xref: /freebsd/sys/vm/vm_page.c (revision 4b50c451720d8b427757a6da1dd2bb4c52cd9e35)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1998 Matthew Dillon.  All Rights Reserved.
7  *
8  * This code is derived from software contributed to Berkeley by
9  * The Mach Operating System project at Carnegie-Mellon University.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
36  */
37 
38 /*-
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  *	Resident memory management module.
67  */
68 
69 #include <sys/cdefs.h>
70 __FBSDID("$FreeBSD$");
71 
72 #include "opt_vm.h"
73 
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/counter.h>
77 #include <sys/domainset.h>
78 #include <sys/kernel.h>
79 #include <sys/limits.h>
80 #include <sys/linker.h>
81 #include <sys/lock.h>
82 #include <sys/malloc.h>
83 #include <sys/mman.h>
84 #include <sys/msgbuf.h>
85 #include <sys/mutex.h>
86 #include <sys/proc.h>
87 #include <sys/rwlock.h>
88 #include <sys/sleepqueue.h>
89 #include <sys/sbuf.h>
90 #include <sys/sched.h>
91 #include <sys/smp.h>
92 #include <sys/sysctl.h>
93 #include <sys/vmmeter.h>
94 #include <sys/vnode.h>
95 
96 #include <vm/vm.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_domainset.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_phys.h>
106 #include <vm/vm_pagequeue.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/vm_extern.h>
111 #include <vm/uma.h>
112 #include <vm/uma_int.h>
113 
114 #include <machine/md_var.h>
115 
116 extern int	uma_startup_count(int);
117 extern void	uma_startup(void *, int);
118 extern int	vmem_startup_count(void);
119 
120 struct vm_domain vm_dom[MAXMEMDOM];
121 
122 DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]);
123 
124 struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT];
125 
126 struct mtx_padalign __exclusive_cache_line vm_domainset_lock;
127 /* The following fields are protected by the domainset lock. */
128 domainset_t __exclusive_cache_line vm_min_domains;
129 domainset_t __exclusive_cache_line vm_severe_domains;
130 static int vm_min_waiters;
131 static int vm_severe_waiters;
132 static int vm_pageproc_waiters;
133 
134 static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0,
135     "VM page statistics");
136 
137 static counter_u64_t pqstate_commit_retries = EARLY_COUNTER;
138 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
139     CTLFLAG_RD, &pqstate_commit_retries,
140     "Number of failed per-page atomic queue state updates");
141 
142 static counter_u64_t queue_ops = EARLY_COUNTER;
143 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
144     CTLFLAG_RD, &queue_ops,
145     "Number of batched queue operations");
146 
147 static counter_u64_t queue_nops = EARLY_COUNTER;
148 SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
149     CTLFLAG_RD, &queue_nops,
150     "Number of batched queue operations with no effects");
151 
152 static void
153 counter_startup(void)
154 {
155 
156 	pqstate_commit_retries = counter_u64_alloc(M_WAITOK);
157 	queue_ops = counter_u64_alloc(M_WAITOK);
158 	queue_nops = counter_u64_alloc(M_WAITOK);
159 }
160 SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
161 
162 /*
163  * bogus page -- for I/O to/from partially complete buffers,
164  * or for paging into sparsely invalid regions.
165  */
166 vm_page_t bogus_page;
167 
168 vm_page_t vm_page_array;
169 long vm_page_array_size;
170 long first_page;
171 
172 static int boot_pages;
173 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
174     &boot_pages, 0,
175     "number of pages allocated for bootstrapping the VM system");
176 
177 static TAILQ_HEAD(, vm_page) blacklist_head;
178 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
179 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
180     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
181 
182 static uma_zone_t fakepg_zone;
183 
184 static void vm_page_alloc_check(vm_page_t m);
185 static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
186     const char *wmesg, bool nonshared, bool locked);
187 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
188 static void vm_page_enqueue(vm_page_t m, uint8_t queue);
189 static bool vm_page_free_prep(vm_page_t m);
190 static void vm_page_free_toq(vm_page_t m);
191 static void vm_page_init(void *dummy);
192 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
193     vm_pindex_t pindex, vm_page_t mpred);
194 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
195     vm_page_t mpred);
196 static void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
197     const uint16_t nflag);
198 static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
199     vm_page_t m_run, vm_paddr_t high);
200 static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
201 static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
202     int req);
203 static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
204     int flags);
205 static void vm_page_zone_release(void *arg, void **store, int cnt);
206 
207 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL);
208 
209 static void
210 vm_page_init(void *dummy)
211 {
212 
213 	fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
214 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
215 	bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ |
216 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
217 }
218 
219 /*
220  * The cache page zone is initialized later since we need to be able to allocate
221  * pages before UMA is fully initialized.
222  */
223 static void
224 vm_page_init_cache_zones(void *dummy __unused)
225 {
226 	struct vm_domain *vmd;
227 	struct vm_pgcache *pgcache;
228 	int cache, domain, maxcache, pool;
229 
230 	maxcache = 0;
231 	TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
232 	maxcache *= mp_ncpus;
233 	for (domain = 0; domain < vm_ndomains; domain++) {
234 		vmd = VM_DOMAIN(domain);
235 		for (pool = 0; pool < VM_NFREEPOOL; pool++) {
236 			pgcache = &vmd->vmd_pgcache[pool];
237 			pgcache->domain = domain;
238 			pgcache->pool = pool;
239 			pgcache->zone = uma_zcache_create("vm pgcache",
240 			    PAGE_SIZE, NULL, NULL, NULL, NULL,
241 			    vm_page_zone_import, vm_page_zone_release, pgcache,
242 			    UMA_ZONE_VM);
243 
244 			/*
245 			 * Limit each pool's zone to 0.1% of the pages in the
246 			 * domain.
247 			 */
248 			cache = maxcache != 0 ? maxcache :
249 			    vmd->vmd_page_count / 1000;
250 			uma_zone_set_maxcache(pgcache->zone, cache);
251 		}
252 	}
253 }
254 SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL);
255 
256 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
257 #if PAGE_SIZE == 32768
258 #ifdef CTASSERT
259 CTASSERT(sizeof(u_long) >= 8);
260 #endif
261 #endif
262 
263 /*
264  *	vm_set_page_size:
265  *
266  *	Sets the page size, perhaps based upon the memory
267  *	size.  Must be called before any use of page-size
268  *	dependent functions.
269  */
270 void
271 vm_set_page_size(void)
272 {
273 	if (vm_cnt.v_page_size == 0)
274 		vm_cnt.v_page_size = PAGE_SIZE;
275 	if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
276 		panic("vm_set_page_size: page size not a power of two");
277 }
278 
279 /*
280  *	vm_page_blacklist_next:
281  *
282  *	Find the next entry in the provided string of blacklist
283  *	addresses.  Entries are separated by space, comma, or newline.
284  *	If an invalid integer is encountered then the rest of the
285  *	string is skipped.  Updates the list pointer to the next
286  *	character, or NULL if the string is exhausted or invalid.
287  */
288 static vm_paddr_t
289 vm_page_blacklist_next(char **list, char *end)
290 {
291 	vm_paddr_t bad;
292 	char *cp, *pos;
293 
294 	if (list == NULL || *list == NULL)
295 		return (0);
296 	if (**list =='\0') {
297 		*list = NULL;
298 		return (0);
299 	}
300 
301 	/*
302 	 * If there's no end pointer then the buffer is coming from
303 	 * the kenv and we know it's null-terminated.
304 	 */
305 	if (end == NULL)
306 		end = *list + strlen(*list);
307 
308 	/* Ensure that strtoq() won't walk off the end */
309 	if (*end != '\0') {
310 		if (*end == '\n' || *end == ' ' || *end  == ',')
311 			*end = '\0';
312 		else {
313 			printf("Blacklist not terminated, skipping\n");
314 			*list = NULL;
315 			return (0);
316 		}
317 	}
318 
319 	for (pos = *list; *pos != '\0'; pos = cp) {
320 		bad = strtoq(pos, &cp, 0);
321 		if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
322 			if (bad == 0) {
323 				if (++cp < end)
324 					continue;
325 				else
326 					break;
327 			}
328 		} else
329 			break;
330 		if (*cp == '\0' || ++cp >= end)
331 			*list = NULL;
332 		else
333 			*list = cp;
334 		return (trunc_page(bad));
335 	}
336 	printf("Garbage in RAM blacklist, skipping\n");
337 	*list = NULL;
338 	return (0);
339 }
340 
341 bool
342 vm_page_blacklist_add(vm_paddr_t pa, bool verbose)
343 {
344 	struct vm_domain *vmd;
345 	vm_page_t m;
346 	int ret;
347 
348 	m = vm_phys_paddr_to_vm_page(pa);
349 	if (m == NULL)
350 		return (true); /* page does not exist, no failure */
351 
352 	vmd = vm_pagequeue_domain(m);
353 	vm_domain_free_lock(vmd);
354 	ret = vm_phys_unfree_page(m);
355 	vm_domain_free_unlock(vmd);
356 	if (ret != 0) {
357 		vm_domain_freecnt_inc(vmd, -1);
358 		TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
359 		if (verbose)
360 			printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa);
361 	}
362 	return (ret);
363 }
364 
365 /*
366  *	vm_page_blacklist_check:
367  *
368  *	Iterate through the provided string of blacklist addresses, pulling
369  *	each entry out of the physical allocator free list and putting it
370  *	onto a list for reporting via the vm.page_blacklist sysctl.
371  */
372 static void
373 vm_page_blacklist_check(char *list, char *end)
374 {
375 	vm_paddr_t pa;
376 	char *next;
377 
378 	next = list;
379 	while (next != NULL) {
380 		if ((pa = vm_page_blacklist_next(&next, end)) == 0)
381 			continue;
382 		vm_page_blacklist_add(pa, bootverbose);
383 	}
384 }
385 
386 /*
387  *	vm_page_blacklist_load:
388  *
389  *	Search for a special module named "ram_blacklist".  It'll be a
390  *	plain text file provided by the user via the loader directive
391  *	of the same name.
392  */
393 static void
394 vm_page_blacklist_load(char **list, char **end)
395 {
396 	void *mod;
397 	u_char *ptr;
398 	u_int len;
399 
400 	mod = NULL;
401 	ptr = NULL;
402 
403 	mod = preload_search_by_type("ram_blacklist");
404 	if (mod != NULL) {
405 		ptr = preload_fetch_addr(mod);
406 		len = preload_fetch_size(mod);
407         }
408 	*list = ptr;
409 	if (ptr != NULL)
410 		*end = ptr + len;
411 	else
412 		*end = NULL;
413 	return;
414 }
415 
416 static int
417 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
418 {
419 	vm_page_t m;
420 	struct sbuf sbuf;
421 	int error, first;
422 
423 	first = 1;
424 	error = sysctl_wire_old_buffer(req, 0);
425 	if (error != 0)
426 		return (error);
427 	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
428 	TAILQ_FOREACH(m, &blacklist_head, listq) {
429 		sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
430 		    (uintmax_t)m->phys_addr);
431 		first = 0;
432 	}
433 	error = sbuf_finish(&sbuf);
434 	sbuf_delete(&sbuf);
435 	return (error);
436 }
437 
438 /*
439  * Initialize a dummy page for use in scans of the specified paging queue.
440  * In principle, this function only needs to set the flag PG_MARKER.
441  * Nonetheless, it write busies the page as a safety precaution.
442  */
443 static void
444 vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
445 {
446 
447 	bzero(marker, sizeof(*marker));
448 	marker->flags = PG_MARKER;
449 	marker->a.flags = aflags;
450 	marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
451 	marker->a.queue = queue;
452 }
453 
454 static void
455 vm_page_domain_init(int domain)
456 {
457 	struct vm_domain *vmd;
458 	struct vm_pagequeue *pq;
459 	int i;
460 
461 	vmd = VM_DOMAIN(domain);
462 	bzero(vmd, sizeof(*vmd));
463 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
464 	    "vm inactive pagequeue";
465 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
466 	    "vm active pagequeue";
467 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
468 	    "vm laundry pagequeue";
469 	*__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) =
470 	    "vm unswappable pagequeue";
471 	vmd->vmd_domain = domain;
472 	vmd->vmd_page_count = 0;
473 	vmd->vmd_free_count = 0;
474 	vmd->vmd_segs = 0;
475 	vmd->vmd_oom = FALSE;
476 	for (i = 0; i < PQ_COUNT; i++) {
477 		pq = &vmd->vmd_pagequeues[i];
478 		TAILQ_INIT(&pq->pq_pl);
479 		mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
480 		    MTX_DEF | MTX_DUPOK);
481 		pq->pq_pdpages = 0;
482 		vm_page_init_marker(&vmd->vmd_markers[i], i, 0);
483 	}
484 	mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF);
485 	mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF);
486 	snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain);
487 
488 	/*
489 	 * inacthead is used to provide FIFO ordering for LRU-bypassing
490 	 * insertions.
491 	 */
492 	vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED);
493 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
494 	    &vmd->vmd_inacthead, plinks.q);
495 
496 	/*
497 	 * The clock pages are used to implement active queue scanning without
498 	 * requeues.  Scans start at clock[0], which is advanced after the scan
499 	 * ends.  When the two clock hands meet, they are reset and scanning
500 	 * resumes from the head of the queue.
501 	 */
502 	vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED);
503 	vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED);
504 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
505 	    &vmd->vmd_clock[0], plinks.q);
506 	TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl,
507 	    &vmd->vmd_clock[1], plinks.q);
508 }
509 
510 /*
511  * Initialize a physical page in preparation for adding it to the free
512  * lists.
513  */
514 static void
515 vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
516 {
517 
518 	m->object = NULL;
519 	m->ref_count = 0;
520 	m->busy_lock = VPB_UNBUSIED;
521 	m->flags = m->a.flags = 0;
522 	m->phys_addr = pa;
523 	m->a.queue = PQ_NONE;
524 	m->psind = 0;
525 	m->segind = segind;
526 	m->order = VM_NFREEORDER;
527 	m->pool = VM_FREEPOOL_DEFAULT;
528 	m->valid = m->dirty = 0;
529 	pmap_page_init(m);
530 }
531 
532 #ifndef PMAP_HAS_PAGE_ARRAY
533 static vm_paddr_t
534 vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
535 {
536 	vm_paddr_t new_end;
537 
538 	/*
539 	 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
540 	 * However, because this page is allocated from KVM, out-of-bounds
541 	 * accesses using the direct map will not be trapped.
542 	 */
543 	*vaddr += PAGE_SIZE;
544 
545 	/*
546 	 * Allocate physical memory for the page structures, and map it.
547 	 */
548 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
549 	vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
550 	    VM_PROT_READ | VM_PROT_WRITE);
551 	vm_page_array_size = page_range;
552 
553 	return (new_end);
554 }
555 #endif
556 
557 /*
558  *	vm_page_startup:
559  *
560  *	Initializes the resident memory module.  Allocates physical memory for
561  *	bootstrapping UMA and some data structures that are used to manage
562  *	physical pages.  Initializes these structures, and populates the free
563  *	page queues.
564  */
565 vm_offset_t
566 vm_page_startup(vm_offset_t vaddr)
567 {
568 	struct vm_phys_seg *seg;
569 	vm_page_t m;
570 	char *list, *listend;
571 	vm_offset_t mapped;
572 	vm_paddr_t end, high_avail, low_avail, new_end, size;
573 	vm_paddr_t page_range __unused;
574 	vm_paddr_t last_pa, pa;
575 	u_long pagecount;
576 	int biggestone, i, segind;
577 #ifdef WITNESS
578 	int witness_size;
579 #endif
580 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
581 	long ii;
582 #endif
583 
584 	vaddr = round_page(vaddr);
585 
586 	vm_phys_early_startup();
587 	biggestone = vm_phys_avail_largest();
588 	end = phys_avail[biggestone+1];
589 
590 	/*
591 	 * Initialize the page and queue locks.
592 	 */
593 	mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF);
594 	for (i = 0; i < PA_LOCK_COUNT; i++)
595 		mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
596 	for (i = 0; i < vm_ndomains; i++)
597 		vm_page_domain_init(i);
598 
599 	/*
600 	 * Allocate memory for use when boot strapping the kernel memory
601 	 * allocator.  Tell UMA how many zones we are going to create
602 	 * before going fully functional.  UMA will add its zones.
603 	 *
604 	 * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP,
605 	 * KMAP ENTRY, MAP ENTRY, VMSPACE.
606 	 */
607 	boot_pages = uma_startup_count(8);
608 
609 #ifndef UMA_MD_SMALL_ALLOC
610 	/* vmem_startup() calls uma_prealloc(). */
611 	boot_pages += vmem_startup_count();
612 	/* vm_map_startup() calls uma_prealloc(). */
613 	boot_pages += howmany(MAX_KMAP,
614 	    slab_ipers(sizeof(struct vm_map), UMA_ALIGN_PTR));
615 
616 	/*
617 	 * Before we are fully boot strapped we need to account for the
618 	 * following allocations:
619 	 *
620 	 * "KMAP ENTRY" from kmem_init()
621 	 * "vmem btag" from vmem_startup()
622 	 * "vmem" from vmem_create()
623 	 * "KMAP" from vm_map_startup()
624 	 *
625 	 * Each needs at least one page per-domain.
626 	 */
627 	boot_pages += 4 * vm_ndomains;
628 #endif
629 	/*
630 	 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
631 	 * manually fetch the value.
632 	 */
633 	TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
634 	new_end = end - (boot_pages * UMA_SLAB_SIZE);
635 	new_end = trunc_page(new_end);
636 	mapped = pmap_map(&vaddr, new_end, end,
637 	    VM_PROT_READ | VM_PROT_WRITE);
638 	bzero((void *)mapped, end - new_end);
639 	uma_startup((void *)mapped, boot_pages);
640 
641 #ifdef WITNESS
642 	witness_size = round_page(witness_startup_count());
643 	new_end -= witness_size;
644 	mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
645 	    VM_PROT_READ | VM_PROT_WRITE);
646 	bzero((void *)mapped, witness_size);
647 	witness_startup((void *)mapped);
648 #endif
649 
650 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
651     defined(__i386__) || defined(__mips__) || defined(__riscv) || \
652     defined(__powerpc64__)
653 	/*
654 	 * Allocate a bitmap to indicate that a random physical page
655 	 * needs to be included in a minidump.
656 	 *
657 	 * The amd64 port needs this to indicate which direct map pages
658 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
659 	 *
660 	 * However, i386 still needs this workspace internally within the
661 	 * minidump code.  In theory, they are not needed on i386, but are
662 	 * included should the sf_buf code decide to use them.
663 	 */
664 	last_pa = 0;
665 	for (i = 0; dump_avail[i + 1] != 0; i += 2)
666 		if (dump_avail[i + 1] > last_pa)
667 			last_pa = dump_avail[i + 1];
668 	page_range = last_pa / PAGE_SIZE;
669 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
670 	new_end -= vm_page_dump_size;
671 	vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
672 	    new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
673 	bzero((void *)vm_page_dump, vm_page_dump_size);
674 #else
675 	(void)last_pa;
676 #endif
677 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
678     defined(__riscv) || defined(__powerpc64__)
679 	/*
680 	 * Include the UMA bootstrap pages, witness pages and vm_page_dump
681 	 * in a crash dump.  When pmap_map() uses the direct map, they are
682 	 * not automatically included.
683 	 */
684 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
685 		dump_add_page(pa);
686 #endif
687 	phys_avail[biggestone + 1] = new_end;
688 #ifdef __amd64__
689 	/*
690 	 * Request that the physical pages underlying the message buffer be
691 	 * included in a crash dump.  Since the message buffer is accessed
692 	 * through the direct map, they are not automatically included.
693 	 */
694 	pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
695 	last_pa = pa + round_page(msgbufsize);
696 	while (pa < last_pa) {
697 		dump_add_page(pa);
698 		pa += PAGE_SIZE;
699 	}
700 #endif
701 	/*
702 	 * Compute the number of pages of memory that will be available for
703 	 * use, taking into account the overhead of a page structure per page.
704 	 * In other words, solve
705 	 *	"available physical memory" - round_page(page_range *
706 	 *	    sizeof(struct vm_page)) = page_range * PAGE_SIZE
707 	 * for page_range.
708 	 */
709 	low_avail = phys_avail[0];
710 	high_avail = phys_avail[1];
711 	for (i = 0; i < vm_phys_nsegs; i++) {
712 		if (vm_phys_segs[i].start < low_avail)
713 			low_avail = vm_phys_segs[i].start;
714 		if (vm_phys_segs[i].end > high_avail)
715 			high_avail = vm_phys_segs[i].end;
716 	}
717 	/* Skip the first chunk.  It is already accounted for. */
718 	for (i = 2; phys_avail[i + 1] != 0; i += 2) {
719 		if (phys_avail[i] < low_avail)
720 			low_avail = phys_avail[i];
721 		if (phys_avail[i + 1] > high_avail)
722 			high_avail = phys_avail[i + 1];
723 	}
724 	first_page = low_avail / PAGE_SIZE;
725 #ifdef VM_PHYSSEG_SPARSE
726 	size = 0;
727 	for (i = 0; i < vm_phys_nsegs; i++)
728 		size += vm_phys_segs[i].end - vm_phys_segs[i].start;
729 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
730 		size += phys_avail[i + 1] - phys_avail[i];
731 #elif defined(VM_PHYSSEG_DENSE)
732 	size = high_avail - low_avail;
733 #else
734 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
735 #endif
736 
737 #ifdef PMAP_HAS_PAGE_ARRAY
738 	pmap_page_array_startup(size / PAGE_SIZE);
739 	biggestone = vm_phys_avail_largest();
740 	end = new_end = phys_avail[biggestone + 1];
741 #else
742 #ifdef VM_PHYSSEG_DENSE
743 	/*
744 	 * In the VM_PHYSSEG_DENSE case, the number of pages can account for
745 	 * the overhead of a page structure per page only if vm_page_array is
746 	 * allocated from the last physical memory chunk.  Otherwise, we must
747 	 * allocate page structures representing the physical memory
748 	 * underlying vm_page_array, even though they will not be used.
749 	 */
750 	if (new_end != high_avail)
751 		page_range = size / PAGE_SIZE;
752 	else
753 #endif
754 	{
755 		page_range = size / (PAGE_SIZE + sizeof(struct vm_page));
756 
757 		/*
758 		 * If the partial bytes remaining are large enough for
759 		 * a page (PAGE_SIZE) without a corresponding
760 		 * 'struct vm_page', then new_end will contain an
761 		 * extra page after subtracting the length of the VM
762 		 * page array.  Compensate by subtracting an extra
763 		 * page from new_end.
764 		 */
765 		if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) {
766 			if (new_end == high_avail)
767 				high_avail -= PAGE_SIZE;
768 			new_end -= PAGE_SIZE;
769 		}
770 	}
771 	end = new_end;
772 	new_end = vm_page_array_alloc(&vaddr, end, page_range);
773 #endif
774 
775 #if VM_NRESERVLEVEL > 0
776 	/*
777 	 * Allocate physical memory for the reservation management system's
778 	 * data structures, and map it.
779 	 */
780 	new_end = vm_reserv_startup(&vaddr, new_end);
781 #endif
782 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
783     defined(__riscv) || defined(__powerpc64__)
784 	/*
785 	 * Include vm_page_array and vm_reserv_array in a crash dump.
786 	 */
787 	for (pa = new_end; pa < end; pa += PAGE_SIZE)
788 		dump_add_page(pa);
789 #endif
790 	phys_avail[biggestone + 1] = new_end;
791 
792 	/*
793 	 * Add physical memory segments corresponding to the available
794 	 * physical pages.
795 	 */
796 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
797 		if (vm_phys_avail_size(i) != 0)
798 			vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
799 
800 	/*
801 	 * Initialize the physical memory allocator.
802 	 */
803 	vm_phys_init();
804 
805 	/*
806 	 * Initialize the page structures and add every available page to the
807 	 * physical memory allocator's free lists.
808 	 */
809 #if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
810 	for (ii = 0; ii < vm_page_array_size; ii++) {
811 		m = &vm_page_array[ii];
812 		vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0);
813 		m->flags = PG_FICTITIOUS;
814 	}
815 #endif
816 	vm_cnt.v_page_count = 0;
817 	for (segind = 0; segind < vm_phys_nsegs; segind++) {
818 		seg = &vm_phys_segs[segind];
819 		for (m = seg->first_page, pa = seg->start; pa < seg->end;
820 		    m++, pa += PAGE_SIZE)
821 			vm_page_init_page(m, pa, segind);
822 
823 		/*
824 		 * Add the segment to the free lists only if it is covered by
825 		 * one of the ranges in phys_avail.  Because we've added the
826 		 * ranges to the vm_phys_segs array, we can assume that each
827 		 * segment is either entirely contained in one of the ranges,
828 		 * or doesn't overlap any of them.
829 		 */
830 		for (i = 0; phys_avail[i + 1] != 0; i += 2) {
831 			struct vm_domain *vmd;
832 
833 			if (seg->start < phys_avail[i] ||
834 			    seg->end > phys_avail[i + 1])
835 				continue;
836 
837 			m = seg->first_page;
838 			pagecount = (u_long)atop(seg->end - seg->start);
839 
840 			vmd = VM_DOMAIN(seg->domain);
841 			vm_domain_free_lock(vmd);
842 			vm_phys_enqueue_contig(m, pagecount);
843 			vm_domain_free_unlock(vmd);
844 			vm_domain_freecnt_inc(vmd, pagecount);
845 			vm_cnt.v_page_count += (u_int)pagecount;
846 
847 			vmd = VM_DOMAIN(seg->domain);
848 			vmd->vmd_page_count += (u_int)pagecount;
849 			vmd->vmd_segs |= 1UL << m->segind;
850 			break;
851 		}
852 	}
853 
854 	/*
855 	 * Remove blacklisted pages from the physical memory allocator.
856 	 */
857 	TAILQ_INIT(&blacklist_head);
858 	vm_page_blacklist_load(&list, &listend);
859 	vm_page_blacklist_check(list, listend);
860 
861 	list = kern_getenv("vm.blacklist");
862 	vm_page_blacklist_check(list, NULL);
863 
864 	freeenv(list);
865 #if VM_NRESERVLEVEL > 0
866 	/*
867 	 * Initialize the reservation management system.
868 	 */
869 	vm_reserv_init();
870 #endif
871 
872 	return (vaddr);
873 }
874 
875 void
876 vm_page_reference(vm_page_t m)
877 {
878 
879 	vm_page_aflag_set(m, PGA_REFERENCED);
880 }
881 
882 static bool
883 vm_page_acquire_flags(vm_page_t m, int allocflags)
884 {
885 	bool locked;
886 
887 	if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
888 		locked = vm_page_trysbusy(m);
889 	else
890 		locked = vm_page_tryxbusy(m);
891 	if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
892 		vm_page_wire(m);
893 	return (locked);
894 }
895 
896 /*
897  *	vm_page_busy_sleep_flags
898  *
899  *	Sleep for busy according to VM_ALLOC_ parameters.
900  */
901 static bool
902 vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wmesg,
903     int allocflags)
904 {
905 
906 	if ((allocflags & VM_ALLOC_NOWAIT) != 0)
907 		return (false);
908 	/*
909 	 * Reference the page before unlocking and
910 	 * sleeping so that the page daemon is less
911 	 * likely to reclaim it.
912 	 */
913 	if ((allocflags & VM_ALLOC_NOCREAT) == 0)
914 		vm_page_aflag_set(m, PGA_REFERENCED);
915 	if (_vm_page_busy_sleep(object, m, wmesg, (allocflags &
916 	    VM_ALLOC_IGN_SBUSY) != 0, true))
917 		VM_OBJECT_WLOCK(object);
918 	if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
919 		return (false);
920 	return (true);
921 }
922 
923 /*
924  *	vm_page_busy_acquire:
925  *
926  *	Acquire the busy lock as described by VM_ALLOC_* flags.  Will loop
927  *	and drop the object lock if necessary.
928  */
929 bool
930 vm_page_busy_acquire(vm_page_t m, int allocflags)
931 {
932 	vm_object_t obj;
933 	bool locked;
934 
935 	/*
936 	 * The page-specific object must be cached because page
937 	 * identity can change during the sleep, causing the
938 	 * re-lock of a different object.
939 	 * It is assumed that a reference to the object is already
940 	 * held by the callers.
941 	 */
942 	obj = m->object;
943 	for (;;) {
944 		if (vm_page_acquire_flags(m, allocflags))
945 			return (true);
946 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
947 			return (false);
948 		if (obj != NULL)
949 			locked = VM_OBJECT_WOWNED(obj);
950 		else
951 			locked = false;
952 		MPASS(locked || vm_page_wired(m));
953 		if (_vm_page_busy_sleep(obj, m, "vmpba",
954 		    (allocflags & VM_ALLOC_SBUSY) != 0, locked))
955 			VM_OBJECT_WLOCK(obj);
956 		if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
957 			return (false);
958 		KASSERT(m->object == obj || m->object == NULL,
959 		    ("vm_page_busy_acquire: page %p does not belong to %p",
960 		    m, obj));
961 	}
962 }
963 
964 /*
965  *	vm_page_busy_downgrade:
966  *
967  *	Downgrade an exclusive busy page into a single shared busy page.
968  */
969 void
970 vm_page_busy_downgrade(vm_page_t m)
971 {
972 	u_int x;
973 
974 	vm_page_assert_xbusied(m);
975 
976 	x = m->busy_lock;
977 	for (;;) {
978 		if (atomic_fcmpset_rel_int(&m->busy_lock,
979 		    &x, VPB_SHARERS_WORD(1)))
980 			break;
981 	}
982 	if ((x & VPB_BIT_WAITERS) != 0)
983 		wakeup(m);
984 }
985 
986 /*
987  *
988  *	vm_page_busy_tryupgrade:
989  *
990  *	Attempt to upgrade a single shared busy into an exclusive busy.
991  */
992 int
993 vm_page_busy_tryupgrade(vm_page_t m)
994 {
995 	u_int ce, x;
996 
997 	vm_page_assert_sbusied(m);
998 
999 	x = m->busy_lock;
1000 	ce = VPB_CURTHREAD_EXCLUSIVE;
1001 	for (;;) {
1002 		if (VPB_SHARERS(x) > 1)
1003 			return (0);
1004 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1005 		    ("vm_page_busy_tryupgrade: invalid lock state"));
1006 		if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
1007 		    ce | (x & VPB_BIT_WAITERS)))
1008 			continue;
1009 		return (1);
1010 	}
1011 }
1012 
1013 /*
1014  *	vm_page_sbusied:
1015  *
1016  *	Return a positive value if the page is shared busied, 0 otherwise.
1017  */
1018 int
1019 vm_page_sbusied(vm_page_t m)
1020 {
1021 	u_int x;
1022 
1023 	x = m->busy_lock;
1024 	return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
1025 }
1026 
1027 /*
1028  *	vm_page_sunbusy:
1029  *
1030  *	Shared unbusy a page.
1031  */
1032 void
1033 vm_page_sunbusy(vm_page_t m)
1034 {
1035 	u_int x;
1036 
1037 	vm_page_assert_sbusied(m);
1038 
1039 	x = m->busy_lock;
1040 	for (;;) {
1041 		if (VPB_SHARERS(x) > 1) {
1042 			if (atomic_fcmpset_int(&m->busy_lock, &x,
1043 			    x - VPB_ONE_SHARER))
1044 				break;
1045 			continue;
1046 		}
1047 		KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
1048 		    ("vm_page_sunbusy: invalid lock state"));
1049 		if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
1050 			continue;
1051 		if ((x & VPB_BIT_WAITERS) == 0)
1052 			break;
1053 		wakeup(m);
1054 		break;
1055 	}
1056 }
1057 
1058 /*
1059  *	vm_page_busy_sleep:
1060  *
1061  *	Sleep if the page is busy, using the page pointer as wchan.
1062  *	This is used to implement the hard-path of busying mechanism.
1063  *
1064  *	If nonshared is true, sleep only if the page is xbusy.
1065  *
1066  *	The object lock must be held on entry and will be released on exit.
1067  */
1068 void
1069 vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared)
1070 {
1071 	vm_object_t obj;
1072 
1073 	obj = m->object;
1074 	VM_OBJECT_ASSERT_LOCKED(obj);
1075 	vm_page_lock_assert(m, MA_NOTOWNED);
1076 
1077 	if (!_vm_page_busy_sleep(obj, m, wmesg, nonshared, true))
1078 		VM_OBJECT_DROP(obj);
1079 }
1080 
1081 /*
1082  *	_vm_page_busy_sleep:
1083  *
1084  *	Internal busy sleep function.
1085  */
1086 static bool
1087 _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg,
1088     bool nonshared, bool locked)
1089 {
1090 	u_int x;
1091 
1092 	/*
1093 	 * If the object is busy we must wait for that to drain to zero
1094 	 * before trying the page again.
1095 	 */
1096 	if (obj != NULL && vm_object_busied(obj)) {
1097 		if (locked)
1098 			VM_OBJECT_DROP(obj);
1099 		vm_object_busy_wait(obj, wmesg);
1100 		return (locked);
1101 	}
1102 	sleepq_lock(m);
1103 	x = m->busy_lock;
1104 	if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) ||
1105 	    ((x & VPB_BIT_WAITERS) == 0 &&
1106 	    !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) {
1107 		sleepq_release(m);
1108 		return (false);
1109 	}
1110 	if (locked)
1111 		VM_OBJECT_DROP(obj);
1112 	DROP_GIANT();
1113 	sleepq_add(m, NULL, wmesg, 0, 0);
1114 	sleepq_wait(m, PVM);
1115 	PICKUP_GIANT();
1116 	return (locked);
1117 }
1118 
1119 /*
1120  *	vm_page_trysbusy:
1121  *
1122  *	Try to shared busy a page.
1123  *	If the operation succeeds 1 is returned otherwise 0.
1124  *	The operation never sleeps.
1125  */
1126 int
1127 vm_page_trysbusy(vm_page_t m)
1128 {
1129 	vm_object_t obj;
1130 	u_int x;
1131 
1132 	obj = m->object;
1133 	x = m->busy_lock;
1134 	for (;;) {
1135 		if ((x & VPB_BIT_SHARED) == 0)
1136 			return (0);
1137 		/*
1138 		 * Reduce the window for transient busies that will trigger
1139 		 * false negatives in vm_page_ps_test().
1140 		 */
1141 		if (obj != NULL && vm_object_busied(obj))
1142 			return (0);
1143 		if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
1144 		    x + VPB_ONE_SHARER))
1145 			break;
1146 	}
1147 
1148 	/* Refetch the object now that we're guaranteed that it is stable. */
1149 	obj = m->object;
1150 	if (obj != NULL && vm_object_busied(obj)) {
1151 		vm_page_sunbusy(m);
1152 		return (0);
1153 	}
1154 	return (1);
1155 }
1156 
1157 /*
1158  *	vm_page_tryxbusy:
1159  *
1160  *	Try to exclusive busy a page.
1161  *	If the operation succeeds 1 is returned otherwise 0.
1162  *	The operation never sleeps.
1163  */
1164 int
1165 vm_page_tryxbusy(vm_page_t m)
1166 {
1167 	vm_object_t obj;
1168 
1169         if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED,
1170             VPB_CURTHREAD_EXCLUSIVE) == 0)
1171 		return (0);
1172 
1173 	obj = m->object;
1174 	if (obj != NULL && vm_object_busied(obj)) {
1175 		vm_page_xunbusy(m);
1176 		return (0);
1177 	}
1178 	return (1);
1179 }
1180 
1181 static void
1182 vm_page_xunbusy_hard_tail(vm_page_t m)
1183 {
1184 	atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1185 	/* Wake the waiter. */
1186 	wakeup(m);
1187 }
1188 
1189 /*
1190  *	vm_page_xunbusy_hard:
1191  *
1192  *	Called when unbusy has failed because there is a waiter.
1193  */
1194 void
1195 vm_page_xunbusy_hard(vm_page_t m)
1196 {
1197 	vm_page_assert_xbusied(m);
1198 	vm_page_xunbusy_hard_tail(m);
1199 }
1200 
1201 void
1202 vm_page_xunbusy_hard_unchecked(vm_page_t m)
1203 {
1204 	vm_page_assert_xbusied_unchecked(m);
1205 	vm_page_xunbusy_hard_tail(m);
1206 }
1207 
1208 /*
1209  * Avoid releasing and reacquiring the same page lock.
1210  */
1211 void
1212 vm_page_change_lock(vm_page_t m, struct mtx **mtx)
1213 {
1214 	struct mtx *mtx1;
1215 
1216 	mtx1 = vm_page_lockptr(m);
1217 	if (*mtx == mtx1)
1218 		return;
1219 	if (*mtx != NULL)
1220 		mtx_unlock(*mtx);
1221 	*mtx = mtx1;
1222 	mtx_lock(mtx1);
1223 }
1224 
1225 /*
1226  *	vm_page_unhold_pages:
1227  *
1228  *	Unhold each of the pages that is referenced by the given array.
1229  */
1230 void
1231 vm_page_unhold_pages(vm_page_t *ma, int count)
1232 {
1233 
1234 	for (; count != 0; count--) {
1235 		vm_page_unwire(*ma, PQ_ACTIVE);
1236 		ma++;
1237 	}
1238 }
1239 
1240 vm_page_t
1241 PHYS_TO_VM_PAGE(vm_paddr_t pa)
1242 {
1243 	vm_page_t m;
1244 
1245 #ifdef VM_PHYSSEG_SPARSE
1246 	m = vm_phys_paddr_to_vm_page(pa);
1247 	if (m == NULL)
1248 		m = vm_phys_fictitious_to_vm_page(pa);
1249 	return (m);
1250 #elif defined(VM_PHYSSEG_DENSE)
1251 	long pi;
1252 
1253 	pi = atop(pa);
1254 	if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1255 		m = &vm_page_array[pi - first_page];
1256 		return (m);
1257 	}
1258 	return (vm_phys_fictitious_to_vm_page(pa));
1259 #else
1260 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1261 #endif
1262 }
1263 
1264 /*
1265  *	vm_page_getfake:
1266  *
1267  *	Create a fictitious page with the specified physical address and
1268  *	memory attribute.  The memory attribute is the only the machine-
1269  *	dependent aspect of a fictitious page that must be initialized.
1270  */
1271 vm_page_t
1272 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1273 {
1274 	vm_page_t m;
1275 
1276 	m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1277 	vm_page_initfake(m, paddr, memattr);
1278 	return (m);
1279 }
1280 
1281 void
1282 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1283 {
1284 
1285 	if ((m->flags & PG_FICTITIOUS) != 0) {
1286 		/*
1287 		 * The page's memattr might have changed since the
1288 		 * previous initialization.  Update the pmap to the
1289 		 * new memattr.
1290 		 */
1291 		goto memattr;
1292 	}
1293 	m->phys_addr = paddr;
1294 	m->a.queue = PQ_NONE;
1295 	/* Fictitious pages don't use "segind". */
1296 	m->flags = PG_FICTITIOUS;
1297 	/* Fictitious pages don't use "order" or "pool". */
1298 	m->oflags = VPO_UNMANAGED;
1299 	m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
1300 	/* Fictitious pages are unevictable. */
1301 	m->ref_count = 1;
1302 	pmap_page_init(m);
1303 memattr:
1304 	pmap_page_set_memattr(m, memattr);
1305 }
1306 
1307 /*
1308  *	vm_page_putfake:
1309  *
1310  *	Release a fictitious page.
1311  */
1312 void
1313 vm_page_putfake(vm_page_t m)
1314 {
1315 
1316 	KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1317 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1318 	    ("vm_page_putfake: bad page %p", m));
1319 	vm_page_xunbusy(m);
1320 	uma_zfree(fakepg_zone, m);
1321 }
1322 
1323 /*
1324  *	vm_page_updatefake:
1325  *
1326  *	Update the given fictitious page to the specified physical address and
1327  *	memory attribute.
1328  */
1329 void
1330 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1331 {
1332 
1333 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
1334 	    ("vm_page_updatefake: bad page %p", m));
1335 	m->phys_addr = paddr;
1336 	pmap_page_set_memattr(m, memattr);
1337 }
1338 
1339 /*
1340  *	vm_page_free:
1341  *
1342  *	Free a page.
1343  */
1344 void
1345 vm_page_free(vm_page_t m)
1346 {
1347 
1348 	m->flags &= ~PG_ZERO;
1349 	vm_page_free_toq(m);
1350 }
1351 
1352 /*
1353  *	vm_page_free_zero:
1354  *
1355  *	Free a page to the zerod-pages queue
1356  */
1357 void
1358 vm_page_free_zero(vm_page_t m)
1359 {
1360 
1361 	m->flags |= PG_ZERO;
1362 	vm_page_free_toq(m);
1363 }
1364 
1365 /*
1366  * Unbusy and handle the page queueing for a page from a getpages request that
1367  * was optionally read ahead or behind.
1368  */
1369 void
1370 vm_page_readahead_finish(vm_page_t m)
1371 {
1372 
1373 	/* We shouldn't put invalid pages on queues. */
1374 	KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m));
1375 
1376 	/*
1377 	 * Since the page is not the actually needed one, whether it should
1378 	 * be activated or deactivated is not obvious.  Empirical results
1379 	 * have shown that deactivating the page is usually the best choice,
1380 	 * unless the page is wanted by another thread.
1381 	 */
1382 	if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1383 		vm_page_activate(m);
1384 	else
1385 		vm_page_deactivate(m);
1386 	vm_page_xunbusy_unchecked(m);
1387 }
1388 
1389 /*
1390  *	vm_page_sleep_if_busy:
1391  *
1392  *	Sleep and release the object lock if the page is busied.
1393  *	Returns TRUE if the thread slept.
1394  *
1395  *	The given page must be unlocked and object containing it must
1396  *	be locked.
1397  */
1398 int
1399 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1400 {
1401 	vm_object_t obj;
1402 
1403 	vm_page_lock_assert(m, MA_NOTOWNED);
1404 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1405 
1406 	/*
1407 	 * The page-specific object must be cached because page
1408 	 * identity can change during the sleep, causing the
1409 	 * re-lock of a different object.
1410 	 * It is assumed that a reference to the object is already
1411 	 * held by the callers.
1412 	 */
1413 	obj = m->object;
1414 	if (vm_page_busied(m) || (obj != NULL && obj->busy)) {
1415 		vm_page_busy_sleep(m, msg, false);
1416 		VM_OBJECT_WLOCK(obj);
1417 		return (TRUE);
1418 	}
1419 	return (FALSE);
1420 }
1421 
1422 /*
1423  *	vm_page_sleep_if_xbusy:
1424  *
1425  *	Sleep and release the object lock if the page is xbusied.
1426  *	Returns TRUE if the thread slept.
1427  *
1428  *	The given page must be unlocked and object containing it must
1429  *	be locked.
1430  */
1431 int
1432 vm_page_sleep_if_xbusy(vm_page_t m, const char *msg)
1433 {
1434 	vm_object_t obj;
1435 
1436 	vm_page_lock_assert(m, MA_NOTOWNED);
1437 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1438 
1439 	/*
1440 	 * The page-specific object must be cached because page
1441 	 * identity can change during the sleep, causing the
1442 	 * re-lock of a different object.
1443 	 * It is assumed that a reference to the object is already
1444 	 * held by the callers.
1445 	 */
1446 	obj = m->object;
1447 	if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) {
1448 		vm_page_busy_sleep(m, msg, true);
1449 		VM_OBJECT_WLOCK(obj);
1450 		return (TRUE);
1451 	}
1452 	return (FALSE);
1453 }
1454 
1455 /*
1456  *	vm_page_dirty_KBI:		[ internal use only ]
1457  *
1458  *	Set all bits in the page's dirty field.
1459  *
1460  *	The object containing the specified page must be locked if the
1461  *	call is made from the machine-independent layer.
1462  *
1463  *	See vm_page_clear_dirty_mask().
1464  *
1465  *	This function should only be called by vm_page_dirty().
1466  */
1467 void
1468 vm_page_dirty_KBI(vm_page_t m)
1469 {
1470 
1471 	/* Refer to this operation by its public name. */
1472 	KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!"));
1473 	m->dirty = VM_PAGE_BITS_ALL;
1474 }
1475 
1476 /*
1477  *	vm_page_insert:		[ internal use only ]
1478  *
1479  *	Inserts the given mem entry into the object and object list.
1480  *
1481  *	The object must be locked.
1482  */
1483 int
1484 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1485 {
1486 	vm_page_t mpred;
1487 
1488 	VM_OBJECT_ASSERT_WLOCKED(object);
1489 	mpred = vm_radix_lookup_le(&object->rtree, pindex);
1490 	return (vm_page_insert_after(m, object, pindex, mpred));
1491 }
1492 
1493 /*
1494  *	vm_page_insert_after:
1495  *
1496  *	Inserts the page "m" into the specified object at offset "pindex".
1497  *
1498  *	The page "mpred" must immediately precede the offset "pindex" within
1499  *	the specified object.
1500  *
1501  *	The object must be locked.
1502  */
1503 static int
1504 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1505     vm_page_t mpred)
1506 {
1507 	vm_page_t msucc;
1508 
1509 	VM_OBJECT_ASSERT_WLOCKED(object);
1510 	KASSERT(m->object == NULL,
1511 	    ("vm_page_insert_after: page already inserted"));
1512 	if (mpred != NULL) {
1513 		KASSERT(mpred->object == object,
1514 		    ("vm_page_insert_after: object doesn't contain mpred"));
1515 		KASSERT(mpred->pindex < pindex,
1516 		    ("vm_page_insert_after: mpred doesn't precede pindex"));
1517 		msucc = TAILQ_NEXT(mpred, listq);
1518 	} else
1519 		msucc = TAILQ_FIRST(&object->memq);
1520 	if (msucc != NULL)
1521 		KASSERT(msucc->pindex > pindex,
1522 		    ("vm_page_insert_after: msucc doesn't succeed pindex"));
1523 
1524 	/*
1525 	 * Record the object/offset pair in this page.
1526 	 */
1527 	m->object = object;
1528 	m->pindex = pindex;
1529 	m->ref_count |= VPRC_OBJREF;
1530 
1531 	/*
1532 	 * Now link into the object's ordered list of backed pages.
1533 	 */
1534 	if (vm_radix_insert(&object->rtree, m)) {
1535 		m->object = NULL;
1536 		m->pindex = 0;
1537 		m->ref_count &= ~VPRC_OBJREF;
1538 		return (1);
1539 	}
1540 	vm_page_insert_radixdone(m, object, mpred);
1541 	return (0);
1542 }
1543 
1544 /*
1545  *	vm_page_insert_radixdone:
1546  *
1547  *	Complete page "m" insertion into the specified object after the
1548  *	radix trie hooking.
1549  *
1550  *	The page "mpred" must precede the offset "m->pindex" within the
1551  *	specified object.
1552  *
1553  *	The object must be locked.
1554  */
1555 static void
1556 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1557 {
1558 
1559 	VM_OBJECT_ASSERT_WLOCKED(object);
1560 	KASSERT(object != NULL && m->object == object,
1561 	    ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1562 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1563 	    ("vm_page_insert_radixdone: page %p is missing object ref", m));
1564 	if (mpred != NULL) {
1565 		KASSERT(mpred->object == object,
1566 		    ("vm_page_insert_radixdone: object doesn't contain mpred"));
1567 		KASSERT(mpred->pindex < m->pindex,
1568 		    ("vm_page_insert_radixdone: mpred doesn't precede pindex"));
1569 	}
1570 
1571 	if (mpred != NULL)
1572 		TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1573 	else
1574 		TAILQ_INSERT_HEAD(&object->memq, m, listq);
1575 
1576 	/*
1577 	 * Show that the object has one more resident page.
1578 	 */
1579 	object->resident_page_count++;
1580 
1581 	/*
1582 	 * Hold the vnode until the last page is released.
1583 	 */
1584 	if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1585 		vhold(object->handle);
1586 
1587 	/*
1588 	 * Since we are inserting a new and possibly dirty page,
1589 	 * update the object's generation count.
1590 	 */
1591 	if (pmap_page_is_write_mapped(m))
1592 		vm_object_set_writeable_dirty(object);
1593 }
1594 
1595 /*
1596  * Do the work to remove a page from its object.  The caller is responsible for
1597  * updating the page's fields to reflect this removal.
1598  */
1599 static void
1600 vm_page_object_remove(vm_page_t m)
1601 {
1602 	vm_object_t object;
1603 	vm_page_t mrem;
1604 
1605 	vm_page_assert_xbusied(m);
1606 	object = m->object;
1607 	VM_OBJECT_ASSERT_WLOCKED(object);
1608 	KASSERT((m->ref_count & VPRC_OBJREF) != 0,
1609 	    ("page %p is missing its object ref", m));
1610 
1611 	/* Deferred free of swap space. */
1612 	if ((m->a.flags & PGA_SWAP_FREE) != 0)
1613 		vm_pager_page_unswapped(m);
1614 
1615 	mrem = vm_radix_remove(&object->rtree, m->pindex);
1616 	KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m));
1617 
1618 	/*
1619 	 * Now remove from the object's list of backed pages.
1620 	 */
1621 	TAILQ_REMOVE(&object->memq, m, listq);
1622 
1623 	/*
1624 	 * And show that the object has one fewer resident page.
1625 	 */
1626 	object->resident_page_count--;
1627 
1628 	/*
1629 	 * The vnode may now be recycled.
1630 	 */
1631 	if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1632 		vdrop(object->handle);
1633 }
1634 
1635 /*
1636  *	vm_page_remove:
1637  *
1638  *	Removes the specified page from its containing object, but does not
1639  *	invalidate any backing storage.  Returns true if the object's reference
1640  *	was the last reference to the page, and false otherwise.
1641  *
1642  *	The object must be locked and the page must be exclusively busied.
1643  *	The exclusive busy will be released on return.  If this is not the
1644  *	final ref and the caller does not hold a wire reference it may not
1645  *	continue to access the page.
1646  */
1647 bool
1648 vm_page_remove(vm_page_t m)
1649 {
1650 	bool dropped;
1651 
1652 	dropped = vm_page_remove_xbusy(m);
1653 	vm_page_xunbusy(m);
1654 
1655 	return (dropped);
1656 }
1657 
1658 /*
1659  *	vm_page_remove_xbusy
1660  *
1661  *	Removes the page but leaves the xbusy held.  Returns true if this
1662  *	removed the final ref and false otherwise.
1663  */
1664 bool
1665 vm_page_remove_xbusy(vm_page_t m)
1666 {
1667 
1668 	vm_page_object_remove(m);
1669 	m->object = NULL;
1670 	return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
1671 }
1672 
1673 /*
1674  *	vm_page_lookup:
1675  *
1676  *	Returns the page associated with the object/offset
1677  *	pair specified; if none is found, NULL is returned.
1678  *
1679  *	The object must be locked.
1680  */
1681 vm_page_t
1682 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1683 {
1684 
1685 	VM_OBJECT_ASSERT_LOCKED(object);
1686 	return (vm_radix_lookup(&object->rtree, pindex));
1687 }
1688 
1689 /*
1690  *	vm_page_find_least:
1691  *
1692  *	Returns the page associated with the object with least pindex
1693  *	greater than or equal to the parameter pindex, or NULL.
1694  *
1695  *	The object must be locked.
1696  */
1697 vm_page_t
1698 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1699 {
1700 	vm_page_t m;
1701 
1702 	VM_OBJECT_ASSERT_LOCKED(object);
1703 	if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1704 		m = vm_radix_lookup_ge(&object->rtree, pindex);
1705 	return (m);
1706 }
1707 
1708 /*
1709  * Returns the given page's successor (by pindex) within the object if it is
1710  * resident; if none is found, NULL is returned.
1711  *
1712  * The object must be locked.
1713  */
1714 vm_page_t
1715 vm_page_next(vm_page_t m)
1716 {
1717 	vm_page_t next;
1718 
1719 	VM_OBJECT_ASSERT_LOCKED(m->object);
1720 	if ((next = TAILQ_NEXT(m, listq)) != NULL) {
1721 		MPASS(next->object == m->object);
1722 		if (next->pindex != m->pindex + 1)
1723 			next = NULL;
1724 	}
1725 	return (next);
1726 }
1727 
1728 /*
1729  * Returns the given page's predecessor (by pindex) within the object if it is
1730  * resident; if none is found, NULL is returned.
1731  *
1732  * The object must be locked.
1733  */
1734 vm_page_t
1735 vm_page_prev(vm_page_t m)
1736 {
1737 	vm_page_t prev;
1738 
1739 	VM_OBJECT_ASSERT_LOCKED(m->object);
1740 	if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) {
1741 		MPASS(prev->object == m->object);
1742 		if (prev->pindex != m->pindex - 1)
1743 			prev = NULL;
1744 	}
1745 	return (prev);
1746 }
1747 
1748 /*
1749  * Uses the page mnew as a replacement for an existing page at index
1750  * pindex which must be already present in the object.
1751  *
1752  * Both pages must be exclusively busied on enter.  The old page is
1753  * unbusied on exit.
1754  *
1755  * A return value of true means mold is now free.  If this is not the
1756  * final ref and the caller does not hold a wire reference it may not
1757  * continue to access the page.
1758  */
1759 static bool
1760 vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1761     vm_page_t mold)
1762 {
1763 	vm_page_t mret;
1764 	bool dropped;
1765 
1766 	VM_OBJECT_ASSERT_WLOCKED(object);
1767 	vm_page_assert_xbusied(mold);
1768 	KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
1769 	    ("vm_page_replace: page %p already in object", mnew));
1770 
1771 	/*
1772 	 * This function mostly follows vm_page_insert() and
1773 	 * vm_page_remove() without the radix, object count and vnode
1774 	 * dance.  Double check such functions for more comments.
1775 	 */
1776 
1777 	mnew->object = object;
1778 	mnew->pindex = pindex;
1779 	atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
1780 	mret = vm_radix_replace(&object->rtree, mnew);
1781 	KASSERT(mret == mold,
1782 	    ("invalid page replacement, mold=%p, mret=%p", mold, mret));
1783 	KASSERT((mold->oflags & VPO_UNMANAGED) ==
1784 	    (mnew->oflags & VPO_UNMANAGED),
1785 	    ("vm_page_replace: mismatched VPO_UNMANAGED"));
1786 
1787 	/* Keep the resident page list in sorted order. */
1788 	TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1789 	TAILQ_REMOVE(&object->memq, mold, listq);
1790 	mold->object = NULL;
1791 
1792 	/*
1793 	 * The object's resident_page_count does not change because we have
1794 	 * swapped one page for another, but the generation count should
1795 	 * change if the page is dirty.
1796 	 */
1797 	if (pmap_page_is_write_mapped(mnew))
1798 		vm_object_set_writeable_dirty(object);
1799 	dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
1800 	vm_page_xunbusy(mold);
1801 
1802 	return (dropped);
1803 }
1804 
1805 void
1806 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
1807     vm_page_t mold)
1808 {
1809 
1810 	vm_page_assert_xbusied(mnew);
1811 
1812 	if (vm_page_replace_hold(mnew, object, pindex, mold))
1813 		vm_page_free(mold);
1814 }
1815 
1816 /*
1817  *	vm_page_rename:
1818  *
1819  *	Move the given memory entry from its
1820  *	current object to the specified target object/offset.
1821  *
1822  *	Note: swap associated with the page must be invalidated by the move.  We
1823  *	      have to do this for several reasons:  (1) we aren't freeing the
1824  *	      page, (2) we are dirtying the page, (3) the VM system is probably
1825  *	      moving the page from object A to B, and will then later move
1826  *	      the backing store from A to B and we can't have a conflict.
1827  *
1828  *	Note: we *always* dirty the page.  It is necessary both for the
1829  *	      fact that we moved it, and because we may be invalidating
1830  *	      swap.
1831  *
1832  *	The objects must be locked.
1833  */
1834 int
1835 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1836 {
1837 	vm_page_t mpred;
1838 	vm_pindex_t opidx;
1839 
1840 	VM_OBJECT_ASSERT_WLOCKED(new_object);
1841 
1842 	KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
1843 	mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1844 	KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1845 	    ("vm_page_rename: pindex already renamed"));
1846 
1847 	/*
1848 	 * Create a custom version of vm_page_insert() which does not depend
1849 	 * by m_prev and can cheat on the implementation aspects of the
1850 	 * function.
1851 	 */
1852 	opidx = m->pindex;
1853 	m->pindex = new_pindex;
1854 	if (vm_radix_insert(&new_object->rtree, m)) {
1855 		m->pindex = opidx;
1856 		return (1);
1857 	}
1858 
1859 	/*
1860 	 * The operation cannot fail anymore.  The removal must happen before
1861 	 * the listq iterator is tainted.
1862 	 */
1863 	m->pindex = opidx;
1864 	vm_page_object_remove(m);
1865 
1866 	/* Return back to the new pindex to complete vm_page_insert(). */
1867 	m->pindex = new_pindex;
1868 	m->object = new_object;
1869 
1870 	vm_page_insert_radixdone(m, new_object, mpred);
1871 	vm_page_dirty(m);
1872 	return (0);
1873 }
1874 
1875 /*
1876  *	vm_page_alloc:
1877  *
1878  *	Allocate and return a page that is associated with the specified
1879  *	object and offset pair.  By default, this page is exclusive busied.
1880  *
1881  *	The caller must always specify an allocation class.
1882  *
1883  *	allocation classes:
1884  *	VM_ALLOC_NORMAL		normal process request
1885  *	VM_ALLOC_SYSTEM		system *really* needs a page
1886  *	VM_ALLOC_INTERRUPT	interrupt time request
1887  *
1888  *	optional allocation flags:
1889  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
1890  *				intends to allocate
1891  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
1892  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
1893  *	VM_ALLOC_NOOBJ		page is not associated with an object and
1894  *				should not be exclusive busy
1895  *	VM_ALLOC_SBUSY		shared busy the allocated page
1896  *	VM_ALLOC_WIRED		wire the allocated page
1897  *	VM_ALLOC_ZERO		prefer a zeroed page
1898  */
1899 vm_page_t
1900 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1901 {
1902 
1903 	return (vm_page_alloc_after(object, pindex, req, object != NULL ?
1904 	    vm_radix_lookup_le(&object->rtree, pindex) : NULL));
1905 }
1906 
1907 vm_page_t
1908 vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain,
1909     int req)
1910 {
1911 
1912 	return (vm_page_alloc_domain_after(object, pindex, domain, req,
1913 	    object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) :
1914 	    NULL));
1915 }
1916 
1917 /*
1918  * Allocate a page in the specified object with the given page index.  To
1919  * optimize insertion of the page into the object, the caller must also specifiy
1920  * the resident page in the object with largest index smaller than the given
1921  * page index, or NULL if no such page exists.
1922  */
1923 vm_page_t
1924 vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex,
1925     int req, vm_page_t mpred)
1926 {
1927 	struct vm_domainset_iter di;
1928 	vm_page_t m;
1929 	int domain;
1930 
1931 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
1932 	do {
1933 		m = vm_page_alloc_domain_after(object, pindex, domain, req,
1934 		    mpred);
1935 		if (m != NULL)
1936 			break;
1937 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
1938 
1939 	return (m);
1940 }
1941 
1942 /*
1943  * Returns true if the number of free pages exceeds the minimum
1944  * for the request class and false otherwise.
1945  */
1946 static int
1947 _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
1948 {
1949 	u_int limit, old, new;
1950 
1951 	if (req_class == VM_ALLOC_INTERRUPT)
1952 		limit = 0;
1953 	else if (req_class == VM_ALLOC_SYSTEM)
1954 		limit = vmd->vmd_interrupt_free_min;
1955 	else
1956 		limit = vmd->vmd_free_reserved;
1957 
1958 	/*
1959 	 * Attempt to reserve the pages.  Fail if we're below the limit.
1960 	 */
1961 	limit += npages;
1962 	old = vmd->vmd_free_count;
1963 	do {
1964 		if (old < limit)
1965 			return (0);
1966 		new = old - npages;
1967 	} while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0);
1968 
1969 	/* Wake the page daemon if we've crossed the threshold. */
1970 	if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old))
1971 		pagedaemon_wakeup(vmd->vmd_domain);
1972 
1973 	/* Only update bitsets on transitions. */
1974 	if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) ||
1975 	    (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe))
1976 		vm_domain_set(vmd);
1977 
1978 	return (1);
1979 }
1980 
1981 int
1982 vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
1983 {
1984 	int req_class;
1985 
1986 	/*
1987 	 * The page daemon is allowed to dig deeper into the free page list.
1988 	 */
1989 	req_class = req & VM_ALLOC_CLASS_MASK;
1990 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1991 		req_class = VM_ALLOC_SYSTEM;
1992 	return (_vm_domain_allocate(vmd, req_class, npages));
1993 }
1994 
1995 vm_page_t
1996 vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain,
1997     int req, vm_page_t mpred)
1998 {
1999 	struct vm_domain *vmd;
2000 	vm_page_t m;
2001 	int flags, pool;
2002 
2003 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2004 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2005 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2006 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2007 	    ("inconsistent object(%p)/req(%x)", object, req));
2008 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2009 	    ("Can't sleep and retry object insertion."));
2010 	KASSERT(mpred == NULL || mpred->pindex < pindex,
2011 	    ("mpred %p doesn't precede pindex 0x%jx", mpred,
2012 	    (uintmax_t)pindex));
2013 	if (object != NULL)
2014 		VM_OBJECT_ASSERT_WLOCKED(object);
2015 
2016 	flags = 0;
2017 	m = NULL;
2018 	pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT;
2019 again:
2020 #if VM_NRESERVLEVEL > 0
2021 	/*
2022 	 * Can we allocate the page from a reservation?
2023 	 */
2024 	if (vm_object_reserv(object) &&
2025 	    (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
2026 	    NULL) {
2027 		domain = vm_phys_domain(m);
2028 		vmd = VM_DOMAIN(domain);
2029 		goto found;
2030 	}
2031 #endif
2032 	vmd = VM_DOMAIN(domain);
2033 	if (vmd->vmd_pgcache[pool].zone != NULL) {
2034 		m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT);
2035 		if (m != NULL) {
2036 			flags |= PG_PCPU_CACHE;
2037 			goto found;
2038 		}
2039 	}
2040 	if (vm_domain_allocate(vmd, req, 1)) {
2041 		/*
2042 		 * If not, allocate it from the free page queues.
2043 		 */
2044 		vm_domain_free_lock(vmd);
2045 		m = vm_phys_alloc_pages(domain, pool, 0);
2046 		vm_domain_free_unlock(vmd);
2047 		if (m == NULL) {
2048 			vm_domain_freecnt_inc(vmd, 1);
2049 #if VM_NRESERVLEVEL > 0
2050 			if (vm_reserv_reclaim_inactive(domain))
2051 				goto again;
2052 #endif
2053 		}
2054 	}
2055 	if (m == NULL) {
2056 		/*
2057 		 * Not allocatable, give up.
2058 		 */
2059 		if (vm_domain_alloc_fail(vmd, object, req))
2060 			goto again;
2061 		return (NULL);
2062 	}
2063 
2064 	/*
2065 	 * At this point we had better have found a good page.
2066 	 */
2067 found:
2068 	vm_page_dequeue(m);
2069 	vm_page_alloc_check(m);
2070 
2071 	/*
2072 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2073 	 */
2074 	if ((req & VM_ALLOC_ZERO) != 0)
2075 		flags |= (m->flags & PG_ZERO);
2076 	if ((req & VM_ALLOC_NODUMP) != 0)
2077 		flags |= PG_NODUMP;
2078 	m->flags = flags;
2079 	m->a.flags = 0;
2080 	m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2081 	    VPO_UNMANAGED : 0;
2082 	m->busy_lock = VPB_UNBUSIED;
2083 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2084 		m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2085 	if ((req & VM_ALLOC_SBUSY) != 0)
2086 		m->busy_lock = VPB_SHARERS_WORD(1);
2087 	if (req & VM_ALLOC_WIRED) {
2088 		vm_wire_add(1);
2089 		m->ref_count = 1;
2090 	}
2091 	m->a.act_count = 0;
2092 
2093 	if (object != NULL) {
2094 		if (vm_page_insert_after(m, object, pindex, mpred)) {
2095 			if (req & VM_ALLOC_WIRED) {
2096 				vm_wire_sub(1);
2097 				m->ref_count = 0;
2098 			}
2099 			KASSERT(m->object == NULL, ("page %p has object", m));
2100 			m->oflags = VPO_UNMANAGED;
2101 			m->busy_lock = VPB_UNBUSIED;
2102 			/* Don't change PG_ZERO. */
2103 			vm_page_free_toq(m);
2104 			if (req & VM_ALLOC_WAITFAIL) {
2105 				VM_OBJECT_WUNLOCK(object);
2106 				vm_radix_wait();
2107 				VM_OBJECT_WLOCK(object);
2108 			}
2109 			return (NULL);
2110 		}
2111 
2112 		/* Ignore device objects; the pager sets "memattr" for them. */
2113 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2114 		    (object->flags & OBJ_FICTITIOUS) == 0)
2115 			pmap_page_set_memattr(m, object->memattr);
2116 	} else
2117 		m->pindex = pindex;
2118 
2119 	return (m);
2120 }
2121 
2122 /*
2123  *	vm_page_alloc_contig:
2124  *
2125  *	Allocate a contiguous set of physical pages of the given size "npages"
2126  *	from the free lists.  All of the physical pages must be at or above
2127  *	the given physical address "low" and below the given physical address
2128  *	"high".  The given value "alignment" determines the alignment of the
2129  *	first physical page in the set.  If the given value "boundary" is
2130  *	non-zero, then the set of physical pages cannot cross any physical
2131  *	address boundary that is a multiple of that value.  Both "alignment"
2132  *	and "boundary" must be a power of two.
2133  *
2134  *	If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
2135  *	then the memory attribute setting for the physical pages is configured
2136  *	to the object's memory attribute setting.  Otherwise, the memory
2137  *	attribute setting for the physical pages is configured to "memattr",
2138  *	overriding the object's memory attribute setting.  However, if the
2139  *	object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
2140  *	memory attribute setting for the physical pages cannot be configured
2141  *	to VM_MEMATTR_DEFAULT.
2142  *
2143  *	The specified object may not contain fictitious pages.
2144  *
2145  *	The caller must always specify an allocation class.
2146  *
2147  *	allocation classes:
2148  *	VM_ALLOC_NORMAL		normal process request
2149  *	VM_ALLOC_SYSTEM		system *really* needs a page
2150  *	VM_ALLOC_INTERRUPT	interrupt time request
2151  *
2152  *	optional allocation flags:
2153  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
2154  *	VM_ALLOC_NODUMP		do not include the page in a kernel core dump
2155  *	VM_ALLOC_NOOBJ		page is not associated with an object and
2156  *				should not be exclusive busy
2157  *	VM_ALLOC_SBUSY		shared busy the allocated page
2158  *	VM_ALLOC_WIRED		wire the allocated page
2159  *	VM_ALLOC_ZERO		prefer a zeroed page
2160  */
2161 vm_page_t
2162 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
2163     u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2164     vm_paddr_t boundary, vm_memattr_t memattr)
2165 {
2166 	struct vm_domainset_iter di;
2167 	vm_page_t m;
2168 	int domain;
2169 
2170 	vm_domainset_iter_page_init(&di, object, pindex, &domain, &req);
2171 	do {
2172 		m = vm_page_alloc_contig_domain(object, pindex, domain, req,
2173 		    npages, low, high, alignment, boundary, memattr);
2174 		if (m != NULL)
2175 			break;
2176 	} while (vm_domainset_iter_page(&di, object, &domain) == 0);
2177 
2178 	return (m);
2179 }
2180 
2181 vm_page_t
2182 vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain,
2183     int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
2184     vm_paddr_t boundary, vm_memattr_t memattr)
2185 {
2186 	struct vm_domain *vmd;
2187 	vm_page_t m, m_ret, mpred;
2188 	u_int busy_lock, flags, oflags;
2189 
2190 	mpred = NULL;	/* XXX: pacify gcc */
2191 	KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
2192 	    (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
2193 	    ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
2194 	    (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
2195 	    ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object,
2196 	    req));
2197 	KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0,
2198 	    ("Can't sleep and retry object insertion."));
2199 	if (object != NULL) {
2200 		VM_OBJECT_ASSERT_WLOCKED(object);
2201 		KASSERT((object->flags & OBJ_FICTITIOUS) == 0,
2202 		    ("vm_page_alloc_contig: object %p has fictitious pages",
2203 		    object));
2204 	}
2205 	KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
2206 
2207 	if (object != NULL) {
2208 		mpred = vm_radix_lookup_le(&object->rtree, pindex);
2209 		KASSERT(mpred == NULL || mpred->pindex != pindex,
2210 		    ("vm_page_alloc_contig: pindex already allocated"));
2211 	}
2212 
2213 	/*
2214 	 * Can we allocate the pages without the number of free pages falling
2215 	 * below the lower bound for the allocation class?
2216 	 */
2217 	m_ret = NULL;
2218 again:
2219 #if VM_NRESERVLEVEL > 0
2220 	/*
2221 	 * Can we allocate the pages from a reservation?
2222 	 */
2223 	if (vm_object_reserv(object) &&
2224 	    (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
2225 	    mpred, npages, low, high, alignment, boundary)) != NULL) {
2226 		domain = vm_phys_domain(m_ret);
2227 		vmd = VM_DOMAIN(domain);
2228 		goto found;
2229 	}
2230 #endif
2231 	vmd = VM_DOMAIN(domain);
2232 	if (vm_domain_allocate(vmd, req, npages)) {
2233 		/*
2234 		 * allocate them from the free page queues.
2235 		 */
2236 		vm_domain_free_lock(vmd);
2237 		m_ret = vm_phys_alloc_contig(domain, npages, low, high,
2238 		    alignment, boundary);
2239 		vm_domain_free_unlock(vmd);
2240 		if (m_ret == NULL) {
2241 			vm_domain_freecnt_inc(vmd, npages);
2242 #if VM_NRESERVLEVEL > 0
2243 			if (vm_reserv_reclaim_contig(domain, npages, low,
2244 			    high, alignment, boundary))
2245 				goto again;
2246 #endif
2247 		}
2248 	}
2249 	if (m_ret == NULL) {
2250 		if (vm_domain_alloc_fail(vmd, object, req))
2251 			goto again;
2252 		return (NULL);
2253 	}
2254 #if VM_NRESERVLEVEL > 0
2255 found:
2256 #endif
2257 	for (m = m_ret; m < &m_ret[npages]; m++) {
2258 		vm_page_dequeue(m);
2259 		vm_page_alloc_check(m);
2260 	}
2261 
2262 	/*
2263 	 * Initialize the pages.  Only the PG_ZERO flag is inherited.
2264 	 */
2265 	flags = 0;
2266 	if ((req & VM_ALLOC_ZERO) != 0)
2267 		flags = PG_ZERO;
2268 	if ((req & VM_ALLOC_NODUMP) != 0)
2269 		flags |= PG_NODUMP;
2270 	oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
2271 	    VPO_UNMANAGED : 0;
2272 	busy_lock = VPB_UNBUSIED;
2273 	if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
2274 		busy_lock = VPB_CURTHREAD_EXCLUSIVE;
2275 	if ((req & VM_ALLOC_SBUSY) != 0)
2276 		busy_lock = VPB_SHARERS_WORD(1);
2277 	if ((req & VM_ALLOC_WIRED) != 0)
2278 		vm_wire_add(npages);
2279 	if (object != NULL) {
2280 		if (object->memattr != VM_MEMATTR_DEFAULT &&
2281 		    memattr == VM_MEMATTR_DEFAULT)
2282 			memattr = object->memattr;
2283 	}
2284 	for (m = m_ret; m < &m_ret[npages]; m++) {
2285 		m->a.flags = 0;
2286 		m->flags = (m->flags | PG_NODUMP) & flags;
2287 		m->busy_lock = busy_lock;
2288 		if ((req & VM_ALLOC_WIRED) != 0)
2289 			m->ref_count = 1;
2290 		m->a.act_count = 0;
2291 		m->oflags = oflags;
2292 		if (object != NULL) {
2293 			if (vm_page_insert_after(m, object, pindex, mpred)) {
2294 				if ((req & VM_ALLOC_WIRED) != 0)
2295 					vm_wire_sub(npages);
2296 				KASSERT(m->object == NULL,
2297 				    ("page %p has object", m));
2298 				mpred = m;
2299 				for (m = m_ret; m < &m_ret[npages]; m++) {
2300 					if (m <= mpred &&
2301 					    (req & VM_ALLOC_WIRED) != 0)
2302 						m->ref_count = 0;
2303 					m->oflags = VPO_UNMANAGED;
2304 					m->busy_lock = VPB_UNBUSIED;
2305 					/* Don't change PG_ZERO. */
2306 					vm_page_free_toq(m);
2307 				}
2308 				if (req & VM_ALLOC_WAITFAIL) {
2309 					VM_OBJECT_WUNLOCK(object);
2310 					vm_radix_wait();
2311 					VM_OBJECT_WLOCK(object);
2312 				}
2313 				return (NULL);
2314 			}
2315 			mpred = m;
2316 		} else
2317 			m->pindex = pindex;
2318 		if (memattr != VM_MEMATTR_DEFAULT)
2319 			pmap_page_set_memattr(m, memattr);
2320 		pindex++;
2321 	}
2322 	return (m_ret);
2323 }
2324 
2325 /*
2326  * Check a page that has been freshly dequeued from a freelist.
2327  */
2328 static void
2329 vm_page_alloc_check(vm_page_t m)
2330 {
2331 
2332 	KASSERT(m->object == NULL, ("page %p has object", m));
2333 	KASSERT(m->a.queue == PQ_NONE &&
2334 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
2335 	    ("page %p has unexpected queue %d, flags %#x",
2336 	    m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
2337 	KASSERT(m->ref_count == 0, ("page %p has references", m));
2338 	KASSERT(!vm_page_busied(m), ("page %p is busy", m));
2339 	KASSERT(m->dirty == 0, ("page %p is dirty", m));
2340 	KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
2341 	    ("page %p has unexpected memattr %d",
2342 	    m, pmap_page_get_memattr(m)));
2343 	KASSERT(m->valid == 0, ("free page %p is valid", m));
2344 }
2345 
2346 /*
2347  * 	vm_page_alloc_freelist:
2348  *
2349  *	Allocate a physical page from the specified free page list.
2350  *
2351  *	The caller must always specify an allocation class.
2352  *
2353  *	allocation classes:
2354  *	VM_ALLOC_NORMAL		normal process request
2355  *	VM_ALLOC_SYSTEM		system *really* needs a page
2356  *	VM_ALLOC_INTERRUPT	interrupt time request
2357  *
2358  *	optional allocation flags:
2359  *	VM_ALLOC_COUNT(number)	the number of additional pages that the caller
2360  *				intends to allocate
2361  *	VM_ALLOC_WIRED		wire the allocated page
2362  *	VM_ALLOC_ZERO		prefer a zeroed page
2363  */
2364 vm_page_t
2365 vm_page_alloc_freelist(int freelist, int req)
2366 {
2367 	struct vm_domainset_iter di;
2368 	vm_page_t m;
2369 	int domain;
2370 
2371 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
2372 	do {
2373 		m = vm_page_alloc_freelist_domain(domain, freelist, req);
2374 		if (m != NULL)
2375 			break;
2376 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
2377 
2378 	return (m);
2379 }
2380 
2381 vm_page_t
2382 vm_page_alloc_freelist_domain(int domain, int freelist, int req)
2383 {
2384 	struct vm_domain *vmd;
2385 	vm_page_t m;
2386 	u_int flags;
2387 
2388 	m = NULL;
2389 	vmd = VM_DOMAIN(domain);
2390 again:
2391 	if (vm_domain_allocate(vmd, req, 1)) {
2392 		vm_domain_free_lock(vmd);
2393 		m = vm_phys_alloc_freelist_pages(domain, freelist,
2394 		    VM_FREEPOOL_DIRECT, 0);
2395 		vm_domain_free_unlock(vmd);
2396 		if (m == NULL)
2397 			vm_domain_freecnt_inc(vmd, 1);
2398 	}
2399 	if (m == NULL) {
2400 		if (vm_domain_alloc_fail(vmd, NULL, req))
2401 			goto again;
2402 		return (NULL);
2403 	}
2404 	vm_page_dequeue(m);
2405 	vm_page_alloc_check(m);
2406 
2407 	/*
2408 	 * Initialize the page.  Only the PG_ZERO flag is inherited.
2409 	 */
2410 	m->a.flags = 0;
2411 	flags = 0;
2412 	if ((req & VM_ALLOC_ZERO) != 0)
2413 		flags = PG_ZERO;
2414 	m->flags &= flags;
2415 	if ((req & VM_ALLOC_WIRED) != 0) {
2416 		vm_wire_add(1);
2417 		m->ref_count = 1;
2418 	}
2419 	/* Unmanaged pages don't use "act_count". */
2420 	m->oflags = VPO_UNMANAGED;
2421 	return (m);
2422 }
2423 
2424 static int
2425 vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
2426 {
2427 	struct vm_domain *vmd;
2428 	struct vm_pgcache *pgcache;
2429 	int i;
2430 
2431 	pgcache = arg;
2432 	vmd = VM_DOMAIN(pgcache->domain);
2433 
2434 	/*
2435 	 * The page daemon should avoid creating extra memory pressure since its
2436 	 * main purpose is to replenish the store of free pages.
2437 	 */
2438 	if (vmd->vmd_severeset || curproc == pageproc ||
2439 	    !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt))
2440 		return (0);
2441 	domain = vmd->vmd_domain;
2442 	vm_domain_free_lock(vmd);
2443 	i = vm_phys_alloc_npages(domain, pgcache->pool, cnt,
2444 	    (vm_page_t *)store);
2445 	vm_domain_free_unlock(vmd);
2446 	if (cnt != i)
2447 		vm_domain_freecnt_inc(vmd, cnt - i);
2448 
2449 	return (i);
2450 }
2451 
2452 static void
2453 vm_page_zone_release(void *arg, void **store, int cnt)
2454 {
2455 	struct vm_domain *vmd;
2456 	struct vm_pgcache *pgcache;
2457 	vm_page_t m;
2458 	int i;
2459 
2460 	pgcache = arg;
2461 	vmd = VM_DOMAIN(pgcache->domain);
2462 	vm_domain_free_lock(vmd);
2463 	for (i = 0; i < cnt; i++) {
2464 		m = (vm_page_t)store[i];
2465 		vm_phys_free_pages(m, 0);
2466 	}
2467 	vm_domain_free_unlock(vmd);
2468 	vm_domain_freecnt_inc(vmd, cnt);
2469 }
2470 
2471 #define	VPSC_ANY	0	/* No restrictions. */
2472 #define	VPSC_NORESERV	1	/* Skip reservations; implies VPSC_NOSUPER. */
2473 #define	VPSC_NOSUPER	2	/* Skip superpages. */
2474 
2475 /*
2476  *	vm_page_scan_contig:
2477  *
2478  *	Scan vm_page_array[] between the specified entries "m_start" and
2479  *	"m_end" for a run of contiguous physical pages that satisfy the
2480  *	specified conditions, and return the lowest page in the run.  The
2481  *	specified "alignment" determines the alignment of the lowest physical
2482  *	page in the run.  If the specified "boundary" is non-zero, then the
2483  *	run of physical pages cannot span a physical address that is a
2484  *	multiple of "boundary".
2485  *
2486  *	"m_end" is never dereferenced, so it need not point to a vm_page
2487  *	structure within vm_page_array[].
2488  *
2489  *	"npages" must be greater than zero.  "m_start" and "m_end" must not
2490  *	span a hole (or discontiguity) in the physical address space.  Both
2491  *	"alignment" and "boundary" must be a power of two.
2492  */
2493 vm_page_t
2494 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2495     u_long alignment, vm_paddr_t boundary, int options)
2496 {
2497 	struct mtx *m_mtx;
2498 	vm_object_t object;
2499 	vm_paddr_t pa;
2500 	vm_page_t m, m_run;
2501 #if VM_NRESERVLEVEL > 0
2502 	int level;
2503 #endif
2504 	int m_inc, order, run_ext, run_len;
2505 
2506 	KASSERT(npages > 0, ("npages is 0"));
2507 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2508 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2509 	m_run = NULL;
2510 	run_len = 0;
2511 	m_mtx = NULL;
2512 	for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2513 		KASSERT((m->flags & PG_MARKER) == 0,
2514 		    ("page %p is PG_MARKER", m));
2515 		KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1,
2516 		    ("fictitious page %p has invalid ref count", m));
2517 
2518 		/*
2519 		 * If the current page would be the start of a run, check its
2520 		 * physical address against the end, alignment, and boundary
2521 		 * conditions.  If it doesn't satisfy these conditions, either
2522 		 * terminate the scan or advance to the next page that
2523 		 * satisfies the failed condition.
2524 		 */
2525 		if (run_len == 0) {
2526 			KASSERT(m_run == NULL, ("m_run != NULL"));
2527 			if (m + npages > m_end)
2528 				break;
2529 			pa = VM_PAGE_TO_PHYS(m);
2530 			if ((pa & (alignment - 1)) != 0) {
2531 				m_inc = atop(roundup2(pa, alignment) - pa);
2532 				continue;
2533 			}
2534 			if (rounddown2(pa ^ (pa + ptoa(npages) - 1),
2535 			    boundary) != 0) {
2536 				m_inc = atop(roundup2(pa, boundary) - pa);
2537 				continue;
2538 			}
2539 		} else
2540 			KASSERT(m_run != NULL, ("m_run == NULL"));
2541 
2542 		vm_page_change_lock(m, &m_mtx);
2543 		m_inc = 1;
2544 retry:
2545 		if (vm_page_wired(m))
2546 			run_ext = 0;
2547 #if VM_NRESERVLEVEL > 0
2548 		else if ((level = vm_reserv_level(m)) >= 0 &&
2549 		    (options & VPSC_NORESERV) != 0) {
2550 			run_ext = 0;
2551 			/* Advance to the end of the reservation. */
2552 			pa = VM_PAGE_TO_PHYS(m);
2553 			m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2554 			    pa);
2555 		}
2556 #endif
2557 		else if ((object = m->object) != NULL) {
2558 			/*
2559 			 * The page is considered eligible for relocation if
2560 			 * and only if it could be laundered or reclaimed by
2561 			 * the page daemon.
2562 			 */
2563 			if (!VM_OBJECT_TRYRLOCK(object)) {
2564 				mtx_unlock(m_mtx);
2565 				VM_OBJECT_RLOCK(object);
2566 				mtx_lock(m_mtx);
2567 				if (m->object != object) {
2568 					/*
2569 					 * The page may have been freed.
2570 					 */
2571 					VM_OBJECT_RUNLOCK(object);
2572 					goto retry;
2573 				}
2574 			}
2575 			/* Don't care: PG_NODUMP, PG_ZERO. */
2576 			if (object->type != OBJT_DEFAULT &&
2577 			    object->type != OBJT_SWAP &&
2578 			    object->type != OBJT_VNODE) {
2579 				run_ext = 0;
2580 #if VM_NRESERVLEVEL > 0
2581 			} else if ((options & VPSC_NOSUPER) != 0 &&
2582 			    (level = vm_reserv_level_iffullpop(m)) >= 0) {
2583 				run_ext = 0;
2584 				/* Advance to the end of the superpage. */
2585 				pa = VM_PAGE_TO_PHYS(m);
2586 				m_inc = atop(roundup2(pa + 1,
2587 				    vm_reserv_size(level)) - pa);
2588 #endif
2589 			} else if (object->memattr == VM_MEMATTR_DEFAULT &&
2590 			    vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) &&
2591 			    !vm_page_wired(m)) {
2592 				/*
2593 				 * The page is allocated but eligible for
2594 				 * relocation.  Extend the current run by one
2595 				 * page.
2596 				 */
2597 				KASSERT(pmap_page_get_memattr(m) ==
2598 				    VM_MEMATTR_DEFAULT,
2599 				    ("page %p has an unexpected memattr", m));
2600 				KASSERT((m->oflags & (VPO_SWAPINPROG |
2601 				    VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2602 				    ("page %p has unexpected oflags", m));
2603 				/* Don't care: PGA_NOSYNC. */
2604 				run_ext = 1;
2605 			} else
2606 				run_ext = 0;
2607 			VM_OBJECT_RUNLOCK(object);
2608 #if VM_NRESERVLEVEL > 0
2609 		} else if (level >= 0) {
2610 			/*
2611 			 * The page is reserved but not yet allocated.  In
2612 			 * other words, it is still free.  Extend the current
2613 			 * run by one page.
2614 			 */
2615 			run_ext = 1;
2616 #endif
2617 		} else if ((order = m->order) < VM_NFREEORDER) {
2618 			/*
2619 			 * The page is enqueued in the physical memory
2620 			 * allocator's free page queues.  Moreover, it is the
2621 			 * first page in a power-of-two-sized run of
2622 			 * contiguous free pages.  Add these pages to the end
2623 			 * of the current run, and jump ahead.
2624 			 */
2625 			run_ext = 1 << order;
2626 			m_inc = 1 << order;
2627 		} else {
2628 			/*
2629 			 * Skip the page for one of the following reasons: (1)
2630 			 * It is enqueued in the physical memory allocator's
2631 			 * free page queues.  However, it is not the first
2632 			 * page in a run of contiguous free pages.  (This case
2633 			 * rarely occurs because the scan is performed in
2634 			 * ascending order.) (2) It is not reserved, and it is
2635 			 * transitioning from free to allocated.  (Conversely,
2636 			 * the transition from allocated to free for managed
2637 			 * pages is blocked by the page lock.) (3) It is
2638 			 * allocated but not contained by an object and not
2639 			 * wired, e.g., allocated by Xen's balloon driver.
2640 			 */
2641 			run_ext = 0;
2642 		}
2643 
2644 		/*
2645 		 * Extend or reset the current run of pages.
2646 		 */
2647 		if (run_ext > 0) {
2648 			if (run_len == 0)
2649 				m_run = m;
2650 			run_len += run_ext;
2651 		} else {
2652 			if (run_len > 0) {
2653 				m_run = NULL;
2654 				run_len = 0;
2655 			}
2656 		}
2657 	}
2658 	if (m_mtx != NULL)
2659 		mtx_unlock(m_mtx);
2660 	if (run_len >= npages)
2661 		return (m_run);
2662 	return (NULL);
2663 }
2664 
2665 /*
2666  *	vm_page_reclaim_run:
2667  *
2668  *	Try to relocate each of the allocated virtual pages within the
2669  *	specified run of physical pages to a new physical address.  Free the
2670  *	physical pages underlying the relocated virtual pages.  A virtual page
2671  *	is relocatable if and only if it could be laundered or reclaimed by
2672  *	the page daemon.  Whenever possible, a virtual page is relocated to a
2673  *	physical address above "high".
2674  *
2675  *	Returns 0 if every physical page within the run was already free or
2676  *	just freed by a successful relocation.  Otherwise, returns a non-zero
2677  *	value indicating why the last attempt to relocate a virtual page was
2678  *	unsuccessful.
2679  *
2680  *	"req_class" must be an allocation class.
2681  */
2682 static int
2683 vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run,
2684     vm_paddr_t high)
2685 {
2686 	struct vm_domain *vmd;
2687 	struct mtx *m_mtx;
2688 	struct spglist free;
2689 	vm_object_t object;
2690 	vm_paddr_t pa;
2691 	vm_page_t m, m_end, m_new;
2692 	int error, order, req;
2693 
2694 	KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2695 	    ("req_class is not an allocation class"));
2696 	SLIST_INIT(&free);
2697 	error = 0;
2698 	m = m_run;
2699 	m_end = m_run + npages;
2700 	m_mtx = NULL;
2701 	for (; error == 0 && m < m_end; m++) {
2702 		KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2703 		    ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2704 
2705 		/*
2706 		 * Avoid releasing and reacquiring the same page lock.
2707 		 */
2708 		vm_page_change_lock(m, &m_mtx);
2709 retry:
2710 		/*
2711 		 * Racily check for wirings.  Races are handled below.
2712 		 */
2713 		if (vm_page_wired(m))
2714 			error = EBUSY;
2715 		else if ((object = m->object) != NULL) {
2716 			/*
2717 			 * The page is relocated if and only if it could be
2718 			 * laundered or reclaimed by the page daemon.
2719 			 */
2720 			if (!VM_OBJECT_TRYWLOCK(object)) {
2721 				mtx_unlock(m_mtx);
2722 				VM_OBJECT_WLOCK(object);
2723 				mtx_lock(m_mtx);
2724 				if (m->object != object) {
2725 					/*
2726 					 * The page may have been freed.
2727 					 */
2728 					VM_OBJECT_WUNLOCK(object);
2729 					goto retry;
2730 				}
2731 			}
2732 			/* Don't care: PG_NODUMP, PG_ZERO. */
2733 			if (object->type != OBJT_DEFAULT &&
2734 			    object->type != OBJT_SWAP &&
2735 			    object->type != OBJT_VNODE)
2736 				error = EINVAL;
2737 			else if (object->memattr != VM_MEMATTR_DEFAULT)
2738 				error = EINVAL;
2739 			else if (vm_page_queue(m) != PQ_NONE &&
2740 			    vm_page_tryxbusy(m) != 0) {
2741 				if (vm_page_wired(m)) {
2742 					vm_page_xunbusy(m);
2743 					error = EBUSY;
2744 					goto unlock;
2745 				}
2746 				KASSERT(pmap_page_get_memattr(m) ==
2747 				    VM_MEMATTR_DEFAULT,
2748 				    ("page %p has an unexpected memattr", m));
2749 				KASSERT(m->oflags == 0,
2750 				    ("page %p has unexpected oflags", m));
2751 				/* Don't care: PGA_NOSYNC. */
2752 				if (!vm_page_none_valid(m)) {
2753 					/*
2754 					 * First, try to allocate a new page
2755 					 * that is above "high".  Failing
2756 					 * that, try to allocate a new page
2757 					 * that is below "m_run".  Allocate
2758 					 * the new page between the end of
2759 					 * "m_run" and "high" only as a last
2760 					 * resort.
2761 					 */
2762 					req = req_class | VM_ALLOC_NOOBJ;
2763 					if ((m->flags & PG_NODUMP) != 0)
2764 						req |= VM_ALLOC_NODUMP;
2765 					if (trunc_page(high) !=
2766 					    ~(vm_paddr_t)PAGE_MASK) {
2767 						m_new = vm_page_alloc_contig(
2768 						    NULL, 0, req, 1,
2769 						    round_page(high),
2770 						    ~(vm_paddr_t)0,
2771 						    PAGE_SIZE, 0,
2772 						    VM_MEMATTR_DEFAULT);
2773 					} else
2774 						m_new = NULL;
2775 					if (m_new == NULL) {
2776 						pa = VM_PAGE_TO_PHYS(m_run);
2777 						m_new = vm_page_alloc_contig(
2778 						    NULL, 0, req, 1,
2779 						    0, pa - 1, PAGE_SIZE, 0,
2780 						    VM_MEMATTR_DEFAULT);
2781 					}
2782 					if (m_new == NULL) {
2783 						pa += ptoa(npages);
2784 						m_new = vm_page_alloc_contig(
2785 						    NULL, 0, req, 1,
2786 						    pa, high, PAGE_SIZE, 0,
2787 						    VM_MEMATTR_DEFAULT);
2788 					}
2789 					if (m_new == NULL) {
2790 						vm_page_xunbusy(m);
2791 						error = ENOMEM;
2792 						goto unlock;
2793 					}
2794 
2795 					/*
2796 					 * Unmap the page and check for new
2797 					 * wirings that may have been acquired
2798 					 * through a pmap lookup.
2799 					 */
2800 					if (object->ref_count != 0 &&
2801 					    !vm_page_try_remove_all(m)) {
2802 						vm_page_xunbusy(m);
2803 						vm_page_free(m_new);
2804 						error = EBUSY;
2805 						goto unlock;
2806 					}
2807 
2808 					/*
2809 					 * Replace "m" with the new page.  For
2810 					 * vm_page_replace(), "m" must be busy
2811 					 * and dequeued.  Finally, change "m"
2812 					 * as if vm_page_free() was called.
2813 					 */
2814 					m_new->a.flags = m->a.flags &
2815 					    ~PGA_QUEUE_STATE_MASK;
2816 					KASSERT(m_new->oflags == VPO_UNMANAGED,
2817 					    ("page %p is managed", m_new));
2818 					m_new->oflags = 0;
2819 					pmap_copy_page(m, m_new);
2820 					m_new->valid = m->valid;
2821 					m_new->dirty = m->dirty;
2822 					m->flags &= ~PG_ZERO;
2823 					vm_page_dequeue(m);
2824 					if (vm_page_replace_hold(m_new, object,
2825 					    m->pindex, m) &&
2826 					    vm_page_free_prep(m))
2827 						SLIST_INSERT_HEAD(&free, m,
2828 						    plinks.s.ss);
2829 
2830 					/*
2831 					 * The new page must be deactivated
2832 					 * before the object is unlocked.
2833 					 */
2834 					vm_page_change_lock(m_new, &m_mtx);
2835 					vm_page_deactivate(m_new);
2836 				} else {
2837 					m->flags &= ~PG_ZERO;
2838 					vm_page_dequeue(m);
2839 					if (vm_page_free_prep(m))
2840 						SLIST_INSERT_HEAD(&free, m,
2841 						    plinks.s.ss);
2842 					KASSERT(m->dirty == 0,
2843 					    ("page %p is dirty", m));
2844 				}
2845 			} else
2846 				error = EBUSY;
2847 unlock:
2848 			VM_OBJECT_WUNLOCK(object);
2849 		} else {
2850 			MPASS(vm_phys_domain(m) == domain);
2851 			vmd = VM_DOMAIN(domain);
2852 			vm_domain_free_lock(vmd);
2853 			order = m->order;
2854 			if (order < VM_NFREEORDER) {
2855 				/*
2856 				 * The page is enqueued in the physical memory
2857 				 * allocator's free page queues.  Moreover, it
2858 				 * is the first page in a power-of-two-sized
2859 				 * run of contiguous free pages.  Jump ahead
2860 				 * to the last page within that run, and
2861 				 * continue from there.
2862 				 */
2863 				m += (1 << order) - 1;
2864 			}
2865 #if VM_NRESERVLEVEL > 0
2866 			else if (vm_reserv_is_page_free(m))
2867 				order = 0;
2868 #endif
2869 			vm_domain_free_unlock(vmd);
2870 			if (order == VM_NFREEORDER)
2871 				error = EINVAL;
2872 		}
2873 	}
2874 	if (m_mtx != NULL)
2875 		mtx_unlock(m_mtx);
2876 	if ((m = SLIST_FIRST(&free)) != NULL) {
2877 		int cnt;
2878 
2879 		vmd = VM_DOMAIN(domain);
2880 		cnt = 0;
2881 		vm_domain_free_lock(vmd);
2882 		do {
2883 			MPASS(vm_phys_domain(m) == domain);
2884 			SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2885 			vm_phys_free_pages(m, 0);
2886 			cnt++;
2887 		} while ((m = SLIST_FIRST(&free)) != NULL);
2888 		vm_domain_free_unlock(vmd);
2889 		vm_domain_freecnt_inc(vmd, cnt);
2890 	}
2891 	return (error);
2892 }
2893 
2894 #define	NRUNS	16
2895 
2896 CTASSERT(powerof2(NRUNS));
2897 
2898 #define	RUN_INDEX(count)	((count) & (NRUNS - 1))
2899 
2900 #define	MIN_RECLAIM	8
2901 
2902 /*
2903  *	vm_page_reclaim_contig:
2904  *
2905  *	Reclaim allocated, contiguous physical memory satisfying the specified
2906  *	conditions by relocating the virtual pages using that physical memory.
2907  *	Returns true if reclamation is successful and false otherwise.  Since
2908  *	relocation requires the allocation of physical pages, reclamation may
2909  *	fail due to a shortage of free pages.  When reclamation fails, callers
2910  *	are expected to perform vm_wait() before retrying a failed allocation
2911  *	operation, e.g., vm_page_alloc_contig().
2912  *
2913  *	The caller must always specify an allocation class through "req".
2914  *
2915  *	allocation classes:
2916  *	VM_ALLOC_NORMAL		normal process request
2917  *	VM_ALLOC_SYSTEM		system *really* needs a page
2918  *	VM_ALLOC_INTERRUPT	interrupt time request
2919  *
2920  *	The optional allocation flags are ignored.
2921  *
2922  *	"npages" must be greater than zero.  Both "alignment" and "boundary"
2923  *	must be a power of two.
2924  */
2925 bool
2926 vm_page_reclaim_contig_domain(int domain, int req, u_long npages,
2927     vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary)
2928 {
2929 	struct vm_domain *vmd;
2930 	vm_paddr_t curr_low;
2931 	vm_page_t m_run, m_runs[NRUNS];
2932 	u_long count, reclaimed;
2933 	int error, i, options, req_class;
2934 
2935 	KASSERT(npages > 0, ("npages is 0"));
2936 	KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2937 	KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2938 	req_class = req & VM_ALLOC_CLASS_MASK;
2939 
2940 	/*
2941 	 * The page daemon is allowed to dig deeper into the free page list.
2942 	 */
2943 	if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2944 		req_class = VM_ALLOC_SYSTEM;
2945 
2946 	/*
2947 	 * Return if the number of free pages cannot satisfy the requested
2948 	 * allocation.
2949 	 */
2950 	vmd = VM_DOMAIN(domain);
2951 	count = vmd->vmd_free_count;
2952 	if (count < npages + vmd->vmd_free_reserved || (count < npages +
2953 	    vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2954 	    (count < npages && req_class == VM_ALLOC_INTERRUPT))
2955 		return (false);
2956 
2957 	/*
2958 	 * Scan up to three times, relaxing the restrictions ("options") on
2959 	 * the reclamation of reservations and superpages each time.
2960 	 */
2961 	for (options = VPSC_NORESERV;;) {
2962 		/*
2963 		 * Find the highest runs that satisfy the given constraints
2964 		 * and restrictions, and record them in "m_runs".
2965 		 */
2966 		curr_low = low;
2967 		count = 0;
2968 		for (;;) {
2969 			m_run = vm_phys_scan_contig(domain, npages, curr_low,
2970 			    high, alignment, boundary, options);
2971 			if (m_run == NULL)
2972 				break;
2973 			curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2974 			m_runs[RUN_INDEX(count)] = m_run;
2975 			count++;
2976 		}
2977 
2978 		/*
2979 		 * Reclaim the highest runs in LIFO (descending) order until
2980 		 * the number of reclaimed pages, "reclaimed", is at least
2981 		 * MIN_RECLAIM.  Reset "reclaimed" each time because each
2982 		 * reclamation is idempotent, and runs will (likely) recur
2983 		 * from one scan to the next as restrictions are relaxed.
2984 		 */
2985 		reclaimed = 0;
2986 		for (i = 0; count > 0 && i < NRUNS; i++) {
2987 			count--;
2988 			m_run = m_runs[RUN_INDEX(count)];
2989 			error = vm_page_reclaim_run(req_class, domain, npages,
2990 			    m_run, high);
2991 			if (error == 0) {
2992 				reclaimed += npages;
2993 				if (reclaimed >= MIN_RECLAIM)
2994 					return (true);
2995 			}
2996 		}
2997 
2998 		/*
2999 		 * Either relax the restrictions on the next scan or return if
3000 		 * the last scan had no restrictions.
3001 		 */
3002 		if (options == VPSC_NORESERV)
3003 			options = VPSC_NOSUPER;
3004 		else if (options == VPSC_NOSUPER)
3005 			options = VPSC_ANY;
3006 		else if (options == VPSC_ANY)
3007 			return (reclaimed != 0);
3008 	}
3009 }
3010 
3011 bool
3012 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
3013     u_long alignment, vm_paddr_t boundary)
3014 {
3015 	struct vm_domainset_iter di;
3016 	int domain;
3017 	bool ret;
3018 
3019 	vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req);
3020 	do {
3021 		ret = vm_page_reclaim_contig_domain(domain, req, npages, low,
3022 		    high, alignment, boundary);
3023 		if (ret)
3024 			break;
3025 	} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
3026 
3027 	return (ret);
3028 }
3029 
3030 /*
3031  * Set the domain in the appropriate page level domainset.
3032  */
3033 void
3034 vm_domain_set(struct vm_domain *vmd)
3035 {
3036 
3037 	mtx_lock(&vm_domainset_lock);
3038 	if (!vmd->vmd_minset && vm_paging_min(vmd)) {
3039 		vmd->vmd_minset = 1;
3040 		DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains);
3041 	}
3042 	if (!vmd->vmd_severeset && vm_paging_severe(vmd)) {
3043 		vmd->vmd_severeset = 1;
3044 		DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains);
3045 	}
3046 	mtx_unlock(&vm_domainset_lock);
3047 }
3048 
3049 /*
3050  * Clear the domain from the appropriate page level domainset.
3051  */
3052 void
3053 vm_domain_clear(struct vm_domain *vmd)
3054 {
3055 
3056 	mtx_lock(&vm_domainset_lock);
3057 	if (vmd->vmd_minset && !vm_paging_min(vmd)) {
3058 		vmd->vmd_minset = 0;
3059 		DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains);
3060 		if (vm_min_waiters != 0) {
3061 			vm_min_waiters = 0;
3062 			wakeup(&vm_min_domains);
3063 		}
3064 	}
3065 	if (vmd->vmd_severeset && !vm_paging_severe(vmd)) {
3066 		vmd->vmd_severeset = 0;
3067 		DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains);
3068 		if (vm_severe_waiters != 0) {
3069 			vm_severe_waiters = 0;
3070 			wakeup(&vm_severe_domains);
3071 		}
3072 	}
3073 
3074 	/*
3075 	 * If pageout daemon needs pages, then tell it that there are
3076 	 * some free.
3077 	 */
3078 	if (vmd->vmd_pageout_pages_needed &&
3079 	    vmd->vmd_free_count >= vmd->vmd_pageout_free_min) {
3080 		wakeup(&vmd->vmd_pageout_pages_needed);
3081 		vmd->vmd_pageout_pages_needed = 0;
3082 	}
3083 
3084 	/* See comments in vm_wait_doms(). */
3085 	if (vm_pageproc_waiters) {
3086 		vm_pageproc_waiters = 0;
3087 		wakeup(&vm_pageproc_waiters);
3088 	}
3089 	mtx_unlock(&vm_domainset_lock);
3090 }
3091 
3092 /*
3093  * Wait for free pages to exceed the min threshold globally.
3094  */
3095 void
3096 vm_wait_min(void)
3097 {
3098 
3099 	mtx_lock(&vm_domainset_lock);
3100 	while (vm_page_count_min()) {
3101 		vm_min_waiters++;
3102 		msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0);
3103 	}
3104 	mtx_unlock(&vm_domainset_lock);
3105 }
3106 
3107 /*
3108  * Wait for free pages to exceed the severe threshold globally.
3109  */
3110 void
3111 vm_wait_severe(void)
3112 {
3113 
3114 	mtx_lock(&vm_domainset_lock);
3115 	while (vm_page_count_severe()) {
3116 		vm_severe_waiters++;
3117 		msleep(&vm_severe_domains, &vm_domainset_lock, PVM,
3118 		    "vmwait", 0);
3119 	}
3120 	mtx_unlock(&vm_domainset_lock);
3121 }
3122 
3123 u_int
3124 vm_wait_count(void)
3125 {
3126 
3127 	return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters);
3128 }
3129 
3130 void
3131 vm_wait_doms(const domainset_t *wdoms)
3132 {
3133 
3134 	/*
3135 	 * We use racey wakeup synchronization to avoid expensive global
3136 	 * locking for the pageproc when sleeping with a non-specific vm_wait.
3137 	 * To handle this, we only sleep for one tick in this instance.  It
3138 	 * is expected that most allocations for the pageproc will come from
3139 	 * kmem or vm_page_grab* which will use the more specific and
3140 	 * race-free vm_wait_domain().
3141 	 */
3142 	if (curproc == pageproc) {
3143 		mtx_lock(&vm_domainset_lock);
3144 		vm_pageproc_waiters++;
3145 		msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP,
3146 		    "pageprocwait", 1);
3147 	} else {
3148 		/*
3149 		 * XXX Ideally we would wait only until the allocation could
3150 		 * be satisfied.  This condition can cause new allocators to
3151 		 * consume all freed pages while old allocators wait.
3152 		 */
3153 		mtx_lock(&vm_domainset_lock);
3154 		if (vm_page_count_min_set(wdoms)) {
3155 			vm_min_waiters++;
3156 			msleep(&vm_min_domains, &vm_domainset_lock,
3157 			    PVM | PDROP, "vmwait", 0);
3158 		} else
3159 			mtx_unlock(&vm_domainset_lock);
3160 	}
3161 }
3162 
3163 /*
3164  *	vm_wait_domain:
3165  *
3166  *	Sleep until free pages are available for allocation.
3167  *	- Called in various places after failed memory allocations.
3168  */
3169 void
3170 vm_wait_domain(int domain)
3171 {
3172 	struct vm_domain *vmd;
3173 	domainset_t wdom;
3174 
3175 	vmd = VM_DOMAIN(domain);
3176 	vm_domain_free_assert_unlocked(vmd);
3177 
3178 	if (curproc == pageproc) {
3179 		mtx_lock(&vm_domainset_lock);
3180 		if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) {
3181 			vmd->vmd_pageout_pages_needed = 1;
3182 			msleep(&vmd->vmd_pageout_pages_needed,
3183 			    &vm_domainset_lock, PDROP | PSWP, "VMWait", 0);
3184 		} else
3185 			mtx_unlock(&vm_domainset_lock);
3186 	} else {
3187 		if (pageproc == NULL)
3188 			panic("vm_wait in early boot");
3189 		DOMAINSET_ZERO(&wdom);
3190 		DOMAINSET_SET(vmd->vmd_domain, &wdom);
3191 		vm_wait_doms(&wdom);
3192 	}
3193 }
3194 
3195 /*
3196  *	vm_wait:
3197  *
3198  *	Sleep until free pages are available for allocation in the
3199  *	affinity domains of the obj.  If obj is NULL, the domain set
3200  *	for the calling thread is used.
3201  *	Called in various places after failed memory allocations.
3202  */
3203 void
3204 vm_wait(vm_object_t obj)
3205 {
3206 	struct domainset *d;
3207 
3208 	d = NULL;
3209 
3210 	/*
3211 	 * Carefully fetch pointers only once: the struct domainset
3212 	 * itself is ummutable but the pointer might change.
3213 	 */
3214 	if (obj != NULL)
3215 		d = obj->domain.dr_policy;
3216 	if (d == NULL)
3217 		d = curthread->td_domain.dr_policy;
3218 
3219 	vm_wait_doms(&d->ds_mask);
3220 }
3221 
3222 /*
3223  *	vm_domain_alloc_fail:
3224  *
3225  *	Called when a page allocation function fails.  Informs the
3226  *	pagedaemon and performs the requested wait.  Requires the
3227  *	domain_free and object lock on entry.  Returns with the
3228  *	object lock held and free lock released.  Returns an error when
3229  *	retry is necessary.
3230  *
3231  */
3232 static int
3233 vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req)
3234 {
3235 
3236 	vm_domain_free_assert_unlocked(vmd);
3237 
3238 	atomic_add_int(&vmd->vmd_pageout_deficit,
3239 	    max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
3240 	if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) {
3241 		if (object != NULL)
3242 			VM_OBJECT_WUNLOCK(object);
3243 		vm_wait_domain(vmd->vmd_domain);
3244 		if (object != NULL)
3245 			VM_OBJECT_WLOCK(object);
3246 		if (req & VM_ALLOC_WAITOK)
3247 			return (EAGAIN);
3248 	}
3249 
3250 	return (0);
3251 }
3252 
3253 /*
3254  *	vm_waitpfault:
3255  *
3256  *	Sleep until free pages are available for allocation.
3257  *	- Called only in vm_fault so that processes page faulting
3258  *	  can be easily tracked.
3259  *	- Sleeps at a lower priority than vm_wait() so that vm_wait()ing
3260  *	  processes will be able to grab memory first.  Do not change
3261  *	  this balance without careful testing first.
3262  */
3263 void
3264 vm_waitpfault(struct domainset *dset, int timo)
3265 {
3266 
3267 	/*
3268 	 * XXX Ideally we would wait only until the allocation could
3269 	 * be satisfied.  This condition can cause new allocators to
3270 	 * consume all freed pages while old allocators wait.
3271 	 */
3272 	mtx_lock(&vm_domainset_lock);
3273 	if (vm_page_count_min_set(&dset->ds_mask)) {
3274 		vm_min_waiters++;
3275 		msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP,
3276 		    "pfault", timo);
3277 	} else
3278 		mtx_unlock(&vm_domainset_lock);
3279 }
3280 
3281 static struct vm_pagequeue *
3282 _vm_page_pagequeue(vm_page_t m, uint8_t queue)
3283 {
3284 
3285 	return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
3286 }
3287 
3288 #ifdef INVARIANTS
3289 static struct vm_pagequeue *
3290 vm_page_pagequeue(vm_page_t m)
3291 {
3292 
3293 	return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
3294 }
3295 #endif
3296 
3297 static __always_inline bool
3298 vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3299 {
3300 	vm_page_astate_t tmp;
3301 
3302 	tmp = *old;
3303 	do {
3304 		if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
3305 			return (true);
3306 		counter_u64_add(pqstate_commit_retries, 1);
3307 	} while (old->_bits == tmp._bits);
3308 
3309 	return (false);
3310 }
3311 
3312 /*
3313  * Do the work of committing a queue state update that moves the page out of
3314  * its current queue.
3315  */
3316 static bool
3317 _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
3318     vm_page_astate_t *old, vm_page_astate_t new)
3319 {
3320 	vm_page_t next;
3321 
3322 	vm_pagequeue_assert_locked(pq);
3323 	KASSERT(vm_page_pagequeue(m) == pq,
3324 	    ("%s: queue %p does not match page %p", __func__, pq, m));
3325 	KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
3326 	    ("%s: invalid queue indices %d %d",
3327 	    __func__, old->queue, new.queue));
3328 
3329 	/*
3330 	 * Once the queue index of the page changes there is nothing
3331 	 * synchronizing with further updates to the page's physical
3332 	 * queue state.  Therefore we must speculatively remove the page
3333 	 * from the queue now and be prepared to roll back if the queue
3334 	 * state update fails.  If the page is not physically enqueued then
3335 	 * we just update its queue index.
3336 	 */
3337 	if ((old->flags & PGA_ENQUEUED) != 0) {
3338 		new.flags &= ~PGA_ENQUEUED;
3339 		next = TAILQ_NEXT(m, plinks.q);
3340 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3341 		vm_pagequeue_cnt_dec(pq);
3342 		if (!vm_page_pqstate_fcmpset(m, old, new)) {
3343 			if (next == NULL)
3344 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3345 			else
3346 				TAILQ_INSERT_BEFORE(next, m, plinks.q);
3347 			vm_pagequeue_cnt_inc(pq);
3348 			return (false);
3349 		} else {
3350 			return (true);
3351 		}
3352 	} else {
3353 		return (vm_page_pqstate_fcmpset(m, old, new));
3354 	}
3355 }
3356 
3357 static bool
3358 vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
3359     vm_page_astate_t new)
3360 {
3361 	struct vm_pagequeue *pq;
3362 	vm_page_astate_t as;
3363 	bool ret;
3364 
3365 	pq = _vm_page_pagequeue(m, old->queue);
3366 
3367 	/*
3368 	 * The queue field and PGA_ENQUEUED flag are stable only so long as the
3369 	 * corresponding page queue lock is held.
3370 	 */
3371 	vm_pagequeue_lock(pq);
3372 	as = vm_page_astate_load(m);
3373 	if (__predict_false(as._bits != old->_bits)) {
3374 		*old = as;
3375 		ret = false;
3376 	} else {
3377 		ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
3378 	}
3379 	vm_pagequeue_unlock(pq);
3380 	return (ret);
3381 }
3382 
3383 /*
3384  * Commit a queue state update that enqueues or requeues a page.
3385  */
3386 static bool
3387 _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
3388     vm_page_astate_t *old, vm_page_astate_t new)
3389 {
3390 	struct vm_domain *vmd;
3391 
3392 	vm_pagequeue_assert_locked(pq);
3393 	KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
3394 	    ("%s: invalid queue indices %d %d",
3395 	    __func__, old->queue, new.queue));
3396 
3397 	new.flags |= PGA_ENQUEUED;
3398 	if (!vm_page_pqstate_fcmpset(m, old, new))
3399 		return (false);
3400 
3401 	if ((old->flags & PGA_ENQUEUED) != 0)
3402 		TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
3403 	else
3404 		vm_pagequeue_cnt_inc(pq);
3405 
3406 	/*
3407 	 * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE.  In particular, if
3408 	 * both flags are set in close succession, only PGA_REQUEUE_HEAD will be
3409 	 * applied, even if it was set first.
3410 	 */
3411 	if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
3412 		vmd = vm_pagequeue_domain(m);
3413 		KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
3414 		    ("%s: invalid page queue for page %p", __func__, m));
3415 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
3416 	} else {
3417 		TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3418 	}
3419 	return (true);
3420 }
3421 
3422 /*
3423  * Commit a queue state update that encodes a request for a deferred queue
3424  * operation.
3425  */
3426 static bool
3427 vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
3428     vm_page_astate_t new)
3429 {
3430 
3431 	KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
3432 	    ("%s: invalid state, queue %d flags %x",
3433 	    __func__, new.queue, new.flags));
3434 
3435 	if (old->_bits != new._bits &&
3436 	    !vm_page_pqstate_fcmpset(m, old, new))
3437 		return (false);
3438 	vm_page_pqbatch_submit(m, new.queue);
3439 	return (true);
3440 }
3441 
3442 /*
3443  * A generic queue state update function.  This handles more cases than the
3444  * specialized functions above.
3445  */
3446 bool
3447 vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
3448 {
3449 
3450 	if (old->_bits == new._bits)
3451 		return (true);
3452 
3453 	if (old->queue != PQ_NONE && new.queue != old->queue) {
3454 		if (!vm_page_pqstate_commit_dequeue(m, old, new))
3455 			return (false);
3456 		if (new.queue != PQ_NONE)
3457 			vm_page_pqbatch_submit(m, new.queue);
3458 	} else {
3459 		if (!vm_page_pqstate_fcmpset(m, old, new))
3460 			return (false);
3461 		if (new.queue != PQ_NONE &&
3462 		    ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
3463 			vm_page_pqbatch_submit(m, new.queue);
3464 	}
3465 	return (true);
3466 }
3467 
3468 /*
3469  * Apply deferred queue state updates to a page.
3470  */
3471 static inline void
3472 vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
3473 {
3474 	vm_page_astate_t new, old;
3475 
3476 	CRITICAL_ASSERT(curthread);
3477 	vm_pagequeue_assert_locked(pq);
3478 	KASSERT(queue < PQ_COUNT,
3479 	    ("%s: invalid queue index %d", __func__, queue));
3480 	KASSERT(pq == _vm_page_pagequeue(m, queue),
3481 	    ("%s: page %p does not belong to queue %p", __func__, m, pq));
3482 
3483 	for (old = vm_page_astate_load(m);;) {
3484 		if (__predict_false(old.queue != queue ||
3485 		    (old.flags & PGA_QUEUE_OP_MASK) == 0)) {
3486 			counter_u64_add(queue_nops, 1);
3487 			break;
3488 		}
3489 		KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0,
3490 		    ("%s: page %p has unexpected queue state", __func__, m));
3491 
3492 		new = old;
3493 		if ((old.flags & PGA_DEQUEUE) != 0) {
3494 			new.flags &= ~PGA_QUEUE_OP_MASK;
3495 			new.queue = PQ_NONE;
3496 			if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
3497 			    m, &old, new))) {
3498 				counter_u64_add(queue_ops, 1);
3499 				break;
3500 			}
3501 		} else {
3502 			new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
3503 			if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
3504 			    m, &old, new))) {
3505 				counter_u64_add(queue_ops, 1);
3506 				break;
3507 			}
3508 		}
3509 	}
3510 }
3511 
3512 static void
3513 vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
3514     uint8_t queue)
3515 {
3516 	int i;
3517 
3518 	for (i = 0; i < bq->bq_cnt; i++)
3519 		vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
3520 	vm_batchqueue_init(bq);
3521 }
3522 
3523 /*
3524  *	vm_page_pqbatch_submit:		[ internal use only ]
3525  *
3526  *	Enqueue a page in the specified page queue's batched work queue.
3527  *	The caller must have encoded the requested operation in the page
3528  *	structure's a.flags field.
3529  */
3530 void
3531 vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
3532 {
3533 	struct vm_batchqueue *bq;
3534 	struct vm_pagequeue *pq;
3535 	int domain;
3536 
3537 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3538 	    ("page %p is unmanaged", m));
3539 	KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
3540 
3541 	domain = vm_phys_domain(m);
3542 	pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue];
3543 
3544 	critical_enter();
3545 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3546 	if (vm_batchqueue_insert(bq, m)) {
3547 		critical_exit();
3548 		return;
3549 	}
3550 	critical_exit();
3551 	vm_pagequeue_lock(pq);
3552 	critical_enter();
3553 	bq = DPCPU_PTR(pqbatch[domain][queue]);
3554 	vm_pqbatch_process(pq, bq, queue);
3555 	vm_pqbatch_process_page(pq, m, queue);
3556 	vm_pagequeue_unlock(pq);
3557 	critical_exit();
3558 }
3559 
3560 /*
3561  *	vm_page_pqbatch_drain:		[ internal use only ]
3562  *
3563  *	Force all per-CPU page queue batch queues to be drained.  This is
3564  *	intended for use in severe memory shortages, to ensure that pages
3565  *	do not remain stuck in the batch queues.
3566  */
3567 void
3568 vm_page_pqbatch_drain(void)
3569 {
3570 	struct thread *td;
3571 	struct vm_domain *vmd;
3572 	struct vm_pagequeue *pq;
3573 	int cpu, domain, queue;
3574 
3575 	td = curthread;
3576 	CPU_FOREACH(cpu) {
3577 		thread_lock(td);
3578 		sched_bind(td, cpu);
3579 		thread_unlock(td);
3580 
3581 		for (domain = 0; domain < vm_ndomains; domain++) {
3582 			vmd = VM_DOMAIN(domain);
3583 			for (queue = 0; queue < PQ_COUNT; queue++) {
3584 				pq = &vmd->vmd_pagequeues[queue];
3585 				vm_pagequeue_lock(pq);
3586 				critical_enter();
3587 				vm_pqbatch_process(pq,
3588 				    DPCPU_PTR(pqbatch[domain][queue]), queue);
3589 				critical_exit();
3590 				vm_pagequeue_unlock(pq);
3591 			}
3592 		}
3593 	}
3594 	thread_lock(td);
3595 	sched_unbind(td);
3596 	thread_unlock(td);
3597 }
3598 
3599 /*
3600  *	vm_page_dequeue_deferred:	[ internal use only ]
3601  *
3602  *	Request removal of the given page from its current page
3603  *	queue.  Physical removal from the queue may be deferred
3604  *	indefinitely.
3605  *
3606  *	The page must be locked.
3607  */
3608 void
3609 vm_page_dequeue_deferred(vm_page_t m)
3610 {
3611 	vm_page_astate_t new, old;
3612 
3613 	old = vm_page_astate_load(m);
3614 	do {
3615 		if (old.queue == PQ_NONE) {
3616 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3617 			    ("%s: page %p has unexpected queue state",
3618 			    __func__, m));
3619 			break;
3620 		}
3621 		new = old;
3622 		new.flags |= PGA_DEQUEUE;
3623 	} while (!vm_page_pqstate_commit_request(m, &old, new));
3624 }
3625 
3626 /*
3627  *	vm_page_dequeue:
3628  *
3629  *	Remove the page from whichever page queue it's in, if any, before
3630  *	returning.
3631  */
3632 void
3633 vm_page_dequeue(vm_page_t m)
3634 {
3635 	vm_page_astate_t new, old;
3636 
3637 	old = vm_page_astate_load(m);
3638 	do {
3639 		if (old.queue == PQ_NONE) {
3640 			KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
3641 			    ("%s: page %p has unexpected queue state",
3642 			    __func__, m));
3643 			break;
3644 		}
3645 		new = old;
3646 		new.flags &= ~PGA_QUEUE_OP_MASK;
3647 		new.queue = PQ_NONE;
3648 	} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
3649 
3650 }
3651 
3652 /*
3653  * Schedule the given page for insertion into the specified page queue.
3654  * Physical insertion of the page may be deferred indefinitely.
3655  */
3656 static void
3657 vm_page_enqueue(vm_page_t m, uint8_t queue)
3658 {
3659 
3660 	KASSERT(m->a.queue == PQ_NONE &&
3661 	    (m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
3662 	    ("%s: page %p is already enqueued", __func__, m));
3663 	KASSERT(m->ref_count > 0,
3664 	    ("%s: page %p does not carry any references", __func__, m));
3665 
3666 	m->a.queue = queue;
3667 	if ((m->a.flags & PGA_REQUEUE) == 0)
3668 		vm_page_aflag_set(m, PGA_REQUEUE);
3669 	vm_page_pqbatch_submit(m, queue);
3670 }
3671 
3672 /*
3673  *	vm_page_free_prep:
3674  *
3675  *	Prepares the given page to be put on the free list,
3676  *	disassociating it from any VM object. The caller may return
3677  *	the page to the free list only if this function returns true.
3678  *
3679  *	The object must be locked.  The page must be locked if it is
3680  *	managed.
3681  */
3682 static bool
3683 vm_page_free_prep(vm_page_t m)
3684 {
3685 
3686 	/*
3687 	 * Synchronize with threads that have dropped a reference to this
3688 	 * page.
3689 	 */
3690 	atomic_thread_fence_acq();
3691 
3692 	if (vm_page_sbusied(m))
3693 		panic("vm_page_free_prep: freeing shared busy page %p", m);
3694 
3695 #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP)
3696 	if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) {
3697 		uint64_t *p;
3698 		int i;
3699 		p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m));
3700 		for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++)
3701 			KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx",
3702 			    m, i, (uintmax_t)*p));
3703 	}
3704 #endif
3705 	if ((m->oflags & VPO_UNMANAGED) == 0) {
3706 		KASSERT(!pmap_page_is_mapped(m),
3707 		    ("vm_page_free_prep: freeing mapped page %p", m));
3708 		KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
3709 		    ("vm_page_free_prep: mapping flags set in page %p", m));
3710 	} else {
3711 		KASSERT(m->a.queue == PQ_NONE,
3712 		    ("vm_page_free_prep: unmanaged page %p is queued", m));
3713 	}
3714 	VM_CNT_INC(v_tfree);
3715 
3716 	if (m->object != NULL) {
3717 		KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
3718 		    ((m->object->flags & OBJ_UNMANAGED) != 0),
3719 		    ("vm_page_free_prep: managed flag mismatch for page %p",
3720 		    m));
3721 		vm_page_object_remove(m);
3722 
3723 		/*
3724 		 * The object reference can be released without an atomic
3725 		 * operation.
3726 		 */
3727 		KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
3728 		    m->ref_count == VPRC_OBJREF,
3729 		    ("vm_page_free_prep: page %p has unexpected ref_count %u",
3730 		    m, m->ref_count));
3731 		m->object = NULL;
3732 		m->ref_count -= VPRC_OBJREF;
3733 		vm_page_xunbusy(m);
3734 	}
3735 
3736 	if (vm_page_xbusied(m))
3737 		panic("vm_page_free_prep: freeing exclusive busy page %p", m);
3738 
3739 	/*
3740 	 * If fictitious remove object association and
3741 	 * return.
3742 	 */
3743 	if ((m->flags & PG_FICTITIOUS) != 0) {
3744 		KASSERT(m->ref_count == 1,
3745 		    ("fictitious page %p is referenced", m));
3746 		KASSERT(m->a.queue == PQ_NONE,
3747 		    ("fictitious page %p is queued", m));
3748 		return (false);
3749 	}
3750 
3751 	/*
3752 	 * Pages need not be dequeued before they are returned to the physical
3753 	 * memory allocator, but they must at least be marked for a deferred
3754 	 * dequeue.
3755 	 */
3756 	if ((m->oflags & VPO_UNMANAGED) == 0)
3757 		vm_page_dequeue_deferred(m);
3758 
3759 	m->valid = 0;
3760 	vm_page_undirty(m);
3761 
3762 	if (m->ref_count != 0)
3763 		panic("vm_page_free_prep: page %p has references", m);
3764 
3765 	/*
3766 	 * Restore the default memory attribute to the page.
3767 	 */
3768 	if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3769 		pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3770 
3771 #if VM_NRESERVLEVEL > 0
3772 	/*
3773 	 * Determine whether the page belongs to a reservation.  If the page was
3774 	 * allocated from a per-CPU cache, it cannot belong to a reservation, so
3775 	 * as an optimization, we avoid the check in that case.
3776 	 */
3777 	if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m))
3778 		return (false);
3779 #endif
3780 
3781 	return (true);
3782 }
3783 
3784 /*
3785  *	vm_page_free_toq:
3786  *
3787  *	Returns the given page to the free list, disassociating it
3788  *	from any VM object.
3789  *
3790  *	The object must be locked.  The page must be locked if it is
3791  *	managed.
3792  */
3793 static void
3794 vm_page_free_toq(vm_page_t m)
3795 {
3796 	struct vm_domain *vmd;
3797 	uma_zone_t zone;
3798 
3799 	if (!vm_page_free_prep(m))
3800 		return;
3801 
3802 	vmd = vm_pagequeue_domain(m);
3803 	zone = vmd->vmd_pgcache[m->pool].zone;
3804 	if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
3805 		uma_zfree(zone, m);
3806 		return;
3807 	}
3808 	vm_domain_free_lock(vmd);
3809 	vm_phys_free_pages(m, 0);
3810 	vm_domain_free_unlock(vmd);
3811 	vm_domain_freecnt_inc(vmd, 1);
3812 }
3813 
3814 /*
3815  *	vm_page_free_pages_toq:
3816  *
3817  *	Returns a list of pages to the free list, disassociating it
3818  *	from any VM object.  In other words, this is equivalent to
3819  *	calling vm_page_free_toq() for each page of a list of VM objects.
3820  *
3821  *	The objects must be locked.  The pages must be locked if it is
3822  *	managed.
3823  */
3824 void
3825 vm_page_free_pages_toq(struct spglist *free, bool update_wire_count)
3826 {
3827 	vm_page_t m;
3828 	int count;
3829 
3830 	if (SLIST_EMPTY(free))
3831 		return;
3832 
3833 	count = 0;
3834 	while ((m = SLIST_FIRST(free)) != NULL) {
3835 		count++;
3836 		SLIST_REMOVE_HEAD(free, plinks.s.ss);
3837 		vm_page_free_toq(m);
3838 	}
3839 
3840 	if (update_wire_count)
3841 		vm_wire_sub(count);
3842 }
3843 
3844 /*
3845  * Mark this page as wired down, preventing reclamation by the page daemon
3846  * or when the containing object is destroyed.
3847  */
3848 void
3849 vm_page_wire(vm_page_t m)
3850 {
3851 	u_int old;
3852 
3853 	KASSERT(m->object != NULL,
3854 	    ("vm_page_wire: page %p does not belong to an object", m));
3855 	if (!vm_page_busied(m) && !vm_object_busied(m->object))
3856 		VM_OBJECT_ASSERT_LOCKED(m->object);
3857 	KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
3858 	    VPRC_WIRE_COUNT(m->ref_count) >= 1,
3859 	    ("vm_page_wire: fictitious page %p has zero wirings", m));
3860 
3861 	old = atomic_fetchadd_int(&m->ref_count, 1);
3862 	KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
3863 	    ("vm_page_wire: counter overflow for page %p", m));
3864 	if (VPRC_WIRE_COUNT(old) == 0) {
3865 		if ((m->oflags & VPO_UNMANAGED) == 0)
3866 			vm_page_aflag_set(m, PGA_DEQUEUE);
3867 		vm_wire_add(1);
3868 	}
3869 }
3870 
3871 /*
3872  * Attempt to wire a mapped page following a pmap lookup of that page.
3873  * This may fail if a thread is concurrently tearing down mappings of the page.
3874  * The transient failure is acceptable because it translates to the
3875  * failure of the caller pmap_extract_and_hold(), which should be then
3876  * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
3877  */
3878 bool
3879 vm_page_wire_mapped(vm_page_t m)
3880 {
3881 	u_int old;
3882 
3883 	old = m->ref_count;
3884 	do {
3885 		KASSERT(old > 0,
3886 		    ("vm_page_wire_mapped: wiring unreferenced page %p", m));
3887 		if ((old & VPRC_BLOCKED) != 0)
3888 			return (false);
3889 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
3890 
3891 	if (VPRC_WIRE_COUNT(old) == 0) {
3892 		if ((m->oflags & VPO_UNMANAGED) == 0)
3893 			vm_page_aflag_set(m, PGA_DEQUEUE);
3894 		vm_wire_add(1);
3895 	}
3896 	return (true);
3897 }
3898 
3899 /*
3900  * Release a wiring reference to a managed page.  If the page still belongs to
3901  * an object, update its position in the page queues to reflect the reference.
3902  * If the wiring was the last reference to the page, free the page.
3903  */
3904 static void
3905 vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
3906 {
3907 	u_int old;
3908 
3909 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3910 	    ("%s: page %p is unmanaged", __func__, m));
3911 
3912 	/*
3913 	 * Update LRU state before releasing the wiring reference.
3914 	 * Use a release store when updating the reference count to
3915 	 * synchronize with vm_page_free_prep().
3916 	 */
3917 	old = m->ref_count;
3918 	do {
3919 		KASSERT(VPRC_WIRE_COUNT(old) > 0,
3920 		    ("vm_page_unwire: wire count underflow for page %p", m));
3921 
3922 		if (old > VPRC_OBJREF + 1) {
3923 			/*
3924 			 * The page has at least one other wiring reference.  An
3925 			 * earlier iteration of this loop may have called
3926 			 * vm_page_release_toq() and cleared PGA_DEQUEUE, so
3927 			 * re-set it if necessary.
3928 			 */
3929 			if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
3930 				vm_page_aflag_set(m, PGA_DEQUEUE);
3931 		} else if (old == VPRC_OBJREF + 1) {
3932 			/*
3933 			 * This is the last wiring.  Clear PGA_DEQUEUE and
3934 			 * update the page's queue state to reflect the
3935 			 * reference.  If the page does not belong to an object
3936 			 * (i.e., the VPRC_OBJREF bit is clear), we only need to
3937 			 * clear leftover queue state.
3938 			 */
3939 			vm_page_release_toq(m, nqueue, false);
3940 		} else if (old == 1) {
3941 			vm_page_aflag_clear(m, PGA_DEQUEUE);
3942 		}
3943 	} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
3944 
3945 	if (VPRC_WIRE_COUNT(old) == 1) {
3946 		vm_wire_sub(1);
3947 		if (old == 1)
3948 			vm_page_free(m);
3949 	}
3950 }
3951 
3952 /*
3953  * Release one wiring of the specified page, potentially allowing it to be
3954  * paged out.
3955  *
3956  * Only managed pages belonging to an object can be paged out.  If the number
3957  * of wirings transitions to zero and the page is eligible for page out, then
3958  * the page is added to the specified paging queue.  If the released wiring
3959  * represented the last reference to the page, the page is freed.
3960  *
3961  * A managed page must be locked.
3962  */
3963 void
3964 vm_page_unwire(vm_page_t m, uint8_t nqueue)
3965 {
3966 
3967 	KASSERT(nqueue < PQ_COUNT,
3968 	    ("vm_page_unwire: invalid queue %u request for page %p",
3969 	    nqueue, m));
3970 
3971 	if ((m->oflags & VPO_UNMANAGED) != 0) {
3972 		if (vm_page_unwire_noq(m) && m->ref_count == 0)
3973 			vm_page_free(m);
3974 		return;
3975 	}
3976 	vm_page_unwire_managed(m, nqueue, false);
3977 }
3978 
3979 /*
3980  * Unwire a page without (re-)inserting it into a page queue.  It is up
3981  * to the caller to enqueue, requeue, or free the page as appropriate.
3982  * In most cases involving managed pages, vm_page_unwire() should be used
3983  * instead.
3984  */
3985 bool
3986 vm_page_unwire_noq(vm_page_t m)
3987 {
3988 	u_int old;
3989 
3990 	old = vm_page_drop(m, 1);
3991 	KASSERT(VPRC_WIRE_COUNT(old) != 0,
3992 	    ("vm_page_unref: counter underflow for page %p", m));
3993 	KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
3994 	    ("vm_page_unref: missing ref on fictitious page %p", m));
3995 
3996 	if (VPRC_WIRE_COUNT(old) > 1)
3997 		return (false);
3998 	if ((m->oflags & VPO_UNMANAGED) == 0)
3999 		vm_page_aflag_clear(m, PGA_DEQUEUE);
4000 	vm_wire_sub(1);
4001 	return (true);
4002 }
4003 
4004 /*
4005  * Ensure that the page ends up in the specified page queue.  If the page is
4006  * active or being moved to the active queue, ensure that its act_count is
4007  * at least ACT_INIT but do not otherwise mess with it.
4008  *
4009  * A managed page must be locked.
4010  */
4011 static __always_inline void
4012 vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
4013 {
4014 	vm_page_astate_t old, new;
4015 
4016 	KASSERT(m->ref_count > 0,
4017 	    ("%s: page %p does not carry any references", __func__, m));
4018 	KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
4019 	    ("%s: invalid flags %x", __func__, nflag));
4020 
4021 	if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
4022 		return;
4023 
4024 	old = vm_page_astate_load(m);
4025 	do {
4026 		if ((old.flags & PGA_DEQUEUE) != 0)
4027 			break;
4028 		new = old;
4029 		new.flags &= ~PGA_QUEUE_OP_MASK;
4030 		if (nqueue == PQ_ACTIVE)
4031 			new.act_count = max(old.act_count, ACT_INIT);
4032 		if (old.queue == nqueue) {
4033 			if (nqueue != PQ_ACTIVE)
4034 				new.flags |= nflag;
4035 		} else {
4036 			new.flags |= nflag;
4037 			new.queue = nqueue;
4038 		}
4039 	} while (!vm_page_pqstate_commit(m, &old, new));
4040 }
4041 
4042 /*
4043  * Put the specified page on the active list (if appropriate).
4044  */
4045 void
4046 vm_page_activate(vm_page_t m)
4047 {
4048 
4049 	vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
4050 }
4051 
4052 /*
4053  * Move the specified page to the tail of the inactive queue, or requeue
4054  * the page if it is already in the inactive queue.
4055  */
4056 void
4057 vm_page_deactivate(vm_page_t m)
4058 {
4059 
4060 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
4061 }
4062 
4063 void
4064 vm_page_deactivate_noreuse(vm_page_t m)
4065 {
4066 
4067 	vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
4068 }
4069 
4070 /*
4071  * Put a page in the laundry, or requeue it if it is already there.
4072  */
4073 void
4074 vm_page_launder(vm_page_t m)
4075 {
4076 
4077 	vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
4078 }
4079 
4080 /*
4081  * Put a page in the PQ_UNSWAPPABLE holding queue.
4082  */
4083 void
4084 vm_page_unswappable(vm_page_t m)
4085 {
4086 
4087 	KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0,
4088 	    ("page %p already unswappable", m));
4089 
4090 	vm_page_dequeue(m);
4091 	vm_page_enqueue(m, PQ_UNSWAPPABLE);
4092 }
4093 
4094 /*
4095  * Release a page back to the page queues in preparation for unwiring.
4096  */
4097 static void
4098 vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
4099 {
4100 	vm_page_astate_t old, new;
4101 	uint16_t nflag;
4102 
4103 	/*
4104 	 * Use a check of the valid bits to determine whether we should
4105 	 * accelerate reclamation of the page.  The object lock might not be
4106 	 * held here, in which case the check is racy.  At worst we will either
4107 	 * accelerate reclamation of a valid page and violate LRU, or
4108 	 * unnecessarily defer reclamation of an invalid page.
4109 	 *
4110 	 * If we were asked to not cache the page, place it near the head of the
4111 	 * inactive queue so that is reclaimed sooner.
4112 	 */
4113 	if (noreuse || m->valid == 0) {
4114 		nqueue = PQ_INACTIVE;
4115 		nflag = PGA_REQUEUE_HEAD;
4116 	} else {
4117 		nflag = PGA_REQUEUE;
4118 	}
4119 
4120 	old = vm_page_astate_load(m);
4121 	do {
4122 		new = old;
4123 
4124 		/*
4125 		 * If the page is already in the active queue and we are not
4126 		 * trying to accelerate reclamation, simply mark it as
4127 		 * referenced and avoid any queue operations.
4128 		 */
4129 		new.flags &= ~PGA_QUEUE_OP_MASK;
4130 		if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE)
4131 			new.flags |= PGA_REFERENCED;
4132 		else {
4133 			new.flags |= nflag;
4134 			new.queue = nqueue;
4135 		}
4136 	} while (!vm_page_pqstate_commit(m, &old, new));
4137 }
4138 
4139 /*
4140  * Unwire a page and either attempt to free it or re-add it to the page queues.
4141  */
4142 void
4143 vm_page_release(vm_page_t m, int flags)
4144 {
4145 	vm_object_t object;
4146 
4147 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4148 	    ("vm_page_release: page %p is unmanaged", m));
4149 
4150 	if ((flags & VPR_TRYFREE) != 0) {
4151 		for (;;) {
4152 			object = (vm_object_t)atomic_load_ptr(&m->object);
4153 			if (object == NULL)
4154 				break;
4155 			/* Depends on type-stability. */
4156 			if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
4157 				break;
4158 			if (object == m->object) {
4159 				vm_page_release_locked(m, flags);
4160 				VM_OBJECT_WUNLOCK(object);
4161 				return;
4162 			}
4163 			VM_OBJECT_WUNLOCK(object);
4164 		}
4165 	}
4166 	vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
4167 }
4168 
4169 /* See vm_page_release(). */
4170 void
4171 vm_page_release_locked(vm_page_t m, int flags)
4172 {
4173 
4174 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4175 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
4176 	    ("vm_page_release_locked: page %p is unmanaged", m));
4177 
4178 	if (vm_page_unwire_noq(m)) {
4179 		if ((flags & VPR_TRYFREE) != 0 &&
4180 		    (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
4181 		    m->dirty == 0 && vm_page_tryxbusy(m)) {
4182 			vm_page_free(m);
4183 		} else {
4184 			vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
4185 		}
4186 	}
4187 }
4188 
4189 static bool
4190 vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
4191 {
4192 	u_int old;
4193 
4194 	KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
4195 	    ("vm_page_try_blocked_op: page %p has no object", m));
4196 	KASSERT(vm_page_busied(m),
4197 	    ("vm_page_try_blocked_op: page %p is not busy", m));
4198 	VM_OBJECT_ASSERT_LOCKED(m->object);
4199 
4200 	old = m->ref_count;
4201 	do {
4202 		KASSERT(old != 0,
4203 		    ("vm_page_try_blocked_op: page %p has no references", m));
4204 		if (VPRC_WIRE_COUNT(old) != 0)
4205 			return (false);
4206 	} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
4207 
4208 	(op)(m);
4209 
4210 	/*
4211 	 * If the object is read-locked, new wirings may be created via an
4212 	 * object lookup.
4213 	 */
4214 	old = vm_page_drop(m, VPRC_BLOCKED);
4215 	KASSERT(!VM_OBJECT_WOWNED(m->object) ||
4216 	    old == (VPRC_BLOCKED | VPRC_OBJREF),
4217 	    ("vm_page_try_blocked_op: unexpected refcount value %u for %p",
4218 	    old, m));
4219 	return (true);
4220 }
4221 
4222 /*
4223  * Atomically check for wirings and remove all mappings of the page.
4224  */
4225 bool
4226 vm_page_try_remove_all(vm_page_t m)
4227 {
4228 
4229 	return (vm_page_try_blocked_op(m, pmap_remove_all));
4230 }
4231 
4232 /*
4233  * Atomically check for wirings and remove all writeable mappings of the page.
4234  */
4235 bool
4236 vm_page_try_remove_write(vm_page_t m)
4237 {
4238 
4239 	return (vm_page_try_blocked_op(m, pmap_remove_write));
4240 }
4241 
4242 /*
4243  * vm_page_advise
4244  *
4245  * 	Apply the specified advice to the given page.
4246  *
4247  *	The object and page must be locked.
4248  */
4249 void
4250 vm_page_advise(vm_page_t m, int advice)
4251 {
4252 
4253 	VM_OBJECT_ASSERT_WLOCKED(m->object);
4254 	if (advice == MADV_FREE)
4255 		/*
4256 		 * Mark the page clean.  This will allow the page to be freed
4257 		 * without first paging it out.  MADV_FREE pages are often
4258 		 * quickly reused by malloc(3), so we do not do anything that
4259 		 * would result in a page fault on a later access.
4260 		 */
4261 		vm_page_undirty(m);
4262 	else if (advice != MADV_DONTNEED) {
4263 		if (advice == MADV_WILLNEED)
4264 			vm_page_activate(m);
4265 		return;
4266 	}
4267 
4268 	if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
4269 		vm_page_dirty(m);
4270 
4271 	/*
4272 	 * Clear any references to the page.  Otherwise, the page daemon will
4273 	 * immediately reactivate the page.
4274 	 */
4275 	vm_page_aflag_clear(m, PGA_REFERENCED);
4276 
4277 	/*
4278 	 * Place clean pages near the head of the inactive queue rather than
4279 	 * the tail, thus defeating the queue's LRU operation and ensuring that
4280 	 * the page will be reused quickly.  Dirty pages not already in the
4281 	 * laundry are moved there.
4282 	 */
4283 	if (m->dirty == 0)
4284 		vm_page_deactivate_noreuse(m);
4285 	else if (!vm_page_in_laundry(m))
4286 		vm_page_launder(m);
4287 }
4288 
4289 static inline int
4290 vm_page_grab_pflags(int allocflags)
4291 {
4292 	int pflags;
4293 
4294 	KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
4295 	    (allocflags & VM_ALLOC_WIRED) != 0,
4296 	    ("vm_page_grab_pflags: the pages must be busied or wired"));
4297 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4298 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4299 	    ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
4300 	    "mismatch"));
4301 	pflags = allocflags &
4302 	    ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
4303 	    VM_ALLOC_NOBUSY);
4304 	if ((allocflags & VM_ALLOC_NOWAIT) == 0)
4305 		pflags |= VM_ALLOC_WAITFAIL;
4306 	if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4307 		pflags |= VM_ALLOC_SBUSY;
4308 
4309 	return (pflags);
4310 }
4311 
4312 /*
4313  * Grab a page, waiting until we are waken up due to the page
4314  * changing state.  We keep on waiting, if the page continues
4315  * to be in the object.  If the page doesn't exist, first allocate it
4316  * and then conditionally zero it.
4317  *
4318  * This routine may sleep.
4319  *
4320  * The object must be locked on entry.  The lock will, however, be released
4321  * and reacquired if the routine sleeps.
4322  */
4323 vm_page_t
4324 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
4325 {
4326 	vm_page_t m;
4327 	int pflags;
4328 
4329 	VM_OBJECT_ASSERT_WLOCKED(object);
4330 	pflags = vm_page_grab_pflags(allocflags);
4331 retrylookup:
4332 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4333 		if (!vm_page_acquire_flags(m, allocflags)) {
4334 			if (vm_page_busy_sleep_flags(object, m, "pgrbwt",
4335 			    allocflags))
4336 				goto retrylookup;
4337 			return (NULL);
4338 		}
4339 		goto out;
4340 	}
4341 	if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4342 		return (NULL);
4343 	m = vm_page_alloc(object, pindex, pflags);
4344 	if (m == NULL) {
4345 		if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4346 			return (NULL);
4347 		goto retrylookup;
4348 	}
4349 	if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
4350 		pmap_zero_page(m);
4351 
4352 out:
4353 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4354 		if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4355 			vm_page_sunbusy(m);
4356 		else
4357 			vm_page_xunbusy(m);
4358 	}
4359 	return (m);
4360 }
4361 
4362 /*
4363  * Grab a page and make it valid, paging in if necessary.  Pages missing from
4364  * their pager are zero filled and validated.  If a VM_ALLOC_COUNT is supplied
4365  * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
4366  * in simultaneously.  Additional pages will be left on a paging queue but
4367  * will neither be wired nor busy regardless of allocflags.
4368  */
4369 int
4370 vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
4371 {
4372 	vm_page_t m;
4373 	vm_page_t ma[VM_INITIAL_PAGEIN];
4374 	bool sleep, xbusy;
4375 	int after, i, pflags, rv;
4376 
4377 	KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
4378 	    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
4379 	    ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
4380 	KASSERT((allocflags &
4381 	    (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
4382 	    ("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
4383 	VM_OBJECT_ASSERT_WLOCKED(object);
4384 	pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY);
4385 	pflags |= VM_ALLOC_WAITFAIL;
4386 
4387 retrylookup:
4388 	xbusy = false;
4389 	if ((m = vm_page_lookup(object, pindex)) != NULL) {
4390 		/*
4391 		 * If the page is fully valid it can only become invalid
4392 		 * with the object lock held.  If it is not valid it can
4393 		 * become valid with the busy lock held.  Therefore, we
4394 		 * may unnecessarily lock the exclusive busy here if we
4395 		 * race with I/O completion not using the object lock.
4396 		 * However, we will not end up with an invalid page and a
4397 		 * shared lock.
4398 		 */
4399 		if (!vm_page_all_valid(m) ||
4400 		    (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) {
4401 			sleep = !vm_page_tryxbusy(m);
4402 			xbusy = true;
4403 		} else
4404 			sleep = !vm_page_trysbusy(m);
4405 		if (sleep) {
4406 			(void)vm_page_busy_sleep_flags(object, m, "pgrbwt",
4407 			    allocflags);
4408 			goto retrylookup;
4409 		}
4410 		if ((allocflags & VM_ALLOC_NOCREAT) != 0 &&
4411 		   !vm_page_all_valid(m)) {
4412 			if (xbusy)
4413 				vm_page_xunbusy(m);
4414 			else
4415 				vm_page_sunbusy(m);
4416 			*mp = NULL;
4417 			return (VM_PAGER_FAIL);
4418 		}
4419 		if ((allocflags & VM_ALLOC_WIRED) != 0)
4420 			vm_page_wire(m);
4421 		if (vm_page_all_valid(m))
4422 			goto out;
4423 	} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
4424 		*mp = NULL;
4425 		return (VM_PAGER_FAIL);
4426 	} else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) {
4427 		xbusy = true;
4428 	} else {
4429 		goto retrylookup;
4430 	}
4431 
4432 	vm_page_assert_xbusied(m);
4433 	MPASS(xbusy);
4434 	if (vm_pager_has_page(object, pindex, NULL, &after)) {
4435 		after = MIN(after, VM_INITIAL_PAGEIN);
4436 		after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
4437 		after = MAX(after, 1);
4438 		ma[0] = m;
4439 		for (i = 1; i < after; i++) {
4440 			if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
4441 				if (ma[i]->valid || !vm_page_tryxbusy(ma[i]))
4442 					break;
4443 			} else {
4444 				ma[i] = vm_page_alloc(object, m->pindex + i,
4445 				    VM_ALLOC_NORMAL);
4446 				if (ma[i] == NULL)
4447 					break;
4448 			}
4449 		}
4450 		after = i;
4451 		rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
4452 		/* Pager may have replaced a page. */
4453 		m = ma[0];
4454 		if (rv != VM_PAGER_OK) {
4455 			if ((allocflags & VM_ALLOC_WIRED) != 0)
4456 				vm_page_unwire_noq(m);
4457 			for (i = 0; i < after; i++) {
4458 				if (!vm_page_wired(ma[i]))
4459 					vm_page_free(ma[i]);
4460 				else
4461 					vm_page_xunbusy(ma[i]);
4462 			}
4463 			*mp = NULL;
4464 			return (rv);
4465 		}
4466 		for (i = 1; i < after; i++)
4467 			vm_page_readahead_finish(ma[i]);
4468 		MPASS(vm_page_all_valid(m));
4469 	} else {
4470 		vm_page_zero_invalid(m, TRUE);
4471 	}
4472 out:
4473 	if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4474 		if (xbusy)
4475 			vm_page_xunbusy(m);
4476 		else
4477 			vm_page_sunbusy(m);
4478 	}
4479 	if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy)
4480 		vm_page_busy_downgrade(m);
4481 	*mp = m;
4482 	return (VM_PAGER_OK);
4483 }
4484 
4485 /*
4486  * Return the specified range of pages from the given object.  For each
4487  * page offset within the range, if a page already exists within the object
4488  * at that offset and it is busy, then wait for it to change state.  If,
4489  * instead, the page doesn't exist, then allocate it.
4490  *
4491  * The caller must always specify an allocation class.
4492  *
4493  * allocation classes:
4494  *	VM_ALLOC_NORMAL		normal process request
4495  *	VM_ALLOC_SYSTEM		system *really* needs the pages
4496  *
4497  * The caller must always specify that the pages are to be busied and/or
4498  * wired.
4499  *
4500  * optional allocation flags:
4501  *	VM_ALLOC_IGN_SBUSY	do not sleep on soft busy pages
4502  *	VM_ALLOC_NOBUSY		do not exclusive busy the page
4503  *	VM_ALLOC_NOWAIT		do not sleep
4504  *	VM_ALLOC_SBUSY		set page to sbusy state
4505  *	VM_ALLOC_WIRED		wire the pages
4506  *	VM_ALLOC_ZERO		zero and validate any invalid pages
4507  *
4508  * If VM_ALLOC_NOWAIT is not specified, this routine may sleep.  Otherwise, it
4509  * may return a partial prefix of the requested range.
4510  */
4511 int
4512 vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags,
4513     vm_page_t *ma, int count)
4514 {
4515 	vm_page_t m, mpred;
4516 	int pflags;
4517 	int i;
4518 
4519 	VM_OBJECT_ASSERT_WLOCKED(object);
4520 	KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0,
4521 	    ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed"));
4522 
4523 	pflags = vm_page_grab_pflags(allocflags);
4524 	if (count == 0)
4525 		return (0);
4526 
4527 	i = 0;
4528 retrylookup:
4529 	m = vm_radix_lookup_le(&object->rtree, pindex + i);
4530 	if (m == NULL || m->pindex != pindex + i) {
4531 		mpred = m;
4532 		m = NULL;
4533 	} else
4534 		mpred = TAILQ_PREV(m, pglist, listq);
4535 	for (; i < count; i++) {
4536 		if (m != NULL) {
4537 			if (!vm_page_acquire_flags(m, allocflags)) {
4538 				if (vm_page_busy_sleep_flags(object, m,
4539 				    "grbmaw", allocflags))
4540 					goto retrylookup;
4541 				break;
4542 			}
4543 		} else {
4544 			if ((allocflags & VM_ALLOC_NOCREAT) != 0)
4545 				break;
4546 			m = vm_page_alloc_after(object, pindex + i,
4547 			    pflags | VM_ALLOC_COUNT(count - i), mpred);
4548 			if (m == NULL) {
4549 				if ((allocflags & VM_ALLOC_NOWAIT) != 0)
4550 					break;
4551 				goto retrylookup;
4552 			}
4553 		}
4554 		if (vm_page_none_valid(m) &&
4555 		    (allocflags & VM_ALLOC_ZERO) != 0) {
4556 			if ((m->flags & PG_ZERO) == 0)
4557 				pmap_zero_page(m);
4558 			vm_page_valid(m);
4559 		}
4560 		if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
4561 			if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
4562 				vm_page_sunbusy(m);
4563 			else
4564 				vm_page_xunbusy(m);
4565 		}
4566 		ma[i] = mpred = m;
4567 		m = vm_page_next(m);
4568 	}
4569 	return (i);
4570 }
4571 
4572 /*
4573  * Mapping function for valid or dirty bits in a page.
4574  *
4575  * Inputs are required to range within a page.
4576  */
4577 vm_page_bits_t
4578 vm_page_bits(int base, int size)
4579 {
4580 	int first_bit;
4581 	int last_bit;
4582 
4583 	KASSERT(
4584 	    base + size <= PAGE_SIZE,
4585 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
4586 	);
4587 
4588 	if (size == 0)		/* handle degenerate case */
4589 		return (0);
4590 
4591 	first_bit = base >> DEV_BSHIFT;
4592 	last_bit = (base + size - 1) >> DEV_BSHIFT;
4593 
4594 	return (((vm_page_bits_t)2 << last_bit) -
4595 	    ((vm_page_bits_t)1 << first_bit));
4596 }
4597 
4598 void
4599 vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
4600 {
4601 
4602 #if PAGE_SIZE == 32768
4603 	atomic_set_64((uint64_t *)bits, set);
4604 #elif PAGE_SIZE == 16384
4605 	atomic_set_32((uint32_t *)bits, set);
4606 #elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
4607 	atomic_set_16((uint16_t *)bits, set);
4608 #elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
4609 	atomic_set_8((uint8_t *)bits, set);
4610 #else		/* PAGE_SIZE <= 8192 */
4611 	uintptr_t addr;
4612 	int shift;
4613 
4614 	addr = (uintptr_t)bits;
4615 	/*
4616 	 * Use a trick to perform a 32-bit atomic on the
4617 	 * containing aligned word, to not depend on the existence
4618 	 * of atomic_{set, clear}_{8, 16}.
4619 	 */
4620 	shift = addr & (sizeof(uint32_t) - 1);
4621 #if BYTE_ORDER == BIG_ENDIAN
4622 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4623 #else
4624 	shift *= NBBY;
4625 #endif
4626 	addr &= ~(sizeof(uint32_t) - 1);
4627 	atomic_set_32((uint32_t *)addr, set << shift);
4628 #endif		/* PAGE_SIZE */
4629 }
4630 
4631 static inline void
4632 vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
4633 {
4634 
4635 #if PAGE_SIZE == 32768
4636 	atomic_clear_64((uint64_t *)bits, clear);
4637 #elif PAGE_SIZE == 16384
4638 	atomic_clear_32((uint32_t *)bits, clear);
4639 #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
4640 	atomic_clear_16((uint16_t *)bits, clear);
4641 #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
4642 	atomic_clear_8((uint8_t *)bits, clear);
4643 #else		/* PAGE_SIZE <= 8192 */
4644 	uintptr_t addr;
4645 	int shift;
4646 
4647 	addr = (uintptr_t)bits;
4648 	/*
4649 	 * Use a trick to perform a 32-bit atomic on the
4650 	 * containing aligned word, to not depend on the existence
4651 	 * of atomic_{set, clear}_{8, 16}.
4652 	 */
4653 	shift = addr & (sizeof(uint32_t) - 1);
4654 #if BYTE_ORDER == BIG_ENDIAN
4655 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4656 #else
4657 	shift *= NBBY;
4658 #endif
4659 	addr &= ~(sizeof(uint32_t) - 1);
4660 	atomic_clear_32((uint32_t *)addr, clear << shift);
4661 #endif		/* PAGE_SIZE */
4662 }
4663 
4664 static inline vm_page_bits_t
4665 vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
4666 {
4667 #if PAGE_SIZE == 32768
4668 	uint64_t old;
4669 
4670 	old = *bits;
4671 	while (atomic_fcmpset_64(bits, &old, newbits) == 0);
4672 	return (old);
4673 #elif PAGE_SIZE == 16384
4674 	uint32_t old;
4675 
4676 	old = *bits;
4677 	while (atomic_fcmpset_32(bits, &old, newbits) == 0);
4678 	return (old);
4679 #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
4680 	uint16_t old;
4681 
4682 	old = *bits;
4683 	while (atomic_fcmpset_16(bits, &old, newbits) == 0);
4684 	return (old);
4685 #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
4686 	uint8_t old;
4687 
4688 	old = *bits;
4689 	while (atomic_fcmpset_8(bits, &old, newbits) == 0);
4690 	return (old);
4691 #else		/* PAGE_SIZE <= 4096*/
4692 	uintptr_t addr;
4693 	uint32_t old, new, mask;
4694 	int shift;
4695 
4696 	addr = (uintptr_t)bits;
4697 	/*
4698 	 * Use a trick to perform a 32-bit atomic on the
4699 	 * containing aligned word, to not depend on the existence
4700 	 * of atomic_{set, swap, clear}_{8, 16}.
4701 	 */
4702 	shift = addr & (sizeof(uint32_t) - 1);
4703 #if BYTE_ORDER == BIG_ENDIAN
4704 	shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
4705 #else
4706 	shift *= NBBY;
4707 #endif
4708 	addr &= ~(sizeof(uint32_t) - 1);
4709 	mask = VM_PAGE_BITS_ALL << shift;
4710 
4711 	old = *bits;
4712 	do {
4713 		new = old & ~mask;
4714 		new |= newbits << shift;
4715 	} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
4716 	return (old >> shift);
4717 #endif		/* PAGE_SIZE */
4718 }
4719 
4720 /*
4721  *	vm_page_set_valid_range:
4722  *
4723  *	Sets portions of a page valid.  The arguments are expected
4724  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4725  *	of any partial chunks touched by the range.  The invalid portion of
4726  *	such chunks will be zeroed.
4727  *
4728  *	(base + size) must be less then or equal to PAGE_SIZE.
4729  */
4730 void
4731 vm_page_set_valid_range(vm_page_t m, int base, int size)
4732 {
4733 	int endoff, frag;
4734 	vm_page_bits_t pagebits;
4735 
4736 	vm_page_assert_busied(m);
4737 	if (size == 0)	/* handle degenerate case */
4738 		return;
4739 
4740 	/*
4741 	 * If the base is not DEV_BSIZE aligned and the valid
4742 	 * bit is clear, we have to zero out a portion of the
4743 	 * first block.
4744 	 */
4745 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4746 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
4747 		pmap_zero_page_area(m, frag, base - frag);
4748 
4749 	/*
4750 	 * If the ending offset is not DEV_BSIZE aligned and the
4751 	 * valid bit is clear, we have to zero out a portion of
4752 	 * the last block.
4753 	 */
4754 	endoff = base + size;
4755 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4756 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
4757 		pmap_zero_page_area(m, endoff,
4758 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4759 
4760 	/*
4761 	 * Assert that no previously invalid block that is now being validated
4762 	 * is already dirty.
4763 	 */
4764 	KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
4765 	    ("vm_page_set_valid_range: page %p is dirty", m));
4766 
4767 	/*
4768 	 * Set valid bits inclusive of any overlap.
4769 	 */
4770 	pagebits = vm_page_bits(base, size);
4771 	if (vm_page_xbusied(m))
4772 		m->valid |= pagebits;
4773 	else
4774 		vm_page_bits_set(m, &m->valid, pagebits);
4775 }
4776 
4777 /*
4778  * Set the page dirty bits and free the invalid swap space if
4779  * present.  Returns the previous dirty bits.
4780  */
4781 vm_page_bits_t
4782 vm_page_set_dirty(vm_page_t m)
4783 {
4784 	vm_page_bits_t old;
4785 
4786 	VM_PAGE_OBJECT_BUSY_ASSERT(m);
4787 
4788 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
4789 		old = m->dirty;
4790 		m->dirty = VM_PAGE_BITS_ALL;
4791 	} else
4792 		old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
4793 	if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
4794 		vm_pager_page_unswapped(m);
4795 
4796 	return (old);
4797 }
4798 
4799 /*
4800  * Clear the given bits from the specified page's dirty field.
4801  */
4802 static __inline void
4803 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
4804 {
4805 
4806 	vm_page_assert_busied(m);
4807 
4808 	/*
4809 	 * If the page is xbusied and not write mapped we are the
4810 	 * only thread that can modify dirty bits.  Otherwise, The pmap
4811 	 * layer can call vm_page_dirty() without holding a distinguished
4812 	 * lock.  The combination of page busy and atomic operations
4813 	 * suffice to guarantee consistency of the page dirty field.
4814 	 */
4815 	if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
4816 		m->dirty &= ~pagebits;
4817 	else
4818 		vm_page_bits_clear(m, &m->dirty, pagebits);
4819 }
4820 
4821 /*
4822  *	vm_page_set_validclean:
4823  *
4824  *	Sets portions of a page valid and clean.  The arguments are expected
4825  *	to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
4826  *	of any partial chunks touched by the range.  The invalid portion of
4827  *	such chunks will be zero'd.
4828  *
4829  *	(base + size) must be less then or equal to PAGE_SIZE.
4830  */
4831 void
4832 vm_page_set_validclean(vm_page_t m, int base, int size)
4833 {
4834 	vm_page_bits_t oldvalid, pagebits;
4835 	int endoff, frag;
4836 
4837 	vm_page_assert_busied(m);
4838 	if (size == 0)	/* handle degenerate case */
4839 		return;
4840 
4841 	/*
4842 	 * If the base is not DEV_BSIZE aligned and the valid
4843 	 * bit is clear, we have to zero out a portion of the
4844 	 * first block.
4845 	 */
4846 	if ((frag = rounddown2(base, DEV_BSIZE)) != base &&
4847 	    (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
4848 		pmap_zero_page_area(m, frag, base - frag);
4849 
4850 	/*
4851 	 * If the ending offset is not DEV_BSIZE aligned and the
4852 	 * valid bit is clear, we have to zero out a portion of
4853 	 * the last block.
4854 	 */
4855 	endoff = base + size;
4856 	if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff &&
4857 	    (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
4858 		pmap_zero_page_area(m, endoff,
4859 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
4860 
4861 	/*
4862 	 * Set valid, clear dirty bits.  If validating the entire
4863 	 * page we can safely clear the pmap modify bit.  We also
4864 	 * use this opportunity to clear the PGA_NOSYNC flag.  If a process
4865 	 * takes a write fault on a MAP_NOSYNC memory area the flag will
4866 	 * be set again.
4867 	 *
4868 	 * We set valid bits inclusive of any overlap, but we can only
4869 	 * clear dirty bits for DEV_BSIZE chunks that are fully within
4870 	 * the range.
4871 	 */
4872 	oldvalid = m->valid;
4873 	pagebits = vm_page_bits(base, size);
4874 	if (vm_page_xbusied(m))
4875 		m->valid |= pagebits;
4876 	else
4877 		vm_page_bits_set(m, &m->valid, pagebits);
4878 #if 0	/* NOT YET */
4879 	if ((frag = base & (DEV_BSIZE - 1)) != 0) {
4880 		frag = DEV_BSIZE - frag;
4881 		base += frag;
4882 		size -= frag;
4883 		if (size < 0)
4884 			size = 0;
4885 	}
4886 	pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
4887 #endif
4888 	if (base == 0 && size == PAGE_SIZE) {
4889 		/*
4890 		 * The page can only be modified within the pmap if it is
4891 		 * mapped, and it can only be mapped if it was previously
4892 		 * fully valid.
4893 		 */
4894 		if (oldvalid == VM_PAGE_BITS_ALL)
4895 			/*
4896 			 * Perform the pmap_clear_modify() first.  Otherwise,
4897 			 * a concurrent pmap operation, such as
4898 			 * pmap_protect(), could clear a modification in the
4899 			 * pmap and set the dirty field on the page before
4900 			 * pmap_clear_modify() had begun and after the dirty
4901 			 * field was cleared here.
4902 			 */
4903 			pmap_clear_modify(m);
4904 		m->dirty = 0;
4905 		vm_page_aflag_clear(m, PGA_NOSYNC);
4906 	} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
4907 		m->dirty &= ~pagebits;
4908 	else
4909 		vm_page_clear_dirty_mask(m, pagebits);
4910 }
4911 
4912 void
4913 vm_page_clear_dirty(vm_page_t m, int base, int size)
4914 {
4915 
4916 	vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
4917 }
4918 
4919 /*
4920  *	vm_page_set_invalid:
4921  *
4922  *	Invalidates DEV_BSIZE'd chunks within a page.  Both the
4923  *	valid and dirty bits for the effected areas are cleared.
4924  */
4925 void
4926 vm_page_set_invalid(vm_page_t m, int base, int size)
4927 {
4928 	vm_page_bits_t bits;
4929 	vm_object_t object;
4930 
4931 	/*
4932 	 * The object lock is required so that pages can't be mapped
4933 	 * read-only while we're in the process of invalidating them.
4934 	 */
4935 	object = m->object;
4936 	VM_OBJECT_ASSERT_WLOCKED(object);
4937 	vm_page_assert_busied(m);
4938 
4939 	if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
4940 	    size >= object->un_pager.vnp.vnp_size)
4941 		bits = VM_PAGE_BITS_ALL;
4942 	else
4943 		bits = vm_page_bits(base, size);
4944 	if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0)
4945 		pmap_remove_all(m);
4946 	KASSERT((bits == 0 && vm_page_all_valid(m)) ||
4947 	    !pmap_page_is_mapped(m),
4948 	    ("vm_page_set_invalid: page %p is mapped", m));
4949 	if (vm_page_xbusied(m)) {
4950 		m->valid &= ~bits;
4951 		m->dirty &= ~bits;
4952 	} else {
4953 		vm_page_bits_clear(m, &m->valid, bits);
4954 		vm_page_bits_clear(m, &m->dirty, bits);
4955 	}
4956 }
4957 
4958 /*
4959  *	vm_page_invalid:
4960  *
4961  *	Invalidates the entire page.  The page must be busy, unmapped, and
4962  *	the enclosing object must be locked.  The object locks protects
4963  *	against concurrent read-only pmap enter which is done without
4964  *	busy.
4965  */
4966 void
4967 vm_page_invalid(vm_page_t m)
4968 {
4969 
4970 	vm_page_assert_busied(m);
4971 	VM_OBJECT_ASSERT_LOCKED(m->object);
4972 	MPASS(!pmap_page_is_mapped(m));
4973 
4974 	if (vm_page_xbusied(m))
4975 		m->valid = 0;
4976 	else
4977 		vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
4978 }
4979 
4980 /*
4981  * vm_page_zero_invalid()
4982  *
4983  *	The kernel assumes that the invalid portions of a page contain
4984  *	garbage, but such pages can be mapped into memory by user code.
4985  *	When this occurs, we must zero out the non-valid portions of the
4986  *	page so user code sees what it expects.
4987  *
4988  *	Pages are most often semi-valid when the end of a file is mapped
4989  *	into memory and the file's size is not page aligned.
4990  */
4991 void
4992 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
4993 {
4994 	int b;
4995 	int i;
4996 
4997 	/*
4998 	 * Scan the valid bits looking for invalid sections that
4999 	 * must be zeroed.  Invalid sub-DEV_BSIZE'd areas ( where the
5000 	 * valid bit may be set ) have already been zeroed by
5001 	 * vm_page_set_validclean().
5002 	 */
5003 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
5004 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
5005 		    (m->valid & ((vm_page_bits_t)1 << i))) {
5006 			if (i > b) {
5007 				pmap_zero_page_area(m,
5008 				    b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
5009 			}
5010 			b = i + 1;
5011 		}
5012 	}
5013 
5014 	/*
5015 	 * setvalid is TRUE when we can safely set the zero'd areas
5016 	 * as being valid.  We can do this if there are no cache consistancy
5017 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
5018 	 */
5019 	if (setvalid)
5020 		vm_page_valid(m);
5021 }
5022 
5023 /*
5024  *	vm_page_is_valid:
5025  *
5026  *	Is (partial) page valid?  Note that the case where size == 0
5027  *	will return FALSE in the degenerate case where the page is
5028  *	entirely invalid, and TRUE otherwise.
5029  *
5030  *	Some callers envoke this routine without the busy lock held and
5031  *	handle races via higher level locks.  Typical callers should
5032  *	hold a busy lock to prevent invalidation.
5033  */
5034 int
5035 vm_page_is_valid(vm_page_t m, int base, int size)
5036 {
5037 	vm_page_bits_t bits;
5038 
5039 	bits = vm_page_bits(base, size);
5040 	return (m->valid != 0 && (m->valid & bits) == bits);
5041 }
5042 
5043 /*
5044  * Returns true if all of the specified predicates are true for the entire
5045  * (super)page and false otherwise.
5046  */
5047 bool
5048 vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m)
5049 {
5050 	vm_object_t object;
5051 	int i, npages;
5052 
5053 	object = m->object;
5054 	if (skip_m != NULL && skip_m->object != object)
5055 		return (false);
5056 	VM_OBJECT_ASSERT_LOCKED(object);
5057 	npages = atop(pagesizes[m->psind]);
5058 
5059 	/*
5060 	 * The physically contiguous pages that make up a superpage, i.e., a
5061 	 * page with a page size index ("psind") greater than zero, will
5062 	 * occupy adjacent entries in vm_page_array[].
5063 	 */
5064 	for (i = 0; i < npages; i++) {
5065 		/* Always test object consistency, including "skip_m". */
5066 		if (m[i].object != object)
5067 			return (false);
5068 		if (&m[i] == skip_m)
5069 			continue;
5070 		if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i]))
5071 			return (false);
5072 		if ((flags & PS_ALL_DIRTY) != 0) {
5073 			/*
5074 			 * Calling vm_page_test_dirty() or pmap_is_modified()
5075 			 * might stop this case from spuriously returning
5076 			 * "false".  However, that would require a write lock
5077 			 * on the object containing "m[i]".
5078 			 */
5079 			if (m[i].dirty != VM_PAGE_BITS_ALL)
5080 				return (false);
5081 		}
5082 		if ((flags & PS_ALL_VALID) != 0 &&
5083 		    m[i].valid != VM_PAGE_BITS_ALL)
5084 			return (false);
5085 	}
5086 	return (true);
5087 }
5088 
5089 /*
5090  * Set the page's dirty bits if the page is modified.
5091  */
5092 void
5093 vm_page_test_dirty(vm_page_t m)
5094 {
5095 
5096 	vm_page_assert_busied(m);
5097 	if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
5098 		vm_page_dirty(m);
5099 }
5100 
5101 void
5102 vm_page_valid(vm_page_t m)
5103 {
5104 
5105 	vm_page_assert_busied(m);
5106 	if (vm_page_xbusied(m))
5107 		m->valid = VM_PAGE_BITS_ALL;
5108 	else
5109 		vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
5110 }
5111 
5112 void
5113 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
5114 {
5115 
5116 	mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
5117 }
5118 
5119 void
5120 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
5121 {
5122 
5123 	mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
5124 }
5125 
5126 int
5127 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
5128 {
5129 
5130 	return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
5131 }
5132 
5133 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
5134 void
5135 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
5136 {
5137 
5138 	vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
5139 }
5140 
5141 void
5142 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
5143 {
5144 
5145 	mtx_assert_(vm_page_lockptr(m), a, file, line);
5146 }
5147 #endif
5148 
5149 #ifdef INVARIANTS
5150 void
5151 vm_page_object_busy_assert(vm_page_t m)
5152 {
5153 
5154 	/*
5155 	 * Certain of the page's fields may only be modified by the
5156 	 * holder of a page or object busy.
5157 	 */
5158 	if (m->object != NULL && !vm_page_busied(m))
5159 		VM_OBJECT_ASSERT_BUSY(m->object);
5160 }
5161 
5162 void
5163 vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits)
5164 {
5165 
5166 	if ((bits & PGA_WRITEABLE) == 0)
5167 		return;
5168 
5169 	/*
5170 	 * The PGA_WRITEABLE flag can only be set if the page is
5171 	 * managed, is exclusively busied or the object is locked.
5172 	 * Currently, this flag is only set by pmap_enter().
5173 	 */
5174 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
5175 	    ("PGA_WRITEABLE on unmanaged page"));
5176 	if (!vm_page_xbusied(m))
5177 		VM_OBJECT_ASSERT_BUSY(m->object);
5178 }
5179 #endif
5180 
5181 #include "opt_ddb.h"
5182 #ifdef DDB
5183 #include <sys/kernel.h>
5184 
5185 #include <ddb/ddb.h>
5186 
5187 DB_SHOW_COMMAND(page, vm_page_print_page_info)
5188 {
5189 
5190 	db_printf("vm_cnt.v_free_count: %d\n", vm_free_count());
5191 	db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count());
5192 	db_printf("vm_cnt.v_active_count: %d\n", vm_active_count());
5193 	db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count());
5194 	db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count());
5195 	db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
5196 	db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
5197 	db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
5198 	db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
5199 }
5200 
5201 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
5202 {
5203 	int dom;
5204 
5205 	db_printf("pq_free %d\n", vm_free_count());
5206 	for (dom = 0; dom < vm_ndomains; dom++) {
5207 		db_printf(
5208     "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n",
5209 		    dom,
5210 		    vm_dom[dom].vmd_page_count,
5211 		    vm_dom[dom].vmd_free_count,
5212 		    vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
5213 		    vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
5214 		    vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt,
5215 		    vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt);
5216 	}
5217 }
5218 
5219 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
5220 {
5221 	vm_page_t m;
5222 	boolean_t phys, virt;
5223 
5224 	if (!have_addr) {
5225 		db_printf("show pginfo addr\n");
5226 		return;
5227 	}
5228 
5229 	phys = strchr(modif, 'p') != NULL;
5230 	virt = strchr(modif, 'v') != NULL;
5231 	if (virt)
5232 		m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
5233 	else if (phys)
5234 		m = PHYS_TO_VM_PAGE(addr);
5235 	else
5236 		m = (vm_page_t)addr;
5237 	db_printf(
5238     "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n"
5239     "  af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
5240 	    m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
5241 	    m->a.queue, m->ref_count, m->a.flags, m->oflags,
5242 	    m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
5243 }
5244 #endif /* DDB */
5245