmap_enter().
5149 	 */
5150 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5151 	    ("PGA_WRITEABLE on unmanaged page"));
5152 	if (!vm_page_xbusied(m))
5153 		VM_OBJECT_ASSERT_BUSY(m->object);
5154 }
5155 #endif
5156 
5157 #include "opt_ddb.h"
5158 #ifdef DDB
5159 #include <sys/kernel.h>
5160 
5161 #include <ddb/ddb.h>
5162 
5163 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5164 {
5165 
5166 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5167 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5168 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5169 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5170 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5171 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5172 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5173 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5174 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5175 }
5176 
5177 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5178 {
5179 	int dom;
5180 
5181 	db_printf("pq_free %d\n", vm_free_count());
5182 	for (dom = 0; dom < vm_ndomains; dom++) {
5183 		db_printf(
5184     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5185 		    dom,
5186 		    vm_dom[dom].vmd_page_count,
5187 		    vm_dom[dom].vmd_free_count,
5188 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5189 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5190 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5191 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5192 	}
5193 }
5194 
5195 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5196 {
5197 	vm_page_t m;
5198 	boolean_t phys, virt;
5199 
5200 	if (!have_addr) {
5201 		db_printf("show pginfo addr\n");
5202 		return;
5203 	}
5204 
5205 	phys = strchr(modif, 'p') != NULL;
5206 	virt = strchr(modif, 'v') != NULL;
5207 	if (virt)
5208 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5209 	else if (phys)
5210 		m = PHYS_TO_VM_PAGE(addr);
5211 	else
5212 		m = (vm_page_t)addr;
5213 	db_printf(
5214     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
5215     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5216 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5217 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5218 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5219 }
5220 #endif /* DDB */
5221